From be13cead34e566bdd561ad3ffc3f645b460e482e Mon Sep 17 00:00:00 2001
From: Michael Tyler <michael.tyler@arm.com>
Date: Tue, 17 Jan 2023 11:04:14 +0000
Subject: 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>
---
 .../generic_indirect.cpp                           | 820 +++++++++++----------
 1 file changed, 411 insertions(+), 409 deletions(-)

(limited to 'src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_3x3_s2_output2x2_mla_depthfirst/generic_indirect.cpp')

diff --git a/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_3x3_s2_output2x2_mla_depthfirst/generic_indirect.cpp b/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_3x3_s2_output2x2_mla_depthfirst/generic_indirect.cpp
index f65633002e..4b24862eb2 100644
--- a/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_3x3_s2_output2x2_mla_depthfirst/generic_indirect.cpp
+++ b/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_3x3_s2_output2x2_mla_depthfirst/generic_indirect.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2021, 2023 Arm Limited.
+ * Copyright (c) 2021 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -88,385 +88,385 @@ void a64_fp32_nhwc_3x3_s2_output2x2_mla_depthfirst_indirect_impl(
 
   __asm__ __volatile__(
     "ldr x21, [%x[params_struct], %[offsetof_args_outptrs]]\n"
-    "mov x26, #0x10\n"  // cntb _, ALL, #1
-    "lsr x25, %x[n_channels], #0x2\n"
-    "ldr x24, [%x[params_struct], %[offsetof_args_params]]\n"
+    "add x16, %x[params_struct], %[offsetof_Args_inptrs]\n"
+    "ldr x15, [%x[params_struct], %[offsetof_args_params]]\n"
     "add x20, %x[params_struct], %[offsetof_args_min]\n"
+    "add x19, %x[params_struct], %[offsetof_args_max]\n"
     "ld1r { v19.4s }, [x20]\n"
-    "add x20, %x[params_struct], %[offsetof_args_max]\n"
-    "ld1r { v18.4s }, [x20]\n"
-    "add x13, %x[params_struct], %[offsetof_Args_inptrs]\n"
-    "ldp x12, x11, [x21, #0x0]\n"
+    "ld1r { v18.4s }, [x19]\n"
+    "mov x14, #0x0\n"
+    "ldp x13, x12, [x21, #0x0]\n"
+    "mov x11, #0x10\n" // cntb _, ALL, #1
     "ldp x10, x9, [x21, #0x10]\n"
-    "mov x28, #0x0\n"
-    "sub x23, XZR, x26\n"
-    "cbz x25, 3f\n"
-    "ldr q17, [x24, #0x0]\n"
-    "ldr q0, [x24, #0x10]\n"
-    "cmp x26, x25, LSL #4\n"
-    "ldr q1, [x24, #0x20]\n"
-    "ldr q2, [x24, #0x30]\n"
-    "ldr q3, [x24, #0x40]\n"
-    "ldr q4, [x24, #0x50]\n"
-    "ldr q5, [x24, #0x60]\n"
-    "ldr q6, [x24, #0x70]\n"
-    "ldr q7, [x24, #0x80]\n"
-    "ldr q8, [x24, #0x90]\n"
-    "add x24, x24, #0xa0\n"
-    "ldp x22, x20, [x13, #0x0]\n"
-    "ldr q9, [x22, x28]\n"
-    "ldr q10, [x20, x28]\n"
-    "ldp x21, x20, [x13, #0x10]\n"
-    "ldr q11, [x21, x28]\n"
-    "ldr q12, [x20, x28]\n"
-    "ldp x22, x21, [x13, #0x20]\n"
-    "ldr q13, [x22, x28]\n"
-    "ldr q14, [x21, x28]\n"
-    "ldp x21, x20, [x13, #0x30]\n"
-    "ldr q15, [x21, x28]\n"
-    "ldr q16, [x20, x28]\n"
+    "sub x28, XZR, x11\n"
+    "lsr x27, %x[n_channels], #0x2\n"
+    "cbz x27, 3f\n"
+    "ldr q17, [x15, #0x0]\n"
+    "ldr q0, [x15, #0x10]\n"
+    "cmp x11, x27, LSL #4\n"
+    "ldr q1, [x15, #0x20]\n"
+    "ldr q2, [x15, #0x30]\n"
+    "ldr q3, [x15, #0x40]\n"
+    "ldr q4, [x15, #0x50]\n"
+    "ldr q5, [x15, #0x60]\n"
+    "ldr q6, [x15, #0x70]\n"
+    "ldr q7, [x15, #0x80]\n"
+    "ldr q8, [x15, #0x90]\n"
+    "add x15, x15, #0xa0\n"
+    "ldp x26, x25, [x16, #0x0]\n"
+    "ldp x24, x23, [x16, #0x10]\n"
+    "ldp x22, x21, [x16, #0x20]\n"
+    "ldr q9, [x26, x14]\n"
+    "ldr q10, [x25, x14]\n"
+    "ldr q11, [x24, x14]\n"
+    "ldr q12, [x23, x14]\n"
+    "ldr q13, [x22, x14]\n"
+    "ldr q14, [x21, x14]\n"
+    "ldp x20, x19, [x16, #0x30]\n"
+    "ldr q15, [x20, x14]\n"
+    "ldr q16, [x19, x14]\n"
     "bge 2f\n"
     "1:"  // Channel loop
-    "mov v28.16b, v17.16b\n fmla v28.4s, v8.4s, v9.4s\n"
-    "mov v29.16b, v17.16b\n fmla v29.4s, v6.4s, v9.4s\n"
-    "ldr x22, [x13, #0x40]\n"
-    "ldr x20, [x13, #0x48]\n"
-    "fmla v28.4s, v0.4s, v10.4s\n"
-    "fmla v29.4s, v1.4s, v12.4s\n"
-    "ldr q12, [x20, x28]\n"
-    "ldr x21, [x13, #0x50]\n"
-    "fmla v28.4s, v1.4s, v11.4s\n"
-    "ldr q11, [x22, x28]\n"
-    "fmla v29.4s, v2.4s, v13.4s\n"
-    "ldr q13, [x21, x28]\n"
-    "fmla v28.4s, v3.4s, v14.4s\n"
-    "fmla v29.4s, v0.4s, v16.4s\n"
-    "ldr x20, [x13, #0x58]\n"
-    "ldr q14, [x20, x28]\n"
-    "fmla v28.4s, v4.4s, v15.4s\n"
