From 4ee8b1599dbaf7634d25607fa5ac96ba3dc6b0f2 Mon Sep 17 00:00:00 2001
From: Georgios Pinitas <georgios.pinitas@arm.com>
Date: Fri, 16 Jul 2021 16:16:43 +0100
Subject: Update GEMM assembly kernels

- Introduce Fp32 kernels with internal calculations in Bfloat16 when
fast_mode is enabled
- Improve kernel selection heuristics

Signed-off-by: Georgios Pinitas <georgios.pinitas@arm.com>
Change-Id: I68a9e7e862b6fd2721b46e0d7cc791091c4ab279
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/5965
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
---
 .../a64_interleaved_bf16fp32_mmla_8x12/generic.cpp | 647 ++++++++-------------
 1 file changed, 255 insertions(+), 392 deletions(-)

(limited to 'src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_bf16fp32_mmla_8x12')

diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_bf16fp32_mmla_8x12/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_bf16fp32_mmla_8x12/generic.cpp
index c476fcf171..94c72a31c9 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_bf16fp32_mmla_8x12/generic.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_bf16fp32_mmla_8x12/generic.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2019-2020 Arm Limited.
+ * Copyright (c) 2021 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -23,406 +23,269 @@
  */
 #ifdef __aarch64__
 
+#include <cstddef>
 #include "../../bfloat.hpp"
-#include "../../asmlib.hpp"
 
 namespace arm_gemm {
 
-void a64_interleaved_bf16fp32_mmla_8x12(const bfloat16 *Apanel, const bfloat16 *Bpanel, float *Cpanel, int ablocks, int bblocks, int K) {
-    const bfloat16 *a_ptr = Apanel;
-    float *c_ptr = Cpanel;
+void a64_interleaved_bf16fp32_mmla_8x12(
+    const bfloat16 *Apanel, const bfloat16 *Bpanel,
+    float *Cpanel, int ablocks, int bblocks, int K) {
 
-    K /= 4;
-    const long loops_count = (K / 2) - 1;
-    const long tails_count = K % 2;
+    struct KernelArgs {
+        size_t bblocks = {};
+        size_t K = {};
+        const bfloat16 *Bpanel = {};
+    } ka;
 
-    for (int yb=0; yb<ablocks; yb++) {
-        const bfloat16 *a_ptr0 = a_ptr;
-        const bfloat16 *b_ptr = Bpanel;
+    ka.bblocks = bblocks;
+    ka.K = (K/4) - 1;
+    ka.Bpanel = Bpanel;
 
-        for (int xb=0; xb<bblocks; xb++) {
-            a_ptr = a_ptr0;
-            long loops = loops_count;
-            long tails = tails_count;
+    __asm__ __volatile__(
 
-            __asm __volatile (
-                "movi v8.4s, #0\n"
-                "ldr q0, [%[a_ptr]]\n"
-                "movi v9.4s, #0\n"
-                "ldr q4, [%[b_ptr]]\n"
-                "movi v10.4s, #0\n"
-                "ldr q1, [%[a_ptr], #0x10]\n"
-                "movi v11.4s, #0\n"
-                "ldr q5, [%[b_ptr], #0x10]\n"
-                "movi v12.4s, #0\n"
-                "ldr q2, [%[a_ptr], #0x20]\n"
-                "movi v13.4s, #0\n"
-                "ldr q6, [%[b_ptr], #0x20]\n"
-                "movi v14.4s, #0\n"
-                "prfm PLDL1KEEP, [%[a_ptr], #0x40]\n"
-                "movi v15.4s, #0\n"
-                "prfm PLDL1KEEP, [%[b_ptr], #0x40]\n"
-                "movi v16.4s, #0\n"
-                "prfm PLDL1KEEP, [%[a_ptr], #0x80]\n"
-                "movi v17.4s, #0\n"
-                "prfm PLDL1KEEP, [%[b_ptr], #0x80]\n"
-                "movi v18.4s, #0\n"
-                "prfm PLDL1KEEP, [%[a_ptr], #0xc0]\n"
-                "movi v19.4s, #0\n"
-                "prfm PLDL1KEEP, [%[b_ptr], #0xc0]\n"
-                "movi v20.4s, #0\n"
-                "prfm PLDL1KEEP, [%[a_ptr], #0x100]\n"
-                "movi v21.4s, #0\n"
-                "prfm PLDL1KEEP, [%[b_ptr], #0x100]\n"
-                "movi v22.4s, #0\n"
-                "prfm PLDL1KEEP, [%[a_ptr], #0x140]\n"
-                "movi v23.4s, #0\n"
-                "prfm PLDL1KEEP, [%[b_ptr], #0x140]\n"
-                "movi v24.4s, #0\n"
-                "prfm PLDL1KEEP, [%[a_ptr], #0x180]\n"
-                "movi v25.4s, #0\n"
-                "prfm PLDL1KEEP, [%[b_ptr], #0x180]\n"
-                "movi v26.4s, #0\n"
-                "prfm PLDL1KEEP, [%[a_ptr], #0x1c0]\n"
-                "movi v27.4s, #0\n"
-                "prfm PLDL1KEEP, [%[b_ptr], #0x1c0]\n"
-                "movi v28.4s, #0\n"
-                "prfm PLDL1KEEP, [%[a_ptr], #0x200]\n"
-                "movi v29.4s, #0\n"
