Revert "Update CPU kernels to remove x19"

This reverts commit 3c59f01c209d2732a15d97d65565ead964787a8b.

Resolves: COMPMID-5817
Change-Id: Ie2443a21854a95db1e3d0cafa2121c0187a5e237
Signed-off-by: Michael Tyler <michael.tyler@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/8974
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Benchmark: Arm Jenkins <bsgcomp@arm.com>

1 files changed, 202 insertions, 202 deletions

diff --git a/src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp16_nhwc_3x3_s1_output2x2_mla_depthfirst/generic_direct.cpp b/src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp16_nhwc_3x3_s1_output2x2_mla_depthfirst/generic_direct.cpp
index 9fd220abf8..955a02de57 100644
--- a/src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp16_nhwc_3x3_s1_output2x2_mla_depthfirst/generic_direct.cpp
+++ b/src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp16_nhwc_3x3_s1_output2x2_mla_depthfirst/generic_direct.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2021, 2023 Arm Limited.
+ * Copyright (c) 2021 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -88,225 +88,225 @@ void sve_fp16_nhwc_3x3_s1_output2x2_mla_depthfirst_direct_impl(
 
   __asm__ __volatile__(
     "ptrue p3.b\n"
-    "mov x10, #0x0\n"
-    "mov x14, #0x0\n"
+    "mov x17, #0x0\n"
+    "mov x16, #0x0\n"
     "1:"  // Tile loop
-    "str x10, [%x[params_struct], %[offsetof_args_tile_i]]\n"
-    "mov x25, #0x2\n"
-    "mov x24, #0x2\n"
-    "str x14, [%x[params_struct], %[offsetof_args_tile_j]]\n"
-    "ldr x23, [%x[params_struct], %[offsetof_args_ld_input_row]]\n"
-    "ldr x22, [%x[params_struct], %[offsetof_args_ld_output_row]]\n"
-    "mul x21, x10, x23\n"  // offset = tile_i * ld_input_row
-    "ldr x13, [%x[params_struct], %[offsetof_args_ld_input_col]]\n"
-    "ldr x12, [%x[params_struct], %[offsetof_args_ld_output_col]]\n"
-    "mul x20, x10, x22\n"  // offset = tile_i * ld_output_row
-    "cnth x11\n"
-    "madd x21, x14, x13, x21\n"  // offset += tile_j * ld_input_col
-    "ldr x10, [%x[params_struct], %[offsetof_args_params]]\n"
-    "ldr x9, [%x[params_struct], %[offsetof_args_inptr]]\n"
-    "whilelt p2.h, XZR, %x[n_channels]\n"
-    "madd x20, x14, x12, x20\n"  // offset += tile_j * ld_output_col
+    "str x17, [%x[params_struct], %[offsetof_args_tile_i]]\n"
+    "mov x23, #0x2\n"
+    "str x16, [%x[params_struct], %[offsetof_args_tile_j]]\n"
+    "mov x15, #0x2\n"
