Update cpu depthwise kernels

Resolves: COMPMID-4688

Signed-off-by: Freddie Liardet <frederick.liardet@arm.com>
Change-Id: I9e22f967f5b7ccaebff2fc49f0253f621d62d820
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/6030
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Pablo Marquez Tello <pablo.tello@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>

1 files changed, 144 insertions, 152 deletions

diff --git a/src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp16_nhwc_5x5_s1_output2x2_mla_depthfirst/generic_direct.cpp b/src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp16_nhwc_5x5_s1_output2x2_mla_depthfirst/generic_direct.cpp
index 6b1564e6c9..1c2e1e27ad 100644
--- a/src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp16_nhwc_5x5_s1_output2x2_mla_depthfirst/generic_direct.cpp
+++ b/src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp16_nhwc_5x5_s1_output2x2_mla_depthfirst/generic_direct.cpp
@@ -25,7 +25,7 @@
 #include <cstddef>
 #include <cstdint>
 
-#if defined(ARM_COMPUTE_ENABLE_SVE) && defined(__ARM_FP16_ARGS)
+#if __aarch64__ && defined(ARM_COMPUTE_ENABLE_SVE) && defined(__ARM_FP16_ARGS)
 
 namespace arm_conv {
 namespace depthwise {
@@ -113,390 +113,382 @@ void sve_fp16_nhwc_5x5_s1_output2x2_mla_depthfirst_direct_impl(
     "add x20, x13, x22, LSL #1\n"
     "ld1rh { z17.h }, p3/Z, [%x[params_struct], %[offsetof_args_max]]\n"
     "add x10, x20, x22, LSL #1\n"
-    "ld1h { z16.h }, p3/Z, [x8]\n" // Load from weights and bias
-    "mov z31.d, z16.d\n"
-    "ld1h { z0.h }, p3/Z, [x8, #1, MUL VL]\n" // Load from weights and bias
+    "ld1h { z16.h }, p3/Z, [x8]\n"
     "add x9, x10, x22, LSL #1\n"
-    "mov z30.d, z16.d\n"
-    "ld1h { z1.h }, p3/Z, [x8, #2, MUL VL]\n" // Load from weights and bias
+    "ld1h { z0.h }, p3/Z, [x8, #1, MUL VL]\n"
     "add x28, x9, x22, LSL #1\n"
-    "mov z29.d, z16.d\n"
-    "ld1h { z2.h }, p3/Z, [x8, #3, MUL VL]\n" // Load from weights and bias
+    "ld1h { z1.h }, p3/Z, [x8, #2, MUL VL]\n"
     "add x27, x28, x22, LSL #1\n"
-    "mov z28.d, z16.d\n"
-    "ld1h { z3.h }, p3/Z, [x8, #4, MUL VL]\n" // Load from weights and bias
+    "ld1h { z2.h }, p3/Z, [x8, #3, MUL VL]\n"
     "add x26, x15, x15\n"
-    "ld1h { z4.h }, p3/Z, [x8, #5, MUL VL]\n" // Load from weights and bias
+    "ld1h { z3.h }, p3/Z, [x8, #4, MUL VL]\n"
     "add x25, x26, x15\n"
-    "mul x19, x5, x21\n" // offset = tile_i * ld_output_row
+    "ld1h { z4.h }, p3/Z, [x8, #5, MUL VL]\n"
     "add x24, x25, x15\n"
+    "mul x19, x5, x21\n" // offset = tile_i * ld_output_row
     "add x23, x24, x15\n"
     "madd x19, x6, x12, x19\n" // offset += tile_j * ld_output_col
     "mul x19, x19, x7\n" // offset *= output_tile_size
     "add x11, x11, x19, LSL #1\n" // outptrs[0] += offset * sizeof(__fp16)
     "add x22, x11, x21, LSL #1\n"
     "whilelt p2.h, XZR, %x[n_channels]\n"
-    "ld1h { z5.h }, p2/Z, [x13]\n" // Load input point (0, 0)
-    "ld1h { z6.h }, p2/Z, [x13, x15, LSL #1]\n" // Load input point (0, 1)
+    "ld1h { z5.h }, p2/Z, [x13]\n"
+    "ld1h { z6.h }, p2/Z, [x13, x15, LSL #1]\n"
     "cmp x16, %x[n_channels]\n"
-    "ld1h { z7.h }, p2/Z, [x20]\n" // Load input point (1, 0)
+    "ld1h { z7.h }, p2/Z, [x20]\n"
     "addvl x8, x8, #6\n"
-    "ld1h { z8.h }, p2/Z, [x20, x15, LSL #1]\n" // Load input point (1, 1)
-    "ld1h { z9.h }, p2/Z, [x13, x26, LSL #1]\n" // Load input point (0, 2)
-    "ld1h { z13.h }, p2/Z, [x20, x26, LSL #1]\n" // Load input point (1, 2)
-    "ld1h { z11.h }, p2/Z, [x13, x25, LSL #1]\n" // Load input point (0, 3)
-    "ld1h { z12.h }, p2/Z, [x13, x24, LSL #1]\n" // Load input point (0, 4)
