From 74921eee924625426429044decefe3673561b174 Mon Sep 17 00:00:00 2001
From: Michael Tyler <michael.tyler@arm.com>
Date: Wed, 12 Apr 2023 17:43:17 +0100
Subject: Update CPU kernel implementations and guard directives

Resolves COMPMID-6023

Change-Id: I868975d14c4f98af6716726feda22405a6a4c891
Signed-off-by: Michael Tyler <michael.tyler@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/9686
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Viet-Hoa Do <viet-hoa.do@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Benchmark: Arm Jenkins <bsgcomp@arm.com>
---
 .../generic_indirect.cpp                           | 462 ++++++++++-----------
 1 file changed, 231 insertions(+), 231 deletions(-)

(limited to 'src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_3x3_s2_output2x2_mla_depthfirst/generic_indirect.cpp')

diff --git a/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_3x3_s2_output2x2_mla_depthfirst/generic_indirect.cpp b/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_3x3_s2_output2x2_mla_depthfirst/generic_indirect.cpp
index 144d11fb39..6ae0b30afd 100644
--- a/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_3x3_s2_output2x2_mla_depthfirst/generic_indirect.cpp
+++ b/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_3x3_s2_output2x2_mla_depthfirst/generic_indirect.cpp
@@ -88,258 +88,258 @@ void a64_fp16_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], #0x3\n"
-    "ldr x24, [%x[params_struct], %[offsetof_args_params]]\n"
+    "mov x25, #0x10\n"  // cntb _, ALL, #1
+    "lsr x24, %x[n_channels], #0x3\n"
+    "ldr x23, [%x[params_struct], %[offsetof_args_params]]\n"
     "add x20, %x[params_struct], %[offsetof_args_min]\n"
-    "ld1r { v19.8h }, [x20]\n"
+    "ld1r { v26.8h }, [x20]\n"
     "add x20, %x[params_struct], %[offsetof_args_max]\n"
-    "ld1r { v18.8h }, [x20]\n"
+    "ld1r { v27.8h }, [x20]\n"
     "add x13, %x[params_struct], %[offsetof_Args_inptrs]\n"
     "ldp x12, x11, [x21, #0x0]\n"
     "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"
+    "sub x22, XZR, x25\n"
+    "cbz x24, 3f\n"
+    "ldr q31, [x23, #0x0]\n"
+    "ldr q0, [x23, #0x10]\n"
+    "cmp x25, x24, LSL #4\n"
+    "ldr q1, [x23, #0x20]\n"
+    "ldr q2, [x23, #0x30]\n"
+    "ldr q3, [x23, #0x40]\n"
+    "ldr q4, [x23, #0x50]\n"
+    "ldr q5, [x23, #0x60]\n"
+    "ldr q6, [x23, #0x70]\n"
+    "ldr q7, [x23, #0x80]\n"
+    "ldr q8, [x23, #0x90]\n"
+    "add x23, x23, #0xa0\n"
+    "ldp x21, x20, [x13, #0x0]\n"
+    "ldr q9, [x21, 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, #0x20]\n"
