aboutsummaryrefslogtreecommitdiff
path: root/src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_s8s32_dot_4x6/a55.cpp
diff options
context:
space:
mode:
authorGeorgios Pinitas <georgios.pinitas@arm.com>2019-10-14 19:03:09 +0100
committerGeorgios Pinitas <georgios.pinitas@arm.com>2019-10-23 12:08:12 +0000
commit48b3ef89de5f21a0169d8416e3d54081f82c7bf8 (patch)
treef857d733ccf446c704823dc7ac796a96eb55095e /src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_s8s32_dot_4x6/a55.cpp
parent1dce3101ef8d77c8cf0af7dfd4af6595a0136b91 (diff)
downloadComputeLibrary-48b3ef89de5f21a0169d8416e3d54081f82c7bf8.tar.gz
COMPMID-2577: Fuse bias addition and activation in gemm assembly kernels
Change-Id: I7f52112d2d05b1ea3d3f3d4b19b8eafab05d6c44 Signed-off-by: Georgios Pinitas <georgios.pinitas@arm.com> Reviewed-on: https://review.mlplatform.org/c/2141 Comments-Addressed: Arm Jenkins <bsgcomp@arm.com> Tested-by: Arm Jenkins <bsgcomp@arm.com> Reviewed-by: Pablo Marquez <pablo.tello@arm.com>
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_s8s32_dot_4x6/a55.cpp')
-rw-r--r--src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_s8s32_dot_4x6/a55.cpp4895
1 files changed, 2523 insertions, 2372 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_s8s32_dot_4x6/a55.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_s8s32_dot_4x6/a55.cpp
index e8ac2d7489..6d71840801 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_s8s32_dot_4x6/a55.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_s8s32_dot_4x6/a55.cpp
@@ -25,15 +25,16 @@
#include <algorithm>
+#include "arm_gemm.hpp"
+
#include <cstdint>
#include "../../asmlib.hpp"
#include "../../utils.hpp"
namespace arm_gemm {
-void a64_smallK_hybrid_s8s32_dot_4x6_a55(const int8_t *A, int lda, const int8_t *B, int32_t *C, int ldc, int32_t beta, int M, int N, int K) {
- UNUSED(beta);
- const long loops_count = (N / 4) - 1;
+void a64_smallK_hybrid_s8s32_dot_4x6_a55(const int8_t *A, int lda, const int8_t *B, int32_t *C, int ldc, int M, int N, int K, const int32_t *, Activation, bool) {
+ const long loops_count = iceildiv(N, (int)4) - 1;
const long ldab = lda * sizeof(int8_t);
const long ldcb = ldc * sizeof(int32_t);
const long odds_count = K % 4;
@@ -97,335 +98,356 @@ void a64_smallK_hybrid_s8s32_dot_4x6_a55(const int8_t *A, int lda, const int8_t
"add a_ptr1, %[a_ptr0], #0x0\n"
"1:\n"
"ldr q0, [%[a_ptr0]], #0x10\n"
- "ldr q4, [a_ptr1], #0x10\n"
- "ldr q8, [a_ptr2], #0x10\n"
- "ldr q12, [a_ptr3], #0x10\n"
- "ldr q16, [a_ptr4], #0x10\n"
- "ldr q20, [a_ptr5], #0x10\n"
+ "ldr q3, [a_ptr1], #0x10\n"
+ "ldr q6, [a_ptr2], #0x10\n"
+ "ldr q9, [a_ptr3], #0x10\n"
+ "ldr q12, [a_ptr4], #0x10\n"
+ "ldr q15, [a_ptr5], #0x10\n"
"ldr q1, [%[a_ptr0]], #0x10\n"
- "ldr q5, [a_ptr1], #0x10\n"
- "ldr q9, [a_ptr2], #0x10\n"
- "ldr q13, [a_ptr3], #0x10\n"
- "ldr q17, [a_ptr4], #0x10\n"
- "ldr q21, [a_ptr5], #0x10\n"
+ "ldr q4, [a_ptr1], #0x10\n"
+ "ldr q7, [a_ptr2], #0x10\n"
+ "ldr q10, [a_ptr3], #0x10\n"
+ "ldr q13, [a_ptr4], #0x10\n"
+ "ldr q16, [a_ptr5], #0x10\n"
"cbnz %[odds], 2f\n"
"ldr s2, [%[a_ptr0]]\n"
- "ldr s6, [a_ptr1]\n"
- "ldr s10, [a_ptr2]\n"
- "ldr s14, [a_ptr3]\n"
- "ldr s18, [a_ptr4]\n"
- "ldr s22, [a_ptr5]\n"
+ "ldr s5, [a_ptr1]\n"
+ "ldr s8, [a_ptr2]\n"
+ "ldr s11, [a_ptr3]\n"
+ "ldr s14, [a_ptr4]\n"
+ "ldr s17, [a_ptr5]\n"
"b 3f\n"
"2:\n"
"subs %[odds], %[odds], #0x1\n"
"b.ne 4f\n"
"ldr b2, [%[a_ptr0]]\n"
- "ldr b6, [a_ptr1]\n"
- "ldr b10, [a_ptr2]\n"
- "ldr b14, [a_ptr3]\n"
- "ldr b18, [a_ptr4]\n"
- "ldr b22, [a_ptr5]\n"
+ "ldr b5, [a_ptr1]\n"
+ "ldr b8, [a_ptr2]\n"
+ "ldr b11, [a_ptr3]\n"
+ "ldr b14, [a_ptr4]\n"
+ "ldr b17, [a_ptr5]\n"
"b 3f\n"
"4:\n"
"ldr h2, [%[a_ptr0]], #0x2\n"
- "ldr h6, [a_ptr1], #0x2\n"
- "ldr h10, [a_ptr2], #0x2\n"
- "ldr h14, [a_ptr3], #0x2\n"
- "ldr h18, [a_ptr4], #0x2\n"
- "ldr h22, [a_ptr5], #0x2\n"
+ "ldr h5, [a_ptr1], #0x2\n"
+ "ldr h8, [a_ptr2], #0x2\n"
+ "ldr h11, [a_ptr3], #0x2\n"
+ "ldr h14, [a_ptr4], #0x2\n"
+ "ldr h17, [a_ptr5], #0x2\n"
"subs %[odds], %[odds], #0x1\n"
"b.ne 5f\n"
"b 3f\n"
"5:\n"
"ld1 {v2.b}[2], [%[a_ptr0]]\n"
- "ld1 {v6.b}[2], [a_ptr1]\n"
- "ld1 {v10.b}[2], [a_ptr2]\n"
- "ld1 {v14.b}[2], [a_ptr3]\n"
- "ld1 {v18.b}[2], [a_ptr4]\n"
- "ld1 {v22.b}[2], [a_ptr5]\n"
+ "ld1 {v5.b}[2], [a_ptr1]\n"
+ "ld1 {v8.b}[2], [a_ptr2]\n"
+ "ld1 {v11.b}[2], [a_ptr3]\n"
+ "ld1 {v14.b}[2], [a_ptr4]\n"
+ "ld1 {v17.b}[2], [a_ptr5]\n"
"3:\n"
"movi v26.4s, #0\n"
- "ldr q24, [%[b_ptr0]]\n"
+ "ldr q18, [%[b_ptr0]]\n"
"movi v27.4s, #0\n"
- "ldr q25, [%[b_ptr0], #0x10]\n"
+ "ldr q19, [%[b_ptr0], #0x10]\n"
"movi v28.4s, #0\n"
- "prfm PLDL1KEEP, [a_ptr5, #0x40]\n"
+ "ldr q20, [%[b_ptr0], #0x20]\n"
"movi v29.4s, #0\n"
- "prfm PLDL1KEEP, [a_ptr5, #0x80]\n"
+ "ldr q21, [%[b_ptr0], #0x30]\n"
"movi v30.4s, #0\n"
- "prfm PLDL1KEEP, [a_ptr5, #0xc0]\n"
+ "ldr q22, [%[b_ptr0], #0x40]\n"
"movi v31.4s, #0\n"
+ "ldr q23, [%[b_ptr0], #0x50]\n"
+ ".inst 0x4f80e25a // sdot v26.4s, v18.16b, v0.4b[0]\n"
+ "ldr q24, [%[b_ptr0], #0x60]\n"
+ ".inst 0x4f83e25b // sdot v27.4s, v18.16b, v3.4b[0]\n"
+ "ldr q25, [%[b_ptr0], #0x70]\n"
+ ".inst 0x4f86e25c // sdot v28.4s, v18.16b, v6.4b[0]\n"
+ "prfm PLDL1KEEP, [a_ptr5, #0x40]\n"
+ ".inst 0x4f89e25d // sdot v29.4s, v18.16b, v9.4b[0]\n"
+ "prfm PLDL1KEEP, [a_ptr5, #0x80]\n"
+ ".inst 0x4f8ce25e // sdot v30.4s, v18.16b, v12.4b[0]\n"
+ "prfm PLDL1KEEP, [a_ptr5, #0xc0]\n"
+ ".inst 0x4f8fe25f // sdot v31.4s, v18.16b, v15.4b[0]\n"
"prfm PLDL1KEEP, [a_ptr5, #0x100]\n"
- ".word 0x4f80e31a // sdot v26.4s, v24.16b, v0.4b[0]\n"
+ ".inst 0x4fa0e27a // sdot v26.4s, v19.16b, v0.4b[1]\n"
"prfm PLDL1KEEP, [a_ptr5, #0x140]\n"
- ".word 0x4f84e31b // sdot v27.4s, v24.16b, v4.4b[0]\n"
+ ".inst 0x4fa3e27b // sdot v27.4s, v19.16b, v3.4b[1]\n"
"prfm PLDL1KEEP, [a_ptr5, #0x180]\n"
- ".word 0x4f88e31c // sdot v28.4s, v24.16b, v8.4b[0]\n"
- "add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f8ce31d // sdot v29.4s, v24.16b, v12.4b[0]\n"
- ".word 0x4f90e31e // sdot v30.4s, v24.16b, v16.4b[0]\n"
- ".word 0x4f94e31f // sdot v31.4s, v24.16b, v20.4b[0]\n"
- "ldr q24, [%[b_ptr0]]\n"
- ".word 0x4fa0e33a // sdot v26.4s, v25.16b, v0.4b[1]\n"
- ".word 0x4fa4e33b // sdot v27.4s, v25.16b, v4.4b[1]\n"
- ".word 0x4fa8e33c // sdot v28.4s, v25.16b, v8.4b[1]\n"
- ".word 0x4face33d // sdot v29.4s, v25.16b, v12.4b[1]\n"
- ".word 0x4fb0e33e // sdot v30.4s, v25.16b, v16.4b[1]\n"
- ".word 0x4fb4e33f // sdot v31.4s, v25.16b, v20.4b[1]\n"
- "ldr q25, [%[b_ptr0], #0x10]\n"
- ".word 0x4f80eb1a // sdot v26.4s, v24.16b, v0.4b[2]\n"
- "add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f84eb1b // sdot v27.4s, v24.16b, v4.4b[2]\n"
- ".word 0x4f88eb1c // sdot v28.4s, v24.16b, v8.4b[2]\n"
- ".word 0x4f8ceb1d // sdot v29.4s, v24.16b, v12.4b[2]\n"
- ".word 0x4f90eb1e // sdot v30.4s, v24.16b, v16.4b[2]\n"
- ".word 0x4f94eb1f // sdot v31.4s, v24.16b, v20.4b[2]\n"
- "ldr q24, [%[b_ptr0]]\n"
- ".word 0x4fa0eb3a // sdot v26.4s, v25.16b, v0.4b[3]\n"
- ".word 0x4fa4eb3b // sdot v27.4s, v25.16b, v4.4b[3]\n"
- ".word 0x4fa8eb3c // sdot v28.4s, v25.16b, v8.4b[3]\n"
- ".word 0x4faceb3d // sdot v29.4s, v25.16b, v12.4b[3]\n"
- ".word 0x4fb0eb3e // sdot v30.4s, v25.16b, v16.4b[3]\n"
- ".word 0x4fb4eb3f // sdot v31.4s, v25.16b, v20.4b[3]\n"
- "ldr q25, [%[b_ptr0], #0x10]\n"
- ".word 0x4f81e31a // sdot v26.4s, v24.16b, v1.4b[0]\n"
- "add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f85e31b // sdot v27.4s, v24.16b, v5.4b[0]\n"
- ".word 0x4f89e31c // sdot v28.4s, v24.16b, v9.4b[0]\n"
- ".word 0x4f8de31d // sdot v29.4s, v24.16b, v13.4b[0]\n"
- ".word 0x4f91e31e // sdot v30.4s, v24.16b, v17.4b[0]\n"
- ".word 0x4f95e31f // sdot v31.4s, v24.16b, v21.4b[0]\n"
- "ldr q24, [%[b_ptr0]]\n"
- ".word 0x4fa1e33a // sdot v26.4s, v25.16b, v1.4b[1]\n"
- ".word 0x4fa5e33b // sdot v27.4s, v25.16b, v5.4b[1]\n"
- ".word 0x4fa9e33c // sdot v28.4s, v25.16b, v9.4b[1]\n"
- ".word 0x4fade33d // sdot v29.4s, v25.16b, v13.4b[1]\n"
- ".word 0x4fb1e33e // sdot v30.4s, v25.16b, v17.4b[1]\n"
- ".word 0x4fb5e33f // sdot v31.4s, v25.16b, v21.4b[1]\n"
- "ldr q25, [%[b_ptr0], #0x10]\n"
- ".word 0x4f81eb1a // sdot v26.4s, v24.16b, v1.4b[2]\n"
- "add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f85eb1b // sdot v27.4s, v24.16b, v5.4b[2]\n"
- ".word 0x4f89eb1c // sdot v28.4s, v24.16b, v9.4b[2]\n"
- ".word 0x4f8deb1d // sdot v29.4s, v24.16b, v13.4b[2]\n"
- ".word 0x4f91eb1e // sdot v30.4s, v24.16b, v17.4b[2]\n"
- ".word 0x4f95eb1f // sdot v31.4s, v24.16b, v21.4b[2]\n"
- "ldr q24, [%[b_ptr0]]\n"
- ".word 0x4fa1eb3a // sdot v26.4s, v25.16b, v1.4b[3]\n"
+ ".inst 0x4fa6e27c // sdot v28.4s, v19.16b, v6.4b[1]\n"
+ "add %[b_ptr0], %[b_ptr0], #0x80\n"
+ ".inst 0x4fa9e27d // sdot v29.4s, v19.16b, v9.4b[1]\n"
+ "ldr q18, [%[b_ptr0]]\n"
+ ".inst 0x4face27e // sdot v30.4s, v19.16b, v12.4b[1]\n"
"add %[b_ptr0], %[b_ptr0], #0x10\n"
- ".word 0x4fa5eb3b // sdot v27.4s, v25.16b, v5.4b[3]\n"
- ".word 0x4fa9eb3c // sdot v28.4s, v25.16b, v9.4b[3]\n"
- ".word 0x4fadeb3d // sdot v29.4s, v25.16b, v13.4b[3]\n"
- ".word 0x4fb1eb3e // sdot v30.4s, v25.16b, v17.4b[3]\n"
- ".word 0x4fb5eb3f // sdot v31.4s, v25.16b, v21.4b[3]\n"
- ".word 0x4f82e31a // sdot v26.4s, v24.16b, v2.4b[0]\n"
- ".word 0x4f86e31b // sdot v27.4s, v24.16b, v6.4b[0]\n"
- ".word 0x4f8ae31c // sdot v28.4s, v24.16b, v10.4b[0]\n"
- ".word 0x4f8ee31d // sdot v29.4s, v24.16b, v14.4b[0]\n"
- ".word 0x4f92e31e // sdot v30.4s, v24.16b, v18.4b[0]\n"
- ".word 0x4f96e31f // sdot v31.4s, v24.16b, v22.4b[0]\n"
+ ".inst 0x4fafe27f // sdot v31.4s, v19.16b, v15.4b[1]\n"
+ ".inst 0x4f80ea9a // sdot v26.4s, v20.16b, v0.4b[2]\n"
+ ".inst 0x4f83ea9b // sdot v27.4s, v20.16b, v3.4b[2]\n"
+ ".inst 0x4f86ea9c // sdot v28.4s, v20.16b, v6.4b[2]\n"
+ ".inst 0x4f89ea9d // sdot v29.4s, v20.16b, v9.4b[2]\n"
+ ".inst 0x4f8cea9e // sdot v30.4s, v20.16b, v12.4b[2]\n"
+ ".inst 0x4f8fea9f // sdot v31.4s, v20.16b, v15.4b[2]\n"
+ ".inst 0x4fa0eaba // sdot v26.4s, v21.16b, v0.4b[3]\n"
+ ".inst 0x4fa3eabb // sdot v27.4s, v21.16b, v3.4b[3]\n"
+ ".inst 0x4fa6eabc // sdot v28.4s, v21.16b, v6.4b[3]\n"
+ ".inst 0x4fa9eabd // sdot v29.4s, v21.16b, v9.4b[3]\n"
+ ".inst 0x4faceabe // sdot v30.4s, v21.16b, v12.4b[3]\n"
+ ".inst 0x4fafeabf // sdot v31.4s, v21.16b, v15.4b[3]\n"
+ ".inst 0x4f81e2da // sdot v26.4s, v22.16b, v1.4b[0]\n"
+ ".inst 0x4f84e2db // sdot v27.4s, v22.16b, v4.4b[0]\n"
+ ".inst 0x4f87e2dc // sdot v28.4s, v22.16b, v7.4b[0]\n"
+ ".inst 0x4f8ae2dd // sdot v29.4s, v22.16b, v10.4b[0]\n"
+ ".inst 0x4f8de2de // sdot v30.4s, v22.16b, v13.4b[0]\n"
+ ".inst 0x4f90e2df // sdot v31.4s, v22.16b, v16.4b[0]\n"
+ ".inst 0x4fa1e2fa // sdot v26.4s, v23.16b, v1.4b[1]\n"
+ ".inst 0x4fa4e2fb // sdot v27.4s, v23.16b, v4.4b[1]\n"
+ ".inst 0x4fa7e2fc // sdot v28.4s, v23.16b, v7.4b[1]\n"
+ ".inst 0x4faae2fd // sdot v29.4s, v23.16b, v10.4b[1]\n"
+ ".inst 0x4fade2fe // sdot v30.4s, v23.16b, v13.4b[1]\n"
+ ".inst 0x4fb0e2ff // sdot v31.4s, v23.16b, v16.4b[1]\n"
+ ".inst 0x4f81eb1a // sdot v26.4s, v24.16b, v1.4b[2]\n"
+ ".inst 0x4f84eb1b // sdot v27.4s, v24.16b, v4.4b[2]\n"
+ ".inst 0x4f87eb1c // sdot v28.4s, v24.16b, v7.4b[2]\n"
+ ".inst 0x4f8aeb1d // sdot v29.4s, v24.16b, v10.4b[2]\n"
+ ".inst 0x4f8deb1e // sdot v30.4s, v24.16b, v13.4b[2]\n"
+ ".inst 0x4f90eb1f // sdot v31.4s, v24.16b, v16.4b[2]\n"