-                "prfm PLDL1KEEP, [%[b_ptr], #0x200]\n"
-                "movi v30.4s, #0\n"
-                "prfm PLDL1KEEP, [%[a_ptr], #0x240]\n"
-                "movi v31.4s, #0\n"
-                "prfm PLDL1KEEP, [%[b_ptr], #0x240]\n"
-                "prfm PLDL1KEEP, [%[a_ptr], #0x280]\n"
-                "prfm PLDL1KEEP, [%[b_ptr], #0x280]\n"
-                "prfm PLDL1KEEP, [%[a_ptr], #0x2c0]\n"
-                "prfm PLDL1KEEP, [%[b_ptr], #0x2c0]\n"
-                "prfm PLDL1KEEP, [%[b_ptr], #0x300]\n"
-                "prfm PLDL1KEEP, [%[b_ptr], #0x340]\n"
-                "prfm PLDL1KEEP, [%[b_ptr], #0x380]\n"
-                "prfm PLDL1KEEP, [%[b_ptr], #0x3c0]\n"
-                "prfm PLDL1KEEP, [%[b_ptr], #0x400]\n"
-                "prfm PLDL1KEEP, [%[b_ptr], #0x440]\n"
-                "add %[a_ptr], %[a_ptr], #0x40\n"
-                "add %[b_ptr], %[b_ptr], #0x40\n"
-                "cbz %[loops], 1f\n"
-                "2:\n"
-                ".inst 0x6e44ec08 // bfmmla v8.4s, v0.8h, v4.8h\n"
-                "ldr q7, [%[b_ptr], #-0x10]\n"
-                ".inst 0x6e44ec2e // bfmmla v14.4s, v1.8h, v4.8h\n"
-                "ldr q3, [%[a_ptr], #-0x10]\n"
-                ".inst 0x6e44ec54 // bfmmla v20.4s, v2.8h, v4.8h\n"
-                "subs %[loops], %[loops], #0x1\n"
-                ".inst 0x6e45ec09 // bfmmla v9.4s, v0.8h, v5.8h\n"
-                "prfm PLDL1KEEP, [%[a_ptr], #0x2c0]\n"
-                ".inst 0x6e44ec7a // bfmmla v26.4s, v3.8h, v4.8h\n"
-                "ldr q4, [%[b_ptr]]\n"
-                ".inst 0x6e45ec2f // bfmmla v15.4s, v1.8h, v5.8h\n"
-                "prfm PLDL1KEEP, [%[b_ptr], #0x440]\n"
-                ".inst 0x6e45ec55 // bfmmla v21.4s, v2.8h, v5.8h\n"
-                "prfm PLDL1KEEP, [%[a_ptr], #0x300]\n"
-                ".inst 0x6e45ec7b // bfmmla v27.4s, v3.8h, v5.8h\n"
-                "ldr q5, [%[b_ptr], #0x10]\n"
-                ".inst 0x6e46ec0a // bfmmla v10.4s, v0.8h, v6.8h\n"
-                "prfm PLDL1KEEP, [%[b_ptr], #0x480]\n"
-                ".inst 0x6e46ec30 // bfmmla v16.4s, v1.8h, v6.8h\n"
-                "prfm PLDL1KEEP, [%[b_ptr], #0x4c0]\n"
-                ".inst 0x6e46ec56 // bfmmla v22.4s, v2.8h, v6.8h\n"
-                ".inst 0x6e46ec7c // bfmmla v28.4s, v3.8h, v6.8h\n"
-                "ldr q6, [%[b_ptr], #0x20]\n"
-                ".inst 0x6e47ec0b // bfmmla v11.4s, v0.8h, v7.8h\n"
-                ".inst 0x6e47ec31 // bfmmla v17.4s, v1.8h, v7.8h\n"
-                ".inst 0x6e47ec57 // bfmmla v23.4s, v2.8h, v7.8h\n"
-                ".inst 0x6e47ec7d // bfmmla v29.4s, v3.8h, v7.8h\n"
-                "ldr q7, [%[b_ptr], #0x30]\n"
-                ".inst 0x6e44ec0c // bfmmla v12.4s, v0.8h, v4.8h\n"
-                ".inst 0x6e44ec32 // bfmmla v18.4s, v1.8h, v4.8h\n"
-                ".inst 0x6e44ec58 // bfmmla v24.4s, v2.8h, v4.8h\n"
-                ".inst 0x6e44ec7e // bfmmla v30.4s, v3.8h, v4.8h\n"
-                "ldr q4, [%[b_ptr], #0x40]\n"
-                ".inst 0x6e45ec0d // bfmmla v13.4s, v0.8h, v5.8h\n"
-                "ldr q0, [%[a_ptr]]\n"
-                ".inst 0x6e45ec33 // bfmmla v19.4s, v1.8h, v5.8h\n"
-                "ldr q1, [%[a_ptr], #0x10]\n"
-                ".inst 0x6e45ec59 // bfmmla v25.4s, v2.8h, v5.8h\n"
-                "ldr q2, [%[a_ptr], #0x20]\n"
-                ".inst 0x6e45ec7f // bfmmla v31.4s, v3.8h, v5.8h\n"
-                "ldr q5, [%[b_ptr], #0x50]\n"
-                ".inst 0x6e46ec08 // bfmmla v8.4s, v0.8h, v6.8h\n"
-                "ldr q3, [%[a_ptr], #0x30]\n"
-                ".inst 0x6e46ec2e // bfmmla v14.4s, v1.8h, v6.8h\n"
-                "add %[a_ptr], %[a_ptr], #0x80\n"
-                ".inst 0x6e46ec54 // bfmmla v20.4s, v2.8h, v6.8h\n"
-                "add %[b_ptr], %[b_ptr], #0xc0\n"
-                ".inst 0x6e46ec7a // bfmmla v26.4s, v3.8h, v6.8h\n"
-                "ldr q6, [%[b_ptr], #-0x60]\n"
-                ".inst 0x6e47ec09 // bfmmla v9.4s, v0.8h, v7.8h\n"
-                ".inst 0x6e47ec2f // bfmmla v15.4s, v1.8h, v7.8h\n"
-                ".inst 0x6e47ec55 // bfmmla v21.4s, v2.8h, v7.8h\n"
-                ".inst 0x6e47ec7b // bfmmla v27.4s, v3.8h, v7.8h\n"
-                "ldr q7, [%[b_ptr], #-0x50]\n"
-                ".inst 0x6e44ec0a // bfmmla v10.4s, v0.8h, v4.8h\n"
-                ".inst 0x6e44ec30 // bfmmla v16.4s, v1.8h, v4.8h\n"