+    "ldr q13, [x21, x28]\n"
+    "ldr q14, [x20, x28]\n"
     "ldp x21, x20, [x13, #0x30]\n"
     "ldr q15, [x21, x28]\n"
     "ldr q16, [x20, x28]\n"
     "bge 2f\n"
     "1:"  // Channel loop
-    "mov v28.16b, v17.16b\n fmla v28.8h, v8.8h, v9.8h\n"
-    "mov v29.16b, v17.16b\n fmla v29.8h, v6.8h, v9.8h\n"
-    "ldr x22, [x13, #0x40]\n"
+    "mov v24.16b, v31.16b\n fmla v24.8h, v8.8h, v9.8h\n"
+    "mov v23.16b, v31.16b\n fmla v23.8h, v6.8h, v9.8h\n"
+    "ldr x21, [x13, #0x40]\n"
     "ldr x20, [x13, #0x48]\n"
-    "fmla v28.8h, v0.8h, v10.8h\n"
-    "fmla v29.8h, v1.8h, v12.8h\n"
-    "ldr q12, [x20, x28]\n"
-    "ldr x21, [x13, #0x50]\n"
-    "fmla v28.8h, v1.8h, v11.8h\n"
-    "ldr q11, [x22, x28]\n"
-    "fmla v29.8h, v2.8h, v13.8h\n"
-    "ldr q13, [x21, x28]\n"
-    "fmla v28.8h, v3.8h, v14.8h\n"
-    "fmla v29.8h, v0.8h, v16.8h\n"
+    "fmla v24.8h, v0.8h, v10.8h\n"
+    "fmla v23.8h, v1.8h, v12.8h\n"
+    "ldr q20, [x20, x28]\n"
+    "ldr x20, [x13, #0x50]\n"
+    "fmla v24.8h, v1.8h, v11.8h\n"
+    "ldr q19, [x21, x28]\n"
+    "fmla v23.8h, v2.8h, v13.8h\n"
+    "ldr q18, [x20, x28]\n"
+    "fmla v24.8h, v3.8h, v14.8h\n"
+    "fmla v23.8h, v0.8h, v16.8h\n"
     "ldr x20, [x13, #0x58]\n"
-    "ldr q14, [x20, x28]\n"
-    "fmla v28.8h, v4.8h, v15.8h\n"
-    "fmla v29.8h, v4.8h, v11.8h\n"
-    "ldr x20, [x13, #0x78]\n"
-    "ldr x22, [x13, #0x60]\n"
-    "ldr q15, [x22, x28]\n"
-    "fmla v28.8h, v2.8h, v16.8h\n"
-    "fmla v29.8h, v5.8h, v12.8h\n"
-    "ldr x22, [x13, #0x80]\n"
-    "ldr q12, [x22, x28]\n"
-    "mov v30.16b, v17.16b\n fmla v30.8h, v2.8h, v9.8h\n"
-    "mov v31.16b, v17.16b\n fmla v31.8h, v0.8h, v9.8h\n"
-    "ldr q17, [x24, #0x0]\n"
-    "fmla v28.8h, v5.8h, v13.8h\n"
-    "fmla v29.8h, v3.8h, v13.8h\n"
-    "ldr q13, [x20, x28]\n"
-    "ldr x21, [x13, #0x68]\n"
-    "ldr q11, [x21, x28]\n"
-    "fmla v30.8h, v3.8h, v14.8h\n"
-    "fmla v31.8h, v4.8h, v13.8h\n"
-    "ldr x20, [x13, #0x88]\n"
-    "ldr q14, [x20, x28]\n"
-    "fmla v30.8h, v0.8h, v15.8h\n"
-    "ldr q0, [x24, #0x10]\n"
-    "fmla v31.8h, v1.8h, v12.8h\n"
-    "ldr x21, [x13, #0x70]\n"
+    "ldr q17, [x20, x28]\n"
+    "fmla v24.8h, v4.8h, v15.8h\n"
+    "fmla v23.8h, v4.8h, v19.8h\n"
+    "ldr x21, [x13, #0x78]\n"
+    "ldr x20, [x13, #0x60]\n"
+    "ldr q22, [x20, x28]\n"
+    "fmla v24.8h, v2.8h, v16.8h\n"
+    "fmla v23.8h, v5.8h, v20.8h\n"
+    "ldr x20, [x13, #0x80]\n"
+    "ldr q21, [x20, x28]\n"
+    "mov v20.16b, v31.16b\n fmla v20.8h, v2.8h, v9.8h\n"
+    "mov v19.16b, v31.16b\n fmla v19.8h, v0.8h, v9.8h\n"
+    "ldr q31, [x23, #0x0]\n"