+ ".inst 0x4fa1eb3a // sdot v26.4s, v25.16b, v1.4b[3]\n"
+ ".inst 0x4fa4eb3b // sdot v27.4s, v25.16b, v4.4b[3]\n"
+ ".inst 0x4fa7eb3c // sdot v28.4s, v25.16b, v7.4b[3]\n"
+ ".inst 0x4faaeb3d // sdot v29.4s, v25.16b, v10.4b[3]\n"
+ ".inst 0x4fadeb3e // sdot v30.4s, v25.16b, v13.4b[3]\n"
+ ".inst 0x4fb0eb3f // sdot v31.4s, v25.16b, v16.4b[3]\n"
+ ".inst 0x4f82e25a // sdot v26.4s, v18.16b, v2.4b[0]\n"
+ ".inst 0x4f85e25b // sdot v27.4s, v18.16b, v5.4b[0]\n"
+ ".inst 0x4f88e25c // sdot v28.4s, v18.16b, v8.4b[0]\n"
+ ".inst 0x4f8be25d // sdot v29.4s, v18.16b, v11.4b[0]\n"
+ ".inst 0x4f8ee25e // sdot v30.4s, v18.16b, v14.4b[0]\n"
+ ".inst 0x4f91e25f // sdot v31.4s, v18.16b, v17.4b[0]\n"
"cbz %[loops], 6f\n"
- "ldr d24, [%[b_ptr0]]\n"
+ "ldr d18, [%[b_ptr0]]\n"
"subs %[loops], %[loops], #0x1\n"
- "ldr temploadreg0, [%[b_ptr0], #0x8]\n"
- "ldr d25, [%[b_ptr0], #0x10]\n"
- "ldr temploadreg1, [%[b_ptr0], #0x18]\n"
- "add %[b_ptr0], %[b_ptr0], #0x20\n"
+ "ldr temploadreg2, [%[b_ptr0], #0x8]\n"
+ "ldr d19, [%[b_ptr0], #0x10]\n"
+ "ldr temploadreg3, [%[b_ptr0], #0x18]\n"
+ "ldr d20, [%[b_ptr0], #0x20]\n"
+ "ldr temploadreg0, [%[b_ptr0], #0x28]\n"
+ "ldr d21, [%[b_ptr0], #0x30]\n"
+ "ldr temploadreg1, [%[b_ptr0], #0x38]\n"
+ "ldr d22, [%[b_ptr0], #0x40]\n"
+ "ins v18.d[1], temploadreg2\n"
+ "ldr temploadreg2, [%[b_ptr0], #0x48]\n"
+ "ldr d23, [%[b_ptr0], #0x50]\n"
+ "ins v19.d[1], temploadreg3\n"
+ "ldr temploadreg3, [%[b_ptr0], #0x58]\n"
+ "ldr d24, [%[b_ptr0], #0x60]\n"
+ "ins v20.d[1], temploadreg0\n"
+ "ldr temploadreg0, [%[b_ptr0], #0x68]\n"
+ "ldr d25, [%[b_ptr0], #0x70]\n"
+ "ins v21.d[1], temploadreg1\n"
+ "ldr temploadreg1, [%[b_ptr0], #0x78]\n"
+ "add %[b_ptr0], %[b_ptr0], #0x80\n"
+ "ins v22.d[1], temploadreg2\n"
+ "ins v23.d[1], temploadreg3\n"
"ins v24.d[1], temploadreg0\n"
"ins v25.d[1], temploadreg1\n"
"b.eq 7f\n"
"8:\n"
- "str q26, [%[c_ptr0]], #0x10\n"
+ "str q26, [%[c_ptr0]]\n"
"subs %[loops], %[loops], #0x1\n"
"movi v26.4s, #0\n"
- "ldr temploadreg0, [%[b_ptr0], #0x8]\n"
- "ldr temploadreg1, [%[b_ptr0], #0x18]\n"
- "prfm PSTL1KEEP, [%[c_ptr0], #0x40]\n"
- "str q27, [c_ptr1], #0x10\n"
+ "ldr temploadreg2, [%[b_ptr0], #0x8]\n"
+ "add %[c_ptr0], %[c_ptr0], #0x10\n"
+ "str q27, [c_ptr1]\n"
+ "add c_ptr1, c_ptr1, #0x10\n"
"movi v27.4s, #0\n"
- ".word 0x4f80e31a // sdot v26.4s, v24.16b, v0.4b[0]\n"
- "str q28, [c_ptr2], #0x10\n"
- "prfm PSTL1KEEP, [c_ptr1, #0x40]\n"
+ "prfm PSTL1KEEP, [%[c_ptr0], #0x40]\n"
+ ".inst 0x4f80e25a // sdot v26.4s, v18.16b, v0.4b[0]\n"
+ "str q28, [c_ptr2]\n"
"movi v28.4s, #0\n"
- ".word 0x4f84e31b // sdot v27.4s, v24.16b, v4.4b[0]\n"
- "str q29, [c_ptr3], #0x10\n"
+ "add c_ptr2, c_ptr2, #0x10\n"
+ ".inst 0x4f83e25b // sdot v27.4s, v18.16b, v3.4b[0]\n"
+ "str q29, [c_ptr3]\n"
"movi v29.4s, #0\n"
- "prfm PSTL1KEEP, [c_ptr2, #0x40]\n"
- ".word 0x4f88e31c // sdot v28.4s, v24.16b, v8.4b[0]\n"
- "str q30, [c_ptr4], #0x10\n"
+ "add c_ptr3, c_ptr3, #0x10\n"
+ ".inst 0x4f86e25c // sdot v28.4s, v18.16b, v6.4b[0]\n"
+ "str q30, [c_ptr4]\n"
"movi v30.4s, #0\n"
- "prfm PSTL1KEEP, [c_ptr3, #0x40]\n"
- ".word 0x4f8ce31d // sdot v29.4s, v24.16b, v12.4b[0]\n"
- "str q31, [c_ptr5], #0x10\n"
+ "add c_ptr4, c_ptr4, #0x10\n"
+ ".inst 0x4f89e25d // sdot v29.4s, v18.16b, v9.4b[0]\n"
+ "str q31, [c_ptr5]\n"
"movi v31.4s, #0\n"
+ "add c_ptr5, c_ptr5, #0x10\n"
+ ".inst 0x4f8ce25e // sdot v30.4s, v18.16b, v12.4b[0]\n"
+ "prfm PSTL1KEEP, [c_ptr1, #0x40]\n"
+ ".inst 0x4f8fe25f // sdot v31.4s, v18.16b, v15.4b[0]\n"
+ "ldr d18, [%[b_ptr0]]\n"
+ ".inst 0x4fa0e27a // sdot v26.4s, v19.16b, v0.4b[1]\n"
+ "add %[b_ptr0], %[b_ptr0], #0x10\n"
+ ".inst 0x4fa3e27b // sdot v27.4s, v19.16b, v3.4b[1]\n"
+ "ins v18.d[1], temploadreg2\n"
+ ".inst 0x4fa6e27c // sdot v28.4s, v19.16b, v6.4b[1]\n"
+ "ldr temploadreg2, [%[b_ptr0], #0x8]\n"
+ ".inst 0x4fa9e27d // sdot v29.4s, v19.16b, v9.4b[1]\n"
+ "ldr temploadreg3, [%[b_ptr0], #0x18]\n"
+ ".inst 0x4face27e // sdot v30.4s, v19.16b, v12.4b[1]\n"
+ "ldr temploadreg0, [%[b_ptr0], #0x28]\n"
+ ".inst 0x4fafe27f // sdot v31.4s, v19.16b, v15.4b[1]\n"
+ "ldr d19, [%[b_ptr0], #0x10]\n"
+ ".inst 0x4f80ea9a // sdot v26.4s, v20.16b, v0.4b[2]\n"
+ "ldr temploadreg1, [%[b_ptr0], #0x38]\n"
+ ".inst 0x4f83ea9b // sdot v27.4s, v20.16b, v3.4b[2]\n"
+ "prfm PSTL1KEEP, [c_ptr2, #0x40]\n"
+ ".inst 0x4f86ea9c // sdot v28.4s, v20.16b, v6.4b[2]\n"
+ "ins v19.d[1], temploadreg3\n"
+ ".inst 0x4f89ea9d // sdot v29.4s, v20.16b, v9.4b[2]\n"
+ "ldr temploadreg3, [%[b_ptr0], #0x58]\n"
+ ".inst 0x4f8cea9e // sdot v30.4s, v20.16b, v12.4b[2]\n"
+ "prfm PSTL1KEEP, [c_ptr3, #0x40]\n"
+ ".inst 0x4f8fea9f // sdot v31.4s, v20.16b, v15.4b[2]\n"
+ "ldr d20, [%[b_ptr0], #0x20]\n"
+ ".inst 0x4fa0eaba // sdot v26.4s, v21.16b, v0.4b[3]\n"
"prfm PSTL1KEEP, [c_ptr4, #0x40]\n"
- ".word 0x4f90e31e // sdot v30.4s, v24.16b, v16.4b[0]\n"
+ ".inst 0x4fa3eabb // sdot v27.4s, v21.16b, v3.4b[3]\n"
+ "ins v20.d[1], temploadreg0\n"
+ ".inst 0x4fa6eabc // sdot v28.4s, v21.16b, v6.4b[3]\n"
+ "ldr temploadreg0, [%[b_ptr0], #0x68]\n"
+ ".inst 0x4fa9eabd // sdot v29.4s, v21.16b, v9.4b[3]\n"
"prfm PSTL1KEEP, [c_ptr5, #0x40]\n"
- ".word 0x4f94e31f // sdot v31.4s, v24.16b, v20.4b[0]\n"
- "ldr d24, [%[b_ptr0]]\n"
- ".word 0x4fa0e33a // sdot v26.4s, v25.16b, v0.4b[1]\n"
- ".word 0x4fa4e33b // sdot v27.4s, v25.16b, v4.4b[1]\n"
- ".word 0x4fa8e33c // sdot v28.4s, v25.16b, v8.4b[1]\n"
- "ins v24.d[1], temploadreg0\n"
- ".word 0x4face33d // sdot v29.4s, v25.16b, v12.4b[1]\n"
- ".word 0x4fb0e33e // sdot v30.4s, v25.16b, v16.4b[1]\n"
- ".word 0x4fb4e33f // sdot v31.4s, v25.16b, v20.4b[1]\n"
- "ldr d25, [%[b_ptr0], #0x10]\n"
- ".word 0x4f80eb1a // sdot v26.4s, v24.16b, v0.4b[2]\n"
- "add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f84eb1b // sdot v27.4s, v24.16b, v4.4b[2]\n"
- "ins v25.d[1], temploadreg1\n"
- ".word 0x4f88eb1c // sdot v28.4s, v24.16b, v8.4b[2]\n"
- "ldr temploadreg0, [%[b_ptr0], #0x8]\n"
- ".word 0x4f8ceb1d // sdot v29.4s, v24.16b, v12.4b[2]\n"
- "ldr temploadreg1, [%[b_ptr0], #0x18]\n"
- ".word 0x4f90eb1e // sdot v30.4s, v24.16b, v16.4b[2]\n"
- ".word 0x4f94eb1f // sdot v31.4s, v24.16b, v20.4b[2]\n"
- "ldr d24, [%[b_ptr0]]\n"
- ".word 0x4fa0eb3a // sdot v26.4s, v25.16b, v0.4b[3]\n"
- ".word 0x4fa4eb3b // sdot v27.4s, v25.16b, v4.4b[3]\n"
- ".word 0x4fa8eb3c // sdot v28.4s, v25.16b, v8.4b[3]\n"
+ ".inst 0x4faceabe // sdot v30.4s, v21.16b, v12.4b[3]\n"
+ ".inst 0x4fafeabf // sdot v31.4s, v21.16b, v15.4b[3]\n"
+ "ldr d21, [%[b_ptr0], #0x30]\n"
+ ".inst 0x4f81e2da // sdot v26.4s, v22.16b, v1.4b[0]\n"
+ ".inst 0x4f84e2db // sdot v27.4s, v22.16b, v4.4b[0]\n"
+ ".inst 0x4f87e2dc // sdot v28.4s, v22.16b, v7.4b[0]\n"
+ "ins v21.d[1], temploadreg1\n"
+ ".inst 0x4f8ae2dd // sdot v29.4s, v22.16b, v10.4b[0]\n"
+ "ldr temploadreg1, [%[b_ptr0], #0x78]\n"
+ ".inst 0x4f8de2de // sdot v30.4s, v22.16b, v13.4b[0]\n"
+ ".inst 0x4f90e2df // sdot v31.4s, v22.16b, v16.4b[0]\n"
+ "ldr d22, [%[b_ptr0], #0x40]\n"
+ ".inst 0x4fa1e2fa // sdot v26.4s, v23.16b, v1.4b[1]\n"
+ ".inst 0x4fa4e2fb // sdot v27.4s, v23.16b, v4.4b[1]\n"
+ ".inst 0x4fa7e2fc // sdot v28.4s, v23.16b, v7.4b[1]\n"
+ ".inst 0x4faae2fd // sdot v29.4s, v23.16b, v10.4b[1]\n"
+ ".inst 0x4fade2fe // sdot v30.4s, v23.16b, v13.4b[1]\n"
+ ".inst 0x4fb0e2ff // sdot v31.4s, v23.16b, v16.4b[1]\n"
+ "ldr d23, [%[b_ptr0], #0x50]\n"
+ ".inst 0x4f81eb1a // sdot v26.4s, v24.16b, v1.4b[2]\n"
+ ".inst 0x4f84eb1b // sdot v27.4s, v24.16b, v4.4b[2]\n"
+ ".inst 0x4f87eb1c // sdot v28.4s, v24.16b, v7.4b[2]\n"
+ "ins v23.d[1], temploadreg3\n"
+ ".inst 0x4f8aeb1d // sdot v29.4s, v24.16b, v10.4b[2]\n"
+ ".inst 0x4f8deb1e // sdot v30.4s, v24.16b, v13.4b[2]\n"
+ ".inst 0x4f90eb1f // sdot v31.4s, v24.16b, v16.4b[2]\n"
+ "ldr d24, [%[b_ptr0], #0x60]\n"
+ ".inst 0x4fa1eb3a // sdot v26.4s, v25.16b, v1.4b[3]\n"
+ ".inst 0x4fa4eb3b // sdot v27.4s, v25.16b, v4.4b[3]\n"
+ ".inst 0x4fa7eb3c // sdot v28.4s, v25.16b, v7.4b[3]\n"
"ins v24.d[1], temploadreg0\n"
- ".word 0x4faceb3d // sdot v29.4s, v25.16b, v12.4b[3]\n"
- ".word 0x4fb0eb3e // sdot v30.4s, v25.16b, v16.4b[3]\n"
- ".word 0x4fb4eb3f // sdot v31.4s, v25.16b, v20.4b[3]\n"
- "ldr d25, [%[b_ptr0], #0x10]\n"
- ".word 0x4f81e31a // sdot v26.4s, v24.16b, v1.4b[0]\n"
- "add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f85e31b // sdot v27.4s, v24.16b, v5.4b[0]\n"
+ ".inst 0x4faaeb3d // sdot v29.4s, v25.16b, v10.4b[3]\n"
+ ".inst 0x4fadeb3e // sdot v30.4s, v25.16b, v13.4b[3]\n"
+ ".inst 0x4fb0eb3f // sdot v31.4s, v25.16b, v16.4b[3]\n"
+ "ldr d25, [%[b_ptr0], #0x70]\n"
+ ".inst 0x4f82e25a // sdot v26.4s, v18.16b, v2.4b[0]\n"
+ ".inst 0x4f85e25b // sdot v27.4s, v18.16b, v5.4b[0]\n"
+ ".inst 0x4f88e25c // sdot v28.4s, v18.16b, v8.4b[0]\n"
"ins v25.d[1], temploadreg1\n"
- ".word 0x4f89e31c // sdot v28.4s, v24.16b, v9.4b[0]\n"
- "ldr temploadreg0, [%[b_ptr0], #0x8]\n"
- ".word 0x4f8de31d // sdot v29.4s, v24.16b, v13.4b[0]\n"
- "ldr temploadreg1, [%[b_ptr0], #0x18]\n"
- ".word 0x4f91e31e // sdot v30.4s, v24.16b, v17.4b[0]\n"
- ".word 0x4f95e31f // sdot v31.4s, v24.16b, v21.4b[0]\n"
- "ldr d24, [%[b_ptr0]]\n"
- ".word 0x4fa1e33a // sdot v26.4s, v25.16b, v1.4b[1]\n"
- ".word 0x4fa5e33b // sdot v27.4s, v25.16b, v5.4b[1]\n"
- ".word 0x4fa9e33c // sdot v28.4s, v25.16b, v9.4b[1]\n"
- "ins v24.d[1], temploadreg0\n"
- ".word 0x4fade33d // sdot v29.4s, v25.16b, v13.4b[1]\n"
- ".word 0x4fb1e33e // sdot v30.4s, v25.16b, v17.4b[1]\n"
- ".word 0x4fb5e33f // sdot v31.4s, v25.16b, v21.4b[1]\n"
- "ldr d25, [%[b_ptr0], #0x10]\n"
- ".word 0x4f81eb1a // sdot v26.4s, v24.16b, v1.4b[2]\n"
- "add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f85eb1b // sdot v27.4s, v24.16b, v5.4b[2]\n"
- "ins v25.d[1], temploadreg1\n"
- ".word 0x4f89eb1c // sdot v28.4s, v24.16b, v9.4b[2]\n"
- "ldr temploadreg0, [%[b_ptr0], #0x8]\n"
- ".word 0x4f8deb1d // sdot v29.4s, v24.16b, v13.4b[2]\n"
- ".word 0x4f91eb1e // sdot v30.4s, v24.16b, v17.4b[2]\n"
- ".word 0x4f95eb1f // sdot v31.4s, v24.16b, v21.4b[2]\n"
- "ldr d24, [%[b_ptr0]]\n"
- ".word 0x4fa1eb3a // sdot v26.4s, v25.16b, v1.4b[3]\n"
- "add %[b_ptr0], %[b_ptr0], #0x10\n"
- ".word 0x4fa5eb3b // sdot v27.4s, v25.16b, v5.4b[3]\n"
- "ins v24.d[1], temploadreg0\n"
- ".word 0x4fa9eb3c // sdot v28.4s, v25.16b, v9.4b[3]\n"
- "ldr temploadreg0, [%[b_ptr0], #0x8]\n"
- ".word 0x4fadeb3d // sdot v29.4s, v25.16b, v13.4b[3]\n"
- "ldr temploadreg1, [%[b_ptr0], #0x18]\n"
- ".word 0x4fb1eb3e // sdot v30.4s, v25.16b, v17.4b[3]\n"
- ".word 0x4fb5eb3f // sdot v31.4s, v25.16b, v21.4b[3]\n"
- "ldr d25, [%[b_ptr0], #0x10]\n"
- ".word 0x4f82e31a // sdot v26.4s, v24.16b, v2.4b[0]\n"
- ".word 0x4f86e31b // sdot v27.4s, v24.16b, v6.4b[0]\n"
- ".word 0x4f8ae31c // sdot v28.4s, v24.16b, v10.4b[0]\n"
- "ins v25.d[1], temploadreg1\n"
- ".word 0x4f8ee31d // sdot v29.4s, v24.16b, v14.4b[0]\n"
- ".word 0x4f92e31e // sdot v30.4s, v24.16b, v18.4b[0]\n"
- ".word 0x4f96e31f // sdot v31.4s, v24.16b, v22.4b[0]\n"
- "ldr d24, [%[b_ptr0]]\n"
- "add %[b_ptr0], %[b_ptr0], #0x20\n"
- "ins v24.d[1], temploadreg0\n"
+ ".inst 0x4f8be25d // sdot v29.4s, v18.16b, v11.4b[0]\n"
+ ".inst 0x4f8ee25e // sdot v30.4s, v18.16b, v14.4b[0]\n"
+ ".inst 0x4f91e25f // sdot v31.4s, v18.16b, v17.4b[0]\n"
+ "ldr d18, [%[b_ptr0]]\n"
+ "ins v18.d[1], temploadreg2\n"
+ "ldr temploadreg2, [%[b_ptr0], #0x48]\n"