-    "fmla v29.4s, v4.4s, v11.4s\n"
-    "ldr x20, [x13, #0x78]\n"
-    "ldr x22, [x13, #0x60]\n"
-    "ldr q15, [x22, x28]\n"
-    "fmla v28.4s, v2.4s, v16.4s\n"
-    "fmla v29.4s, v5.4s, v12.4s\n"
-    "ldr x22, [x13, #0x80]\n"
-    "ldr q12, [x22, x28]\n"
-    "mov v30.16b, v17.16b\n fmla v30.4s, v2.4s, v9.4s\n"
-    "mov v31.16b, v17.16b\n fmla v31.4s, v0.4s, v9.4s\n"
-    "ldr q17, [x24, #0x0]\n"
-    "fmla v28.4s, v5.4s, v13.4s\n"
-    "fmla v29.4s, v3.4s, v13.4s\n"
-    "ldr q13, [x20, x28]\n"
-    "ldr x21, [x13, #0x68]\n"
-    "ldr q11, [x21, x28]\n"
-    "fmla v30.4s, v3.4s, v14.4s\n"
-    "fmla v31.4s, v4.4s, v13.4s\n"
-    "ldr x20, [x13, #0x88]\n"
-    "ldr q14, [x20, x28]\n"
-    "fmla v30.4s, v0.4s, v15.4s\n"
-    "ldr q0, [x24, #0x10]\n"
-    "fmla v31.4s, v1.4s, v12.4s\n"
-    "ldr x21, [x13, #0x70]\n"
-    "ldr q16, [x21, x28]\n"
+    "mov v31.16b, v17.16b\n fmla v31.4s, v8.4s, v9.4s\n"
+    "ldr x26, [x16, #0x40]\n"
+    "add x28, x28, #0x10\n"
+    "mov v30.16b, v17.16b\n fmla v30.4s, v6.4s, v9.4s\n"
+    "ldr x25, [x16, #0x48]\n"
+    "mov v29.16b, v17.16b\n fmla v29.4s, v2.4s, v9.4s\n"
+    "ldr x24, [x16, #0x50]\n"
+    "mov v28.16b, v17.16b\n fmla v28.4s, v0.4s, v9.4s\n"
+    "ldr x23, [x16, #0x58]\n"
+    "ldr x22, [x16, #0x60]\n"
+    "fmla v31.4s, v0.4s, v10.4s\n"
+    "ldr x21, [x16, #0x68]\n"
+    "fmla v30.4s, v1.4s, v12.4s\n"
+    "ldr q12, [x25, x14]\n"
+    "fmla v31.4s, v1.4s, v11.4s\n"
+    "ldr q11, [x26, x14]\n"
+    "ldr x20, [x16, #0x70]\n"
+    "fmla v30.4s, v2.4s, v13.4s\n"
+    "ldr q13, [x24, x14]\n"
+    "fmla v31.4s, v3.4s, v14.4s\n"
+    "ldr q14, [x23, x14]\n"
+    "ldr x19, [x16, #0x78]\n"
+    "fmla v30.4s, v0.4s, v16.4s\n"
+    "ldr x26, [x16, #0x80]\n"
+    "fmla v31.4s, v4.4s, v15.4s\n"
+    "ldr q15, [x22, x14]\n"
+    "fmla v29.4s, v3.4s, v14.4s\n"
+    "ldr x25, [x16, #0x88]\n"
     "fmla v30.4s, v4.4s, v11.4s\n"
-    "fmla v31.4s, v5.4s, v14.4s\n"
-    "ldr q4, [x24, #0x50]\n"
-    "ldr x20, [x13, #0x98]\n"
-    "fmla v28.4s, v6.4s, v15.4s\n"
-    "fmla v30.4s, v1.4s, v16.4s\n"
-    "ldr q11, [x20, x28]\n"
-    "ldr q1, [x24, #0x20]\n"
-    "fmla v31.4s, v2.4s, v11.4s\n"
-    "fmla v28.4s, v7.4s, v16.4s\n"
-    "ldr q2, [x24, #0x30]\n"
-    "ldr x21, [x13, #0x90]\n"
-    "fmla v29.4s, v7.4s, v12.4s\n"
-    "fmla v29.4s, v8.4s, v11.4s\n"
-    "ldr q15, [x21, x28]\n"
-    "ldr x21, [x13, #0xa8]\n"
-    "fmla v30.4s, v6.4s, v15.4s\n"
-    "fmax v28.4s, v28.4s, v19.4s\n"
-    "ldr q16, [x21, x28]\n"
-    "ldr x22, [x13, #0xa0]\n"
-    "fmla v31.4s, v3.4s, v16.4s\n"
-    "fmax v29.4s, v29.4s, v19.4s\n"
-    "ldr q13, [x22, x28]\n"
-    "ldr q3, [x24, #0x40]\n"
-    "fmla v30.4s, v7.4s, v13.4s\n"
-    "fmla v30.4s, v5.4s, v16.4s\n"
-    "ldr q5, [x24, #0x60]\n"
-    "ldr x21, [x13, #0xb0]\n"
-    "add x23, x23, #0x10\n"
-    "fmin v28.4s, v28.4s, v18.4s\n"
-    "ldr q14, [x21, x28]\n"
-    "ldr x20, [x13, #0xb8]\n"
-    "fmla v31.4s, v7.4s, v14.4s\n"
-    "fmin v29.4s, v29.4s, v18.4s\n"
-    "ldr q15, [x20, x28]\n"
-    "ldr q7, [x24, #0x80]\n"
+    "ldr q11, [x21, x14]\n"
+    "ldr x24, [x16, #0x90]\n"
+    "fmla v31.4s, v2.4s, v16.4s\n"
+    "ldr q16, [x20, x14]\n"
+    "fmla v29.4s, v0.4s, v15.4s\n"
+    "ldr q14, [x25, x14]\n"
+    "fmla v30.4s, v5.4s, v12.4s\n"
+    "ldr q12, [x26, x14]\n"
+    "ldr x23, [x16, #0x98]\n"
+    "fmla v31.4s, v5.4s, v13.4s\n"
+    "ldr x22, [x16, #0xa0]\n"
+    "fmla v29.4s, v4.4s, v11.4s\n"
+    "ldr q11, [x23, x14]\n"
+    "fmla v30.4s, v3.4s, v13.4s\n"
+    "ldr q13, [x19, x14]\n"
+    "ldr x21, [x16, #0xa8]\n"
     "fmla v31.4s, v6.4s, v15.4s\n"
-    "fmla v30.4s, v8.4s, v15.4s\n"
-    "ldr q6, [x24, #0x70]\n"
-    "ldr x22, [x13, #0xc0]\n"
-    "fmax v30.4s, v30.4s, v19.4s\n"
-    "fmin v30.4s, v30.4s, v18.4s\n"
-    "ldr q11, [x22, x28]\n"
-    "fmla v31.4s, v8.4s, v11.4s\n"
-    "ldr q8, [x24, #0x90]\n"
+    "ldr q15, [x24, x14]\n"
+    "fmla v29.4s, v1.4s, v16.4s\n"
+    "ldr x20, [x16, #0xb0]\n"
+    "fmla v30.4s, v7.4s, v12.4s\n"
+    "ldr x19, [x16, #0xb8]\n"
+    "fmla v28.4s, v4.4s, v13.4s\n"
+    "ldr q13, [x22, x14]\n"
+    "ldr x26, [x16, #0xc0]\n"
+    "fmla v31.4s, v7.4s, v16.4s\n"
+    "fmla v29.4s, v6.4s, v15.4s\n"
+    "ldr q16, [x21, x14]\n"
+    "fmla v30.4s, v8.4s, v11.4s\n"
+    "ldr q15, [x19, x14]\n"
+    "fmla v28.4s, v1.4s, v12.4s\n"
+    "ldr q17, [x15, #0x0]\n"
