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, 80 insertions, 112 deletions

diff --git a/src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_nhwc_3x3_s1_output4x4_mla_depthfirst/generic_direct.cpp b/src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_nhwc_3x3_s1_output4x4_mla_depthfirst/generic_direct.cpp
index 7df8e481c0..84b4b3b72b 100644
--- a/src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_nhwc_3x3_s1_output4x4_mla_depthfirst/generic_direct.cpp
+++ b/src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_nhwc_3x3_s1_output4x4_mla_depthfirst/generic_direct.cpp
@@ -25,7 +25,7 @@
 #include <cstddef>
 #include <cstdint>
 
-#if defined(ARM_COMPUTE_ENABLE_SVE)
+#if __aarch64__ && defined(ARM_COMPUTE_ENABLE_SVE)
 
 namespace arm_conv {
 namespace depthwise {
@@ -114,45 +114,29 @@ void sve_fp32_nhwc_3x3_s1_output4x4_mla_depthfirst_direct_impl(
     "ld1rw { z14.s }, p3/Z, [%x[params_struct], %[offsetof_args_max]]\n"
     "add x14, x15, x22, LSL #2\n"
     "ld1w { z13.s }, p3/Z, [x4]\n"
-    "mov z31.d, z13.d\n"
-    "ld1w { z0.s }, p3/Z, [x4, #1, MUL VL]\n"
     "add x13, x14, x22, LSL #2\n"
-    "mov z30.d, z13.d\n"
-    "ld1w { z1.s }, p3/Z, [x4, #2, MUL VL]\n"
+    "ld1w { z0.s }, p3/Z, [x4, #1, MUL VL]\n"
     "add x12, x13, x22, LSL #2\n"
-    "mov z29.d, z13.d\n"
-    "ld1w { z2.s }, p3/Z, [x4, #3, MUL VL]\n"
+    "ld1w { z1.s }, p3/Z, [x4, #2, MUL VL]\n"
     "add x11, x12, x22, LSL #2\n"
-    "mov z28.d, z13.d\n"
-    "ld1w { z3.s }, p3/Z, [x4, #4, MUL VL]\n"
+    "ld1w { z2.s }, p3/Z, [x4, #3, MUL VL]\n"
     "add x10, x7, x7\n"
-    "mov z27.d, z13.d\n"
-    "ld1w { z4.s }, p3/Z, [x4, #5, MUL VL]\n"
+    "ld1w { z3.s }, p3/Z, [x4, #4, MUL VL]\n"
     "add x9, x10, x7\n"
-    "mov z26.d, z13.d\n"
-    "ld1w { z5.s }, p3/Z, [x4, #6, MUL VL]\n"
+    "ld1w { z4.s }, p3/Z, [x4, #5, MUL VL]\n"
     "add x28, x9, x7\n"
-    "mov z25.d, z13.d\n"
-    "ld1w { z6.s }, p3/Z, [x4, #7, MUL VL]\n"
+    "ld1w { z5.s }, p3/Z, [x4, #6, MUL VL]\n"
     "add x27, x28, x7\n"
-    "mov z24.d, z13.d\n"
+    "ld1w { z6.s }, p3/Z, [x4, #7, MUL VL]\n"
     "mul x19, x2, x20\n" // offset = tile_i * ld_output_row
-    "mov z23.d, z13.d\n"
+    "add x26, x17, x17\n"
     "madd x19, x3, x17, x19\n" // offset += tile_j * ld_output_col
-    "mov z22.d, z13.d\n"
     "mul x19, x19, x23\n" // offset *= output_tile_size
-    "mov z21.d, z13.d\n"
     "add x16, x16, x19, LSL #2\n" // outptrs[0] += offset * sizeof(float)
-    "mov z20.d, z13.d\n"