+ "add %[b_ptr0], %[b_ptr0], #0x80\n"
+ "ins v22.d[1], temploadreg2\n"
"b.ne 8b\n"
"7:\n"
- "str q26, [%[c_ptr0]], #0x10\n"
+ "str q26, [%[c_ptr0]]\n"
+ "add %[c_ptr0], %[c_ptr0], #0x10\n"
"movi v26.4s, #0\n"
- "str q27, [c_ptr1], #0x10\n"
+ "str q27, [c_ptr1]\n"
+ "add c_ptr1, c_ptr1, #0x10\n"
"movi v27.4s, #0\n"
- ".word 0x4f80e31a // sdot v26.4s, v24.16b, v0.4b[0]\n"
- "str q28, [c_ptr2], #0x10\n"
+ ".inst 0x4f80e25a // sdot v26.4s, v18.16b, v0.4b[0]\n"
+ "str q28, [c_ptr2]\n"
"movi v28.4s, #0\n"
- ".word 0x4f84e31b // sdot v27.4s, v24.16b, v4.4b[0]\n"
- ".word 0x4fa0e33a // sdot v26.4s, v25.16b, v0.4b[1]\n"
- "str q29, [c_ptr3], #0x10\n"
+ "add c_ptr2, c_ptr2, #0x10\n"
+ ".inst 0x4f83e25b // sdot v27.4s, v18.16b, v3.4b[0]\n"
+ "str q29, [c_ptr3]\n"
"movi v29.4s, #0\n"
- ".word 0x4f88e31c // sdot v28.4s, v24.16b, v8.4b[0]\n"
- ".word 0x4fa4e33b // sdot v27.4s, v25.16b, v4.4b[1]\n"
- "str q30, [c_ptr4], #0x10\n"
+ "add c_ptr3, c_ptr3, #0x10\n"
+ ".inst 0x4f86e25c // sdot v28.4s, v18.16b, v6.4b[0]\n"
+ "str q30, [c_ptr4]\n"
"movi v30.4s, #0\n"
- ".word 0x4f8ce31d // sdot v29.4s, v24.16b, v12.4b[0]\n"
- ".word 0x4fa8e33c // sdot v28.4s, v25.16b, v8.4b[1]\n"
- "str q31, [c_ptr5], #0x10\n"
+ "add c_ptr4, c_ptr4, #0x10\n"
+ ".inst 0x4f89e25d // sdot v29.4s, v18.16b, v9.4b[0]\n"
+ "str q31, [c_ptr5]\n"
"movi v31.4s, #0\n"
- ".word 0x4f90e31e // sdot v30.4s, v24.16b, v16.4b[0]\n"
- ".word 0x4face33d // sdot v29.4s, v25.16b, v12.4b[1]\n"
- ".word 0x4f94e31f // sdot v31.4s, v24.16b, v20.4b[0]\n"
- "ldr q24, [%[b_ptr0]]\n"
- ".word 0x4fb0e33e // sdot v30.4s, v25.16b, v16.4b[1]\n"
- ".word 0x4fb4e33f // sdot v31.4s, v25.16b, v20.4b[1]\n"
- "ldr q25, [%[b_ptr0], #0x10]\n"
- ".word 0x4f80eb1a // sdot v26.4s, v24.16b, v0.4b[2]\n"
- "add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f84eb1b // sdot v27.4s, v24.16b, v4.4b[2]\n"
- ".word 0x4f88eb1c // sdot v28.4s, v24.16b, v8.4b[2]\n"
- ".word 0x4f8ceb1d // sdot v29.4s, v24.16b, v12.4b[2]\n"
- ".word 0x4f90eb1e // sdot v30.4s, v24.16b, v16.4b[2]\n"
- ".word 0x4f94eb1f // sdot v31.4s, v24.16b, v20.4b[2]\n"
- "ldr q24, [%[b_ptr0]]\n"
- ".word 0x4fa0eb3a // sdot v26.4s, v25.16b, v0.4b[3]\n"
- ".word 0x4fa4eb3b // sdot v27.4s, v25.16b, v4.4b[3]\n"
- ".word 0x4fa8eb3c // sdot v28.4s, v25.16b, v8.4b[3]\n"
- ".word 0x4faceb3d // sdot v29.4s, v25.16b, v12.4b[3]\n"
- ".word 0x4fb0eb3e // sdot v30.4s, v25.16b, v16.4b[3]\n"
- ".word 0x4fb4eb3f // sdot v31.4s, v25.16b, v20.4b[3]\n"
- "ldr q25, [%[b_ptr0], #0x10]\n"
- ".word 0x4f81e31a // sdot v26.4s, v24.16b, v1.4b[0]\n"
- "add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f85e31b // sdot v27.4s, v24.16b, v5.4b[0]\n"
- ".word 0x4f89e31c // sdot v28.4s, v24.16b, v9.4b[0]\n"
- ".word 0x4f8de31d // sdot v29.4s, v24.16b, v13.4b[0]\n"
- ".word 0x4f91e31e // sdot v30.4s, v24.16b, v17.4b[0]\n"
- ".word 0x4f95e31f // sdot v31.4s, v24.16b, v21.4b[0]\n"
- "ldr q24, [%[b_ptr0]]\n"
- ".word 0x4fa1e33a // sdot v26.4s, v25.16b, v1.4b[1]\n"
- ".word 0x4fa5e33b // sdot v27.4s, v25.16b, v5.4b[1]\n"
- ".word 0x4fa9e33c // sdot v28.4s, v25.16b, v9.4b[1]\n"
- ".word 0x4fade33d // sdot v29.4s, v25.16b, v13.4b[1]\n"
- ".word 0x4fb1e33e // sdot v30.4s, v25.16b, v17.4b[1]\n"
- ".word 0x4fb5e33f // sdot v31.4s, v25.16b, v21.4b[1]\n"
- "ldr q25, [%[b_ptr0], #0x10]\n"
- ".word 0x4f81eb1a // sdot v26.4s, v24.16b, v1.4b[2]\n"
- "add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f85eb1b // sdot v27.4s, v24.16b, v5.4b[2]\n"
- ".word 0x4f89eb1c // sdot v28.4s, v24.16b, v9.4b[2]\n"
- ".word 0x4f8deb1d // sdot v29.4s, v24.16b, v13.4b[2]\n"
- ".word 0x4f91eb1e // sdot v30.4s, v24.16b, v17.4b[2]\n"
- ".word 0x4f95eb1f // sdot v31.4s, v24.16b, v21.4b[2]\n"
- "ldr q24, [%[b_ptr0]]\n"
- ".word 0x4fa1eb3a // sdot v26.4s, v25.16b, v1.4b[3]\n"
+ "add c_ptr5, c_ptr5, #0x10\n"
+ ".inst 0x4f8ce25e // sdot v30.4s, v18.16b, v12.4b[0]\n"
+ ".inst 0x4fa0e27a // sdot v26.4s, v19.16b, v0.4b[1]\n"
+ ".inst 0x4f8fe25f // sdot v31.4s, v18.16b, v15.4b[0]\n"
+ "ldr q18, [%[b_ptr0]]\n"
+ ".inst 0x4fa3e27b // sdot v27.4s, v19.16b, v3.4b[1]\n"
"add %[b_ptr0], %[b_ptr0], #0x10\n"
- ".word 0x4fa5eb3b // sdot v27.4s, v25.16b, v5.4b[3]\n"
- ".word 0x4fa9eb3c // sdot v28.4s, v25.16b, v9.4b[3]\n"
- ".word 0x4fadeb3d // sdot v29.4s, v25.16b, v13.4b[3]\n"
- ".word 0x4fb1eb3e // sdot v30.4s, v25.16b, v17.4b[3]\n"
- ".word 0x4fb5eb3f // sdot v31.4s, v25.16b, v21.4b[3]\n"
- ".word 0x4f82e31a // sdot v26.4s, v24.16b, v2.4b[0]\n"
- ".word 0x4f86e31b // sdot v27.4s, v24.16b, v6.4b[0]\n"
- ".word 0x4f8ae31c // sdot v28.4s, v24.16b, v10.4b[0]\n"
- ".word 0x4f8ee31d // sdot v29.4s, v24.16b, v14.4b[0]\n"
- ".word 0x4f92e31e // sdot v30.4s, v24.16b, v18.4b[0]\n"
- ".word 0x4f96e31f // sdot v31.4s, v24.16b, v22.4b[0]\n"
+ ".inst 0x4fa6e27c // sdot v28.4s, v19.16b, v6.4b[1]\n"
+ ".inst 0x4fa9e27d // sdot v29.4s, v19.16b, v9.4b[1]\n"
+ ".inst 0x4face27e // sdot v30.4s, v19.16b, v12.4b[1]\n"
+ ".inst 0x4fafe27f // sdot v31.4s, v19.16b, v15.4b[1]\n"
+ ".inst 0x4f80ea9a // sdot v26.4s, v20.16b, v0.4b[2]\n"
+ ".inst 0x4f83ea9b // sdot v27.4s, v20.16b, v3.4b[2]\n"
+ ".inst 0x4f86ea9c // sdot v28.4s, v20.16b, v6.4b[2]\n"
+ ".inst 0x4f89ea9d // sdot v29.4s, v20.16b, v9.4b[2]\n"
+ ".inst 0x4f8cea9e // sdot v30.4s, v20.16b, v12.4b[2]\n"
+ ".inst 0x4f8fea9f // sdot v31.4s, v20.16b, v15.4b[2]\n"
+ ".inst 0x4fa0eaba // sdot v26.4s, v21.16b, v0.4b[3]\n"
+ ".inst 0x4fa3eabb // sdot v27.4s, v21.16b, v3.4b[3]\n"
+ ".inst 0x4fa6eabc // sdot v28.4s, v21.16b, v6.4b[3]\n"
+ ".inst 0x4fa9eabd // sdot v29.4s, v21.16b, v9.4b[3]\n"
+ ".inst 0x4faceabe // sdot v30.4s, v21.16b, v12.4b[3]\n"
+ ".inst 0x4fafeabf // sdot v31.4s, v21.16b, v15.4b[3]\n"
+ ".inst 0x4f81e2da // sdot v26.4s, v22.16b, v1.4b[0]\n"
+ ".inst 0x4f84e2db // sdot v27.4s, v22.16b, v4.4b[0]\n"
+ ".inst 0x4f87e2dc // sdot v28.4s, v22.16b, v7.4b[0]\n"
+ ".inst 0x4f8ae2dd // sdot v29.4s, v22.16b, v10.4b[0]\n"
+ ".inst 0x4f8de2de // sdot v30.4s, v22.16b, v13.4b[0]\n"
+ ".inst 0x4f90e2df // sdot v31.4s, v22.16b, v16.4b[0]\n"
+ ".inst 0x4fa1e2fa // sdot v26.4s, v23.16b, v1.4b[1]\n"
+ ".inst 0x4fa4e2fb // sdot v27.4s, v23.16b, v4.4b[1]\n"
+ ".inst 0x4fa7e2fc // sdot v28.4s, v23.16b, v7.4b[1]\n"
+ ".inst 0x4faae2fd // sdot v29.4s, v23.16b, v10.4b[1]\n"
+ ".inst 0x4fade2fe // sdot v30.4s, v23.16b, v13.4b[1]\n"
+ ".inst 0x4fb0e2ff // sdot v31.4s, v23.16b, v16.4b[1]\n"
+ ".inst 0x4f81eb1a // sdot v26.4s, v24.16b, v1.4b[2]\n"
+ ".inst 0x4f84eb1b // sdot v27.4s, v24.16b, v4.4b[2]\n"
+ ".inst 0x4f87eb1c // sdot v28.4s, v24.16b, v7.4b[2]\n"
+ ".inst 0x4f8aeb1d // sdot v29.4s, v24.16b, v10.4b[2]\n"
+ ".inst 0x4f8deb1e // sdot v30.4s, v24.16b, v13.4b[2]\n"
+ ".inst 0x4f90eb1f // sdot v31.4s, v24.16b, v16.4b[2]\n"
+ ".inst 0x4fa1eb3a // sdot v26.4s, v25.16b, v1.4b[3]\n"
+ ".inst 0x4fa4eb3b // sdot v27.4s, v25.16b, v4.4b[3]\n"
+ ".inst 0x4fa7eb3c // sdot v28.4s, v25.16b, v7.4b[3]\n"
+ ".inst 0x4faaeb3d // sdot v29.4s, v25.16b, v10.4b[3]\n"
+ ".inst 0x4fadeb3e // sdot v30.4s, v25.16b, v13.4b[3]\n"
+ ".inst 0x4fb0eb3f // sdot v31.4s, v25.16b, v16.4b[3]\n"
+ ".inst 0x4f82e25a // sdot v26.4s, v18.16b, v2.4b[0]\n"
+ ".inst 0x4f85e25b // sdot v27.4s, v18.16b, v5.4b[0]\n"
+ ".inst 0x4f88e25c // sdot v28.4s, v18.16b, v8.4b[0]\n"
+ ".inst 0x4f8be25d // sdot v29.4s, v18.16b, v11.4b[0]\n"
+ ".inst 0x4f8ee25e // sdot v30.4s, v18.16b, v14.4b[0]\n"
+ ".inst 0x4f91e25f // sdot v31.4s, v18.16b, v17.4b[0]\n"
"6:\n"
"str q26, [%[c_ptr0]]\n"
+ "add %[c_ptr0], %[c_ptr0], #0x10\n"
"str q27, [c_ptr1]\n"
+ "add c_ptr1, c_ptr1, #0x10\n"
"str q28, [c_ptr2]\n"
+ "add c_ptr2, c_ptr2, #0x10\n"
"str q29, [c_ptr3]\n"
+ "add c_ptr3, c_ptr3, #0x10\n"
"str q30, [c_ptr4]\n"
+ "add c_ptr4, c_ptr4, #0x10\n"
"str q31, [c_ptr5]\n"
+ "add c_ptr5, c_ptr5, #0x10\n"
".unreq a_ptr1\n"
".unreq a_ptr2\n"
".unreq a_ptr3\n"
@@ -493,364 +515,385 @@ void a64_smallK_hybrid_s8s32_dot_4x6_a55(const int8_t *A, int lda, const int8_t
"add a_ptr1, %[a_ptr0], #0x0\n"
"1:\n"
"ldr q0, [%[a_ptr0]], #0x10\n"
- "ldr q4, [a_ptr1], #0x10\n"
- "ldr q8, [a_ptr2], #0x10\n"
- "ldr q12, [a_ptr3], #0x10\n"
- "ldr q16, [a_ptr4], #0x10\n"
- "ldr q20, [a_ptr5], #0x10\n"
+ "ldr q3, [a_ptr1], #0x10\n"
+ "ldr q6, [a_ptr2], #0x10\n"
+ "ldr q9, [a_ptr3], #0x10\n"
+ "ldr q12, [a_ptr4], #0x10\n"
+ "ldr q15, [a_ptr5], #0x10\n"
"ldr q1, [%[a_ptr0]], #0x10\n"
- "ldr q5, [a_ptr1], #0x10\n"
- "ldr q9, [a_ptr2], #0x10\n"
- "ldr q13, [a_ptr3], #0x10\n"
- "ldr q17, [a_ptr4], #0x10\n"
- "ldr q21, [a_ptr5], #0x10\n"
+ "ldr q4, [a_ptr1], #0x10\n"
+ "ldr q7, [a_ptr2], #0x10\n"
+ "ldr q10, [a_ptr3], #0x10\n"
+ "ldr q13, [a_ptr4], #0x10\n"
+ "ldr q16, [a_ptr5], #0x10\n"
"cbnz %[odds], 2f\n"
"ldr d2, [%[a_ptr0]]\n"
- "ldr d6, [a_ptr1]\n"
- "ldr d10, [a_ptr2]\n"
- "ldr d14, [a_ptr3]\n"
- "ldr d18, [a_ptr4]\n"
- "ldr d22, [a_ptr5]\n"
+ "ldr d5, [a_ptr1]\n"
+ "ldr d8, [a_ptr2]\n"
+ "ldr d11, [a_ptr3]\n"
+ "ldr d14, [a_ptr4]\n"
+ "ldr d17, [a_ptr5]\n"
"b 3f\n"
"2:\n"
"ldr s2, [%[a_ptr0]], #0x4\n"
- "ldr s6, [a_ptr1], #0x4\n"
- "ldr s10, [a_ptr2], #0x4\n"
- "ldr s14, [a_ptr3], #0x4\n"
- "ldr s18, [a_ptr4], #0x4\n"
- "ldr s22, [a_ptr5], #0x4\n"
+ "ldr s5, [a_ptr1], #0x4\n"
+ "ldr s8, [a_ptr2], #0x4\n"
+ "ldr s11, [a_ptr3], #0x4\n"
+ "ldr s14, [a_ptr4], #0x4\n"
+ "ldr s17, [a_ptr5], #0x4\n"
"subs %[odds], %[odds], #0x1\n"
"b.ne 4f\n"
"ld1 {v2.b}[4], [%[a_ptr0]]\n"
- "ld1 {v6.b}[4], [a_ptr1]\n"
- "ld1 {v10.b}[4], [a_ptr2]\n"
- "ld1 {v14.b}[4], [a_ptr3]\n"
- "ld1 {v18.b}[4], [a_ptr4]\n"
- "ld1 {v22.b}[4], [a_ptr5]\n"
+ "ld1 {v5.b}[4], [a_ptr1]\n"
+ "ld1 {v8.b}[4], [a_ptr2]\n"
+ "ld1 {v11.b}[4], [a_ptr3]\n"
+ "ld1 {v14.b}[4], [a_ptr4]\n"
+ "ld1 {v17.b}[4], [a_ptr5]\n"
"b 3f\n"
"4:\n"
"ld1 {v2.h}[2], [%[a_ptr0]], #2\n"
- "ld1 {v6.h}[2], [a_ptr1], #2\n"
- "ld1 {v10.h}[2], [a_ptr2], #2\n"
- "ld1 {v14.h}[2], [a_ptr3], #2\n"
- "ld1 {v18.h}[2], [a_ptr4], #2\n"
- "ld1 {v22.h}[2], [a_ptr5], #2\n"
+ "ld1 {v5.h}[2], [a_ptr1], #2\n"
+ "ld1 {v8.h}[2], [a_ptr2], #2\n"
+ "ld1 {v11.h}[2], [a_ptr3], #2\n"
+ "ld1 {v14.h}[2], [a_ptr4], #2\n"
+ "ld1 {v17.h}[2], [a_ptr5], #2\n"
"subs %[odds], %[odds], #0x1\n"
"b.ne 5f\n"
"b 3f\n"
"5:\n"
"ld1 {v2.b}[6], [%[a_ptr0]]\n"
- "ld1 {v6.b}[6], [a_ptr1]\n"
- "ld1 {v10.b}[6], [a_ptr2]\n"
- "ld1 {v14.b}[6], [a_ptr3]\n"
- "ld1 {v18.b}[6], [a_ptr4]\n"
- "ld1 {v22.b}[6], [a_ptr5]\n"
+ "ld1 {v5.b}[6], [a_ptr1]\n"
+ "ld1 {v8.b}[6], [a_ptr2]\n"
+ "ld1 {v11.b}[6], [a_ptr3]\n"
+ "ld1 {v14.b}[6], [a_ptr4]\n"
+ "ld1 {v17.b}[6], [a_ptr5]\n"
"3:\n"
"movi v26.4s, #0\n"
- "ldr q24, [%[b_ptr0]]\n"
+ "ldr q18, [%[b_ptr0]]\n"
"movi v27.4s, #0\n"
- "ldr q25, [%[b_ptr0], #0x10]\n"
+ "ldr q19, [%[b_ptr0], #0x10]\n"
"movi v28.4s, #0\n"
- "prfm PLDL1KEEP, [a_ptr5, #0x40]\n"
+ "ldr q20, [%[b_ptr0], #0x20]\n"
"movi v29.4s, #0\n"
- "prfm PLDL1KEEP, [a_ptr5, #0x80]\n"
+ "ldr q21, [%[b_ptr0], #0x30]\n"
"movi v30.4s, #0\n"
- "prfm PLDL1KEEP, [a_ptr5, #0xc0]\n"
+ "ldr q22, [%[b_ptr0], #0x40]\n"
"movi v31.4s, #0\n"
+ "ldr q23, [%[b_ptr0], #0x50]\n"
+ ".inst 0x4f80e25a // sdot v26.4s, v18.16b, v0.4b[0]\n"
+ "ldr q24, [%[b_ptr0], #0x60]\n"
+ ".inst 0x4f83e25b // sdot v27.4s, v18.16b, v3.4b[0]\n"
+ "ldr q25, [%[b_ptr0], #0x70]\n"
+ ".inst 0x4f86e25c // sdot v28.4s, v18.16b, v6.4b[0]\n"
+ "prfm PLDL1KEEP, [a_ptr5, #0x40]\n"
+ ".inst 0x4f89e25d // sdot v29.4s, v18.16b, v9.4b[0]\n"