-    "ld1h { z10.h }, p2/Z, [x20, x23, LSL #1]\n" // Load input point (1, 5)
-    "ld1h { z14.h }, p2/Z, [x10]\n" // Load input point (2, 0)
+    "ld1h { z8.h }, p2/Z, [x20, x15, LSL #1]\n"
+    "ld1h { z9.h }, p2/Z, [x13, x26, LSL #1]\n"
+    "ld1h { z13.h }, p2/Z, [x20, x26, LSL #1]\n"
+    "ld1h { z11.h }, p2/Z, [x13, x25, LSL #1]\n"
+    "ld1h { z12.h }, p2/Z, [x13, x24, LSL #1]\n"
+    "ld1h { z10.h }, p2/Z, [x20, x23, LSL #1]\n"
+    "ld1h { z14.h }, p2/Z, [x10]\n"
     "bge 3f\n"
     "2:"  // Tile loop: Channel loop
-    "fmla z31.h, p3/M, z0.h, z5.h\n"
-    "ld1h { z5.h }, p2/Z, [x20, x25, LSL #1]\n" // Load input point (1, 3)
+    "movprfx z31, z16\n fmla z31.h, p3/M, z0.h, z5.h\n"
+    "ld1h { z5.h }, p2/Z, [x20, x25, LSL #1]\n"
     "whilelt p1.h, x16, %x[n_channels]\n"
-    "fmla z30.h, p3/M, z0.h, z6.h\n"
+    "movprfx z30, z16\n fmla z30.h, p3/M, z0.h, z6.h\n"
     "inch x14\n"
-    "fmla z29.h, p3/M, z0.h, z7.h\n"
+    "movprfx z29, z16\n fmla z29.h, p3/M, z0.h, z7.h\n"
     "mov p0.b, p2.b\n"
-    "fmla z28.h, p3/M, z0.h, z8.h\n"
-    "ld1h { z0.h }, p3/Z, [x8]\n" // Load from weights and bias
+    "movprfx z28, z16\n fmla z28.h, p3/M, z0.h, z8.h\n"
+    "ld1h { z0.h }, p3/Z, [x8]\n"
     "inch x17\n"
     "fmla z31.h, p3/M, z1.h, z6.h\n"
-    "ld1h { z6.h }, p2/Z, [x20, x24, LSL #1]\n" // Load input point (1, 4)
+    "ld1h { z6.h }, p2/Z, [x20, x24, LSL #1]\n"
     "addvl x20, x20, #1\n"
     "fmla z30.h, p3/M, z1.h, z9.h\n"
     "inch x16\n"
     "fmla z29.h, p3/M, z1.h, z8.h\n"
     "fmla z28.h, p3/M, z1.h, z13.h\n"
-    "ld1h { z1.h }, p3/Z, [x8, #1, MUL VL]\n" // Load from weights and bias
+    "ld1h { z1.h }, p3/Z, [x8, #1, MUL VL]\n"
     "fmla z31.h, p3/M, z2.h, z9.h\n"
-    "ld1h { z9.h }, p2/Z, [x13, x23, LSL #1]\n" // Load input point (0, 5)
+    "ld1h { z9.h }, p2/Z, [x13, x23, LSL #1]\n"
     "addvl x13, x13, #1\n"
     "fmla z30.h, p3/M, z2.h, z11.h\n"
     "fmla z29.h, p3/M, z2.h, z13.h\n"
     "fmla z28.h, p3/M, z2.h, z5.h\n"
-    "ld1h { z2.h }, p3/Z, [x8, #2, MUL VL]\n" // Load from weights and bias
+    "ld1h { z2.h }, p3/Z, [x8, #2, MUL VL]\n"
     "fmla z31.h, p3/M, z3.h, z11.h\n"
-    "ld1h { z11.h }, p2/Z, [x10, x15, LSL #1]\n" // Load input point (2, 1)
+    "ld1h { z11.h }, p2/Z, [x10, x15, LSL #1]\n"
     "fmla z30.h, p3/M, z3.h, z12.h\n"
     "fmla z29.h, p3/M, z3.h, z5.h\n"
     "fmla z28.h, p3/M, z3.h, z6.h\n"
-    "ld1h { z3.h }, p3/Z, [x8, #3, MUL VL]\n" // Load from weights and bias
+    "ld1h { z3.h }, p3/Z, [x8, #3, MUL VL]\n"
     "fmla z31.h, p3/M, z4.h, z12.h\n"
-    "ld1h { z12.h }, p2/Z, [x10, x26, LSL #1]\n" // Load input point (2, 2)
+    "ld1h { z12.h }, p2/Z, [x10, x26, LSL #1]\n"
     "fmla z30.h, p3/M, z4.h, z9.h\n"
-    "ld1h { z9.h }, p2/Z, [x10, x25, LSL #1]\n" // Load input point (2, 3)
+    "ld1h { z9.h }, p2/Z, [x10, x25, LSL #1]\n"
     "fmla z29.h, p3/M, z4.h, z6.h\n"
     "fmla z28.h, p3/M, z4.h, z10.h\n"
-    "ld1h { z4.h }, p3/Z, [x8, #4, MUL VL]\n" // Load from weights and bias
+    "ld1h { z4.h }, p3/Z, [x8, #4, MUL VL]\n"
     "fmla z31.h, p3/M, z0.h, z7.h\n"
-    "ld1h { z7.h }, p1/Z, [x20]\n" // Load input point (1, 0)
+    "ld1h { z7.h }, p1/Z, [x20]\n"
     "fmla z30.h, p3/M, z0.h, z8.h\n"
     "fmla z29.h, p3/M, z0.h, z14.h\n"
     "fmla z28.h, p3/M, z0.h, z11.h\n"
-    "ld1h { z0.h }, p3/Z, [x8, #5, MUL VL]\n" // Load from weights and bias
+    "ld1h { z0.h }, p3/Z, [x8, #5, MUL VL]\n"
     "fmla z31.h, p3/M, z1.h, z8.h\n"
-    "ld1h { z8.h }, p2/Z, [x10, x23, LSL #1]\n" // Load input point (2, 5)
+    "ld1h { z8.h }, p2/Z, [x10, x23, LSL #1]\n"
     "fmla z30.h, p3/M, z1.h, z13.h\n"
     "fmla z29.h, p3/M, z1.h, z11.h\n"
