aboutsummaryrefslogtreecommitdiff
path: root/src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_s8_12x8/generic.cpp
diff options
context:
space:
mode:
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_s8_12x8/generic.cpp')
-rw-r--r--src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_s8_12x8/generic.cpp254
1 files changed, 121 insertions, 133 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_s8_12x8/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_s8_12x8/generic.cpp
index 19225ddb34..a4e23a2bed 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_s8_12x8/generic.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_s8_12x8/generic.cpp
@@ -1,5 +1,5 @@
/*
- * Copyright (c) 2017-2018 ARM Limited.
+ * Copyright (c) 2017-2019 ARM Limited.
*
* SPDX-License-Identifier: MIT
*
@@ -27,10 +27,6 @@
#include "../../asmlib.hpp"
-#ifdef NO_DOT_IN_TOOLCHAIN
-#include "dot_toolchain_support.h"
-#endif
-
namespace arm_gemm {
void a64_gemm_s8_12x8(const int8_t *Apanel, const int8_t *Bpanel, int32_t *Cpanel, int ablocks, int bblocks, int K) {
@@ -56,11 +52,6 @@ void a64_gemm_s8_12x8(const int8_t *Apanel, const int8_t *Bpanel, int32_t *Cpane
register int32x4_t a0a asm("v5");
register int32x4_t a1a asm("v6");
__asm __volatile (
-#ifdef NO_DOT_IN_TOOLCHAIN
- _DECLARE_SDOT
-#else
- ".arch armv8.2-a+dotprod\n"
-#endif
// Initialize result registers, load initial operands, prime prefetches.
"movi v8.4s, #0x0\n"
"ldr %q[a0], [%[a_ptr]]\n"
@@ -106,76 +97,76 @@ void a64_gemm_s8_12x8(const int8_t *Apanel, const int8_t *Bpanel, int32_t *Cpane
// Loop proper
"1:\n"
- "sdot v8.4s , %[b0].16b, %[a0].4b[0]\n"
- "sdot v9.4s , %[b0].16b, %[a0].4b[1]\n"
+ ".word 0x4f80e048 // sdot v8.4s , %[b0].16b, %[a0].4b[0]\n"
+ ".word 0x4fa0e049 // sdot v9.4s , %[b0].16b, %[a0].4b[1]\n"
"ldr %q[b2], [%[b_ptr], #32]\n"
- "sdot v10.4s, %[b0].16b, %[a0].4b[2]\n"
- "sdot v11.4s, %[b0].16b, %[a0].4b[3]\n"
+ ".word 0x4f80e84a // sdot v10.4s, %[b0].16b, %[a0].4b[2]\n"
+ ".word 0x4fa0e84b // sdot v11.4s, %[b0].16b, %[a0].4b[3]\n"
"ldr %q[a0a], [%[a_ptr], #32]\n"
- "sdot v12.4s, %[b0].16b, %[a1].4b[0]\n"
