aboutsummaryrefslogtreecommitdiff
path: root/src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_u8_12x8/generic.cpp
diff options
context:
space:
mode:
authorGeorgios Pinitas <georgios.pinitas@arm.com>2019-09-25 17:52:07 +0100
committerGeorgios Pinitas <georgios.pinitas@arm.com>2019-09-26 16:04:22 +0000
commitbc88c621bd89cb93b503e2d8386383ff9ae6412a (patch)
tree713da19d959791fe69d8d9aa3a5a19af561ce030 /src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_u8_12x8/generic.cpp
parentdc200dc1328b34c85554ec7db3b48bcc6410d18f (diff)
downloadComputeLibrary-bc88c621bd89cb93b503e2d8386383ff9ae6412a.tar.gz
COMPMID-2452: Replace DOT assemble macro with the corresponding word directives.
Change-Id: Id602cfe2d00deed6d994ba4c90cdc5914a8e6016 Signed-off-by: Georgios Pinitas <georgios.pinitas@arm.com> Reviewed-on: https://review.mlplatform.org/c/1987 Comments-Addressed: Arm Jenkins <bsgcomp@arm.com> Reviewed-by: Pablo Marquez <pablo.tello@arm.com>
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_u8_12x8/generic.cpp')
-rw-r--r--src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_u8_12x8/generic.cpp254
1 files changed, 121 insertions, 133 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_u8_12x8/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_u8_12x8/generic.cpp
index 80dd87356d..eda6386fef 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_u8_12x8/generic.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_u8_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_u8_12x8(const uint8_t *Apanel, const uint8_t *Bpanel, uint32_t *Cpanel, int ablocks, int bblocks, int K) {
@@ -56,11 +52,6 @@ void a64_gemm_u8_12x8(const uint8_t *Apanel, const uint8_t *Bpanel, uint32_t *Cp
register uint8x16_t a0a asm("v5");
register uint8x16_t a1a asm("v6");
__asm __volatile (
-#ifdef NO_DOT_IN_TOOLCHAIN
- _DECLARE_UDOT
-#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_u8_12x8(const uint8_t *Apanel, const uint8_t *Bpanel, uint32_t *Cp
// Loop proper
"1:\n"
- "udot v8.4s , %[b0].16b, %[a0].4b[0]\n"
- "udot v9.4s , %[b0].16b, %[a0].4b[1]\n"
+ ".word 0x6f80e048 // udot v8.4s , %[b0].16b, %[a0].4b[0]\n"
+ ".word 0x6fa0e049 // udot v9.4s , %[b0].16b, %[a0].4b[1]\n"
"ldr %q[b2], [%[b_ptr], #32]\n"
- "udot v10.4s, %[b0].16b, %[a0].4b[2]\n"
- "udot v11.4s, %[b0].16b, %[a0].4b[3]\n"
+ ".word 0x6f80e84a // udot v10.4s, %[b0].16b, %[a0].4b[2]\n"
+ ".word 0x6fa0e84b // udot v11.4s, %[b0].16b, %[a0].4b[3]\n"