     "fmla z28.h, p3/M, z1.h, z12.h\n"
-    "ld1h { z1.h }, p3/Z, [x8, #6, MUL VL]\n" // Load from weights and bias
+    "ld1h { z1.h }, p3/Z, [x8, #6, MUL VL]\n"
     "fmla z31.h, p3/M, z2.h, z13.h\n"
-    "ld1h { z13.h }, p2/Z, [x10, x24, LSL #1]\n" // Load input point (2, 4)
+    "ld1h { z13.h }, p2/Z, [x10, x24, LSL #1]\n"
     "addvl x10, x10, #1\n"
     "fmla z30.h, p3/M, z2.h, z5.h\n"
     "fmla z29.h, p3/M, z2.h, z12.h\n"
     "fmla z28.h, p3/M, z2.h, z9.h\n"
-    "ld1h { z2.h }, p3/Z, [x8, #7, MUL VL]\n" // Load from weights and bias
+    "ld1h { z2.h }, p3/Z, [x8, #7, MUL VL]\n"
     "addvl x8, x8, #16\n"
     "fmla z31.h, p3/M, z3.h, z5.h\n"
-    "ld1h { z5.h }, p2/Z, [x9]\n" // Load input point (3, 0)
-    "ld1h { z16.h }, p3/Z, [x8, #4, MUL VL]\n" // Load from weights and bias
+    "ld1h { z5.h }, p2/Z, [x9]\n"
+    "ld1h { z16.h }, p3/Z, [x8, #4, MUL VL]\n"
     "fmla z30.h, p3/M, z3.h, z6.h\n"
     "fmla z29.h, p3/M, z3.h, z9.h\n"
     "fmla z28.h, p3/M, z3.h, z13.h\n"
-    "ld1h { z3.h }, p3/Z, [x8, #-8, MUL VL]\n" // Load from weights and bias
+    "ld1h { z3.h }, p3/Z, [x8, #-8, MUL VL]\n"
     "fmla z31.h, p3/M, z4.h, z6.h\n"
-    "ld1h { z6.h }, p2/Z, [x9, x15, LSL #1]\n" // Load input point (3, 1)
+    "ld1h { z6.h }, p2/Z, [x9, x15, LSL #1]\n"
     "fmla z30.h, p3/M, z4.h, z10.h\n"
-    "ld1h { z10.h }, p2/Z, [x9, x26, LSL #1]\n" // Load input point (3, 2)
+    "ld1h { z10.h }, p2/Z, [x9, x26, LSL #1]\n"
     "fmla z29.h, p3/M, z4.h, z13.h\n"
     "fmla z28.h, p3/M, z4.h, z8.h\n"
-    "ld1h { z4.h }, p3/Z, [x8, #-7, MUL VL]\n" // Load from weights and bias
+    "ld1h { z4.h }, p3/Z, [x8, #-7, MUL VL]\n"
     "fmla z31.h, p3/M, z0.h, z14.h\n"
-    "ld1h { z14.h }, p2/Z, [x9, x23, LSL #1]\n" // Load input point (3, 5)
+    "ld1h { z14.h }, p2/Z, [x9, x23, LSL #1]\n"
     "fmla z30.h, p3/M, z0.h, z11.h\n"
     "fmla z29.h, p3/M, z0.h, z5.h\n"
     "fmla z28.h, p3/M, z0.h, z6.h\n"
-    "ld1h { z0.h }, p3/Z, [x8, #-6, MUL VL]\n" // Load from weights and bias
+    "ld1h { z0.h }, p3/Z, [x8, #-6, MUL VL]\n"
     "fmla z31.h, p3/M, z1.h, z11.h\n"
-    "ld1h { z11.h }, p2/Z, [x9, x25, LSL #1]\n" // Load input point (3, 3)
+    "ld1h { z11.h }, p2/Z, [x9, x25, LSL #1]\n"
     "fmla z30.h, p3/M, z1.h, z12.h\n"
     "fmla z29.h, p3/M, z1.h, z6.h\n"
     "fmla z28.h, p3/M, z1.h, z10.h\n"
-    "ld1h { z1.h }, p3/Z, [x8, #-5, MUL VL]\n" // Load from weights and bias
+    "ld1h { z1.h }, p3/Z, [x8, #-5, MUL VL]\n"
     "fmla z31.h, p3/M, z2.h, z12.h\n"
-    "ld1h { z12.h }, p2/Z, [x9, x24, LSL #1]\n" // Load input point (3, 4)
+    "ld1h { z12.h }, p2/Z, [x9, x24, LSL #1]\n"
     "addvl x9, x9, #1\n"
     "fmla z30.h, p3/M, z2.h, z9.h\n"
     "fmla z29.h, p3/M, z2.h, z10.h\n"
     "fmla z28.h, p3/M, z2.h, z11.h\n"
-    "ld1h { z2.h }, p3/Z, [x8, #-4, MUL VL]\n" // Load from weights and bias
+    "ld1h { z2.h }, p3/Z, [x8, #-4, MUL VL]\n"
     "fmla z31.h, p3/M, z3.h, z9.h\n"
-    "ld1h { z9.h }, p2/Z, [x28]\n" // Load input point (4, 0)
+    "ld1h { z9.h }, p2/Z, [x28]\n"
     "fmla z30.h, p3/M, z3.h, z13.h\n"
     "fmla z29.h, p3/M, z3.h, z11.h\n"
     "fmla z28.h, p3/M, z3.h, z12.h\n"
-    "ld1h { z3.h }, p3/Z, [x8, #-3, MUL VL]\n" // Load from weights and bias
+    "ld1h { z3.h }, p3/Z, [x8, #-3, MUL VL]\n"
     "fmla z31.h, p3/M, z4.h, z13.h\n"
-    "ld1h { z13.h }, p2/Z, [x28, x15, LSL #1]\n" // Load input point (4, 1)
+    "ld1h { z13.h }, p2/Z, [x28, x15, LSL #1]\n"
     "fmla z30.h, p3/M, z4.h, z8.h\n"
-    "ld1h { z8.h }, p2/Z, [x28, x24, LSL #1]\n" // Load input point (4, 4)
+    "ld1h { z8.h }, p2/Z, [x28, x24, LSL #1]\n"
     "fmla z29.h, p3/M, z4.h, z12.h\n"