- "sdot v13.4s, %[b0].16b, %[a1].4b[1]\n"
+ ".word 0x4f81e04c // sdot v12.4s, %[b0].16b, %[a1].4b[0]\n"
+ ".word 0x4fa1e04d // sdot v13.4s, %[b0].16b, %[a1].4b[1]\n"
"ldr %q[a1a], [%[a_ptr], #48]\n"
- "sdot v14.4s, %[b0].16b, %[a1].4b[2]\n"
- "sdot v15.4s, %[b0].16b, %[a1].4b[3]\n"
+ ".word 0x4f81e84e // sdot v14.4s, %[b0].16b, %[a1].4b[2]\n"
+ ".word 0x4fa1e84f // sdot v15.4s, %[b0].16b, %[a1].4b[3]\n"
"ldr %q[b0], [%[b_ptr], #48]\n"
- "sdot v16.4s, %[b1].16b, %[a0].4b[0]\n"
- "sdot v17.4s, %[b1].16b, %[a0].4b[1]\n"
+ ".word 0x4f80e070 // sdot v16.4s, %[b1].16b, %[a0].4b[0]\n"
+ ".word 0x4fa0e071 // sdot v17.4s, %[b1].16b, %[a0].4b[1]\n"
ASM_PREFETCH("[%[a_ptr], #320]")
- "sdot v18.4s, %[b1].16b, %[a0].4b[2]\n"
- "sdot v19.4s, %[b1].16b, %[a0].4b[3]\n"
- "sdot v20.4s, %[b1].16b, %[a1].4b[0]\n"
- "sdot v21.4s, %[b1].16b, %[a1].4b[1]\n"
- "sdot v22.4s, %[b1].16b, %[a1].4b[2]\n"
- "sdot v23.4s, %[b1].16b, %[a1].4b[3]\n"
+ ".word 0x4f80e872 // sdot v18.4s, %[b1].16b, %[a0].4b[2]\n"
+ ".word 0x4fa0e873 // sdot v19.4s, %[b1].16b, %[a0].4b[3]\n"
+ ".word 0x4f81e074 // sdot v20.4s, %[b1].16b, %[a1].4b[0]\n"
+ ".word 0x4fa1e075 // sdot v21.4s, %[b1].16b, %[a1].4b[1]\n"
+ ".word 0x4f81e876 // sdot v22.4s, %[b1].16b, %[a1].4b[2]\n"
+ ".word 0x4fa1e877 // sdot v23.4s, %[b1].16b, %[a1].4b[3]\n"
"ldr %q[b1], [%[b_ptr], #64]\n"
- "sdot v24.4s, %[b2].16b, %[a0].4b[0]\n"
- "sdot v25.4s, %[b2].16b, %[a0].4b[1]\n"
+ ".word 0x4f80e098 // sdot v24.4s, %[b2].16b, %[a0].4b[0]\n"
+ ".word 0x4fa0e099 // sdot v25.4s, %[b2].16b, %[a0].4b[1]\n"
ASM_PREFETCH("[%[b_ptr], #448]")
- "sdot v26.4s, %[b2].16b, %[a0].4b[2]\n"
- "sdot v27.4s, %[b2].16b, %[a0].4b[3]\n"
- "sdot v28.4s, %[b2].16b, %[a1].4b[0]\n"
- "sdot v29.4s, %[b2].16b, %[a1].4b[1]\n"
- "sdot v30.4s, %[b2].16b, %[a1].4b[2]\n"
- "sdot v31.4s, %[b2].16b, %[a1].4b[3]\n"
+ ".word 0x4f80e89a // sdot v26.4s, %[b2].16b, %[a0].4b[2]\n"
+ ".word 0x4fa0e89b // sdot v27.4s, %[b2].16b, %[a0].4b[3]\n"
+ ".word 0x4f81e09c // sdot v28.4s, %[b2].16b, %[a1].4b[0]\n"
+ ".word 0x4fa1e09d // sdot v29.4s, %[b2].16b, %[a1].4b[1]\n"
+ ".word 0x4f81e89e // sdot v30.4s, %[b2].16b, %[a1].4b[2]\n"