-    "add x26, x16, x20, LSL #2\n"
-    "mov z19.d, z13.d\n"
-    "add x25, x26, x20, LSL #2\n"
-    "mov z18.d, z13.d\n"
+    "add x25, x16, x20, LSL #2\n"
     "add x24, x25, x20, LSL #2\n"
-    "mov z17.d, z13.d\n"
-    "add x23, x17, x17\n"
-    "mov z16.d, z13.d\n"
-    "add x22, x23, x17\n"
+    "add x23, x24, x20, LSL #2\n"
+    "add x22, x26, x17\n"
     "whilelt p2.s, XZR, %x[n_channels]\n"
     "ld1w { z9.s }, p2/Z, [x14, x10, LSL #2]\n"
     "ld1w { z10.s }, p2/Z, [x8]\n"
@@ -165,39 +149,38 @@ void sve_fp32_nhwc_3x3_s1_output4x4_mla_depthfirst_direct_impl(
     "ld1w { z12.s }, p2/Z, [x14, x9, LSL #2]\n"
     "bge 3f\n"
     "2:"  // Tile loop: Channel loop
-    "fmla z31.s, p3/M, z8.s, z9.s\n"
-    "ld1w { z13.s }, p3/Z, [x4]\n"
+    "movprfx z31, z13\n fmla z31.s, p3/M, z8.s, z9.s\n"
     "whilelt p1.s, x6, %x[n_channels]\n"
-    "fmla z30.s, p3/M, z7.s, z9.s\n"
+    "movprfx z30, z13\n fmla z30.s, p3/M, z7.s, z9.s\n"
     "incw x21\n"
-    "fmla z29.s, p3/M, z6.s, z9.s\n"
+    "movprfx z29, z13\n fmla z29.s, p3/M, z6.s, z9.s\n"
     "mov p0.b, p2.b\n"
-    "fmla z27.s, p3/M, z5.s, z9.s\n"
+    "movprfx z27, z13\n fmla z27.s, p3/M, z5.s, z9.s\n"
     "incw x5\n"
-    "fmla z26.s, p3/M, z4.s, z9.s\n"
+    "movprfx z26, z13\n fmla z26.s, p3/M, z4.s, z9.s\n"
     "incw x6\n"
-    "fmla z25.s, p3/M, z3.s, z9.s\n"
-    "fmla z23.s, p3/M, z2.s, z9.s\n"
-    "fmla z22.s, p3/M, z1.s, z9.s\n"
-    "fmla z21.s, p3/M, z0.s, z9.s\n"
+    "movprfx z25, z13\n fmla z25.s, p3/M, z3.s, z9.s\n"
+    "movprfx z23, z13\n fmla z23.s, p3/M, z2.s, z9.s\n"
+    "movprfx z22, z13\n fmla z22.s, p3/M, z1.s, z9.s\n"
+    "movprfx z21, z13\n fmla z21.s, p3/M, z0.s, z9.s\n"
     "ld1w { z9.s }, p2/Z, [x13, x10, LSL #2]\n"
     "fmla z31.s, p3/M, z0.s, z10.s\n"
     "ld1w { z10.s }, p2/Z, [x11]\n"
-    "fmla z28.s, p3/M, z2.s, z11.s\n"
+    "movprfx z28, z13\n fmla z28.s, p3/M, z2.s, z11.s\n"
     "ld1w { z11.s }, p2/Z, [x11, x27, LSL #2]\n"
     "fmla z30.s, p3/M, z8.s, z12.s\n"
     "fmla z29.s, p3/M, z7.s, z12.s\n"
     "fmla z26.s, p3/M, z5.s, z12.s\n"
     "fmla z28.s, p3/M, z6.s, z12.s\n"
     "fmla z25.s, p3/M, z4.s, z12.s\n"
-    "fmla z24.s, p3/M, z3.s, z12.s\n"
+    "movprfx z24, z13\n fmla z24.s, p3/M, z3.s, z12.s\n"
     "fmla z22.s, p3/M, z2.s, z12.s\n"
     "fmla z21.s, p3/M, z1.s, z12.s\n"
-    "fmla z20.s, p3/M, z0.s, z12.s\n"
+    "movprfx z20, z13\n fmla z20.s, p3/M, z0.s, z12.s\n"
     "ld1w { z12.s }, p2/Z, [x8, x7, LSL #2]\n"
-    "fmla z19.s, p3/M, z6.s, z10.s\n"