     "fmla z28.h, p3/M, z4.h, z14.h\n"
-    "ld1h { z4.h }, p3/Z, [x8, #-2, MUL VL]\n" // Load from weights and bias
+    "ld1h { z4.h }, p3/Z, [x8, #-2, MUL VL]\n"
     "fmla z31.h, p3/M, z0.h, z5.h\n"
-    "ld1h { z5.h }, p2/Z, [x28, x26, LSL #1]\n" // Load input point (4, 2)
+    "ld1h { z5.h }, p2/Z, [x28, x26, LSL #1]\n"
     "fmla z30.h, p3/M, z0.h, z6.h\n"
     "fmla z29.h, p3/M, z0.h, z9.h\n"
     "fmla z28.h, p3/M, z0.h, z13.h\n"
-    "ld1h { z0.h }, p3/Z, [x8, #-1, MUL VL]\n" // Load from weights and bias
+    "ld1h { z0.h }, p3/Z, [x8, #-1, MUL VL]\n"
     "fmla z31.h, p3/M, z1.h, z6.h\n"
-    "ld1h { z6.h }, p2/Z, [x28, x25, LSL #1]\n" // Load input point (4, 3)
+    "ld1h { z6.h }, p2/Z, [x28, x25, LSL #1]\n"
     "fmla z30.h, p3/M, z1.h, z10.h\n"
     "fmla z29.h, p3/M, z1.h, z13.h\n"
     "fmla z28.h, p3/M, z1.h, z5.h\n"
-    "ld1h { z1.h }, p3/Z, [x8]\n" // Load from weights and bias
+    "ld1h { z1.h }, p3/Z, [x8]\n"
     "fmla z31.h, p3/M, z2.h, z10.h\n"
-    "ld1h { z10.h }, p2/Z, [x28, x23, LSL #1]\n" // Load input point (4, 5)
+    "ld1h { z10.h }, p2/Z, [x28, x23, LSL #1]\n"
     "addvl x28, x28, #1\n"
     "fmla z30.h, p3/M, z2.h, z11.h\n"
     "fmla z29.h, p3/M, z2.h, z5.h\n"
     "fmla z28.h, p3/M, z2.h, z6.h\n"
-    "ld1h { z2.h }, p3/Z, [x8, #1, MUL VL]\n" // Load from weights and bias
+    "ld1h { z2.h }, p3/Z, [x8, #1, MUL VL]\n"
     "fmla z31.h, p3/M, z3.h, z11.h\n"
-    "ld1h { z11.h }, p2/Z, [x27]\n" // Load input point (5, 0)
+    "ld1h { z11.h }, p2/Z, [x27]\n"
     "fmla z30.h, p3/M, z3.h, z12.h\n"
     "fmla z29.h, p3/M, z3.h, z6.h\n"
     "fmla z28.h, p3/M, z3.h, z8.h\n"
-    "ld1h { z3.h }, p3/Z, [x8, #2, MUL VL]\n" // Load from weights and bias
+    "ld1h { z3.h }, p3/Z, [x8, #2, MUL VL]\n"
     "fmla z31.h, p3/M, z4.h, z12.h\n"
-    "ld1h { z12.h }, p2/Z, [x27, x15, LSL #1]\n" // Load input point (5, 1)
+    "ld1h { z12.h }, p2/Z, [x27, x15, LSL #1]\n"
     "fmla z30.h, p3/M, z4.h, z14.h\n"
-    "ld1h { z14.h }, p1/Z, [x10]\n" // Load input point (2, 0)
+    "ld1h { z14.h }, p1/Z, [x10]\n"
     "fmla z29.h, p3/M, z4.h, z8.h\n"
     "fmla z28.h, p3/M, z4.h, z10.h\n"
-    "ld1h { z4.h }, p3/Z, [x8, #3, MUL VL]\n" // Load from weights and bias
+    "ld1h { z4.h }, p3/Z, [x8, #3, MUL VL]\n"
     "fmla z31.h, p3/M, z0.h, z9.h\n"
-    "ld1h { z9.h }, p2/Z, [x27, x26, LSL #1]\n" // Load input point (5, 2)
+    "ld1h { z9.h }, p2/Z, [x27, x26, LSL #1]\n"
     "fmla z30.h, p3/M, z0.h, z13.h\n"
     "fmla z29.h, p3/M, z0.h, z11.h\n"
-    "ld1h { z11.h }, p2/Z, [x27, x25, LSL #1]\n" // Load input point (5, 3)
+    "ld1h { z11.h }, p2/Z, [x27, x25, LSL #1]\n"
     "fmla z28.h, p3/M, z0.h, z12.h\n"
-    "ld1h { z0.h }, p3/Z, [x8, #5, MUL VL]\n" // Load from weights and bias
+    "ld1h { z0.h }, p3/Z, [x8, #5, MUL VL]\n"
     "fmla z31.h, p3/M, z1.h, z13.h\n"
-    "ld1h { z13.h }, p1/Z, [x20, x26, LSL #1]\n" // Load input point (1, 2)
+    "ld1h { z13.h }, p1/Z, [x20, x26, LSL #1]\n"
     "fmla z30.h, p3/M, z1.h, z5.h\n"
     "fmla z29.h, p3/M, z1.h, z12.h\n"
-    "ld1h { z12.h }, p2/Z, [x27, x24, LSL #1]\n" // Load input point (5, 4)
+    "ld1h { z12.h }, p2/Z, [x27, x24, LSL #1]\n"
     "fmla z28.h, p3/M, z1.h, z9.h\n"
-    "ld1h { z1.h }, p3/Z, [x8, #6, MUL VL]\n" // Load from weights and bias
+    "ld1h { z1.h }, p3/Z, [x8, #6, MUL VL]\n"
     "fmla z31.h, p3/M, z2.h, z5.h\n"
-    "ld1h { z5.h }, p1/Z, [x13]\n" // Load input point (0, 0)
+    "ld1h { z5.h }, p1/Z, [x13]\n"
     "fmla z30.h, p3/M, z2.h, z6.h\n"
     "fmla z29.h, p3/M, z2.h, z9.h\n"
-    "ld1h { z9.h }, p2/Z, [x27, x23, LSL #1]\n" // Load input point (5, 5)