+    "fmla v24.8h, v5.8h, v18.8h\n"
+    "fmla v23.8h, v3.8h, v18.8h\n"
     "ldr q16, [x21, x28]\n"
-    "fmla v30.8h, v4.8h, v11.8h\n"
-    "fmla v31.8h, v5.8h, v14.8h\n"
-    "ldr q4, [x24, #0x50]\n"
+    "ldr x20, [x13, #0x68]\n"
+    "ldr q18, [x20, x28]\n"
+    "fmla v20.8h, v3.8h, v17.8h\n"
+    "fmla v19.8h, v4.8h, v16.8h\n"
+    "ldr x20, [x13, #0x88]\n"
+    "ldr q16, [x20, x28]\n"
+    "fmla v20.8h, v0.8h, v22.8h\n"
+    "ldr q0, [x23, #0x10]\n"
+    "fmla v19.8h, v1.8h, v21.8h\n"
+    "ldr x20, [x13, #0x70]\n"
+    "ldr q17, [x20, x28]\n"
+    "fmla v20.8h, v4.8h, v18.8h\n"
+    "fmla v19.8h, v5.8h, v16.8h\n"
+    "ldr q4, [x23, #0x50]\n"
     "ldr x20, [x13, #0x98]\n"
-    "fmla v28.8h, v6.8h, v15.8h\n"
-    "fmla v30.8h, v1.8h, v16.8h\n"
-    "ldr q11, [x20, x28]\n"
-    "ldr q1, [x24, #0x20]\n"
-    "fmla v31.8h, v2.8h, v11.8h\n"
-    "fmla v28.8h, v7.8h, v16.8h\n"
-    "ldr q2, [x24, #0x30]\n"
-    "ldr x21, [x13, #0x90]\n"
-    "fmla v29.8h, v7.8h, v12.8h\n"
-    "fmla v29.8h, v8.8h, v11.8h\n"
-    "ldr q15, [x21, x28]\n"
-    "ldr x21, [x13, #0xa8]\n"
-    "fmla v30.8h, v6.8h, v15.8h\n"
-    "fmax v28.8h, v28.8h, v19.8h\n"
-    "ldr q16, [x21, x28]\n"
-    "ldr x22, [x13, #0xa0]\n"
-    "fmla v31.8h, v3.8h, v16.8h\n"
-    "fmax v29.8h, v29.8h, v19.8h\n"
-    "ldr q13, [x22, x28]\n"
-    "ldr q3, [x24, #0x40]\n"
-    "fmla v30.8h, v7.8h, v13.8h\n"
-    "fmla v30.8h, v5.8h, v16.8h\n"
-    "ldr q5, [x24, #0x60]\n"
-    "ldr x21, [x13, #0xb0]\n"
-    "add x23, x23, #0x10\n"
-    "fmin v28.8h, v28.8h, v18.8h\n"
-    "ldr q14, [x21, x28]\n"
+    "fmla v24.8h, v6.8h, v22.8h\n"
+    "fmla v20.8h, v1.8h, v17.8h\n"
+    "ldr q16, [x20, x28]\n"
+    "ldr q1, [x23, #0x20]\n"
+    "fmla v19.8h, v2.8h, v16.8h\n"
+    "fmla v24.8h, v7.8h, v17.8h\n"
+    "ldr q2, [x23, #0x30]\n"
+    "ldr x20, [x13, #0x90]\n"
+    "fmla v23.8h, v7.8h, v21.8h\n"
+    "fmla v23.8h, v8.8h, v16.8h\n"
+    "ldr q16, [x20, x28]\n"
+    "ldr x20, [x13, #0xa8]\n"
+    "fmla v20.8h, v6.8h, v16.8h\n"
+    "fmax v24.8h, v24.8h, v26.8h\n"
+    "ldr q17, [x20, x28]\n"
+    "ldr x20, [x13, #0xa0]\n"
+    "fmla v19.8h, v3.8h, v17.8h\n"
+    "fmax v23.8h, v23.8h, v26.8h\n"
+    "ldr q16, [x20, x28]\n"
+    "ldr q3, [x23, #0x40]\n"
+    "fmla v20.8h, v7.8h, v16.8h\n"
+    "fmla v20.8h, v5.8h, v17.8h\n"
+    "ldr q5, [x23, #0x60]\n"
+    "ldr x20, [x13, #0xb0]\n"
+    "add x22, x22, #0x10\n"
+    "fmin v24.8h, v24.8h, v27.8h\n"
+    "ldr q16, [x20, x28]\n"
     "ldr x20, [x13, #0xb8]\n"
-    "fmla v31.8h, v7.8h, v14.8h\n"
-    "fmin v29.8h, v29.8h, v18.8h\n"
