Update CPU kernels to remove x19

Resolves: COMPMID-5805
Signed-off-by: Michael Tyler <michael.tyler@arm.com>
Change-Id: I250f64531e209625e4ff176dd5a552c1c34bc484
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/8909
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Gunes Bayir <gunes.bayir@arm.com>
Reviewed-by: Viet-Hoa Do <viet-hoa.do@arm.com>
Benchmark: Arm Jenkins <bsgcomp@arm.com>

1 files changed, 224 insertions, 224 deletions

diff --git a/src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_nhwc_3x3_s2_output2x2_mla_depthfirst/generic_direct.cpp b/src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_nhwc_3x3_s2_output2x2_mla_depthfirst/generic_direct.cpp
index ac33dcbce5..5a1f309b88 100644
--- a/src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_nhwc_3x3_s2_output2x2_mla_depthfirst/generic_direct.cpp
+++ b/src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_nhwc_3x3_s2_output2x2_mla_depthfirst/generic_direct.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2021 Arm Limited.
+ * Copyright (c) 2021, 2023 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -88,246 +88,246 @@ void sve_fp32_nhwc_3x3_s2_output2x2_mla_depthfirst_direct_impl(
 
   __asm__ __volatile__(
     "ptrue p3.b\n"
-    "mov x7, #0x0\n"
-    "mov x8, #0x0\n"
+    "mov x11, #0x0\n"
+    "mov x16, #0x0\n"
     "1:"  // Tile loop
-    "str x7, [%x[params_struct], %[offsetof_args_tile_i]]\n"
-    "mov x23, #0x4\n"
-    "str x8, [%x[params_struct], %[offsetof_args_tile_j]]\n"
-    "mov x17, #0x2\n"
-    "ldr x16, [%x[params_struct], %[offsetof_args_params]]\n"
-    "mov x15, #0x0\n"
-    "ldr x22, [%x[params_struct], %[offsetof_args_ld_input_row]]\n"
-    "cntw x14\n"
-    "ldr x13, [%x[params_struct], %[offsetof_args_ld_input_col]]\n"
-    "sub x12, XZR, x14\n"
-    "ldr x21, [%x[params_struct], %[offsetof_args_inptr]]\n"
-    "mul x19, x7, x22\n" // offset = tile_i * ld_input_row
-    "ldr x20, [%x[params_struct], %[offsetof_args_ld_output_row]]\n"
-    "madd x19, x8, x13, x19\n" // offset += tile_j * ld_input_col
-    "ldr x11, [%x[params_struct], %[offsetof_args_ld_output_col]]\n"
-    "mul x19, x19, x23\n" // offset *= kernel_stride * output_size
-    "ldr x10, [%x[params_struct], %[offsetof_args_outptr]]\n"
-    "add x21, x21, x19, LSL #2\n" // inptr[0] += offset * sizeof(float)
-    "ld1rw { z19.s }, p3/Z, [%x[params_struct], %[offsetof_args_min]]\n"
-    "add x9, x21, x22, LSL #2\n"
-    "ld1rw { z18.s }, p3/Z, [%x[params_struct], %[offsetof_args_max]]\n"
-    "add x28, x9, x22, LSL #2\n"
-    "ld1w { z17.s }, p3/Z, [x16]\n"
-    "add x27, x28, x22, LSL #2\n"
-    "ld1w { z0.s }, p3/Z, [x16, #1, MUL VL]\n"