+    "ld1h { z9.h }, p2/Z, [x27, x23, LSL #1]\n"
     "whilelt p2.h, x17, %x[n_channels]\n"
     "fmla z28.h, p3/M, z2.h, z11.h\n"
-    "ld1h { z2.h }, p3/Z, [x8, #7, MUL VL]\n" // Load from weights and bias
+    "ld1h { z2.h }, p3/Z, [x8, #7, MUL VL]\n"
     "addvl x27, x27, #1\n"
     "fmla z31.h, p3/M, z3.h, z6.h\n"
-    "ld1h { z6.h }, p1/Z, [x13, x15, LSL #1]\n" // Load input point (0, 1)
+    "ld1h { z6.h }, p1/Z, [x13, x15, LSL #1]\n"
     "addvl x8, x8, #16\n"
     "fmla z30.h, p3/M, z3.h, z8.h\n"
     "cmp x16, %x[n_channels]\n"
     "fmla z29.h, p3/M, z3.h, z11.h\n"
-    "ld1h { z11.h }, p1/Z, [x13, x25, LSL #1]\n" // Load input point (0, 3)
+    "ld1h { z11.h }, p1/Z, [x13, x25, LSL #1]\n"
     "fmla z28.h, p3/M, z3.h, z12.h\n"
-    "ld1h { z3.h }, p3/Z, [x8, #-8, MUL VL]\n" // Load from weights and bias
+    "ld1h { z3.h }, p3/Z, [x8, #-8, MUL VL]\n"
     "fmla z31.h, p3/M, z4.h, z8.h\n"
-    "ld1h { z8.h }, p1/Z, [x20, x15, LSL #1]\n" // Load input point (1, 1)
+    "ld1h { z8.h }, p1/Z, [x20, x15, LSL #1]\n"
     "fmla z30.h, p3/M, z4.h, z10.h\n"
-    "ld1h { z10.h }, p1/Z, [x20, x23, LSL #1]\n" // Load input point (1, 5)
+    "ld1h { z10.h }, p1/Z, [x20, x23, LSL #1]\n"
     "fmla z29.h, p3/M, z4.h, z12.h\n"
-    "ld1h { z12.h }, p1/Z, [x13, x24, LSL #1]\n" // Load input point (0, 4)
+    "ld1h { z12.h }, p1/Z, [x13, x24, LSL #1]\n"
     "fmla z28.h, p3/M, z4.h, z9.h\n"
-    "ld1h { z9.h }, p1/Z, [x13, x26, LSL #1]\n" // Load input point (0, 2)
-    "ld1h { z4.h }, p3/Z, [x8, #-7, MUL VL]\n" // Load from weights and bias
+    "ld1h { z9.h }, p1/Z, [x13, x26, LSL #1]\n"
+    "ld1h { z4.h }, p3/Z, [x8, #-7, MUL VL]\n"
     "fmax z31.h, p3/M, z31.h, z18.h\n"
     "addvl x8, x8, #-6\n"
     "fmax z30.h, p3/M, z30.h, z18.h\n"
     "fmax z29.h, p3/M, z29.h, z18.h\n"
     "fmax z28.h, p3/M, z28.h, z18.h\n"
     "fmin z31.h, p3/M, z31.h, z17.h\n"
-    "st1h { z31.h }, p0, [x11]\n" // Store output point (0, 0)
-    "mov z31.d, z16.d\n"
+    "st1h { z31.h }, p0, [x11]\n"
     "fmin z30.h, p3/M, z30.h, z17.h\n"
-    "st1h { z30.h }, p0, [x11, x12, LSL #1]\n" // Store output point (0, 1)
-    "mov z30.d, z16.d\n"
-    "addvl x11, x11, #1\n"
     "fmin z29.h, p3/M, z29.h, z17.h\n"
-    "st1h { z29.h }, p0, [x22]\n" // Store output point (1, 0)
-    "mov z29.d, z16.d\n"
+    "st1h { z30.h }, p0, [x11, x12, LSL #1]\n"
     "fmin z28.h, p3/M, z28.h, z17.h\n"
-    "st1h { z28.h }, p0, [x22, x12, LSL #1]\n" // Store output point (1, 1)
-    "mov z28.d, z16.d\n"
+    "addvl x11, x11, #1\n"
+    "st1h { z29.h }, p0, [x22]\n"
+    "st1h { z28.h }, p0, [x22, x12, LSL #1]\n"
     "addvl x22, x22, #1\n"
     "blt 2b\n"
     "3:"  // Tile loop: Channel tail
-    "fmla z31.h, p3/M, z0.h, z5.h\n"
-    "ld1h { z5.h }, p2/Z, [x20, x25, LSL #1]\n" // Load input point (1, 3)
+    "movprfx z31, z16\n fmla z31.h, p3/M, z0.h, z5.h\n"
+    "ld1h { z5.h }, p2/Z, [x20, x25, LSL #1]\n"
     "mov p0.b, p2.b\n"
-    "fmla z30.h, p3/M, z0.h, z6.h\n"
+    "movprfx z30, z16\n fmla z30.h, p3/M, z0.h, z6.h\n"
     "ldr x5, [%x[params_struct], %[offsetof_args_tile_i]]\n"
     "add x21, x5, #0x1\n"
-    "fmla z29.h, p3/M, z0.h, z7.h\n"
+    "movprfx z29, z16\n fmla z29.h, p3/M, z0.h, z7.h\n"
     "ldr x6, [%x[params_struct], %[offsetof_args_tile_j]]\n"
-    "fmla z28.h, p3/M, z0.h, z8.h\n"
-    "ld1h { z0.h }, p3/Z, [x8]\n" // Load from weights and bias
+    "movprfx z28, z16\n fmla z28.h, p3/M, z0.h, z8.h\n"
+    "ld1h { z0.h }, p3/Z, [x8]\n"
     "add x6, x6, #0x1\n"
     "fmla z31.h, p3/M, z1.h, z6.h\n"
-    "ld1h { z6.h }, p2/Z, [x20, x24, LSL #1]\n" // Load input point (1, 4)
+    "ld1h { z6.h }, p2/Z, [x20, x24, LSL #1]\n"