+    "movprfx z19, z13\n fmla z19.s, p3/M, z6.s, z10.s\n"
     "ld1w { z10.s }, p2/Z, [x13, x9, LSL #2]\n"
-    "fmla z16.s, p3/M, z8.s, z11.s\n"
+    "movprfx z16, z13\n fmla z16.s, p3/M, z8.s, z11.s\n"
     "ld1w { z11.s }, p2/Z, [x8, x28, LSL #2]\n"
     "fmla z27.s, p3/M, z8.s, z9.s\n"
     "fmla z26.s, p3/M, z7.s, z9.s\n"
@@ -206,10 +189,11 @@ void sve_fp32_nhwc_3x3_s1_output4x4_mla_depthfirst_direct_impl(
     "fmla z22.s, p3/M, z4.s, z9.s\n"
     "fmla z21.s, p3/M, z3.s, z9.s\n"
     "fmla z19.s, p3/M, z2.s, z9.s\n"
-    "fmla z18.s, p3/M, z1.s, z9.s\n"
-    "fmla z17.s, p3/M, z0.s, z9.s\n"
+    "movprfx z18, z13\n fmla z18.s, p3/M, z1.s, z9.s\n"
+    "movprfx z17, z13\n fmla z17.s, p3/M, z0.s, z9.s\n"
     "ld1w { z9.s }, p2/Z, [x15]\n"
     "fmla z31.s, p3/M, z1.s, z12.s\n"
+    "ld1w { z13.s }, p3/Z, [x4]\n"
     "fmla z30.s, p3/M, z0.s, z12.s\n"
     "ld1w { z12.s }, p2/Z, [x15, x27, LSL #2]\n"
     "fmla z29.s, p3/M, z2.s, z11.s\n"
@@ -375,109 +359,93 @@ void sve_fp32_nhwc_3x3_s1_output4x4_mla_depthfirst_direct_impl(
     "addvl x4, x4, #-6\n"
     "fmin z31.s, p3/M, z31.s, z14.s\n"
     "st1w { z31.s }, p0, [x16]\n"
-    "mov z31.d, z13.d\n"
     "fmin z30.s, p3/M, z30.s, z14.s\n"
-    "st1w { z30.s }, p0, [x16, x17, LSL #2]\n"
-    "mov z30.d, z13.d\n"
     "fmin z29.s, p3/M, z29.s, z14.s\n"
-    "st1w { z29.s }, p0, [x16, x23, LSL #2]\n"
-    "mov z29.d, z13.d\n"
+    "st1w { z30.s }, p0, [x16, x17, LSL #2]\n"
     "fmax z28.s, p3/M, z28.s, z15.s\n"
     "fmax z27.s, p3/M, z27.s, z15.s\n"
+    "st1w { z29.s }, p0, [x16, x26, LSL #2]\n"
     "fmax z26.s, p3/M, z26.s, z15.s\n"
     "fmax z25.s, p3/M, z25.s, z15.s\n"
+    "fmax z24.s, p3/M, z24.s, z15.s\n"
     "fmin z28.s, p3/M, z28.s, z14.s\n"
     "st1w { z28.s }, p0, [x16, x22, LSL #2]\n"
-    "mov z28.d, z13.d\n"
-    "addvl x16, x16, #1\n"
     "fmin z27.s, p3/M, z27.s, z14.s\n"
-    "st1w { z27.s }, p0, [x26]\n"
-    "mov z27.d, z13.d\n"
+    "addvl x16, x16, #1\n"
     "fmin z26.s, p3/M, z26.s, z14.s\n"
-    "st1w { z26.s }, p0, [x26, x17, LSL #2]\n"
-    "mov z26.d, z13.d\n"
+    "st1w { z27.s }, p0, [x25]\n"
     "fmin z25.s, p3/M, z25.s, z14.s\n"
-    "st1w { z25.s }, p0, [x26, x23, LSL #2]\n"
-    "mov z25.d, z13.d\n"
-    "fmax z24.s, p3/M, z24.s, z15.s\n"
+    "fmin z24.s, p3/M, z24.s, z14.s\n"
+    "st1w { z26.s }, p0, [x25, x17, LSL #2]\n"
     "fmax z23.s, p3/M, z23.s, z15.s\n"
+    "st1w { z25.s }, p0, [x25, x26, LSL #2]\n"
     "fmax z22.s, p3/M, z22.s, z15.s\n"