-    "add x26, x27, x22, LSL #2\n"
-    "ld1w { z1.s }, p3/Z, [x16, #2, MUL VL]\n"
-    "add x25, x13, x13\n"
-    "ld1w { z2.s }, p3/Z, [x16, #3, MUL VL]\n"
-    "add x24, x25, x13\n"
-    "ld1w { z3.s }, p3/Z, [x16, #4, MUL VL]\n"
-    "add x23, x24, x13\n"
-    "ld1w { z4.s }, p3/Z, [x16, #5, MUL VL]\n"
-    "mul x19, x7, x20\n" // offset = tile_i * ld_output_row
-    "ld1w { z5.s }, p3/Z, [x16, #6, MUL VL]\n"
-    "madd x19, x8, x11, x19\n" // offset += tile_j * ld_output_col
-    "ld1w { z6.s }, p3/Z, [x16, #7, MUL VL]\n"
-    "mul x19, x19, x17\n" // offset *= output_tile_size
+    "str x11, [%x[params_struct], %[offsetof_args_tile_i]]\n"
+    "mov x25, #0x4\n"
+    "mov x24, #0x2\n"
+    "str x16, [%x[params_struct], %[offsetof_args_tile_j]]\n"
+    "ldr x23, [%x[params_struct], %[offsetof_args_ld_input_row]]\n"
+    "ldr x15, [%x[params_struct], %[offsetof_args_ld_input_col]]\n"
+    "mul x22, x11, x23\n"  // offset = tile_i * ld_input_row
+    "ldr x21, [%x[params_struct], %[offsetof_args_ld_output_row]]\n"
+    "madd x22, x16, x15, x22\n"  // offset += tile_j * ld_input_col
+    "ldr x14, [%x[params_struct], %[offsetof_args_ld_output_col]]\n"
+    "cntw x13\n"
+    "mul x20, x11, x21\n"  // offset = tile_i * ld_output_row
+    "ldr x12, [%x[params_struct], %[offsetof_args_inptr]]\n"
+    "ldr x11, [%x[params_struct], %[offsetof_args_params]]\n"
+    "add x10, x15, x15\n"
+    "mul x22, x22, x25\n"  // offset *= kernel_stride * output_size
+    "add x12, x12, x22, LSL #2\n"  // inptr[0] += offset * sizeof(float)
+    "ldr x9, [%x[params_struct], %[offsetof_args_outptr]]\n"
+    "add x28, x12, x23, LSL #2\n"
+    "madd x20, x16, x14, x20\n"  // offset += tile_j * ld_output_col
     "whilelt p2.s, XZR, %x[n_channels]\n"
-    "ld1w { z9.s }, p2/Z, [x28, x25, LSL #2]\n"
-    "ld1w { z10.s }, p2/Z, [x21]\n"
-    "add x10, x10, x19, LSL #2\n" // outptrs[0] += offset * sizeof(float)
-    "ld1w { z11.s }, p2/Z, [x21, x13, LSL #2]\n"
-    "add x22, x10, x20, LSL #2\n"
-    "ld1w { z12.s }, p2/Z, [x21, x24, LSL #2]\n"
-    "addvl x16, x16, #16\n"
-    "ld1w { z13.s }, p2/Z, [x21, x23, LSL #2]\n"
-    "cmp x14, %x[n_channels]\n"
-    "ld1w { z7.s }, p3/Z, [x16, #-8, MUL VL]\n"
-    "ld1w { z8.s }, p3/Z, [x16, #-7, MUL VL]\n"
-    "addvl x16, x16, #-6\n"
-    "ld1w { z14.s }, p2/Z, [x9]\n"
-    "ld1w { z15.s }, p2/Z, [x9, x13, LSL #2]\n"
-    "ld1w { z16.s }, p2/Z, [x21, x25, LSL #2]\n"
+    "ld1w { z19.s }, p3/Z, [x11]\n"
+    "ld1w { z0.s }, p3/Z, [x11, #1, MUL VL]\n"
+    "mul x20, x20, x24\n"  // offset *= output_tile_size
+    "ld1w { z1.s }, p3/Z, [x11, #2, MUL VL]\n"
+    "ld1w { z2.s }, p3/Z, [x11, #3, MUL VL]\n"