-    "ldr q15, [x20, x28]\n"
-    "ldr q7, [x24, #0x80]\n"
-    "fmla v31.8h, v6.8h, v15.8h\n"
-    "fmla v30.8h, v8.8h, v15.8h\n"
-    "ldr q6, [x24, #0x70]\n"
-    "ldr x22, [x13, #0xc0]\n"
-    "fmax v30.8h, v30.8h, v19.8h\n"
-    "fmin v30.8h, v30.8h, v18.8h\n"
-    "ldr q11, [x22, x28]\n"
-    "fmla v31.8h, v8.8h, v11.8h\n"
-    "ldr q8, [x24, #0x90]\n"
-    "fmax v31.8h, v31.8h, v19.8h\n"
-    "ldp x22, x20, [x13, #0x0]\n"
-    "ldr q9, [x22, x26]\n"
-    "fmin v31.8h, v31.8h, v18.8h\n"
+    "fmla v19.8h, v7.8h, v16.8h\n"
+    "fmin v23.8h, v23.8h, v27.8h\n"
+    "ldr q16, [x20, x28]\n"
+    "ldr q7, [x23, #0x80]\n"
+    "fmla v19.8h, v6.8h, v16.8h\n"
+    "fmla v20.8h, v8.8h, v16.8h\n"
+    "ldr q6, [x23, #0x70]\n"
+    "ldr x20, [x13, #0xc0]\n"
+    "fmax v20.8h, v20.8h, v26.8h\n"
+    "fmin v20.8h, v20.8h, v27.8h\n"
+    "ldr q16, [x20, x28]\n"
+    "fmla v19.8h, v8.8h, v16.8h\n"
+    "ldr q8, [x23, #0x90]\n"
+    "fmax v19.8h, v19.8h, v26.8h\n"
+    "ldp x21, x20, [x13, #0x0]\n"
+    "ldr q9, [x21, x25]\n"
+    "fmin v19.8h, v19.8h, v27.8h\n"
     "add x28, x28, #0x10\n"
-    "ldr q10, [x20, x26]\n"
+    "ldr q10, [x20, x25]\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"
+    "str q24, [x12, x22]\n"
+    "add x23, x23, #0xa0\n"
+    "ldr q11, [x21, x25]\n"
+    "ldr q12, [x20, x25]\n"
+    "str q23, [x11, x22]\n"
+    "ldp x21, x20, [x13, #0x20]\n"
+    "ldr q13, [x21, x25]\n"
+    "str q20, [x10, x22]\n"
+    "ldr q14, [x20, x25]\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 q19, [x9, x22]\n"
+    "ldr q15, [x21, x25]\n"
+    "ldr q16, [x20, x25]\n"
+    "add x25, x25, #0x10\n"
+    "cmp x25, x24, LSL #4\n"
     "blt 1b\n"
     "2:"  // Channel tail
-    "mov v28.16b, v17.16b\n fmla v28.8h, v8.8h, v9.8h\n"
-    "mov v29.16b, v17.16b\n fmla v29.8h, v6.8h, v9.8h\n"
-    "ldr x22, [x13, #0x40]\n"
+    "mov v25.16b, v31.16b\n fmla v25.8h, v8.8h, v9.8h\n"
+    "mov v24.16b, v31.16b\n fmla v24.8h, v6.8h, v9.8h\n"
+    "ldr x21, [x13, #0x40]\n"
     "ldr x20, [x13, #0x48]\n"
-    "fmla v28.8h, v0.8h, v10.8h\n"
-    "fmla v29.8h, v1.8h, v12.8h\n"
-    "ldr q12, [x20, x28]\n"
-    "ldr x21, [x13, #0x50]\n"
-    "fmla v28.8h, v1.8h, v11.8h\n"
-    "ldr q11, [x22, x28]\n"
-    "fmla v29.8h, v2.8h, v13.8h\n"
-    "ldr q13, [x21, x28]\n"
-    "fmla v28.8h, v3.8h, v14.8h\n"
-    "fmla v29.8h, v0.8h, v16.8h\n"
+    "fmla v25.8h, v0.8h, v10.8h\n"
+    "fmla v24.8h, v1.8h, v12.8h\n"
+    "ldr q20, [x20, x28]\n"
+    "ldr x20, [x13, #0x50]\n"
+    "fmla v25.8h, v1.8h, v11.8h\n"
+    "ldr q18, [x21, x28]\n"