"ldr %q[a0a], [%[a_ptr], #32]\n"
- "udot v12.4s, %[b0].16b, %[a1].4b[0]\n"
- "udot v13.4s, %[b0].16b, %[a1].4b[1]\n"
+ ".word 0x6f81e04c // udot v12.4s, %[b0].16b, %[a1].4b[0]\n"
+ ".word 0x6fa1e04d // udot v13.4s, %[b0].16b, %[a1].4b[1]\n"
"ldr %q[a1a], [%[a_ptr], #48]\n"
- "udot v14.4s, %[b0].16b, %[a1].4b[2]\n"
- "udot v15.4s, %[b0].16b, %[a1].4b[3]\n"
+ ".word 0x6f81e84e // udot v14.4s, %[b0].16b, %[a1].4b[2]\n"
+ ".word 0x6fa1e84f // udot v15.4s, %[b0].16b, %[a1].4b[3]\n"
"ldr %q[b0], [%[b_ptr], #48]\n"
- "udot v16.4s, %[b1].16b, %[a0].4b[0]\n"
- "udot v17.4s, %[b1].16b, %[a0].4b[1]\n"
+ ".word 0x6f80e070 // udot v16.4s, %[b1].16b, %[a0].4b[0]\n"
+ ".word 0x6fa0e071 // udot v17.4s, %[b1].16b, %[a0].4b[1]\n"
ASM_PREFETCH("[%[a_ptr], #320]")
- "udot v18.4s, %[b1].16b, %[a0].4b[2]\n"
- "udot v19.4s, %[b1].16b, %[a0].4b[3]\n"
- "udot v20.4s, %[b1].16b, %[a1].4b[0]\n"
- "udot v21.4s, %[b1].16b, %[a1].4b[1]\n"
- "udot v22.4s, %[b1].16b, %[a1].4b[2]\n"
- "udot v23.4s, %[b1].16b, %[a1].4b[3]\n"
+ ".word 0x6f80e872 // udot v18.4s, %[b1].16b, %[a0].4b[2]\n"
+ ".word 0x6fa0e873 // udot v19.4s, %[b1].16b, %[a0].4b[3]\n"
+ ".word 0x6f81e074 // udot v20.4s, %[b1].16b, %[a1].4b[0]\n"
+ ".word 0x6fa1e075 // udot v21.4s, %[b1].16b, %[a1].4b[1]\n"
+ ".word 0x6f81e876 // udot v22.4s, %[b1].16b, %[a1].4b[2]\n"
+ ".word 0x6fa1e877 // udot v23.4s, %[b1].16b, %[a1].4b[3]\n"
"ldr %q[b1], [%[b_ptr], #64]\n"
- "udot v24.4s, %[b2].16b, %[a0].4b[0]\n"
- "udot v25.4s, %[b2].16b, %[a0].4b[1]\n"
+ ".word 0x6f80e098 // udot v24.4s, %[b2].16b, %[a0].4b[0]\n"
+ ".word 0x6fa0e099 // udot v25.4s, %[b2].16b, %[a0].4b[1]\n"
ASM_PREFETCH("[%[b_ptr], #448]")
- "udot v26.4s, %[b2].16b, %[a0].4b[2]\n"
- "udot v27.4s, %[b2].16b, %[a0].4b[3]\n"
- "udot v28.4s, %[b2].16b, %[a1].4b[0]\n"
- "udot v29.4s, %[b2].16b, %[a1].4b[1]\n"
- "udot v30.4s, %[b2].16b, %[a1].4b[2]\n"
- "udot v31.4s, %[b2].16b, %[a1].4b[3]\n"
+ ".word 0x6f80e89a // udot v26.4s, %[b2].16b, %[a0].4b[2]\n"
+ ".word 0x6fa0e89b // udot v27.4s, %[b2].16b, %[a0].4b[3]\n"
+ ".word 0x6f81e09c // udot v28.4s, %[b2].16b, %[a1].4b[0]\n"
+ ".word 0x6fa1e09d // udot v29.4s, %[b2].16b, %[a1].4b[1]\n"
+ ".word 0x6f81e89e // udot v30.4s, %[b2].16b, %[a1].4b[2]\n"
+ ".word 0x6fa1e89f // udot v31.4s, %[b2].16b, %[a1].4b[3]\n"
"ldr %q[b2], [%[b_ptr], #80]\n"
- "udot v8.4s , %[b0].16b, %[a0a].4b[0]\n"
- "udot v9.4s , %[b0].16b, %[a0a].4b[1]\n"
+ ".word 0x6f85e048 // udot v8.4s , %[b0].16b, %[a0a].4b[0]\n"