+    "add x27, x28, x23, LSL #2\n"
+    "ld1w { z3.s }, p3/Z, [x11, #4, MUL VL]\n"
+    "ld1w { z4.s }, p3/Z, [x11, #5, MUL VL]\n"
+    "add x26, x10, x15\n"
+    "add x25, x27, x23, LSL #2\n"
+    "ld1w { z5.s }, p3/Z, [x11, #6, MUL VL]\n"
+    "ld1w { z6.s }, p3/Z, [x11, #7, MUL VL]\n"
+    "addvl x11, x11, #16\n"
+    "add x24, x26, x15\n"
+    "add x9, x9, x20, LSL #2\n"  // outptrs[0] += offset * sizeof(float)
+    "cmp x13, %x[n_channels]\n"
+    "ld1rw { z18.s }, p3/Z, [%x[params_struct], %[offsetof_args_min]]\n"
+    "ld1rw { z17.s }, p3/Z, [%x[params_struct], %[offsetof_args_max]]\n"
+    "add x23, x25, x23, LSL #2\n"
+    "add x22, x9, x21, LSL #2\n"
+    "ld1w { z7.s }, p3/Z, [x11, #-8, MUL VL]\n"
+    "ld1w { z8.s }, p3/Z, [x11, #-7, MUL VL]\n"
+    "mov x21, #0x0\n"
+    "sub x20, XZR, x13\n"
+    "ld1w { z9.s }, p2/Z, [x27, x10, LSL #2]\n"
+    "ld1w { z10.s }, p2/Z, [x12]\n"
+    "ld1w { z11.s }, p2/Z, [x12, x15, LSL #2]\n"
+    "ld1w { z12.s }, p2/Z, [x12, x26, LSL #2]\n"
+    "addvl x11, x11, #-6\n"
+    "ld1w { z13.s }, p2/Z, [x12, x24, LSL #2]\n"
+    "ld1w { z14.s }, p2/Z, [x28]\n"
+    "ld1w { z15.s }, p2/Z, [x28, x15, LSL #2]\n"
+    "ld1w { z16.s }, p2/Z, [x12, x10, LSL #2]\n"
     "bge 3f\n"
     "2:"  // Tile loop: Channel loop
-    "movprfx z31, z17\n fmla z31.s, p3/M, z8.s, z9.s\n"
-    "whilelt p1.s, x14, %x[n_channels]\n"
-    "movprfx z30, z17\n fmla z30.s, p3/M, z6.s, z9.s\n"
-    "incw x12\n"
-    "movprfx z29, z17\n fmla z29.s, p3/M, z2.s, z9.s\n"
-    "mov p0.b, p2.b\n"
-    "movprfx z28, z17\n fmla z28.s, p3/M, z0.s, z9.s\n"
-    "ld1w { z17.s }, p3/Z, [x16]\n"
-    "incw x15\n"
-    "fmla z31.s, p3/M, z0.s, z10.s\n"
-    "addvl x21, x21, #1\n"
-    "ld1w { z10.s }, p1/Z, [x21]\n"
-    "fmla z30.s, p3/M, z1.s, z12.s\n"
-    "incw x14\n"
-    "fmla z31.s, p3/M, z1.s, z11.s\n"
-    "ld1w { z11.s }, p2/Z, [x9, x24, LSL #2]\n"
-    "ld1w { z12.s }, p2/Z, [x9, x23, LSL #2]\n"
-    "fmla z30.s, p3/M, z2.s, z13.s\n"
-    "ld1w { z13.s }, p2/Z, [x9, x25, LSL #2]\n"
-    "fmla z31.s, p3/M, z3.s, z14.s\n"
-    "ld1w { z14.s }, p2/Z, [x27]\n"
-    "addvl x9, x9, #1\n"
-    "fmla z30.s, p3/M, z0.s, z16.s\n"
-    "fmla z29.s, p3/M, z3.s, z14.s\n"
-    "ld1w { z14.s }, p2/Z, [x27, x23, LSL #2]\n"
-    "fmla z31.s, p3/M, z4.s, z15.s\n"
-    "ld1w { z15.s }, p2/Z, [x28]\n"
-    "fmla z30.s, p3/M, z4.s, z11.s\n"
-    "ld1w { z11.s }, p2/Z, [x27, x13, LSL #2]\n"
-    "fmla z29.s, p3/M, z0.s, z15.s\n"
-    "ld1w { z0.s }, p3/Z, [x16, #1, MUL VL]\n"
-    "fmla z31.s, p3/M, z2.s, z16.s\n"
-    "ld1w { z16.s }, p2/Z, [x28, x13, LSL #2]\n"
-    "fmla z30.s, p3/M, z5.s, z12.s\n"