+ "prfm PLDL1KEEP, [a_ptr5, #0x80]\n"
+ ".inst 0x4f8ce25e // sdot v30.4s, v18.16b, v12.4b[0]\n"
+ "prfm PLDL1KEEP, [a_ptr5, #0xc0]\n"
+ ".inst 0x4f8fe25f // sdot v31.4s, v18.16b, v15.4b[0]\n"
"prfm PLDL1KEEP, [a_ptr5, #0x100]\n"
- ".word 0x4f80e31a // sdot v26.4s, v24.16b, v0.4b[0]\n"
+ ".inst 0x4fa0e27a // sdot v26.4s, v19.16b, v0.4b[1]\n"
"prfm PLDL1KEEP, [a_ptr5, #0x140]\n"
- ".word 0x4f84e31b // sdot v27.4s, v24.16b, v4.4b[0]\n"
+ ".inst 0x4fa3e27b // sdot v27.4s, v19.16b, v3.4b[1]\n"
"prfm PLDL1KEEP, [a_ptr5, #0x180]\n"
- ".word 0x4f88e31c // sdot v28.4s, v24.16b, v8.4b[0]\n"
- "add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f8ce31d // sdot v29.4s, v24.16b, v12.4b[0]\n"
- ".word 0x4f90e31e // sdot v30.4s, v24.16b, v16.4b[0]\n"
- ".word 0x4f94e31f // sdot v31.4s, v24.16b, v20.4b[0]\n"
- "ldr q24, [%[b_ptr0]]\n"
- ".word 0x4fa0e33a // sdot v26.4s, v25.16b, v0.4b[1]\n"
- ".word 0x4fa4e33b // sdot v27.4s, v25.16b, v4.4b[1]\n"
- ".word 0x4fa8e33c // sdot v28.4s, v25.16b, v8.4b[1]\n"
- ".word 0x4face33d // sdot v29.4s, v25.16b, v12.4b[1]\n"
- ".word 0x4fb0e33e // sdot v30.4s, v25.16b, v16.4b[1]\n"
- ".word 0x4fb4e33f // sdot v31.4s, v25.16b, v20.4b[1]\n"
- "ldr q25, [%[b_ptr0], #0x10]\n"
- ".word 0x4f80eb1a // sdot v26.4s, v24.16b, v0.4b[2]\n"
- "add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f84eb1b // sdot v27.4s, v24.16b, v4.4b[2]\n"
- ".word 0x4f88eb1c // sdot v28.4s, v24.16b, v8.4b[2]\n"
- ".word 0x4f8ceb1d // sdot v29.4s, v24.16b, v12.4b[2]\n"
- ".word 0x4f90eb1e // sdot v30.4s, v24.16b, v16.4b[2]\n"
- ".word 0x4f94eb1f // sdot v31.4s, v24.16b, v20.4b[2]\n"
- "ldr q24, [%[b_ptr0]]\n"
- ".word 0x4fa0eb3a // sdot v26.4s, v25.16b, v0.4b[3]\n"
- ".word 0x4fa4eb3b // sdot v27.4s, v25.16b, v4.4b[3]\n"
- ".word 0x4fa8eb3c // sdot v28.4s, v25.16b, v8.4b[3]\n"
- ".word 0x4faceb3d // sdot v29.4s, v25.16b, v12.4b[3]\n"
- ".word 0x4fb0eb3e // sdot v30.4s, v25.16b, v16.4b[3]\n"
- ".word 0x4fb4eb3f // sdot v31.4s, v25.16b, v20.4b[3]\n"
- "ldr q25, [%[b_ptr0], #0x10]\n"
- ".word 0x4f81e31a // sdot v26.4s, v24.16b, v1.4b[0]\n"
- "add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f85e31b // sdot v27.4s, v24.16b, v5.4b[0]\n"
- ".word 0x4f89e31c // sdot v28.4s, v24.16b, v9.4b[0]\n"
- ".word 0x4f8de31d // sdot v29.4s, v24.16b, v13.4b[0]\n"
- ".word 0x4f91e31e // sdot v30.4s, v24.16b, v17.4b[0]\n"
- ".word 0x4f95e31f // sdot v31.4s, v24.16b, v21.4b[0]\n"
- "ldr q24, [%[b_ptr0]]\n"
- ".word 0x4fa1e33a // sdot v26.4s, v25.16b, v1.4b[1]\n"
- ".word 0x4fa5e33b // sdot v27.4s, v25.16b, v5.4b[1]\n"
- ".word 0x4fa9e33c // sdot v28.4s, v25.16b, v9.4b[1]\n"
- ".word 0x4fade33d // sdot v29.4s, v25.16b, v13.4b[1]\n"
- ".word 0x4fb1e33e // sdot v30.4s, v25.16b, v17.4b[1]\n"
- ".word 0x4fb5e33f // sdot v31.4s, v25.16b, v21.4b[1]\n"
- "ldr q25, [%[b_ptr0], #0x10]\n"
- ".word 0x4f81eb1a // sdot v26.4s, v24.16b, v1.4b[2]\n"
- "add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f85eb1b // sdot v27.4s, v24.16b, v5.4b[2]\n"
- ".word 0x4f89eb1c // sdot v28.4s, v24.16b, v9.4b[2]\n"
- ".word 0x4f8deb1d // sdot v29.4s, v24.16b, v13.4b[2]\n"
- ".word 0x4f91eb1e // sdot v30.4s, v24.16b, v17.4b[2]\n"
- ".word 0x4f95eb1f // sdot v31.4s, v24.16b, v21.4b[2]\n"
- "ldr q24, [%[b_ptr0]]\n"
- ".word 0x4fa1eb3a // sdot v26.4s, v25.16b, v1.4b[3]\n"
- ".word 0x4fa5eb3b // sdot v27.4s, v25.16b, v5.4b[3]\n"
- ".word 0x4fa9eb3c // sdot v28.4s, v25.16b, v9.4b[3]\n"
- ".word 0x4fadeb3d // sdot v29.4s, v25.16b, v13.4b[3]\n"
- ".word 0x4fb1eb3e // sdot v30.4s, v25.16b, v17.4b[3]\n"
- ".word 0x4fb5eb3f // sdot v31.4s, v25.16b, v21.4b[3]\n"
- "ldr q25, [%[b_ptr0], #0x10]\n"
- ".word 0x4f82e31a // sdot v26.4s, v24.16b, v2.4b[0]\n"
- "add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f86e31b // sdot v27.4s, v24.16b, v6.4b[0]\n"
- ".word 0x4f8ae31c // sdot v28.4s, v24.16b, v10.4b[0]\n"
- ".word 0x4f8ee31d // sdot v29.4s, v24.16b, v14.4b[0]\n"
- ".word 0x4f92e31e // sdot v30.4s, v24.16b, v18.4b[0]\n"
- ".word 0x4f96e31f // sdot v31.4s, v24.16b, v22.4b[0]\n"
- ".word 0x4fa2e33a // sdot v26.4s, v25.16b, v2.4b[1]\n"
- ".word 0x4fa6e33b // sdot v27.4s, v25.16b, v6.4b[1]\n"
- ".word 0x4faae33c // sdot v28.4s, v25.16b, v10.4b[1]\n"
- ".word 0x4faee33d // sdot v29.4s, v25.16b, v14.4b[1]\n"
- ".word 0x4fb2e33e // sdot v30.4s, v25.16b, v18.4b[1]\n"
- ".word 0x4fb6e33f // sdot v31.4s, v25.16b, v22.4b[1]\n"
+ ".inst 0x4fa6e27c // sdot v28.4s, v19.16b, v6.4b[1]\n"
+ "add %[b_ptr0], %[b_ptr0], #0x80\n"
+ ".inst 0x4fa9e27d // sdot v29.4s, v19.16b, v9.4b[1]\n"
+ "ldr q18, [%[b_ptr0]]\n"
+ ".inst 0x4face27e // sdot v30.4s, v19.16b, v12.4b[1]\n"
+ ".inst 0x4fafe27f // sdot v31.4s, v19.16b, v15.4b[1]\n"
+ "ldr q19, [%[b_ptr0], #0x10]\n"
+ ".inst 0x4f80ea9a // sdot v26.4s, v20.16b, v0.4b[2]\n"
+ "add %[b_ptr0], %[b_ptr0], #0x20\n"
+ ".inst 0x4f83ea9b // sdot v27.4s, v20.16b, v3.4b[2]\n"
+ ".inst 0x4f86ea9c // sdot v28.4s, v20.16b, v6.4b[2]\n"
+ ".inst 0x4f89ea9d // sdot v29.4s, v20.16b, v9.4b[2]\n"
+ ".inst 0x4f8cea9e // sdot v30.4s, v20.16b, v12.4b[2]\n"
+ ".inst 0x4f8fea9f // sdot v31.4s, v20.16b, v15.4b[2]\n"
+ ".inst 0x4fa0eaba // sdot v26.4s, v21.16b, v0.4b[3]\n"
+ ".inst 0x4fa3eabb // sdot v27.4s, v21.16b, v3.4b[3]\n"
+ ".inst 0x4fa6eabc // sdot v28.4s, v21.16b, v6.4b[3]\n"
+ ".inst 0x4fa9eabd // sdot v29.4s, v21.16b, v9.4b[3]\n"
+ ".inst 0x4faceabe // sdot v30.4s, v21.16b, v12.4b[3]\n"
+ ".inst 0x4fafeabf // sdot v31.4s, v21.16b, v15.4b[3]\n"
+ ".inst 0x4f81e2da // sdot v26.4s, v22.16b, v1.4b[0]\n"
+ ".inst 0x4f84e2db // sdot v27.4s, v22.16b, v4.4b[0]\n"
+ ".inst 0x4f87e2dc // sdot v28.4s, v22.16b, v7.4b[0]\n"
+ ".inst 0x4f8ae2dd // sdot v29.4s, v22.16b, v10.4b[0]\n"
+ ".inst 0x4f8de2de // sdot v30.4s, v22.16b, v13.4b[0]\n"
+ ".inst 0x4f90e2df // sdot v31.4s, v22.16b, v16.4b[0]\n"
+ ".inst 0x4fa1e2fa // sdot v26.4s, v23.16b, v1.4b[1]\n"
+ ".inst 0x4fa4e2fb // sdot v27.4s, v23.16b, v4.4b[1]\n"
+ ".inst 0x4fa7e2fc // sdot v28.4s, v23.16b, v7.4b[1]\n"
+ ".inst 0x4faae2fd // sdot v29.4s, v23.16b, v10.4b[1]\n"
+ ".inst 0x4fade2fe // sdot v30.4s, v23.16b, v13.4b[1]\n"
+ ".inst 0x4fb0e2ff // sdot v31.4s, v23.16b, v16.4b[1]\n"
+ ".inst 0x4f81eb1a // sdot v26.4s, v24.16b, v1.4b[2]\n"
+ ".inst 0x4f84eb1b // sdot v27.4s, v24.16b, v4.4b[2]\n"
+ ".inst 0x4f87eb1c // sdot v28.4s, v24.16b, v7.4b[2]\n"
+ ".inst 0x4f8aeb1d // sdot v29.4s, v24.16b, v10.4b[2]\n"
+ ".inst 0x4f8deb1e // sdot v30.4s, v24.16b, v13.4b[2]\n"
+ ".inst 0x4f90eb1f // sdot v31.4s, v24.16b, v16.4b[2]\n"
+ ".inst 0x4fa1eb3a // sdot v26.4s, v25.16b, v1.4b[3]\n"
+ ".inst 0x4fa4eb3b // sdot v27.4s, v25.16b, v4.4b[3]\n"
+ ".inst 0x4fa7eb3c // sdot v28.4s, v25.16b, v7.4b[3]\n"
+ ".inst 0x4faaeb3d // sdot v29.4s, v25.16b, v10.4b[3]\n"
+ ".inst 0x4fadeb3e // sdot v30.4s, v25.16b, v13.4b[3]\n"
+ ".inst 0x4fb0eb3f // sdot v31.4s, v25.16b, v16.4b[3]\n"
+ ".inst 0x4f82e25a // sdot v26.4s, v18.16b, v2.4b[0]\n"
+ ".inst 0x4f85e25b // sdot v27.4s, v18.16b, v5.4b[0]\n"
+ ".inst 0x4f88e25c // sdot v28.4s, v18.16b, v8.4b[0]\n"
+ ".inst 0x4f8be25d // sdot v29.4s, v18.16b, v11.4b[0]\n"
+ ".inst 0x4f8ee25e // sdot v30.4s, v18.16b, v14.4b[0]\n"
+ ".inst 0x4f91e25f // sdot v31.4s, v18.16b, v17.4b[0]\n"
+ ".inst 0x4fa2e27a // sdot v26.4s, v19.16b, v2.4b[1]\n"
+ ".inst 0x4fa5e27b // sdot v27.4s, v19.16b, v5.4b[1]\n"
+ ".inst 0x4fa8e27c // sdot v28.4s, v19.16b, v8.4b[1]\n"
+ ".inst 0x4fabe27d // sdot v29.4s, v19.16b, v11.4b[1]\n"
+ ".inst 0x4faee27e // sdot v30.4s, v19.16b, v14.4b[1]\n"
+ ".inst 0x4fb1e27f // sdot v31.4s, v19.16b, v17.4b[1]\n"
"cbz %[loops], 6f\n"
- "ldr d24, [%[b_ptr0]]\n"
+ "ldr d18, [%[b_ptr0]]\n"
"subs %[loops], %[loops], #0x1\n"
- "ldr temploadreg0, [%[b_ptr0], #0x8]\n"
- "ldr d25, [%[b_ptr0], #0x10]\n"
- "ldr temploadreg1, [%[b_ptr0], #0x18]\n"
- "add %[b_ptr0], %[b_ptr0], #0x20\n"
+ "ldr temploadreg2, [%[b_ptr0], #0x8]\n"
+ "ldr d19, [%[b_ptr0], #0x10]\n"
+ "ldr temploadreg3, [%[b_ptr0], #0x18]\n"
+ "ldr d20, [%[b_ptr0], #0x20]\n"
+ "ldr temploadreg0, [%[b_ptr0], #0x28]\n"
+ "ldr d21, [%[b_ptr0], #0x30]\n"
+ "ldr temploadreg1, [%[b_ptr0], #0x38]\n"
+ "ldr d22, [%[b_ptr0], #0x40]\n"
+ "ins v18.d[1], temploadreg2\n"
+ "ldr temploadreg2, [%[b_ptr0], #0x48]\n"
+ "ldr d23, [%[b_ptr0], #0x50]\n"
+ "ins v19.d[1], temploadreg3\n"
+ "ldr temploadreg3, [%[b_ptr0], #0x58]\n"
+ "ldr d24, [%[b_ptr0], #0x60]\n"
+ "ins v20.d[1], temploadreg0\n"
+ "ldr temploadreg0, [%[b_ptr0], #0x68]\n"
+ "ldr d25, [%[b_ptr0], #0x70]\n"
+ "ins v21.d[1], temploadreg1\n"
+ "ldr temploadreg1, [%[b_ptr0], #0x78]\n"
+ "add %[b_ptr0], %[b_ptr0], #0x80\n"
+ "ins v22.d[1], temploadreg2\n"
+ "ins v23.d[1], temploadreg3\n"
"ins v24.d[1], temploadreg0\n"
+ "ins v25.d[1], temploadreg1\n"
"b.eq 7f\n"
"8:\n"
- "str q26, [%[c_ptr0]], #0x10\n"
+ "str q26, [%[c_ptr0]]\n"
"subs %[loops], %[loops], #0x1\n"
"movi v26.4s, #0\n"
- "ins v25.d[1], temploadreg1\n"
- "ldr temploadreg0, [%[b_ptr0], #0x8]\n"
- "prfm PSTL1KEEP, [%[c_ptr0], #0x40]\n"
- "str q27, [c_ptr1], #0x10\n"
+ "ldr temploadreg2, [%[b_ptr0], #0x8]\n"
+ "ldr temploadreg3, [%[b_ptr0], #0x18]\n"
+ "add %[c_ptr0], %[c_ptr0], #0x10\n"
+ "str q27, [c_ptr1]\n"
+ "add c_ptr1, c_ptr1, #0x10\n"
"movi v27.4s, #0\n"
- "ldr temploadreg1, [%[b_ptr0], #0x18]\n"
- ".word 0x4f80e31a // sdot v26.4s, v24.16b, v0.4b[0]\n"
- "str q28, [c_ptr2], #0x10\n"
- "prfm PSTL1KEEP, [c_ptr1, #0x40]\n"
+ "prfm PSTL1KEEP, [%[c_ptr0], #0x40]\n"
+ ".inst 0x4f80e25a // sdot v26.4s, v18.16b, v0.4b[0]\n"
+ "str q28, [c_ptr2]\n"
"movi v28.4s, #0\n"
- ".word 0x4f84e31b // sdot v27.4s, v24.16b, v4.4b[0]\n"
- "str q29, [c_ptr3], #0x10\n"
+ "add c_ptr2, c_ptr2, #0x10\n"
+ ".inst 0x4f83e25b // sdot v27.4s, v18.16b, v3.4b[0]\n"
+ "str q29, [c_ptr3]\n"
"movi v29.4s, #0\n"
- "prfm PSTL1KEEP, [c_ptr2, #0x40]\n"
- ".word 0x4f88e31c // sdot v28.4s, v24.16b, v8.4b[0]\n"
- "str q30, [c_ptr4], #0x10\n"
+ "add c_ptr3, c_ptr3, #0x10\n"
+ ".inst 0x4f86e25c // sdot v28.4s, v18.16b, v6.4b[0]\n"
+ "str q30, [c_ptr4]\n"
"movi v30.4s, #0\n"
- "prfm PSTL1KEEP, [c_ptr3, #0x40]\n"
- ".word 0x4f8ce31d // sdot v29.4s, v24.16b, v12.4b[0]\n"
- "str q31, [c_ptr5], #0x10\n"
+ "add c_ptr4, c_ptr4, #0x10\n"
+ ".inst 0x4f89e25d // sdot v29.4s, v18.16b, v9.4b[0]\n"
+ "str q31, [c_ptr5]\n"
"movi v31.4s, #0\n"
+ "add c_ptr5, c_ptr5, #0x10\n"
+ ".inst 0x4f8ce25e // sdot v30.4s, v18.16b, v12.4b[0]\n"
+ "prfm PSTL1KEEP, [c_ptr1, #0x40]\n"
+ ".inst 0x4f8fe25f // sdot v31.4s, v18.16b, v15.4b[0]\n"
+ "ldr d18, [%[b_ptr0]]\n"
+ ".inst 0x4fa0e27a // sdot v26.4s, v19.16b, v0.4b[1]\n"
+ "prfm PSTL1KEEP, [c_ptr2, #0x40]\n"
+ ".inst 0x4fa3e27b // sdot v27.4s, v19.16b, v3.4b[1]\n"
+ "ins v18.d[1], temploadreg2\n"
+ ".inst 0x4fa6e27c // sdot v28.4s, v19.16b, v6.4b[1]\n"
+ "prfm PSTL1KEEP, [c_ptr3, #0x40]\n"
+ ".inst 0x4fa9e27d // sdot v29.4s, v19.16b, v9.4b[1]\n"
"prfm PSTL1KEEP, [c_ptr4, #0x40]\n"
- ".word 0x4f90e31e // sdot v30.4s, v24.16b, v16.4b[0]\n"
+ ".inst 0x4face27e // sdot v30.4s, v19.16b, v12.4b[1]\n"
"prfm PSTL1KEEP, [c_ptr5, #0x40]\n"
- ".word 0x4f94e31f // sdot v31.4s, v24.16b, v20.4b[0]\n"
- "ldr d24, [%[b_ptr0]]\n"
- ".word 0x4fa0e33a // sdot v26.4s, v25.16b, v0.4b[1]\n"
- ".word 0x4fa4e33b // sdot v27.4s, v25.16b, v4.4b[1]\n"
- ".word 0x4fa8e33c // sdot v28.4s, v25.16b, v8.4b[1]\n"
- "ins v24.d[1], temploadreg0\n"
- ".word 0x4face33d // sdot v29.4s, v25.16b, v12.4b[1]\n"