+    "ldr q0, [x15, #0x10]\n"
+    "fmla v29.4s, v7.4s, v13.4s\n"
     "fmax v31.4s, v31.4s, v19.4s\n"
-    "ldp x22, x20, [x13, #0x0]\n"
-    "ldr q9, [x22, x26]\n"
+    "ldr q1, [x15, #0x20]\n"
+    "fmax v30.4s, v30.4s, v19.4s\n"
+    "ldr q4, [x15, #0x50]\n"
+    "fmla v28.4s, v5.4s, v14.4s\n"
+    "ldr q14, [x20, x14]\n"
     "fmin v31.4s, v31.4s, v18.4s\n"
-    "add x28, x28, #0x10\n"
-    "ldr q10, [x20, x26]\n"
-    "ldp x21, x20, [x13, #0x10]\n"
-    "str q28, [x12, x23]\n"
-    "add x24, x24, #0xa0\n"
-    "ldr q11, [x21, x26]\n"
-    "ldr q12, [x20, x26]\n"
-    "str q29, [x11, x23]\n"
-    "ldp x22, x21, [x13, #0x20]\n"
-    "ldr q13, [x22, x26]\n"
-    "str q30, [x10, x23]\n"
-    "ldr q14, [x21, x26]\n"
-    "ldp x21, x20, [x13, #0x30]\n"
-    "str q31, [x9, x23]\n"
-    "ldr q15, [x21, x26]\n"
-    "ldr q16, [x20, x26]\n"
-    "add x26, x26, #0x10\n"
-    "cmp x26, x25, LSL #4\n"
+    "str q31, [x13, x28]\n"
+    "fmla v28.4s, v2.4s, v11.4s\n"
+    "fmla v29.4s, v5.4s, v16.4s\n"
+    "ldr q11, [x26, x14]\n"
+    "add x14, x14, #0x10\n"
+    "fmin v30.4s, v30.4s, v18.4s\n"
+    "ldp x26, x25, [x16, #0x0]\n"
+    "ldp x24, x23, [x16, #0x10]\n"
+    "fmla v28.4s, v3.4s, v16.4s\n"
+    "ldp x22, x21, [x16, #0x20]\n"
+    "fmla v29.4s, v8.4s, v15.4s\n"
+    "ldr q9, [x26, x11]\n"
+    "ldr q10, [x25, x11]\n"
+    "fmla v28.4s, v7.4s, v14.4s\n"
+    "ldr q12, [x23, x11]\n"
+    "fmax v29.4s, v29.4s, v19.4s\n"
+    "ldr q13, [x22, x11]\n"
+    "ldr q14, [x21, x11]\n"
+    "fmin v29.4s, v29.4s, v18.4s\n"
+    "ldp x20, x19, [x16, #0x30]\n"
+    "str q30, [x12, x28]\n"
+    "fmla v28.4s, v6.4s, v15.4s\n"
+    "ldr q2, [x15, #0x30]\n"
+    "fmla v28.4s, v8.4s, v11.4s\n"
+    "ldr q11, [x24, x11]\n"
+    "ldr q15, [x20, x11]\n"
+    "fmax v28.4s, v28.4s, v19.4s\n"
+    "ldr q16, [x19, x11]\n"
+    "add x11, x11, #0x10\n"
+    "fmin v28.4s, v28.4s, v18.4s\n"
+    "str q29, [x10, x28]\n"
+    "cmp x11, x27, LSL #4\n"
+    "ldr q3, [x15, #0x40]\n"
+    "ldr q5, [x15, #0x60]\n"
+    "ldr q6, [x15, #0x70]\n"
+    "str q28, [x9, x28]\n"
+    "ldr q7, [x15, #0x80]\n"
+    "ldr q8, [x15, #0x90]\n"
+    "add x15, x15, #0xa0\n"
     "blt 1b\n"
     "2:"  // Channel tail
-    "mov v28.16b, v17.16b\n fmla v28.4s, v8.4s, v9.4s\n"
-    "mov v29.16b, v17.16b\n fmla v29.4s, v6.4s, v9.4s\n"
-    "ldr x22, [x13, #0x40]\n"
-    "ldr x20, [x13, #0x48]\n"
-    "fmla v28.4s, v0.4s, v10.4s\n"
-    "fmla v29.4s, v1.4s, v12.4s\n"
-    "ldr q12, [x20, x28]\n"
-    "ldr x21, [x13, #0x50]\n"
-    "fmla v28.4s, v1.4s, v11.4s\n"
-    "ldr q11, [x22, x28]\n"
-    "fmla v29.4s, v2.4s, v13.4s\n"
-    "ldr q13, [x21, x28]\n"
-    "fmla v28.4s, v3.4s, v14.4s\n"
-    "fmla v29.4s, v0.4s, v16.4s\n"
-    "ldr x20, [x13, #0x58]\n"
-    "ldr q14, [x20, x28]\n"
-    "fmla v28.4s, v4.4s, v15.4s\n"
-    "fmla v29.4s, v4.4s, v11.4s\n"
-    "ldr x20, [x13, #0x78]\n"
-    "ldr x22, [x13, #0x60]\n"
-    "ldr q15, [x22, x28]\n"
-    "fmla v28.4s, v2.4s, v16.4s\n"
-    "fmla v29.4s, v5.4s, v12.4s\n"
-    "ldr x22, [x13, #0x80]\n"
-    "ldr q12, [x22, x28]\n"
-    "mov v30.16b, v17.16b\n fmla v30.4s, v2.4s, v9.4s\n"
-    "mov v31.16b, v17.16b\n fmla v31.4s, v0.4s, v9.4s\n"
-    "ldr x21, [x13, #0x68]\n"
-    "ldr q11, [x21, x28]\n"
-    "fmla v28.4s, v5.4s, v13.4s\n"
-    "fmla v29.4s, v3.4s, v13.4s\n"
-    "ldr q13, [x20, x28]\n"
-    "fmla v30.4s, v3.4s, v14.4s\n"
-    "fmla v31.4s, v4.4s, v13.4s\n"
-    "ldr x20, [x13, #0x88]\n"
-    "ldr q14, [x20, x28]\n"
-    "fmla v30.4s, v0.4s, v15.4s\n"
-    "fmla v31.4s, v1.4s, v12.4s\n"
-    "ldr x21, [x13, #0x70]\n"
-    "ldr q16, [x21, x28]\n"
-    "ldr x20, [x13, #0x98]\n"
+    "mov v31.16b, v17.16b\n fmla v31.4s, v8.4s, v9.4s\n"
+    "ldr x26, [x16, #0x40]\n"
+    "add x28, x28, #0x10\n"
+    "mov v30.16b, v17.16b\n fmla v30.4s, v6.4s, v9.4s\n"
+    "ldr x25, [x16, #0x48]\n"
+    "mov v29.16b, v17.16b\n fmla v29.4s, v2.4s, v9.4s\n"
+    "ldr x24, [x16, #0x50]\n"
+    "mov v28.16b, v17.16b\n fmla v28.4s, v0.4s, v9.4s\n"
+    "ldr x23, [x16, #0x58]\n"
+    "ldr x22, [x16, #0x60]\n"
+    "fmla v31.4s, v0.4s, v10.4s\n"
+    "ldr x21, [x16, #0x68]\n"
+    "fmla v30.4s, v1.4s, v12.4s\n"
+    "ldr q12, [x25, x14]\n"
+    "fmla v31.4s, v1.4s, v11.4s\n"
+    "ldr q11, [x26, x14]\n"
+    "ldr x20, [x16, #0x70]\n"
+    "fmla v30.4s, v2.4s, v13.4s\n"
+    "ldr q13, [x24, x14]\n"