-                ".inst 0x6e44ec56 // bfmmla v22.4s, v2.8h, v4.8h\n"
-                ".inst 0x6e44ec7c // bfmmla v28.4s, v3.8h, v4.8h\n"
-                "ldr q4, [%[b_ptr], #-0x40]\n"
-                ".inst 0x6e45ec0b // bfmmla v11.4s, v0.8h, v5.8h\n"
-                ".inst 0x6e45ec31 // bfmmla v17.4s, v1.8h, v5.8h\n"
-                ".inst 0x6e45ec57 // bfmmla v23.4s, v2.8h, v5.8h\n"
-                ".inst 0x6e45ec7d // bfmmla v29.4s, v3.8h, v5.8h\n"
-                "ldr q5, [%[b_ptr], #-0x30]\n"
-                ".inst 0x6e46ec0c // bfmmla v12.4s, v0.8h, v6.8h\n"
-                ".inst 0x6e46ec32 // bfmmla v18.4s, v1.8h, v6.8h\n"
-                ".inst 0x6e46ec58 // bfmmla v24.4s, v2.8h, v6.8h\n"
-                ".inst 0x6e46ec7e // bfmmla v30.4s, v3.8h, v6.8h\n"
-                "ldr q6, [%[b_ptr], #-0x20]\n"
-                ".inst 0x6e47ec0d // bfmmla v13.4s, v0.8h, v7.8h\n"
-                "ldr q0, [%[a_ptr], #-0x40]\n"
-                ".inst 0x6e47ec33 // bfmmla v19.4s, v1.8h, v7.8h\n"
-                "ldr q1, [%[a_ptr], #-0x30]\n"
-                ".inst 0x6e47ec59 // bfmmla v25.4s, v2.8h, v7.8h\n"
-                "ldr q2, [%[a_ptr], #-0x20]\n"
-                ".inst 0x6e47ec7f // bfmmla v31.4s, v3.8h, v7.8h\n"
-                "b.ne 2b\n"
-                "1:\n"
-                "cbz %[tails], 3f\n"
-                ".inst 0x6e44ec08 // bfmmla v8.4s, v0.8h, v4.8h\n"
-                "ldr q7, [%[b_ptr], #-0x10]\n"
-                ".inst 0x6e44ec2e // bfmmla v14.4s, v1.8h, v4.8h\n"
-                "ldr q3, [%[a_ptr], #-0x10]\n"
-                ".inst 0x6e44ec54 // bfmmla v20.4s, v2.8h, v4.8h\n"
-                ".inst 0x6e45ec09 // bfmmla v9.4s, v0.8h, v5.8h\n"
-                ".inst 0x6e45ec2f // bfmmla v15.4s, v1.8h, v5.8h\n"
-                ".inst 0x6e44ec7a // bfmmla v26.4s, v3.8h, v4.8h\n"
-                "ldr q4, [%[b_ptr]]\n"
-                ".inst 0x6e45ec55 // bfmmla v21.4s, v2.8h, v5.8h\n"
-                ".inst 0x6e45ec7b // bfmmla v27.4s, v3.8h, v5.8h\n"
-                "ldr q5, [%[b_ptr], #0x10]\n"
-                ".inst 0x6e46ec0a // bfmmla v10.4s, v0.8h, v6.8h\n"
-                ".inst 0x6e46ec30 // bfmmla v16.4s, v1.8h, v6.8h\n"
-                ".inst 0x6e46ec56 // bfmmla v22.4s, v2.8h, v6.8h\n"
-                ".inst 0x6e46ec7c // bfmmla v28.4s, v3.8h, v6.8h\n"
-                "ldr q6, [%[b_ptr], #0x20]\n"
-                ".inst 0x6e47ec0b // bfmmla v11.4s, v0.8h, v7.8h\n"
-                ".inst 0x6e47ec31 // bfmmla v17.4s, v1.8h, v7.8h\n"
-                ".inst 0x6e47ec57 // bfmmla v23.4s, v2.8h, v7.8h\n"
-                ".inst 0x6e47ec7d // bfmmla v29.4s, v3.8h, v7.8h\n"
-                "ldr q7, [%[b_ptr], #0x30]\n"
-                ".inst 0x6e44ec0c // bfmmla v12.4s, v0.8h, v4.8h\n"
-                ".inst 0x6e44ec32 // bfmmla v18.4s, v1.8h, v4.8h\n"
-                ".inst 0x6e44ec58 // bfmmla v24.4s, v2.8h, v4.8h\n"
-                ".inst 0x6e44ec7e // bfmmla v30.4s, v3.8h, v4.8h\n"
-                "ldr q4, [%[b_ptr], #0x40]\n"
-                ".inst 0x6e45ec0d // bfmmla v13.4s, v0.8h, v5.8h\n"
-                "ldr q0, [%[a_ptr]]\n"
-                ".inst 0x6e45ec33 // bfmmla v19.4s, v1.8h, v5.8h\n"
-                "ldr q1, [%[a_ptr], #0x10]\n"
-                ".inst 0x6e45ec59 // bfmmla v25.4s, v2.8h, v5.8h\n"
-                "ldr q2, [%[a_ptr], #0x20]\n"
-                ".inst 0x6e45ec7f // bfmmla v31.4s, v3.8h, v5.8h\n"
-                "ldr q5, [%[b_ptr], #0x50]\n"
-                ".inst 0x6e46ec08 // bfmmla v8.4s, v0.8h, v6.8h\n"
-                "ldr q3, [%[a_ptr], #0x30]\n"
-                ".inst 0x6e46ec2e // bfmmla v14.4s, v1.8h, v6.8h\n"
-                "add %[a_ptr], %[a_ptr], #0x80\n"
-                ".inst 0x6e46ec54 // bfmmla v20.4s, v2.8h, v6.8h\n"
-                "add %[b_ptr], %[b_ptr], #0xe0\n"
-                ".inst 0x6e46ec7a // bfmmla v26.4s, v3.8h, v6.8h\n"
-                "ldr q6, [%[b_ptr], #-0x80]\n"
-                ".inst 0x6e47ec09 // bfmmla v9.4s, v0.8h, v7.8h\n"
-                ".inst 0x6e47ec2f // bfmmla v15.4s, v1.8h, v7.8h\n"
-                ".inst 0x6e47ec55 // bfmmla v21.4s, v2.8h, v7.8h\n"
-                ".inst 0x6e47ec7b // bfmmla v27.4s, v3.8h, v7.8h\n"
-                "ldr q7, [%[b_ptr], #-0x70]\n"