- ".word 0x4fb0e33e // sdot v30.4s, v25.16b, v16.4b[1]\n"
- ".word 0x4fb4e33f // sdot v31.4s, v25.16b, v20.4b[1]\n"
- "ldr d25, [%[b_ptr0], #0x10]\n"
- ".word 0x4f80eb1a // sdot v26.4s, v24.16b, v0.4b[2]\n"
- "add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f84eb1b // sdot v27.4s, v24.16b, v4.4b[2]\n"
- "ins v25.d[1], temploadreg1\n"
- ".word 0x4f88eb1c // sdot v28.4s, v24.16b, v8.4b[2]\n"
- "ldr temploadreg0, [%[b_ptr0], #0x8]\n"
- ".word 0x4f8ceb1d // sdot v29.4s, v24.16b, v12.4b[2]\n"
- "ldr temploadreg1, [%[b_ptr0], #0x18]\n"
- ".word 0x4f90eb1e // sdot v30.4s, v24.16b, v16.4b[2]\n"
- ".word 0x4f94eb1f // sdot v31.4s, v24.16b, v20.4b[2]\n"
- "ldr d24, [%[b_ptr0]]\n"
- ".word 0x4fa0eb3a // sdot v26.4s, v25.16b, v0.4b[3]\n"
- ".word 0x4fa4eb3b // sdot v27.4s, v25.16b, v4.4b[3]\n"
- ".word 0x4fa8eb3c // sdot v28.4s, v25.16b, v8.4b[3]\n"
+ ".inst 0x4fafe27f // sdot v31.4s, v19.16b, v15.4b[1]\n"
+ "ldr d19, [%[b_ptr0], #0x10]\n"
+ ".inst 0x4f80ea9a // sdot v26.4s, v20.16b, v0.4b[2]\n"
+ "add %[b_ptr0], %[b_ptr0], #0x20\n"
+ ".inst 0x4f83ea9b // sdot v27.4s, v20.16b, v3.4b[2]\n"
+ "ins v19.d[1], temploadreg3\n"
+ ".inst 0x4f86ea9c // sdot v28.4s, v20.16b, v6.4b[2]\n"
+ "ldr temploadreg2, [%[b_ptr0], #0x8]\n"
+ ".inst 0x4f89ea9d // sdot v29.4s, v20.16b, v9.4b[2]\n"
+ "ldr temploadreg3, [%[b_ptr0], #0x18]\n"
+ ".inst 0x4f8cea9e // sdot v30.4s, v20.16b, v12.4b[2]\n"
+ "ldr temploadreg0, [%[b_ptr0], #0x28]\n"
+ ".inst 0x4f8fea9f // sdot v31.4s, v20.16b, v15.4b[2]\n"
+ "ldr d20, [%[b_ptr0], #0x20]\n"
+ ".inst 0x4fa0eaba // sdot v26.4s, v21.16b, v0.4b[3]\n"
+ "ldr temploadreg1, [%[b_ptr0], #0x38]\n"
+ ".inst 0x4fa3eabb // sdot v27.4s, v21.16b, v3.4b[3]\n"
+ ".inst 0x4fa6eabc // sdot v28.4s, v21.16b, v6.4b[3]\n"
+ "ins v20.d[1], temploadreg0\n"
+ ".inst 0x4fa9eabd // sdot v29.4s, v21.16b, v9.4b[3]\n"
+ "ldr temploadreg0, [%[b_ptr0], #0x68]\n"
+ ".inst 0x4faceabe // sdot v30.4s, v21.16b, v12.4b[3]\n"
+ ".inst 0x4fafeabf // sdot v31.4s, v21.16b, v15.4b[3]\n"
+ "ldr d21, [%[b_ptr0], #0x30]\n"
+ ".inst 0x4f81e2da // sdot v26.4s, v22.16b, v1.4b[0]\n"
+ ".inst 0x4f84e2db // sdot v27.4s, v22.16b, v4.4b[0]\n"
+ ".inst 0x4f87e2dc // sdot v28.4s, v22.16b, v7.4b[0]\n"
+ "ins v21.d[1], temploadreg1\n"
+ ".inst 0x4f8ae2dd // sdot v29.4s, v22.16b, v10.4b[0]\n"
+ "ldr temploadreg1, [%[b_ptr0], #0x78]\n"
+ ".inst 0x4f8de2de // sdot v30.4s, v22.16b, v13.4b[0]\n"
+ ".inst 0x4f90e2df // sdot v31.4s, v22.16b, v16.4b[0]\n"
+ "ldr d22, [%[b_ptr0], #0x40]\n"
+ ".inst 0x4fa1e2fa // sdot v26.4s, v23.16b, v1.4b[1]\n"
+ ".inst 0x4fa4e2fb // sdot v27.4s, v23.16b, v4.4b[1]\n"
+ ".inst 0x4fa7e2fc // sdot v28.4s, v23.16b, v7.4b[1]\n"
+ ".inst 0x4faae2fd // sdot v29.4s, v23.16b, v10.4b[1]\n"
+ ".inst 0x4fade2fe // sdot v30.4s, v23.16b, v13.4b[1]\n"
+ ".inst 0x4fb0e2ff // sdot v31.4s, v23.16b, v16.4b[1]\n"
+ "ldr d23, [%[b_ptr0], #0x50]\n"
+ ".inst 0x4f81eb1a // sdot v26.4s, v24.16b, v1.4b[2]\n"
+ ".inst 0x4f84eb1b // sdot v27.4s, v24.16b, v4.4b[2]\n"
+ ".inst 0x4f87eb1c // sdot v28.4s, v24.16b, v7.4b[2]\n"
+ ".inst 0x4f8aeb1d // sdot v29.4s, v24.16b, v10.4b[2]\n"
+ ".inst 0x4f8deb1e // sdot v30.4s, v24.16b, v13.4b[2]\n"
+ ".inst 0x4f90eb1f // sdot v31.4s, v24.16b, v16.4b[2]\n"
+ "ldr d24, [%[b_ptr0], #0x60]\n"
+ ".inst 0x4fa1eb3a // sdot v26.4s, v25.16b, v1.4b[3]\n"
+ ".inst 0x4fa4eb3b // sdot v27.4s, v25.16b, v4.4b[3]\n"
+ ".inst 0x4fa7eb3c // sdot v28.4s, v25.16b, v7.4b[3]\n"
"ins v24.d[1], temploadreg0\n"
- ".word 0x4faceb3d // sdot v29.4s, v25.16b, v12.4b[3]\n"
- ".word 0x4fb0eb3e // sdot v30.4s, v25.16b, v16.4b[3]\n"
- ".word 0x4fb4eb3f // sdot v31.4s, v25.16b, v20.4b[3]\n"
- "ldr d25, [%[b_ptr0], #0x10]\n"
- ".word 0x4f81e31a // sdot v26.4s, v24.16b, v1.4b[0]\n"
- "add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f85e31b // sdot v27.4s, v24.16b, v5.4b[0]\n"
+ ".inst 0x4faaeb3d // sdot v29.4s, v25.16b, v10.4b[3]\n"
+ ".inst 0x4fadeb3e // sdot v30.4s, v25.16b, v13.4b[3]\n"
+ ".inst 0x4fb0eb3f // sdot v31.4s, v25.16b, v16.4b[3]\n"
+ "ldr d25, [%[b_ptr0], #0x70]\n"
+ ".inst 0x4f82e25a // sdot v26.4s, v18.16b, v2.4b[0]\n"
+ ".inst 0x4f85e25b // sdot v27.4s, v18.16b, v5.4b[0]\n"
+ ".inst 0x4f88e25c // sdot v28.4s, v18.16b, v8.4b[0]\n"
"ins v25.d[1], temploadreg1\n"
- ".word 0x4f89e31c // sdot v28.4s, v24.16b, v9.4b[0]\n"
- "ldr temploadreg0, [%[b_ptr0], #0x8]\n"
- ".word 0x4f8de31d // sdot v29.4s, v24.16b, v13.4b[0]\n"
- "ldr temploadreg1, [%[b_ptr0], #0x18]\n"
- ".word 0x4f91e31e // sdot v30.4s, v24.16b, v17.4b[0]\n"
- ".word 0x4f95e31f // sdot v31.4s, v24.16b, v21.4b[0]\n"
- "ldr d24, [%[b_ptr0]]\n"
- ".word 0x4fa1e33a // sdot v26.4s, v25.16b, v1.4b[1]\n"
- ".word 0x4fa5e33b // sdot v27.4s, v25.16b, v5.4b[1]\n"
- ".word 0x4fa9e33c // sdot v28.4s, v25.16b, v9.4b[1]\n"
- "ins v24.d[1], temploadreg0\n"
- ".word 0x4fade33d // sdot v29.4s, v25.16b, v13.4b[1]\n"
- ".word 0x4fb1e33e // sdot v30.4s, v25.16b, v17.4b[1]\n"
- ".word 0x4fb5e33f // sdot v31.4s, v25.16b, v21.4b[1]\n"
- "ldr d25, [%[b_ptr0], #0x10]\n"
- ".word 0x4f81eb1a // sdot v26.4s, v24.16b, v1.4b[2]\n"
- "add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f85eb1b // sdot v27.4s, v24.16b, v5.4b[2]\n"
- "ins v25.d[1], temploadreg1\n"
- ".word 0x4f89eb1c // sdot v28.4s, v24.16b, v9.4b[2]\n"
- "ldr temploadreg0, [%[b_ptr0], #0x8]\n"
- ".word 0x4f8deb1d // sdot v29.4s, v24.16b, v13.4b[2]\n"
- "ldr temploadreg1, [%[b_ptr0], #0x18]\n"
- ".word 0x4f91eb1e // sdot v30.4s, v24.16b, v17.4b[2]\n"
- ".word 0x4f95eb1f // sdot v31.4s, v24.16b, v21.4b[2]\n"
- "ldr d24, [%[b_ptr0]]\n"
- ".word 0x4fa1eb3a // sdot v26.4s, v25.16b, v1.4b[3]\n"
- ".word 0x4fa5eb3b // sdot v27.4s, v25.16b, v5.4b[3]\n"
- ".word 0x4fa9eb3c // sdot v28.4s, v25.16b, v9.4b[3]\n"
- "ins v24.d[1], temploadreg0\n"
- ".word 0x4fadeb3d // sdot v29.4s, v25.16b, v13.4b[3]\n"
- ".word 0x4fb1eb3e // sdot v30.4s, v25.16b, v17.4b[3]\n"
- ".word 0x4fb5eb3f // sdot v31.4s, v25.16b, v21.4b[3]\n"
- "ldr d25, [%[b_ptr0], #0x10]\n"
- ".word 0x4f82e31a // sdot v26.4s, v24.16b, v2.4b[0]\n"
- "add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f86e31b // sdot v27.4s, v24.16b, v6.4b[0]\n"
- "ins v25.d[1], temploadreg1\n"
- ".word 0x4f8ae31c // sdot v28.4s, v24.16b, v10.4b[0]\n"
- "ldr temploadreg0, [%[b_ptr0], #0x8]\n"
- ".word 0x4f8ee31d // sdot v29.4s, v24.16b, v14.4b[0]\n"
- "ldr temploadreg1, [%[b_ptr0], #0x18]\n"
- ".word 0x4f92e31e // sdot v30.4s, v24.16b, v18.4b[0]\n"
- ".word 0x4f96e31f // sdot v31.4s, v24.16b, v22.4b[0]\n"
- "ldr d24, [%[b_ptr0]]\n"
- ".word 0x4fa2e33a // sdot v26.4s, v25.16b, v2.4b[1]\n"
- ".word 0x4fa6e33b // sdot v27.4s, v25.16b, v6.4b[1]\n"
- ".word 0x4faae33c // sdot v28.4s, v25.16b, v10.4b[1]\n"
- "ins v24.d[1], temploadreg0\n"
- ".word 0x4faee33d // sdot v29.4s, v25.16b, v14.4b[1]\n"
- ".word 0x4fb2e33e // sdot v30.4s, v25.16b, v18.4b[1]\n"
- ".word 0x4fb6e33f // sdot v31.4s, v25.16b, v22.4b[1]\n"
- "ldr d25, [%[b_ptr0], #0x10]\n"
- "add %[b_ptr0], %[b_ptr0], #0x20\n"
+ ".inst 0x4f8be25d // sdot v29.4s, v18.16b, v11.4b[0]\n"
+ ".inst 0x4f8ee25e // sdot v30.4s, v18.16b, v14.4b[0]\n"
+ ".inst 0x4f91e25f // sdot v31.4s, v18.16b, v17.4b[0]\n"
+ "ldr d18, [%[b_ptr0]]\n"
+ ".inst 0x4fa2e27a // sdot v26.4s, v19.16b, v2.4b[1]\n"
+ ".inst 0x4fa5e27b // sdot v27.4s, v19.16b, v5.4b[1]\n"
+ ".inst 0x4fa8e27c // sdot v28.4s, v19.16b, v8.4b[1]\n"
+ "ins v18.d[1], temploadreg2\n"
+ ".inst 0x4fabe27d // sdot v29.4s, v19.16b, v11.4b[1]\n"
+ "ldr temploadreg2, [%[b_ptr0], #0x48]\n"
+ ".inst 0x4faee27e // sdot v30.4s, v19.16b, v14.4b[1]\n"
+ ".inst 0x4fb1e27f // sdot v31.4s, v19.16b, v17.4b[1]\n"
+ "ldr d19, [%[b_ptr0], #0x10]\n"
+ "ins v22.d[1], temploadreg2\n"
+ "ins v19.d[1], temploadreg3\n"
+ "ldr temploadreg3, [%[b_ptr0], #0x58]\n"
+ "add %[b_ptr0], %[b_ptr0], #0x80\n"
+ "ins v23.d[1], temploadreg3\n"
"b.ne 8b\n"
"7:\n"
- "str q26, [%[c_ptr0]], #0x10\n"
+ "str q26, [%[c_ptr0]]\n"
+ "add %[c_ptr0], %[c_ptr0], #0x10\n"
"movi v26.4s, #0\n"
- "ins v25.d[1], temploadreg1\n"
- "str q27, [c_ptr1], #0x10\n"
+ "str q27, [c_ptr1]\n"
+ "add c_ptr1, c_ptr1, #0x10\n"
"movi v27.4s, #0\n"
- ".word 0x4f80e31a // sdot v26.4s, v24.16b, v0.4b[0]\n"
- "str q28, [c_ptr2], #0x10\n"
+ ".inst 0x4f80e25a // sdot v26.4s, v18.16b, v0.4b[0]\n"
+ "str q28, [c_ptr2]\n"
"movi v28.4s, #0\n"
- ".word 0x4f84e31b // sdot v27.4s, v24.16b, v4.4b[0]\n"
- ".word 0x4fa0e33a // sdot v26.4s, v25.16b, v0.4b[1]\n"
- "str q29, [c_ptr3], #0x10\n"
+ "add c_ptr2, c_ptr2, #0x10\n"
+ ".inst 0x4f83e25b // sdot v27.4s, v18.16b, v3.4b[0]\n"
+ "str q29, [c_ptr3]\n"
"movi v29.4s, #0\n"
- ".word 0x4f88e31c // sdot v28.4s, v24.16b, v8.4b[0]\n"
- ".word 0x4fa4e33b // sdot v27.4s, v25.16b, v4.4b[1]\n"
- "str q30, [c_ptr4], #0x10\n"
+ "add c_ptr3, c_ptr3, #0x10\n"
+ ".inst 0x4f86e25c // sdot v28.4s, v18.16b, v6.4b[0]\n"
+ "str q30, [c_ptr4]\n"
"movi v30.4s, #0\n"
- ".word 0x4f8ce31d // sdot v29.4s, v24.16b, v12.4b[0]\n"
- ".word 0x4fa8e33c // sdot v28.4s, v25.16b, v8.4b[1]\n"
- "str q31, [c_ptr5], #0x10\n"
+ "add c_ptr4, c_ptr4, #0x10\n"
+ ".inst 0x4f89e25d // sdot v29.4s, v18.16b, v9.4b[0]\n"
+ "str q31, [c_ptr5]\n"
"movi v31.4s, #0\n"
- ".word 0x4f90e31e // sdot v30.4s, v24.16b, v16.4b[0]\n"
- ".word 0x4face33d // sdot v29.4s, v25.16b, v12.4b[1]\n"
- ".word 0x4f94e31f // sdot v31.4s, v24.16b, v20.4b[0]\n"
- "ldr q24, [%[b_ptr0]]\n"
- ".word 0x4fb0e33e // sdot v30.4s, v25.16b, v16.4b[1]\n"
- ".word 0x4fb4e33f // sdot v31.4s, v25.16b, v20.4b[1]\n"
- "ldr q25, [%[b_ptr0], #0x10]\n"
- ".word 0x4f80eb1a // sdot v26.4s, v24.16b, v0.4b[2]\n"
- "add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f84eb1b // sdot v27.4s, v24.16b, v4.4b[2]\n"
- ".word 0x4f88eb1c // sdot v28.4s, v24.16b, v8.4b[2]\n"
- ".word 0x4f8ceb1d // sdot v29.4s, v24.16b, v12.4b[2]\n"
- ".word 0x4f90eb1e // sdot v30.4s, v24.16b, v16.4b[2]\n"
- ".word 0x4f94eb1f // sdot v31.4s, v24.16b, v20.4b[2]\n"
- "ldr q24, [%[b_ptr0]]\n"
- ".word 0x4fa0eb3a // sdot v26.4s, v25.16b, v0.4b[3]\n"
- ".word 0x4fa4eb3b // sdot v27.4s, v25.16b, v4.4b[3]\n"
- ".word 0x4fa8eb3c // sdot v28.4s, v25.16b, v8.4b[3]\n"
- ".word 0x4faceb3d // sdot v29.4s, v25.16b, v12.4b[3]\n"
- ".word 0x4fb0eb3e // sdot v30.4s, v25.16b, v16.4b[3]\n"
- ".word 0x4fb4eb3f // sdot v31.4s, v25.16b, v20.4b[3]\n"
- "ldr q25, [%[b_ptr0], #0x10]\n"
- ".word 0x4f81e31a // sdot v26.4s, v24.16b, v1.4b[0]\n"
- "add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f85e31b // sdot v27.4s, v24.16b, v5.4b[0]\n"
- ".word 0x4f89e31c // sdot v28.4s, v24.16b, v9.4b[0]\n"
- ".word 0x4f8de31d // sdot v29.4s, v24.16b, v13.4b[0]\n"
- ".word 0x4f91e31e // sdot v30.4s, v24.16b, v17.4b[0]\n"
- ".word 0x4f95e31f // sdot v31.4s, v24.16b, v21.4b[0]\n"
- "ldr q24, [%[b_ptr0]]\n"
- ".word 0x4fa1e33a // sdot v26.4s, v25.16b, v1.4b[1]\n"
- ".word 0x4fa5e33b // sdot v27.4s, v25.16b, v5.4b[1]\n"
- ".word 0x4fa9e33c // sdot v28.4s, v25.16b, v9.4b[1]\n"
- ".word 0x4fade33d // sdot v29.4s, v25.16b, v13.4b[1]\n"
- ".word 0x4fb1e33e // sdot v30.4s, v25.16b, v17.4b[1]\n"
- ".word 0x4fb5e33f // sdot v31.4s, v25.16b, v21.4b[1]\n"
- "ldr q25, [%[b_ptr0], #0x10]\n"
- ".word 0x4f81eb1a // sdot v26.4s, v24.16b, v1.4b[2]\n"
- "add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f85eb1b // sdot v27.4s, v24.16b, v5.4b[2]\n"