+    "ldr x14, [%x[params_struct], %[offsetof_args_params]]\n"
+    "mov x13, #0x0\n"
+    "ldr x22, [%x[params_struct], %[offsetof_args_ld_input_row]]\n"
+    "cnth x12\n"
+    "ldr x11, [%x[params_struct], %[offsetof_args_ld_input_col]]\n"
+    "sub x21, XZR, x12\n"
+    "ldr x10, [%x[params_struct], %[offsetof_args_inptr]]\n"
+    "mul x19, x17, x22\n" // offset = tile_i * ld_input_row
+    "ldr x20, [%x[params_struct], %[offsetof_args_ld_output_row]]\n"
+    "madd x19, x16, x11, x19\n" // offset += tile_j * ld_input_col
+    "ldr x9, [%x[params_struct], %[offsetof_args_ld_output_col]]\n"
+    "mul x19, x19, x23\n" // offset *= kernel_stride * output_size
     "ldr x28, [%x[params_struct], %[offsetof_args_outptr]]\n"
-    "ld1h { z18.h }, p3/Z, [x10]\n"
-    "add x27, x13, x13\n"
-    "mul x21, x21, x25\n"  // offset *= kernel_stride * output_size
-    "add x9, x9, x21, LSL #1\n" // inptr[0] += offset * sizeof(__fp16)
-    "ld1h { z0.h }, p3/Z, [x10, #1, MUL VL]\n"
-    "ld1h { z1.h }, p3/Z, [x10, #2, MUL VL]\n"
-    "mul x20, x20, x24\n"  // offset *= output_tile_size
-    "ld1h { z2.h }, p3/Z, [x10, #3, MUL VL]\n"
-    "ld1h { z3.h }, p3/Z, [x10, #4, MUL VL]\n"
-    "add x26, x9, x23, LSL #1\n"
-    "ld1h { z4.h }, p3/Z, [x10, #5, MUL VL]\n"
-    "ld1h { z5.h }, p3/Z, [x10, #6, MUL VL]\n"
-    "add x25, x26, x23, LSL #1\n"
-    "add x24, x27, x13\n"
-    "ld1h { z6.h }, p3/Z, [x10, #7, MUL VL]\n"
-    "addvl x10, x10, #16\n"
-    "add x28, x28, x20, LSL #1\n"  // outptrs[0] += offset * sizeof(__fp16)
-    "ld1rh { z17.h }, p3/Z, [%x[params_struct], %[offsetof_args_min]]\n"
-    "cmp x11, %x[n_channels]\n"
-    "add x23, x25, x23, LSL #1\n"
-    "ld1rh { z16.h }, p3/Z, [%x[params_struct], %[offsetof_args_max]]\n"
-    "ld1h { z7.h }, p3/Z, [x10, #-8, MUL VL]\n"
-    "add x22, x28, x22, LSL #1\n"
-    "mov x21, #0x0\n"
-    "ld1h { z8.h }, p3/Z, [x10, #-7, MUL VL]\n"
-    "ld1h { z9.h }, p2/Z, [x26, x13, LSL #1]\n"
-    "sub x20, XZR, x11\n"
-    "ld1h { z10.h }, p2/Z, [x9]\n"
-    "ld1h { z11.h }, p2/Z, [x9, x24, LSL #1]\n"
-    "addvl x10, x10, #-6\n"
-    "ld1h { z12.h }, p2/Z, [x26, x27, LSL #1]\n"
-    "ld1h { z13.h }, p2/Z, [x25, x13, LSL #1]\n"
+    "add x10, x10, x19, LSL #1\n" // inptr[0] += offset * sizeof(__fp16)
+    "ld1rh { z18.h }, p3/Z, [%x[params_struct], %[offsetof_args_min]]\n"
+    "add x27, x10, x22, LSL #1\n"
+    "ld1rh { z17.h }, p3/Z, [%x[params_struct], %[offsetof_args_max]]\n"
