Update Neon™ depthwise kernel

- Reduce duplication and simplify overall structure.
- Improve multi-threaded performance by sharing more data
  in lower-level caches.

Partially Resolves: COMPMID-5054
Signed-off-by: Ramy Elgammal <ramy.elgammal@arm.com>
Change-Id: Iac747f39b21c540122fa75218762631c4d787911
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/7449
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Andrew Mundy
Reviewed-by: Sheri Zhang <sheri.zhang@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>

1 files changed, 857 insertions, 883 deletions

diff --git a/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8q_nhwc_3x3_s2_output2x2_mla_depthfirst/generic.cpp b/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8q_nhwc_3x3_s2_output2x2_mla_depthfirst/generic.cpp
index 8d22836e5b..872f665b40 100644
--- a/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8q_nhwc_3x3_s2_output2x2_mla_depthfirst/generic.cpp
+++ b/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8q_nhwc_3x3_s2_output2x2_mla_depthfirst/generic.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2021 Arm Limited.
+ * Copyright (c) 2021-2022 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -46,7 +46,7 @@ void a64_u8q_nhwc_3x3_s2_output2x2_mla_depthfirst_impl(
   struct Params
   {
     long unsigned int n_channels;
-    const uint8_t *weights;
+    const void *weights;
     const int32_t *bias;
     const arm_gemm::Requantize32 *requant;
     const int32_t *const requant_muls;
@@ -57,7 +57,7 @@ void a64_u8q_nhwc_3x3_s2_output2x2_mla_depthfirst_impl(
     Params(
       long unsigned int n_channels,
       const uint8_t *const *inptrs_raw,
-      const uint8_t *const weights,
+      const void *const weights,
       const int32_t *const bias,
       const arm_gemm::Requantize32 &qp,
       const int32_t *const requant_muls,
@@ -100,75 +100,75 @@ void a64_u8q_nhwc_3x3_s2_output2x2_mla_depthfirst_impl(
                       requant_muls, requant_shifts, outptrs);
 
   __asm__ __volatile__(
-    "ldr x4, [%x[params], %[offsetof_Params_n_channels]]\n"
-    "mov x5, #0x0\n"
-    "ldr x6, [%x[params], %[offsetof_Params_weights]]\n"
-    "mov x7, #0x0\n"
-    "ldr x22, [%x[params], %[offsetof_Params_requant]]\n"
-    "add x8, %x[params], %[offsetof_Params_inptrs]\n"
-    "ldr x17, [%x[params], %[offsetof_Params_requant_muls]]\n"
-    "lsr x16, x4, #0x3\n"
-    "ldr x15, [%x[params], %[offsetof_Params_requant_shifts]]\n"
-    "add x19, x22, %[offsetof_Requantize32_a_offset]\n"
-    "ldr x21, [%x[params], %[offsetof_Params_outptrs]]\n"
-    "add x20, x22, %[offsetof_Requantize32_b_offset]\n"
-    "ld1r { v12.16b }, [x19]\n"
-    "add x19, x22, %[offsetof_Requantize32_c_offset]\n"
-    "ld1r { v13.16b }, [x20]\n"
-    "add x20, x22, %[offsetof_Requantize32_minval]\n"
-    "ld1r { v11.4s }, [x19]\n"
-    "add x19, x22, %[offsetof_Requantize32_maxval]\n"
-    "ld1r { v19.4s }, [x20]\n"
-    "ld1r { v14.4s }, [x19]\n"
-    "ldp x14, x13, [x21, #0x0]\n"
-    "ldp x12, x11, [x21, #0x10]\n"
+    "ldr x19, [%x[params], %[offsetof_Params_requant]]\n"
+    "ldr x8, [%x[params], %[offsetof_Params_n_channels]]\n"
+    "add x24, x19, %[offsetof_Requantize32_a_offset]\n"
+    "add x23, x19, %[offsetof_Requantize32_b_offset]\n"
+    "ldr x22, [%x[params], %[offsetof_Params_outptrs]]\n"
+    "add x21, x19, %[offsetof_Requantize32_c_offset]\n"
+    "add x20, x19, %[offsetof_Requantize32_minval]\n"
+    "ldr x17, [%x[params], %[offsetof_Params_weights]]\n"
+    "add x19, x19, %[offsetof_Requantize32_maxval]\n"
+    "ld1r { v12.16b }, [x24]\n"
+    "ld1r { v13.16b }, [x23]\n"
+    "lsr x16, x8, #0x3\n"
+    "ld1r { v11.8h }, [x21]\n"
+    "ld1r { v17.8h }, [x20]\n"
+    "mov x15, #0x0\n"
+    "mov x14, #0x0\n"
+    "ld1r { v14.8h }, [x19]\n"
+    "ldr x13, [%x[params], %[offsetof_Params_requant_muls]]\n"
+    "add x12, %x[params], %[offsetof_Params_inptrs]\n"
+    "ldr x11, [%x[params], %[offsetof_Params_requant_shifts]]\n"
+    "ldp x10, x9, [x22, #0x0]\n"
+    "ldp x28, x27, [x22, #0x10]\n"
     "cbz x16, 3f\n"
-    "subs x16, x16, #0x1\n"
     "ldr x19, [%x[params], %[offsetof_Params_bias]]\n"
     "ldr q15, [x19, #0x0]\n"
-    "mov v20.16b, v15.16b\n"
+    "subs x16, x16, #0x1\n"
+    "mov v9.16b, v15.16b\n"
     "ldr q10, [x19, #0x10]\n"
     "add x19, x19, #0x20\n"
-    "mov v16.16b, v15.16b\n"
     "str x19, [%x[params], %[offsetof_Params_bias]]\n"
-    "mov v17.16b, v15.16b\n"
-    "ldr d0, [x6, #0x0]\n"
+    "ldr d0, [x17, #0x0]\n"
+    "ldr d1, [x17, #0x8]\n"
+    "ldr d2, [x17, #0x10]\n"
+    "mov v16.16b, v10.16b\n"
+    "mov v22.16b, v15.16b\n"
+    "ldr d3, [x17, #0x18]\n"
+    "ldr d4, [x17, #0x20]\n"
+    "mov v21.16b, v10.16b\n"
+    "mov v23.16b, v15.16b\n"
+    "ldr d5, [x17, #0x28]\n"
+    "ldr d6, [x17, #0x30]\n"
+    "mov v18.16b, v10.16b\n"
     "usubl v0.8h, v0.8b, v13.8b\n"
-    "mov v23.16b, v10.16b\n"
-    "ldr d1, [x6, #0x8]\n"
-    "mov v22.16b, v10.16b\n"
-    "ldr d2, [x6, #0x10]\n"
+    "ldr d7, [x17, #0x38]\n"
+    "ldr d8, [x17, #0x40]\n"
     "usubl v1.8h, v1.8b, v13.8b\n"
-    "mov v18.16b, v10.16b\n"
-    "ldr d3, [x6, #0x18]\n"
-    "ldr d4, [x6, #0x20]\n"
     "usubl v2.8h, v2.8b, v13.8b\n"
-    "ldr d5, [x6, #0x28]\n"
+    "ldp x26, x25, [x12, #0x0]\n"
+    "ldp x24, x23, [x12, #0x10]\n"
     "usubl v3.8h, v3.8b, v13.8b\n"
-    "ldr d6, [x6, #0x30]\n"
-    "ldr d7, [x6, #0x38]\n"
     "usubl v4.8h, v4.8b, v13.8b\n"
-    "ldr d8, [x6, #0x40]\n"
+    "ldp x22, x21, [x12, #0x20]\n"
+    "ldp x20, x19, [x12, #0x30]\n"
     "usubl v5.8h, v5.8b, v13.8b\n"
-    "ldp x26, x25, [x8, #0x0]\n"
     "usubl v6.8h, v6.8b, v13.8b\n"
-    "ldp x24, x23, [x8, #0x10]\n"
+    "ldr d31, [x26, x15]\n"
+    "ldr d30, [x25, x15]\n"
     "usubl v7.8h, v7.8b, v13.8b\n"
     "usubl v8.8h, v8.8b, v13.8b\n"
-    "ldp x22, x21, [x8, #0x20]\n"
-    "ldp x20, x19, [x8, #0x30]\n"
-    "ldr d31, [x26, x5]\n"
+    "ldr d29, [x24, x15]\n"
+    "ldr d28, [x23, x15]\n"
     "usubl v31.8h, v31.8b, v12.8b\n"
-    "ldr d30, [x25, x5]\n"
-    "ldr d29, [x24, x5]\n"
     "usubl v30.8h, v30.8b, v12.8b\n"
-    "ldr d28, [x23, x5]\n"
-    "ldr d27, [x22, x5]\n"
+    "ldr d27, [x22, x15]\n"
+    "ldr d26, [x21, x15]\n"
     "usubl v29.8h, v29.8b, v12.8b\n"
-    "ldr d26, [x21, x5]\n"
     "usubl v28.8h, v28.8b, v12.8b\n"
-    "ldr d25, [x20, x5]\n"
-    "ldr d24, [x19, x5]\n"
+    "ldr d25, [x20, x15]\n"
+    "ldr d24, [x19, x15]\n"
     "usubl v27.8h, v27.8b, v12.8b\n"
     "usubl v26.8h, v26.8b, v12.8b\n"
     "usubl v25.8h, v25.8b, v12.8b\n"
@@ -176,259 +176,251 @@ void a64_u8q_nhwc_3x3_s2_output2x2_mla_depthfirst_impl(
     "beq 2f\n"
     "1:"  // Loop
     "smlal v15.4s, v31.4h, v8.4h\n"
-    "ldr x23, [x8, #0x40]\n"
-    "add x6, x6, #0x48\n"
     "smlal2 v10.4s, v31.8h, v8.8h\n"
-    "ldr x22, [x8, #0x48]\n"
-    "subs x16, x16, #0x1\n"
-    "smlal v20.4s, v31.4h, v6.4h\n"
-    "ldr x21, [x8, #0x50]\n"
-    "smlal2 v23.4s, v31.8h, v6.8h\n"
-    "ldr x20, [x8, #0x58]\n"
-    "smlal v16.4s, v31.4h, v2.4h\n"
-    "ldr x19, [x8, #0x60]\n"
-    "smlal2 v22.4s, v31.8h, v2.8h\n"
-    "ldr x10, [x8, #0x68]\n"
-    "smlal v17.4s, v31.4h, v0.4h\n"
-    "ldr x9, [x8, #0x70]\n"
-    "smlal2 v18.4s, v31.8h, v0.8h\n"
-    "ldr x28, [x8, #0x78]\n"
+    "ldr x24, [x12, #0x40]\n"
+    "ldr x23, [x12, #0x48]\n"
+    "smlal v9.4s, v31.4h, v6.4h\n"
+    "smlal2 v16.4s, v31.8h, v6.8h\n"
+    "ldr x21, [x12, #0x50]\n"
+    "ldr x19, [x12, #0x58]\n"
     "smlal v15.4s, v30.4h, v0.4h\n"
-    "ldr x27, [x8, #0x80]\n"
     "smlal2 v10.4s, v30.8h, v0.8h\n"
-    "ldr x26, [x8, #0x88]\n"
-    "smlal v20.4s, v28.4h, v1.4h\n"
-    "ldr x25, [x8, #0x90]\n"
-    "smlal2 v23.4s, v28.8h, v1.8h\n"
-    "ldr d28, [x22, x5]\n"
+    "ldr x22, [x12, #0x78]\n"
+    "ldr x20, [x12, #0x60]\n"
+    "smlal v9.4s, v28.4h, v1.4h\n"
+    "smlal2 v16.4s, v28.8h, v1.8h\n"
+    "ldr d28, [x23, x15]\n"
     "usubl v28.8h, v28.8b, v12.8b\n"
     "smlal v15.4s, v29.4h, v1.4h\n"
-    "ldr x24, [x8, #0x98]\n"
     "smlal2 v10.4s, v29.8h, v1.8h\n"
-    "ldr d29, [x23, x5]\n"
+    "ldr d29, [x24, x15]\n"
     "usubl v29.8h, v29.8b, v12.8b\n"
-    "smlal v20.4s, v27.4h, v2.4h\n"
-    "ldr x23, [x8, #0xa0]\n"
-    "smlal2 v23.4s, v27.8h, v2.8h\n"
-    "ldr d27, [x21, x5]\n"
+    "smlal v9.4s, v27.4h, v2.4h\n"
+    "smlal2 v16.4s, v27.8h, v2.8h\n"
+    "ldr d27, [x21, x15]\n"
     "usubl v27.8h, v27.8b, v12.8b\n"
     "smlal v15.4s, v26.4h, v3.4h\n"
-    "ldr x22, [x8, #0xa8]\n"
     "smlal2 v10.4s, v26.8h, v3.8h\n"
-    "ldr d26, [x20, x5]\n"
+    "ldr d26, [x19, x15]\n"
     "usubl v26.8h, v26.8b, v12.8b\n"
+    "smlal v9.4s, v24.4h, v0.4h\n"
+    "smlal2 v16.4s, v24.8h, v0.8h\n"
+    "ldr x21, [x12, #0x80]\n"
+    "ldr x19, [x12, #0x68]\n"
     "smlal v15.4s, v25.4h, v4.4h\n"
-    "ldr x21, [x8, #0xb0]\n"
     "smlal2 v10.4s, v25.8h, v4.8h\n"
-    "ldr d25, [x19, x5]\n"
+    "ldr d25, [x20, x15]\n"
     "usubl v25.8h, v25.8b, v12.8b\n"
+    "smlal v9.4s, v29.4h, v4.4h\n"
+    "smlal2 v16.4s, v29.8h, v4.8h\n"
+    "ldr x20, [x12, #0x88]\n"
+    "ldr d29, [x19, x15]\n"
     "smlal v15.4s, v24.4h, v2.4h\n"
-    "ldr x20, [x8, #0xb8]\n"
     "smlal2 v10.4s, v24.8h, v2.8h\n"
-    "ldr x19, [x8, #0xc0]\n"
-    "smlal v20.4s, v24.4h, v0.4h\n"
-    "ldr q21, [x17, #0x0]\n"
-    "smlal2 v23.4s, v24.8h, v0.8h\n"
-    "ldr d24, [x9, x5]\n"
-    "usubl v24.8h, v24.8b, v12.8b\n"
-    "smlal v20.4s, v29.4h, v4.4h\n"
-    "ldr q30, [x15, #0x0]\n"
-    "smlal2 v23.4s, v29.8h, v4.8h\n"
-    "ldr d29, [x10, x5]\n"
+    "ldr x19, [x12, #0x70]\n"
     "usubl v29.8h, v29.8b, v12.8b\n"
-    "smlal v20.4s, v28.4h, v5.4h\n"