+    "fmla v31.4s, v3.4s, v14.4s\n"
+    "ldr q14, [x23, x14]\n"
+    "ldr x19, [x16, #0x78]\n"
+    "fmla v30.4s, v0.4s, v16.4s\n"
+    "ldr x26, [x16, #0x80]\n"
+    "fmla v31.4s, v4.4s, v15.4s\n"
+    "ldr q15, [x22, x14]\n"
+    "fmla v29.4s, v3.4s, v14.4s\n"
+    "ldr x25, [x16, #0x88]\n"
     "fmla v30.4s, v4.4s, v11.4s\n"
-    "ldr q11, [x20, x28]\n"
-    "fmla v31.4s, v5.4s, v14.4s\n"
-    "fmla v28.4s, v6.4s, v15.4s\n"
-    "ldr x21, [x13, #0x90]\n"
-    "ldr q15, [x21, x28]\n"
-    "fmla v30.4s, v1.4s, v16.4s\n"
-    "ldr x21, [x13, #0xa8]\n"
-    "fmla v31.4s, v2.4s, v11.4s\n"
-    "fmla v28.4s, v7.4s, v16.4s\n"
-    "ldr q16, [x21, x28]\n"
-    "ldr x22, [x13, #0xa0]\n"
-    "ldr q13, [x22, x28]\n"
-    "fmla v30.4s, v6.4s, v15.4s\n"
-    "fmla v31.4s, v3.4s, v16.4s\n"
-    "ldr x21, [x13, #0xb0]\n"
-    "ldr q14, [x21, x28]\n"
-    "fmla v30.4s, v7.4s, v13.4s\n"
-    "fmla v31.4s, v7.4s, v14.4s\n"
-    "ldr x20, [x13, #0xb8]\n"
-    "ldr q15, [x20, x28]\n"
-    "fmla v29.4s, v7.4s, v12.4s\n"
-    "fmla v30.4s, v5.4s, v16.4s\n"
-    "ldr x22, [x13, #0xc0]\n"
+    "ldr q11, [x21, x14]\n"
+    "ldr x24, [x16, #0x90]\n"
+    "fmla v31.4s, v2.4s, v16.4s\n"
+    "ldr q16, [x20, x14]\n"
+    "fmla v29.4s, v0.4s, v15.4s\n"
+    "ldr q14, [x25, x14]\n"
+    "fmla v30.4s, v5.4s, v12.4s\n"
+    "ldr q12, [x26, x14]\n"
+    "ldr x23, [x16, #0x98]\n"
+    "fmla v31.4s, v5.4s, v13.4s\n"
+    "ldr x22, [x16, #0xa0]\n"
+    "fmla v29.4s, v4.4s, v11.4s\n"
+    "ldr q11, [x23, x14]\n"
+    "fmla v30.4s, v3.4s, v13.4s\n"
+    "ldr q13, [x19, x14]\n"
+    "ldr x21, [x16, #0xa8]\n"
     "fmla v31.4s, v6.4s, v15.4s\n"
-    "fmla v29.4s, v8.4s, v11.4s\n"
-    "ldr q11, [x22, x28]\n"
-    "fmla v30.4s, v8.4s, v15.4s\n"
-    "fmla v31.4s, v8.4s, v11.4s\n"
-    "fmax v28.4s, v28.4s, v19.4s\n"
-    "add x23, x23, #0x10\n"
-    "fmax v29.4s, v29.4s, v19.4s\n"
-    "fmax v30.4s, v30.4s, v19.4s\n"
-    "add x28, x28, #0x10\n"
+    "ldr q15, [x24, x14]\n"
+    "fmla v29.4s, v1.4s, v16.4s\n"
+    "ldr x20, [x16, #0xb0]\n"
+    "fmla v30.4s, v7.4s, v12.4s\n"
+    "ldr x19, [x16, #0xb8]\n"
+    "fmla v28.4s, v4.4s, v13.4s\n"
+    "ldr q13, [x22, x14]\n"
+    "ldr x26, [x16, #0xc0]\n"
+    "fmla v31.4s, v7.4s, v16.4s\n"
+    "fmla v29.4s, v6.4s, v15.4s\n"
+    "ldr q16, [x21, x14]\n"
+    "fmla v30.4s, v8.4s, v11.4s\n"
+    "ldr q15, [x19, x14]\n"
+    "fmla v28.4s, v1.4s, v12.4s\n"
+    "fmla v29.4s, v7.4s, v13.4s\n"
     "fmax v31.4s, v31.4s, v19.4s\n"
-    "fmin v28.4s, v28.4s, v18.4s\n"
-    "str q28, [x12, x23]\n"
-    "fmin v29.4s, v29.4s, v18.4s\n"
-    "fmin v30.4s, v30.4s, v18.4s\n"
-    "str q29, [x11, x23]\n"
+    "fmax v30.4s, v30.4s, v19.4s\n"
+    "fmla v28.4s, v5.4s, v14.4s\n"
+    "ldr q14, [x20, x14]\n"
     "fmin v31.4s, v31.4s, v18.4s\n"
-    "str q30, [x10, x23]\n"
-    "str q31, [x9, x23]\n"
+    "str q31, [x13, x28]\n"
+    "fmla v28.4s, v2.4s, v11.4s\n"
+    "fmla v29.4s, v5.4s, v16.4s\n"
+    "ldr q11, [x26, x14]\n"
+    "add x14, x14, #0x10\n"
+    "fmin v30.4s, v30.4s, v18.4s\n"
+    "str q30, [x12, x28]\n"
+    "fmla v28.4s, v3.4s, v16.4s\n"
+    "fmla v29.4s, v8.4s, v15.4s\n"
+    "fmla v28.4s, v7.4s, v14.4s\n"
+    "fmax v29.4s, v29.4s, v19.4s\n"
+    "fmin v29.4s, v29.4s, v18.4s\n"
+    "str q29, [x10, x28]\n"
+    "fmla v28.4s, v6.4s, v15.4s\n"
+    "fmla v28.4s, v8.4s, v11.4s\n"
+    "fmax v28.4s, v28.4s, v19.4s\n"
+    "fmin v28.4s, v28.4s, v18.4s\n"
+    "str q28, [x9, x28]\n"
     "3:"  // Oddments
     "tst %x[n_channels], #0x3\n"
     "beq 42f\n"
-    "ldr q17, [x24, #0x0]\n"
-    "ldr q0, [x24, #0x10]\n"
-    "mov x23, x28\n"
-    "add x12, x12, x23\n"
-    "ldr q1, [x24, #0x20]\n"
-    "ldr q2, [x24, #0x30]\n"
-    "add x11, x11, x23\n"
-    "add x10, x10, x23\n"
-    "ldr q3, [x24, #0x40]\n"
-    "ldr q4, [x24, #0x50]\n"
-    "add x9, x9, x23\n"
-    "ldr q5, [x24, #0x60]\n"
-    "ldr q6, [x24, #0x70]\n"
-    "ldr q7, [x24, #0x80]\n"
-    "ldr q8, [x24, #0x90]\n"
-    "ldr x27, [x13, #0x0]\n"
-    "ldr x26, [x13, #0x8]\n"
-    "add x27, x27, x28\n"
-    "add x26, x26, x28\n"
-    "ldr x25, [x13, #0x10]\n"
-    "ldr x24, [x13, #0x18]\n"
-    "add x25, x25, x28\n"
-    "add x24, x24, x28\n"
-    "ldr x23, [x13, #0x20]\n"
-    "ldr x22, [x13, #0x28]\n"
-    "add x23, x23, x28\n"
-    "add x22, x22, x28\n"
-    "ldr x21, [x13, #0x30]\n"
-    "ldr x20, [x13, #0x38]\n"