+    "add x26, x27, x22, LSL #1\n"
+    "ld1h { z16.h }, p3/Z, [x14]\n"
+    "add x25, x26, x22, LSL #1\n"
+    "ld1h { z0.h }, p3/Z, [x14, #1, MUL VL]\n"
+    "add x24, x11, x11\n"
+    "ld1h { z1.h }, p3/Z, [x14, #2, MUL VL]\n"
+    "add x23, x24, x11\n"
+    "ld1h { z2.h }, p3/Z, [x14, #3, MUL VL]\n"
+    "mul x19, x17, x20\n" // offset = tile_i * ld_output_row
+    "ld1h { z3.h }, p3/Z, [x14, #4, MUL VL]\n"
+    "madd x19, x16, x9, x19\n" // offset += tile_j * ld_output_col
+    "ld1h { z4.h }, p3/Z, [x14, #5, MUL VL]\n"
+    "mul x19, x19, x15\n" // offset *= output_tile_size
+    "ld1h { z5.h }, p3/Z, [x14, #6, MUL VL]\n"
+    "add x28, x28, x19, LSL #1\n" // outptrs[0] += offset * sizeof(__fp16)
+    "ld1h { z6.h }, p3/Z, [x14, #7, MUL VL]\n"
+    "add x22, x28, x20, LSL #1\n"
+    "whilelt p2.h, XZR, %x[n_channels]\n"
+    "ld1h { z9.h }, p2/Z, [x27, x11, LSL #1]\n"
+    "ld1h { z10.h }, p2/Z, [x10]\n"
+    "addvl x14, x14, #16\n"
+    "ld1h { z11.h }, p2/Z, [x10, x23, LSL #1]\n"
+    "cmp x12, %x[n_channels]\n"
+    "ld1h { z7.h }, p3/Z, [x14, #-8, MUL VL]\n"
+    "ld1h { z8.h }, p3/Z, [x14, #-7, MUL VL]\n"
+    "addvl x14, x14, #-6\n"
+    "ld1h { z12.h }, p2/Z, [x27, x24, LSL #1]\n"
+    "ld1h { z13.h }, p2/Z, [x26, x11, LSL #1]\n"
     "bge 3f\n"
     "2:"  // Tile loop: Channel loop
-    "movprfx z28, z18\n fmla z28.h, p3/M, z4.h, z9.h\n"
-    "movprfx z29, z18\n fmla z29.h, p3/M, z3.h, z9.h\n"
-    "whilelt p1.h, x11, %x[n_channels]\n"
+    "movprfx z31, z16\n fmla z31.h, p3/M, z4.h, z9.h\n"
+    "whilelt p1.h, x12, %x[n_channels]\n"
+    "movprfx z30, z16\n fmla z30.h, p3/M, z3.h, z9.h\n"
     "inch x21\n"
-    "movprfx z30, z18\n fmla z30.h, p3/M, z1.h, z9.h\n"
-    "movprfx z31, z18\n fmla z31.h, p3/M, z0.h, z9.h\n"
-    "ld1h { z9.h }, p2/Z, [x23]\n"
-    "inch x11\n"
-    "fmla z28.h, p3/M, z0.h, z10.h\n"
-    "fmla z29.h, p3/M, z2.h, z11.h\n"
-    "ld1h { z11.h }, p2/Z, [x23, x24, LSL #1]\n"
-    "ld1h { z10.h }, p2/Z, [x25, x27, LSL #1]\n"
-    "fmla z30.h, p3/M, z2.h, z12.h\n"
-    "fmla z31.h, p3/M, z1.h, z12.h\n"
+    "movprfx z29, z16\n fmla z29.h, p3/M, z1.h, z9.h\n"
     "mov p0.b, p2.b\n"
-    "ld1h { z18.h }, p3/Z, [x10]\n"
-    "fmla z28.h, p3/M, z5.h, z12.h\n"
-    "fmla z29.h, p3/M, z4.h, z12.h\n"
-    "ld1h { z12.h }, p2/Z, [x9, x13, LSL #1]\n"
-    "inch x20\n"
-    "fmla z30.h, p3/M, z6.h, z9.h\n"