-    "ldr q31, [x17, #0x10]\n"
-    "smlal2 v23.4s, v28.8h, v5.8h\n"
-    "ldr d28, [x27, x5]\n"
-    "add x17, x17, #0x20\n"
+    "smlal v9.4s, v28.4h, v5.4h\n"
+    "smlal2 v16.4s, v28.8h, v5.8h\n"
+    "ldr d28, [x21, x15]\n"
+    "usubl v28.8h, v28.8b, v12.8b\n"
+    "smlal v22.4s, v31.4h, v2.4h\n"
+    "smlal2 v21.4s, v31.8h, v2.8h\n"
+    "ldr x24, [x12, #0x98]\n"
+    "ldr d24, [x19, x15]\n"
     "smlal v15.4s, v27.4h, v5.4h\n"
-    "ldr q9, [x15, #0x10]\n"
-    "add x15, x15, #0x20\n"
     "smlal2 v10.4s, v27.8h, v5.8h\n"
-    "usubl v28.8h, v28.8b, v12.8b\n"
-    "smlal v20.4s, v27.4h, v3.4h\n"
-    "smlal2 v23.4s, v27.8h, v3.8h\n"
-    "ldr d27, [x28, x5]\n"
+    "usubl v24.8h, v24.8b, v12.8b\n"
+    "ldr x23, [x12, #0x90]\n"
+    "smlal v9.4s, v27.4h, v3.4h\n"
+    "smlal2 v16.4s, v27.8h, v3.8h\n"
+    "ldr d27, [x22, x15]\n"
     "usubl v27.8h, v27.8b, v12.8b\n"
-    "smlal v16.4s, v26.4h, v3.4h\n"
-    "smlal2 v22.4s, v26.8h, v3.8h\n"
-    "ldr d26, [x26, x5]\n"
+    "smlal v23.4s, v31.4h, v0.4h\n"
+    "smlal v22.4s, v26.4h, v3.4h\n"
+    "ldr x22, [x12, #0xa8]\n"
+    "ldr x19, [x12, #0xa0]\n"
+    "smlal2 v21.4s, v26.8h, v3.8h\n"
+    "smlal2 v18.4s, v31.8h, v0.8h\n"
+    "ldr d26, [x20, x15]\n"
     "usubl v26.8h, v26.8b, v12.8b\n"
+    "smlal v23.4s, v27.4h, v4.4h\n"
+    "smlal v22.4s, v25.4h, v0.4h\n"
+    "ldr x21, [x12, #0xb0]\n"
+    "ldr x20, [x12, #0xb8]\n"
+    "smlal2 v21.4s, v25.8h, v0.8h\n"
+    "smlal2 v18.4s, v27.8h, v4.8h\n"
+    "ldr d27, [x19, x15]\n"
+    "usubl v27.8h, v27.8b, v12.8b\n"
+    "smlal v23.4s, v28.4h, v1.4h\n"
     "smlal v15.4s, v25.4h, v6.4h\n"
+    "ldr x19, [x12, #0xc0]\n"
+    "ldr q19, [x13, #0x0]\n"
     "smlal2 v10.4s, v25.8h, v6.8h\n"
-    "smlal v16.4s, v25.4h, v0.4h\n"
-    "smlal2 v22.4s, v25.8h, v0.8h\n"
-    "ldr d25, [x25, x5]\n"
+    "smlal v22.4s, v29.4h, v4.4h\n"
+    "ldr d25, [x23, x15]\n"
     "usubl v25.8h, v25.8b, v12.8b\n"
-    "smlal v16.4s, v29.4h, v4.4h\n"
-    "smlal2 v22.4s, v29.8h, v4.8h\n"
-    "ldr d29, [x24, x5]\n"
+    "smlal2 v21.4s, v29.8h, v4.8h\n"
+    "ldr d29, [x24, x15]\n"
+    "smlal2 v18.4s, v28.8h, v1.8h\n"
     "usubl v29.8h, v29.8b, v12.8b\n"
+    "smlal v23.4s, v26.4h, v5.4h\n"
     "smlal v15.4s, v24.4h, v7.4h\n"
+    "ldr q0, [x11, #0x0]\n"
+    "ldr q4, [x13, #0x10]\n"
     "smlal2 v10.4s, v24.8h, v7.8h\n"
-    "smlal v16.4s, v24.4h, v1.4h\n"
-    "smlal2 v22.4s, v24.8h, v1.8h\n"
-    "ldr d24, [x22, x5]\n"
-    "usubl v24.8h, v24.8b, v12.8b\n"
-    "smlal v17.4s, v27.4h, v4.4h\n"
-    "smlal2 v18.4s, v27.8h, v4.8h\n"
-    "ldr d27, [x23, x5]\n"
-    "usubl v27.8h, v27.8b, v12.8b\n"
-    "smlal v20.4s, v28.4h, v7.4h\n"
-    "smlal2 v23.4s, v28.8h, v7.8h\n"
-    "smlal v17.4s, v28.4h, v1.4h\n"
-    "smlal2 v18.4s, v28.8h, v1.8h\n"
-    "smlal v16.4s, v25.4h, v6.4h\n"
-    "smlal2 v22.4s, v25.8h, v6.8h\n"
-    "ldr d25, [x20, x5]\n"
-    "usubl v25.8h, v25.8b, v12.8b\n"
-    "smlal v17.4s, v26.4h, v5.4h\n"
+    "smlal v22.4s, v24.4h, v1.4h\n"
+    "sqdmulh v15.4s, v15.4s, v19.4s\n"
+    "ldr q31, [x11, #0x10]\n"
+    "smlal2 v21.4s, v24.8h, v1.8h\n"
+    "ldr d24, [x22, x15]\n"
     "smlal2 v18.4s, v26.8h, v5.8h\n"
-    "ldr d26, [x21, x5]\n"
-    "usubl v26.8h, v26.8b, v12.8b\n"
-    "smlal v20.4s, v29.4h, v8.4h\n"
-    "smlal2 v23.4s, v29.8h, v8.8h\n"
-    "smlal v17.4s, v29.4h, v2.4h\n"
+    "usubl v24.8h, v24.8b, v12.8b\n"
+    "smlal v23.4s, v29.4h, v2.4h\n"
+    "ldr d26, [x21, x15]\n"
     "smlal2 v18.4s, v29.8h, v2.8h\n"
-    "ldr d29, [x19, x5]\n"
-    "add x5, x5, #0x8\n"
-    "smlal v16.4s, v27.4h, v7.4h\n"
-    "usubl v29.8h, v29.8b, v12.8b\n"
-    "smlal2 v22.4s, v27.8h, v7.8h\n"
-    "smlal v17.4s, v24.4h, v3.4h\n"
-    "smlal v16.4s, v24.4h, v5.4h\n"
+    "usubl v26.8h, v26.8b, v12.8b\n"
+    "smlal v22.4s, v25.4h, v6.4h\n"
+    "smlal v23.4s, v24.4h, v3.4h\n"
+    "and v30.16b, v15.16b, v0.16b\n"
+    "add x17, x17, #0x48\n"
+    "smlal v9.4s, v28.4h, v7.4h\n"
+    "smlal2 v16.4s, v28.8h, v7.8h\n"
+    "sqdmulh v10.4s, v10.4s, v4.4s\n"
+    "subs x16, x16, #0x1\n"
+    "smlal2 v21.4s, v25.8h, v6.8h\n"
+    "ldr d25, [x20, x15]\n"
     "smlal2 v18.4s, v24.8h, v3.8h\n"
-    "sqrdmulh v15.4s, v15.4s, v21.4s\n"
-    "smlal2 v22.4s, v24.8h, v5.8h\n"
-    "smlal v17.4s, v26.4h, v7.4h\n"
+    "usubl v25.8h, v25.8b, v12.8b\n"
+    "smlal v22.4s, v27.4h, v7.4h\n"
+    "smlal v23.4s, v26.4h, v7.4h\n"
+    "sshr v30.4s, v30.4s, #0x1f\n"
+    "add x13, x13, #0x20\n"
+    "smlal v9.4s, v29.4h, v8.4h\n"
+    "smlal2 v16.4s, v29.8h, v8.8h\n"
+    "ldr d29, [x19, x15]\n"
+    "usubl v29.8h, v29.8b, v12.8b\n"
+    "smlal2 v21.4s, v27.8h, v7.8h\n"
     "smlal2 v18.4s, v26.8h, v7.8h\n"
-    "smlal v16.4s, v25.4h, v8.4h\n"
-    "smlal2 v22.4s, v25.8h, v8.8h\n"
-    "smlal v17.4s, v25.4h, v6.4h\n"
+    "sqdmulh v9.4s, v9.4s, v19.4s\n"
+    "add x15, x15, #0x8\n"
+    "smlal v22.4s, v24.4h, v5.4h\n"
+    "smlal v23.4s, v25.4h, v6.4h\n"
+    "and v28.16b, v9.16b, v0.16b\n"
+    "add x11, x11, #0x20\n"
+    "smlal2 v21.4s, v24.8h, v5.8h\n"
     "smlal2 v18.4s, v25.8h, v6.8h\n"
-    "and v26.16b, v15.16b, v30.16b\n"
-    "sshr v26.4s, v26.4s, #0x1f\n"
-    "smlal v17.4s, v29.4h, v8.4h\n"
+    "sqdmulh v16.4s, v16.4s, v4.4s\n"
+    "smlal v22.4s, v25.4h, v8.4h\n"
+    "smlal v23.4s, v29.4h, v8.4h\n"
+    "sqdmulh v22.4s, v22.4s, v19.4s\n"
+    "smlal2 v21.4s, v25.8h, v8.8h\n"
     "smlal2 v18.4s, v29.8h, v8.8h\n"
-    "sqrdmulh v10.4s, v10.4s, v31.4s\n"
-    "sqrdmulh v20.4s, v20.4s, v21.4s\n"
-    "sqrdmulh v23.4s, v23.4s, v31.4s\n"
-    "sqrdmulh v16.4s, v16.4s, v21.4s\n"
-    "sqadd v15.4s, v15.4s, v26.4s\n"
-    "and v8.16b, v10.16b, v9.16b\n"
-    "sshr v8.4s, v8.4s, #0x1f\n"
-    "srshl v15.4s, v15.4s, v30.4s\n"
-    "and v4.16b, v20.16b, v30.16b\n"
+    "sqdmulh v23.4s, v23.4s, v19.4s\n"
+    "and v29.16b, v22.16b, v0.16b\n"
+    "sqdmulh v21.4s, v21.4s, v4.4s\n"
+    "and v20.16b, v23.16b, v0.16b\n"
+    "sqdmulh v18.4s, v18.4s, v4.4s\n"
+    "and v19.16b, v10.16b, v31.16b\n"
+    "sshr v28.4s, v28.4s, #0x1f\n"
+    "and v4.16b, v16.16b, v31.16b\n"
+    "sshr v29.4s, v29.4s, #0x1f\n"
+    "and v5.16b, v21.16b, v31.16b\n"
+    "sshr v20.4s, v20.4s, #0x1f\n"
+    "and v26.16b, v18.16b, v31.16b\n"
+    "sqadd v15.4s, v15.4s, v30.4s\n"
+    "sshr v19.4s, v19.4s, #0x1f\n"
+    "sqadd v9.4s, v9.4s, v28.4s\n"
     "sshr v4.4s, v4.4s, #0x1f\n"
-    "and v2.16b, v23.16b, v9.16b\n"
-    "and v1.16b, v16.16b, v30.16b\n"
-    "sshr v2.4s, v2.4s, #0x1f\n"
-    "add v15.4s, v15.4s, v11.4s\n"
-    "sqadd v10.4s, v10.4s, v8.4s\n"
-    "sshr v1.4s, v1.4s, #0x1f\n"
-    "sqrdmulh v22.4s, v22.4s, v31.4s\n"
-    "sqadd v20.4s, v20.4s, v4.4s\n"
-    "smin v15.4s, v15.4s, v14.4s\n"
-    "srshl v10.4s, v10.4s, v9.4s\n"
-    "sqadd v23.4s, v23.4s, v2.4s\n"
-    "smax v15.4s, v15.4s, v19.4s\n"
-    "srshl v20.4s, v20.4s, v30.4s\n"
-    "add v10.4s, v10.4s, v11.4s\n"
-    "srshl v23.4s, v23.4s, v9.4s\n"
-    "sqadd v16.4s, v16.4s, v1.4s\n"
-    "smin v10.4s, v10.4s, v14.4s\n"
-    "add v20.4s, v20.4s, v11.4s\n"
-    "add v23.4s, v23.4s, v11.4s\n"
-    "smax v10.4s, v10.4s, v19.4s\n"
-    "smin v20.4s, v20.4s, v14.4s\n"
-    "smin v23.4s, v23.4s, v14.4s\n"
-    "uzp1 v15.16b, v15.16b, v10.16b\n"
-    "smax v20.4s, v20.4s, v19.4s\n"
+    "sqadd v22.4s, v22.4s, v29.4s\n"
+    "sshr v5.4s, v5.4s, #0x1f\n"
+    "sqadd v23.4s, v23.4s, v20.4s\n"
+    "sshr v26.4s, v26.4s, #0x1f\n"
+    "srshl v15.4s, v15.4s, v0.4s\n"
+    "sqadd v10.4s, v10.4s, v19.4s\n"
+    "srshl v9.4s, v9.4s, v0.4s\n"
+    "sqadd v16.4s, v16.4s, v4.4s\n"
+    "srshl v22.4s, v22.4s, v0.4s\n"
+    "sqadd v21.4s, v21.4s, v5.4s\n"
+    "srshl v23.4s, v23.4s, v0.4s\n"
+    "sqadd v18.4s, v18.4s, v26.4s\n"
+    "srshl v10.4s, v10.4s, v31.4s\n"
+    "sqxtn v15.4h, v15.4s\n"
+    "srshl v16.4s, v16.4s, v31.4s\n"
+    "sqxtn v9.4h, v9.4s\n"
+    "srshl v21.4s, v21.4s, v31.4s\n"
+    "sqxtn v22.4h, v22.4s\n"
+    "srshl v18.4s, v18.4s, v31.4s\n"
+    "sqxtn v23.4h, v23.4s\n"
+    "sqxtn2 v15.8h, v10.4s\n"
+    "sqxtn2 v9.8h, v16.4s\n"
+    "sqxtn2 v22.8h, v21.4s\n"
+    "sqxtn2 v23.8h, v18.4s\n"
+    "sqadd v15.8h, v15.8h, v11.8h\n"
+    "sqadd v9.8h, v9.8h, v11.8h\n"
+    "sqadd v22.8h, v22.8h, v11.8h\n"
+    "sqadd v23.8h, v23.8h, v11.8h\n"
+    "smax v15.8h, v15.8h, v17.8h\n"
+    "smax v9.8h, v9.8h, v17.8h\n"
+    "smax v22.8h, v22.8h, v17.8h\n"
+    "smax v23.8h, v23.8h, v17.8h\n"
+    "smin v15.8h, v15.8h, v14.8h\n"
+    "smin v9.8h, v9.8h, v14.8h\n"
+    "smin v22.8h, v22.8h, v14.8h\n"
+    "smin v23.8h, v23.8h, v14.8h\n"
     "uzp1 v15.16b, v15.16b, v15.16b\n"
-    "str d15, [x14, x7]\n"
-    "smax v23.4s, v23.4s, v19.4s\n"
-    "srshl v16.4s, v16.4s, v30.4s\n"
-    "and v24.16b, v22.16b, v9.16b\n"
-    "sshr v24.4s, v24.4s, #0x1f\n"
-    "uzp1 v20.16b, v20.16b, v23.16b\n"
-    "add v16.4s, v16.4s, v11.4s\n"
-    "sqrdmulh v17.4s, v17.4s, v21.4s\n"
-    "uzp1 v20.16b, v20.16b, v20.16b\n"
-    "str d20, [x13, x7]\n"
-    "smin v16.4s, v16.4s, v14.4s\n"
-    "sqrdmulh v18.4s, v18.4s, v31.4s\n"
-    "sqadd v22.4s, v22.4s, v24.4s\n"
-    "and v2.16b, v17.16b, v30.16b\n"
-    "sshr v2.4s, v2.4s, #0x1f\n"
-    "smax v16.4s, v16.4s, v19.4s\n"
-    "srshl v22.4s, v22.4s, v9.4s\n"
-    "and v31.16b, v18.16b, v9.16b\n"
-    "sshr v31.4s, v31.4s, #0x1f\n"
-    "add v22.4s, v22.4s, v11.4s\n"
-    "sqadd v17.4s, v17.4s, v2.4s\n"
-    "smin v22.4s, v22.4s, v14.4s\n"
-    "srshl v17.4s, v17.4s, v30.4s\n"
-    "sqadd v18.4s, v18.4s, v31.4s\n"
-    "smax v22.4s, v22.4s, v19.4s\n"
-    "uzp1 v16.16b, v16.16b, v22.16b\n"
-    "add v17.4s, v17.4s, v11.4s\n"
-    "srshl v18.4s, v18.4s, v9.4s\n"