     "fmla z30.h, p3/M, z1.h, z9.h\n"
     "ldr x20, [%x[params_struct], %[offsetof_args_n_tile_rows]]\n"
     "fmla z29.h, p3/M, z1.h, z8.h\n"
     "ldr x19, [%x[params_struct], %[offsetof_args_n_tile_cols]]\n"
     "cmp x6, x19\n"
     "fmla z28.h, p3/M, z1.h, z13.h\n"
-    "ld1h { z1.h }, p3/Z, [x8, #1, MUL VL]\n" // Load from weights and bias
+    "ld1h { z1.h }, p3/Z, [x8, #1, MUL VL]\n"
     "fmla z31.h, p3/M, z2.h, z9.h\n"
-    "ld1h { z9.h }, p2/Z, [x13, x23, LSL #1]\n" // Load input point (0, 5)
+    "ld1h { z9.h }, p2/Z, [x13, x23, LSL #1]\n"
     "csel x6, x6, XZR, LT\n"
     "fmla z30.h, p3/M, z2.h, z11.h\n"
     "csel x5, x5, x21, LT\n"
     "fmla z29.h, p3/M, z2.h, z13.h\n"
     "cmp x5, x20\n"
     "fmla z28.h, p3/M, z2.h, z5.h\n"
-    "ld1h { z2.h }, p3/Z, [x8, #2, MUL VL]\n" // Load from weights and bias
+    "ld1h { z2.h }, p3/Z, [x8, #2, MUL VL]\n"
     "fmla z31.h, p3/M, z3.h, z11.h\n"
-    "ld1h { z11.h }, p2/Z, [x10, x15, LSL #1]\n" // Load input point (2, 1)
+    "ld1h { z11.h }, p2/Z, [x10, x15, LSL #1]\n"
     "fmla z30.h, p3/M, z3.h, z12.h\n"
     "fmla z29.h, p3/M, z3.h, z5.h\n"
     "fmla z28.h, p3/M, z3.h, z6.h\n"
-    "ld1h { z3.h }, p3/Z, [x8, #3, MUL VL]\n" // Load from weights and bias
+    "ld1h { z3.h }, p3/Z, [x8, #3, MUL VL]\n"
     "fmla z31.h, p3/M, z4.h, z12.h\n"
-    "ld1h { z12.h }, p2/Z, [x10, x26, LSL #1]\n" // Load input point (2, 2)
+    "ld1h { z12.h }, p2/Z, [x10, x26, LSL #1]\n"
     "fmla z30.h, p3/M, z4.h, z9.h\n"
-    "ld1h { z9.h }, p2/Z, [x10, x25, LSL #1]\n" // Load input point (2, 3)
+    "ld1h { z9.h }, p2/Z, [x10, x25, LSL #1]\n"
     "fmla z29.h, p3/M, z4.h, z6.h\n"
     "fmla z28.h, p3/M, z4.h, z10.h\n"
-    "ld1h { z4.h }, p3/Z, [x8, #4, MUL VL]\n" // Load from weights and bias
+    "ld1h { z4.h }, p3/Z, [x8, #4, MUL VL]\n"
     "fmla z31.h, p3/M, z0.h, z7.h\n"
     "fmla z30.h, p3/M, z0.h, z8.h\n"
     "fmla z29.h, p3/M, z0.h, z14.h\n"
     "fmla z28.h, p3/M, z0.h, z11.h\n"
-    "ld1h { z0.h }, p3/Z, [x8, #5, MUL VL]\n" // Load from weights and bias
+    "ld1h { z0.h }, p3/Z, [x8, #5, MUL VL]\n"
     "fmla z31.h, p3/M, z1.h, z8.h\n"
-    "ld1h { z8.h }, p2/Z, [x10, x23, LSL #1]\n" // Load input point (2, 5)
+    "ld1h { z8.h }, p2/Z, [x10, x23, LSL #1]\n"
     "fmla z30.h, p3/M, z1.h, z13.h\n"
     "fmla z29.h, p3/M, z1.h, z11.h\n"
     "fmla z28.h, p3/M, z1.h, z12.h\n"
-    "ld1h { z1.h }, p3/Z, [x8, #6, MUL VL]\n" // Load from weights and bias
+    "ld1h { z1.h }, p3/Z, [x8, #6, MUL VL]\n"
     "fmla z31.h, p3/M, z2.h, z13.h\n"
-    "ld1h { z13.h }, p2/Z, [x10, x24, LSL #1]\n" // Load input point (2, 4)
+    "ld1h { z13.h }, p2/Z, [x10, x24, LSL #1]\n"
     "fmla z30.h, p3/M, z2.h, z5.h\n"
     "fmla z29.h, p3/M, z2.h, z12.h\n"
     "fmla z28.h, p3/M, z2.h, z9.h\n"
-    "ld1h { z2.h }, p3/Z, [x8, #7, MUL VL]\n" // Load from weights and bias
+    "ld1h { z2.h }, p3/Z, [x8, #7, MUL VL]\n"
     "addvl x8, x8, #16\n"
     "fmla z31.h, p3/M, z3.h, z5.h\n"
-    "ld1h { z5.h }, p2/Z, [x9]\n" // Load input point (3, 0)
+    "ld1h { z5.h }, p2/Z, [x9]\n"
     "fmla z30.h, p3/M, z3.h, z6.h\n"
     "fmla z29.h, p3/M, z3.h, z9.h\n"
     "fmla z28.h, p3/M, z3.h, z13.h\n"
-    "ld1h { z3.h }, p3/Z, [x8, #-8, MUL VL]\n" // Load from weights and bias
+    "ld1h { z3.h }, p3/Z, [x8, #-8, MUL VL]\n"
     "fmla z31.h, p3/M, z4.h, z6.h\n"
-    "ld1h { z6.h }, p2/Z, [x9, x15, LSL #1]\n" // Load input point (3, 1)
+    "ld1h { z6.h }, p2/Z, [x9, x15, LSL #1]\n"
     "fmla z30.h, p3/M, z4.h, z10.h\n"
-    "ld1h { z10.h }, p2/Z, [x9, x26, LSL #1]\n" // Load input point (3, 2)
+    "ld1h { z10.h }, p2/Z, [x9, x26, LSL #1]\n"