-    "fmla z31.h, p3/M, z3.h, z13.h\n"
-    "ld1h { z9.h }, p2/Z, [x9, x27, LSL #1]\n"
-    "addvl x9, x9, #1\n"
-    "fmla z28.h, p3/M, z7.h, z13.h\n"
-    "fmla z29.h, p3/M, z6.h, z13.h\n"
-    "fmla z30.h, p3/M, z4.h, z13.h\n"
-    "fmla z31.h, p3/M, z8.h, z11.h\n"
-    "ld1h { z11.h }, p2/Z, [x26]\n"
+    "movprfx z28, z16\n fmla z28.h, p3/M, z0.h, z9.h\n"
+    "ld1h { z9.h }, p2/Z, [x25]\n"
+    "inch x13\n"
+    "fmla z31.h, p3/M, z0.h, z10.h\n"
+    "ld1h { z10.h }, p2/Z, [x26, x24, LSL #1]\n"
+    "inch x12\n"
+    "fmla z30.h, p3/M, z2.h, z11.h\n"
+    "ld1h { z11.h }, p2/Z, [x25, x23, LSL #1]\n"
+    "fmla z29.h, p3/M, z2.h, z12.h\n"
+    "ld1h { z16.h }, p3/Z, [x14]\n"
     "fmla z28.h, p3/M, z1.h, z12.h\n"
-    "fmla z29.h, p3/M, z0.h, z12.h\n"
-    "ld1h { z12.h }, p2/Z, [x26, x24, LSL #1]\n"
+    "fmla z31.h, p3/M, z5.h, z12.h\n"
+    "fmla z30.h, p3/M, z4.h, z12.h\n"
+    "ld1h { z12.h }, p2/Z, [x10, x11, LSL #1]\n"
+    "fmla z29.h, p3/M, z6.h, z9.h\n"
+    "ld1h { z9.h }, p2/Z, [x10, x24, LSL #1]\n"
+    "addvl x10, x10, #1\n"
+    "fmla z28.h, p3/M, z3.h, z13.h\n"
+    "fmla z31.h, p3/M, z7.h, z13.h\n"
+    "fmla z30.h, p3/M, z6.h, z13.h\n"
+    "fmla z29.h, p3/M, z4.h, z13.h\n"
+    "fmla z28.h, p3/M, z8.h, z11.h\n"
+    "ld1h { z11.h }, p2/Z, [x27]\n"
+    "fmla z31.h, p3/M, z1.h, z12.h\n"
+    "fmla z30.h, p3/M, z0.h, z12.h\n"
+    "ld1h { z12.h }, p2/Z, [x27, x23, LSL #1]\n"
+    "addvl x27, x27, #1\n"
+    "fmla z29.h, p3/M, z5.h, z10.h\n"
+    "fmla z28.h, p3/M, z4.h, z10.h\n"
+    "ld1h { z4.h }, p3/Z, [x14, #5, MUL VL]\n"
+    "fmla z31.h, p3/M, z2.h, z9.h\n"
+    "fmla z30.h, p3/M, z1.h, z9.h\n"
+    "ld1h { z9.h }, p2/Z, [x26]\n"
+    "ld1h { z1.h }, p3/Z, [x14, #2, MUL VL]\n"
+    "fmla z29.h, p3/M, z0.h, z11.h\n"
+    "ld1h { z0.h }, p3/Z, [x14, #1, MUL VL]\n"
+    "fmla z28.h, p3/M, z2.h, z12.h\n"
+    "ld1h { z2.h }, p3/Z, [x14, #3, MUL VL]\n"
+    "fmla z31.h, p3/M, z8.h, z10.h\n"
+    "fmla z30.h, p3/M, z7.h, z10.h\n"
+    "ld1h { z10.h }, p2/Z, [x26, x23, LSL #1]\n"
     "addvl x26, x26, #1\n"
-    "fmla z30.h, p3/M, z5.h, z10.h\n"
-    "fmla z31.h, p3/M, z4.h, z10.h\n"
-    "ld1h { z4.h }, p3/Z, [x10, #5, MUL VL]\n"
-    "fmla z28.h, p3/M, z2.h, z9.h\n"
-    "fmla z29.h, p3/M, z1.h, z9.h\n"
-    "ld1h { z9.h }, p2/Z, [x25]\n"
-    "ld1h { z1.h }, p3/Z, [x10, #2, MUL VL]\n"
-    "fmla z30.h, p3/M, z0.h, z11.h\n"
-    "fmla z31.h, p3/M, z2.h, z12.h\n"