     "fmax z21.s, p3/M, z21.s, z15.s\n"
-    "fmin z24.s, p3/M, z24.s, z14.s\n"
-    "st1w { z24.s }, p0, [x26, x22, LSL #2]\n"
-    "mov z24.d, z13.d\n"
-    "addvl x26, x26, #1\n"
+    "st1w { z24.s }, p0, [x25, x22, LSL #2]\n"
+    "addvl x25, x25, #1\n"
     "fmin z23.s, p3/M, z23.s, z14.s\n"
-    "st1w { z23.s }, p0, [x25]\n"
-    "mov z23.d, z13.d\n"
+    "st1w { z23.s }, p0, [x24]\n"
     "fmin z22.s, p3/M, z22.s, z14.s\n"
-    "st1w { z22.s }, p0, [x25, x17, LSL #2]\n"
-    "mov z22.d, z13.d\n"
     "fmin z21.s, p3/M, z21.s, z14.s\n"
-    "st1w { z21.s }, p0, [x25, x23, LSL #2]\n"
-    "mov z21.d, z13.d\n"
+    "st1w { z22.s }, p0, [x24, x17, LSL #2]\n"
     "fmax z20.s, p3/M, z20.s, z15.s\n"
     "fmax z19.s, p3/M, z19.s, z15.s\n"
+    "st1w { z21.s }, p0, [x24, x26, LSL #2]\n"
     "fmax z18.s, p3/M, z18.s, z15.s\n"
     "fmax z17.s, p3/M, z17.s, z15.s\n"
+    "fmax z16.s, p3/M, z16.s, z15.s\n"
     "fmin z20.s, p3/M, z20.s, z14.s\n"
-    "st1w { z20.s }, p0, [x25, x22, LSL #2]\n"
-    "mov z20.d, z13.d\n"
-    "addvl x25, x25, #1\n"
+    "st1w { z20.s }, p0, [x24, x22, LSL #2]\n"
     "fmin z19.s, p3/M, z19.s, z14.s\n"
-    "st1w { z19.s }, p0, [x24]\n"
-    "mov z19.d, z13.d\n"
+    "addvl x24, x24, #1\n"
     "fmin z18.s, p3/M, z18.s, z14.s\n"
-    "st1w { z18.s }, p0, [x24, x17, LSL #2]\n"
-    "mov z18.d, z13.d\n"
+    "st1w { z19.s }, p0, [x23]\n"
     "fmin z17.s, p3/M, z17.s, z14.s\n"
-    "st1w { z17.s }, p0, [x24, x23, LSL #2]\n"
-    "mov z17.d, z13.d\n"
-    "fmax z16.s, p3/M, z16.s, z15.s\n"
     "fmin z16.s, p3/M, z16.s, z14.s\n"
-    "st1w { z16.s }, p0, [x24, x22, LSL #2]\n"
-    "mov z16.d, z13.d\n"
-    "addvl x24, x24, #1\n"
+    "st1w { z18.s }, p0, [x23, x17, LSL #2]\n"
+    "st1w { z17.s }, p0, [x23, x26, LSL #2]\n"
+    "st1w { z16.s }, p0, [x23, x22, LSL #2]\n"
+    "addvl x23, x23, #1\n"
     "blt 2b\n"
     "3:"  // Tile loop: Channel tail
-    "fmla z31.s, p3/M, z8.s, z9.s\n"
+    "movprfx z31, z13\n fmla z31.s, p3/M, z8.s, z9.s\n"
     "ldr x2, [%x[params_struct], %[offsetof_args_tile_i]]\n"
     "mov p0.b, p2.b\n"
-    "fmla z30.s, p3/M, z7.s, z9.s\n"
+    "movprfx z30, z13\n fmla z30.s, p3/M, z7.s, z9.s\n"
     "ldr x3, [%x[params_struct], %[offsetof_args_tile_j]]\n"
     "add x21, x2, #0x1\n"
-    "fmla z29.s, p3/M, z6.s, z9.s\n"
+    "movprfx z29, z13\n fmla z29.s, p3/M, z6.s, z9.s\n"
     "ldr x20, [%x[params_struct], %[offsetof_args_n_tile_rows]]\n"
-    "fmla z27.s, p3/M, z5.s, z9.s\n"