-    "uzp1 v16.16b, v16.16b, v16.16b\n"
-    "str d16, [x12, x7]\n"
-    "smin v17.4s, v17.4s, v14.4s\n"
-    "add v18.4s, v18.4s, v11.4s\n"
-    "smax v17.4s, v17.4s, v19.4s\n"
-    "smin v18.4s, v18.4s, v14.4s\n"
-    "smax v18.4s, v18.4s, v19.4s\n"
-    "uzp1 v17.16b, v17.16b, v18.16b\n"
-    "uzp1 v17.16b, v17.16b, v17.16b\n"
-    "str d17, [x11, x7]\n"
-    "add x7, x7, #0x8\n"
+    "str d15, [x10, x14]\n"
+    "uzp1 v9.16b, v9.16b, v9.16b\n"
+    "uzp1 v22.16b, v22.16b, v22.16b\n"
+    "str d9, [x9, x14]\n"
+    "uzp1 v23.16b, v23.16b, v23.16b\n"
+    "str d22, [x28, x14]\n"
+    "str d23, [x27, x14]\n"
     "ldr x19, [%x[params], %[offsetof_Params_bias]]\n"
     "ldr q15, [x19, #0x0]\n"
-    "mov v20.16b, v15.16b\n"
+    "add x14, x14, #0x8\n"
     "ldr q10, [x19, #0x10]\n"
     "add x19, x19, #0x20\n"
-    "mov v16.16b, v15.16b\n"
     "str x19, [%x[params], %[offsetof_Params_bias]]\n"
-    "mov v17.16b, v15.16b\n"
-    "ldr d0, [x6, #0x0]\n"
+    "ldr d0, [x17, #0x0]\n"
+    "ldr d1, [x17, #0x8]\n"
+    "ldr d2, [x17, #0x10]\n"
+    "mov v9.16b, v15.16b\n"
+    "mov v16.16b, v10.16b\n"
+    "ldr d3, [x17, #0x18]\n"
+    "ldr d4, [x17, #0x20]\n"
+    "mov v22.16b, v15.16b\n"
+    "mov v21.16b, v10.16b\n"
+    "ldr d5, [x17, #0x28]\n"
+    "ldr d6, [x17, #0x30]\n"
+    "mov v23.16b, v15.16b\n"
+    "mov v18.16b, v10.16b\n"
+    "ldr d7, [x17, #0x38]\n"
+    "ldr d8, [x17, #0x40]\n"
     "usubl v0.8h, v0.8b, v13.8b\n"
-    "mov v23.16b, v10.16b\n"
-    "ldr d1, [x6, #0x8]\n"
-    "mov v22.16b, v10.16b\n"
-    "ldr d2, [x6, #0x10]\n"
     "usubl v1.8h, v1.8b, v13.8b\n"
-    "mov v18.16b, v10.16b\n"
-    "ldr d3, [x6, #0x18]\n"
-    "ldr d4, [x6, #0x20]\n"
+    "ldp x26, x25, [x12, #0x0]\n"
+    "ldp x24, x23, [x12, #0x10]\n"
     "usubl v2.8h, v2.8b, v13.8b\n"
-    "ldr d5, [x6, #0x28]\n"
     "usubl v3.8h, v3.8b, v13.8b\n"
-    "ldr d6, [x6, #0x30]\n"
-    "ldr d7, [x6, #0x38]\n"
+    "ldp x22, x21, [x12, #0x20]\n"
+    "ldp x20, x19, [x12, #0x30]\n"
     "usubl v4.8h, v4.8b, v13.8b\n"
-    "ldr d8, [x6, #0x40]\n"
     "usubl v5.8h, v5.8b, v13.8b\n"
-    "ldp x26, x25, [x8, #0x0]\n"
+    "ldr d31, [x26, x15]\n"
+    "ldr d30, [x25, x15]\n"
     "usubl v6.8h, v6.8b, v13.8b\n"
-    "ldp x24, x23, [x8, #0x10]\n"
     "usubl v7.8h, v7.8b, v13.8b\n"
+    "ldr d29, [x24, x15]\n"
+    "ldr d28, [x23, x15]\n"
     "usubl v8.8h, v8.8b, v13.8b\n"
-    "ldp x22, x21, [x8, #0x20]\n"
-    "ldp x20, x19, [x8, #0x30]\n"
-    "ldr d31, [x26, x5]\n"
     "usubl v31.8h, v31.8b, v12.8b\n"
-    "ldr d30, [x25, x5]\n"
-    "ldr d29, [x24, x5]\n"
+    "ldr d27, [x22, x15]\n"
+    "ldr d26, [x21, x15]\n"
     "usubl v30.8h, v30.8b, v12.8b\n"
-    "ldr d28, [x23, x5]\n"
-    "ldr d27, [x22, x5]\n"
     "usubl v29.8h, v29.8b, v12.8b\n"
-    "ldr d26, [x21, x5]\n"
+    "ldr d25, [x20, x15]\n"
+    "ldr d24, [x19, x15]\n"
     "usubl v28.8h, v28.8b, v12.8b\n"
-    "ldr d25, [x20, x5]\n"
-    "ldr d24, [x19, x5]\n"
     "usubl v27.8h, v27.8b, v12.8b\n"
     "usubl v26.8h, v26.8b, v12.8b\n"
     "usubl v25.8h, v25.8b, v12.8b\n"
@@ -436,275 +428,267 @@ void a64_u8q_nhwc_3x3_s2_output2x2_mla_depthfirst_impl(
     "bgt 1b\n"
     "2:"  // Tail
     "smlal v15.4s, v31.4h, v8.4h\n"
-    "ldr x23, [x8, #0x40]\n"
-    "tst x4, #0x7\n"
     "smlal2 v10.4s, v31.8h, v8.8h\n"
-    "ldr x22, [x8, #0x48]\n"
-    "smlal v20.4s, v31.4h, v6.4h\n"
-    "ldr x21, [x8, #0x50]\n"
-    "smlal2 v23.4s, v31.8h, v6.8h\n"
-    "ldr x20, [x8, #0x58]\n"
-    "smlal v16.4s, v31.4h, v2.4h\n"
-    "ldr x19, [x8, #0x60]\n"
-    "smlal2 v22.4s, v31.8h, v2.8h\n"
-    "ldr x10, [x8, #0x68]\n"
-    "smlal v17.4s, v31.4h, v0.4h\n"
-    "ldr x9, [x8, #0x70]\n"
-    "smlal2 v18.4s, v31.8h, v0.8h\n"
-    "ldr x28, [x8, #0x78]\n"
+    "ldr x24, [x12, #0x40]\n"
+    "ldr x23, [x12, #0x48]\n"
+    "smlal v9.4s, v31.4h, v6.4h\n"
+    "smlal2 v16.4s, v31.8h, v6.8h\n"
+    "ldr x21, [x12, #0x50]\n"
+    "ldr x19, [x12, #0x58]\n"
     "smlal v15.4s, v30.4h, v0.4h\n"
-    "ldr x27, [x8, #0x80]\n"
     "smlal2 v10.4s, v30.8h, v0.8h\n"
-    "ldr x26, [x8, #0x88]\n"
-    "smlal v20.4s, v28.4h, v1.4h\n"
-    "ldr x25, [x8, #0x90]\n"
-    "smlal2 v23.4s, v28.8h, v1.8h\n"
-    "ldr d28, [x22, x5]\n"
+    "ldr x22, [x12, #0x78]\n"
+    "ldr x20, [x12, #0x60]\n"
+    "smlal v9.4s, v28.4h, v1.4h\n"
+    "smlal2 v16.4s, v28.8h, v1.8h\n"
+    "ldr d28, [x23, x15]\n"
     "usubl v28.8h, v28.8b, v12.8b\n"
     "smlal v15.4s, v29.4h, v1.4h\n"
-    "ldr x24, [x8, #0x98]\n"
     "smlal2 v10.4s, v29.8h, v1.8h\n"
-    "ldr d29, [x23, x5]\n"
+    "ldr d29, [x24, x15]\n"
     "usubl v29.8h, v29.8b, v12.8b\n"
-    "smlal v20.4s, v27.4h, v2.4h\n"
-    "ldr x23, [x8, #0xa0]\n"
-    "smlal2 v23.4s, v27.8h, v2.8h\n"
-    "ldr d27, [x21, x5]\n"
+    "smlal v9.4s, v27.4h, v2.4h\n"
+    "smlal2 v16.4s, v27.8h, v2.8h\n"
+    "ldr d27, [x21, x15]\n"
     "usubl v27.8h, v27.8b, v12.8b\n"
     "smlal v15.4s, v26.4h, v3.4h\n"
-    "ldr x22, [x8, #0xa8]\n"
     "smlal2 v10.4s, v26.8h, v3.8h\n"
-    "ldr d26, [x20, x5]\n"
+    "ldr d26, [x19, x15]\n"
     "usubl v26.8h, v26.8b, v12.8b\n"
+    "smlal v9.4s, v24.4h, v0.4h\n"
+    "smlal2 v16.4s, v24.8h, v0.8h\n"
+    "ldr x21, [x12, #0x80]\n"
+    "ldr x19, [x12, #0x68]\n"
     "smlal v15.4s, v25.4h, v4.4h\n"
-    "ldr x21, [x8, #0xb0]\n"
     "smlal2 v10.4s, v25.8h, v4.8h\n"
-    "ldr d25, [x19, x5]\n"
+    "ldr d25, [x20, x15]\n"
     "usubl v25.8h, v25.8b, v12.8b\n"
+    "smlal v9.4s, v29.4h, v4.4h\n"
+    "smlal2 v16.4s, v29.8h, v4.8h\n"
+    "ldr x20, [x12, #0x88]\n"
+    "ldr d29, [x19, x15]\n"
     "smlal v15.4s, v24.4h, v2.4h\n"
-    "ldr x20, [x8, #0xb8]\n"
     "smlal2 v10.4s, v24.8h, v2.8h\n"
-    "ldr x19, [x8, #0xc0]\n"
-    "smlal v20.4s, v24.4h, v0.4h\n"
-    "ldr q21, [x17, #0x0]\n"
-    "smlal2 v23.4s, v24.8h, v0.8h\n"
-    "ldr d24, [x9, x5]\n"
-    "usubl v24.8h, v24.8b, v12.8b\n"
-    "smlal v20.4s, v29.4h, v4.4h\n"
-    "ldr q30, [x15, #0x0]\n"
-    "smlal2 v23.4s, v29.8h, v4.8h\n"
-    "ldr d29, [x10, x5]\n"
+    "ldr x19, [x12, #0x70]\n"
     "usubl v29.8h, v29.8b, v12.8b\n"
-    "smlal v20.4s, v28.4h, v5.4h\n"
-    "ldr q31, [x17, #0x10]\n"
-    "smlal2 v23.4s, v28.8h, v5.8h\n"
-    "ldr d28, [x27, x5]\n"
-    "add x17, x17, #0x20\n"
+    "smlal v9.4s, v28.4h, v5.4h\n"
+    "smlal2 v16.4s, v28.8h, v5.8h\n"
+    "ldr d28, [x21, x15]\n"
+    "usubl v28.8h, v28.8b, v12.8b\n"
+    "smlal v22.4s, v31.4h, v2.4h\n"
+    "smlal2 v21.4s, v31.8h, v2.8h\n"
+    "ldr x24, [x12, #0x98]\n"
+    "ldr d24, [x19, x15]\n"
     "smlal v15.4s, v27.4h, v5.4h\n"
-    "ldr q9, [x15, #0x10]\n"
-    "add x15, x15, #0x20\n"
     "smlal2 v10.4s, v27.8h, v5.8h\n"
-    "usubl v28.8h, v28.8b, v12.8b\n"
-    "smlal v20.4s, v27.4h, v3.4h\n"
-    "smlal2 v23.4s, v27.8h, v3.8h\n"
-    "ldr d27, [x28, x5]\n"
+    "usubl v24.8h, v24.8b, v12.8b\n"
+    "ldr x23, [x12, #0x90]\n"
+    "smlal v9.4s, v27.4h, v3.4h\n"
+    "smlal2 v16.4s, v27.8h, v3.8h\n"
+    "ldr d27, [x22, x15]\n"
     "usubl v27.8h, v27.8b, v12.8b\n"
-    "smlal v16.4s, v26.4h, v3.4h\n"
-    "smlal2 v22.4s, v26.8h, v3.8h\n"
-    "ldr d26, [x26, x5]\n"
+    "smlal v23.4s, v31.4h, v0.4h\n"
+    "smlal v22.4s, v26.4h, v3.4h\n"
+    "ldr x22, [x12, #0xa8]\n"
+    "ldr x19, [x12, #0xa0]\n"
+    "smlal2 v21.4s, v26.8h, v3.8h\n"
+    "smlal2 v18.4s, v31.8h, v0.8h\n"
+    "ldr d26, [x20, x15]\n"
     "usubl v26.8h, v26.8b, v12.8b\n"
+    "smlal v23.4s, v27.4h, v4.4h\n"
+    "smlal v22.4s, v25.4h, v0.4h\n"
+    "ldr x21, [x12, #0xb0]\n"
+    "ldr x20, [x12, #0xb8]\n"
+    "smlal2 v21.4s, v25.8h, v0.8h\n"
+    "smlal2 v18.4s, v27.8h, v4.8h\n"
+    "ldr d27, [x19, x15]\n"
+    "usubl v27.8h, v27.8b, v12.8b\n"
+    "smlal v23.4s, v28.4h, v1.4h\n"
     "smlal v15.4s, v25.4h, v6.4h\n"
+    "ldr x19, [x12, #0xc0]\n"
+    "ldr q19, [x13, #0x0]\n"
     "smlal2 v10.4s, v25.8h, v6.8h\n"
-    "smlal v16.4s, v25.4h, v0.4h\n"
-    "smlal2 v22.4s, v25.8h, v0.8h\n"
-    "ldr d25, [x25, x5]\n"
+    "smlal v22.4s, v29.4h, v4.4h\n"
+    "ldr d25, [x23, x15]\n"
     "usubl v25.8h, v25.8b, v12.8b\n"
-    "smlal v16.4s, v29.4h, v4.4h\n"
-    "smlal2 v22.4s, v29.8h, v4.8h\n"
-    "ldr d29, [x24, x5]\n"
+    "smlal2 v21.4s, v29.8h, v4.8h\n"
+    "ldr d29, [x24, x15]\n"
+    "smlal2 v18.4s, v28.8h, v1.8h\n"
     "usubl v29.8h, v29.8b, v12.8b\n"
+    "smlal v23.4s, v26.4h, v5.4h\n"
     "smlal v15.4s, v24.4h, v7.4h\n"
+    "ldr q0, [x11, #0x0]\n"
+    "ldr q4, [x13, #0x10]\n"
     "smlal2 v10.4s, v24.8h, v7.8h\n"
-    "smlal v16.4s, v24.4h, v1.4h\n"
-    "smlal2 v22.4s, v24.8h, v1.8h\n"
-    "ldr d24, [x22, x5]\n"
-    "usubl v24.8h, v24.8b, v12.8b\n"
-    "smlal v17.4s, v27.4h, v4.4h\n"
-    "smlal2 v18.4s, v27.8h, v4.8h\n"
-    "ldr d27, [x23, x5]\n"
-    "usubl v27.8h, v27.8b, v12.8b\n"
-    "smlal v20.4s, v28.4h, v7.4h\n"
-    "smlal2 v23.4s, v28.8h, v7.8h\n"
-    "smlal v17.4s, v28.4h, v1.4h\n"
-    "smlal2 v18.4s, v28.8h, v1.8h\n"
-    "smlal v16.4s, v25.4h, v6.4h\n"
-    "smlal2 v22.4s, v25.8h, v6.8h\n"
-    "ldr d25, [x20, x5]\n"
-    "usubl v25.8h, v25.8b, v12.8b\n"
-    "smlal v17.4s, v26.4h, v5.4h\n"
+    "smlal v22.4s, v24.4h, v1.4h\n"
+    "sqdmulh v15.4s, v15.4s, v19.4s\n"
+    "ldr q31, [x11, #0x10]\n"
+    "smlal2 v21.4s, v24.8h, v1.8h\n"
+    "ldr d24, [x22, x15]\n"
     "smlal2 v18.4s, v26.8h, v5.8h\n"
-    "ldr d26, [x21, x5]\n"
-    "usubl v26.8h, v26.8b, v12.8b\n"
-    "smlal v20.4s, v29.4h, v8.4h\n"
-    "smlal2 v23.4s, v29.8h, v8.8h\n"
-    "smlal v17.4s, v29.4h, v2.4h\n"
+    "usubl v24.8h, v24.8b, v12.8b\n"
+    "smlal v23.4s, v29.4h, v2.4h\n"
+    "ldr d26, [x21, x15]\n"
     "smlal2 v18.4s, v29.8h, v2.8h\n"
-    "ldr d29, [x19, x5]\n"