-    "ld1h { z0.h }, p3/Z, [x10, #1, MUL VL]\n"
-    "ld1h { z2.h }, p3/Z, [x10, #3, MUL VL]\n"
-    "fmla z28.h, p3/M, z8.h, z10.h\n"
-    "fmla z29.h, p3/M, z7.h, z10.h\n"
-    "ld1h { z10.h }, p2/Z, [x25, x24, LSL #1]\n"
+    "fmla z29.h, p3/M, z3.h, z9.h\n"
+    "ld1h { z13.h }, p1/Z, [x26, x11, LSL #1]\n"
+    "fmla z31.h, p3/M, z3.h, z11.h\n"
+    "ld1h { z11.h }, p2/Z, [x25, x11, LSL #1]\n"
+    "fmla z28.h, p3/M, z5.h, z10.h\n"
+    "ld1h { z3.h }, p3/Z, [x14, #4, MUL VL]\n"
+    "fmla z30.h, p3/M, z5.h, z12.h\n"
+    "ld1h { z12.h }, p2/Z, [x25, x24, LSL #1]\n"
+    "whilelt p2.h, x13, %x[n_channels]\n"
+    "fmla z29.h, p3/M, z7.h, z11.h\n"
+    "ld1h { z5.h }, p3/Z, [x14, #6, MUL VL]\n"
     "addvl x25, x25, #1\n"
-    "fmla z30.h, p3/M, z3.h, z9.h\n"
-    "fmla z31.h, p3/M, z5.h, z10.h\n"
-    "ld1h { z13.h }, p1/Z, [x25, x13, LSL #1]\n"
-    "fmla z28.h, p3/M, z3.h, z11.h\n"
-    "ld1h { z11.h }, p2/Z, [x23, x13, LSL #1]\n"
-    "fmla z29.h, p3/M, z5.h, z12.h\n"
-    "ld1h { z12.h }, p2/Z, [x23, x27, LSL #1]\n"
-    "fmla z30.h, p3/M, z7.h, z11.h\n"
-    "fmla z31.h, p3/M, z6.h, z11.h\n"
-    "ld1h { z3.h }, p3/Z, [x10, #4, MUL VL]\n"
-    "ld1h { z5.h }, p3/Z, [x10, #6, MUL VL]\n"
-    "fmla z28.h, p3/M, z6.h, z9.h\n"
-    "fmla z29.h, p3/M, z8.h, z10.h\n"
-    "fmax z28.h, p3/M, z28.h, z17.h\n"
-    "fmax z29.h, p3/M, z29.h, z17.h\n"
-    "fmla z30.h, p3/M, z8.h, z12.h\n"
-    "fmla z31.h, p3/M, z7.h, z12.h\n"
-    "fmax z30.h, p3/M, z30.h, z17.h\n"
-    "fmax z31.h, p3/M, z31.h, z17.h\n"
-    "ld1h { z6.h }, p3/Z, [x10, #7, MUL VL]\n"
-    "addvl x10, x10, #16\n"
-    "whilelt p2.h, x21, %x[n_channels]\n"
-    "ld1h { z9.h }, p1/Z, [x26, x13, LSL #1]\n"
-    "cmp x11, %x[n_channels]\n"
-    "fmin z28.h, p3/M, z28.h, z16.h\n"
-    "ld1h { z10.h }, p1/Z, [x9]\n"
-    "ld1h { z11.h }, p1/Z, [x9, x24, LSL #1]\n"
-    "fmin z29.h, p3/M, z29.h, z16.h\n"
-    "fmin z30.h, p3/M, z30.h, z16.h\n"
-    "ld1h { z12.h }, p1/Z, [x26, x27, LSL #1]\n"
-    "st1h { z28.h }, p0, [x28]\n"
-    "fmin z31.h, p3/M, z31.h, z16.h\n"
-    "addvl x23, x23, #1\n"
-    "st1h { z29.h }, p0, [x28, x12, LSL #1]\n"
-    "ld1h { z7.h }, p3/Z, [x10, #-8, MUL VL]\n"
-    "st1h { z30.h }, p0, [x22]\n"
+    "fmla z31.h, p3/M, z6.h, z9.h\n"
+    "ld1h { z9.h }, p1/Z, [x27, x11, LSL #1]\n"
+    "cmp x12, %x[n_channels]\n"
+    "fmla z30.h, p3/M, z8.h, z10.h\n"
+    "ld1h { z10.h }, p1/Z, [x10]\n"