-                ".inst 0x6e44ec0a // bfmmla v10.4s, v0.8h, v4.8h\n"
-                ".inst 0x6e44ec30 // bfmmla v16.4s, v1.8h, v4.8h\n"
-                ".inst 0x6e44ec56 // bfmmla v22.4s, v2.8h, v4.8h\n"
-                ".inst 0x6e44ec7c // bfmmla v28.4s, v3.8h, v4.8h\n"
-                "ldr q4, [%[b_ptr], #-0x60]\n"
-                ".inst 0x6e45ec0b // bfmmla v11.4s, v0.8h, v5.8h\n"
-                ".inst 0x6e45ec31 // bfmmla v17.4s, v1.8h, v5.8h\n"
-                ".inst 0x6e45ec57 // bfmmla v23.4s, v2.8h, v5.8h\n"
-                ".inst 0x6e45ec7d // bfmmla v29.4s, v3.8h, v5.8h\n"
-                "ldr q5, [%[b_ptr], #-0x50]\n"
-                ".inst 0x6e46ec0c // bfmmla v12.4s, v0.8h, v6.8h\n"
-                ".inst 0x6e46ec32 // bfmmla v18.4s, v1.8h, v6.8h\n"
-                ".inst 0x6e46ec58 // bfmmla v24.4s, v2.8h, v6.8h\n"
-                ".inst 0x6e46ec7e // bfmmla v30.4s, v3.8h, v6.8h\n"
-                "ldr q6, [%[b_ptr], #-0x40]\n"
-                ".inst 0x6e47ec0d // bfmmla v13.4s, v0.8h, v7.8h\n"
-                "ldr q0, [%[a_ptr], #-0x40]\n"
-                ".inst 0x6e47ec33 // bfmmla v19.4s, v1.8h, v7.8h\n"
-                "ldr q1, [%[a_ptr], #-0x30]\n"
-                ".inst 0x6e47ec59 // bfmmla v25.4s, v2.8h, v7.8h\n"
-                "ldr q2, [%[a_ptr], #-0x20]\n"
-                ".inst 0x6e47ec7f // bfmmla v31.4s, v3.8h, v7.8h\n"
-                "ldr q7, [%[b_ptr], #-0x30]\n"
-                ".inst 0x6e44ec08 // bfmmla v8.4s, v0.8h, v4.8h\n"
-                "ldr q3, [%[a_ptr], #-0x10]\n"
-                ".inst 0x6e44ec2e // bfmmla v14.4s, v1.8h, v4.8h\n"
-                ".inst 0x6e44ec54 // bfmmla v20.4s, v2.8h, v4.8h\n"
-                ".inst 0x6e45ec09 // bfmmla v9.4s, v0.8h, v5.8h\n"
-                ".inst 0x6e44ec7a // bfmmla v26.4s, v3.8h, v4.8h\n"
-                "ldr q4, [%[b_ptr], #-0x20]\n"
-                ".inst 0x6e45ec2f // bfmmla v15.4s, v1.8h, v5.8h\n"
-                ".inst 0x6e45ec55 // bfmmla v21.4s, v2.8h, v5.8h\n"
-                ".inst 0x6e45ec7b // bfmmla v27.4s, v3.8h, v5.8h\n"
-                "ldr q5, [%[b_ptr], #-0x10]\n"
-                ".inst 0x6e46ec0a // bfmmla v10.4s, v0.8h, v6.8h\n"
-                ".inst 0x6e46ec30 // bfmmla v16.4s, v1.8h, v6.8h\n"
-                ".inst 0x6e46ec56 // bfmmla v22.4s, v2.8h, v6.8h\n"
-                ".inst 0x6e46ec7c // bfmmla v28.4s, v3.8h, v6.8h\n"
-                "uzp1 v6.2d, v14.2d, v15.2d\n"
-                ".inst 0x6e47ec0b // bfmmla v11.4s, v0.8h, v7.8h\n"
-                ".inst 0x6e47ec31 // bfmmla v17.4s, v1.8h, v7.8h\n"
-                ".inst 0x6e47ec57 // bfmmla v23.4s, v2.8h, v7.8h\n"
-                ".inst 0x6e47ec7d // bfmmla v29.4s, v3.8h, v7.8h\n"
-                ".inst 0x6e44ec0c // bfmmla v12.4s, v0.8h, v4.8h\n"
-                "uzp1 v7.2d, v16.2d, v17.2d\n"
-                ".inst 0x6e44ec32 // bfmmla v18.4s, v1.8h, v4.8h\n"
-                ".inst 0x6e44ec58 // bfmmla v24.4s, v2.8h, v4.8h\n"
-                ".inst 0x6e44ec7e // bfmmla v30.4s, v3.8h, v4.8h\n"
-                "uzp2 v4.2d, v10.2d, v11.2d\n"
-                ".inst 0x6e45ec0d // bfmmla v13.4s, v0.8h, v5.8h\n"
-                "uzp1 v0.2d, v8.2d, v9.2d\n"
-                ".inst 0x6e45ec33 // bfmmla v19.4s, v1.8h, v5.8h\n"
-                "uzp1 v1.2d, v10.2d, v11.2d\n"
-                ".inst 0x6e45ec59 // bfmmla v25.4s, v2.8h, v5.8h\n"
-                "str q0, [%[c_ptr]]\n"
-                "uzp1 v2.2d, v12.2d, v13.2d\n"
-                "uzp1 v0.2d, v18.2d, v19.2d\n"
-                ".inst 0x6e45ec7f // bfmmla v31.4s, v3.8h, v5.8h\n"
-                "str q1, [%[c_ptr], #0x10]\n"
-                "uzp2 v3.2d, v8.2d, v9.2d\n"
-                "uzp2 v5.2d, v12.2d, v13.2d\n"
-                "uzp2 v1.2d, v14.2d, v15.2d\n"
-                "str q2, [%[c_ptr], #0x20]\n"
-                "b 4f\n"
-                "3:\n"
-                ".inst 0x6e44ec08 // bfmmla v8.4s, v0.8h, v4.8h\n"
-                "ldr q7, [%[b_ptr], #-0x10]\n"
-                ".inst 0x6e44ec2e // bfmmla v14.4s, v1.8h, v4.8h\n"
-                "ldr q3, [%[a_ptr], #-0x10]\n"
-                ".inst 0x6e44ec54 // bfmmla v20.4s, v2.8h, v4.8h\n"
-                "add %[a_ptr], %[a_ptr], #0x40\n"
-                ".inst 0x6e45ec09 // bfmmla v9.4s, v0.8h, v5.8h\n"
-                "add %[b_ptr], %[b_ptr], #0x80\n"