+ ".word 0x6fa5e049 // udot v9.4s , %[b0].16b, %[a0a].4b[1]\n"
"ldr %q[a0], [%[a_ptr], #64]\n"
- "udot v10.4s, %[b0].16b, %[a0a].4b[2]\n"
- "udot v11.4s, %[b0].16b, %[a0a].4b[3]\n"
- "udot v12.4s, %[b0].16b, %[a1a].4b[0]\n"
+ ".word 0x6f85e84a // udot v10.4s, %[b0].16b, %[a0a].4b[2]\n"
+ ".word 0x6fa5e84b // udot v11.4s, %[b0].16b, %[a0a].4b[3]\n"
+ ".word 0x6f86e04c // udot v12.4s, %[b0].16b, %[a1a].4b[0]\n"
"ldr %q[a1], [%[a_ptr], #80]\n"
- "udot v13.4s, %[b0].16b, %[a1a].4b[1]\n"
- "udot v14.4s, %[b0].16b, %[a1a].4b[2]\n"
- "udot v15.4s, %[b0].16b, %[a1a].4b[3]\n"
+ ".word 0x6fa6e04d // udot v13.4s, %[b0].16b, %[a1a].4b[1]\n"
+ ".word 0x6f86e84e // udot v14.4s, %[b0].16b, %[a1a].4b[2]\n"
+ ".word 0x6fa6e84f // udot v15.4s, %[b0].16b, %[a1a].4b[3]\n"
"ldr %q[b0], [%[b_ptr], #96]\n"
- "udot v16.4s, %[b1].16b, %[a0a].4b[0]\n"
- "udot v17.4s, %[b1].16b, %[a0a].4b[1]\n"
+ ".word 0x6f85e070 // udot v16.4s, %[b1].16b, %[a0a].4b[0]\n"
+ ".word 0x6fa5e071 // udot v17.4s, %[b1].16b, %[a0a].4b[1]\n"
ASM_PREFETCH("[%[b_ptr], #512]")
- "udot v18.4s, %[b1].16b, %[a0a].4b[2]\n"
- "udot v19.4s, %[b1].16b, %[a0a].4b[3]\n"
- "udot v20.4s, %[b1].16b, %[a1a].4b[0]\n"
- "udot v21.4s, %[b1].16b, %[a1a].4b[1]\n"
- "udot v22.4s, %[b1].16b, %[a1a].4b[2]\n"
- "udot v23.4s, %[b1].16b, %[a1a].4b[3]\n"
+ ".word 0x6f85e872 // udot v18.4s, %[b1].16b, %[a0a].4b[2]\n"
+ ".word 0x6fa5e873 // udot v19.4s, %[b1].16b, %[a0a].4b[3]\n"
+ ".word 0x6f86e074 // udot v20.4s, %[b1].16b, %[a1a].4b[0]\n"
+ ".word 0x6fa6e075 // udot v21.4s, %[b1].16b, %[a1a].4b[1]\n"
+ ".word 0x6f86e876 // udot v22.4s, %[b1].16b, %[a1a].4b[2]\n"
+ ".word 0x6fa6e877 // udot v23.4s, %[b1].16b, %[a1a].4b[3]\n"
"ldr %q[b1], [%[b_ptr], #112]\n"
- "udot v24.4s, %[b2].16b, %[a0a].4b[0]\n"
- "udot v25.4s, %[b2].16b, %[a0a].4b[1]\n"
+ ".word 0x6f85e098 // udot v24.4s, %[b2].16b, %[a0a].4b[0]\n"
+ ".word 0x6fa5e099 // udot v25.4s, %[b2].16b, %[a0a].4b[1]\n"
"add %[a_ptr], %[a_ptr], #64\n"
- "udot v26.4s, %[b2].16b, %[a0a].4b[2]\n"
- "udot v27.4s, %[b2].16b, %[a0a].4b[3]\n"
+ ".word 0x6f85e89a // udot v26.4s, %[b2].16b, %[a0a].4b[2]\n"
+ ".word 0x6fa5e89b // udot v27.4s, %[b2].16b, %[a0a].4b[3]\n"
"add %[b_ptr], %[b_ptr], #96\n"
- "udot v28.4s, %[b2].16b, %[a1a].4b[0]\n"
- "udot v29.4s, %[b2].16b, %[a1a].4b[1]\n"
+ ".word 0x6f86e09c // udot v28.4s, %[b2].16b, %[a1a].4b[0]\n"
+ ".word 0x6fa6e09d // udot v29.4s, %[b2].16b, %[a1a].4b[1]\n"
"subs %w[k], %w[k], #1\n"