+    "fmla v24.8h, v2.8h, v13.8h\n"
+    "ldr q19, [x20, x28]\n"
+    "fmla v25.8h, v3.8h, v14.8h\n"
+    "fmla v24.8h, v0.8h, v16.8h\n"
     "ldr x20, [x13, #0x58]\n"
-    "ldr q14, [x20, x28]\n"
-    "fmla v28.8h, v4.8h, v15.8h\n"
-    "fmla v29.8h, v4.8h, v11.8h\n"
-    "ldr x20, [x13, #0x78]\n"
-    "ldr x22, [x13, #0x60]\n"
-    "ldr q15, [x22, x28]\n"
-    "fmla v28.8h, v2.8h, v16.8h\n"
-    "fmla v29.8h, v5.8h, v12.8h\n"
-    "ldr x22, [x13, #0x80]\n"
-    "ldr q12, [x22, x28]\n"
-    "mov v30.16b, v17.16b\n fmla v30.8h, v2.8h, v9.8h\n"
-    "mov v31.16b, v17.16b\n fmla v31.8h, v0.8h, v9.8h\n"
-    "ldr x21, [x13, #0x68]\n"
-    "ldr q11, [x21, x28]\n"
-    "fmla v28.8h, v5.8h, v13.8h\n"
-    "fmla v29.8h, v3.8h, v13.8h\n"
-    "ldr q13, [x20, x28]\n"
-    "fmla v30.8h, v3.8h, v14.8h\n"
-    "fmla v31.8h, v4.8h, v13.8h\n"
-    "ldr x20, [x13, #0x88]\n"
-    "ldr q14, [x20, x28]\n"
-    "fmla v30.8h, v0.8h, v15.8h\n"
-    "fmla v31.8h, v1.8h, v12.8h\n"
-    "ldr x21, [x13, #0x70]\n"
+    "ldr q17, [x20, x28]\n"
+    "fmla v25.8h, v4.8h, v15.8h\n"
+    "fmla v24.8h, v4.8h, v18.8h\n"
+    "ldr x21, [x13, #0x78]\n"
+    "ldr x20, [x13, #0x60]\n"
+    "ldr q23, [x20, x28]\n"
+    "fmla v25.8h, v2.8h, v16.8h\n"
+    "fmla v24.8h, v5.8h, v20.8h\n"
+    "ldr x20, [x13, #0x80]\n"
+    "ldr q22, [x20, x28]\n"
+    "mov v21.16b, v31.16b\n fmla v21.8h, v2.8h, v9.8h\n"
+    "mov v20.16b, v31.16b\n fmla v20.8h, v0.8h, v9.8h\n"
+    "ldr x20, [x13, #0x68]\n"
+    "ldr q18, [x20, x28]\n"
+    "fmla v25.8h, v5.8h, v19.8h\n"
+    "fmla v24.8h, v3.8h, v19.8h\n"
     "ldr q16, [x21, x28]\n"
+    "fmla v21.8h, v3.8h, v17.8h\n"
+    "fmla v20.8h, v4.8h, v16.8h\n"
+    "ldr x20, [x13, #0x88]\n"
+    "ldr q16, [x20, x28]\n"
+    "fmla v21.8h, v0.8h, v23.8h\n"
+    "fmla v20.8h, v1.8h, v22.8h\n"
+    "ldr x20, [x13, #0x70]\n"
+    "ldr q17, [x20, x28]\n"
     "ldr x20, [x13, #0x98]\n"
-    "fmla v30.8h, v4.8h, v11.8h\n"
-    "ldr q11, [x20, x28]\n"
-    "fmla v31.8h, v5.8h, v14.8h\n"
-    "fmla v28.8h, v6.8h, v15.8h\n"
-    "ldr x21, [x13, #0x90]\n"
-    "ldr q15, [x21, x28]\n"
-    "fmla v30.8h, v1.8h, v16.8h\n"
-    "ldr x21, [x13, #0xa8]\n"
-    "fmla v31.8h, v2.8h, v11.8h\n"
-    "fmla v28.8h, v7.8h, v16.8h\n"
-    "ldr q16, [x21, x28]\n"
-    "ldr x22, [x13, #0xa0]\n"
-    "ldr q13, [x22, x28]\n"
-    "fmla v30.8h, v6.8h, v15.8h\n"
-    "fmla v31.8h, v3.8h, v16.8h\n"
-    "ldr x21, [x13, #0xb0]\n"
-    "ldr q14, [x21, x28]\n"
-    "fmla v30.8h, v7.8h, v13.8h\n"
-    "fmla v31.8h, v7.8h, v14.8h\n"
+    "fmla v21.8h, v4.8h, v18.8h\n"