-                ".inst 0x6e44ec7a // bfmmla v26.4s, v3.8h, v4.8h\n"
-                "ldr q4, [%[b_ptr], #-0x80]\n"
-                ".inst 0x6e45ec2f // bfmmla v15.4s, v1.8h, v5.8h\n"
-                ".inst 0x6e45ec55 // bfmmla v21.4s, v2.8h, v5.8h\n"
-                ".inst 0x6e45ec7b // bfmmla v27.4s, v3.8h, v5.8h\n"
-                "ldr q5, [%[b_ptr], #-0x70]\n"
-                ".inst 0x6e46ec0a // bfmmla v10.4s, v0.8h, v6.8h\n"
-                ".inst 0x6e46ec30 // bfmmla v16.4s, v1.8h, v6.8h\n"
-                ".inst 0x6e46ec56 // bfmmla v22.4s, v2.8h, v6.8h\n"
-                ".inst 0x6e46ec7c // bfmmla v28.4s, v3.8h, v6.8h\n"
-                "ldr q6, [%[b_ptr], #-0x60]\n"
-                ".inst 0x6e47ec0b // bfmmla v11.4s, v0.8h, v7.8h\n"
-                ".inst 0x6e47ec31 // bfmmla v17.4s, v1.8h, v7.8h\n"
-                ".inst 0x6e47ec57 // bfmmla v23.4s, v2.8h, v7.8h\n"
-                ".inst 0x6e47ec7d // bfmmla v29.4s, v3.8h, v7.8h\n"
-                "ldr q7, [%[b_ptr], #-0x50]\n"
-                ".inst 0x6e44ec0c // bfmmla v12.4s, v0.8h, v4.8h\n"
-                ".inst 0x6e44ec32 // bfmmla v18.4s, v1.8h, v4.8h\n"
-                ".inst 0x6e44ec58 // bfmmla v24.4s, v2.8h, v4.8h\n"
-                ".inst 0x6e44ec7e // bfmmla v30.4s, v3.8h, v4.8h\n"
-                "ldr q4, [%[b_ptr], #-0x40]\n"
-                ".inst 0x6e45ec0d // bfmmla v13.4s, v0.8h, v5.8h\n"
-                "ldr q0, [%[a_ptr], #-0x40]\n"
-                ".inst 0x6e45ec33 // bfmmla v19.4s, v1.8h, v5.8h\n"
-                "ldr q1, [%[a_ptr], #-0x30]\n"
-                ".inst 0x6e45ec59 // bfmmla v25.4s, v2.8h, v5.8h\n"
-                "ldr q2, [%[a_ptr], #-0x20]\n"
-                ".inst 0x6e45ec7f // bfmmla v31.4s, v3.8h, v5.8h\n"
-                "ldr q5, [%[b_ptr], #-0x30]\n"
-                ".inst 0x6e46ec08 // bfmmla v8.4s, v0.8h, v6.8h\n"
-                "ldr q3, [%[a_ptr], #-0x10]\n"
-                ".inst 0x6e46ec2e // bfmmla v14.4s, v1.8h, v6.8h\n"
-                ".inst 0x6e46ec54 // bfmmla v20.4s, v2.8h, v6.8h\n"
-                ".inst 0x6e47ec09 // bfmmla v9.4s, v0.8h, v7.8h\n"
-                ".inst 0x6e46ec7a // bfmmla v26.4s, v3.8h, v6.8h\n"
-                "ldr q6, [%[b_ptr], #-0x20]\n"
-                ".inst 0x6e47ec2f // bfmmla v15.4s, v1.8h, v7.8h\n"
-                ".inst 0x6e47ec55 // bfmmla v21.4s, v2.8h, v7.8h\n"
-                ".inst 0x6e47ec7b // bfmmla v27.4s, v3.8h, v7.8h\n"
-                "ldr q7, [%[b_ptr], #-0x10]\n"
-                ".inst 0x6e44ec0a // bfmmla v10.4s, v0.8h, v4.8h\n"
-                ".inst 0x6e44ec30 // bfmmla v16.4s, v1.8h, v4.8h\n"
-                ".inst 0x6e44ec56 // bfmmla v22.4s, v2.8h, v4.8h\n"
-                ".inst 0x6e44ec7c // bfmmla v28.4s, v3.8h, v4.8h\n"
-                ".inst 0x6e45ec0b // bfmmla v11.4s, v0.8h, v5.8h\n"
-                ".inst 0x6e45ec31 // bfmmla v17.4s, v1.8h, v5.8h\n"
-                ".inst 0x6e45ec57 // bfmmla v23.4s, v2.8h, v5.8h\n"
-                ".inst 0x6e45ec7d // bfmmla v29.4s, v3.8h, v5.8h\n"
-                "uzp2 v4.2d, v10.2d, v11.2d\n"
-                ".inst 0x6e46ec0c // bfmmla v12.4s, v0.8h, v6.8h\n"
-                ".inst 0x6e46ec32 // bfmmla v18.4s, v1.8h, v6.8h\n"
-                ".inst 0x6e46ec58 // bfmmla v24.4s, v2.8h, v6.8h\n"
-                ".inst 0x6e46ec7e // bfmmla v30.4s, v3.8h, v6.8h\n"
-                "uzp1 v6.2d, v14.2d, v15.2d\n"
-                ".inst 0x6e47ec0d // bfmmla v13.4s, v0.8h, v7.8h\n"
-                "uzp1 v0.2d, v8.2d, v9.2d\n"
-                ".inst 0x6e47ec33 // bfmmla v19.4s, v1.8h, v7.8h\n"
-                "uzp1 v1.2d, v10.2d, v11.2d\n"
-                "uzp2 v5.2d, v12.2d, v13.2d\n"
-                "str q0, [%[c_ptr]]\n"
-                ".inst 0x6e47ec59 // bfmmla v25.4s, v2.8h, v7.8h\n"
-                "uzp1 v2.2d, v12.2d, v13.2d\n"
-                "uzp1 v0.2d, v18.2d, v19.2d\n"
-                "str q1, [%[c_ptr], #0x10]\n"
-                "uzp2 v1.2d, v14.2d, v15.2d\n"
-                ".inst 0x6e47ec7f // bfmmla v31.4s, v3.8h, v7.8h\n"
-                "uzp2 v3.2d, v8.2d, v9.2d\n"
-                "str q2, [%[c_ptr], #0x20]\n"
-                "uzp1 v7.2d, v16.2d, v17.2d\n"
-                "4:\n"
-                "uzp2 v2.2d, v16.2d, v17.2d\n"
-                "str q3, [%[c_ptr], #0x30]\n"