+ ".word 0x4fa1e89f // sdot v31.4s, %[b2].16b, %[a1].4b[3]\n"
"ldr %q[b2], [%[b_ptr], #80]\n"
- "sdot v8.4s , %[b0].16b, %[a0a].4b[0]\n"
- "sdot v9.4s , %[b0].16b, %[a0a].4b[1]\n"
+ ".word 0x4f85e048 // sdot v8.4s , %[b0].16b, %[a0a].4b[0]\n"
+ ".word 0x4fa5e049 // sdot v9.4s , %[b0].16b, %[a0a].4b[1]\n"
"ldr %q[a0], [%[a_ptr], #64]\n"
- "sdot v10.4s, %[b0].16b, %[a0a].4b[2]\n"
- "sdot v11.4s, %[b0].16b, %[a0a].4b[3]\n"
- "sdot v12.4s, %[b0].16b, %[a1a].4b[0]\n"
+ ".word 0x4f85e84a // sdot v10.4s, %[b0].16b, %[a0a].4b[2]\n"
+ ".word 0x4fa5e84b // sdot v11.4s, %[b0].16b, %[a0a].4b[3]\n"
+ ".word 0x4f86e04c // sdot v12.4s, %[b0].16b, %[a1a].4b[0]\n"
"ldr %q[a1], [%[a_ptr], #80]\n"
- "sdot v13.4s, %[b0].16b, %[a1a].4b[1]\n"
- "sdot v14.4s, %[b0].16b, %[a1a].4b[2]\n"
- "sdot v15.4s, %[b0].16b, %[a1a].4b[3]\n"
+ ".word 0x4fa6e04d // sdot v13.4s, %[b0].16b, %[a1a].4b[1]\n"
+ ".word 0x4f86e84e // sdot v14.4s, %[b0].16b, %[a1a].4b[2]\n"
+ ".word 0x4fa6e84f // sdot v15.4s, %[b0].16b, %[a1a].4b[3]\n"
"ldr %q[b0], [%[b_ptr], #96]\n"
- "sdot v16.4s, %[b1].16b, %[a0a].4b[0]\n"
- "sdot v17.4s, %[b1].16b, %[a0a].4b[1]\n"
+ ".word 0x4f85e070 // sdot v16.4s, %[b1].16b, %[a0a].4b[0]\n"
+ ".word 0x4fa5e071 // sdot v17.4s, %[b1].16b, %[a0a].4b[1]\n"
ASM_PREFETCH("[%[b_ptr], #512]")
- "sdot v18.4s, %[b1].16b, %[a0a].4b[2]\n"
- "sdot v19.4s, %[b1].16b, %[a0a].4b[3]\n"
- "sdot v20.4s, %[b1].16b, %[a1a].4b[0]\n"
- "sdot v21.4s, %[b1].16b, %[a1a].4b[1]\n"
- "sdot v22.4s, %[b1].16b, %[a1a].4b[2]\n"
- "sdot v23.4s, %[b1].16b, %[a1a].4b[3]\n"
+ ".word 0x4f85e872 // sdot v18.4s, %[b1].16b, %[a0a].4b[2]\n"
+ ".word 0x4fa5e873 // sdot v19.4s, %[b1].16b, %[a0a].4b[3]\n"
+ ".word 0x4f86e074 // sdot v20.4s, %[b1].16b, %[a1a].4b[0]\n"
+ ".word 0x4fa6e075 // sdot v21.4s, %[b1].16b, %[a1a].4b[1]\n"
+ ".word 0x4f86e876 // sdot v22.4s, %[b1].16b, %[a1a].4b[2]\n"
+ ".word 0x4fa6e877 // sdot v23.4s, %[b1].16b, %[a1a].4b[3]\n"
"ldr %q[b1], [%[b_ptr], #112]\n"
- "sdot v24.4s, %[b2].16b, %[a0a].4b[0]\n"
- "sdot v25.4s, %[b2].16b, %[a0a].4b[1]\n"
+ ".word 0x4f85e098 // sdot v24.4s, %[b2].16b, %[a0a].4b[0]\n"