- "udot v30.4s, %[b2].16b, %[a1a].4b[2]\n"
- "udot v31.4s, %[b2].16b, %[a1a].4b[3]\n"
+ ".word 0x6f86e89e // udot v30.4s, %[b2].16b, %[a1a].4b[2]\n"
+ ".word 0x6fa6e89f // udot 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_u8_12x8(const uint8_t *Apanel, const uint8_t *Bpanel, uint32_t *Cp
"cbnz %w[oddk], 2f\n"
// Detached final iteration (even K)
- "udot v8.4s , %[b0].16b, %[a0].4b[0]\n"
- "udot v9.4s , %[b0].16b, %[a0].4b[1]\n"
+ ".word 0x6f80e048 // udot v8.4s , %[b0].16b, %[a0].4b[0]\n"
+ ".word 0x6fa0e049 // udot v9.4s , %[b0].16b, %[a0].4b[1]\n"
"ldr %q[b2], [%[b_ptr], #32]\n"
- "udot v10.4s, %[b0].16b, %[a0].4b[2]\n"
- "udot v11.4s, %[b0].16b, %[a0].4b[3]\n"
+ ".word 0x6f80e84a // udot v10.4s, %[b0].16b, %[a0].4b[2]\n"
+ ".word 0x6fa0e84b // udot v11.4s, %[b0].16b, %[a0].4b[3]\n"
"ldr %q[a0a], [%[a_ptr], #32]\n"
- "udot v12.4s, %[b0].16b, %[a1].4b[0]\n"
- "udot v13.4s, %[b0].16b, %[a1].4b[1]\n"
+ ".word 0x6f81e04c // udot v12.4s, %[b0].16b, %[a1].4b[0]\n"
+ ".word 0x6fa1e04d // udot v13.4s, %[b0].16b, %[a1].4b[1]\n"
"ldr %q[a1a], [%[a_ptr], #48]\n"
- "udot v14.4s, %[b0].16b, %[a1].4b[2]\n"
- "udot v15.4s, %[b0].16b, %[a1].4b[3]\n"
+ ".word 0x6f81e84e // udot v14.4s, %[b0].16b, %[a1].4b[2]\n"
+ ".word 0x6fa1e84f // udot v15.4s, %[b0].16b, %[a1].4b[3]\n"
"ldr %q[b0], [%[b_ptr], #48]\n"
- "udot v16.4s, %[b1].16b, %[a0].4b[0]\n"
- "udot v17.4s, %[b1].16b, %[a0].4b[1]\n"
- "udot v18.4s, %[b1].16b, %[a0].4b[2]\n"
- "udot v19.4s, %[b1].16b, %[a0].4b[3]\n"
- "udot v20.4s, %[b1].16b, %[a1].4b[0]\n"
- "udot v21.4s, %[b1].16b, %[a1].4b[1]\n"
- "udot v22.4s, %[b1].16b, %[a1].4b[2]\n"
- "udot v23.4s, %[b1].16b, %[a1].4b[3]\n"
+ ".word 0x6f80e070 // udot v16.4s, %[b1].16b, %[a0].4b[0]\n"
+ ".word 0x6fa0e071 // udot v17.4s, %[b1].16b, %[a0].4b[1]\n"
+ ".word 0x6f80e872 // udot v18.4s, %[b1].16b, %[a0].4b[2]\n"
+ ".word 0x6fa0e873 // udot v19.4s, %[b1].16b, %[a0].4b[3]\n"
+ ".word 0x6f81e074 // udot v20.4s, %[b1].16b, %[a1].4b[0]\n"
+ ".word 0x6fa1e075 // udot v21.4s, %[b1].16b, %[a1].4b[1]\n"
+ ".word 0x6f81e876 // udot v22.4s, %[b1].16b, %[a1].4b[2]\n"
+ ".word 0x6fa1e877 // udot v23.4s, %[b1].16b, %[a1].4b[3]\n"
"ldr %q[b1], [%[b_ptr], #64]\n"
- "udot v24.4s, %[b2].16b, %[a0].4b[0]\n"
- "udot v25.4s, %[b2].16b, %[a0].4b[1]\n"
+ ".word 0x6f80e098 // udot v24.4s, %[b2].16b, %[a0].4b[0]\n"
+ ".word 0x6fa0e099 // udot v25.4s, %[b2].16b, %[a0].4b[1]\n"
"add %[a_ptr], %[a_ptr], #64\n"
- "udot v26.4s, %[b2].16b, %[a0].4b[2]\n"