-                "uzp2 v3.2d, v18.2d, v19.2d\n"
-                "str q4, [%[c_ptr], #0x40]\n"
-                "uzp1 v4.2d, v20.2d, v21.2d\n"
-                "str q5, [%[c_ptr], #0x50]\n"
-                "uzp1 v5.2d, v22.2d, v23.2d\n"
-                "str q6, [%[c_ptr], #0x60]\n"
-                "uzp1 v6.2d, v24.2d, v25.2d\n"
-                "str q7, [%[c_ptr], #0x70]\n"
-                "uzp2 v7.2d, v20.2d, v21.2d\n"
-                "str q0, [%[c_ptr], #0x80]\n"
-                "uzp2 v0.2d, v22.2d, v23.2d\n"
-                "str q1, [%[c_ptr], #0x90]\n"
-                "uzp2 v1.2d, v24.2d, v25.2d\n"
-                "str q2, [%[c_ptr], #0xa0]\n"
-                "uzp1 v2.2d, v26.2d, v27.2d\n"
-                "str q3, [%[c_ptr], #0xb0]\n"
-                "uzp1 v3.2d, v28.2d, v29.2d\n"
-                "str q4, [%[c_ptr], #0xc0]\n"
-                "uzp1 v4.2d, v30.2d, v31.2d\n"
-                "str q5, [%[c_ptr], #0xd0]\n"
-                "uzp2 v5.2d, v26.2d, v27.2d\n"
-                "str q6, [%[c_ptr], #0xe0]\n"
-                "uzp2 v6.2d, v28.2d, v29.2d\n"
-                "str q7, [%[c_ptr], #0xf0]\n"
-                "uzp2 v7.2d, v30.2d, v31.2d\n"
-                "str q0, [%[c_ptr], #0x100]\n"
-                "str q1, [%[c_ptr], #0x110]\n"
-                "str q2, [%[c_ptr], #0x120]\n"
-                "str q3, [%[c_ptr], #0x130]\n"
-                "str q4, [%[c_ptr], #0x140]\n"
-                "str q5, [%[c_ptr], #0x150]\n"
-                "str q6, [%[c_ptr], #0x160]\n"
-                "str q7, [%[c_ptr], #0x170]\n"
-                "add %[c_ptr], %[c_ptr], #0x180\n"
-            : [a_ptr] "+r" (a_ptr), [b_ptr] "+r" (b_ptr), [c_ptr] "+r" (c_ptr),
-              [loops] "+r" (loops), [tails] "+r" (tails)
-            :
-            : "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", "cc", "memory"
-            );
-        }
-    }
+      "1:"  // Height loop
+      "ldr x22, [%x[args_ptr], %[offsetof_bblocks]]\n"
+      "mov x21, %x[Apanel]\n"
+      "ldr x20, [%x[args_ptr], %[offsetof_Bpanel]]\n"
+      "2:"  // Width loop
+      "ldr x19, [%x[args_ptr], %[offsetof_K]]\n"
+      "mov %x[Apanel], x21\n"
+      "cmp x19, #0x2\n"
+      "movi v8.16b, #0x0\n"
+      "movi v9.16b, #0x0\n"
+      "ldr q4, [x20, #0x0]\n"
+      "movi v10.16b, #0x0\n"
+      "movi v11.16b, #0x0\n"
+      "ldr q0, [%x[Apanel], #0x0]\n"
+      "movi v12.16b, #0x0\n"
+      "movi v13.16b, #0x0\n"
+      "ldr q1, [%x[Apanel], #0x10]\n"
+      "movi v14.16b, #0x0\n"
+      "movi v15.16b, #0x0\n"
+      "ldr q5, [x20, #0x10]\n"
+      "movi v16.16b, #0x0\n"
+      "movi v17.16b, #0x0\n"
+      "ldr q2, [%x[Apanel], #0x20]\n"
+      "movi v18.16b, #0x0\n"
+      "movi v19.16b, #0x0\n"
+      "add x20, x20, #0x20\n"
+      "movi v20.16b, #0x0\n"
+      "movi v21.16b, #0x0\n"
+      "add %x[Apanel], %x[Apanel], #0x30\n"
+      "movi v22.16b, #0x0\n"
+      "movi v23.16b, #0x0\n"
+      "movi v24.16b, #0x0\n"
+      "movi v25.16b, #0x0\n"
+      "movi v26.16b, #0x0\n"
+      "movi v27.16b, #0x0\n"
+      "movi v28.16b, #0x0\n"
+      "movi v29.16b, #0x0\n"
+      "movi v30.16b, #0x0\n"
+      "movi v31.16b, #0x0\n"
+      "blt 4f\n"
+      "3:"  // main loop head
+      "ldr q3, [%x[Apanel], #0x0]\n"
+      ".inst 0x6e44ec08  // bfmmla v8.4s, v0.8h, v4.8h\n"
+      ".inst 0x6e44ec2e  // bfmmla v14.4s, v1.8h, v4.8h\n"
+      ".inst 0x6e45ec0b  // bfmmla v11.4s, v0.8h, v5.8h\n"
+      ".inst 0x6e45ec31  // bfmmla v17.4s, v1.8h, v5.8h\n"
+      "ldr q6, [x20, #0x0]\n"
+      ".inst 0x6e44ec54  // bfmmla v20.4s, v2.8h, v4.8h\n"
+      ".inst 0x6e45ec57  // bfmmla v23.4s, v2.8h, v5.8h\n"
+      "ldr q7, [x20, #0x10]\n"
+      ".inst 0x6e44ec7a  // bfmmla v26.4s, v3.8h, v4.8h\n"
+      ".inst 0x6e45ec7d  // bfmmla v29.4s, v3.8h, v5.8h\n"
+      "ldr q4, [x20, #0x20]\n"
+      "ldr q5, [x20, #0x30]\n"
+      ".inst 0x6e46ec09  // bfmmla v9.4s, v0.8h, v6.8h\n"
+      ".inst 0x6e46ec2f  // bfmmla v15.4s, v1.8h, v6.8h\n"
+      ".inst 0x6e46ec55  // bfmmla v21.4s, v2.8h, v6.8h\n"
+      ".inst 0x6e46ec7b  // bfmmla v27.4s, v3.8h, v6.8h\n"
+      "ldr q6, [x20, #0x40]\n"
+      ".inst 0x6e47ec0c  // bfmmla v12.4s, v0.8h, v7.8h\n"
+      ".inst 0x6e44ec0a  // bfmmla v10.4s, v0.8h, v4.8h\n"
+      "sub x19, x19, #0x2\n"
+      ".inst 0x6e45ec0d  // bfmmla v13.4s, v0.8h, v5.8h\n"