-    "add x21, x21, x28\n"
-    "add x20, x20, x28\n"
+    "ldr q17, [x15, #0x0]\n"
+    "ldr q0, [x15, #0x10]\n"
+    "mov x28, x14\n"
+    "ldr q1, [x15, #0x20]\n"
+    "add x13, x13, x28\n"
+    "ldr q2, [x15, #0x30]\n"
+    "add x12, x12, x28\n"
+    "ldr q3, [x15, #0x40]\n"
+    "add x10, x10, x28\n"
+    "ldr q4, [x15, #0x50]\n"
+    "add x9, x9, x28\n"
+    "ldr q5, [x15, #0x60]\n"
+    "ldr q6, [x15, #0x70]\n"
+    "ldr q7, [x15, #0x80]\n"
+    "ldr q8, [x15, #0x90]\n"
+    "ldr x26, [x16, #0x0]\n"
+    "ldr x25, [x16, #0x8]\n"
+    "ldr x24, [x16, #0x10]\n"
+    "add x26, x26, x14\n"
+    "ldr x23, [x16, #0x18]\n"
+    "add x25, x25, x14\n"
+    "ldr x22, [x16, #0x20]\n"
+    "add x24, x24, x14\n"
+    "ldr x21, [x16, #0x28]\n"
+    "add x23, x23, x14\n"
+    "ldr x20, [x16, #0x30]\n"
+    "add x22, x22, x14\n"
+    "ldr x19, [x16, #0x38]\n"
+    "add x21, x21, x14\n"
+    "add x20, x20, x14\n"
+    "add x19, x19, x14\n"
     "tbz %x[n_channels], #1, 4f\n"
-    "ld1 { v9.d }[0], [x27], #0x8\n"
-    "ld1 { v10.d }[0], [x26], #0x8\n"
-    "ld1 { v11.d }[0], [x25], #0x8\n"
-    "ld1 { v12.d }[0], [x24], #0x8\n"
-    "ld1 { v13.d }[0], [x23], #0x8\n"
-    "ld1 { v14.d }[0], [x22], #0x8\n"
-    "ld1 { v15.d }[0], [x21], #0x8\n"
-    "ld1 { v16.d }[0], [x20], #0x8\n"
+    "ld1 { v9.d }[0], [x26], #0x8\n"
+    "ld1 { v10.d }[0], [x25], #0x8\n"
+    "ld1 { v11.d }[0], [x24], #0x8\n"
+    "ld1 { v12.d }[0], [x23], #0x8\n"
+    "ld1 { v13.d }[0], [x22], #0x8\n"
+    "ld1 { v14.d }[0], [x21], #0x8\n"
+    "ld1 { v15.d }[0], [x20], #0x8\n"
+    "ld1 { v16.d }[0], [x19], #0x8\n"
     "tbz %x[n_channels], #0, 5f\n"
-    "ld1 { v9.s }[2], [x27], #0x4\n"
-    "ld1 { v10.s }[2], [x26], #0x4\n"
-    "ld1 { v11.s }[2], [x25], #0x4\n"
-    "ld1 { v12.s }[2], [x24], #0x4\n"
-    "ld1 { v13.s }[2], [x23], #0x4\n"
-    "ld1 { v14.s }[2], [x22], #0x4\n"
-    "ld1 { v15.s }[2], [x21], #0x4\n"
-    "ld1 { v16.s }[2], [x20], #0x4\n"
+    "ld1 { v9.s }[2], [x26], #0x4\n"
+    "ld1 { v10.s }[2], [x25], #0x4\n"
+    "ld1 { v11.s }[2], [x24], #0x4\n"
+    "ld1 { v12.s }[2], [x23], #0x4\n"
+    "ld1 { v13.s }[2], [x22], #0x4\n"
+    "ld1 { v14.s }[2], [x21], #0x4\n"
+    "ld1 { v15.s }[2], [x20], #0x4\n"
+    "ld1 { v16.s }[2], [x19], #0x4\n"
     "b 5f\n"
     "4:"  // Oddments: Load inputs (2, 2), (0, 0), (0, 1), (0, 3), (0, 4), (1, 0), (1, 1), (0, 2): Bit 1: Unset
-    "ld1 { v9.s }[0], [x27], #0x4\n"
-    "ld1 { v10.s }[0], [x26], #0x4\n"
-    "ld1 { v11.s }[0], [x25], #0x4\n"
-    "ld1 { v12.s }[0], [x24], #0x4\n"
-    "ld1 { v13.s }[0], [x23], #0x4\n"
-    "ld1 { v14.s }[0], [x22], #0x4\n"
-    "ld1 { v15.s }[0], [x21], #0x4\n"
-    "ld1 { v16.s }[0], [x20], #0x4\n"
+    "ld1 { v9.s }[0], [x26], #0x4\n"
+    "ld1 { v10.s }[0], [x25], #0x4\n"
+    "ld1 { v11.s }[0], [x24], #0x4\n"
+    "ld1 { v12.s }[0], [x23], #0x4\n"
+    "ld1 { v13.s }[0], [x22], #0x4\n"
+    "ld1 { v14.s }[0], [x21], #0x4\n"
+    "ld1 { v15.s }[0], [x20], #0x4\n"
+    "ld1 { v16.s }[0], [x19], #0x4\n"
     "5:"  // Oddments: Load inputs (2, 2), (0, 0), (0, 1), (0, 3), (0, 4), (1, 0), (1, 1), (0, 2): Bit 1: End
-    "mov v28.16b, v17.16b\n fmla v28.4s, v8.4s, v9.4s\n"
-    "fmla v28.4s, v0.4s, v10.4s\n"
-    "ldr x20, [x13, #0x40]\n"
-    "add x20, x20, x28\n"
-    "mov v29.16b, v17.16b\n fmla v29.4s, v6.4s, v9.4s\n"
-    "fmla v28.4s, v1.4s, v11.4s\n"
-    "fmla v29.4s, v1.4s, v12.4s\n"
-    "fmla v28.4s, v3.4s, v14.4s\n"
-    "fmla v29.4s, v2.4s, v13.4s\n"
-    "fmla v28.4s, v4.4s, v15.4s\n"
-    "mov v30.16b, v17.16b\n fmla v30.4s, v2.4s, v9.4s\n"
-    "mov v31.16b, v17.16b\n fmla v31.4s, v0.4s, v9.4s\n"
-    "fmla v28.4s, v2.4s, v16.4s\n"
-    "fmla v29.4s, v0.4s, v16.4s\n"
+    "mov v31.16b, v17.16b\n fmla v31.4s, v8.4s, v9.4s\n"
+    "ldr x26, [x16, #0x40]\n"
+    "add x26, x26, x14\n"
+    "mov v30.16b, v17.16b\n fmla v30.4s, v6.4s, v9.4s\n"
+    "mov v29.16b, v17.16b\n fmla v29.4s, v2.4s, v9.4s\n"
+    "mov v28.16b, v17.16b\n fmla v28.4s, v0.4s, v9.4s\n"
+    "fmla v31.4s, v0.4s, v10.4s\n"
+    "fmla v30.4s, v1.4s, v12.4s\n"
+    "fmla v31.4s, v1.4s, v11.4s\n"
+    "fmla v30.4s, v2.4s, v13.4s\n"
+    "fmla v31.4s, v3.4s, v14.4s\n"
+    "fmla v30.4s, v0.4s, v16.4s\n"
+    "fmla v31.4s, v4.4s, v15.4s\n"
+    "fmla v31.4s, v2.4s, v16.4s\n"
     "tbz %x[n_channels], #1, 6f\n"
-    "ld1 { v11.d }[0], [x20], #0x8\n"