-    "add x5, x5, #0x8\n"
-    "smlal v16.4s, v27.4h, v7.4h\n"
-    "usubl v29.8h, v29.8b, v12.8b\n"
-    "smlal2 v22.4s, v27.8h, v7.8h\n"
-    "smlal v17.4s, v24.4h, v3.4h\n"
-    "smlal v16.4s, v24.4h, v5.4h\n"
+    "usubl v26.8h, v26.8b, v12.8b\n"
+    "smlal v22.4s, v25.4h, v6.4h\n"
+    "smlal v23.4s, v24.4h, v3.4h\n"
+    "and v30.16b, v15.16b, v0.16b\n"
+    "tst x8, #0x7\n"
+    "smlal v9.4s, v28.4h, v7.4h\n"
+    "smlal2 v16.4s, v28.8h, v7.8h\n"
+    "sqdmulh v10.4s, v10.4s, v4.4s\n"
+    "add x13, x13, #0x20\n"
+    "smlal2 v21.4s, v25.8h, v6.8h\n"
+    "ldr d25, [x20, x15]\n"
     "smlal2 v18.4s, v24.8h, v3.8h\n"
-    "sqrdmulh v15.4s, v15.4s, v21.4s\n"
-    "smlal2 v22.4s, v24.8h, v5.8h\n"
-    "smlal v17.4s, v26.4h, v7.4h\n"
+    "usubl v25.8h, v25.8b, v12.8b\n"
+    "smlal v22.4s, v27.4h, v7.4h\n"
+    "smlal v23.4s, v26.4h, v7.4h\n"
+    "sshr v30.4s, v30.4s, #0x1f\n"
+    "add x11, x11, #0x20\n"
+    "smlal v9.4s, v29.4h, v8.4h\n"
+    "smlal2 v16.4s, v29.8h, v8.8h\n"
+    "ldr d29, [x19, x15]\n"
+    "usubl v29.8h, v29.8b, v12.8b\n"
+    "smlal2 v21.4s, v27.8h, v7.8h\n"
     "smlal2 v18.4s, v26.8h, v7.8h\n"
-    "smlal v16.4s, v25.4h, v8.4h\n"
-    "smlal2 v22.4s, v25.8h, v8.8h\n"
-    "smlal v17.4s, v25.4h, v6.4h\n"
+    "sqdmulh v9.4s, v9.4s, v19.4s\n"
+    "add x15, x15, #0x8\n"
+    "smlal v22.4s, v24.4h, v5.4h\n"
+    "smlal v23.4s, v25.4h, v6.4h\n"
+    "and v28.16b, v9.16b, v0.16b\n"
+    "smlal2 v21.4s, v24.8h, v5.8h\n"
     "smlal2 v18.4s, v25.8h, v6.8h\n"
-    "and v26.16b, v15.16b, v30.16b\n"
-    "sshr v26.4s, v26.4s, #0x1f\n"
-    "smlal v17.4s, v29.4h, v8.4h\n"
+    "sqdmulh v16.4s, v16.4s, v4.4s\n"
+    "smlal v22.4s, v25.4h, v8.4h\n"
+    "smlal v23.4s, v29.4h, v8.4h\n"
+    "sqdmulh v22.4s, v22.4s, v19.4s\n"
+    "smlal2 v21.4s, v25.8h, v8.8h\n"
     "smlal2 v18.4s, v29.8h, v8.8h\n"
-    "sqrdmulh v10.4s, v10.4s, v31.4s\n"
-    "sqrdmulh v20.4s, v20.4s, v21.4s\n"
-    "sqrdmulh v23.4s, v23.4s, v31.4s\n"
-    "sqrdmulh v16.4s, v16.4s, v21.4s\n"
-    "sqadd v15.4s, v15.4s, v26.4s\n"
-    "and v8.16b, v10.16b, v9.16b\n"
-    "sshr v8.4s, v8.4s, #0x1f\n"
-    "srshl v15.4s, v15.4s, v30.4s\n"
-    "and v4.16b, v20.16b, v30.16b\n"
+    "sqdmulh v23.4s, v23.4s, v19.4s\n"
+    "and v29.16b, v22.16b, v0.16b\n"
+    "sqdmulh v21.4s, v21.4s, v4.4s\n"
+    "and v20.16b, v23.16b, v0.16b\n"
+    "sqdmulh v18.4s, v18.4s, v4.4s\n"
+    "and v19.16b, v10.16b, v31.16b\n"
+    "sshr v28.4s, v28.4s, #0x1f\n"
+    "and v4.16b, v16.16b, v31.16b\n"
+    "sshr v29.4s, v29.4s, #0x1f\n"
+    "and v5.16b, v21.16b, v31.16b\n"
+    "sshr v20.4s, v20.4s, #0x1f\n"
+    "and v26.16b, v18.16b, v31.16b\n"
+    "sqadd v15.4s, v15.4s, v30.4s\n"
+    "sshr v19.4s, v19.4s, #0x1f\n"
+    "sqadd v9.4s, v9.4s, v28.4s\n"
     "sshr v4.4s, v4.4s, #0x1f\n"
-    "and v2.16b, v23.16b, v9.16b\n"
-    "and v1.16b, v16.16b, v30.16b\n"
-    "sshr v2.4s, v2.4s, #0x1f\n"
-    "add v15.4s, v15.4s, v11.4s\n"
-    "sqadd v10.4s, v10.4s, v8.4s\n"
-    "sshr v1.4s, v1.4s, #0x1f\n"
-    "sqrdmulh v22.4s, v22.4s, v31.4s\n"
-    "sqadd v20.4s, v20.4s, v4.4s\n"
-    "smin v15.4s, v15.4s, v14.4s\n"
-    "srshl v10.4s, v10.4s, v9.4s\n"
-    "sqadd v23.4s, v23.4s, v2.4s\n"
-    "smax v15.4s, v15.4s, v19.4s\n"
-    "srshl v20.4s, v20.4s, v30.4s\n"
-    "add v10.4s, v10.4s, v11.4s\n"
-    "srshl v23.4s, v23.4s, v9.4s\n"
-    "sqadd v16.4s, v16.4s, v1.4s\n"
-    "smin v10.4s, v10.4s, v14.4s\n"
-    "add v20.4s, v20.4s, v11.4s\n"
-    "add v23.4s, v23.4s, v11.4s\n"
-    "smax v10.4s, v10.4s, v19.4s\n"
-    "smin v20.4s, v20.4s, v14.4s\n"
-    "smin v23.4s, v23.4s, v14.4s\n"
-    "uzp1 v15.16b, v15.16b, v10.16b\n"
-    "smax v20.4s, v20.4s, v19.4s\n"
+    "sqadd v22.4s, v22.4s, v29.4s\n"
+    "sshr v5.4s, v5.4s, #0x1f\n"
+    "sqadd v23.4s, v23.4s, v20.4s\n"
+    "sshr v26.4s, v26.4s, #0x1f\n"
+    "srshl v15.4s, v15.4s, v0.4s\n"
+    "sqadd v10.4s, v10.4s, v19.4s\n"
+    "srshl v9.4s, v9.4s, v0.4s\n"
+    "sqadd v16.4s, v16.4s, v4.4s\n"
+    "srshl v22.4s, v22.4s, v0.4s\n"
+    "sqadd v21.4s, v21.4s, v5.4s\n"
+    "srshl v23.4s, v23.4s, v0.4s\n"
+    "sqadd v18.4s, v18.4s, v26.4s\n"
+    "srshl v10.4s, v10.4s, v31.4s\n"
+    "sqxtn v15.4h, v15.4s\n"
+    "srshl v16.4s, v16.4s, v31.4s\n"
+    "sqxtn v9.4h, v9.4s\n"
+    "srshl v21.4s, v21.4s, v31.4s\n"
+    "sqxtn v22.4h, v22.4s\n"
+    "srshl v18.4s, v18.4s, v31.4s\n"
+    "sqxtn v23.4h, v23.4s\n"
+    "sqxtn2 v15.8h, v10.4s\n"
+    "sqxtn2 v9.8h, v16.4s\n"
+    "sqxtn2 v22.8h, v21.4s\n"
+    "sqxtn2 v23.8h, v18.4s\n"
+    "sqadd v15.8h, v15.8h, v11.8h\n"
+    "sqadd v9.8h, v9.8h, v11.8h\n"
+    "sqadd v22.8h, v22.8h, v11.8h\n"
+    "sqadd v23.8h, v23.8h, v11.8h\n"
+    "smax v15.8h, v15.8h, v17.8h\n"
+    "smax v9.8h, v9.8h, v17.8h\n"
+    "smax v22.8h, v22.8h, v17.8h\n"
+    "smax v23.8h, v23.8h, v17.8h\n"
+    "smin v15.8h, v15.8h, v14.8h\n"
+    "smin v9.8h, v9.8h, v14.8h\n"
+    "smin v22.8h, v22.8h, v14.8h\n"
+    "smin v23.8h, v23.8h, v14.8h\n"
     "uzp1 v15.16b, v15.16b, v15.16b\n"
-    "str d15, [x14, x7]\n"
-    "smax v23.4s, v23.4s, v19.4s\n"
-    "srshl v16.4s, v16.4s, v30.4s\n"
-    "and v24.16b, v22.16b, v9.16b\n"
-    "sshr v24.4s, v24.4s, #0x1f\n"
-    "uzp1 v20.16b, v20.16b, v23.16b\n"
-    "add v16.4s, v16.4s, v11.4s\n"
-    "sqrdmulh v17.4s, v17.4s, v21.4s\n"
-    "uzp1 v20.16b, v20.16b, v20.16b\n"
-    "str d20, [x13, x7]\n"
-    "smin v16.4s, v16.4s, v14.4s\n"
-    "sqrdmulh v18.4s, v18.4s, v31.4s\n"
-    "sqadd v22.4s, v22.4s, v24.4s\n"
-    "and v2.16b, v17.16b, v30.16b\n"
-    "sshr v2.4s, v2.4s, #0x1f\n"
-    "smax v16.4s, v16.4s, v19.4s\n"
-    "srshl v22.4s, v22.4s, v9.4s\n"
-    "and v31.16b, v18.16b, v9.16b\n"
-    "sshr v31.4s, v31.4s, #0x1f\n"
-    "add v22.4s, v22.4s, v11.4s\n"
-    "sqadd v17.4s, v17.4s, v2.4s\n"
-    "smin v22.4s, v22.4s, v14.4s\n"
-    "srshl v17.4s, v17.4s, v30.4s\n"
-    "sqadd v18.4s, v18.4s, v31.4s\n"
-    "smax v22.4s, v22.4s, v19.4s\n"
-    "uzp1 v16.16b, v16.16b, v22.16b\n"
-    "add v17.4s, v17.4s, v11.4s\n"
-    "srshl v18.4s, v18.4s, v9.4s\n"
-    "uzp1 v16.16b, v16.16b, v16.16b\n"
-    "str d16, [x12, x7]\n"
-    "smin v17.4s, v17.4s, v14.4s\n"
-    "add v18.4s, v18.4s, v11.4s\n"
-    "smax v17.4s, v17.4s, v19.4s\n"
-    "smin v18.4s, v18.4s, v14.4s\n"
-    "smax v18.4s, v18.4s, v19.4s\n"
-    "uzp1 v17.16b, v17.16b, v18.16b\n"
-    "uzp1 v17.16b, v17.16b, v17.16b\n"
-    "str d17, [x11, x7]\n"
-    "add x7, x7, #0x8\n"
+    "str d15, [x10, x14]\n"
+    "uzp1 v9.16b, v9.16b, v9.16b\n"
+    "uzp1 v22.16b, v22.16b, v22.16b\n"
+    "str d9, [x9, x14]\n"
+    "uzp1 v23.16b, v23.16b, v23.16b\n"
+    "str d22, [x28, x14]\n"
+    "str d23, [x27, x14]\n"
+    "add x14, x14, #0x8\n"
     "beq 88f\n"
-    "add x6, x6, #0x48\n"
+    "add x17, x17, #0x48\n"
     "3:"  // Oddments
     "ldr x19, [%x[params], %[offsetof_Params_bias]]\n"
-    "tbz x4, #2, 5f\n"
+    "tbz x8, #2, 5f\n"
     "ld1 { v15.4s }, [x19], #0x10\n"
-    "tbz x4, #1, 4f\n"
+    "tbz x8, #1, 4f\n"
     "ld1 { v10.d }[0], [x19], #0x8\n"
-    "tbz x4, #0, 7f\n"
+    "tbz x8, #0, 7f\n"
     "ld1 { v10.s }[2], [x19]\n"
     "b 7f\n"
     "4:"  // Oddments: Load bias: Bit 2: Bit 1: Unset
-    "tbz x4, #0, 7f\n"
+    "tbz x8, #0, 7f\n"
     "ld1 { v10.s }[0], [x19]\n"
     "b 7f\n"
     "5:"  // Oddments: Load bias: Bit 2: Unset
-    "tbz x4, #1, 6f\n"
+    "tbz x8, #1, 6f\n"
     "ld1 { v15.d }[0], [x19], #0x8\n"
-    "tbz x4, #0, 7f\n"
+    "tbz x8, #0, 7f\n"
     "ld1 { v15.s }[2], [x19]\n"
     "b 7f\n"
     "6:"  // Oddments: Load bias: Bit 2: Unset: Bit 1: Unset
-    "tbz x4, #0, 7f\n"
+    "tbz x8, #0, 7f\n"
     "ld1 { v15.s }[0], [x19]\n"
     "7:"  // Oddments: Load bias: Bit 2: End
-    "mov v20.16b, v15.16b\n"
-    "ldr d0, [x6, #0x0]\n"
-    "mov v23.16b, v10.16b\n"
-    "ldr d1, [x6, #0x8]\n"
-    "mov v16.16b, v15.16b\n"
-    "ldr d2, [x6, #0x10]\n"
-    "mov v22.16b, v10.16b\n"
-    "ldr d3, [x6, #0x18]\n"
-    "mov v17.16b, v15.16b\n"
-    "ldr d4, [x6, #0x20]\n"
-    "usubl v0.8h, v0.8b, v13.8b\n"
+    "ldr d0, [x17, #0x0]\n"
+    "ldr d1, [x17, #0x8]\n"
+    "mov v9.16b, v15.16b\n"
+    "mov v16.16b, v10.16b\n"
+    "ldr d2, [x17, #0x10]\n"
+    "ldr d3, [x17, #0x18]\n"
+    "mov v22.16b, v15.16b\n"
+    "mov v21.16b, v10.16b\n"
+    "ldr d4, [x17, #0x20]\n"
+    "ldr d5, [x17, #0x28]\n"
+    "mov v23.16b, v15.16b\n"
     "mov v18.16b, v10.16b\n"
-    "ldr d5, [x6, #0x28]\n"
+    "ldr d6, [x17, #0x30]\n"
+    "ldr d7, [x17, #0x38]\n"
+    "usubl v0.8h, v0.8b, v13.8b\n"
     "usubl v1.8h, v1.8b, v13.8b\n"
-    "ldr d6, [x6, #0x30]\n"
+    "ldr d8, [x17, #0x40]\n"
+    "ldp x26, x25, [x12, #0x0]\n"
     "usubl v2.8h, v2.8b, v13.8b\n"
-    "ldr d7, [x6, #0x38]\n"
     "usubl v3.8h, v3.8b, v13.8b\n"
-    "ldr d8, [x6, #0x40]\n"
+    "ldp x24, x23, [x12, #0x10]\n"
+    "ldp x22, x21, [x12, #0x20]\n"
     "usubl v4.8h, v4.8b, v13.8b\n"
-    "ldp x26, x25, [x8, #0x0]\n"
     "usubl v5.8h, v5.8b, v13.8b\n"
-    "ldp x24, x23, [x8, #0x10]\n"
+    "ldp x20, x19, [x12, #0x30]\n"
     "usubl v6.8h, v6.8b, v13.8b\n"
     "usubl v7.8h, v7.8b, v13.8b\n"
-    "ldp x22, x21, [x8, #0x20]\n"
     "usubl v8.8h, v8.8b, v13.8b\n"
-    "ldp x20, x19, [x8, #0x30]\n"
-    "add x26, x26, x5\n"
-    "add x25, x25, x5\n"
-    "add x24, x24, x5\n"
-    "add x23, x23, x5\n"
-    "add x22, x22, x5\n"
-    "add x21, x21, x5\n"
-    "add x20, x20, x5\n"
-    "add x19, x19, x5\n"
-    "tbz x4, #2, 9f\n"
+    "add x26, x26, x15\n"
+    "add x25, x25, x15\n"
+    "add x24, x24, x15\n"
+    "add x23, x23, x15\n"
+    "add x22, x22, x15\n"
+    "add x21, x21, x15\n"
+    "add x20, x20, x15\n"