+    "ldr q19, [x20, x28]\n"
+    "fmla v20.8h, v5.8h, v16.8h\n"
+    "fmla v25.8h, v6.8h, v23.8h\n"
+    "ldr x20, [x13, #0x90]\n"
+    "ldr q16, [x20, x28]\n"
+    "fmla v21.8h, v1.8h, v17.8h\n"
+    "ldr x20, [x13, #0xa8]\n"
+    "fmla v20.8h, v2.8h, v19.8h\n"
+    "fmla v25.8h, v7.8h, v17.8h\n"
+    "ldr q18, [x20, x28]\n"
+    "ldr x20, [x13, #0xa0]\n"
+    "ldr q17, [x20, x28]\n"
+    "fmla v21.8h, v6.8h, v16.8h\n"
+    "fmla v20.8h, v3.8h, v18.8h\n"
+    "ldr x20, [x13, #0xb0]\n"
+    "ldr q16, [x20, x28]\n"
+    "fmla v21.8h, v7.8h, v17.8h\n"
+    "fmla v20.8h, v7.8h, v16.8h\n"
     "ldr x20, [x13, #0xb8]\n"
-    "ldr q15, [x20, x28]\n"
-    "fmla v29.8h, v7.8h, v12.8h\n"
-    "fmla v30.8h, v5.8h, v16.8h\n"
-    "ldr x22, [x13, #0xc0]\n"
-    "fmla v31.8h, v6.8h, v15.8h\n"
-    "fmla v29.8h, v8.8h, v11.8h\n"
-    "ldr q11, [x22, x28]\n"
-    "fmla v30.8h, v8.8h, v15.8h\n"
-    "fmla v31.8h, v8.8h, v11.8h\n"
-    "fmax v28.8h, v28.8h, v19.8h\n"
-    "add x23, x23, #0x10\n"
-    "fmax v29.8h, v29.8h, v19.8h\n"
-    "fmax v30.8h, v30.8h, v19.8h\n"
+    "ldr q17, [x20, x28]\n"
+    "fmla v24.8h, v7.8h, v22.8h\n"
+    "fmla v21.8h, v5.8h, v18.8h\n"
+    "ldr x20, [x13, #0xc0]\n"
+    "fmla v20.8h, v6.8h, v17.8h\n"
+    "fmla v24.8h, v8.8h, v19.8h\n"
+    "ldr q16, [x20, x28]\n"
+    "fmla v21.8h, v8.8h, v17.8h\n"
+    "fmla v20.8h, v8.8h, v16.8h\n"
+    "fmax v25.8h, v25.8h, v26.8h\n"
+    "add x22, x22, #0x10\n"
+    "fmax v24.8h, v24.8h, v26.8h\n"
+    "fmax v21.8h, v21.8h, v26.8h\n"
     "add x28, x28, #0x10\n"
-    "fmax v31.8h, v31.8h, v19.8h\n"
-    "fmin v28.8h, v28.8h, v18.8h\n"
-    "str q28, [x12, x23]\n"
-    "fmin v29.8h, v29.8h, v18.8h\n"
-    "fmin v30.8h, v30.8h, v18.8h\n"
-    "str q29, [x11, x23]\n"
-    "fmin v31.8h, v31.8h, v18.8h\n"
-    "str q30, [x10, x23]\n"
-    "str q31, [x9, x23]\n"
+    "fmax v20.8h, v20.8h, v26.8h\n"
+    "fmin v25.8h, v25.8h, v27.8h\n"
+    "str q25, [x12, x22]\n"
+    "fmin v24.8h, v24.8h, v27.8h\n"
+    "fmin v21.8h, v21.8h, v27.8h\n"
+    "str q24, [x11, x22]\n"
+    "fmin v20.8h, v20.8h, v27.8h\n"
+    "str q21, [x10, x22]\n"
+    "str q20, [x9, x22]\n"
     "3:"  // Oddments
     "tst %x[n_channels], #0x7\n"
     "beq 80f\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 q31, [x23, #0x0]\n"
+    "ldr q0, [x23, #0x10]\n"
+    "mov x20, x28\n"
+    "add x12, x12, x20\n"
+    "ldr q1, [x23, #0x20]\n"
+    "ldr q2, [x23, #0x30]\n"
+    "add x11, x11, x20\n"
+    "add x10, x10, x20\n"
+    "ldr q3, [x23, #0x40]\n"
+    "ldr q4, [x23, #0x50]\n"
+    "add x9, x9, x20\n"