- "udot v27.4s, %[b2].16b, %[a0].4b[3]\n"
- "udot v28.4s, %[b2].16b, %[a1].4b[0]\n"
- "udot v29.4s, %[b2].16b, %[a1].4b[1]\n"
- "udot v30.4s, %[b2].16b, %[a1].4b[2]\n"
- "udot v31.4s, %[b2].16b, %[a1].4b[3]\n"
+ ".word 0x6f80e89a // udot v26.4s, %[b2].16b, %[a0].4b[2]\n"
+ ".word 0x6fa0e89b // udot v27.4s, %[b2].16b, %[a0].4b[3]\n"
+ ".word 0x6f81e09c // udot v28.4s, %[b2].16b, %[a1].4b[0]\n"
+ ".word 0x6fa1e09d // udot v29.4s, %[b2].16b, %[a1].4b[1]\n"
+ ".word 0x6f81e89e // udot v30.4s, %[b2].16b, %[a1].4b[2]\n"
+ ".word 0x6fa1e89f // udot v31.4s, %[b2].16b, %[a1].4b[3]\n"
"ldr %q[b2], [%[b_ptr], #80]\n"
- "udot v8.4s , %[b0].16b, %[a0a].4b[0]\n"
+ ".word 0x6f85e048 // udot v8.4s , %[b0].16b, %[a0a].4b[0]\n"
- "udot v16.4s, %[b1].16b, %[a0a].4b[0]\n"
+ ".word 0x6f85e070 // udot v16.4s, %[b1].16b, %[a0a].4b[0]\n"
"add %[b_ptr], %[b_ptr], #96\n"
- "udot v9.4s , %[b0].16b, %[a0a].4b[1]\n"
+ ".word 0x6fa5e049 // udot v9.4s , %[b0].16b, %[a0a].4b[1]\n"
"str q8, [%[c_ptr], #0]\n"
- "udot v17.4s, %[b1].16b, %[a0a].4b[1]\n"
+ ".word 0x6fa5e071 // udot v17.4s, %[b1].16b, %[a0a].4b[1]\n"
"str q16, [%[c_ptr], #16]\n"
- "udot v24.4s, %[b2].16b, %[a0a].4b[0]\n"
+ ".word 0x6f85e098 // udot v24.4s, %[b2].16b, %[a0a].4b[0]\n"
"str q24, [%[c_ptr], #32]\n"
- "udot v25.4s, %[b2].16b, %[a0a].4b[1]\n"
+ ".word 0x6fa5e099 // udot v25.4s, %[b2].16b, %[a0a].4b[1]\n"
"str q9, [%[c_ptr], #48]\n"
- "udot v10.4s, %[b0].16b, %[a0a].4b[2]\n"
+ ".word 0x6f85e84a // udot v10.4s, %[b0].16b, %[a0a].4b[2]\n"
"str q17, [%[c_ptr], #64]\n"
- "udot v18.4s, %[b1].16b, %[a0a].4b[2]\n"
+ ".word 0x6f85e872 // udot v18.4s, %[b1].16b, %[a0a].4b[2]\n"
"str q25, [%[c_ptr], #80]\n"
- "udot v26.4s, %[b2].16b, %[a0a].4b[2]\n"
+ ".word 0x6f85e89a // udot v26.4s, %[b2].16b, %[a0a].4b[2]\n"
"str q10, [%[c_ptr], #96]\n"
- "udot v11.4s, %[b0].16b, %[a0a].4b[3]\n"
+ ".word 0x6fa5e84b // udot v11.4s, %[b0].16b, %[a0a].4b[3]\n"
"str q18, [%[c_ptr], #112]\n"
- "udot v19.4s, %[b1].16b, %[a0a].4b[3]\n"
+ ".word 0x6fa5e873 // udot v19.4s, %[b1].16b, %[a0a].4b[3]\n"
"str q26, [%[c_ptr], #128]\n"
- "udot v27.4s, %[b2].16b, %[a0a].4b[3]\n"
+ ".word 0x6fa5e89b // udot v27.4s, %[b2].16b, %[a0a].4b[3]\n"
"str q11, [%[c_ptr], #144]\n"
- "udot v12.4s, %[b0].16b, %[a1a].4b[0]\n"
+ ".word 0x6f86e04c // udot v12.4s, %[b0].16b, %[a1a].4b[0]\n"
"str q19, [%[c_ptr], #160]\n"
- "udot v20.4s, %[b1].16b, %[a1a].4b[0]\n"
+ ".word 0x6f86e074 // udot v20.4s, %[b1].16b, %[a1a].4b[0]\n"