+    "add x19, x19, x15\n"
+    "tbz x8, #2, 9f\n"
     "ld1 { v31.s }[0], [x26], #0x4\n"
     "ld1 { v30.s }[0], [x25], #0x4\n"
     "ld1 { v29.s }[0], [x24], #0x4\n"
@@ -713,7 +697,7 @@ void a64_u8q_nhwc_3x3_s2_output2x2_mla_depthfirst_impl(
     "ld1 { v26.s }[0], [x21], #0x4\n"
     "ld1 { v25.s }[0], [x20], #0x4\n"
     "ld1 { v24.s }[0], [x19], #0x4\n"
-    "tbz x4, #1, 8f\n"
+    "tbz x8, #1, 8f\n"
     "ld1 { v31.h }[2], [x26], #0x2\n"
     "ld1 { v30.h }[2], [x25], #0x2\n"
     "ld1 { v29.h }[2], [x24], #0x2\n"
@@ -722,7 +706,7 @@ void a64_u8q_nhwc_3x3_s2_output2x2_mla_depthfirst_impl(
     "ld1 { v26.h }[2], [x21], #0x2\n"
     "ld1 { v25.h }[2], [x20], #0x2\n"
     "ld1 { v24.h }[2], [x19], #0x2\n"
-    "tbz x4, #0, 11f\n"
+    "tbz x8, #0, 11f\n"
     "ld1 { v31.b }[6], [x26]\n"
     "ld1 { v30.b }[6], [x25]\n"
     "ld1 { v29.b }[6], [x24]\n"
@@ -733,7 +717,7 @@ void a64_u8q_nhwc_3x3_s2_output2x2_mla_depthfirst_impl(
     "ld1 { v24.b }[6], [x19]\n"
     "b 11f\n"
     "8:"  // Oddments: Initial loads: Bit 2: Bit 1: Unset
-    "tbz x4, #0, 11f\n"
+    "tbz x8, #0, 11f\n"
     "ld1 { v31.b }[4], [x26]\n"
     "ld1 { v30.b }[4], [x25]\n"
     "ld1 { v29.b }[4], [x24]\n"
@@ -744,7 +728,7 @@ void a64_u8q_nhwc_3x3_s2_output2x2_mla_depthfirst_impl(
     "ld1 { v24.b }[4], [x19]\n"
     "b 11f\n"
     "9:"  // Oddments: Initial loads: Bit 2: Unset
-    "tbz x4, #1, 10f\n"
+    "tbz x8, #1, 10f\n"
     "ld1 { v31.h }[0], [x26], #0x2\n"
     "ld1 { v30.h }[0], [x25], #0x2\n"
     "ld1 { v29.h }[0], [x24], #0x2\n"
@@ -753,7 +737,7 @@ void a64_u8q_nhwc_3x3_s2_output2x2_mla_depthfirst_impl(
     "ld1 { v26.h }[0], [x21], #0x2\n"
     "ld1 { v25.h }[0], [x20], #0x2\n"
     "ld1 { v24.h }[0], [x19], #0x2\n"
-    "tbz x4, #0, 11f\n"
+    "tbz x8, #0, 11f\n"
     "ld1 { v31.b }[2], [x26]\n"
     "ld1 { v30.b }[2], [x25]\n"
     "ld1 { v29.b }[2], [x24]\n"
@@ -764,7 +748,7 @@ void a64_u8q_nhwc_3x3_s2_output2x2_mla_depthfirst_impl(
     "ld1 { v24.b }[2], [x19]\n"
     "b 11f\n"
     "10:"  // Oddments: Initial loads: Bit 2: Unset: Bit 1: Unset
-    "tbz x4, #0, 11f\n"
+    "tbz x8, #0, 11f\n"
     "ld1 { v31.b }[0], [x26]\n"
     "ld1 { v30.b }[0], [x25]\n"
     "ld1 { v29.b }[0], [x24]\n"
@@ -774,646 +758,636 @@ void a64_u8q_nhwc_3x3_s2_output2x2_mla_depthfirst_impl(
     "ld1 { v25.b }[0], [x20]\n"
     "ld1 { v24.b }[0], [x19]\n"
     "11:"  // Oddments: Initial loads: Bit 2: End
-    "ldr x23, [x8, #0x40]\n"
     "usubl v31.8h, v31.8b, v12.8b\n"
     "smlal v15.4s, v31.4h, v8.4h\n"
-    "usubl v30.8h, v30.8b, v12.8b\n"
     "smlal2 v10.4s, v31.8h, v8.8h\n"
-    "usubl v29.8h, v29.8b, v12.8b\n"
-    "smlal v20.4s, v31.4h, v6.4h\n"
-    "usubl v28.8h, v28.8b, v12.8b\n"
-    "smlal2 v23.4s, v31.8h, v6.8h\n"
-    "usubl v27.8h, v27.8b, v12.8b\n"
-    "smlal v16.4s, v31.4h, v2.4h\n"
-    "usubl v26.8h, v26.8b, v12.8b\n"
-    "smlal2 v22.4s, v31.8h, v2.8h\n"
-    "usubl v25.8h, v25.8b, v12.8b\n"
-    "smlal v17.4s, v31.4h, v0.4h\n"
-    "usubl v24.8h, v24.8b, v12.8b\n"
-    "smlal2 v18.4s, v31.8h, v0.8h\n"
-    "add x23, x23, x5\n"
+    "ldr x24, [x12, #0x40]\n"
+    "usubl v30.8h, v30.8b, v12.8b\n"
     "smlal v15.4s, v30.4h, v0.4h\n"
     "smlal2 v10.4s, v30.8h, v0.8h\n"
-    "smlal v20.4s, v28.4h, v1.4h\n"
-    "smlal2 v23.4s, v28.8h, v1.8h\n"
+    "add x24, x24, x15\n"
+    "usubl v29.8h, v29.8b, v12.8b\n"
+    "smlal v9.4s, v31.4h, v6.4h\n"
+    "smlal2 v16.4s, v31.8h, v6.8h\n"
     "smlal v15.4s, v29.4h, v1.4h\n"
     "smlal2 v10.4s, v29.8h, v1.8h\n"
-    "smlal v20.4s, v27.4h, v2.4h\n"
-    "smlal2 v23.4s, v27.8h, v2.8h\n"
+    "usubl v28.8h, v28.8b, v12.8b\n"
+    "usubl v26.8h, v26.8b, v12.8b\n"
+    "smlal v9.4s, v28.4h, v1.4h\n"
+    "smlal2 v16.4s, v28.8h, v1.8h\n"
     "smlal v15.4s, v26.4h, v3.4h\n"
     "smlal2 v10.4s, v26.8h, v3.8h\n"
-    "smlal v20.4s, v24.4h, v0.4h\n"
-    "smlal2 v23.4s, v24.8h, v0.8h\n"
+    "usubl v27.8h, v27.8b, v12.8b\n"
+    "usubl v25.8h, v25.8b, v12.8b\n"
+    "smlal v9.4s, v27.4h, v2.4h\n"
+    "smlal2 v16.4s, v27.8h, v2.8h\n"
     "smlal v15.4s, v25.4h, v4.4h\n"
     "smlal2 v10.4s, v25.8h, v4.8h\n"
+    "usubl v24.8h, v24.8b, v12.8b\n"
+    "smlal v22.4s, v31.4h, v2.4h\n"
+    "smlal2 v21.4s, v31.8h, v2.8h\n"
+    "smlal v23.4s, v31.4h, v0.4h\n"
+    "smlal2 v18.4s, v31.8h, v0.8h\n"
     "smlal v15.4s, v24.4h, v2.4h\n"
     "smlal2 v10.4s, v24.8h, v2.8h\n"
-    "tbz x4, #2, 13f\n"
-    "ld1 { v29.s }[0], [x23], #0x4\n"
-    "tbz x4, #1, 12f\n"
-    "ld1 { v29.h }[2], [x23], #0x2\n"
-    "tbz x4, #0, 15f\n"
-    "ld1 { v29.b }[6], [x23]\n"
+    "smlal v9.4s, v24.4h, v0.4h\n"
+    "smlal2 v16.4s, v24.8h, v0.8h\n"
+    "tbz x8, #2, 13f\n"
+    "ld1 { v29.s }[0], [x24], #0x4\n"
+    "tbz x8, #1, 12f\n"
+    "ld1 { v29.h }[2], [x24], #0x2\n"
+    "tbz x8, #0, 15f\n"
+    "ld1 { v29.b }[6], [x24]\n"
     "b 15f\n"
     "12:"  // Oddments: Load (1, 3): Bit 2: Bit 1: Unset
-    "tbz x4, #0, 15f\n"
-    "ld1 { v29.b }[4], [x23]\n"
+    "tbz x8, #0, 15f\n"
+    "ld1 { v29.b }[4], [x24]\n"
     "b 15f\n"
     "13:"  // Oddments: Load (1, 3): Bit 2: Unset
-    "tbz x4, #1, 14f\n"
-    "ld1 { v29.h }[0], [x23], #0x2\n"
-    "tbz x4, #0, 15f\n"
-    "ld1 { v29.b }[2], [x23]\n"
+    "tbz x8, #1, 14f\n"
+    "ld1 { v29.h }[0], [x24], #0x2\n"
+    "tbz x8, #0, 15f\n"
+    "ld1 { v29.b }[2], [x24]\n"
     "b 15f\n"
     "14:"  // Oddments: Load (1, 3): Bit 2: Unset: Bit 1: Unset
-    "tbz x4, #0, 15f\n"
-    "ld1 { v29.b }[0], [x23]\n"
+    "tbz x8, #0, 15f\n"
+    "ld1 { v29.b }[0], [x24]\n"
     "15:"  // Oddments: Load (1, 3): Bit 2: End
-    "ldr x22, [x8, #0x48]\n"
     "usubl v29.8h, v29.8b, v12.8b\n"
-    "smlal v20.4s, v29.4h, v4.4h\n"
-    "smlal2 v23.4s, v29.8h, v4.8h\n"
-    "add x22, x22, x5\n"
-    "tbz x4, #2, 17f\n"
-    "ld1 { v28.s }[0], [x22], #0x4\n"
-    "tbz x4, #1, 16f\n"
-    "ld1 { v28.h }[2], [x22], #0x2\n"
-    "tbz x4, #0, 19f\n"
-    "ld1 { v28.b }[6], [x22]\n"
+    "ldr x23, [x12, #0x48]\n"
+    "smlal v9.4s, v29.4h, v4.4h\n"
+    "smlal2 v16.4s, v29.8h, v4.8h\n"
+    "add x23, x23, x15\n"
+    "tbz x8, #2, 17f\n"
+    "ld1 { v28.s }[0], [x23], #0x4\n"
+    "tbz x8, #1, 16f\n"
+    "ld1 { v28.h }[2], [x23], #0x2\n"
+    "tbz x8, #0, 19f\n"
+    "ld1 { v28.b }[6], [x23]\n"
     "b 19f\n"
     "16:"  // Oddments: Load (1, 4): Bit 2: Bit 1: Unset
-    "tbz x4, #0, 19f\n"
-    "ld1 { v28.b }[4], [x22]\n"
+    "tbz x8, #0, 19f\n"
+    "ld1 { v28.b }[4], [x23]\n"
     "b 19f\n"
     "17:"  // Oddments: Load (1, 4): Bit 2: Unset
-    "tbz x4, #1, 18f\n"
-    "ld1 { v28.h }[0], [x22], #0x2\n"
-    "tbz x4, #0, 19f\n"
-    "ld1 { v28.b }[2], [x22]\n"
+    "tbz x8, #1, 18f\n"
+    "ld1 { v28.h }[0], [x23], #0x2\n"
+    "tbz x8, #0, 19f\n"
+    "ld1 { v28.b }[2], [x23]\n"
     "b 19f\n"
     "18:"  // Oddments: Load (1, 4): Bit 2: Unset: Bit 1: Unset
-    "tbz x4, #0, 19f\n"
-    "ld1 { v28.b }[0], [x22]\n"
+    "tbz x8, #0, 19f\n"
+    "ld1 { v28.b }[0], [x23]\n"
     "19:"  // Oddments: Load (1, 4): Bit 2: End
-    "ldr x21, [x8, #0x50]\n"
     "usubl v28.8h, v28.8b, v12.8b\n"
-    "smlal v20.4s, v28.4h, v5.4h\n"
-    "smlal2 v23.4s, v28.8h, v5.8h\n"
-    "add x21, x21, x5\n"
-    "tbz x4, #2, 21f\n"
+    "ldr x21, [x12, #0x50]\n"
+    "smlal v9.4s, v28.4h, v5.4h\n"
+    "smlal2 v16.4s, v28.8h, v5.8h\n"
+    "add x21, x21, x15\n"
+    "tbz x8, #2, 21f\n"
     "ld1 { v27.s }[0], [x21], #0x4\n"
-    "tbz x4, #1, 20f\n"
+    "tbz x8, #1, 20f\n"
     "ld1 { v27.h }[2], [x21], #0x2\n"
-    "tbz x4, #0, 23f\n"
+    "tbz x8, #0, 23f\n"
     "ld1 { v27.b }[6], [x21]\n"
     "b 23f\n"
     "20:"  // Oddments: Load (1, 2): Bit 2: Bit 1: Unset
-    "tbz x4, #0, 23f\n"
+    "tbz x8, #0, 23f\n"
     "ld1 { v27.b }[4], [x21]\n"
     "b 23f\n"
     "21:"  // Oddments: Load (1, 2): Bit 2: Unset
-    "tbz x4, #1, 22f\n"
+    "tbz x8, #1, 22f\n"
     "ld1 { v27.h }[0], [x21], #0x2\n"
-    "tbz x4, #0, 23f\n"
+    "tbz x8, #0, 23f\n"
     "ld1 { v27.b }[2], [x21]\n"
     "b 23f\n"
     "22:"  // Oddments: Load (1, 2): Bit 2: Unset: Bit 1: Unset
-    "tbz x4, #0, 23f\n"
+    "tbz x8, #0, 23f\n"
     "ld1 { v27.b }[0], [x21]\n"
     "23:"  // Oddments: Load (1, 2): Bit 2: End
-    "ldr x20, [x8, #0x58]\n"
     "usubl v27.8h, v27.8b, v12.8b\n"
+    "ldr x19, [x12, #0x58]\n"
     "smlal v15.4s, v27.4h, v5.4h\n"
     "smlal2 v10.4s, v27.8h, v5.8h\n"
-    "add x20, x20, x5\n"
-    "smlal v20.4s, v27.4h, v3.4h\n"
-    "smlal2 v23.4s, v27.8h, v3.8h\n"
-    "tbz x4, #2, 25f\n"
-    "ld1 { v26.s }[0], [x20], #0x4\n"
-    "tbz x4, #1, 24f\n"
-    "ld1 { v26.h }[2], [x20], #0x2\n"
-    "tbz x4, #0, 27f\n"
-    "ld1 { v26.b }[6], [x20]\n"
+    "smlal v9.4s, v27.4h, v3.4h\n"
+    "smlal2 v16.4s, v27.8h, v3.8h\n"
+    "add x19, x19, x15\n"
+    "tbz x8, #2, 25f\n"
+    "ld1 { v26.s }[0], [x19], #0x4\n"
+    "tbz x8, #1, 24f\n"
+    "ld1 { v26.h }[2], [x19], #0x2\n"
+    "tbz x8, #0, 27f\n"
+    "ld1 { v26.b }[6], [x19]\n"
     "b 27f\n"
     "24:"  // Oddments: Load (3, 0): Bit 2: Bit 1: Unset
-    "tbz x4, #0, 27f\n"
-    "ld1 { v26.b }[4], [x20]\n"
+    "tbz x8, #0, 27f\n"
+    "ld1 { v26.b }[4], [x19]\n"
     "b 27f\n"
     "25:"  // Oddments: Load (3, 0): Bit 2: Unset
-    "tbz x4, #1, 26f\n"
-    "ld1 { v26.h }[0], [x20], #0x2\n"
-    "tbz x4, #0, 27f\n"
-    "ld1 { v26.b }[2], [x20]\n"
+    "tbz x8, #1, 26f\n"
+    "ld1 { v26.h }[0], [x19], #0x2\n"
+    "tbz x8, #0, 27f\n"
+    "ld1 { v26.b }[2], [x19]\n"
     "b 27f\n"
     "26:"  // Oddments: Load (3, 0): Bit 2: Unset: Bit 1: Unset
-    "tbz x4, #0, 27f\n"
-    "ld1 { v26.b }[0], [x20]\n"
+    "tbz x8, #0, 27f\n"
+    "ld1 { v26.b }[0], [x19]\n"