+ ".word 0x4fa5e099 // sdot v25.4s, %[b2].16b, %[a0a].4b[1]\n"
"add %[a_ptr], %[a_ptr], #64\n"
- "sdot v26.4s, %[b2].16b, %[a0a].4b[2]\n"
- "sdot v27.4s, %[b2].16b, %[a0a].4b[3]\n"
+ ".word 0x4f85e89a // sdot v26.4s, %[b2].16b, %[a0a].4b[2]\n"
+ ".word 0x4fa5e89b // sdot v27.4s, %[b2].16b, %[a0a].4b[3]\n"
"add %[b_ptr], %[b_ptr], #96\n"
- "sdot v28.4s, %[b2].16b, %[a1a].4b[0]\n"
- "sdot v29.4s, %[b2].16b, %[a1a].4b[1]\n"
+ ".word 0x4f86e09c // sdot v28.4s, %[b2].16b, %[a1a].4b[0]\n"
+ ".word 0x4fa6e09d // sdot v29.4s, %[b2].16b, %[a1a].4b[1]\n"
"subs %w[k], %w[k], #1\n"
- "sdot v30.4s, %[b2].16b, %[a1a].4b[2]\n"
- "sdot v31.4s, %[b2].16b, %[a1a].4b[3]\n"
+ ".word 0x4f86e89e // sdot v30.4s, %[b2].16b, %[a1a].4b[2]\n"
+ ".word 0x4fa6e89f // sdot v31.4s, %[b2].16b, %[a1a].4b[3]\n"
"bne 1b\n"
// Target to use when K is 1 or 2 (i.e. zero iterations of main loop)
@@ -185,153 +176,153 @@ void a64_gemm_s8_12x8(const int8_t *Apanel, const int8_t *Bpanel, int32_t *Cpane
"cbnz %w[oddk], 2f\n"
// Detached final iteration (even K)
- "sdot v8.4s , %[b0].16b, %[a0].4b[0]\n"
- "sdot v9.4s , %[b0].16b, %[a0].4b[1]\n"
+ ".word 0x4f80e048 // sdot v8.4s , %[b0].16b, %[a0].4b[0]\n"
+ ".word 0x4fa0e049 // sdot v9.4s , %[b0].16b, %[a0].4b[1]\n"
"ldr %q[b2], [%[b_ptr], #32]\n"
- "sdot v10.4s, %[b0].16b, %[a0].4b[2]\n"
- "sdot v11.4s, %[b0].16b, %[a0].4b[3]\n"
+ ".word 0x4f80e84a // sdot v10.4s, %[b0].16b, %[a0].4b[2]\n"
+ ".word 0x4fa0e84b // sdot v11.4s, %[b0].16b, %[a0].4b[3]\n"
"ldr %q[a0a], [%[a_ptr], #32]\n"
- "sdot v12.4s, %[b0].16b, %[a1].4b[0]\n"
- "sdot v13.4s, %[b0].16b, %[a1].4b[1]\n"
+ ".word 0x4f81e04c // sdot v12.4s, %[b0].16b, %[a1].4b[0]\n"
+ ".word 0x4fa1e04d // sdot v13.4s, %[b0].16b, %[a1].4b[1]\n"
"ldr %q[a1a], [%[a_ptr], #48]\n"
- "sdot v14.4s, %[b0].16b, %[a1].4b[2]\n"
- "sdot v15.4s, %[b0].16b, %[a1].4b[3]\n"
+ ".word 0x4f81e84e // sdot v14.4s, %[b0].16b, %[a1].4b[2]\n"
+ ".word 0x4fa1e84f // sdot v15.4s, %[b0].16b, %[a1].4b[3]\n"
"ldr %q[b0], [%[b_ptr], #48]\n"
- "sdot v16.4s, %[b1].16b, %[a0].4b[0]\n"
- "sdot v17.4s, %[b1].16b, %[a0].4b[1]\n"
- "sdot v18.4s, %[b1].16b, %[a0].4b[2]\n"