+      ".inst 0x6e47ec32  // bfmmla v18.4s, v1.8h, v7.8h\n"
+      "ldr q0, [%x[Apanel], #0x10]\n"
+      ".inst 0x6e44ec30  // bfmmla v16.4s, v1.8h, v4.8h\n"
+      ".inst 0x6e45ec33  // bfmmla v19.4s, v1.8h, v5.8h\n"
+      "ldr q1, [%x[Apanel], #0x20]\n"
+      ".inst 0x6e47ec58  // bfmmla v24.4s, v2.8h, v7.8h\n"
+      ".inst 0x6e47ec7e  // bfmmla v30.4s, v3.8h, v7.8h\n"
+      "ldr q7, [x20, #0x50]\n"
+      ".inst 0x6e44ec56  // bfmmla v22.4s, v2.8h, v4.8h\n"
+      ".inst 0x6e45ec59  // bfmmla v25.4s, v2.8h, v5.8h\n"
+      "ldr q2, [%x[Apanel], #0x30]\n"
+      ".inst 0x6e44ec7c  // bfmmla v28.4s, v3.8h, v4.8h\n"
+      ".inst 0x6e45ec7f  // bfmmla v31.4s, v3.8h, v5.8h\n"
+      "ldr q3, [%x[Apanel], #0x40]\n"
+      ".inst 0x6e46ec08  // bfmmla v8.4s, v0.8h, v6.8h\n"
+      ".inst 0x6e46ec2e  // bfmmla v14.4s, v1.8h, v6.8h\n"
+      "ldr q4, [x20, #0x60]\n"
+      ".inst 0x6e47ec0b  // bfmmla v11.4s, v0.8h, v7.8h\n"
+      ".inst 0x6e47ec31  // bfmmla v17.4s, v1.8h, v7.8h\n"
+      "ldr q5, [x20, #0x70]\n"
+      ".inst 0x6e46ec54  // bfmmla v20.4s, v2.8h, v6.8h\n"
+      ".inst 0x6e47ec57  // bfmmla v23.4s, v2.8h, v7.8h\n"
+      "cmp x19, #0x2\n"
+      ".inst 0x6e46ec7a  // bfmmla v26.4s, v3.8h, v6.8h\n"
+      ".inst 0x6e47ec7d  // bfmmla v29.4s, v3.8h, v7.8h\n"
+      "ldr q6, [x20, #0x80]\n"
+      "ldr q7, [x20, #0x90]\n"
+      ".inst 0x6e44ec09  // bfmmla v9.4s, v0.8h, v4.8h\n"
+      ".inst 0x6e44ec2f  // bfmmla v15.4s, v1.8h, v4.8h\n"
+      ".inst 0x6e44ec55  // bfmmla v21.4s, v2.8h, v4.8h\n"
+      ".inst 0x6e44ec7b  // bfmmla v27.4s, v3.8h, v4.8h\n"
+      "ldr q4, [x20, #0xa0]\n"
+      ".inst 0x6e45ec0c  // bfmmla v12.4s, v0.8h, v5.8h\n"
+      ".inst 0x6e46ec0a  // bfmmla v10.4s, v0.8h, v6.8h\n"
+      ".inst 0x6e47ec0d  // bfmmla v13.4s, v0.8h, v7.8h\n"
+      ".inst 0x6e45ec32  // bfmmla v18.4s, v1.8h, v5.8h\n"
+      "ldr q0, [%x[Apanel], #0x50]\n"
+      ".inst 0x6e46ec30  // bfmmla v16.4s, v1.8h, v6.8h\n"
+      ".inst 0x6e47ec33  // bfmmla v19.4s, v1.8h, v7.8h\n"
+      "ldr q1, [%x[Apanel], #0x60]\n"
+      ".inst 0x6e45ec58  // bfmmla v24.4s, v2.8h, v5.8h\n"
+      ".inst 0x6e45ec7e  // bfmmla v30.4s, v3.8h, v5.8h\n"
+      "ldr q5, [x20, #0xb0]\n"
+      ".inst 0x6e46ec56  // bfmmla v22.4s, v2.8h, v6.8h\n"
+      ".inst 0x6e47ec59  // bfmmla v25.4s, v2.8h, v7.8h\n"
+      "ldr q2, [%x[Apanel], #0x70]\n"
+      ".inst 0x6e46ec7c  // bfmmla v28.4s, v3.8h, v6.8h\n"
+      ".inst 0x6e47ec7f  // bfmmla v31.4s, v3.8h, v7.8h\n"
+      "add %x[Apanel], %x[Apanel], #0x80\n"
+      "add x20, x20, #0xc0\n"
+      "bge 3b\n"
+      "4:"  // main loop skip
+      "ldr q3, [%x[Apanel], #0x0]\n"
+      ".inst 0x6e44ec08  // bfmmla v8.4s, v0.8h, v4.8h\n"
+      ".inst 0x6e44ec2e  // bfmmla v14.4s, v1.8h, v4.8h\n"
+      ".inst 0x6e45ec0b  // bfmmla v11.4s, v0.8h, v5.8h\n"
+      ".inst 0x6e45ec31  // bfmmla v17.4s, v1.8h, v5.8h\n"
+      "ldr q6, [x20, #0x0]\n"
+      ".inst 0x6e44ec54  // bfmmla v20.4s, v2.8h, v4.8h\n"
+      ".inst 0x6e45ec57  // bfmmla v23.4s, v2.8h, v5.8h\n"
+      "ldr q7, [x20, #0x10]\n"
+      ".inst 0x6e44ec7a  // bfmmla v26.4s, v3.8h, v4.8h\n"
+      ".inst 0x6e45ec7d  // bfmmla v29.4s, v3.8h, v5.8h\n"
+      "ldr q4, [x20, #0x20]\n"
+      "ldr q5, [x20, #0x30]\n"
+      ".inst 0x6e46ec09  // bfmmla v9.4s, v0.8h, v6.8h\n"
+      ".inst 0x6e46ec2f  // bfmmla v15.4s, v1.8h, v6.8h\n"
+      ".inst 0x6e46ec55  // bfmmla v21.4s, v2.8h, v6.8h\n"
+      ".inst 0x6e46ec7b  // bfmmla v27.4s, v3.8h, v6.8h\n"
+      "add %x[Apanel], %x[Apanel], #0x10\n"
+      ".inst 0x6e47ec0c  // bfmmla v12.4s, v0.8h, v7.8h\n"
+      ".inst 0x6e44ec0a  // bfmmla v10.4s, v0.8h, v4.8h\n"
+      "add x20, x20, #0x40\n"
+      ".inst 0x6e45ec0d  // bfmmla v13.4s, v0.8h, v5.8h\n"
+      ".inst 0x6e47ec32  // bfmmla v18.4s, v1.8h, v7.8h\n"
+      ".inst 0x6e44ec30  // bfmmla v16.4s, v1.8h, v4.8h\n"
+      ".inst 0x6e45ec33  // bfmmla v19.4s, v1.8h, v5.8h\n"
+      ".inst 0x6e47ec58  // bfmmla v24.4s, v2.8h, v7.8h\n"
+      ".inst 0x6e47ec7e  // bfmmla v30.4s, v3.8h, v7.8h\n"
+      ".inst 0x6e44ec56  // bfmmla v22.4s, v2.8h, v4.8h\n"
+      ".inst 0x6e45ec59  // bfmmla v25.4s, v2.8h, v5.8h\n"
+      ".inst 0x6e44ec7c  // bfmmla v28.4s, v3.8h, v4.8h\n"
+      ".inst 0x6e45ec7f  // bfmmla v31.4s, v3.8h, v5.8h\n"
+      "cbz x19, 5f\n"
+      "ldr q6, [x20, #0x0]\n"