     "27:"  // Oddments: Load (3, 0): Bit 2: End
-    "ldr x19, [x8, #0x60]\n"
     "usubl v26.8h, v26.8b, v12.8b\n"
-    "smlal v16.4s, v26.4h, v3.4h\n"
-    "smlal2 v22.4s, v26.8h, v3.8h\n"
-    "add x19, x19, x5\n"
-    "tbz x4, #2, 29f\n"
-    "ld1 { v25.s }[0], [x19], #0x4\n"
-    "tbz x4, #1, 28f\n"
-    "ld1 { v25.h }[2], [x19], #0x2\n"
-    "tbz x4, #0, 31f\n"
-    "ld1 { v25.b }[6], [x19]\n"
+    "ldr x20, [x12, #0x60]\n"
+    "smlal v22.4s, v26.4h, v3.4h\n"
+    "smlal2 v21.4s, v26.8h, v3.8h\n"
+    "add x20, x20, x15\n"
+    "tbz x8, #2, 29f\n"
+    "ld1 { v25.s }[0], [x20], #0x4\n"
+    "tbz x8, #1, 28f\n"
+    "ld1 { v25.h }[2], [x20], #0x2\n"
+    "tbz x8, #0, 31f\n"
+    "ld1 { v25.b }[6], [x20]\n"
     "b 31f\n"
     "28:"  // Oddments: Load (2, 0): Bit 2: Bit 1: Unset
-    "tbz x4, #0, 31f\n"
-    "ld1 { v25.b }[4], [x19]\n"
+    "tbz x8, #0, 31f\n"
+    "ld1 { v25.b }[4], [x20]\n"
     "b 31f\n"
     "29:"  // Oddments: Load (2, 0): Bit 2: Unset
-    "tbz x4, #1, 30f\n"
-    "ld1 { v25.h }[0], [x19], #0x2\n"
-    "tbz x4, #0, 31f\n"
-    "ld1 { v25.b }[2], [x19]\n"
+    "tbz x8, #1, 30f\n"
+    "ld1 { v25.h }[0], [x20], #0x2\n"
+    "tbz x8, #0, 31f\n"
+    "ld1 { v25.b }[2], [x20]\n"
     "b 31f\n"
     "30:"  // Oddments: Load (2, 0): Bit 2: Unset: Bit 1: Unset
-    "tbz x4, #0, 31f\n"
-    "ld1 { v25.b }[0], [x19]\n"
+    "tbz x8, #0, 31f\n"
+    "ld1 { v25.b }[0], [x20]\n"
     "31:"  // Oddments: Load (2, 0): Bit 2: End
-    "ldr x10, [x8, #0x68]\n"
     "usubl v25.8h, v25.8b, v12.8b\n"
+    "ldr x19, [x12, #0x68]\n"
     "smlal v15.4s, v25.4h, v6.4h\n"
     "smlal2 v10.4s, v25.8h, v6.8h\n"
-    "add x10, x10, x5\n"
-    "smlal v16.4s, v25.4h, v0.4h\n"
-    "smlal2 v22.4s, v25.8h, v0.8h\n"
-    "tbz x4, #2, 33f\n"
-    "ld1 { v29.s }[0], [x10], #0x4\n"
-    "tbz x4, #1, 32f\n"
-    "ld1 { v29.h }[2], [x10], #0x2\n"
-    "tbz x4, #0, 35f\n"
-    "ld1 { v29.b }[6], [x10]\n"
+    "smlal v22.4s, v25.4h, v0.4h\n"
+    "smlal2 v21.4s, v25.8h, v0.8h\n"
+    "add x19, x19, x15\n"
+    "tbz x8, #2, 33f\n"
+    "ld1 { v29.s }[0], [x19], #0x4\n"
+    "tbz x8, #1, 32f\n"
+    "ld1 { v29.h }[2], [x19], #0x2\n"
+    "tbz x8, #0, 35f\n"
+    "ld1 { v29.b }[6], [x19]\n"
     "b 35f\n"
     "32:"  // Oddments: Load (3, 1): Bit 2: Bit 1: Unset
-    "tbz x4, #0, 35f\n"
-    "ld1 { v29.b }[4], [x10]\n"
+    "tbz x8, #0, 35f\n"
+    "ld1 { v29.b }[4], [x19]\n"
     "b 35f\n"
     "33:"  // Oddments: Load (3, 1): Bit 2: Unset
-    "tbz x4, #1, 34f\n"
-    "ld1 { v29.h }[0], [x10], #0x2\n"
-    "tbz x4, #0, 35f\n"
-    "ld1 { v29.b }[2], [x10]\n"
+    "tbz x8, #1, 34f\n"
+    "ld1 { v29.h }[0], [x19], #0x2\n"
+    "tbz x8, #0, 35f\n"
+    "ld1 { v29.b }[2], [x19]\n"
     "b 35f\n"
     "34:"  // Oddments: Load (3, 1): Bit 2: Unset: Bit 1: Unset
-    "tbz x4, #0, 35f\n"
-    "ld1 { v29.b }[0], [x10]\n"
+    "tbz x8, #0, 35f\n"
+    "ld1 { v29.b }[0], [x19]\n"
     "35:"  // Oddments: Load (3, 1): Bit 2: End
-    "ldr x9, [x8, #0x70]\n"
     "usubl v29.8h, v29.8b, v12.8b\n"
-    "smlal v16.4s, v29.4h, v4.4h\n"
-    "smlal2 v22.4s, v29.8h, v4.8h\n"
-    "add x9, x9, x5\n"
-    "tbz x4, #2, 37f\n"
-    "ld1 { v24.s }[0], [x9], #0x4\n"
-    "tbz x4, #1, 36f\n"
-    "ld1 { v24.h }[2], [x9], #0x2\n"
-    "tbz x4, #0, 39f\n"
-    "ld1 { v24.b }[6], [x9]\n"
+    "ldr x19, [x12, #0x70]\n"
+    "smlal v22.4s, v29.4h, v4.4h\n"
+    "smlal2 v21.4s, v29.8h, v4.8h\n"
+    "add x19, x19, x15\n"
+    "tbz x8, #2, 37f\n"
+    "ld1 { v24.s }[0], [x19], #0x4\n"
+    "tbz x8, #1, 36f\n"
+    "ld1 { v24.h }[2], [x19], #0x2\n"
+    "tbz x8, #0, 39f\n"
+    "ld1 { v24.b }[6], [x19]\n"
     "b 39f\n"
     "36:"  // Oddments: Load (2, 1): Bit 2: Bit 1: Unset
-    "tbz x4, #0, 39f\n"
-    "ld1 { v24.b }[4], [x9]\n"
+    "tbz x8, #0, 39f\n"
+    "ld1 { v24.b }[4], [x19]\n"
     "b 39f\n"
     "37:"  // Oddments: Load (2, 1): Bit 2: Unset
-    "tbz x4, #1, 38f\n"
-    "ld1 { v24.h }[0], [x9], #0x2\n"
-    "tbz x4, #0, 39f\n"
-    "ld1 { v24.b }[2], [x9]\n"
+    "tbz x8, #1, 38f\n"
+    "ld1 { v24.h }[0], [x19], #0x2\n"
+    "tbz x8, #0, 39f\n"
+    "ld1 { v24.b }[2], [x19]\n"
     "b 39f\n"
     "38:"  // Oddments: Load (2, 1): Bit 2: Unset: Bit 1: Unset
-    "tbz x4, #0, 39f\n"
-    "ld1 { v24.b }[0], [x9]\n"
+    "tbz x8, #0, 39f\n"
+    "ld1 { v24.b }[0], [x19]\n"
     "39:"  // Oddments: Load (2, 1): Bit 2: End
-    "ldr x28, [x8, #0x78]\n"
     "usubl v24.8h, v24.8b, v12.8b\n"
+    "ldr x22, [x12, #0x78]\n"
     "smlal v15.4s, v24.4h, v7.4h\n"
     "smlal2 v10.4s, v24.8h, v7.8h\n"
-    "add x28, x28, x5\n"
-    "smlal v16.4s, v24.4h, v1.4h\n"
-    "smlal2 v22.4s, v24.8h, v1.8h\n"
-    "tbz x4, #2, 41f\n"
-    "ld1 { v27.s }[0], [x28], #0x4\n"
-    "tbz x4, #1, 40f\n"
-    "ld1 { v27.h }[2], [x28], #0x2\n"
-    "tbz x4, #0, 43f\n"
-    "ld1 { v27.b }[6], [x28]\n"
+    "smlal v22.4s, v24.4h, v1.4h\n"
+    "smlal2 v21.4s, v24.8h, v1.8h\n"
+    "add x22, x22, x15\n"
+    "tbz x8, #2, 41f\n"
+    "ld1 { v27.s }[0], [x22], #0x4\n"
+    "tbz x8, #1, 40f\n"
+    "ld1 { v27.h }[2], [x22], #0x2\n"
+    "tbz x8, #0, 43f\n"
+    "ld1 { v27.b }[6], [x22]\n"
     "b 43f\n"
     "40:"  // Oddments: Load (3, 3): Bit 2: Bit 1: Unset
-    "tbz x4, #0, 43f\n"
-    "ld1 { v27.b }[4], [x28]\n"
+    "tbz x8, #0, 43f\n"
+    "ld1 { v27.b }[4], [x22]\n"
     "b 43f\n"
     "41:"  // Oddments: Load (3, 3): Bit 2: Unset
-    "tbz x4, #1, 42f\n"
-    "ld1 { v27.h }[0], [x28], #0x2\n"
-    "tbz x4, #0, 43f\n"
-    "ld1 { v27.b }[2], [x28]\n"
+    "tbz x8, #1, 42f\n"
+    "ld1 { v27.h }[0], [x22], #0x2\n"
+    "tbz x8, #0, 43f\n"
+    "ld1 { v27.b }[2], [x22]\n"
     "b 43f\n"
     "42:"  // Oddments: Load (3, 3): Bit 2: Unset: Bit 1: Unset
-    "tbz x4, #0, 43f\n"
-    "ld1 { v27.b }[0], [x28]\n"
+    "tbz x8, #0, 43f\n"
+    "ld1 { v27.b }[0], [x22]\n"
     "43:"  // Oddments: Load (3, 3): Bit 2: End
-    "ldr x27, [x8, #0x80]\n"
     "usubl v27.8h, v27.8b, v12.8b\n"
-    "smlal v17.4s, v27.4h, v4.4h\n"
+    "ldr x21, [x12, #0x80]\n"
+    "smlal v23.4s, v27.4h, v4.4h\n"
     "smlal2 v18.4s, v27.8h, v4.8h\n"
-    "add x27, x27, x5\n"
-    "tbz x4, #2, 45f\n"
-    "ld1 { v28.s }[0], [x27], #0x4\n"
-    "tbz x4, #1, 44f\n"
-    "ld1 { v28.h }[2], [x27], #0x2\n"
-    "tbz x4, #0, 47f\n"
-    "ld1 { v28.b }[6], [x27]\n"
+    "add x21, x21, x15\n"
+    "tbz x8, #2, 45f\n"
+    "ld1 { v28.s }[0], [x21], #0x4\n"
+    "tbz x8, #1, 44f\n"
+    "ld1 { v28.h }[2], [x21], #0x2\n"
+    "tbz x8, #0, 47f\n"
+    "ld1 { v28.b }[6], [x21]\n"
     "b 47f\n"
     "44:"  // Oddments: Load (2, 3): Bit 2: Bit 1: Unset
-    "tbz x4, #0, 47f\n"
-    "ld1 { v28.b }[4], [x27]\n"
+    "tbz x8, #0, 47f\n"
+    "ld1 { v28.b }[4], [x21]\n"
     "b 47f\n"
     "45:"  // Oddments: Load (2, 3): Bit 2: Unset
-    "tbz x4, #1, 46f\n"
-    "ld1 { v28.h }[0], [x27], #0x2\n"
-    "tbz x4, #0, 47f\n"
-    "ld1 { v28.b }[2], [x27]\n"
+    "tbz x8, #1, 46f\n"
+    "ld1 { v28.h }[0], [x21], #0x2\n"
+    "tbz x8, #0, 47f\n"
+    "ld1 { v28.b }[2], [x21]\n"
     "b 47f\n"
     "46:"  // Oddments: Load (2, 3): Bit 2: Unset: Bit 1: Unset
-    "tbz x4, #0, 47f\n"
-    "ld1 { v28.b }[0], [x27]\n"
+    "tbz x8, #0, 47f\n"
+    "ld1 { v28.b }[0], [x21]\n"
     "47:"  // Oddments: Load (2, 3): Bit 2: End
-    "ldr x26, [x8, #0x88]\n"
     "usubl v28.8h, v28.8b, v12.8b\n"
-    "smlal v20.4s, v28.4h, v7.4h\n"
-    "smlal2 v23.4s, v28.8h, v7.8h\n"
-    "add x26, x26, x5\n"
-    "smlal v17.4s, v28.4h, v1.4h\n"
+    "ldr x20, [x12, #0x88]\n"
+    "smlal v9.4s, v28.4h, v7.4h\n"
+    "smlal2 v16.4s, v28.8h, v7.8h\n"
+    "smlal v23.4s, v28.4h, v1.4h\n"
     "smlal2 v18.4s, v28.8h, v1.8h\n"
-    "tbz x4, #2, 49f\n"
-    "ld1 { v26.s }[0], [x26], #0x4\n"
-    "tbz x4, #1, 48f\n"
-    "ld1 { v26.h }[2], [x26], #0x2\n"
-    "tbz x4, #0, 51f\n"
-    "ld1 { v26.b }[6], [x26]\n"
+    "add x20, x20, x15\n"
+    "tbz x8, #2, 49f\n"
+    "ld1 { v26.s }[0], [x20], #0x4\n"
+    "tbz x8, #1, 48f\n"
+    "ld1 { v26.h }[2], [x20], #0x2\n"
+    "tbz x8, #0, 51f\n"
+    "ld1 { v26.b }[6], [x20]\n"
     "b 51f\n"
     "48:"  // Oddments: Load (3, 4): Bit 2: Bit 1: Unset
-    "tbz x4, #0, 51f\n"
-    "ld1 { v26.b }[4], [x26]\n"
+    "tbz x8, #0, 51f\n"
+    "ld1 { v26.b }[4], [x20]\n"
     "b 51f\n"
     "49:"  // Oddments: Load (3, 4): Bit 2: Unset
-    "tbz x4, #1, 50f\n"
-    "ld1 { v26.h }[0], [x26], #0x2\n"
-    "tbz x4, #0, 51f\n"
-    "ld1 { v26.b }[2], [x26]\n"
+    "tbz x8, #1, 50f\n"
+    "ld1 { v26.h }[0], [x20], #0x2\n"
+    "tbz x8, #0, 51f\n"
+    "ld1 { v26.b }[2], [x20]\n"
     "b 51f\n"
     "50:"  // Oddments: Load (3, 4): Bit 2: Unset: Bit 1: Unset
-    "tbz x4, #0, 51f\n"
-    "ld1 { v26.b }[0], [x26]\n"
+    "tbz x8, #0, 51f\n"
+    "ld1 { v26.b }[0], [x20]\n"
     "51:"  // Oddments: Load (3, 4): Bit 2: End
-    "ldr x25, [x8, #0x90]\n"
     "usubl v26.8h, v26.8b, v12.8b\n"
-    "smlal v17.4s, v26.4h, v5.4h\n"
+    "ldr x23, [x12, #0x90]\n"
+    "smlal v23.4s, v26.4h, v5.4h\n"
     "smlal2 v18.4s, v26.8h, v5.8h\n"
-    "add x25, x25, x5\n"
-    "tbz x4, #2, 53f\n"
-    "ld1 { v25.s }[0], [x25], #0x4\n"
-    "tbz x4, #1, 52f\n"
-    "ld1 { v25.h }[2], [x25], #0x2\n"
-    "tbz x4, #0, 55f\n"
-    "ld1 { v25.b }[6], [x25]\n"
+    "add x23, x23, x15\n"
+    "tbz x8, #2, 53f\n"
+    "ld1 { v25.s }[0], [x23], #0x4\n"
+    "tbz x8, #1, 52f\n"
+    "ld1 { v25.h }[2], [x23], #0x2\n"
+    "tbz x8, #0, 55f\n"
+    "ld1 { v25.b }[6], [x23]\n"
     "b 55f\n"
     "52:"  // Oddments: Load (4, 0): Bit 2: Bit 1: Unset
-    "tbz x4, #0, 55f\n"
-    "ld1 { v25.b }[4], [x25]\n"
+    "tbz x8, #0, 55f\n"
+    "ld1 { v25.b }[4], [x23]\n"