     "fmla z29.h, p3/M, z4.h, z13.h\n"
     "fmla z28.h, p3/M, z4.h, z8.h\n"
-    "ld1h { z4.h }, p3/Z, [x8, #-7, MUL VL]\n" // Load from weights and bias
+    "ld1h { z4.h }, p3/Z, [x8, #-7, MUL VL]\n"
     "fmla z31.h, p3/M, z0.h, z14.h\n"
-    "ld1h { z14.h }, p2/Z, [x9, x23, LSL #1]\n" // Load input point (3, 5)
+    "ld1h { z14.h }, p2/Z, [x9, x23, LSL #1]\n"
     "fmla z30.h, p3/M, z0.h, z11.h\n"
     "fmla z29.h, p3/M, z0.h, z5.h\n"
     "fmla z28.h, p3/M, z0.h, z6.h\n"
-    "ld1h { z0.h }, p3/Z, [x8, #-6, MUL VL]\n" // Load from weights and bias
+    "ld1h { z0.h }, p3/Z, [x8, #-6, MUL VL]\n"
     "fmla z31.h, p3/M, z1.h, z11.h\n"
-    "ld1h { z11.h }, p2/Z, [x9, x25, LSL #1]\n" // Load input point (3, 3)
+    "ld1h { z11.h }, p2/Z, [x9, x25, LSL #1]\n"
     "fmla z30.h, p3/M, z1.h, z12.h\n"
     "fmla z29.h, p3/M, z1.h, z6.h\n"
     "fmla z28.h, p3/M, z1.h, z10.h\n"
-    "ld1h { z1.h }, p3/Z, [x8, #-5, MUL VL]\n" // Load from weights and bias
+    "ld1h { z1.h }, p3/Z, [x8, #-5, MUL VL]\n"
     "fmla z31.h, p3/M, z2.h, z12.h\n"
-    "ld1h { z12.h }, p2/Z, [x9, x24, LSL #1]\n" // Load input point (3, 4)
+    "ld1h { z12.h }, p2/Z, [x9, x24, LSL #1]\n"
     "fmla z30.h, p3/M, z2.h, z9.h\n"
     "fmla z29.h, p3/M, z2.h, z10.h\n"
     "fmla z28.h, p3/M, z2.h, z11.h\n"
-    "ld1h { z2.h }, p3/Z, [x8, #-4, MUL VL]\n" // Load from weights and bias
+    "ld1h { z2.h }, p3/Z, [x8, #-4, MUL VL]\n"
     "fmla z31.h, p3/M, z3.h, z9.h\n"
-    "ld1h { z9.h }, p2/Z, [x28]\n" // Load input point (4, 0)
+    "ld1h { z9.h }, p2/Z, [x28]\n"
     "fmla z30.h, p3/M, z3.h, z13.h\n"
     "fmla z29.h, p3/M, z3.h, z11.h\n"
     "fmla z28.h, p3/M, z3.h, z12.h\n"
-    "ld1h { z3.h }, p3/Z, [x8, #-3, MUL VL]\n" // Load from weights and bias
+    "ld1h { z3.h }, p3/Z, [x8, #-3, MUL VL]\n"
     "fmla z31.h, p3/M, z4.h, z13.h\n"
-    "ld1h { z13.h }, p2/Z, [x28, x15, LSL #1]\n" // Load input point (4, 1)
+    "ld1h { z13.h }, p2/Z, [x28, x15, LSL #1]\n"
     "fmla z30.h, p3/M, z4.h, z8.h\n"
-    "ld1h { z8.h }, p2/Z, [x28, x24, LSL #1]\n" // Load input point (4, 4)
+    "ld1h { z8.h }, p2/Z, [x28, x24, LSL #1]\n"
     "fmla z29.h, p3/M, z4.h, z12.h\n"
     "fmla z28.h, p3/M, z4.h, z14.h\n"
-    "ld1h { z4.h }, p3/Z, [x8, #-2, MUL VL]\n" // Load from weights and bias
+    "ld1h { z4.h }, p3/Z, [x8, #-2, MUL VL]\n"
     "fmla z31.h, p3/M, z0.h, z5.h\n"
-    "ld1h { z5.h }, p2/Z, [x28, x26, LSL #1]\n" // Load input point (4, 2)
+    "ld1h { z5.h }, p2/Z, [x28, x26, LSL #1]\n"
     "fmla z30.h, p3/M, z0.h, z6.h\n"
     "fmla z29.h, p3/M, z0.h, z9.h\n"
     "fmla z28.h, p3/M, z0.h, z13.h\n"
-    "ld1h { z0.h }, p3/Z, [x8, #-1, MUL VL]\n" // Load from weights and bias
+    "ld1h { z0.h }, p3/Z, [x8, #-1, MUL VL]\n"
     "fmla z31.h, p3/M, z1.h, z6.h\n"
-    "ld1h { z6.h }, p2/Z, [x28, x25, LSL #1]\n" // Load input point (4, 3)
+    "ld1h { z6.h }, p2/Z, [x28, x25, LSL #1]\n"
     "fmla z30.h, p3/M, z1.h, z10.h\n"
     "fmla z29.h, p3/M, z1.h, z13.h\n"
     "fmla z28.h, p3/M, z1.h, z5.h\n"
-    "ld1h { z1.h }, p3/Z, [x8]\n" // Load from weights and bias
+    "ld1h { z1.h }, p3/Z, [x8]\n"
     "fmla z31.h, p3/M, z2.h, z10.h\n"
-    "ld1h { z10.h }, p2/Z, [x28, x23, LSL #1]\n" // Load input point (4, 5)
+    "ld1h { z10.h }, p2/Z, [x28, x23, LSL #1]\n"
     "fmla z30.h, p3/M, z2.h, z11.h\n"
     "fmla z29.h, p3/M, z2.h, z5.h\n"
     "fmla z28.h, p3/M, z2.h, z6.h\n"
-    "ld1h { z2.h }, p3/Z, [x8, #1, MUL VL]\n" // Load from weights and bias
+    "ld1h { z2.h }, p3/Z, [x8, #1, MUL VL]\n"
     "fmla z31.h, p3/M, z3.h, z11.h\n"
-    "ld1h { z11.h }, p2/Z, [x27]\n" // Load input point (5, 0)