- "sdot v19.4s, %[b1].16b, %[a0].4b[3]\n"
- "sdot v20.4s, %[b1].16b, %[a1].4b[0]\n"
- "sdot v21.4s, %[b1].16b, %[a1].4b[1]\n"
- "sdot v22.4s, %[b1].16b, %[a1].4b[2]\n"
- "sdot v23.4s, %[b1].16b, %[a1].4b[3]\n"
+ ".word 0x4f80e070 // sdot v16.4s, %[b1].16b, %[a0].4b[0]\n"
+ ".word 0x4fa0e071 // sdot v17.4s, %[b1].16b, %[a0].4b[1]\n"
+ ".word 0x4f80e872 // sdot v18.4s, %[b1].16b, %[a0].4b[2]\n"
+ ".word 0x4fa0e873 // sdot v19.4s, %[b1].16b, %[a0].4b[3]\n"
+ ".word 0x4f81e074 // sdot v20.4s, %[b1].16b, %[a1].4b[0]\n"
+ ".word 0x4fa1e075 // sdot v21.4s, %[b1].16b, %[a1].4b[1]\n"
+ ".word 0x4f81e876 // sdot v22.4s, %[b1].16b, %[a1].4b[2]\n"
+ ".word 0x4fa1e877 // sdot v23.4s, %[b1].16b, %[a1].4b[3]\n"
"ldr %q[b1], [%[b_ptr], #64]\n"
- "sdot v24.4s, %[b2].16b, %[a0].4b[0]\n"
- "sdot v25.4s, %[b2].16b, %[a0].4b[1]\n"
+ ".word 0x4f80e098 // sdot v24.4s, %[b2].16b, %[a0].4b[0]\n"
+ ".word 0x4fa0e099 // sdot v25.4s, %[b2].16b, %[a0].4b[1]\n"
"add %[a_ptr], %[a_ptr], #64\n"
- "sdot v26.4s, %[b2].16b, %[a0].4b[2]\n"
- "sdot v27.4s, %[b2].16b, %[a0].4b[3]\n"
- "sdot v28.4s, %[b2].16b, %[a1].4b[0]\n"
- "sdot v29.4s, %[b2].16b, %[a1].4b[1]\n"
- "sdot v30.4s, %[b2].16b, %[a1].4b[2]\n"
- "sdot v31.4s, %[b2].16b, %[a1].4b[3]\n"
+ ".word 0x4f80e89a // sdot v26.4s, %[b2].16b, %[a0].4b[2]\n"
+ ".word 0x4fa0e89b // sdot v27.4s, %[b2].16b, %[a0].4b[3]\n"
+ ".word 0x4f81e09c // sdot v28.4s, %[b2].16b, %[a1].4b[0]\n"
+ ".word 0x4fa1e09d // sdot v29.4s, %[b2].16b, %[a1].4b[1]\n"
+ ".word 0x4f81e89e // sdot v30.4s, %[b2].16b, %[a1].4b[2]\n"
+ ".word 0x4fa1e89f // sdot v31.4s, %[b2].16b, %[a1].4b[3]\n"
"ldr %q[b2], [%[b_ptr], #80]\n"
- "sdot v8.4s , %[b0].16b, %[a0a].4b[0]\n"
+ ".word 0x4f85e048 // sdot v8.4s , %[b0].16b, %[a0a].4b[0]\n"
- "sdot v16.4s, %[b1].16b, %[a0a].4b[0]\n"
+ ".word 0x4f85e070 // sdot v16.4s, %[b1].16b, %[a0a].4b[0]\n"
"add %[b_ptr], %[b_ptr], #96\n"
- "sdot v9.4s , %[b0].16b, %[a0a].4b[1]\n"
+ ".word 0x4fa5e049 // sdot v9.4s , %[b0].16b, %[a0a].4b[1]\n"
"str q8, [%[c_ptr], #0]\n"
- "sdot v17.4s, %[b1].16b, %[a0a].4b[1]\n"
+ ".word 0x4fa5e071 // sdot v17.4s, %[b1].16b, %[a0a].4b[1]\n"
"str q16, [%[c_ptr], #16]\n"