     "b 55f\n"
     "53:"  // Oddments: Load (4, 0): Bit 2: Unset
-    "tbz x4, #1, 54f\n"
-    "ld1 { v25.h }[0], [x25], #0x2\n"
-    "tbz x4, #0, 55f\n"
-    "ld1 { v25.b }[2], [x25]\n"
+    "tbz x8, #1, 54f\n"
+    "ld1 { v25.h }[0], [x23], #0x2\n"
+    "tbz x8, #0, 55f\n"
+    "ld1 { v25.b }[2], [x23]\n"
     "b 55f\n"
     "54:"  // Oddments: Load (4, 0): Bit 2: Unset: Bit 1: Unset
-    "tbz x4, #0, 55f\n"
-    "ld1 { v25.b }[0], [x25]\n"
+    "tbz x8, #0, 55f\n"
+    "ld1 { v25.b }[0], [x23]\n"
     "55:"  // Oddments: Load (4, 0): Bit 2: End
-    "ldr x24, [x8, #0x98]\n"
     "usubl v25.8h, v25.8b, v12.8b\n"
-    "smlal v16.4s, v25.4h, v6.4h\n"
-    "smlal2 v22.4s, v25.8h, v6.8h\n"
-    "add x24, x24, x5\n"
-    "tbz x4, #2, 57f\n"
+    "ldr x24, [x12, #0x98]\n"
+    "smlal v22.4s, v25.4h, v6.4h\n"
+    "smlal2 v21.4s, v25.8h, v6.8h\n"
+    "add x24, x24, x15\n"
+    "tbz x8, #2, 57f\n"
     "ld1 { v29.s }[0], [x24], #0x4\n"
-    "tbz x4, #1, 56f\n"
+    "tbz x8, #1, 56f\n"
     "ld1 { v29.h }[2], [x24], #0x2\n"
-    "tbz x4, #0, 59f\n"
+    "tbz x8, #0, 59f\n"
     "ld1 { v29.b }[6], [x24]\n"
     "b 59f\n"
     "56:"  // Oddments: Load (2, 4): Bit 2: Bit 1: Unset
-    "tbz x4, #0, 59f\n"
+    "tbz x8, #0, 59f\n"
     "ld1 { v29.b }[4], [x24]\n"
     "b 59f\n"
     "57:"  // Oddments: Load (2, 4): Bit 2: Unset
-    "tbz x4, #1, 58f\n"
+    "tbz x8, #1, 58f\n"
     "ld1 { v29.h }[0], [x24], #0x2\n"
-    "tbz x4, #0, 59f\n"
+    "tbz x8, #0, 59f\n"
     "ld1 { v29.b }[2], [x24]\n"
     "b 59f\n"
     "58:"  // Oddments: Load (2, 4): Bit 2: Unset: Bit 1: Unset
-    "tbz x4, #0, 59f\n"
+    "tbz x8, #0, 59f\n"
     "ld1 { v29.b }[0], [x24]\n"
     "59:"  // Oddments: Load (2, 4): Bit 2: End
-    "ldr x23, [x8, #0xa0]\n"
     "usubl v29.8h, v29.8b, v12.8b\n"
-    "smlal v20.4s, v29.4h, v8.4h\n"
-    "smlal2 v23.4s, v29.8h, v8.8h\n"
-    "add x23, x23, x5\n"
-    "smlal v17.4s, v29.4h, v2.4h\n"
+    "ldr x19, [x12, #0xa0]\n"
+    "smlal v9.4s, v29.4h, v8.4h\n"
+    "smlal2 v16.4s, v29.8h, v8.8h\n"
+    "smlal v23.4s, v29.4h, v2.4h\n"
     "smlal2 v18.4s, v29.8h, v2.8h\n"
-    "tbz x4, #2, 61f\n"
-    "ld1 { v27.s }[0], [x23], #0x4\n"
-    "tbz x4, #1, 60f\n"
-    "ld1 { v27.h }[2], [x23], #0x2\n"
-    "tbz x4, #0, 63f\n"
-    "ld1 { v27.b }[6], [x23]\n"
+    "add x19, x19, x15\n"
+    "tbz x8, #2, 61f\n"
+    "ld1 { v27.s }[0], [x19], #0x4\n"
+    "tbz x8, #1, 60f\n"
+    "ld1 { v27.h }[2], [x19], #0x2\n"
+    "tbz x8, #0, 63f\n"
+    "ld1 { v27.b }[6], [x19]\n"
     "b 63f\n"
     "60:"  // Oddments: Load (4, 1): Bit 2: Bit 1: Unset
-    "tbz x4, #0, 63f\n"
-    "ld1 { v27.b }[4], [x23]\n"
+    "tbz x8, #0, 63f\n"
+    "ld1 { v27.b }[4], [x19]\n"
     "b 63f\n"
     "61:"  // Oddments: Load (4, 1): Bit 2: Unset
-    "tbz x4, #1, 62f\n"
-    "ld1 { v27.h }[0], [x23], #0x2\n"
-    "tbz x4, #0, 63f\n"
-    "ld1 { v27.b }[2], [x23]\n"
+    "tbz x8, #1, 62f\n"
+    "ld1 { v27.h }[0], [x19], #0x2\n"
+    "tbz x8, #0, 63f\n"
+    "ld1 { v27.b }[2], [x19]\n"
     "b 63f\n"
     "62:"  // Oddments: Load (4, 1): Bit 2: Unset: Bit 1: Unset
-    "tbz x4, #0, 63f\n"
-    "ld1 { v27.b }[0], [x23]\n"
+    "tbz x8, #0, 63f\n"
+    "ld1 { v27.b }[0], [x19]\n"
     "63:"  // Oddments: Load (4, 1): Bit 2: End
-    "ldr x22, [x8, #0xa8]\n"
     "usubl v27.8h, v27.8b, v12.8b\n"
-    "smlal v16.4s, v27.4h, v7.4h\n"
-    "smlal2 v22.4s, v27.8h, v7.8h\n"
-    "add x22, x22, x5\n"
-    "tbz x4, #2, 65f\n"
+    "ldr x22, [x12, #0xa8]\n"
+    "smlal v22.4s, v27.4h, v7.4h\n"
+    "smlal2 v21.4s, v27.8h, v7.8h\n"
+    "add x22, x22, x15\n"
+    "tbz x8, #2, 65f\n"
     "ld1 { v24.s }[0], [x22], #0x4\n"
-    "tbz x4, #1, 64f\n"
+    "tbz x8, #1, 64f\n"
     "ld1 { v24.h }[2], [x22], #0x2\n"
-    "tbz x4, #0, 67f\n"
+    "tbz x8, #0, 67f\n"
     "ld1 { v24.b }[6], [x22]\n"
     "b 67f\n"
     "64:"  // Oddments: Load (3, 2): Bit 2: Bit 1: Unset
-    "tbz x4, #0, 67f\n"
+    "tbz x8, #0, 67f\n"
     "ld1 { v24.b }[4], [x22]\n"
     "b 67f\n"
     "65:"  // Oddments: Load (3, 2): Bit 2: Unset
-    "tbz x4, #1, 66f\n"
+    "tbz x8, #1, 66f\n"
     "ld1 { v24.h }[0], [x22], #0x2\n"
-    "tbz x4, #0, 67f\n"
+    "tbz x8, #0, 67f\n"
     "ld1 { v24.b }[2], [x22]\n"
     "b 67f\n"
     "66:"  // Oddments: Load (3, 2): Bit 2: Unset: Bit 1: Unset
-    "tbz x4, #0, 67f\n"
+    "tbz x8, #0, 67f\n"
     "ld1 { v24.b }[0], [x22]\n"
     "67:"  // Oddments: Load (3, 2): Bit 2: End
-    "ldr x21, [x8, #0xb0]\n"
     "usubl v24.8h, v24.8b, v12.8b\n"
-    "smlal v16.4s, v24.4h, v5.4h\n"
-    "smlal2 v22.4s, v24.8h, v5.8h\n"
-    "add x21, x21, x5\n"
-    "smlal v17.4s, v24.4h, v3.4h\n"
+    "ldr x21, [x12, #0xb0]\n"
+    "smlal v22.4s, v24.4h, v5.4h\n"
+    "smlal2 v21.4s, v24.8h, v5.8h\n"
+    "smlal v23.4s, v24.4h, v3.4h\n"
     "smlal2 v18.4s, v24.8h, v3.8h\n"
-    "tbz x4, #2, 69f\n"
+    "add x21, x21, x15\n"
+    "tbz x8, #2, 69f\n"
     "ld1 { v26.s }[0], [x21], #0x4\n"
-    "tbz x4, #1, 68f\n"
+    "tbz x8, #1, 68f\n"
     "ld1 { v26.h }[2], [x21], #0x2\n"
-    "tbz x4, #0, 71f\n"
+    "tbz x8, #0, 71f\n"
     "ld1 { v26.b }[6], [x21]\n"
     "b 71f\n"
     "68:"  // Oddments: Load (4, 3): Bit 2: Bit 1: Unset
-    "tbz x4, #0, 71f\n"
+    "tbz x8, #0, 71f\n"
     "ld1 { v26.b }[4], [x21]\n"
     "b 71f\n"
     "69:"  // Oddments: Load (4, 3): Bit 2: Unset
-    "tbz x4, #1, 70f\n"
+    "tbz x8, #1, 70f\n"
     "ld1 { v26.h }[0], [x21], #0x2\n"
-    "tbz x4, #0, 71f\n"
+    "tbz x8, #0, 71f\n"
     "ld1 { v26.b }[2], [x21]\n"
     "b 71f\n"
     "70:"  // Oddments: Load (4, 3): Bit 2: Unset: Bit 1: Unset
-    "tbz x4, #0, 71f\n"
+    "tbz x8, #0, 71f\n"
     "ld1 { v26.b }[0], [x21]\n"
     "71:"  // Oddments: Load (4, 3): Bit 2: End
-    "ldr x20, [x8, #0xb8]\n"
     "usubl v26.8h, v26.8b, v12.8b\n"
-    "smlal v17.4s, v26.4h, v7.4h\n"
+    "ldr x20, [x12, #0xb8]\n"
+    "smlal v23.4s, v26.4h, v7.4h\n"
     "smlal2 v18.4s, v26.8h, v7.8h\n"
-    "add x20, x20, x5\n"
-    "tbz x4, #2, 73f\n"
+    "add x20, x20, x15\n"
+    "tbz x8, #2, 73f\n"
     "ld1 { v25.s }[0], [x20], #0x4\n"
-    "tbz x4, #1, 72f\n"
+    "tbz x8, #1, 72f\n"
     "ld1 { v25.h }[2], [x20], #0x2\n"
-    "tbz x4, #0, 75f\n"
+    "tbz x8, #0, 75f\n"
     "ld1 { v25.b }[6], [x20]\n"
     "b 75f\n"
     "72:"  // Oddments: Load (4, 2): Bit 2: Bit 1: Unset
-    "tbz x4, #0, 75f\n"
+    "tbz x8, #0, 75f\n"
     "ld1 { v25.b }[4], [x20]\n"
     "b 75f\n"
     "73:"  // Oddments: Load (4, 2): Bit 2: Unset
-    "tbz x4, #1, 74f\n"
+    "tbz x8, #1, 74f\n"
     "ld1 { v25.h }[0], [x20], #0x2\n"
-    "tbz x4, #0, 75f\n"
+    "tbz x8, #0, 75f\n"
     "ld1 { v25.b }[2], [x20]\n"
     "b 75f\n"
     "74:"  // Oddments: Load (4, 2): Bit 2: Unset: Bit 1: Unset
-    "tbz x4, #0, 75f\n"
+    "tbz x8, #0, 75f\n"
     "ld1 { v25.b }[0], [x20]\n"
     "75:"  // Oddments: Load (4, 2): Bit 2: End
-    "ldr x19, [x8, #0xc0]\n"
     "usubl v25.8h, v25.8b, v12.8b\n"
-    "smlal v16.4s, v25.4h, v8.4h\n"
-    "smlal2 v22.4s, v25.8h, v8.8h\n"
-    "add x19, x19, x5\n"
-    "smlal v17.4s, v25.4h, v6.4h\n"
+    "ldr x19, [x12, #0xc0]\n"
+    "smlal v22.4s, v25.4h, v8.4h\n"
+    "smlal2 v21.4s, v25.8h, v8.8h\n"
+    "smlal v23.4s, v25.4h, v6.4h\n"
     "smlal2 v18.4s, v25.8h, v6.8h\n"
-    "tbz x4, #2, 77f\n"
+    "add x19, x19, x15\n"
+    "tbz x8, #2, 77f\n"
     "ld1 { v29.s }[0], [x19], #0x4\n"
-    "tbz x4, #1, 76f\n"
+    "tbz x8, #1, 76f\n"
     "ld1 { v29.h }[2], [x19], #0x2\n"
-    "tbz x4, #0, 79f\n"
+    "tbz x8, #0, 79f\n"
     "ld1 { v29.b }[6], [x19]\n"
     "b 79f\n"
     "76:"  // Oddments: Load (4, 4): Bit 2: Bit 1: Unset
-    "tbz x4, #0, 79f\n"
+    "tbz x8, #0, 79f\n"
     "ld1 { v29.b }[4], [x19]\n"
     "b 79f\n"
     "77:"  // Oddments: Load (4, 4): Bit 2: Unset
-    "tbz x4, #1, 78f\n"
+    "tbz x8, #1, 78f\n"
     "ld1 { v29.h }[0], [x19], #0x2\n"
-    "tbz x4, #0, 79f\n"
+    "tbz x8, #0, 79f\n"
     "ld1 { v29.b }[2], [x19]\n"
     "b 79f\n"
     "78:"  // Oddments: Load (4, 4): Bit 2: Unset: Bit 1: Unset
-    "tbz x4, #0, 79f\n"
+    "tbz x8, #0, 79f\n"
     "ld1 { v29.b }[0], [x19]\n"
     "79:"  // Oddments: Load (4, 4): Bit 2: End
     "usubl v29.8h, v29.8b, v12.8b\n"
-    "smlal v17.4s, v29.4h, v8.4h\n"
+    "smlal v23.4s, v29.4h, v8.4h\n"
     "smlal2 v18.4s, v29.8h, v8.8h\n"
-    "tbz x4, #2, 81f\n"
-    "ld1 { v21.4s }, [x17], #0x10\n"
-    "ld1 { v30.4s }, [x15], #0x10\n"
-    "tbz x4, #1, 80f\n"
-    "ld1 { v31.d }[0], [x17], #0x8\n"
-    "ld1 { v9.d }[0], [x15], #0x8\n"
-    "tbz x4, #0, 83f\n"
-    "ld1 { v31.s }[2], [x17]\n"
-    "ld1 { v9.s }[2], [x15]\n"
+    "tbz x8, #2, 81f\n"
+    "ld1 { v19.4s }, [x13], #0x10\n"
+    "ld1 { v0.4s }, [x11], #0x10\n"
+    "tbz x8, #1, 80f\n"
+    "ld1 { v4.d }[0], [x13], #0x8\n"
+    "ld1 { v31.d }[0], [x11], #0x8\n"
+    "tbz x8, #0, 83f\n"
+    "ld1 { v4.s }[2], [x13]\n"
+    "ld1 { v31.s }[2], [x11]\n"
     "b 83f\n"
     "80:"  // Oddments: Load requant params: Bit 2: Bit 1: Unset
-    "tbz x4, #0, 83f\n"
-    "ld1 { v31.s }[0], [x17]\n"
-    "ld1 { v9.s }[0], [x15]\n"
+    "tbz x8, #0, 83f\n"
+    "ld1 { v4.s }[0], [x13]\n"
+    "ld1 { v31.s }[0], [x11]\n"
     "b 83f\n"
     "81:"  // Oddments: Load requant params: Bit 2: Unset
-    "tbz x4, #1, 82f\n"
-    "ld1 { v21.d }[0], [x17], #0x8\n"
-    "ld1 { v30.d }[0], [x15], #0x8\n"
-    "tbz x4, #0, 83f\n"
-    "ld1 { v21.s }[2], [x17]\n"
-    "ld1 { v30.s }[2], [x15]\n"
+    "tbz x8, #1, 82f\n"
+    "ld1 { v19.d }[0], [x13], #0x8\n"
+    "ld1 { v0.d }[0], [x11], #0x8\n"
+    "tbz x8, #0, 83f\n"
+    "ld1 { v19.s }[2], [x13]\n"
+    "ld1 { v0.s }[2], [x11]\n"
     "b 83f\n"