- ".word 0x4f89eb1c // sdot v28.4s, v24.16b, v9.4b[2]\n"
- ".word 0x4f8deb1d // sdot v29.4s, v24.16b, v13.4b[2]\n"
- ".word 0x4f91eb1e // sdot v30.4s, v24.16b, v17.4b[2]\n"
- ".word 0x4f95eb1f // sdot v31.4s, v24.16b, v21.4b[2]\n"
- "ldr q24, [%[b_ptr0]]\n"
- ".word 0x4fa1eb3a // sdot v26.4s, v25.16b, v1.4b[3]\n"
- ".word 0x4fa5eb3b // sdot v27.4s, v25.16b, v5.4b[3]\n"
- ".word 0x4fa9eb3c // sdot v28.4s, v25.16b, v9.4b[3]\n"
- ".word 0x4fadeb3d // sdot v29.4s, v25.16b, v13.4b[3]\n"
- ".word 0x4fb1eb3e // sdot v30.4s, v25.16b, v17.4b[3]\n"
- ".word 0x4fb5eb3f // sdot v31.4s, v25.16b, v21.4b[3]\n"
- "ldr q25, [%[b_ptr0], #0x10]\n"
- ".word 0x4f82e31a // sdot v26.4s, v24.16b, v2.4b[0]\n"
- "add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f86e31b // sdot v27.4s, v24.16b, v6.4b[0]\n"
- ".word 0x4f8ae31c // sdot v28.4s, v24.16b, v10.4b[0]\n"
- ".word 0x4f8ee31d // sdot v29.4s, v24.16b, v14.4b[0]\n"
- ".word 0x4f92e31e // sdot v30.4s, v24.16b, v18.4b[0]\n"
- ".word 0x4f96e31f // sdot v31.4s, v24.16b, v22.4b[0]\n"
- ".word 0x4fa2e33a // sdot v26.4s, v25.16b, v2.4b[1]\n"
- ".word 0x4fa6e33b // sdot v27.4s, v25.16b, v6.4b[1]\n"
- ".word 0x4faae33c // sdot v28.4s, v25.16b, v10.4b[1]\n"
- ".word 0x4faee33d // sdot v29.4s, v25.16b, v14.4b[1]\n"
- ".word 0x4fb2e33e // sdot v30.4s, v25.16b, v18.4b[1]\n"
- ".word 0x4fb6e33f // sdot v31.4s, v25.16b, v22.4b[1]\n"
+ "add c_ptr5, c_ptr5, #0x10\n"
+ ".inst 0x4f8ce25e // sdot v30.4s, v18.16b, v12.4b[0]\n"
+ ".inst 0x4fa0e27a // sdot v26.4s, v19.16b, v0.4b[1]\n"
+ ".inst 0x4f8fe25f // sdot v31.4s, v18.16b, v15.4b[0]\n"
+ "ldr q18, [%[b_ptr0]]\n"
+ ".inst 0x4fa3e27b // sdot v27.4s, v19.16b, v3.4b[1]\n"
+ ".inst 0x4fa6e27c // sdot v28.4s, v19.16b, v6.4b[1]\n"
+ ".inst 0x4fa9e27d // sdot v29.4s, v19.16b, v9.4b[1]\n"
+ ".inst 0x4face27e // sdot v30.4s, v19.16b, v12.4b[1]\n"
+ ".inst 0x4fafe27f // sdot v31.4s, v19.16b, v15.4b[1]\n"
+ "ldr q19, [%[b_ptr0], #0x10]\n"
+ ".inst 0x4f80ea9a // sdot v26.4s, v20.16b, v0.4b[2]\n"
+ "add %[b_ptr0], %[b_ptr0], #0x20\n"
+ ".inst 0x4f83ea9b // sdot v27.4s, v20.16b, v3.4b[2]\n"
+ ".inst 0x4f86ea9c // sdot v28.4s, v20.16b, v6.4b[2]\n"
+ ".inst 0x4f89ea9d // sdot v29.4s, v20.16b, v9.4b[2]\n"
+ ".inst 0x4f8cea9e // sdot v30.4s, v20.16b, v12.4b[2]\n"
+ ".inst 0x4f8fea9f // sdot v31.4s, v20.16b, v15.4b[2]\n"
+ ".inst 0x4fa0eaba // sdot v26.4s, v21.16b, v0.4b[3]\n"
+ ".inst 0x4fa3eabb // sdot v27.4s, v21.16b, v3.4b[3]\n"
+ ".inst 0x4fa6eabc // sdot v28.4s, v21.16b, v6.4b[3]\n"
+ ".inst 0x4fa9eabd // sdot v29.4s, v21.16b, v9.4b[3]\n"
+ ".inst 0x4faceabe // sdot v30.4s, v21.16b, v12.4b[3]\n"
+ ".inst 0x4fafeabf // sdot v31.4s, v21.16b, v15.4b[3]\n"
+ ".inst 0x4f81e2da // sdot v26.4s, v22.16b, v1.4b[0]\n"
+ ".inst 0x4f84e2db // sdot v27.4s, v22.16b, v4.4b[0]\n"
+ ".inst 0x4f87e2dc // sdot v28.4s, v22.16b, v7.4b[0]\n"
+ ".inst 0x4f8ae2dd // sdot v29.4s, v22.16b, v10.4b[0]\n"
+ ".inst 0x4f8de2de // sdot v30.4s, v22.16b, v13.4b[0]\n"
+ ".inst 0x4f90e2df // sdot v31.4s, v22.16b, v16.4b[0]\n"
+ ".inst 0x4fa1e2fa // sdot v26.4s, v23.16b, v1.4b[1]\n"
+ ".inst 0x4fa4e2fb // sdot v27.4s, v23.16b, v4.4b[1]\n"
+ ".inst 0x4fa7e2fc // sdot v28.4s, v23.16b, v7.4b[1]\n"
+ ".inst 0x4faae2fd // sdot v29.4s, v23.16b, v10.4b[1]\n"
+ ".inst 0x4fade2fe // sdot v30.4s, v23.16b, v13.4b[1]\n"
+ ".inst 0x4fb0e2ff // sdot v31.4s, v23.16b, v16.4b[1]\n"
+ ".inst 0x4f81eb1a // sdot v26.4s, v24.16b, v1.4b[2]\n"
+ ".inst 0x4f84eb1b // sdot v27.4s, v24.16b, v4.4b[2]\n"
+ ".inst 0x4f87eb1c // sdot v28.4s, v24.16b, v7.4b[2]\n"
+ ".inst 0x4f8aeb1d // sdot v29.4s, v24.16b, v10.4b[2]\n"
+ ".inst 0x4f8deb1e // sdot v30.4s, v24.16b, v13.4b[2]\n"
+ ".inst 0x4f90eb1f // sdot v31.4s, v24.16b, v16.4b[2]\n"
+ ".inst 0x4fa1eb3a // sdot v26.4s, v25.16b, v1.4b[3]\n"
+ ".inst 0x4fa4eb3b // sdot v27.4s, v25.16b, v4.4b[3]\n"
+ ".inst 0x4fa7eb3c // sdot v28.4s, v25.16b, v7.4b[3]\n"
+ ".inst 0x4faaeb3d // sdot v29.4s, v25.16b, v10.4b[3]\n"
+ ".inst 0x4fadeb3e // sdot v30.4s, v25.16b, v13.4b[3]\n"
+ ".inst 0x4fb0eb3f // sdot v31.4s, v25.16b, v16.4b[3]\n"
+ ".inst 0x4f82e25a // sdot v26.4s, v18.16b, v2.4b[0]\n"
+ ".inst 0x4f85e25b // sdot v27.4s, v18.16b, v5.4b[0]\n"
+ ".inst 0x4f88e25c // sdot v28.4s, v18.16b, v8.4b[0]\n"
+ ".inst 0x4f8be25d // sdot v29.4s, v18.16b, v11.4b[0]\n"
+ ".inst 0x4f8ee25e // sdot v30.4s, v18.16b, v14.4b[0]\n"
+ ".inst 0x4f91e25f // sdot v31.4s, v18.16b, v17.4b[0]\n"
+ ".inst 0x4fa2e27a // sdot v26.4s, v19.16b, v2.4b[1]\n"
+ ".inst 0x4fa5e27b // sdot v27.4s, v19.16b, v5.4b[1]\n"
+ ".inst 0x4fa8e27c // sdot v28.4s, v19.16b, v8.4b[1]\n"
+ ".inst 0x4fabe27d // sdot v29.4s, v19.16b, v11.4b[1]\n"
+ ".inst 0x4faee27e // sdot v30.4s, v19.16b, v14.4b[1]\n"
+ ".inst 0x4fb1e27f // sdot v31.4s, v19.16b, v17.4b[1]\n"
"6:\n"
"str q26, [%[c_ptr0]]\n"
+ "add %[c_ptr0], %[c_ptr0], #0x10\n"
"str q27, [c_ptr1]\n"
+ "add c_ptr1, c_ptr1, #0x10\n"
"str q28, [c_ptr2]\n"
+ "add c_ptr2, c_ptr2, #0x10\n"
"str q29, [c_ptr3]\n"
+ "add c_ptr3, c_ptr3, #0x10\n"
"str q30, [c_ptr4]\n"
+ "add c_ptr4, c_ptr4, #0x10\n"
"str q31, [c_ptr5]\n"
+ "add c_ptr5, c_ptr5, #0x10\n"
".unreq a_ptr1\n"
".unreq a_ptr2\n"
".unreq a_ptr3\n"
@@ -918,390 +961,408 @@ void a64_smallK_hybrid_s8s32_dot_4x6_a55(const int8_t *A, int lda, const int8_t
"add a_ptr1, %[a_ptr0], #0x0\n"
"1:\n"
"ldr q0, [%[a_ptr0]], #0x10\n"
- "ldr q4, [a_ptr1], #0x10\n"
- "ldr q8, [a_ptr2], #0x10\n"
- "ldr q12, [a_ptr3], #0x10\n"
- "ldr q16, [a_ptr4], #0x10\n"
- "ldr q20, [a_ptr5], #0x10\n"
+ "ldr q3, [a_ptr1], #0x10\n"
+ "ldr q6, [a_ptr2], #0x10\n"
+ "ldr q9, [a_ptr3], #0x10\n"
+ "ldr q12, [a_ptr4], #0x10\n"
+ "ldr q15, [a_ptr5], #0x10\n"
"ldr q1, [%[a_ptr0]], #0x10\n"
- "ldr q5, [a_ptr1], #0x10\n"
- "ldr q9, [a_ptr2], #0x10\n"
- "ldr q13, [a_ptr3], #0x10\n"
+ "ldr q4, [a_ptr1], #0x10\n"
+ "ldr q7, [a_ptr2], #0x10\n"
+ "ldr q10, [a_ptr3], #0x10\n"
"ldr d2, [%[a_ptr0]], #0x8\n"
- "ldr q17, [a_ptr4], #0x10\n"
- "ldr d6, [a_ptr1], #0x8\n"
- "ldr q21, [a_ptr5], #0x10\n"
- "ldr d10, [a_ptr2], #0x8\n"
- "ldr d14, [a_ptr3], #0x8\n"
- "ldr d18, [a_ptr4], #0x8\n"
- "ldr d22, [a_ptr5], #0x8\n"
+ "ldr q13, [a_ptr4], #0x10\n"
+ "ldr d5, [a_ptr1], #0x8\n"
+ "ldr q16, [a_ptr5], #0x10\n"
+ "ldr d8, [a_ptr2], #0x8\n"
+ "ldr d11, [a_ptr3], #0x8\n"
+ "ldr d14, [a_ptr4], #0x8\n"
+ "ldr d17, [a_ptr5], #0x8\n"
"cbnz %[odds], 2f\n"
"ld1 {v2.s}[2], [%[a_ptr0]]\n"
- "ld1 {v6.s}[2], [a_ptr1]\n"
- "ld1 {v10.s}[2], [a_ptr2]\n"
- "ld1 {v14.s}[2], [a_ptr3]\n"
- "ld1 {v18.s}[2], [a_ptr4]\n"
- "ld1 {v22.s}[2], [a_ptr5]\n"
+ "ld1 {v5.s}[2], [a_ptr1]\n"
+ "ld1 {v8.s}[2], [a_ptr2]\n"
+ "ld1 {v11.s}[2], [a_ptr3]\n"
+ "ld1 {v14.s}[2], [a_ptr4]\n"
+ "ld1 {v17.s}[2], [a_ptr5]\n"
"b 3f\n"
"2:\n"
"subs %[odds], %[odds], #0x1\n"
"b.ne 4f\n"
"ld1 {v2.b}[8], [%[a_ptr0]]\n"
- "ld1 {v6.b}[8], [a_ptr1]\n"
- "ld1 {v10.b}[8], [a_ptr2]\n"
- "ld1 {v14.b}[8], [a_ptr3]\n"
- "ld1 {v18.b}[8], [a_ptr4]\n"
- "ld1 {v22.b}[8], [a_ptr5]\n"
+ "ld1 {v5.b}[8], [a_ptr1]\n"
+ "ld1 {v8.b}[8], [a_ptr2]\n"
+ "ld1 {v11.b}[8], [a_ptr3]\n"
+ "ld1 {v14.b}[8], [a_ptr4]\n"
+ "ld1 {v17.b}[8], [a_ptr5]\n"
"b 3f\n"
"4:\n"
"ld1 {v2.h}[4], [%[a_ptr0]], #2\n"
- "ld1 {v6.h}[4], [a_ptr1], #2\n"
- "ld1 {v10.h}[4], [a_ptr2], #2\n"
- "ld1 {v14.h}[4], [a_ptr3], #2\n"
- "ld1 {v18.h}[4], [a_ptr4], #2\n"
- "ld1 {v22.h}[4], [a_ptr5], #2\n"
+ "ld1 {v5.h}[4], [a_ptr1], #2\n"
+ "ld1 {v8.h}[4], [a_ptr2], #2\n"
+ "ld1 {v11.h}[4], [a_ptr3], #2\n"
+ "ld1 {v14.h}[4], [a_ptr4], #2\n"
+ "ld1 {v17.h}[4], [a_ptr5], #2\n"
"subs %[odds], %[odds], #0x1\n"
"b.ne 5f\n"
"b 3f\n"
"5:\n"
"ld1 {v2.b}[10], [%[a_ptr0]]\n"
- "ld1 {v6.b}[10], [a_ptr1]\n"
- "ld1 {v10.b}[10], [a_ptr2]\n"
- "ld1 {v14.b}[10], [a_ptr3]\n"
- "ld1 {v18.b}[10], [a_ptr4]\n"
- "ld1 {v22.b}[10], [a_ptr5]\n"
+ "ld1 {v5.b}[10], [a_ptr1]\n"
+ "ld1 {v8.b}[10], [a_ptr2]\n"
+ "ld1 {v11.b}[10], [a_ptr3]\n"
+ "ld1 {v14.b}[10], [a_ptr4]\n"
+ "ld1 {v17.b}[10], [a_ptr5]\n"
"3:\n"
"movi v26.4s, #0\n"
- "ldr q24, [%[b_ptr0]]\n"
+ "ldr q18, [%[b_ptr0]]\n"
"movi v27.4s, #0\n"
- "ldr q25, [%[b_ptr0], #0x10]\n"
+ "ldr q19, [%[b_ptr0], #0x10]\n"
"movi v28.4s, #0\n"
- "prfm PLDL1KEEP, [a_ptr5, #0x40]\n"
+ "ldr q20, [%[b_ptr0], #0x20]\n"
"movi v29.4s, #0\n"
- "prfm PLDL1KEEP, [a_ptr5, #0x80]\n"
+ "ldr q21, [%[b_ptr0], #0x30]\n"
"movi v30.4s, #0\n"
- "prfm PLDL1KEEP, [a_ptr5, #0xc0]\n"
+ "ldr q22, [%[b_ptr0], #0x40]\n"
"movi v31.4s, #0\n"
+ "ldr q23, [%[b_ptr0], #0x50]\n"
+ ".inst 0x4f80e25a // sdot v26.4s, v18.16b, v0.4b[0]\n"
+ "ldr q24, [%[b_ptr0], #0x60]\n"
+ ".inst 0x4f83e25b // sdot v27.4s, v18.16b, v3.4b[0]\n"
+ "ldr q25, [%[b_ptr0], #0x70]\n"
+ ".inst 0x4f86e25c // sdot v28.4s, v18.16b, v6.4b[0]\n"
+ "prfm PLDL1KEEP, [a_ptr5, #0x40]\n"
+ ".inst 0x4f89e25d // sdot v29.4s, v18.16b, v9.4b[0]\n"
+ "prfm PLDL1KEEP, [a_ptr5, #0x80]\n"
+ ".inst 0x4f8ce25e // sdot v30.4s, v18.16b, v12.4b[0]\n"
+ "prfm PLDL1KEEP, [a_ptr5, #0xc0]\n"
+ ".inst 0x4f8fe25f // sdot v31.4s, v18.16b, v15.4b[0]\n"
"prfm PLDL1KEEP, [a_ptr5, #0x100]\n"
- ".word 0x4f80e31a // sdot v26.4s, v24.16b, v0.4b[0]\n"
+ ".inst 0x4fa0e27a // sdot v26.4s, v19.16b, v0.4b[1]\n"
"prfm PLDL1KEEP, [a_ptr5, #0x140]\n"
- ".word 0x4f84e31b // sdot v27.4s, v24.16b, v4.4b[0]\n"
+ ".inst 0x4fa3e27b // sdot v27.4s, v19.16b, v3.4b[1]\n"
"prfm PLDL1KEEP, [a_ptr5, #0x180]\n"
- ".word 0x4f88e31c // sdot v28.4s, v24.16b, v8.4b[0]\n"
- "add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f8ce31d // sdot v29.4s, v24.16b, v12.4b[0]\n"
- ".word 0x4f90e31e // sdot v30.4s, v24.16b, v16.4b[0]\n"
- ".word 0x4f94e31f // sdot v31.4s, v24.16b, v20.4b[0]\n"
- "ldr q24, [%[b_ptr0]]\n"
- ".word 0x4fa0e33a // sdot v26.4s, v25.16b, v0.4b[1]\n"
- ".word 0x4fa4e33b // sdot v27.4s, v25.16b, v4.4b[1]\n"
- ".word 0x4fa8e33c // sdot v28.4s, v25.16b, v8.4b[1]\n"
- ".word 0x4face33d // sdot v29.4s, v25.16b, v12.4b[1]\n"
- ".word 0x4fb0e33e // sdot v30.4s, v25.16b, v16.4b[1]\n"
- ".word 0x4fb4e33f // sdot v31.4s, v25.16b, v20.4b[1]\n"
- "ldr q25, [%[b_ptr0], #0x10]\n"
- ".word 0x4f80eb1a // sdot v26.4s, v24.16b, v0.4b[2]\n"
- "add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f84eb1b // sdot v27.4s, v24.16b, v4.4b[2]\n"
- ".word 0x4f88eb1c // sdot v28.4s, v24.16b, v8.4b[2]\n"
- ".word 0x4f8ceb1d // sdot v29.4s, v24.16b, v12.4b[2]\n"
- ".word 0x4f90eb1e // sdot v30.4s, v24.16b, v16.4b[2]\n"
- ".word 0x4f94eb1f // sdot v31.4s, v24.16b, v20.4b[2]\n"
- "ldr q24, [%[b_ptr0]]\n"
- ".word 0x4fa0eb3a // sdot v26.4s, v25.16b, v0.4b[3]\n"
- ".word 0x4fa4eb3b // sdot v27.4s, v25.16b, v4.4b[3]\n"
- ".word 0x4fa8eb3c // sdot v28.4s, v25.16b, v8.4b[3]\n"
- ".word 0x4faceb3d // sdot v29.4s, v25.16b, v12.4b[3]\n"
- ".word 0x4fb0eb3e // sdot v30.4s, v25.16b, v16.4b[3]\n"
- ".word 0x4fb4eb3f // sdot v31.4s, v25.16b, v20.4b[3]\n"
- "ldr q25, [%[b_ptr0], #0x10]\n"
- ".word 0x4f81e31a // sdot v26.4s, v24.16b, v1.4b[0]\n"
- "add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f85e31b // sdot v27.4s, v24.16b, v5.4b[0]\n"
- ".word 0x4f89e31c // sdot v28.4s, v24.16b, v9.4b[0]\n"
- ".word 0x4f8de31d // sdot v29.4s, v24.16b, v13.4b[0]\n"
- ".word 0x4f91e31e // sdot v30.4s, v24.16b, v17.4b[0]\n"
- ".word 0x4f95e31f // sdot v31.4s, v24.16b, v21.4b[0]\n"
- "ldr q24, [%[b_ptr0]]\n"
- ".word 0x4fa1e33a // sdot v26.4s, v25.16b, v1.4b[1]\n"
- ".word 0x4fa5e33b // sdot v27.4s, v25.16b, v5.4b[1]\n"