- "sdot v24.4s, %[b2].16b, %[a0a].4b[0]\n"
+ ".word 0x4f85e098 // sdot v24.4s, %[b2].16b, %[a0a].4b[0]\n"
"str q24, [%[c_ptr], #32]\n"
- "sdot v25.4s, %[b2].16b, %[a0a].4b[1]\n"
+ ".word 0x4fa5e099 // sdot v25.4s, %[b2].16b, %[a0a].4b[1]\n"
"str q9, [%[c_ptr], #48]\n"
- "sdot v10.4s, %[b0].16b, %[a0a].4b[2]\n"
+ ".word 0x4f85e84a // sdot v10.4s, %[b0].16b, %[a0a].4b[2]\n"
"str q17, [%[c_ptr], #64]\n"
- "sdot v18.4s, %[b1].16b, %[a0a].4b[2]\n"
+ ".word 0x4f85e872 // sdot v18.4s, %[b1].16b, %[a0a].4b[2]\n"
"str q25, [%[c_ptr], #80]\n"
- "sdot v26.4s, %[b2].16b, %[a0a].4b[2]\n"
+ ".word 0x4f85e89a // sdot v26.4s, %[b2].16b, %[a0a].4b[2]\n"
"str q10, [%[c_ptr], #96]\n"
- "sdot v11.4s, %[b0].16b, %[a0a].4b[3]\n"
+ ".word 0x4fa5e84b // sdot v11.4s, %[b0].16b, %[a0a].4b[3]\n"
"str q18, [%[c_ptr], #112]\n"
- "sdot v19.4s, %[b1].16b, %[a0a].4b[3]\n"
+ ".word 0x4fa5e873 // sdot v19.4s, %[b1].16b, %[a0a].4b[3]\n"
"str q26, [%[c_ptr], #128]\n"
- "sdot v27.4s, %[b2].16b, %[a0a].4b[3]\n"
+ ".word 0x4fa5e89b // sdot v27.4s, %[b2].16b, %[a0a].4b[3]\n"
"str q11, [%[c_ptr], #144]\n"
- "sdot v12.4s, %[b0].16b, %[a1a].4b[0]\n"
+ ".word 0x4f86e04c // sdot v12.4s, %[b0].16b, %[a1a].4b[0]\n"
"str q19, [%[c_ptr], #160]\n"
- "sdot v20.4s, %[b1].16b, %[a1a].4b[0]\n"
+ ".word 0x4f86e074 // sdot v20.4s, %[b1].16b, %[a1a].4b[0]\n"
"str q27, [%[c_ptr], #176]\n"
- "sdot v28.4s, %[b2].16b, %[a1a].4b[0]\n"
+ ".word 0x4f86e09c // sdot v28.4s, %[b2].16b, %[a1a].4b[0]\n"
"str q12, [%[c_ptr], #192]\n"
- "sdot v13.4s, %[b0].16b, %[a1a].4b[1]\n"
+ ".word 0x4fa6e04d // sdot v13.4s, %[b0].16b, %[a1a].4b[1]\n"
"str q20, [%[c_ptr], #208]\n"
- "sdot v21.4s, %[b1].16b, %[a1a].4b[1]\n"
+ ".word 0x4fa6e075 // sdot v21.4s, %[b1].16b, %[a1a].4b[1]\n"
"str q28, [%[c_ptr], #224]\n"
- "sdot v29.4s, %[b2].16b, %[a1a].4b[1]\n"
+ ".word 0x4fa6e09d // sdot v29.4s, %[b2].16b, %[a1a].4b[1]\n"
"str q13, [%[c_ptr], #240]\n"
- "sdot v14.4s, %[b0].16b, %[a1a].4b[2]\n"
+ ".word 0x4f86e84e // sdot v14.4s, %[b0].16b, %[a1a].4b[2]\n"
"str q21, [%[c_ptr], #256]\n"
- "sdot v22.4s, %[b1].16b, %[a1a].4b[2]\n"
+ ".word 0x4f86e876 // sdot v22.4s, %[b1].16b, %[a1a].4b[2]\n"