+    "movprfx z28, z19\n fmla z28.s, p3/M, z8.s, z9.s\n"
+    "movprfx z29, z19\n fmla z29.s, p3/M, z6.s, z9.s\n"
+    "whilelt p1.s, x13, %x[n_channels]\n"
+    "incw x21\n"
+    "fmla z28.s, p3/M, z0.s, z10.s\n"
+    "fmla z29.s, p3/M, z1.s, z12.s\n"
     "ld1w { z12.s }, p2/Z, [x28, x24, LSL #2]\n"
+    "incw x13\n"
+    "fmla z28.s, p3/M, z1.s, z11.s\n"
+    "fmla z29.s, p3/M, z2.s, z13.s\n"
+    "ld1w { z11.s }, p2/Z, [x28, x26, LSL #2]\n"
+    "ld1w { z13.s }, p2/Z, [x28, x10, LSL #2]\n"
+    "fmla z28.s, p3/M, z3.s, z14.s\n"
+    "fmla z29.s, p3/M, z0.s, z16.s\n"
+    "ld1w { z14.s }, p2/Z, [x25]\n"
+    "mov p0.b, p2.b\n"
+    "fmla z28.s, p3/M, z4.s, z15.s\n"
     "fmla z29.s, p3/M, z4.s, z11.s\n"
-    "ld1w { z11.s }, p2/Z, [x28, x23, LSL #2]\n"
+    "ld1w { z15.s }, p2/Z, [x27]\n"
+    "ld1w { z11.s }, p2/Z, [x25, x15, LSL #2]\n"
+    "fmla z28.s, p3/M, z2.s, z16.s\n"
+    "fmla z29.s, p3/M, z5.s, z12.s\n"
+    "ld1w { z12.s }, p2/Z, [x27, x26, LSL #2]\n"
+    "ld1w { z16.s }, p2/Z, [x27, x15, LSL #2]\n"
+    "movprfx z30, z19\n fmla z30.s, p3/M, z2.s, z9.s\n"
+    "movprfx z31, z19\n fmla z31.s, p3/M, z0.s, z9.s\n"
+    "addvl x12, x12, #1\n"
     "addvl x28, x28, #1\n"
-    "fmla z31.s, p3/M, z5.s, z13.s\n"
-    "ld1w { z9.s }, p1/Z, [x28, x25, LSL #2]\n"
-    "fmla z30.s, p3/M, z3.s, z13.s\n"
-    "ld1w { z13.s }, p2/Z, [x27, x24, LSL #2]\n"
-    "fmla z29.s, p3/M, z1.s, z16.s\n"
-    "fmla z31.s, p3/M, z6.s, z15.s\n"
-    "ld1w { z15.s }, p2/Z, [x26]\n"
-    "fmla z28.s, p3/M, z4.s, z13.s\n"
-    "ld1w { z13.s }, p2/Z, [x26, x13, LSL #2]\n"
-    "fmla z30.s, p3/M, z7.s, z12.s\n"
-    "ld1w { z4.s }, p3/Z, [x16, #5, MUL VL]\n"
-    "fmla z29.s, p3/M, z6.s, z15.s\n"
-    "ld1w { z15.s }, p2/Z, [x26, x25, LSL #2]\n"
-    "fmla z31.s, p3/M, z7.s, z16.s\n"
-    "ld1w { z16.s }, p2/Z, [x27, x25, LSL #2]\n"
-    "addvl x27, x27, #1\n"
-    "fmla z28.s, p3/M, z1.s, z12.s\n"
-    "ld1w { z12.s }, p1/Z, [x21, x24, LSL #2]\n"
-    "fmla z30.s, p3/M, z8.s, z11.s\n"
-    "ld1w { z1.s }, p3/Z, [x16, #2, MUL VL]\n"
-    "fmla z29.s, p3/M, z7.s, z13.s\n"
-    "ld1w { z13.s }, p1/Z, [x21, x23, LSL #2]\n"
-    "fmax z31.s, p3/M, z31.s, z19.s\n"
-    "fmla z28.s, p3/M, z5.s, z14.s\n"
-    "ld1w { z14.s }, p2/Z, [x26, x24, LSL #2]\n"
-    "fmax z30.s, p3/M, z30.s, z19.s\n"
-    "fmla z29.s, p3/M, z5.s, z16.s\n"
-    "ld1w { z5.s }, p3/Z, [x16, #6, MUL VL]\n"
-    "fmin z31.s, p3/M, z31.s, z18.s\n"
-    "st1w { z31.s }, p0, [x10]\n"
-    "fmla z28.s, p3/M, z2.s, z11.s\n"
-    "fmla z29.s, p3/M, z8.s, z15.s\n"
-    "ld1w { z11.s }, p2/Z, [x26, x23, LSL #2]\n"
-    "whilelt p2.s, x15, %x[n_channels]\n"
-    "fmin z30.s, p3/M, z30.s, z18.s\n"