- ".word 0x4fa9e33c // sdot v28.4s, v25.16b, v9.4b[1]\n"
- ".word 0x4fade33d // sdot v29.4s, v25.16b, v13.4b[1]\n"
- ".word 0x4fb1e33e // sdot v30.4s, v25.16b, v17.4b[1]\n"
- ".word 0x4fb5e33f // sdot v31.4s, v25.16b, v21.4b[1]\n"
- "ldr q25, [%[b_ptr0], #0x10]\n"
- ".word 0x4f81eb1a // sdot v26.4s, v24.16b, v1.4b[2]\n"
- "add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f85eb1b // sdot v27.4s, v24.16b, v5.4b[2]\n"
- ".word 0x4f89eb1c // sdot v28.4s, v24.16b, v9.4b[2]\n"
- ".word 0x4f8deb1d // sdot v29.4s, v24.16b, v13.4b[2]\n"
- ".word 0x4f91eb1e // sdot v30.4s, v24.16b, v17.4b[2]\n"
- ".word 0x4f95eb1f // sdot v31.4s, v24.16b, v21.4b[2]\n"
- "ldr q24, [%[b_ptr0]]\n"
- ".word 0x4fa1eb3a // sdot v26.4s, v25.16b, v1.4b[3]\n"
- ".word 0x4fa5eb3b // sdot v27.4s, v25.16b, v5.4b[3]\n"
- ".word 0x4fa9eb3c // sdot v28.4s, v25.16b, v9.4b[3]\n"
- ".word 0x4fadeb3d // sdot v29.4s, v25.16b, v13.4b[3]\n"
- ".word 0x4fb1eb3e // sdot v30.4s, v25.16b, v17.4b[3]\n"
- ".word 0x4fb5eb3f // sdot v31.4s, v25.16b, v21.4b[3]\n"
- "ldr q25, [%[b_ptr0], #0x10]\n"
- ".word 0x4f82e31a // sdot v26.4s, v24.16b, v2.4b[0]\n"
- "add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f86e31b // sdot v27.4s, v24.16b, v6.4b[0]\n"
- ".word 0x4f8ae31c // sdot v28.4s, v24.16b, v10.4b[0]\n"
- ".word 0x4f8ee31d // sdot v29.4s, v24.16b, v14.4b[0]\n"
- ".word 0x4f92e31e // sdot v30.4s, v24.16b, v18.4b[0]\n"
- ".word 0x4f96e31f // sdot v31.4s, v24.16b, v22.4b[0]\n"
- "ldr q24, [%[b_ptr0]]\n"
- ".word 0x4fa2e33a // sdot v26.4s, v25.16b, v2.4b[1]\n"
- "add %[b_ptr0], %[b_ptr0], #0x10\n"
- ".word 0x4fa6e33b // sdot v27.4s, v25.16b, v6.4b[1]\n"
- ".word 0x4faae33c // sdot v28.4s, v25.16b, v10.4b[1]\n"
- ".word 0x4faee33d // sdot v29.4s, v25.16b, v14.4b[1]\n"
- ".word 0x4fb2e33e // sdot v30.4s, v25.16b, v18.4b[1]\n"
- ".word 0x4fb6e33f // sdot v31.4s, v25.16b, v22.4b[1]\n"
- ".word 0x4f82eb1a // sdot v26.4s, v24.16b, v2.4b[2]\n"
- ".word 0x4f86eb1b // sdot v27.4s, v24.16b, v6.4b[2]\n"
- ".word 0x4f8aeb1c // sdot v28.4s, v24.16b, v10.4b[2]\n"
- ".word 0x4f8eeb1d // sdot v29.4s, v24.16b, v14.4b[2]\n"
- ".word 0x4f92eb1e // sdot v30.4s, v24.16b, v18.4b[2]\n"
- ".word 0x4f96eb1f // sdot v31.4s, v24.16b, v22.4b[2]\n"
+ ".inst 0x4fa6e27c // sdot v28.4s, v19.16b, v6.4b[1]\n"
+ "add %[b_ptr0], %[b_ptr0], #0x80\n"
+ ".inst 0x4fa9e27d // sdot v29.4s, v19.16b, v9.4b[1]\n"
+ "ldr q18, [%[b_ptr0]]\n"
+ ".inst 0x4face27e // sdot v30.4s, v19.16b, v12.4b[1]\n"
+ ".inst 0x4fafe27f // sdot v31.4s, v19.16b, v15.4b[1]\n"
+ "ldr q19, [%[b_ptr0], #0x10]\n"
+ ".inst 0x4f80ea9a // sdot v26.4s, v20.16b, v0.4b[2]\n"
+ ".inst 0x4f83ea9b // sdot v27.4s, v20.16b, v3.4b[2]\n"
+ ".inst 0x4f86ea9c // sdot v28.4s, v20.16b, v6.4b[2]\n"
+ ".inst 0x4f89ea9d // sdot v29.4s, v20.16b, v9.4b[2]\n"
+ ".inst 0x4f8cea9e // sdot v30.4s, v20.16b, v12.4b[2]\n"
+ ".inst 0x4f8fea9f // sdot v31.4s, v20.16b, v15.4b[2]\n"
+ "ldr q20, [%[b_ptr0], #0x20]\n"
+ ".inst 0x4fa0eaba // sdot v26.4s, v21.16b, v0.4b[3]\n"
+ "add %[b_ptr0], %[b_ptr0], #0x30\n"
+ ".inst 0x4fa3eabb // sdot v27.4s, v21.16b, v3.4b[3]\n"
+ ".inst 0x4fa6eabc // sdot v28.4s, v21.16b, v6.4b[3]\n"
+ ".inst 0x4fa9eabd // sdot v29.4s, v21.16b, v9.4b[3]\n"
+ ".inst 0x4faceabe // sdot v30.4s, v21.16b, v12.4b[3]\n"
+ ".inst 0x4fafeabf // sdot v31.4s, v21.16b, v15.4b[3]\n"
+ ".inst 0x4f81e2da // sdot v26.4s, v22.16b, v1.4b[0]\n"
+ ".inst 0x4f84e2db // sdot v27.4s, v22.16b, v4.4b[0]\n"
+ ".inst 0x4f87e2dc // sdot v28.4s, v22.16b, v7.4b[0]\n"
+ ".inst 0x4f8ae2dd // sdot v29.4s, v22.16b, v10.4b[0]\n"
+ ".inst 0x4f8de2de // sdot v30.4s, v22.16b, v13.4b[0]\n"
+ ".inst 0x4f90e2df // sdot v31.4s, v22.16b, v16.4b[0]\n"
+ ".inst 0x4fa1e2fa // sdot v26.4s, v23.16b, v1.4b[1]\n"
+ ".inst 0x4fa4e2fb // sdot v27.4s, v23.16b, v4.4b[1]\n"
+ ".inst 0x4fa7e2fc // sdot v28.4s, v23.16b, v7.4b[1]\n"
+ ".inst 0x4faae2fd // sdot v29.4s, v23.16b, v10.4b[1]\n"
+ ".inst 0x4fade2fe // sdot v30.4s, v23.16b, v13.4b[1]\n"
+ ".inst 0x4fb0e2ff // sdot v31.4s, v23.16b, v16.4b[1]\n"
+ ".inst 0x4f81eb1a // sdot v26.4s, v24.16b, v1.4b[2]\n"
+ ".inst 0x4f84eb1b // sdot v27.4s, v24.16b, v4.4b[2]\n"
+ ".inst 0x4f87eb1c // sdot v28.4s, v24.16b, v7.4b[2]\n"
+ ".inst 0x4f8aeb1d // sdot v29.4s, v24.16b, v10.4b[2]\n"
+ ".inst 0x4f8deb1e // sdot v30.4s, v24.16b, v13.4b[2]\n"
+ ".inst 0x4f90eb1f // sdot v31.4s, v24.16b, v16.4b[2]\n"
+ ".inst 0x4fa1eb3a // sdot v26.4s, v25.16b, v1.4b[3]\n"
+ ".inst 0x4fa4eb3b // sdot v27.4s, v25.16b, v4.4b[3]\n"
+ ".inst 0x4fa7eb3c // sdot v28.4s, v25.16b, v7.4b[3]\n"
+ ".inst 0x4faaeb3d // sdot v29.4s, v25.16b, v10.4b[3]\n"
+ ".inst 0x4fadeb3e // sdot v30.4s, v25.16b, v13.4b[3]\n"
+ ".inst 0x4fb0eb3f // sdot v31.4s, v25.16b, v16.4b[3]\n"
+ ".inst 0x4f82e25a // sdot v26.4s, v18.16b, v2.4b[0]\n"
+ ".inst 0x4f85e25b // sdot v27.4s, v18.16b, v5.4b[0]\n"
+ ".inst 0x4f88e25c // sdot v28.4s, v18.16b, v8.4b[0]\n"
+ ".inst 0x4f8be25d // sdot v29.4s, v18.16b, v11.4b[0]\n"
+ ".inst 0x4f8ee25e // sdot v30.4s, v18.16b, v14.4b[0]\n"
+ ".inst 0x4f91e25f // sdot v31.4s, v18.16b, v17.4b[0]\n"
+ ".inst 0x4fa2e27a // sdot v26.4s, v19.16b, v2.4b[1]\n"
+ ".inst 0x4fa5e27b // sdot v27.4s, v19.16b, v5.4b[1]\n"
+ ".inst 0x4fa8e27c // sdot v28.4s, v19.16b, v8.4b[1]\n"
+ ".inst 0x4fabe27d // sdot v29.4s, v19.16b, v11.4b[1]\n"
+ ".inst 0x4faee27e // sdot v30.4s, v19.16b, v14.4b[1]\n"
+ ".inst 0x4fb1e27f // sdot v31.4s, v19.16b, v17.4b[1]\n"
+ ".inst 0x4f82ea9a // sdot v26.4s, v20.16b, v2.4b[2]\n"
+ ".inst 0x4f85ea9b // sdot v27.4s, v20.16b, v5.4b[2]\n"
+ ".inst 0x4f88ea9c // sdot v28.4s, v20.16b, v8.4b[2]\n"
+ ".inst 0x4f8bea9d // sdot v29.4s, v20.16b, v11.4b[2]\n"
+ ".inst 0x4f8eea9e // sdot v30.4s, v20.16b, v14.4b[2]\n"
+ ".inst 0x4f91ea9f // sdot v31.4s, v20.16b, v17.4b[2]\n"
"cbz %[loops], 6f\n"
- "ldr d24, [%[b_ptr0]]\n"
+ "ldr d18, [%[b_ptr0]]\n"
"subs %[loops], %[loops], #0x1\n"
- "ldr temploadreg0, [%[b_ptr0], #0x8]\n"
- "ldr d25, [%[b_ptr0], #0x10]\n"
- "ldr temploadreg1, [%[b_ptr0], #0x18]\n"
- "add %[b_ptr0], %[b_ptr0], #0x20\n"
+ "ldr temploadreg2, [%[b_ptr0], #0x8]\n"
+ "ldr d19, [%[b_ptr0], #0x10]\n"
+ "ldr temploadreg3, [%[b_ptr0], #0x18]\n"
+ "ldr d20, [%[b_ptr0], #0x20]\n"
+ "ldr temploadreg0, [%[b_ptr0], #0x28]\n"
+ "ldr d21, [%[b_ptr0], #0x30]\n"
+ "ldr temploadreg1, [%[b_ptr0], #0x38]\n"
+ "ldr d22, [%[b_ptr0], #0x40]\n"
+ "ins v18.d[1], temploadreg2\n"
+ "ldr temploadreg2, [%[b_ptr0], #0x48]\n"
+ "ldr d23, [%[b_ptr0], #0x50]\n"
+ "ins v19.d[1], temploadreg3\n"
+ "ldr temploadreg3, [%[b_ptr0], #0x58]\n"
+ "ldr d24, [%[b_ptr0], #0x60]\n"
+ "ins v20.d[1], temploadreg0\n"
+ "ldr temploadreg0, [%[b_ptr0], #0x68]\n"
+ "ldr d25, [%[b_ptr0], #0x70]\n"
+ "ins v21.d[1], temploadreg1\n"
+ "ldr temploadreg1, [%[b_ptr0], #0x78]\n"
+ "add %[b_ptr0], %[b_ptr0], #0x80\n"
+ "ins v22.d[1], temploadreg2\n"
+ "ins v23.d[1], temploadreg3\n"
"ins v24.d[1], temploadreg0\n"
"ins v25.d[1], temploadreg1\n"
"b.eq 7f\n"
"8:\n"
- "str q26, [%[c_ptr0]], #0x10\n"
+ "str q26, [%[c_ptr0]]\n"
"subs %[loops], %[loops], #0x1\n"
"movi v26.4s, #0\n"
- "ldr temploadreg0, [%[b_ptr0], #0x8]\n"
- "ldr temploadreg1, [%[b_ptr0], #0x18]\n"
- "prfm PSTL1KEEP, [%[c_ptr0], #0x40]\n"
- "str q27, [c_ptr1], #0x10\n"
+ "ldr temploadreg2, [%[b_ptr0], #0x8]\n"
+ "ldr temploadreg3, [%[b_ptr0], #0x18]\n"
+ "add %[c_ptr0], %[c_ptr0], #0x10\n"
+ "str q27, [c_ptr1]\n"
+ "add c_ptr1, c_ptr1, #0x10\n"
"movi v27.4s, #0\n"
- ".word 0x4f80e31a // sdot v26.4s, v24.16b, v0.4b[0]\n"
- "str q28, [c_ptr2], #0x10\n"
- "prfm PSTL1KEEP, [c_ptr1, #0x40]\n"
+ "ldr temploadreg0, [%[b_ptr0], #0x28]\n"
+ ".inst 0x4f80e25a // sdot v26.4s, v18.16b, v0.4b[0]\n"
+ "str q28, [c_ptr2]\n"
"movi v28.4s, #0\n"
- ".word 0x4f84e31b // sdot v27.4s, v24.16b, v4.4b[0]\n"
- "str q29, [c_ptr3], #0x10\n"
+ "add c_ptr2, c_ptr2, #0x10\n"
+ ".inst 0x4f83e25b // sdot v27.4s, v18.16b, v3.4b[0]\n"
+ "str q29, [c_ptr3]\n"
"movi v29.4s, #0\n"
- "prfm PSTL1KEEP, [c_ptr2, #0x40]\n"
- ".word 0x4f88e31c // sdot v28.4s, v24.16b, v8.4b[0]\n"
- "str q30, [c_ptr4], #0x10\n"
+ "add c_ptr3, c_ptr3, #0x10\n"
+ ".inst 0x4f86e25c // sdot v28.4s, v18.16b, v6.4b[0]\n"
+ "str q30, [c_ptr4]\n"
"movi v30.4s, #0\n"
- "prfm PSTL1KEEP, [c_ptr3, #0x40]\n"
- ".word 0x4f8ce31d // sdot v29.4s, v24.16b, v12.4b[0]\n"
- "str q31, [c_ptr5], #0x10\n"
+ "add c_ptr4, c_ptr4, #0x10\n"
+ ".inst 0x4f89e25d // sdot v29.4s, v18.16b, v9.4b[0]\n"
+ "str q31, [c_ptr5]\n"
"movi v31.4s, #0\n"
+ "add c_ptr5, c_ptr5, #0x10\n"
+ ".inst 0x4f8ce25e // sdot v30.4s, v18.16b, v12.4b[0]\n"
+ "prfm PSTL1KEEP, [%[c_ptr0], #0x40]\n"
+ ".inst 0x4f8fe25f // sdot v31.4s, v18.16b, v15.4b[0]\n"
+ "ldr d18, [%[b_ptr0]]\n"
+ ".inst 0x4fa0e27a // sdot v26.4s, v19.16b, v0.4b[1]\n"
+ "prfm PSTL1KEEP, [c_ptr1, #0x40]\n"
+ ".inst 0x4fa3e27b // sdot v27.4s, v19.16b, v3.4b[1]\n"
+ "ins v18.d[1], temploadreg2\n"
+ ".inst 0x4fa6e27c // sdot v28.4s, v19.16b, v6.4b[1]\n"
+ "prfm PSTL1KEEP, [c_ptr2, #0x40]\n"
+ ".inst 0x4fa9e27d // sdot v29.4s, v19.16b, v9.4b[1]\n"
+ "prfm PSTL1KEEP, [c_ptr3, #0x40]\n"
+ ".inst 0x4face27e // sdot v30.4s, v19.16b, v12.4b[1]\n"
"prfm PSTL1KEEP, [c_ptr4, #0x40]\n"
- ".word 0x4f90e31e // sdot v30.4s, v24.16b, v16.4b[0]\n"
+ ".inst 0x4fafe27f // sdot v31.4s, v19.16b, v15.4b[1]\n"
+ "ldr d19, [%[b_ptr0], #0x10]\n"
+ ".inst 0x4f80ea9a // sdot v26.4s, v20.16b, v0.4b[2]\n"
"prfm PSTL1KEEP, [c_ptr5, #0x40]\n"
- ".word 0x4f94e31f // sdot v31.4s, v24.16b, v20.4b[0]\n"
- "ldr d24, [%[b_ptr0]]\n"
- ".word 0x4fa0e33a // sdot v26.4s, v25.16b, v0.4b[1]\n"
- ".word 0x4fa4e33b // sdot v27.4s, v25.16b, v4.4b[1]\n"
- ".word 0x4fa8e33c // sdot v28.4s, v25.16b, v8.4b[1]\n"
- "ins v24.d[1], temploadreg0\n"
- ".word 0x4face33d // sdot v29.4s, v25.16b, v12.4b[1]\n"
- ".word 0x4fb0e33e // sdot v30.4s, v25.16b, v16.4b[1]\n"
- ".word 0x4fb4e33f // sdot v31.4s, v25.16b, v20.4b[1]\n"
- "ldr d25, [%[b_ptr0], #0x10]\n"
- ".word 0x4f80eb1a // sdot v26.4s, v24.16b, v0.4b[2]\n"
- "add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f84eb1b // sdot v27.4s, v24.16b, v4.4b[2]\n"
- "ins v25.d[1], temploadreg1\n"
- ".word 0x4f88eb1c // sdot v28.4s, v24.16b, v8.4b[2]\n"
- "ldr temploadreg0, [%[b_ptr0], #0x8]\n"
- ".word 0x4f8ceb1d // sdot v29.4s, v24.16b, v12.4b[2]\n"
- "ldr temploadreg1, [%[b_ptr0], #0x18]\n"
- ".word 0x4f90eb1e // sdot v30.4s, v24.16b, v16.4b[2]\n"
- ".word 0x4f94eb1f // sdot v31.4s, v24.16b, v20.4b[2]\n"
- "ldr d24, [%[b_ptr0]]\n"
- ".word 0x4fa0eb3a // sdot v26.4s, v25.16b, v0.4b[3]\n"
- ".word 0x4fa4eb3b // sdot v27.4s, v25.16b, v4.4b[3]\n"
- ".word 0x4fa8eb3c // sdot v28.4s, v25.16b, v8.4b[3]\n"
- "ins v24.d[1], temploadreg0\n"
- ".word 0x4faceb3d // sdot v29.4s, v25.16b, v12.4b[3]\n"
- ".word 0x4fb0eb3e // sdot v30.4s, v25.16b, v16.4b[3]\n"
- ".word 0x4fb4eb3f // sdot v31.4s, v25.16b, v20.4b[3]\n"
- "ldr d25, [%[b_ptr0], #0x10]\n"
- ".word 0x4f81e31a // sdot v26.4s, v24.16b, v1.4b[0]\n"
- "add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f85e31b // sdot v27.4s, v24.16b, v5.4b[0]\n"
+ ".inst 0x4f83ea9b // sdot v27.4s, v20.16b, v3.4b[2]\n"
+ "ins v19.d[1], temploadreg3\n"
+ ".inst 0x4f86ea9c // sdot v28.4s, v20.16b, v6.4b[2]\n"
+ ".inst 0x4f89ea9d // sdot v29.4s, v20.16b, v9.4b[2]\n"
+ ".inst 0x4f8cea9e // sdot v30.4s, v20.16b, v12.4b[2]\n"
+ ".inst 0x4f8fea9f // sdot v31.4s, v20.16b, v15.4b[2]\n"