+    "ld1 { v11.d }[0], [x26], #0x8\n"
     "tbz %x[n_channels], #0, 7f\n"
-    "ld1 { v11.s }[2], [x20], #0x4\n"
+    "ld1 { v11.s }[2], [x26], #0x4\n"
     "b 7f\n"
     "6:"  // Oddments: Load input (1, 3): Bit 1: Unset
-    "ld1 { v11.s }[0], [x20], #0x4\n"
+    "ld1 { v11.s }[0], [x26], #0x4\n"
     "7:"  // Oddments: Load input (1, 3): Bit 1: End
-    "ldr x20, [x13, #0x48]\n"
-    "fmla v29.4s, v4.4s, v11.4s\n"
-    "add x20, x20, x28\n"
+    "fmla v30.4s, v4.4s, v11.4s\n"
+    "ldr x25, [x16, #0x48]\n"
+    "add x25, x25, x14\n"
     "tbz %x[n_channels], #1, 8f\n"
-    "ld1 { v12.d }[0], [x20], #0x8\n"
+    "ld1 { v12.d }[0], [x25], #0x8\n"
     "tbz %x[n_channels], #0, 9f\n"
-    "ld1 { v12.s }[2], [x20], #0x4\n"
+    "ld1 { v12.s }[2], [x25], #0x4\n"
     "b 9f\n"
     "8:"  // Oddments: Load input (1, 4): Bit 1: Unset
-    "ld1 { v12.s }[0], [x20], #0x4\n"
+    "ld1 { v12.s }[0], [x25], #0x4\n"
     "9:"  // Oddments: Load input (1, 4): Bit 1: End
-    "ldr x20, [x13, #0x50]\n"
-    "fmla v29.4s, v5.4s, v12.4s\n"
-    "add x20, x20, x28\n"
+    "fmla v30.4s, v5.4s, v12.4s\n"
+    "ldr x24, [x16, #0x50]\n"
+    "add x24, x24, x14\n"
     "tbz %x[n_channels], #1, 10f\n"
-    "ld1 { v13.d }[0], [x20], #0x8\n"
+    "ld1 { v13.d }[0], [x24], #0x8\n"
     "tbz %x[n_channels], #0, 11f\n"
-    "ld1 { v13.s }[2], [x20], #0x4\n"
+    "ld1 { v13.s }[2], [x24], #0x4\n"
     "b 11f\n"
     "10:"  // Oddments: Load input (1, 2): Bit 1: Unset
-    "ld1 { v13.s }[0], [x20], #0x4\n"
+    "ld1 { v13.s }[0], [x24], #0x4\n"
     "11:"  // Oddments: Load input (1, 2): Bit 1: End
-    "ldr x20, [x13, #0x58]\n"
-    "fmla v28.4s, v5.4s, v13.4s\n"
-    "fmla v29.4s, v3.4s, v13.4s\n"
-    "add x20, x20, x28\n"
+    "fmla v31.4s, v5.4s, v13.4s\n"
+    "ldr x23, [x16, #0x58]\n"
+    "fmla v30.4s, v3.4s, v13.4s\n"
+    "add x23, x23, x14\n"
     "tbz %x[n_channels], #1, 12f\n"
-    "ld1 { v14.d }[0], [x20], #0x8\n"
+    "ld1 { v14.d }[0], [x23], #0x8\n"
     "tbz %x[n_channels], #0, 13f\n"
-    "ld1 { v14.s }[2], [x20], #0x4\n"
+    "ld1 { v14.s }[2], [x23], #0x4\n"
     "b 13f\n"
     "12:"  // Oddments: Load input (3, 0): Bit 1: Unset
-    "ld1 { v14.s }[0], [x20], #0x4\n"
+    "ld1 { v14.s }[0], [x23], #0x4\n"
     "13:"  // Oddments: Load input (3, 0): Bit 1: End
-    "ldr x20, [x13, #0x60]\n"
-    "fmla v30.4s, v3.4s, v14.4s\n"
-    "add x20, x20, x28\n"
+    "fmla v29.4s, v3.4s, v14.4s\n"
+    "ldr x22, [x16, #0x60]\n"
+    "add x22, x22, x14\n"
     "tbz %x[n_channels], #1, 14f\n"
-    "ld1 { v15.d }[0], [x20], #0x8\n"
+    "ld1 { v15.d }[0], [x22], #0x8\n"
     "tbz %x[n_channels], #0, 15f\n"
-    "ld1 { v15.s }[2], [x20], #0x4\n"
+    "ld1 { v15.s }[2], [x22], #0x4\n"
     "b 15f\n"
     "14:"  // Oddments: Load input (2, 0): Bit 1: Unset
-    "ld1 { v15.s }[0], [x20], #0x4\n"
+    "ld1 { v15.s }[0], [x22], #0x4\n"
     "15:"  // Oddments: Load input (2, 0): Bit 1: End
-    "ldr x20, [x13, #0x68]\n"
-    "fmla v28.4s, v6.4s, v15.4s\n"
-    "fmla v30.4s, v0.4s, v15.4s\n"
-    "add x20, x20, x28\n"
+    "fmla v31.4s, v6.4s, v15.4s\n"
+    "ldr x21, [x16, #0x68]\n"
+    "fmla v29.4s, v0.4s, v15.4s\n"
+    "add x21, x21, x14\n"
     "tbz %x[n_channels], #1, 16f\n"
-    "ld1 { v11.d }[0], [x20], #0x8\n"
+    "ld1 { v11.d }[0], [x21], #0x8\n"
     "tbz %x[n_channels], #0, 17f\n"
-    "ld1 { v11.s }[2], [x20], #0x4\n"
+    "ld1 { v11.s }[2], [x21], #0x4\n"
     "b 17f\n"
     "16:"  // Oddments: Load input (3, 1): Bit 1: Unset
-    "ld1 { v11.s }[0], [x20], #0x4\n"
+    "ld1 { v11.s }[0], [x21], #0x4\n"
     "17:"  // Oddments: Load input (3, 1): Bit 1: End
-    "ldr x20, [x13, #0x70]\n"
-    "fmla v30.4s, v4.4s, v11.4s\n"
-    "add x20, x20, x28\n"
+    "fmla v29.4s, v4.4s, v11.4s\n"
+    "ldr x20, [x16, #0x70]\n"
+    "add x20, x20, x14\n"
     "tbz %x[n_channels], #1, 18f\n"
     "ld1 { v16.d }[0], [x20], #0x8\n"
     "tbz %x[n_channels], #0, 19f\n"
@@ -475,90 +475,90 @@ void a64_fp32_nhwc_3x3_s2_output2x2_mla_depthfirst_indirect_impl(
     "18:"  // Oddments: Load input (2, 1): Bit 1: Unset
     "ld1 { v16.s }[0], [x20], #0x4\n"
     "19:"  // Oddments: Load input (2, 1): Bit 1: End
-    "ldr x20, [x13, #0x78]\n"
-    "fmla v28.4s, v7.4s, v16.4s\n"
-    "fmla v30.4s, v1.4s, v16.4s\n"
-    "add x20, x20, x28\n"
+    "fmla v31.4s, v7.4s, v16.4s\n"
+    "ldr x19, [x16, #0x78]\n"
+    "fmla v29.4s, v1.4s, v16.4s\n"