"str q27, [%[c_ptr], #176]\n"
- "udot v28.4s, %[b2].16b, %[a1a].4b[0]\n"
+ ".word 0x6f86e09c // udot v28.4s, %[b2].16b, %[a1a].4b[0]\n"
"str q12, [%[c_ptr], #192]\n"
- "udot v13.4s, %[b0].16b, %[a1a].4b[1]\n"
+ ".word 0x6fa6e04d // udot v13.4s, %[b0].16b, %[a1a].4b[1]\n"
"str q20, [%[c_ptr], #208]\n"
- "udot v21.4s, %[b1].16b, %[a1a].4b[1]\n"
+ ".word 0x6fa6e075 // udot v21.4s, %[b1].16b, %[a1a].4b[1]\n"
"str q28, [%[c_ptr], #224]\n"
- "udot v29.4s, %[b2].16b, %[a1a].4b[1]\n"
+ ".word 0x6fa6e09d // udot v29.4s, %[b2].16b, %[a1a].4b[1]\n"
"str q13, [%[c_ptr], #240]\n"
- "udot v14.4s, %[b0].16b, %[a1a].4b[2]\n"
+ ".word 0x6f86e84e // udot v14.4s, %[b0].16b, %[a1a].4b[2]\n"
"str q21, [%[c_ptr], #256]\n"
- "udot v22.4s, %[b1].16b, %[a1a].4b[2]\n"
+ ".word 0x6f86e876 // udot v22.4s, %[b1].16b, %[a1a].4b[2]\n"
"str q29, [%[c_ptr], #272]\n"
- "udot v30.4s, %[b2].16b, %[a1a].4b[2]\n"
+ ".word 0x6f86e89e // udot v30.4s, %[b2].16b, %[a1a].4b[2]\n"
"str q14, [%[c_ptr], #288]\n"
- "udot v15.4s, %[b0].16b, %[a1a].4b[3]\n"
+ ".word 0x6fa6e84f // udot v15.4s, %[b0].16b, %[a1a].4b[3]\n"
"str q22, [%[c_ptr], #304]\n"
- "udot v23.4s, %[b1].16b, %[a1a].4b[3]\n"
+ ".word 0x6fa6e877 // udot v23.4s, %[b1].16b, %[a1a].4b[3]\n"
"str q30, [%[c_ptr], #320]\n"
- "udot v31.4s, %[b2].16b, %[a1a].4b[3]\n"
+ ".word 0x6fa6e89f // udot v31.4s, %[b2].16b, %[a1a].4b[3]\n"
"str q15, [%[c_ptr], #336]\n"
"b 3f\n"
// Detached final iteration (odd K)
"2:\n"
- "udot v8.4s , %[b0].16b, %[a0].4b[0]\n"
+ ".word 0x6f80e048 // udot v8.4s , %[b0].16b, %[a0].4b[0]\n"
"ldr %q[b2], [%[b_ptr], #32]\n"
- "udot v16.4s, %[b1].16b, %[a0].4b[0]\n"
- "udot v9.4s , %[b0].16b, %[a0].4b[1]\n"
+ ".word 0x6f80e070 // udot v16.4s, %[b1].16b, %[a0].4b[0]\n"
+ ".word 0x6fa0e049 // udot v9.4s , %[b0].16b, %[a0].4b[1]\n"
"str q8, [%[c_ptr], #0]\n"
- "udot v17.4s, %[b1].16b, %[a0].4b[1]\n"
+ ".word 0x6fa0e071 // udot v17.4s, %[b1].16b, %[a0].4b[1]\n"
"str q16, [%[c_ptr], #16]\n"
- "udot v24.4s, %[b2].16b, %[a0].4b[0]\n"
+ ".word 0x6f80e098 // udot 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"
- "udot v25.4s, %[b2].16b, %[a0].4b[1]\n"
+ ".word 0x6fa0e099 // udot v25.4s, %[b2].16b, %[a0].4b[1]\n"
"str q9, [%[c_ptr], #48]\n"
- "udot v10.4s, %[b0].16b, %[a0].4b[2]\n"
+ ".word 0x6f80e84a // udot v10.4s, %[b0].16b, %[a0].4b[2]\n"
"str q17, [%[c_ptr], #64]\n"
- "udot v18.4s, %[b1].16b, %[a0].4b[2]\n"
+ ".word 0x6f80e872 // udot v18.4s, %[b1].16b, %[a0].4b[2]\n"