-    "ld1w { z2.s }, p3/Z, [x16, #3, MUL VL]\n"
-    "addvl x26, x26, #1\n"
-    "fmla z28.s, p3/M, z3.s, z16.s\n"
-    "ld1w { z16.s }, p1/Z, [x21, x25, LSL #2]\n"
-    "cmp x14, %x[n_channels]\n"
-    "fmax z29.s, p3/M, z29.s, z19.s\n"
-    "ld1w { z3.s }, p3/Z, [x16, #4, MUL VL]\n"
-    "st1w { z30.s }, p0, [x10, x11, LSL #2]\n"
-    "fmla z28.s, p3/M, z7.s, z14.s\n"
-    "ld1w { z14.s }, p1/Z, [x9]\n"
-    "addvl x10, x10, #1\n"
-    "fmin z29.s, p3/M, z29.s, z18.s\n"
-    "st1w { z29.s }, p0, [x22]\n"
+    "fmla z28.s, p3/M, z5.s, z13.s\n"
+    "fmla z29.s, p3/M, z3.s, z13.s\n"
+    "ld1w { z13.s }, p2/Z, [x25, x26, LSL #2]\n"
+    "ld1w { z19.s }, p3/Z, [x11]\n"
+    "fmla z30.s, p3/M, z3.s, z14.s\n"
+    "fmla z31.s, p3/M, z4.s, z13.s\n"
+    "ld1w { z14.s }, p2/Z, [x25, x24, LSL #2]\n"
+    "ld1w { z13.s }, p2/Z, [x23, x15, LSL #2]\n"
+    "fmla z30.s, p3/M, z0.s, z15.s\n"
+    "fmla z31.s, p3/M, z1.s, z12.s\n"
+    "ld1w { z0.s }, p3/Z, [x11, #1, MUL VL]\n"
+    "incw x20\n"
+    "fmla z30.s, p3/M, z4.s, z11.s\n"
+    "fmla z31.s, p3/M, z5.s, z14.s\n"
+    "ld1w { z11.s }, p2/Z, [x27, x24, LSL #2]\n"
+    "ld1w { z14.s }, p2/Z, [x23, x26, LSL #2]\n"
     "fmla z28.s, p3/M, z6.s, z15.s\n"
-    "ld1w { z15.s }, p1/Z, [x9, x13, LSL #2]\n"
-    "fmla z28.s, p3/M, z8.s, z11.s\n"
-    "ld1w { z11.s }, p1/Z, [x21, x13, LSL #2]\n"
-    "ld1w { z6.s }, p3/Z, [x16, #7, MUL VL]\n"
-    "fmax z28.s, p3/M, z28.s, z19.s\n"
-    "addvl x16, x16, #16\n"
-    "ld1w { z7.s }, p3/Z, [x16, #-8, MUL VL]\n"
-    "fmin z28.s, p3/M, z28.s, z18.s\n"
-    "ld1w { z8.s }, p3/Z, [x16, #-7, MUL VL]\n"
-    "addvl x16, x16, #-6\n"
-    "st1w { z28.s }, p0, [x22, x11, LSL #2]\n"
+    "fmla z30.s, p3/M, z1.s, z16.s\n"
+    "ld1w { z15.s }, p2/Z, [x23]\n"
+    "addvl x27, x27, #1\n"
+    "fmla z31.s, p3/M, z2.s, z11.s\n"
+    "fmla z28.s, p3/M, z7.s, z16.s\n"
+    "ld1w { z16.s }, p2/Z, [x25, x10, LSL #2]\n"
+    "fmax z28.s, p3/M, z28.s, z18.s\n"
+    "fmla z30.s, p3/M, z6.s, z15.s\n"
+    "fmla z31.s, p3/M, z3.s, z16.s\n"
+    "ld1w { z15.s }, p2/Z, [x23, x10, LSL #2]\n"
+    "ld1w { z1.s }, p3/Z, [x11, #2, MUL VL]\n"
+    "fmla z30.s, p3/M, z7.s, z13.s\n"
+    "fmla z31.s, p3/M, z7.s, z14.s\n"
+    "ld1w { z2.s }, p3/Z, [x11, #3, MUL VL]\n"
+    "ld1w { z3.s }, p3/Z, [x11, #4, MUL VL]\n"
+    "fmla z29.s, p3/M, z7.s, z12.s\n"
+    "fmla z30.s, p3/M, z5.s, z16.s\n"
+    "ld1w { z4.s }, p3/Z, [x11, #5, MUL VL]\n"
+    "ld1w { z5.s }, p3/Z, [x11, #6, MUL VL]\n"
+    "fmla z31.s, p3/M, z6.s, z15.s\n"
+    "fmla z29.s, p3/M, z8.s, z11.s\n"
+    "ld1w { z11.s }, p2/Z, [x23, x24, LSL #2]\n"
+    "fmax z29.s, p3/M, z29.s, z18.s\n"