+    "add x19, x19, x14\n"
     "tbz %x[n_channels], #1, 20f\n"
-    "ld1 { v13.d }[0], [x20], #0x8\n"
+    "ld1 { v13.d }[0], [x19], #0x8\n"
     "tbz %x[n_channels], #0, 21f\n"
-    "ld1 { v13.s }[2], [x20], #0x4\n"
+    "ld1 { v13.s }[2], [x19], #0x4\n"
     "b 21f\n"
     "20:"  // Oddments: Load input (3, 3): Bit 1: Unset
-    "ld1 { v13.s }[0], [x20], #0x4\n"
+    "ld1 { v13.s }[0], [x19], #0x4\n"
     "21:"  // Oddments: Load input (3, 3): Bit 1: End
-    "ldr x20, [x13, #0x80]\n"
-    "fmla v31.4s, v4.4s, v13.4s\n"
-    "add x20, x20, x28\n"
+    "fmla v28.4s, v4.4s, v13.4s\n"
+    "ldr x26, [x16, #0x80]\n"
+    "add x26, x26, x14\n"
     "tbz %x[n_channels], #1, 22f\n"
-    "ld1 { v12.d }[0], [x20], #0x8\n"
+    "ld1 { v12.d }[0], [x26], #0x8\n"
     "tbz %x[n_channels], #0, 23f\n"
-    "ld1 { v12.s }[2], [x20], #0x4\n"
+    "ld1 { v12.s }[2], [x26], #0x4\n"
     "b 23f\n"
     "22:"  // Oddments: Load input (2, 3): Bit 1: Unset
-    "ld1 { v12.s }[0], [x20], #0x4\n"
+    "ld1 { v12.s }[0], [x26], #0x4\n"
     "23:"  // Oddments: Load input (2, 3): Bit 1: End
-    "ldr x20, [x13, #0x88]\n"
-    "fmla v29.4s, v7.4s, v12.4s\n"
-    "fmla v31.4s, v1.4s, v12.4s\n"
-    "add x20, x20, x28\n"
+    "fmla v30.4s, v7.4s, v12.4s\n"
+    "ldr x25, [x16, #0x88]\n"
+    "fmla v28.4s, v1.4s, v12.4s\n"
+    "add x25, x25, x14\n"
     "tbz %x[n_channels], #1, 24f\n"
-    "ld1 { v14.d }[0], [x20], #0x8\n"
+    "ld1 { v14.d }[0], [x25], #0x8\n"
     "tbz %x[n_channels], #0, 25f\n"
-    "ld1 { v14.s }[2], [x20], #0x4\n"
+    "ld1 { v14.s }[2], [x25], #0x4\n"
     "b 25f\n"
     "24:"  // Oddments: Load input (3, 4): Bit 1: Unset
-    "ld1 { v14.s }[0], [x20], #0x4\n"
+    "ld1 { v14.s }[0], [x25], #0x4\n"
     "25:"  // Oddments: Load input (3, 4): Bit 1: End
-    "ldr x20, [x13, #0x90]\n"
-    "fmla v31.4s, v5.4s, v14.4s\n"
-    "add x20, x20, x28\n"
+    "fmla v28.4s, v5.4s, v14.4s\n"
+    "ldr x24, [x16, #0x90]\n"
+    "add x24, x24, x14\n"
     "tbz %x[n_channels], #1, 26f\n"
-    "ld1 { v15.d }[0], [x20], #0x8\n"
+    "ld1 { v15.d }[0], [x24], #0x8\n"
     "tbz %x[n_channels], #0, 27f\n"
-    "ld1 { v15.s }[2], [x20], #0x4\n"
+    "ld1 { v15.s }[2], [x24], #0x4\n"
     "b 27f\n"
     "26:"  // Oddments: Load input (4, 0): Bit 1: Unset
-    "ld1 { v15.s }[0], [x20], #0x4\n"
+    "ld1 { v15.s }[0], [x24], #0x4\n"
     "27:"  // Oddments: Load input (4, 0): Bit 1: End
-    "ldr x20, [x13, #0x98]\n"
-    "fmla v30.4s, v6.4s, v15.4s\n"
-    "add x20, x20, x28\n"
+    "fmla v29.4s, v6.4s, v15.4s\n"
+    "ldr x23, [x16, #0x98]\n"
+    "add x23, x23, x14\n"
     "tbz %x[n_channels], #1, 28f\n"
-    "ld1 { v11.d }[0], [x20], #0x8\n"
+    "ld1 { v11.d }[0], [x23], #0x8\n"
     "tbz %x[n_channels], #0, 29f\n"
-    "ld1 { v11.s }[2], [x20], #0x4\n"
+    "ld1 { v11.s }[2], [x23], #0x4\n"
     "b 29f\n"
     "28:"  // Oddments: Load input (2, 4): Bit 1: Unset
-    "ld1 { v11.s }[0], [x20], #0x4\n"
+    "ld1 { v11.s }[0], [x23], #0x4\n"
     "29:"  // Oddments: Load input (2, 4): Bit 1: End
-    "ldr x20, [x13, #0xa0]\n"
-    "fmla v29.4s, v8.4s, v11.4s\n"
-    "fmla v31.4s, v2.4s, v11.4s\n"
-    "add x20, x20, x28\n"
+    "fmla v30.4s, v8.4s, v11.4s\n"
+    "ldr x22, [x16, #0xa0]\n"
+    "fmla v28.4s, v2.4s, v11.4s\n"
+    "add x22, x22, x14\n"
     "tbz %x[n_channels], #1, 30f\n"
-    "ld1 { v13.d }[0], [x20], #0x8\n"
+    "ld1 { v13.d }[0], [x22], #0x8\n"
     "tbz %x[n_channels], #0, 31f\n"
-    "ld1 { v13.s }[2], [x20], #0x4\n"
+    "ld1 { v13.s }[2], [x22], #0x4\n"
     "b 31f\n"
     "30:"  // Oddments: Load input (4, 1): Bit 1: Unset
-    "ld1 { v13.s }[0], [x20], #0x4\n"
+    "ld1 { v13.s }[0], [x22], #0x4\n"
     "31:"  // Oddments: Load input (4, 1): Bit 1: End
-    "ldr x20, [x13, #0xa8]\n"
-    "fmla v30.4s, v7.4s, v13.4s\n"
-    "add x20, x20, x28\n"
+    "fmla v29.4s, v7.4s, v13.4s\n"
+    "ldr x21, [x16, #0xa8]\n"
+    "add x21, x21, x14\n"
     "tbz %x[n_channels], #1, 32f\n"
-    "ld1 { v16.d }[0], [x20], #0x8\n"
+    "ld1 { v16.d }[0], [x21], #0x8\n"
     "tbz %x[n_channels], #0, 33f\n"
-    "ld1 { v16.s }[2], [x20], #0x4\n"
+    "ld1 { v16.s }[2], [x21], #0x4\n"
     "b 33f\n"
     "32:"  // Oddments: Load input (3, 2): Bit 1: Unset
-    "ld1 { v16.s }[0], [x20], #0x4\n"
+    "ld1 { v16.s }[0], [x21], #0x4\n"
     "33:"  // Oddments: Load input (3, 2): Bit 1: End