"str q25, [%[c_ptr], #80]\n"
- "udot v26.4s, %[b2].16b, %[a0].4b[2]\n"
+ ".word 0x6f80e89a // udot v26.4s, %[b2].16b, %[a0].4b[2]\n"
"str q10, [%[c_ptr], #96]\n"
- "udot v11.4s, %[b0].16b, %[a0].4b[3]\n"
+ ".word 0x6fa0e84b // udot v11.4s, %[b0].16b, %[a0].4b[3]\n"
"str q18, [%[c_ptr], #112]\n"
- "udot v19.4s, %[b1].16b, %[a0].4b[3]\n"
+ ".word 0x6fa0e873 // udot v19.4s, %[b1].16b, %[a0].4b[3]\n"
"str q26, [%[c_ptr], #128]\n"
- "udot v27.4s, %[b2].16b, %[a0].4b[3]\n"
+ ".word 0x6fa0e89b // udot v27.4s, %[b2].16b, %[a0].4b[3]\n"
"str q11, [%[c_ptr], #144]\n"
- "udot v12.4s, %[b0].16b, %[a1].4b[0]\n"
+ ".word 0x6f81e04c // udot v12.4s, %[b0].16b, %[a1].4b[0]\n"
"str q19, [%[c_ptr], #160]\n"
- "udot v20.4s, %[b1].16b, %[a1].4b[0]\n"
+ ".word 0x6f81e074 // udot v20.4s, %[b1].16b, %[a1].4b[0]\n"
"str q27, [%[c_ptr], #176]\n"
- "udot v28.4s, %[b2].16b, %[a1].4b[0]\n"
+ ".word 0x6f81e09c // udot v28.4s, %[b2].16b, %[a1].4b[0]\n"
"str q12, [%[c_ptr], #192]\n"
- "udot v13.4s, %[b0].16b, %[a1].4b[1]\n"
+ ".word 0x6fa1e04d // udot v13.4s, %[b0].16b, %[a1].4b[1]\n"
"str q20, [%[c_ptr], #208]\n"
- "udot v21.4s, %[b1].16b, %[a1].4b[1]\n"
+ ".word 0x6fa1e075 // udot v21.4s, %[b1].16b, %[a1].4b[1]\n"
"str q28, [%[c_ptr], #224]\n"
- "udot v29.4s, %[b2].16b, %[a1].4b[1]\n"
+ ".word 0x6fa1e09d // udot v29.4s, %[b2].16b, %[a1].4b[1]\n"
"str q13, [%[c_ptr], #240]\n"
- "udot v14.4s, %[b0].16b, %[a1].4b[2]\n"
+ ".word 0x6f81e84e // udot v14.4s, %[b0].16b, %[a1].4b[2]\n"
"str q21, [%[c_ptr], #256]\n"
- "udot v22.4s, %[b1].16b, %[a1].4b[2]\n"
+ ".word 0x6f81e876 // udot v22.4s, %[b1].16b, %[a1].4b[2]\n"
"str q29, [%[c_ptr], #272]\n"
- "udot v30.4s, %[b2].16b, %[a1].4b[2]\n"
+ ".word 0x6f81e89e // udot v30.4s, %[b2].16b, %[a1].4b[2]\n"
"str q14, [%[c_ptr], #288]\n"
- "udot v15.4s, %[b0].16b, %[a1].4b[3]\n"
+ ".word 0x6fa1e84f // udot v15.4s, %[b0].16b, %[a1].4b[3]\n"
"str q22, [%[c_ptr], #304]\n"
- "udot v23.4s, %[b1].16b, %[a1].4b[3]\n"
+ ".word 0x6fa1e877 // udot v23.4s, %[b1].16b, %[a1].4b[3]\n"
"str q30, [%[c_ptr], #320]\n"
- "udot v31.4s, %[b2].16b, %[a1].4b[3]\n"
+ ".word 0x6fa1e89f // udot v31.4s, %[b2].16b, %[a1].4b[3]\n"
"str q15, [%[c_ptr], #336]\n"
@@ -341,9 +332,6 @@ void a64_gemm_u8_12x8(const uint8_t *Apanel, const uint8_t *Bpanel, uint32_t *Cp
"str q31, [%[c_ptr], #368]\n"
"add %[c_ptr], %[c_ptr], #384\n"
-#ifdef NO_DOT_IN_TOOLCHAIN
- ".purgem udot\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),