+    "movprfx z27, z13\n fmla z27.s, p3/M, z5.s, z9.s\n"
     "ldr x19, [%x[params_struct], %[offsetof_args_n_tile_cols]]\n"
     "add x3, x3, #0x1\n"
-    "fmla z26.s, p3/M, z4.s, z9.s\n"
+    "movprfx z26, z13\n fmla z26.s, p3/M, z4.s, z9.s\n"
     "cmp x3, x19\n"
-    "fmla z25.s, p3/M, z3.s, z9.s\n"
-    "fmla z23.s, p3/M, z2.s, z9.s\n"
+    "movprfx z25, z13\n fmla z25.s, p3/M, z3.s, z9.s\n"
+    "movprfx z23, z13\n fmla z23.s, p3/M, z2.s, z9.s\n"
     "csel x3, x3, XZR, LT\n"
-    "fmla z22.s, p3/M, z1.s, z9.s\n"
+    "movprfx z22, z13\n fmla z22.s, p3/M, z1.s, z9.s\n"
     "csel x2, x2, x21, LT\n"
-    "fmla z21.s, p3/M, z0.s, z9.s\n"
+    "movprfx z21, z13\n fmla z21.s, p3/M, z0.s, z9.s\n"
     "ld1w { z9.s }, p2/Z, [x13, x10, LSL #2]\n"
     "cmp x2, x20\n"
     "fmla z31.s, p3/M, z0.s, z10.s\n"
     "ld1w { z10.s }, p2/Z, [x11]\n"
-    "fmla z28.s, p3/M, z2.s, z11.s\n"
+    "movprfx z28, z13\n fmla z28.s, p3/M, z2.s, z11.s\n"
     "ld1w { z11.s }, p2/Z, [x11, x27, LSL #2]\n"
     "fmla z30.s, p3/M, z8.s, z12.s\n"
     "fmla z29.s, p3/M, z7.s, z12.s\n"
     "fmla z26.s, p3/M, z5.s, z12.s\n"
     "fmla z28.s, p3/M, z6.s, z12.s\n"
     "fmla z25.s, p3/M, z4.s, z12.s\n"
-    "fmla z24.s, p3/M, z3.s, z12.s\n"
+    "movprfx z24, z13\n fmla z24.s, p3/M, z3.s, z12.s\n"
     "fmla z22.s, p3/M, z2.s, z12.s\n"
     "fmla z21.s, p3/M, z1.s, z12.s\n"
-    "fmla z20.s, p3/M, z0.s, z12.s\n"
+    "movprfx z20, z13\n fmla z20.s, p3/M, z0.s, z12.s\n"
     "ld1w { z12.s }, p2/Z, [x8, x7, LSL #2]\n"
-    "fmla z19.s, p3/M, z6.s, z10.s\n"
+    "movprfx z19, z13\n fmla z19.s, p3/M, z6.s, z10.s\n"
     "ld1w { z10.s }, p2/Z, [x13, x9, LSL #2]\n"
-    "fmla z16.s, p3/M, z8.s, z11.s\n"
+    "movprfx z16, z13\n fmla z16.s, p3/M, z8.s, z11.s\n"
     "ld1w { z11.s }, p2/Z, [x8, x28, LSL #2]\n"
     "fmla z27.s, p3/M, z8.s, z9.s\n"
     "fmla z26.s, p3/M, z7.s, z9.s\n"
@@ -486,8 +454,8 @@ void sve_fp32_nhwc_3x3_s1_output4x4_mla_depthfirst_direct_impl(
     "fmla z22.s, p3/M, z4.s, z9.s\n"
     "fmla z21.s, p3/M, z3.s, z9.s\n"
     "fmla z19.s, p3/M, z2.s, z9.s\n"
-    "fmla z18.s, p3/M, z1.s, z9.s\n"
-    "fmla z17.s, p3/M, z0.s, z9.s\n"
+    "movprfx z18, z13\n fmla z18.s, p3/M, z1.s, z9.s\n"
+    "movprfx z17, z13\n fmla z17.s, p3/M, z0.s, z9.s\n"
     "ld1w { z9.s }, p2/Z, [x15]\n"
     "fmla z31.s, p3/M, z1.s, z12.s\n"
     "fmla z30.s, p3/M, z0.s, z12.s\n"
@@ -638,43 +606,43 @@ void sve_fp32_nhwc_3x3_s1_output4x4_mla_depthfirst_direct_impl(
     "st1w { z30.s }, p0, [x16, x17, LSL #2]\n"