     "82:"  // Oddments: Load requant params: Bit 2: Unset: Bit 1: Unset
-    "tbz x4, #0, 83f\n"
-    "ld1 { v21.s }[0], [x17]\n"
-    "ld1 { v30.s }[0], [x15]\n"
+    "tbz x8, #0, 83f\n"
+    "ld1 { v19.s }[0], [x13]\n"
+    "ld1 { v0.s }[0], [x11]\n"
     "83:"  // Oddments: Load requant params: Bit 2: End
-    "sqrdmulh v15.4s, v15.4s, v21.4s\n"
-    "add x14, x14, x7\n"
-    "sqrdmulh v10.4s, v10.4s, v31.4s\n"
-    "add x13, x13, x7\n"
-    "sqrdmulh v20.4s, v20.4s, v21.4s\n"
-    "add x12, x12, x7\n"
-    "sqrdmulh v23.4s, v23.4s, v31.4s\n"
-    "add x11, x11, x7\n"
-    "sqrdmulh v16.4s, v16.4s, v21.4s\n"
-    "and v26.16b, v15.16b, v30.16b\n"
-    "sshr v26.4s, v26.4s, #0x1f\n"
-    "and v8.16b, v10.16b, v9.16b\n"
-    "and v4.16b, v20.16b, v30.16b\n"
-    "sshr v8.4s, v8.4s, #0x1f\n"
-    "and v2.16b, v23.16b, v9.16b\n"
-    "and v1.16b, v16.16b, v30.16b\n"
+    "sqdmulh v15.4s, v15.4s, v19.4s\n"
+    "sqdmulh v9.4s, v9.4s, v19.4s\n"
+    "add x10, x10, x14\n"
+    "add x9, x9, x14\n"
+    "sqdmulh v22.4s, v22.4s, v19.4s\n"
+    "sqdmulh v23.4s, v23.4s, v19.4s\n"
+    "add x28, x28, x14\n"
+    "add x27, x27, x14\n"
+    "and v30.16b, v15.16b, v0.16b\n"
+    "sqdmulh v10.4s, v10.4s, v4.4s\n"
+    "and v28.16b, v9.16b, v0.16b\n"
+    "sqdmulh v16.4s, v16.4s, v4.4s\n"
+    "and v29.16b, v22.16b, v0.16b\n"
+    "sqdmulh v21.4s, v21.4s, v4.4s\n"
+    "and v20.16b, v23.16b, v0.16b\n"
+    "sqdmulh v18.4s, v18.4s, v4.4s\n"
+    "sshr v30.4s, v30.4s, #0x1f\n"
+    "and v19.16b, v10.16b, v31.16b\n"
+    "sshr v28.4s, v28.4s, #0x1f\n"
+    "and v4.16b, v16.16b, v31.16b\n"
+    "sshr v29.4s, v29.4s, #0x1f\n"
+    "and v5.16b, v21.16b, v31.16b\n"
+    "sshr v20.4s, v20.4s, #0x1f\n"
+    "and v26.16b, v18.16b, v31.16b\n"
+    "sqadd v15.4s, v15.4s, v30.4s\n"
+    "sshr v19.4s, v19.4s, #0x1f\n"
+    "sqadd v9.4s, v9.4s, v28.4s\n"
     "sshr v4.4s, v4.4s, #0x1f\n"
-    "sqrdmulh v22.4s, v22.4s, v31.4s\n"
-    "sshr v2.4s, v2.4s, #0x1f\n"
-    "sqadd v15.4s, v15.4s, v26.4s\n"
-    "sqrdmulh v17.4s, v17.4s, v21.4s\n"
-    "sshr v1.4s, v1.4s, #0x1f\n"
-    "sqrdmulh v18.4s, v18.4s, v31.4s\n"
-    "sqadd v10.4s, v10.4s, v8.4s\n"
-    "sqadd v20.4s, v20.4s, v4.4s\n"
-    "srshl v15.4s, v15.4s, v30.4s\n"
-    "sqadd v23.4s, v23.4s, v2.4s\n"
-    "srshl v10.4s, v10.4s, v9.4s\n"
-    "srshl v20.4s, v20.4s, v30.4s\n"
-    "add v15.4s, v15.4s, v11.4s\n"
-    "srshl v23.4s, v23.4s, v9.4s\n"
-    "add v10.4s, v10.4s, v11.4s\n"
-    "smin v15.4s, v15.4s, v14.4s\n"
-    "add v20.4s, v20.4s, v11.4s\n"
-    "smin v10.4s, v10.4s, v14.4s\n"
-    "smax v15.4s, v15.4s, v19.4s\n"
-    "smin v20.4s, v20.4s, v14.4s\n"
-    "smax v10.4s, v10.4s, v19.4s\n"
-    "add v23.4s, v23.4s, v11.4s\n"
-    "smax v20.4s, v20.4s, v19.4s\n"
-    "uzp1 v15.16b, v15.16b, v10.16b\n"
-    "smin v23.4s, v23.4s, v14.4s\n"
+    "sqadd v22.4s, v22.4s, v29.4s\n"
+    "sshr v5.4s, v5.4s, #0x1f\n"
+    "sqadd v23.4s, v23.4s, v20.4s\n"
+    "sshr v26.4s, v26.4s, #0x1f\n"
+    "srshl v15.4s, v15.4s, v0.4s\n"
+    "sqadd v10.4s, v10.4s, v19.4s\n"
+    "srshl v9.4s, v9.4s, v0.4s\n"
+    "sqadd v16.4s, v16.4s, v4.4s\n"
+    "srshl v22.4s, v22.4s, v0.4s\n"
+    "sqadd v21.4s, v21.4s, v5.4s\n"
+    "srshl v23.4s, v23.4s, v0.4s\n"
+    "sqadd v18.4s, v18.4s, v26.4s\n"
+    "srshl v10.4s, v10.4s, v31.4s\n"
+    "sqxtn v15.4h, v15.4s\n"
+    "srshl v16.4s, v16.4s, v31.4s\n"
+    "sqxtn v9.4h, v9.4s\n"
+    "srshl v21.4s, v21.4s, v31.4s\n"
+    "sqxtn v22.4h, v22.4s\n"
+    "srshl v18.4s, v18.4s, v31.4s\n"
+    "sqxtn v23.4h, v23.4s\n"
+    "sqxtn2 v15.8h, v10.4s\n"
+    "sqxtn2 v9.8h, v16.4s\n"
+    "sqxtn2 v22.8h, v21.4s\n"
+    "sqxtn2 v23.8h, v18.4s\n"
+    "sqadd v15.8h, v15.8h, v11.8h\n"
+    "sqadd v9.8h, v9.8h, v11.8h\n"
+    "sqadd v22.8h, v22.8h, v11.8h\n"
+    "sqadd v23.8h, v23.8h, v11.8h\n"
+    "smax v15.8h, v15.8h, v17.8h\n"
+    "smax v9.8h, v9.8h, v17.8h\n"
+    "smax v22.8h, v22.8h, v17.8h\n"
+    "smax v23.8h, v23.8h, v17.8h\n"
+    "smin v15.8h, v15.8h, v14.8h\n"
+    "smin v9.8h, v9.8h, v14.8h\n"
+    "smin v22.8h, v22.8h, v14.8h\n"
+    "smin v23.8h, v23.8h, v14.8h\n"
     "uzp1 v15.16b, v15.16b, v15.16b\n"
-    "sqadd v16.4s, v16.4s, v1.4s\n"
-    "smax v23.4s, v23.4s, v19.4s\n"
-    "and v24.16b, v22.16b, v9.16b\n"
-    "sshr v24.4s, v24.4s, #0x1f\n"
-    "uzp1 v20.16b, v20.16b, v23.16b\n"
-    "srshl v16.4s, v16.4s, v30.4s\n"
-    "and v2.16b, v17.16b, v30.16b\n"
-    "sshr v2.4s, v2.4s, #0x1f\n"
-    "uzp1 v20.16b, v20.16b, v20.16b\n"
-    "add v16.4s, v16.4s, v11.4s\n"
-    "sqadd v22.4s, v22.4s, v24.4s\n"
-    "and v31.16b, v18.16b, v9.16b\n"
-    "sshr v31.4s, v31.4s, #0x1f\n"
-    "smin v16.4s, v16.4s, v14.4s\n"
-    "srshl v22.4s, v22.4s, v9.4s\n"
-    "sqadd v17.4s, v17.4s, v2.4s\n"
-    "smax v16.4s, v16.4s, v19.4s\n"
-    "add v22.4s, v22.4s, v11.4s\n"
-    "srshl v17.4s, v17.4s, v30.4s\n"
-    "sqadd v18.4s, v18.4s, v31.4s\n"
-    "smin v22.4s, v22.4s, v14.4s\n"
-    "add v17.4s, v17.4s, v11.4s\n"
-    "srshl v18.4s, v18.4s, v9.4s\n"
-    "smax v22.4s, v22.4s, v19.4s\n"
-    "smin v17.4s, v17.4s, v14.4s\n"
-    "uzp1 v16.16b, v16.16b, v22.16b\n"
-    "add v18.4s, v18.4s, v11.4s\n"
-    "uzp1 v16.16b, v16.16b, v16.16b\n"
-    "smax v17.4s, v17.4s, v19.4s\n"
-    "smin v18.4s, v18.4s, v14.4s\n"
-    "smax v18.4s, v18.4s, v19.4s\n"
-    "uzp1 v17.16b, v17.16b, v18.16b\n"
-    "uzp1 v17.16b, v17.16b, v17.16b\n"
-    "tbz x4, #2, 85f\n"
-    "st1 { v15.s }[0], [x14], #0x4\n"
-    "st1 { v20.s }[0], [x13], #0x4\n"
-    "st1 { v16.s }[0], [x12], #0x4\n"
-    "st1 { v17.s }[0], [x11], #0x4\n"
-    "tbz x4, #1, 84f\n"
-    "st1 { v15.h }[2], [x14], #0x2\n"
-    "st1 { v20.h }[2], [x13], #0x2\n"
-    "st1 { v16.h }[2], [x12], #0x2\n"
-    "st1 { v17.h }[2], [x11], #0x2\n"
-    "tbz x4, #0, 87f\n"
-    "st1 { v15.b }[6], [x14], #0x1\n"
-    "st1 { v20.b }[6], [x13], #0x1\n"
-    "st1 { v16.b }[6], [x12], #0x1\n"
-    "st1 { v17.b }[6], [x11], #0x1\n"
+    "uzp1 v9.16b, v9.16b, v9.16b\n"
+    "uzp1 v22.16b, v22.16b, v22.16b\n"
+    "uzp1 v23.16b, v23.16b, v23.16b\n"
+    "tbz x8, #2, 85f\n"
+    "st1 { v15.s }[0], [x10], #0x4\n"
+    "st1 { v9.s }[0], [x9], #0x4\n"
+    "st1 { v22.s }[0], [x28], #0x4\n"
+    "st1 { v23.s }[0], [x27], #0x4\n"
+    "tbz x8, #1, 84f\n"
+    "st1 { v15.h }[2], [x10], #0x2\n"
+    "st1 { v9.h }[2], [x9], #0x2\n"
+    "st1 { v22.h }[2], [x28], #0x2\n"
+    "st1 { v23.h }[2], [x27], #0x2\n"
+    "tbz x8, #0, 87f\n"
+    "st1 { v15.b }[6], [x10], #0x1\n"
+    "st1 { v9.b }[6], [x9], #0x1\n"
+    "st1 { v22.b }[6], [x28], #0x1\n"
+    "st1 { v23.b }[6], [x27], #0x1\n"
     "b 87f\n"
     "84:"  // Oddments: Bit 2: Bit 1: Unset
-    "tbz x4, #0, 87f\n"
-    "st1 { v15.b }[4], [x14], #0x1\n"
-    "st1 { v20.b }[4], [x13], #0x1\n"
-    "st1 { v16.b }[4], [x12], #0x1\n"
-    "st1 { v17.b }[4], [x11], #0x1\n"
+    "tbz x8, #0, 87f\n"
+    "st1 { v15.b }[4], [x10], #0x1\n"
+    "st1 { v9.b }[4], [x9], #0x1\n"
+    "st1 { v22.b }[4], [x28], #0x1\n"
+    "st1 { v23.b }[4], [x27], #0x1\n"
     "b 87f\n"
     "85:"  // Oddments: Bit 2: Unset
-    "tbz x4, #1, 86f\n"
-    "st1 { v15.h }[0], [x14], #0x2\n"
-    "st1 { v20.h }[0], [x13], #0x2\n"
-    "st1 { v16.h }[0], [x12], #0x2\n"
-    "st1 { v17.h }[0], [x11], #0x2\n"
-    "tbz x4, #0, 87f\n"
-    "st1 { v15.b }[2], [x14], #0x1\n"
-    "st1 { v20.b }[2], [x13], #0x1\n"
-    "st1 { v16.b }[2], [x12], #0x1\n"
-    "st1 { v17.b }[2], [x11], #0x1\n"
+    "tbz x8, #1, 86f\n"
+    "st1 { v15.h }[0], [x10], #0x2\n"
+    "st1 { v9.h }[0], [x9], #0x2\n"
+    "st1 { v22.h }[0], [x28], #0x2\n"
+    "st1 { v23.h }[0], [x27], #0x2\n"
+    "tbz x8, #0, 87f\n"
+    "st1 { v15.b }[2], [x10], #0x1\n"
+    "st1 { v9.b }[2], [x9], #0x1\n"
+    "st1 { v22.b }[2], [x28], #0x1\n"
+    "st1 { v23.b }[2], [x27], #0x1\n"
     "b 87f\n"
     "86:"  // Oddments: Bit 2: Unset: Bit 1: Unset
-    "tbz x4, #0, 87f\n"
-    "st1 { v15.b }[0], [x14], #0x1\n"
-    "st1 { v20.b }[0], [x13], #0x1\n"
-    "st1 { v16.b }[0], [x12], #0x1\n"
-    "st1 { v17.b }[0], [x11], #0x1\n"
+    "tbz x8, #0, 87f\n"
+    "st1 { v15.b }[0], [x10], #0x1\n"
+    "st1 { v9.b }[0], [x9], #0x1\n"
+    "st1 { v22.b }[0], [x28], #0x1\n"
+    "st1 { v23.b }[0], [x27], #0x1\n"
     "87:"  // Oddments: Bit 2: End
-
     "88:"  // End
-
     :
     : [offsetof_Params_bias] "I" (offsetof(Params, bias)), [offsetof_Params_inptrs] "I" (offsetof(Params, inptrs)), [offsetof_Params_n_channels] "I" (offsetof(Params, n_channels)), [offsetof_Params_outptrs] "I" (offsetof(Params, outptrs)), [offsetof_Params_requant] "I" (offsetof(Params, requant)), [offsetof_Params_requant_muls] "I" (offsetof(Params, requant_muls)), [offsetof_Params_requant_shifts] "I" (offsetof(Params, requant_shifts)), [offsetof_Params_weights] "I" (offsetof(Params, weights)), [offsetof_Requantize32_a_offset] "I" (offsetof(arm_gemm::Requantize32, a_offset)), [offsetof_Requantize32_b_offset] "I" (offsetof(arm_gemm::Requantize32, b_offset)), [offsetof_Requantize32_c_offset] "I" (offsetof(arm_gemm::Requantize32, c_offset)), [offsetof_Requantize32_maxval] "I" (offsetof(arm_gemm::Requantize32, maxval)), [offsetof_Requantize32_minval] "I" (offsetof(arm_gemm::Requantize32, minval)), [params] "r" (&params)
-    : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x4", "x5", "x6", "x7", "x8", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x16", "x17", "x19", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28"
+    : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x8", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x16", "x17", "x19", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28"
   );
 }