+      "ldr q0, [%x[Apanel], #0x0]\n"
+      ".inst 0x6e46ec08  // bfmmla v8.4s, v0.8h, v6.8h\n"
+      "ldr q1, [%x[Apanel], #0x10]\n"
+      "ldr q7, [x20, #0x10]\n"
+      ".inst 0x6e46ec2e  // bfmmla v14.4s, v1.8h, v6.8h\n"
+      "ldr q2, [%x[Apanel], #0x20]\n"
+      "ldr q3, [%x[Apanel], #0x30]\n"
+      ".inst 0x6e47ec0b  // bfmmla v11.4s, v0.8h, v7.8h\n"
+      ".inst 0x6e47ec31  // bfmmla v17.4s, v1.8h, v7.8h\n"
+      ".inst 0x6e46ec54  // bfmmla v20.4s, v2.8h, v6.8h\n"
+      "ldr q4, [x20, #0x20]\n"
+      ".inst 0x6e47ec57  // bfmmla v23.4s, v2.8h, v7.8h\n"
+      ".inst 0x6e46ec7a  // bfmmla v26.4s, v3.8h, v6.8h\n"
+      "ldr q5, [x20, #0x30]\n"
+      ".inst 0x6e47ec7d  // bfmmla v29.4s, v3.8h, v7.8h\n"
+      "ldr q6, [x20, #0x40]\n"
+      "ldr q7, [x20, #0x50]\n"
+      ".inst 0x6e44ec09  // bfmmla v9.4s, v0.8h, v4.8h\n"
+      ".inst 0x6e44ec2f  // bfmmla v15.4s, v1.8h, v4.8h\n"
+      "add x20, x20, #0x60\n"
+      ".inst 0x6e44ec55  // bfmmla v21.4s, v2.8h, v4.8h\n"
+      ".inst 0x6e44ec7b  // bfmmla v27.4s, v3.8h, v4.8h\n"
+      "add %x[Apanel], %x[Apanel], #0x40\n"
+      ".inst 0x6e45ec0c  // bfmmla v12.4s, v0.8h, v5.8h\n"
+      ".inst 0x6e46ec0a  // bfmmla v10.4s, v0.8h, v6.8h\n"
+      ".inst 0x6e47ec0d  // bfmmla v13.4s, v0.8h, v7.8h\n"
+      ".inst 0x6e45ec32  // bfmmla v18.4s, v1.8h, v5.8h\n"
+      ".inst 0x6e46ec30  // bfmmla v16.4s, v1.8h, v6.8h\n"
+      ".inst 0x6e47ec33  // bfmmla v19.4s, v1.8h, v7.8h\n"
+      ".inst 0x6e45ec58  // bfmmla v24.4s, v2.8h, v5.8h\n"
+      ".inst 0x6e45ec7e  // bfmmla v30.4s, v3.8h, v5.8h\n"
+      ".inst 0x6e46ec56  // bfmmla v22.4s, v2.8h, v6.8h\n"
+      ".inst 0x6e47ec59  // bfmmla v25.4s, v2.8h, v7.8h\n"
+      ".inst 0x6e46ec7c  // bfmmla v28.4s, v3.8h, v6.8h\n"
+      ".inst 0x6e47ec7f  // bfmmla v31.4s, v3.8h, v7.8h\n"
+      "5:"  // multiply loop done
+      "subs x22, x22, #0x1\n"
+      "uzp1 v4.2d, v8.2d, v11.2d\n"
+      "uzp2 v8.2d, v8.2d, v11.2d\n"
+      "uzp1 v11.2d, v9.2d, v12.2d\n"
+      "uzp2 v9.2d, v9.2d, v12.2d\n"
+      "str q4, [%x[Cpanel], #0x0]\n"
+      "uzp1 v12.2d, v10.2d, v13.2d\n"
+      "uzp2 v10.2d, v10.2d, v13.2d\n"
+      "str q11, [%x[Cpanel], #0x10]\n"
+      "str q12, [%x[Cpanel], #0x20]\n"
+      "uzp1 v13.2d, v14.2d, v17.2d\n"
+      "uzp2 v14.2d, v14.2d, v17.2d\n"
+      "str q8, [%x[Cpanel], #0x30]\n"
+      "uzp1 v17.2d, v15.2d, v18.2d\n"
+      "uzp2 v15.2d, v15.2d, v18.2d\n"
+      "str q9, [%x[Cpanel], #0x40]\n"
+      "uzp1 v18.2d, v16.2d, v19.2d\n"
+      "uzp2 v16.2d, v16.2d, v19.2d\n"
+      "str q10, [%x[Cpanel], #0x50]\n"
+      "uzp1 v19.2d, v20.2d, v23.2d\n"
+      "uzp2 v20.2d, v20.2d, v23.2d\n"
+      "str q13, [%x[Cpanel], #0x60]\n"
+      "uzp1 v23.2d, v21.2d, v24.2d\n"
+      "uzp2 v21.2d, v21.2d, v24.2d\n"
+      "str q17, [%x[Cpanel], #0x70]\n"
+      "uzp1 v24.2d, v22.2d, v25.2d\n"
+      "uzp2 v22.2d, v22.2d, v25.2d\n"
+      "str q18, [%x[Cpanel], #0x80]\n"
+      "uzp1 v25.2d, v26.2d, v29.2d\n"
+      "uzp2 v26.2d, v26.2d, v29.2d\n"
+      "str q14, [%x[Cpanel], #0x90]\n"
+      "uzp1 v29.2d, v27.2d, v30.2d\n"
+      "uzp2 v27.2d, v27.2d, v30.2d\n"
+      "str q15, [%x[Cpanel], #0xa0]\n"
+      "uzp1 v30.2d, v28.2d, v31.2d\n"
+      "uzp2 v28.2d, v28.2d, v31.2d\n"
+      "str q16, [%x[Cpanel], #0xb0]\n"
+      "str q19, [%x[Cpanel], #0xc0]\n"
+      "str q23, [%x[Cpanel], #0xd0]\n"
+      "str q24, [%x[Cpanel], #0xe0]\n"
+      "str q20, [%x[Cpanel], #0xf0]\n"
+      "str q21, [%x[Cpanel], #0x100]\n"
+      "str q22, [%x[Cpanel], #0x110]\n"
+      "str q25, [%x[Cpanel], #0x120]\n"
+      "str q29, [%x[Cpanel], #0x130]\n"
+      "str q30, [%x[Cpanel], #0x140]\n"
+      "str q26, [%x[Cpanel], #0x150]\n"
+      "str q27, [%x[Cpanel], #0x160]\n"
+      "str q28, [%x[Cpanel], #0x170]\n"
+      "add %x[Cpanel], %x[Cpanel], #0x180\n"
+      "bgt 2b\n"
+      "subs %x[ablocks], %x[ablocks], #0x1\n"
+      "bne 1b\n"
+      : [Apanel] "+&r" (Apanel), [Cpanel] "+&r" (Cpanel), [ablocks] "+&r" (ablocks)
+      : [args_ptr] "r" (&ka), [offsetof_Bpanel] "I" (offsetof(KernelArgs, Bpanel)), [offsetof_K] "I" (offsetof(KernelArgs, K)), [offsetof_bblocks] "I" (offsetof(KernelArgs, bblocks))
+      : "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", "x19", "x20", "x21", "x22"
+    );
 }
 
 } // namespace arm_gemm
-
 #endif // __aarch64__
-- 
cgit v1.2.1