-    "ldr x20, [x13, #0xb0]\n"
-    "fmla v30.4s, v5.4s, v16.4s\n"
-    "fmla v31.4s, v3.4s, v16.4s\n"
-    "add x20, x20, x28\n"
+    "fmla v29.4s, v5.4s, v16.4s\n"
+    "ldr x20, [x16, #0xb0]\n"
+    "fmla v28.4s, v3.4s, v16.4s\n"
+    "add x20, x20, x14\n"
     "tbz %x[n_channels], #1, 34f\n"
     "ld1 { v14.d }[0], [x20], #0x8\n"
     "tbz %x[n_channels], #0, 35f\n"
@@ -567,59 +567,61 @@ void a64_fp32_nhwc_3x3_s2_output2x2_mla_depthfirst_indirect_impl(
     "34:"  // Oddments: Load input (4, 3): Bit 1: Unset
     "ld1 { v14.s }[0], [x20], #0x4\n"
     "35:"  // Oddments: Load input (4, 3): Bit 1: End
-    "ldr x20, [x13, #0xb8]\n"
-    "fmla v31.4s, v7.4s, v14.4s\n"
-    "add x20, x20, x28\n"
+    "fmla v28.4s, v7.4s, v14.4s\n"
+    "ldr x19, [x16, #0xb8]\n"
+    "add x19, x19, x14\n"
     "tbz %x[n_channels], #1, 36f\n"
-    "ld1 { v15.d }[0], [x20], #0x8\n"
+    "ld1 { v15.d }[0], [x19], #0x8\n"
     "tbz %x[n_channels], #0, 37f\n"
-    "ld1 { v15.s }[2], [x20], #0x4\n"
+    "ld1 { v15.s }[2], [x19], #0x4\n"
     "b 37f\n"
     "36:"  // Oddments: Load input (4, 2): Bit 1: Unset
-    "ld1 { v15.s }[0], [x20], #0x4\n"
+    "ld1 { v15.s }[0], [x19], #0x4\n"
     "37:"  // Oddments: Load input (4, 2): Bit 1: End
-    "ldr x20, [x13, #0xc0]\n"
-    "fmla v30.4s, v8.4s, v15.4s\n"
-    "fmla v31.4s, v6.4s, v15.4s\n"
-    "add x20, x20, x28\n"
+    "fmla v29.4s, v8.4s, v15.4s\n"
+    "ldr x26, [x16, #0xc0]\n"
+    "fmla v28.4s, v6.4s, v15.4s\n"
+    "add x26, x26, x14\n"
     "tbz %x[n_channels], #1, 38f\n"
-    "ld1 { v11.d }[0], [x20], #0x8\n"
+    "ld1 { v11.d }[0], [x26], #0x8\n"
     "tbz %x[n_channels], #0, 39f\n"
-    "ld1 { v11.s }[2], [x20], #0x4\n"
+    "ld1 { v11.s }[2], [x26], #0x4\n"
     "b 39f\n"
     "38:"  // Oddments: Load input (4, 4): Bit 1: Unset
-    "ld1 { v11.s }[0], [x20], #0x4\n"
+    "ld1 { v11.s }[0], [x26], #0x4\n"
     "39:"  // Oddments: Load input (4, 4): Bit 1: End
-    "fmla v31.4s, v8.4s, v11.4s\n"
-    "fmax v28.4s, v28.4s, v19.4s\n"
-    "fmax v29.4s, v29.4s, v19.4s\n"
-    "fmax v30.4s, v30.4s, v19.4s\n"
+    "fmla v28.4s, v8.4s, v11.4s\n"
     "fmax v31.4s, v31.4s, v19.4s\n"
-    "fmin v28.4s, v28.4s, v18.4s\n"
-    "fmin v29.4s, v29.4s, v18.4s\n"
-    "fmin v30.4s, v30.4s, v18.4s\n"
+    "fmax v30.4s, v30.4s, v19.4s\n"
+    "fmax v29.4s, v29.4s, v19.4s\n"
     "fmin v31.4s, v31.4s, v18.4s\n"
+    "fmin v30.4s, v30.4s, v18.4s\n"
+    "fmin v29.4s, v29.4s, v18.4s\n"
+    "fmax v28.4s, v28.4s, v19.4s\n"
+    "fmin v28.4s, v28.4s, v18.4s\n"
     "tbz %x[n_channels], #1, 40f\n"
-    "st1 { v28.d }[0], [x12], #0x8\n"
-    "st1 { v29.d }[0], [x11], #0x8\n"
-    "st1 { v30.d }[0], [x10], #0x8\n"
-    "st1 { v31.d }[0], [x9], #0x8\n"
+    "st1 { v31.d }[0], [x13], #0x8\n"
+    "st1 { v30.d }[0], [x12], #0x8\n"
+    "st1 { v29.d }[0], [x10], #0x8\n"
+    "st1 { v28.d }[0], [x9], #0x8\n"
     "tbz %x[n_channels], #0, 41f\n"
-    "st1 { v28.s }[2], [x12], #0x4\n"
-    "st1 { v29.s }[2], [x11], #0x4\n"
-    "st1 { v30.s }[2], [x10], #0x4\n"
-    "st1 { v31.s }[2], [x9], #0x4\n"
+    "st1 { v31.s }[2], [x13], #0x4\n"
+    "st1 { v30.s }[2], [x12], #0x4\n"
+    "st1 { v29.s }[2], [x10], #0x4\n"
+    "st1 { v28.s }[2], [x9], #0x4\n"
     "b 41f\n"
     "40:"  // Oddments: Store: Bit 1: Unset
-    "st1 { v28.s }[0], [x12], #0x4\n"
-    "st1 { v29.s }[0], [x11], #0x4\n"
-    "st1 { v30.s }[0], [x10], #0x4\n"
-    "st1 { v31.s }[0], [x9], #0x4\n"
+    "st1 { v31.s }[0], [x13], #0x4\n"
+    "st1 { v30.s }[0], [x12], #0x4\n"
+    "st1 { v29.s }[0], [x10], #0x4\n"
+    "st1 { v28.s }[0], [x9], #0x4\n"
     "41:"  // Oddments: Store: Bit 1: End
+
     "42:"  // End
+
     :
     : [n_channels] "r" ((unsigned long) n_channels), [offsetof_Args_inptrs] "I" (offsetof(Args, inptrs)), [offsetof_args_max] "I" (offsetof(Args, max)), [offsetof_args_min] "I" (offsetof(Args, min)), [offsetof_args_outptrs] "I" (offsetof(Args, outptrs)), [offsetof_args_params] "I" (offsetof(Args, params)), [params_struct] "r" (&params_struct)
-    : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v28", "v29", "v30", "v31", "x9", "x10", "x11", "x12", "x13", "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", "v28", "v29", "v30", "v31", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x16", "x19", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28"
   );
 }
 
-- 
cgit v1.2.1