+    "fmla z28.h, p3/M, z6.h, z11.h\n"
+    "ld1h { z11.h }, p1/Z, [x10, x23, LSL #1]\n"
+    "ld1h { z6.h }, p3/Z, [x14, #7, MUL VL]\n"
+    "fmla z29.h, p3/M, z8.h, z12.h\n"
+    "addvl x14, x14, #16\n"
+    "fmax z31.h, p3/M, z31.h, z18.h\n"
+    "ld1h { z8.h }, p3/Z, [x14, #-7, MUL VL]\n"
+    "fmla z28.h, p3/M, z7.h, z12.h\n"
+    "ld1h { z12.h }, p1/Z, [x27, x24, LSL #1]\n"
+    "fmax z30.h, p3/M, z30.h, z18.h\n"
+    "ld1h { z7.h }, p3/Z, [x14, #-8, MUL VL]\n"
+    "addvl x14, x14, #-6\n"
+    "fmax z29.h, p3/M, z29.h, z18.h\n"
+    "fmin z31.h, p3/M, z31.h, z17.h\n"
+    "st1h { z31.h }, p0, [x28]\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, [x28, x9, LSL #1]\n"
+    "fmax z28.h, p3/M, z28.h, z18.h\n"
     "addvl x28, x28, #1\n"
-    "ld1h { z8.h }, p3/Z, [x10, #-7, MUL VL]\n"
-    "addvl x10, x10, #-6\n"
-    "st1h { z31.h }, p0, [x22, x12, LSL #1]\n"
+    "fmin z28.h, p3/M, z28.h, z17.h\n"
+    "st1h { z29.h }, p0, [x22]\n"
+    "st1h { z28.h }, p0, [x22, x9, LSL #1]\n"
     "addvl x22, x22, #1\n"
     "blt 2b\n"
     "3:"  // Tile loop: Channel tail
-    "movprfx z28, z18\n fmla z28.h, p3/M, z4.h, z9.h\n"
-    "movprfx z29, z18\n fmla z29.h, p3/M, z3.h, z9.h\n"
-    "ldr x14, [%x[params_struct], %[offsetof_args_tile_j]]\n"
-    "ldr x10, [%x[params_struct], %[offsetof_args_tile_i]]\n"
-    "movprfx z30, z18\n fmla z30.h, p3/M, z1.h, z9.h\n"
-    "movprfx z31, z18\n fmla z31.h, p3/M, z0.h, z9.h\n"
-    "ld1h { z9.h }, p2/Z, [x23]\n"
-    "ldr x20, [%x[params_struct], %[offsetof_args_n_tile_cols]]\n"
-    "fmla z28.h, p3/M, z0.h, z10.h\n"
-    "fmla z29.h, p3/M, z2.h, z11.h\n"
-    "ld1h { z11.h }, p2/Z, [x23, x24, LSL #1]\n"
-    "ld1h { z10.h }, p2/Z, [x25, x27, LSL #1]\n"
-    "fmla z30.h, p3/M, z2.h, z12.h\n"
-    "fmla z31.h, p3/M, z1.h, z12.h\n"
-    "add x14, x14, #0x1\n"
-    "cmp x14, x20\n"
-    "fmla z28.h, p3/M, z5.h, z12.h\n"
-    "fmla z29.h, p3/M, z4.h, z12.h\n"
-    "ld1h { z12.h }, p2/Z, [x9, x13, LSL #1]\n"
-    "add x21, x10, #0x1\n"
-    "fmla z30.h, p3/M, z6.h, z9.h\n"
-    "fmla z31.h, p3/M, z3.h, z13.h\n"
-    "ld1h { z9.h }, p2/Z, [x9, x27, LSL #1]\n"
-    "ldr x20, [%x[params_struct], %[offsetof_args_n_tile_rows]]\n"
-    "fmla z28.h, p3/M, z7.h, z13.h\n"
-    "fmla z29.h, p3/M, z6.h, z13.h\n"
-    "csel x10, x10, x21, LT\n"
+    "movprfx z31, z16\n fmla z31.h, p3/M, z4.h, z9.h\n"
+    "ldr x17, [%x[params_struct], %[offsetof_args_tile_i]]\n"