"str q29, [%[c_ptr], #272]\n"
- "sdot v30.4s, %[b2].16b, %[a1a].4b[2]\n"
+ ".word 0x4f86e89e // sdot v30.4s, %[b2].16b, %[a1a].4b[2]\n"
"str q14, [%[c_ptr], #288]\n"
- "sdot v15.4s, %[b0].16b, %[a1a].4b[3]\n"
+ ".word 0x4fa6e84f // sdot v15.4s, %[b0].16b, %[a1a].4b[3]\n"
"str q22, [%[c_ptr], #304]\n"
- "sdot v23.4s, %[b1].16b, %[a1a].4b[3]\n"
+ ".word 0x4fa6e877 // sdot v23.4s, %[b1].16b, %[a1a].4b[3]\n"
"str q30, [%[c_ptr], #320]\n"
- "sdot v31.4s, %[b2].16b, %[a1a].4b[3]\n"
+ ".word 0x4fa6e89f // sdot v31.4s, %[b2].16b, %[a1a].4b[3]\n"
"str q15, [%[c_ptr], #336]\n"
"b 3f\n"
// Detached final iteration (odd K)
"2:\n"
- "sdot v8.4s , %[b0].16b, %[a0].4b[0]\n"
+ ".word 0x4f80e048 // sdot v8.4s , %[b0].16b, %[a0].4b[0]\n"
"ldr %q[b2], [%[b_ptr], #32]\n"
- "sdot v16.4s, %[b1].16b, %[a0].4b[0]\n"
- "sdot v9.4s , %[b0].16b, %[a0].4b[1]\n"
+ ".word 0x4f80e070 // sdot v16.4s, %[b1].16b, %[a0].4b[0]\n"
+ ".word 0x4fa0e049 // sdot v9.4s , %[b0].16b, %[a0].4b[1]\n"
"str q8, [%[c_ptr], #0]\n"
- "sdot v17.4s, %[b1].16b, %[a0].4b[1]\n"
+ ".word 0x4fa0e071 // sdot v17.4s, %[b1].16b, %[a0].4b[1]\n"
"str q16, [%[c_ptr], #16]\n"
- "sdot v24.4s, %[b2].16b, %[a0].4b[0]\n"
+ ".word 0x4f80e098 // sdot v24.4s, %[b2].16b, %[a0].4b[0]\n"
"add %[b_ptr], %[b_ptr], #48\n"
"add %[a_ptr], %[a_ptr], #32\n"
"str q24, [%[c_ptr], #32]\n"
- "sdot v25.4s, %[b2].16b, %[a0].4b[1]\n"
+ ".word 0x4fa0e099 // sdot v25.4s, %[b2].16b, %[a0].4b[1]\n"
"str q9, [%[c_ptr], #48]\n"
- "sdot v10.4s, %[b0].16b, %[a0].4b[2]\n"
+ ".word 0x4f80e84a // sdot v10.4s, %[b0].16b, %[a0].4b[2]\n"
"str q17, [%[c_ptr], #64]\n"
- "sdot v18.4s, %[b1].16b, %[a0].4b[2]\n"
+ ".word 0x4f80e872 // sdot v18.4s, %[b1].16b, %[a0].4b[2]\n"
"str q25, [%[c_ptr], #80]\n"
- "sdot v26.4s, %[b2].16b, %[a0].4b[2]\n"
+ ".word 0x4f80e89a // sdot v26.4s, %[b2].16b, %[a0].4b[2]\n"
"str q10, [%[c_ptr], #96]\n"
- "sdot v11.4s, %[b0].16b, %[a0].4b[3]\n"
+ ".word 0x4fa0e84b // sdot v11.4s, %[b0].16b, %[a0].4b[3]\n"
"str q18, [%[c_ptr], #112]\n"
- "sdot v19.4s, %[b1].16b, %[a0].4b[3]\n"
+ ".word 0x4fa0e873 // sdot v19.4s, %[b1].16b, %[a0].4b[3]\n"