+    "ld1h { z11.h }, p2/Z, [x27]\n"
     "fmla z30.h, p3/M, z3.h, z12.h\n"
     "fmla z29.h, p3/M, z3.h, z6.h\n"
     "fmla z28.h, p3/M, z3.h, z8.h\n"
-    "ld1h { z3.h }, p3/Z, [x8, #2, MUL VL]\n" // Load from weights and bias
+    "ld1h { z3.h }, p3/Z, [x8, #2, MUL VL]\n"
     "fmla z31.h, p3/M, z4.h, z12.h\n"
-    "ld1h { z12.h }, p2/Z, [x27, x15, LSL #1]\n" // Load input point (5, 1)
+    "ld1h { z12.h }, p2/Z, [x27, x15, LSL #1]\n"
     "fmla z30.h, p3/M, z4.h, z14.h\n"
     "fmla z29.h, p3/M, z4.h, z8.h\n"
     "fmla z28.h, p3/M, z4.h, z10.h\n"
-    "ld1h { z4.h }, p3/Z, [x8, #3, MUL VL]\n" // Load from weights and bias
+    "ld1h { z4.h }, p3/Z, [x8, #3, MUL VL]\n"
     "fmla z31.h, p3/M, z0.h, z9.h\n"
-    "ld1h { z9.h }, p2/Z, [x27, x26, LSL #1]\n" // Load input point (5, 2)
+    "ld1h { z9.h }, p2/Z, [x27, x26, LSL #1]\n"
     "fmla z30.h, p3/M, z0.h, z13.h\n"
     "fmla z29.h, p3/M, z0.h, z11.h\n"
-    "ld1h { z11.h }, p2/Z, [x27, x25, LSL #1]\n" // Load input point (5, 3)
+    "ld1h { z11.h }, p2/Z, [x27, x25, LSL #1]\n"
     "fmla z28.h, p3/M, z0.h, z12.h\n"
     "fmla z31.h, p3/M, z1.h, z13.h\n"
     "fmla z30.h, p3/M, z1.h, z5.h\n"
     "fmla z29.h, p3/M, z1.h, z12.h\n"
-    "ld1h { z12.h }, p2/Z, [x27, x24, LSL #1]\n" // Load input point (5, 4)
+    "ld1h { z12.h }, p2/Z, [x27, x24, LSL #1]\n"
     "fmla z28.h, p3/M, z1.h, z9.h\n"
     "fmla z31.h, p3/M, z2.h, z5.h\n"
     "fmla z30.h, p3/M, z2.h, z6.h\n"
     "fmla z29.h, p3/M, z2.h, z9.h\n"
-    "ld1h { z9.h }, p2/Z, [x27, x23, LSL #1]\n" // Load input point (5, 5)
+    "ld1h { z9.h }, p2/Z, [x27, x23, LSL #1]\n"
     "fmla z28.h, p3/M, z2.h, z11.h\n"
     "fmla z31.h, p3/M, z3.h, z6.h\n"
     "fmla z30.h, p3/M, z3.h, z8.h\n"
@@ -511,13 +503,13 @@ void sve_fp16_nhwc_5x5_s1_output2x2_mla_depthfirst_direct_impl(
     "fmax z29.h, p3/M, z29.h, z18.h\n"
     "fmax z28.h, p3/M, z28.h, z18.h\n"
     "fmin z31.h, p3/M, z31.h, z17.h\n"
-    "st1h { z31.h }, p0, [x11]\n" // Store output point (0, 0)
+    "st1h { z31.h }, p0, [x11]\n"
     "fmin z30.h, p3/M, z30.h, z17.h\n"
     "fmin z29.h, p3/M, z29.h, z17.h\n"
-    "st1h { z30.h }, p0, [x11, x12, LSL #1]\n" // Store output point (0, 1)
+    "st1h { z30.h }, p0, [x11, x12, LSL #1]\n"
     "fmin z28.h, p3/M, z28.h, z17.h\n"
-    "st1h { z29.h }, p0, [x22]\n" // Store output point (1, 0)
-    "st1h { z28.h }, p0, [x22, x12, LSL #1]\n" // Store output point (1, 1)
+    "st1h { z29.h }, p0, [x22]\n"
+    "st1h { z28.h }, p0, [x22, x12, LSL #1]\n"
     "blt 1b\n"
     :
     : [n_channels] "r" ((unsigned long) n_channels), [offsetof_args_inptr] "I" (offsetof(Args, inptr)), [offsetof_args_ld_input_col] "I" (offsetof(Args, ld_input_col)), [offsetof_args_ld_input_row] "I" (offsetof(Args, ld_input_row)), [offsetof_args_ld_output_col] "I" (offsetof(Args, ld_output_col)), [offsetof_args_ld_output_row] "I" (offsetof(Args, ld_output_row)), [offsetof_args_max] "I" (offsetof(Args, max)), [offsetof_args_min] "I" (offsetof(Args, min)), [offsetof_args_n_tile_cols] "I" (offsetof(Args, n_tile_cols)), [offsetof_args_n_tile_rows] "I" (offsetof(Args, n_tile_rows)), [offsetof_args_outptr] "I" (offsetof(Args, outptr)), [offsetof_args_params] "I" (offsetof(Args, params)), [offsetof_args_tile_i] "I" (offsetof(Args, tile_i)), [offsetof_args_tile_j] "I" (offsetof(Args, tile_j)), [params_struct] "r" (&params_struct)
@@ -528,4 +520,4 @@ void sve_fp16_nhwc_5x5_s1_output2x2_mla_depthfirst_direct_impl(
 }  // namespace depthwise
 }  // namespace arm_conv
 
-#endif  // defined(ARM_COMPUTE_ENABLE_SVE) && defined(__ARM_FP16_ARGS)
+#endif  // __aarch64__ && defined(ARM_COMPUTE_ENABLE_SVE) && defined(__ARM_FP16_ARGS)