+    "fmla z30.s, p3/M, z8.s, z15.s\n"
+    "fmla z31.s, p3/M, z8.s, z11.s\n"
+    "fmax z30.s, p3/M, z30.s, z18.s\n"
+    "fmax z31.s, p3/M, z31.s, z18.s\n"
+    "ld1w { z6.s }, p3/Z, [x11, #7, MUL VL]\n"
+    "addvl x11, x11, #16\n"
+    "whilelt p2.s, x21, %x[n_channels]\n"
+    "ld1w { z9.s }, p1/Z, [x27, x10, LSL #2]\n"
+    "cmp x13, %x[n_channels]\n"
+    "fmin z28.s, p3/M, z28.s, z17.s\n"
+    "ld1w { z10.s }, p1/Z, [x12]\n"
+    "ld1w { z11.s }, p1/Z, [x12, x15, LSL #2]\n"
+    "fmin z29.s, p3/M, z29.s, z17.s\n"
+    "fmin z30.s, p3/M, z30.s, z17.s\n"
+    "ld1w { z12.s }, p1/Z, [x12, x26, LSL #2]\n"
+    "ld1w { z13.s }, p1/Z, [x12, x24, LSL #2]\n"
+    "fmin z31.s, p3/M, z31.s, z17.s\n"
+    "addvl x25, x25, #1\n"
+    "ld1w { z14.s }, p1/Z, [x28]\n"
+    "ld1w { z15.s }, p1/Z, [x28, x15, LSL #2]\n"
+    "addvl x23, x23, #1\n"
+    "ld1w { z16.s }, p1/Z, [x12, x10, LSL #2]\n"
+    "st1w { z28.s }, p0, [x9]\n"
+    "ld1w { z7.s }, p3/Z, [x11, #-8, MUL VL]\n"
+    "st1w { z29.s }, p0, [x9, x14, LSL #2]\n"
+    "addvl x9, x9, #1\n"
+    "ld1w { z8.s }, p3/Z, [x11, #-7, MUL VL]\n"
+    "addvl x11, x11, #-6\n"
+    "st1w { z30.s }, p0, [x22]\n"
+    "st1w { z31.s }, p0, [x22, x14, LSL #2]\n"
     "addvl x22, x22, #1\n"
     "blt 2b\n"
     "3:"  // Tile loop: Channel tail
-    "movprfx z31, z17\n fmla z31.s, p3/M, z8.s, z9.s\n"
-    "ldr x7, [%x[params_struct], %[offsetof_args_tile_i]]\n"
-    "mov p0.b, p2.b\n"
-    "movprfx z30, z17\n fmla z30.s, p3/M, z6.s, z9.s\n"
-    "ldr x8, [%x[params_struct], %[offsetof_args_tile_j]]\n"
-    "add x21, x7, #0x1\n"
-    "movprfx z29, z17\n fmla z29.s, p3/M, z2.s, z9.s\n"
-    "ldr x20, [%x[params_struct], %[offsetof_args_n_tile_rows]]\n"
-    "movprfx z28, z17\n fmla z28.s, p3/M, z0.s, z9.s\n"
-    "ldr x19, [%x[params_struct], %[offsetof_args_n_tile_cols]]\n"
-    "add x8, x8, #0x1\n"
-    "fmla z31.s, p3/M, z0.s, z10.s\n"
-    "cmp x8, x19\n"
-    "fmla z30.s, p3/M, z1.s, z12.s\n"
-    "ld1w { z12.s }, p2/Z, [x9, x23, LSL #2]\n"
-    "fmla z31.s, p3/M, z1.s, z11.s\n"
-    "ld1w { z11.s }, p2/Z, [x9, x24, LSL #2]\n"
-    "csel x8, x8, XZR, LT\n"
-    "fmla z30.s, p3/M, z2.s, z13.s\n"
-    "ld1w { z13.s }, p2/Z, [x9, x25, LSL #2]\n"
-    "csel x7, x7, x21, LT\n"
-    "fmla z31.s, p3/M, z3.s, z14.s\n"
-    "ld1w { z14.s }, p2/Z, [x27]\n"
-    "cmp x7, x20\n"
-    "fmla z30.s, p3/M, z0.s, z16.s\n"
-    "fmla z29.s, p3/M, z3.s, z14.s\n"
-    "ld1w { z14.s }, p2/Z, [x27, x23, LSL #2]\n"
-    "fmla z31.s, p3/M, z4.s, z15.s\n"
-    "ld1w { z15.s }, p2/Z, [x28]\n"
-    "fmla z30.s, p3/M, z4.s, z11.s\n"
-    "ld1w { z11.s }, p2/Z, [x27, x13, LSL #2]\n"
-    "fmla z29.s, p3/M, z0.s, z15.s\n"