     "mov p0.b, p2.b\n"
-    "fmla z30.h, p3/M, z4.h, z13.h\n"
-    "fmla z31.h, p3/M, z8.h, z11.h\n"
-    "ld1h { z11.h }, p2/Z, [x26]\n"
-    "csel x14, x14, XZR, LT\n"
-    "fmla z28.h, p3/M, z1.h, z12.h\n"
-    "fmla z29.h, p3/M, z0.h, z12.h\n"
-    "ld1h { z12.h }, p2/Z, [x26, x24, LSL #1]\n"
-    "cmp x10, x20\n"
-    "fmla z30.h, p3/M, z5.h, z10.h\n"
-    "fmla z31.h, p3/M, z4.h, z10.h\n"
-    "fmla z28.h, p3/M, z2.h, z9.h\n"
-    "fmla z29.h, p3/M, z1.h, z9.h\n"
+    "movprfx z30, z16\n fmla z30.h, p3/M, z3.h, z9.h\n"
+    "ldr x16, [%x[params_struct], %[offsetof_args_tile_j]]\n"
+    "add x21, x17, #0x1\n"
+    "movprfx z29, z16\n fmla z29.h, p3/M, z1.h, z9.h\n"
+    "ldr x20, [%x[params_struct], %[offsetof_args_n_tile_rows]]\n"
+    "movprfx z28, z16\n fmla z28.h, p3/M, z0.h, z9.h\n"
     "ld1h { z9.h }, p2/Z, [x25]\n"
-    "fmla z30.h, p3/M, z0.h, z11.h\n"
-    "fmla z31.h, p3/M, z2.h, z12.h\n"
-    "fmla z28.h, p3/M, z8.h, z10.h\n"
-    "fmla z29.h, p3/M, z7.h, z10.h\n"
-    "ld1h { z10.h }, p2/Z, [x25, x24, LSL #1]\n"
-    "fmla z30.h, p3/M, z3.h, z9.h\n"
-    "fmla z31.h, p3/M, z5.h, z10.h\n"
-    "fmla z28.h, p3/M, z3.h, z11.h\n"
-    "ld1h { z11.h }, p2/Z, [x23, x13, LSL #1]\n"
-    "fmla z29.h, p3/M, z5.h, z12.h\n"
-    "ld1h { z12.h }, p2/Z, [x23, x27, LSL #1]\n"
-    "fmla z30.h, p3/M, z7.h, z11.h\n"
-    "fmla z31.h, p3/M, z6.h, z11.h\n"
-    "fmla z28.h, p3/M, z6.h, z9.h\n"
-    "fmla z29.h, p3/M, z8.h, z10.h\n"
-    "fmax z28.h, p3/M, z28.h, z17.h\n"
-    "fmax z29.h, p3/M, z29.h, z17.h\n"
-    "fmla z30.h, p3/M, z8.h, z12.h\n"
-    "fmla z31.h, p3/M, z7.h, z12.h\n"
-    "fmax z30.h, p3/M, z30.h, z17.h\n"
-    "fmax z31.h, p3/M, z31.h, z17.h\n"
-    "fmin z28.h, p3/M, z28.h, z16.h\n"
-    "fmin z29.h, p3/M, z29.h, z16.h\n"
-    "st1h { z28.h }, p0, [x28]\n"
-    "fmin z30.h, p3/M, z30.h, z16.h\n"
-    "fmin z31.h, p3/M, z31.h, z16.h\n"
-    "st1h { z29.h }, p0, [x28, x12, LSL #1]\n"
-    "st1h { z30.h }, p0, [x22]\n"
-    "st1h { z31.h }, p0, [x22, x12, LSL #1]\n"
+    "add x16, x16, #0x1\n"
+    "fmla z31.h, p3/M, z0.h, z10.h\n"
+    "ld1h { z10.h }, p2/Z, [x26, x24, LSL #1]\n"
+    "fmla z30.h, p3/M, z2.h, z11.h\n"
+    "ld1h { z11.h }, p2/Z, [x25, x23, LSL #1]\n"
+    "ldr x19, [%x[params_struct], %[offsetof_args_n_tile_cols]]\n"
+    "fmla z29.h, p3/M, z2.h, z12.h\n"
+    "cmp x16, x19\n"
+    "fmla z31.h, p3/M, z5.h, z12.h\n"