"str q26, [%[c_ptr], #128]\n"
- "sdot v27.4s, %[b2].16b, %[a0].4b[3]\n"
+ ".word 0x4fa0e89b // sdot v27.4s, %[b2].16b, %[a0].4b[3]\n"
"str q11, [%[c_ptr], #144]\n"
- "sdot v12.4s, %[b0].16b, %[a1].4b[0]\n"
+ ".word 0x4f81e04c // sdot v12.4s, %[b0].16b, %[a1].4b[0]\n"
"str q19, [%[c_ptr], #160]\n"
- "sdot v20.4s, %[b1].16b, %[a1].4b[0]\n"
+ ".word 0x4f81e074 // sdot v20.4s, %[b1].16b, %[a1].4b[0]\n"
"str q27, [%[c_ptr], #176]\n"
- "sdot v28.4s, %[b2].16b, %[a1].4b[0]\n"
+ ".word 0x4f81e09c // sdot v28.4s, %[b2].16b, %[a1].4b[0]\n"
"str q12, [%[c_ptr], #192]\n"
- "sdot v13.4s, %[b0].16b, %[a1].4b[1]\n"
+ ".word 0x4fa1e04d // sdot v13.4s, %[b0].16b, %[a1].4b[1]\n"
"str q20, [%[c_ptr], #208]\n"
- "sdot v21.4s, %[b1].16b, %[a1].4b[1]\n"
+ ".word 0x4fa1e075 // sdot v21.4s, %[b1].16b, %[a1].4b[1]\n"
"str q28, [%[c_ptr], #224]\n"
- "sdot v29.4s, %[b2].16b, %[a1].4b[1]\n"
+ ".word 0x4fa1e09d // sdot v29.4s, %[b2].16b, %[a1].4b[1]\n"
"str q13, [%[c_ptr], #240]\n"
- "sdot v14.4s, %[b0].16b, %[a1].4b[2]\n"
+ ".word 0x4f81e84e // sdot v14.4s, %[b0].16b, %[a1].4b[2]\n"
"str q21, [%[c_ptr], #256]\n"
- "sdot v22.4s, %[b1].16b, %[a1].4b[2]\n"
+ ".word 0x4f81e876 // sdot v22.4s, %[b1].16b, %[a1].4b[2]\n"
"str q29, [%[c_ptr], #272]\n"
- "sdot v30.4s, %[b2].16b, %[a1].4b[2]\n"
+ ".word 0x4f81e89e // sdot v30.4s, %[b2].16b, %[a1].4b[2]\n"
"str q14, [%[c_ptr], #288]\n"
- "sdot v15.4s, %[b0].16b, %[a1].4b[3]\n"
+ ".word 0x4fa1e84f // sdot v15.4s, %[b0].16b, %[a1].4b[3]\n"
"str q22, [%[c_ptr], #304]\n"
- "sdot v23.4s, %[b1].16b, %[a1].4b[3]\n"
+ ".word 0x4fa1e877 // sdot v23.4s, %[b1].16b, %[a1].4b[3]\n"
"str q30, [%[c_ptr], #320]\n"
- "sdot v31.4s, %[b2].16b, %[a1].4b[3]\n"
+ ".word 0x4fa1e89f // sdot v31.4s, %[b2].16b, %[a1].4b[3]\n"
"str q15, [%[c_ptr], #336]\n"
@@ -341,9 +332,6 @@ void a64_gemm_s8_12x8(const int8_t *Apanel, const int8_t *Bpanel, int32_t *Cpane
"str q31, [%[c_ptr], #368]\n"
"add %[c_ptr], %[c_ptr], #384\n"
-#ifdef NO_DOT_IN_TOOLCHAIN
- ".purgem sdot\n"
-#endif
:
[a_ptr] "+r" (a_ptr), [b_ptr] "+r" (b_ptr), [c_ptr] "+r" (c_ptr),
[a0] "+w" (a0), [a1] "+w" (a1), [a0a] "+w" (a0a), [a1a] "+w" (a1a),