     "fmin z28.s, p3/M, z28.s, z14.s\n"
     "fmax z27.s, p3/M, z27.s, z15.s\n"
-    "st1w { z29.s }, p0, [x16, x23, LSL #2]\n"
+    "st1w { z29.s }, p0, [x16, x26, LSL #2]\n"
     "fmax z26.s, p3/M, z26.s, z15.s\n"
     "st1w { z28.s }, p0, [x16, x22, LSL #2]\n"
     "fmin z27.s, p3/M, z27.s, z14.s\n"
     "fmax z25.s, p3/M, z25.s, z15.s\n"
-    "st1w { z27.s }, p0, [x26]\n"
+    "st1w { z27.s }, p0, [x25]\n"
     "fmin z26.s, p3/M, z26.s, z14.s\n"
     "fmin z25.s, p3/M, z25.s, z14.s\n"
-    "st1w { z26.s }, p0, [x26, x17, LSL #2]\n"
+    "st1w { z26.s }, p0, [x25, x17, LSL #2]\n"
     "fmax z24.s, p3/M, z24.s, z15.s\n"
     "fmax z23.s, p3/M, z23.s, z15.s\n"
-    "st1w { z25.s }, p0, [x26, x23, LSL #2]\n"
+    "st1w { z25.s }, p0, [x25, x26, LSL #2]\n"
     "fmax z22.s, p3/M, z22.s, z15.s\n"
     "fmax z21.s, p3/M, z21.s, z15.s\n"
     "fmax z20.s, p3/M, z20.s, z15.s\n"
     "fmin z24.s, p3/M, z24.s, z14.s\n"
-    "st1w { z24.s }, p0, [x26, x22, LSL #2]\n"
+    "st1w { z24.s }, p0, [x25, x22, LSL #2]\n"
     "fmin z23.s, p3/M, z23.s, z14.s\n"
     "fmin z22.s, p3/M, z22.s, z14.s\n"
-    "st1w { z23.s }, p0, [x25]\n"
+    "st1w { z23.s }, p0, [x24]\n"
     "fmin z21.s, p3/M, z21.s, z14.s\n"
     "fmin z20.s, p3/M, z20.s, z14.s\n"
-    "st1w { z22.s }, p0, [x25, x17, LSL #2]\n"
+    "st1w { z22.s }, p0, [x24, x17, LSL #2]\n"
     "fmax z19.s, p3/M, z19.s, z15.s\n"
-    "st1w { z21.s }, p0, [x25, x23, LSL #2]\n"
+    "st1w { z21.s }, p0, [x24, x26, LSL #2]\n"
     "fmax z18.s, p3/M, z18.s, z15.s\n"
     "fmax z17.s, p3/M, z17.s, z15.s\n"
-    "st1w { z20.s }, p0, [x25, x22, LSL #2]\n"
+    "st1w { z20.s }, p0, [x24, x22, LSL #2]\n"
     "fmin z19.s, p3/M, z19.s, z14.s\n"
-    "st1w { z19.s }, p0, [x24]\n"
+    "st1w { z19.s }, p0, [x23]\n"
     "fmin z18.s, p3/M, z18.s, z14.s\n"
     "fmin z17.s, p3/M, z17.s, z14.s\n"
-    "st1w { z18.s }, p0, [x24, x17, LSL #2]\n"
+    "st1w { z18.s }, p0, [x23, x17, LSL #2]\n"
     "fmax z16.s, p3/M, z16.s, z15.s\n"
-    "st1w { z17.s }, p0, [x24, x23, LSL #2]\n"
+    "st1w { z17.s }, p0, [x23, x26, LSL #2]\n"
     "fmin z16.s, p3/M, z16.s, z14.s\n"
-    "st1w { z16.s }, p0, [x24, x22, LSL #2]\n"
+    "st1w { z16.s }, p0, [x23, x22, 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)
@@ -685,4 +653,4 @@ void sve_fp32_nhwc_3x3_s1_output4x4_mla_depthfirst_direct_impl(
 }  // namespace depthwise
 }  // namespace arm_conv
 
-#endif  // defined(ARM_COMPUTE_ENABLE_SVE)
+#endif  // __aarch64__ && defined(ARM_COMPUTE_ENABLE_SVE)