+    "fmla z30.h, p3/M, z4.h, z12.h\n"
+    "csel x16, x16, XZR, LT\n"
+    "fmla z28.h, p3/M, z1.h, z12.h\n"
+    "ld1h { z12.h }, p2/Z, [x10, x11, LSL #1]\n"
+    "csel x17, x17, x21, LT\n"
+    "fmla z29.h, p3/M, z6.h, z9.h\n"
+    "ld1h { z9.h }, p2/Z, [x10, x24, LSL #1]\n"
+    "cmp x17, x20\n"
+    "fmla z31.h, p3/M, z7.h, z13.h\n"
+    "fmla z30.h, p3/M, z6.h, z13.h\n"
+    "fmla z28.h, p3/M, z3.h, z13.h\n"
+    "fmla z29.h, p3/M, z4.h, z13.h\n"
+    "fmla z31.h, p3/M, z1.h, z12.h\n"
+    "fmla z30.h, p3/M, z0.h, z12.h\n"
+    "ld1h { z12.h }, p2/Z, [x27, x23, LSL #1]\n"
+    "fmla z28.h, p3/M, z8.h, z11.h\n"
+    "ld1h { z11.h }, p2/Z, [x27]\n"
+    "fmla z29.h, p3/M, z5.h, z10.h\n"
+    "fmla z31.h, p3/M, z2.h, z9.h\n"
+    "fmla z30.h, p3/M, z1.h, z9.h\n"
+    "ld1h { z9.h }, p2/Z, [x26]\n"
+    "fmla z28.h, p3/M, z4.h, z10.h\n"
+    "fmla z29.h, p3/M, z0.h, z11.h\n"
+    "fmla z31.h, p3/M, z8.h, z10.h\n"
+    "fmla z30.h, p3/M, z7.h, z10.h\n"
+    "ld1h { z10.h }, p2/Z, [x26, x23, LSL #1]\n"
+    "fmla z28.h, p3/M, z2.h, z12.h\n"
+    "fmla z29.h, p3/M, z3.h, z9.h\n"
+    "fmla z31.h, p3/M, z3.h, z11.h\n"
+    "ld1h { z11.h }, p2/Z, [x25, x11, LSL #1]\n"
+    "fmla z30.h, p3/M, z5.h, z12.h\n"
+    "ld1h { z12.h }, p2/Z, [x25, x24, LSL #1]\n"
+    "fmla z28.h, p3/M, z5.h, z10.h\n"
+    "fmla z29.h, p3/M, z7.h, z11.h\n"
+    "fmla z31.h, p3/M, z6.h, z9.h\n"
+    "fmla z30.h, p3/M, z8.h, z10.h\n"
+    "fmla z28.h, p3/M, z6.h, z11.h\n"
+    "fmla z29.h, p3/M, z8.h, z12.h\n"
+    "fmax z31.h, p3/M, z31.h, z18.h\n"
+    "fmax z30.h, p3/M, z30.h, z18.h\n"
+    "fmla z28.h, p3/M, z7.h, z12.h\n"
+    "fmax z29.h, p3/M, z29.h, z18.h\n"
+    "fmin z31.h, p3/M, z31.h, z17.h\n"
+    "st1h { z31.h }, p0, [x28]\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, [x28, x9, LSL #1]\n"
+    "fmax z28.h, p3/M, z28.h, z18.h\n"
+    "st1h { z29.h }, p0, [x22]\n"
+    "fmin z28.h, p3/M, z28.h, z17.h\n"
+    "st1h { z28.h }, p0, [x22, x9, 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)
-    : "cc", "memory", "p0", "p1", "p2", "p3", "x9", "x10", "x11", "x12", "x13", "x14", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z16", "z17", "z18", "z28", "z29", "z30", "z31"
+    : "cc", "memory", "p0", "p1", "p2", "p3", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x16", "x17", "x19", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z16", "z17", "z18", "z28", "z29", "z30", "z31"
   );
 }