-    "fmla z31.s, p3/M, z2.s, z16.s\n"
-    "ld1w { z16.s }, p2/Z, [x28, x13, LSL #2]\n"
-    "fmla z30.s, p3/M, z5.s, z12.s\n"
+    "movprfx z28, z19\n fmla z28.s, p3/M, z8.s, z9.s\n"
+    "movprfx z29, z19\n fmla z29.s, p3/M, z6.s, z9.s\n"
+    "ldr x16, [%x[params_struct], %[offsetof_args_tile_j]]\n"
+    "ldr x11, [%x[params_struct], %[offsetof_args_tile_i]]\n"
+    "fmla z28.s, p3/M, z0.s, z10.s\n"
+    "fmla z29.s, p3/M, z1.s, z12.s\n"
     "ld1w { z12.s }, p2/Z, [x28, x24, LSL #2]\n"
+    "ldr x20, [%x[params_struct], %[offsetof_args_n_tile_cols]]\n"
+    "fmla z28.s, p3/M, z1.s, z11.s\n"
+    "fmla z29.s, p3/M, z2.s, z13.s\n"
+    "ld1w { z11.s }, p2/Z, [x28, x26, LSL #2]\n"
+    "ld1w { z13.s }, p2/Z, [x28, x10, LSL #2]\n"
+    "fmla z28.s, p3/M, z3.s, z14.s\n"
+    "fmla z29.s, p3/M, z0.s, z16.s\n"
+    "ld1w { z14.s }, p2/Z, [x25]\n"
+    "add x16, x16, #0x1\n"
+    "fmla z28.s, p3/M, z4.s, z15.s\n"
     "fmla z29.s, p3/M, z4.s, z11.s\n"
-    "ld1w { z11.s }, p2/Z, [x28, x23, LSL #2]\n"
-    "fmla z31.s, p3/M, z5.s, z13.s\n"
-    "fmla z30.s, p3/M, z3.s, z13.s\n"
-    "ld1w { z13.s }, p2/Z, [x27, x24, LSL #2]\n"
-    "fmla z29.s, p3/M, z1.s, z16.s\n"
-    "fmla z31.s, p3/M, z6.s, z15.s\n"
-    "ld1w { z15.s }, p2/Z, [x26]\n"
-    "fmla z28.s, p3/M, z4.s, z13.s\n"
-    "ld1w { z13.s }, p2/Z, [x26, x13, LSL #2]\n"
-    "fmla z30.s, p3/M, z7.s, z12.s\n"
-    "fmla z29.s, p3/M, z6.s, z15.s\n"
-    "ld1w { z15.s }, p2/Z, [x26, x25, LSL #2]\n"
-    "fmla z31.s, p3/M, z7.s, z16.s\n"
-    "ld1w { z16.s }, p2/Z, [x27, x25, LSL #2]\n"
-    "fmla z28.s, p3/M, z1.s, z12.s\n"
-    "fmla z30.s, p3/M, z8.s, z11.s\n"
-    "fmla z29.s, p3/M, z7.s, z13.s\n"
-    "fmax z31.s, p3/M, z31.s, z19.s\n"
-    "fmla z28.s, p3/M, z5.s, z14.s\n"
-    "ld1w { z14.s }, p2/Z, [x26, x24, LSL #2]\n"
-    "fmax z30.s, p3/M, z30.s, z19.s\n"
-    "fmla z29.s, p3/M, z5.s, z16.s\n"
-    "fmin z31.s, p3/M, z31.s, z18.s\n"
-    "st1w { z31.s }, p0, [x10]\n"
-    "fmla z28.s, p3/M, z2.s, z11.s\n"
-    "fmla z29.s, p3/M, z8.s, z15.s\n"
-    "ld1w { z11.s }, p2/Z, [x26, x23, LSL #2]\n"
-    "fmin z30.s, p3/M, z30.s, z18.s\n"
-    "st1w { z30.s }, p0, [x10, x11, LSL #2]\n"
-    "fmla z28.s, p3/M, z3.s, z16.s\n"
-    "fmax z29.s, p3/M, z29.s, z19.s\n"
-    "fmla z28.s, p3/M, z7.s, z14.s\n"
-    "fmin z29.s, p3/M, z29.s, z18.s\n"
-    "st1w { z29.s }, p0, [x22]\n"
+    "ld1w { z15.s }, p2/Z, [x27]\n"
+    "ld1w { z11.s }, p2/Z, [x25, x15, LSL #2]\n"
+    "fmla z28.s, p3/M, z2.s, z16.s\n"
+    "fmla z29.s, p3/M, z5.s, z12.s\n"
+    "ld1w { z12.s }, p2/Z, [x27, x26, LSL #2]\n"
+    "ld1w { z16.s }, p2/Z, [x27, x15, LSL #2]\n"