+    "ldr q5, [x23, #0x60]\n"
+    "ldr q6, [x23, #0x70]\n"
+    "ldr q7, [x23, #0x80]\n"
+    "ldr q8, [x23, #0x90]\n"
     "ldr x27, [x13, #0x0]\n"
     "ldr x26, [x13, #0x8]\n"
     "add x27, x27, x28\n"
@@ -425,18 +425,18 @@ void a64_fp16_nhwc_3x3_s2_output2x2_mla_depthfirst_indirect_impl(
     "ld1 { v15.h }[0], [x21], #0x2\n"
     "ld1 { v16.h }[0], [x20], #0x2\n"
     "7:"  // Oddments: Load inputs (2, 2), (0, 0), (0, 1), (0, 3), (0, 4), (1, 0), (1, 1), (0, 2): Bit 2: End
-    "mov v28.16b, v17.16b\n fmla v28.8h, v8.8h, v9.8h\n"
+    "mov v28.16b, v31.16b\n fmla v28.8h, v8.8h, v9.8h\n"
     "fmla v28.8h, v0.8h, v10.8h\n"
     "ldr x20, [x13, #0x40]\n"
     "add x20, x20, x28\n"
-    "mov v29.16b, v17.16b\n fmla v29.8h, v6.8h, v9.8h\n"
+    "mov v29.16b, v31.16b\n fmla v29.8h, v6.8h, v9.8h\n"
     "fmla v28.8h, v1.8h, v11.8h\n"
     "fmla v29.8h, v1.8h, v12.8h\n"
     "fmla v28.8h, v3.8h, v14.8h\n"
     "fmla v29.8h, v2.8h, v13.8h\n"
     "fmla v28.8h, v4.8h, v15.8h\n"
-    "mov v30.16b, v17.16b\n fmla v30.8h, v2.8h, v9.8h\n"
-    "mov v31.16b, v17.16b\n fmla v31.8h, v0.8h, v9.8h\n"
+    "mov v30.16b, v31.16b\n fmla v30.8h, v2.8h, v9.8h\n"
+    "fmla v31.8h, v0.8h, v9.8h\n"
     "fmla v28.8h, v2.8h, v16.8h\n"
     "fmla v29.8h, v0.8h, v16.8h\n"
     "tbz %x[n_channels], #2, 9f\n"
@@ -835,14 +835,14 @@ void a64_fp16_nhwc_3x3_s2_output2x2_mla_depthfirst_indirect_impl(
     "ld1 { v11.h }[0], [x20], #0x2\n"
     "75:"  // Oddments: Load input (4, 4): Bit 2: End
     "fmla v31.8h, v8.8h, v11.8h\n"
-    "fmax v28.8h, v28.8h, v19.8h\n"
-    "fmax v29.8h, v29.8h, v19.8h\n"
-    "fmax v30.8h, v30.8h, v19.8h\n"
-    "fmax v31.8h, v31.8h, v19.8h\n"
-    "fmin v28.8h, v28.8h, v18.8h\n"
-    "fmin v29.8h, v29.8h, v18.8h\n"
-    "fmin v30.8h, v30.8h, v18.8h\n"
-    "fmin v31.8h, v31.8h, v18.8h\n"
+    "fmax v28.8h, v28.8h, v26.8h\n"
+    "fmax v29.8h, v29.8h, v26.8h\n"
+    "fmax v30.8h, v30.8h, v26.8h\n"
+    "fmax v31.8h, v31.8h, v26.8h\n"
+    "fmin v28.8h, v28.8h, v27.8h\n"
+    "fmin v29.8h, v29.8h, v27.8h\n"
+    "fmin v30.8h, v30.8h, v27.8h\n"
+    "fmin v31.8h, v31.8h, v27.8h\n"
     "tbz %x[n_channels], #2, 77f\n"
     "st1 { v28.d }[0], [x12], #0x8\n"
     "st1 { v29.d }[0], [x11], #0x8\n"
@@ -887,7 +887,7 @@ void a64_fp16_nhwc_3x3_s2_output2x2_mla_depthfirst_indirect_impl(
     "80:"  // 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", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x9", "x10", "x11", "x12", "x13", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28"
   );
 }
 
-- 
cgit v1.2.1