+    "movprfx z30, z19\n fmla z30.s, p3/M, z2.s, z9.s\n"
+    "movprfx z31, z19\n fmla z31.s, p3/M, z0.s, z9.s\n"
+    "cmp x16, x20\n"
+    "add x21, x11, #0x1\n"
+    "fmla z28.s, p3/M, z5.s, z13.s\n"
+    "fmla z29.s, p3/M, z3.s, z13.s\n"
+    "ld1w { z13.s }, p2/Z, [x25, x26, LSL #2]\n"
+    "ldr x20, [%x[params_struct], %[offsetof_args_n_tile_rows]]\n"
+    "fmla z30.s, p3/M, z3.s, z14.s\n"
+    "fmla z31.s, p3/M, z4.s, z13.s\n"
+    "ld1w { z14.s }, p2/Z, [x25, x24, LSL #2]\n"
+    "ld1w { z13.s }, p2/Z, [x23, x15, LSL #2]\n"
+    "fmla z30.s, p3/M, z0.s, z15.s\n"
+    "fmla z31.s, p3/M, z1.s, z12.s\n"
+    "csel x11, x11, x21, LT\n"
+    "mov p0.b, p2.b\n"
+    "fmla z30.s, p3/M, z4.s, z11.s\n"
+    "fmla z31.s, p3/M, z5.s, z14.s\n"
+    "ld1w { z11.s }, p2/Z, [x27, x24, LSL #2]\n"
+    "ld1w { z14.s }, p2/Z, [x23, x26, LSL #2]\n"
     "fmla z28.s, p3/M, z6.s, z15.s\n"
-    "fmla z28.s, p3/M, z8.s, z11.s\n"
-    "fmax z28.s, p3/M, z28.s, z19.s\n"
-    "fmin z28.s, p3/M, z28.s, z18.s\n"
-    "st1w { z28.s }, p0, [x22, x11, LSL #2]\n"
+    "fmla z30.s, p3/M, z1.s, z16.s\n"
+    "ld1w { z15.s }, p2/Z, [x23]\n"
+    "csel x16, x16, XZR, LT\n"
+    "fmla z31.s, p3/M, z2.s, z11.s\n"
+    "fmla z28.s, p3/M, z7.s, z16.s\n"
+    "ld1w { z16.s }, p2/Z, [x25, x10, LSL #2]\n"
+    "fmax z28.s, p3/M, z28.s, z18.s\n"
+    "fmla z30.s, p3/M, z6.s, z15.s\n"
+    "fmla z31.s, p3/M, z3.s, z16.s\n"
+    "ld1w { z15.s }, p2/Z, [x23, x10, LSL #2]\n"
+    "cmp x11, x20\n"
+    "fmla z30.s, p3/M, z7.s, z13.s\n"
+    "fmla z31.s, p3/M, z7.s, z14.s\n"
+    "fmin z28.s, p3/M, z28.s, z17.s\n"
+    "st1w { z28.s }, p0, [x9]\n"
+    "fmla z29.s, p3/M, z7.s, z12.s\n"
+    "fmla z30.s, p3/M, z5.s, z16.s\n"
+    "fmla z31.s, p3/M, z6.s, z15.s\n"
+    "fmla z29.s, p3/M, z8.s, z11.s\n"
+    "ld1w { z11.s }, p2/Z, [x23, x24, LSL #2]\n"
+    "fmax z29.s, p3/M, z29.s, z18.s\n"
+    "fmla z30.s, p3/M, z8.s, z15.s\n"
+    "fmla z31.s, p3/M, z8.s, z11.s\n"
+    "fmax z30.s, p3/M, z30.s, z18.s\n"
+    "fmax z31.s, p3/M, z31.s, z18.s\n"
+    "fmin z29.s, p3/M, z29.s, z17.s\n"
+    "fmin z30.s, p3/M, z30.s, z17.s\n"
+    "st1w { z29.s }, p0, [x9, x14, LSL #2]\n"
+    "fmin z31.s, p3/M, z31.s, z17.s\n"
+    "st1w { z30.s }, p0, [x22]\n"
+    "st1w { z31.s }, p0, [x22, x14, LSL #2]\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", "x7", "x8", "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", "z14", "z15", "z16", "z17", "z18", "z19", "z28", "z29", "z30", "z31"
+    : "cc", "memory", "p0", "p1", "p2", "p3", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x16", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z28", "z29", "z30", "z31"
   );
 }