+ "ldr d20, [%[b_ptr0], #0x20]\n"
+ ".inst 0x4fa0eaba // sdot v26.4s, v21.16b, v0.4b[3]\n"
+ "add %[b_ptr0], %[b_ptr0], #0x30\n"
+ ".inst 0x4fa3eabb // sdot v27.4s, v21.16b, v3.4b[3]\n"
+ "ins v20.d[1], temploadreg0\n"
+ ".inst 0x4fa6eabc // sdot v28.4s, v21.16b, v6.4b[3]\n"
+ "ldr temploadreg2, [%[b_ptr0], #0x8]\n"
+ ".inst 0x4fa9eabd // sdot v29.4s, v21.16b, v9.4b[3]\n"
+ "ldr temploadreg3, [%[b_ptr0], #0x18]\n"
+ ".inst 0x4faceabe // sdot v30.4s, v21.16b, v12.4b[3]\n"
+ "ldr temploadreg0, [%[b_ptr0], #0x28]\n"
+ ".inst 0x4fafeabf // sdot v31.4s, v21.16b, v15.4b[3]\n"
+ "ldr d21, [%[b_ptr0], #0x30]\n"
+ ".inst 0x4f81e2da // sdot v26.4s, v22.16b, v1.4b[0]\n"
+ "ldr temploadreg1, [%[b_ptr0], #0x38]\n"
+ ".inst 0x4f84e2db // sdot v27.4s, v22.16b, v4.4b[0]\n"
+ ".inst 0x4f87e2dc // sdot v28.4s, v22.16b, v7.4b[0]\n"
+ ".inst 0x4f8ae2dd // sdot v29.4s, v22.16b, v10.4b[0]\n"
+ "ins v21.d[1], temploadreg1\n"
+ ".inst 0x4f8de2de // sdot v30.4s, v22.16b, v13.4b[0]\n"
+ "ldr temploadreg1, [%[b_ptr0], #0x78]\n"
+ ".inst 0x4f90e2df // sdot v31.4s, v22.16b, v16.4b[0]\n"
+ "ldr d22, [%[b_ptr0], #0x40]\n"
+ ".inst 0x4fa1e2fa // sdot v26.4s, v23.16b, v1.4b[1]\n"
+ ".inst 0x4fa4e2fb // sdot v27.4s, v23.16b, v4.4b[1]\n"
+ ".inst 0x4fa7e2fc // sdot v28.4s, v23.16b, v7.4b[1]\n"
+ ".inst 0x4faae2fd // sdot v29.4s, v23.16b, v10.4b[1]\n"
+ ".inst 0x4fade2fe // sdot v30.4s, v23.16b, v13.4b[1]\n"
+ ".inst 0x4fb0e2ff // sdot v31.4s, v23.16b, v16.4b[1]\n"
+ "ldr d23, [%[b_ptr0], #0x50]\n"
+ ".inst 0x4f81eb1a // sdot v26.4s, v24.16b, v1.4b[2]\n"
+ ".inst 0x4f84eb1b // sdot v27.4s, v24.16b, v4.4b[2]\n"
+ ".inst 0x4f87eb1c // sdot v28.4s, v24.16b, v7.4b[2]\n"
+ ".inst 0x4f8aeb1d // sdot v29.4s, v24.16b, v10.4b[2]\n"
+ ".inst 0x4f8deb1e // sdot v30.4s, v24.16b, v13.4b[2]\n"
+ ".inst 0x4f90eb1f // sdot v31.4s, v24.16b, v16.4b[2]\n"
+ "ldr d24, [%[b_ptr0], #0x60]\n"
+ ".inst 0x4fa1eb3a // sdot v26.4s, v25.16b, v1.4b[3]\n"
+ ".inst 0x4fa4eb3b // sdot v27.4s, v25.16b, v4.4b[3]\n"
+ ".inst 0x4fa7eb3c // sdot v28.4s, v25.16b, v7.4b[3]\n"
+ ".inst 0x4faaeb3d // sdot v29.4s, v25.16b, v10.4b[3]\n"
+ ".inst 0x4fadeb3e // sdot v30.4s, v25.16b, v13.4b[3]\n"
+ ".inst 0x4fb0eb3f // sdot v31.4s, v25.16b, v16.4b[3]\n"
+ "ldr d25, [%[b_ptr0], #0x70]\n"
+ ".inst 0x4f82e25a // sdot v26.4s, v18.16b, v2.4b[0]\n"
+ ".inst 0x4f85e25b // sdot v27.4s, v18.16b, v5.4b[0]\n"
+ ".inst 0x4f88e25c // sdot v28.4s, v18.16b, v8.4b[0]\n"
"ins v25.d[1], temploadreg1\n"
- ".word 0x4f89e31c // sdot v28.4s, v24.16b, v9.4b[0]\n"
- "ldr temploadreg0, [%[b_ptr0], #0x8]\n"
- ".word 0x4f8de31d // sdot v29.4s, v24.16b, v13.4b[0]\n"
- "ldr temploadreg1, [%[b_ptr0], #0x18]\n"
- ".word 0x4f91e31e // sdot v30.4s, v24.16b, v17.4b[0]\n"
- ".word 0x4f95e31f // sdot v31.4s, v24.16b, v21.4b[0]\n"
- "ldr d24, [%[b_ptr0]]\n"
- ".word 0x4fa1e33a // sdot v26.4s, v25.16b, v1.4b[1]\n"
- ".word 0x4fa5e33b // sdot v27.4s, v25.16b, v5.4b[1]\n"
- ".word 0x4fa9e33c // sdot v28.4s, v25.16b, v9.4b[1]\n"
- "ins v24.d[1], temploadreg0\n"
- ".word 0x4fade33d // sdot v29.4s, v25.16b, v13.4b[1]\n"
- ".word 0x4fb1e33e // sdot v30.4s, v25.16b, v17.4b[1]\n"
- ".word 0x4fb5e33f // sdot v31.4s, v25.16b, v21.4b[1]\n"
- "ldr d25, [%[b_ptr0], #0x10]\n"
- ".word 0x4f81eb1a // sdot v26.4s, v24.16b, v1.4b[2]\n"
- "add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f85eb1b // sdot v27.4s, v24.16b, v5.4b[2]\n"
- "ins v25.d[1], temploadreg1\n"
- ".word 0x4f89eb1c // sdot v28.4s, v24.16b, v9.4b[2]\n"
- "ldr temploadreg0, [%[b_ptr0], #0x8]\n"
- ".word 0x4f8deb1d // sdot v29.4s, v24.16b, v13.4b[2]\n"
- "ldr temploadreg1, [%[b_ptr0], #0x18]\n"
- ".word 0x4f91eb1e // sdot v30.4s, v24.16b, v17.4b[2]\n"
- ".word 0x4f95eb1f // sdot v31.4s, v24.16b, v21.4b[2]\n"
- "ldr d24, [%[b_ptr0]]\n"
- ".word 0x4fa1eb3a // sdot v26.4s, v25.16b, v1.4b[3]\n"
- ".word 0x4fa5eb3b // sdot v27.4s, v25.16b, v5.4b[3]\n"
- ".word 0x4fa9eb3c // sdot v28.4s, v25.16b, v9.4b[3]\n"
- "ins v24.d[1], temploadreg0\n"
- ".word 0x4fadeb3d // sdot v29.4s, v25.16b, v13.4b[3]\n"
- ".word 0x4fb1eb3e // sdot v30.4s, v25.16b, v17.4b[3]\n"
- ".word 0x4fb5eb3f // sdot v31.4s, v25.16b, v21.4b[3]\n"
- "ldr d25, [%[b_ptr0], #0x10]\n"
- ".word 0x4f82e31a // sdot v26.4s, v24.16b, v2.4b[0]\n"
- "add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f86e31b // sdot v27.4s, v24.16b, v6.4b[0]\n"
- "ins v25.d[1], temploadreg1\n"
- ".word 0x4f8ae31c // sdot v28.4s, v24.16b, v10.4b[0]\n"
- "ldr temploadreg0, [%[b_ptr0], #0x8]\n"
- ".word 0x4f8ee31d // sdot v29.4s, v24.16b, v14.4b[0]\n"
- ".word 0x4f92e31e // sdot v30.4s, v24.16b, v18.4b[0]\n"
- ".word 0x4f96e31f // sdot v31.4s, v24.16b, v22.4b[0]\n"
- "ldr d24, [%[b_ptr0]]\n"
- ".word 0x4fa2e33a // sdot v26.4s, v25.16b, v2.4b[1]\n"
- "add %[b_ptr0], %[b_ptr0], #0x10\n"
- ".word 0x4fa6e33b // sdot v27.4s, v25.16b, v6.4b[1]\n"
- "ins v24.d[1], temploadreg0\n"
- ".word 0x4faae33c // sdot v28.4s, v25.16b, v10.4b[1]\n"
- "ldr temploadreg0, [%[b_ptr0], #0x8]\n"
- ".word 0x4faee33d // sdot v29.4s, v25.16b, v14.4b[1]\n"
- "ldr temploadreg1, [%[b_ptr0], #0x18]\n"
- ".word 0x4fb2e33e // sdot v30.4s, v25.16b, v18.4b[1]\n"
- ".word 0x4fb6e33f // sdot v31.4s, v25.16b, v22.4b[1]\n"
- "ldr d25, [%[b_ptr0], #0x10]\n"
- ".word 0x4f82eb1a // sdot v26.4s, v24.16b, v2.4b[2]\n"
- ".word 0x4f86eb1b // sdot v27.4s, v24.16b, v6.4b[2]\n"
- ".word 0x4f8aeb1c // sdot v28.4s, v24.16b, v10.4b[2]\n"
- "ins v25.d[1], temploadreg1\n"
- ".word 0x4f8eeb1d // sdot v29.4s, v24.16b, v14.4b[2]\n"
- ".word 0x4f92eb1e // sdot v30.4s, v24.16b, v18.4b[2]\n"
- ".word 0x4f96eb1f // sdot v31.4s, v24.16b, v22.4b[2]\n"
- "ldr d24, [%[b_ptr0]]\n"
- "add %[b_ptr0], %[b_ptr0], #0x20\n"
+ ".inst 0x4f8be25d // sdot v29.4s, v18.16b, v11.4b[0]\n"
+ ".inst 0x4f8ee25e // sdot v30.4s, v18.16b, v14.4b[0]\n"
+ ".inst 0x4f91e25f // sdot v31.4s, v18.16b, v17.4b[0]\n"
+ "ldr d18, [%[b_ptr0]]\n"
+ ".inst 0x4fa2e27a // sdot v26.4s, v19.16b, v2.4b[1]\n"
+ ".inst 0x4fa5e27b // sdot v27.4s, v19.16b, v5.4b[1]\n"
+ ".inst 0x4fa8e27c // sdot v28.4s, v19.16b, v8.4b[1]\n"
+ "ins v18.d[1], temploadreg2\n"
+ ".inst 0x4fabe27d // sdot v29.4s, v19.16b, v11.4b[1]\n"
+ "ldr temploadreg2, [%[b_ptr0], #0x48]\n"
+ ".inst 0x4faee27e // sdot v30.4s, v19.16b, v14.4b[1]\n"
+ ".inst 0x4fb1e27f // sdot v31.4s, v19.16b, v17.4b[1]\n"
+ "ldr d19, [%[b_ptr0], #0x10]\n"
+ ".inst 0x4f82ea9a // sdot v26.4s, v20.16b, v2.4b[2]\n"
+ "ins v22.d[1], temploadreg2\n"
+ ".inst 0x4f85ea9b // sdot v27.4s, v20.16b, v5.4b[2]\n"
+ ".inst 0x4f88ea9c // sdot v28.4s, v20.16b, v8.4b[2]\n"
+ "ins v19.d[1], temploadreg3\n"
+ ".inst 0x4f8bea9d // sdot v29.4s, v20.16b, v11.4b[2]\n"
+ "ldr temploadreg3, [%[b_ptr0], #0x58]\n"
+ ".inst 0x4f8eea9e // sdot v30.4s, v20.16b, v14.4b[2]\n"
+ ".inst 0x4f91ea9f // sdot v31.4s, v20.16b, v17.4b[2]\n"
+ "ldr d20, [%[b_ptr0], #0x20]\n"
+ "ins v23.d[1], temploadreg3\n"
+ "ins v20.d[1], temploadreg0\n"
+ "ldr temploadreg0, [%[b_ptr0], #0x68]\n"
+ "add %[b_ptr0], %[b_ptr0], #0x80\n"
"ins v24.d[1], temploadreg0\n"
"b.ne 8b\n"
"7:\n"
- "str q26, [%[c_ptr0]], #0x10\n"
+ "str q26, [%[c_ptr0]]\n"
+ "add %[c_ptr0], %[c_ptr0], #0x10\n"
"movi v26.4s, #0\n"
- "str q27, [c_ptr1], #0x10\n"
+ "str q27, [c_ptr1]\n"
+ "add c_ptr1, c_ptr1, #0x10\n"
"movi v27.4s, #0\n"
- ".word 0x4f80e31a // sdot v26.4s, v24.16b, v0.4b[0]\n"
- "str q28, [c_ptr2], #0x10\n"
+ ".inst 0x4f80e25a // sdot v26.4s, v18.16b, v0.4b[0]\n"
+ "str q28, [c_ptr2]\n"
"movi v28.4s, #0\n"
- ".word 0x4f84e31b // sdot v27.4s, v24.16b, v4.4b[0]\n"
- ".word 0x4fa0e33a // sdot v26.4s, v25.16b, v0.4b[1]\n"
- "str q29, [c_ptr3], #0x10\n"
+ "add c_ptr2, c_ptr2, #0x10\n"
+ ".inst 0x4f83e25b // sdot v27.4s, v18.16b, v3.4b[0]\n"
+ "str q29, [c_ptr3]\n"
"movi v29.4s, #0\n"
- ".word 0x4f88e31c // sdot v28.4s, v24.16b, v8.4b[0]\n"
- ".word 0x4fa4e33b // sdot v27.4s, v25.16b, v4.4b[1]\n"
- "str q30, [c_ptr4], #0x10\n"
+ "add c_ptr3, c_ptr3, #0x10\n"
+ ".inst 0x4f86e25c // sdot v28.4s, v18.16b, v6.4b[0]\n"
+ "str q30, [c_ptr4]\n"
"movi v30.4s, #0\n"
- ".word 0x4f8ce31d // sdot v29.4s, v24.16b, v12.4b[0]\n"
- ".word 0x4fa8e33c // sdot v28.4s, v25.16b, v8.4b[1]\n"
- "str q31, [c_ptr5], #0x10\n"
+ "add c_ptr4, c_ptr4, #0x10\n"
+ ".inst 0x4f89e25d // sdot v29.4s, v18.16b, v9.4b[0]\n"
+ "str q31, [c_ptr5]\n"
"movi v31.4s, #0\n"
- ".word 0x4f90e31e // sdot v30.4s, v24.16b, v16.4b[0]\n"
- ".word 0x4face33d // sdot v29.4s, v25.16b, v12.4b[1]\n"
- ".word 0x4f94e31f // sdot v31.4s, v24.16b, v20.4b[0]\n"
- "ldr q24, [%[b_ptr0]]\n"
- ".word 0x4fb0e33e // sdot v30.4s, v25.16b, v16.4b[1]\n"
- ".word 0x4fb4e33f // sdot v31.4s, v25.16b, v20.4b[1]\n"
- "ldr q25, [%[b_ptr0], #0x10]\n"
- ".word 0x4f80eb1a // sdot v26.4s, v24.16b, v0.4b[2]\n"
- "add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f84eb1b // sdot v27.4s, v24.16b, v4.4b[2]\n"
- ".word 0x4f88eb1c // sdot v28.4s, v24.16b, v8.4b[2]\n"
- ".word 0x4f8ceb1d // sdot v29.4s, v24.16b, v12.4b[2]\n"
- ".word 0x4f90eb1e // sdot v30.4s, v24.16b, v16.4b[2]\n"
- ".word 0x4f94eb1f // sdot v31.4s, v24.16b, v20.4b[2]\n"
- "ldr q24, [%[b_ptr0]]\n"
- ".word 0x4fa0eb3a // sdot v26.4s, v25.16b, v0.4b[3]\n"
- ".word 0x4fa4eb3b // sdot v27.4s, v25.16b, v4.4b[3]\n"
- ".word 0x4fa8eb3c // sdot v28.4s, v25.16b, v8.4b[3]\n"
- ".word 0x4faceb3d // sdot v29.4s, v25.16b, v12.4b[3]\n"
- ".word 0x4fb0eb3e // sdot v30.4s, v25.16b, v16.4b[3]\n"
- ".word 0x4fb4eb3f // sdot v31.4s, v25.16b, v20.4b[3]\n"
- "ldr q25, [%[b_ptr0], #0x10]\n"
- ".word 0x4f81e31a // sdot v26.4s, v24.16b, v1.4b[0]\n"
- "add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f85e31b // sdot v27.4s, v24.16b, v5.4b[0]\n"
- ".word 0x4f89e31c // sdot v28.4s, v24.16b, v9.4b[0]\n"
- ".word 0x4f8de31d // sdot v29.4s, v24.16b, v13.4b[0]\n"
- ".word 0x4f91e31e // sdot v30.4s, v24.16b, v17.4b[0]\n"
- ".word 0x4f95e31f // sdot v31.4s, v24.16b, v21.4b[0]\n"
- "ldr q24, [%[b_ptr0]]\n"
- ".word 0x4fa1e33a // sdot v26.4s, v25.16b, v1.4b[1]\n"
- ".word 0x4fa5e33b // sdot v27.4s, v25.16b, v5.4b[1]\n"
- ".word 0x4fa9e33c // sdot v28.4s, v25.16b, v9.4b[1]\n"
- ".word 0x4fade33d // sdot v29.4s, v25.16b, v13.4b[1]\n"
- ".word 0x4fb1e33e // sdot v30.4s, v25.16b, v17.4b[1]\n"
- ".word 0x4fb5e33f // sdot v31.4s, v25.16b, v21.4b[1]\n"
- "ldr q25, [%[b_ptr0], #0x10]\n"
- ".word 0x4f81eb1a // sdot v26.4s, v24.16b, v1.4b[2]\n"
- "add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f85eb1b // sdot v27.4s, v24.16b, v5.4b[2]\n"
- ".word 0x4f89eb1c // sdot v28.4s, v24.16b, v9.4b[2]\n"
- ".word 0x4f8deb1d // sdot v29.4s, v24.16b, v13.4b[2]\n"
- ".word 0x4f91eb1e // sdot v30.4s, v24.16b, v17.4b[2]\n"
- ".word 0x4f95eb1f // sdot v31.4s, v24.16b, v21.4b[2]\n"
- "ldr q24, [%[b_ptr0]]\n"
- ".word 0x4fa1eb3a // sdot v26.4s, v25.16b, v1.4b[3]\n"
- ".word 0x4fa5eb3b // sdot v27.4s, v25.16b, v5.4b[3]\n"
- ".word 0x4fa9eb3c // sdot v28.4s, v25.16b, v9.4b[3]\n"
- ".word 0x4fadeb3d // sdot v29.4s, v25.16b, v13.4b[3]\n"
- ".word 0x4fb1eb3e // sdot v30.4s, v25.16b, v17.4b[3]\n"
- ".word 0x4fb5eb3f // sdot v31.4s, v25.16b, v21.4b[3]\n"
- "ldr q25, [%[b_ptr0], #0x10]\n"
- ".word 0x4f82e31a // sdot v26.4s, v24.16b, v2.4b[0]\n"
- "add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f86e31b // sdot v27.4s, v24.16b, v6.4b[0]\n"
- ".word 0x4f8ae31c // sdot v28.4s, v24.16b, v10.4b[0]\n"
- ".word 0x4f8ee31d // sdot v29.4s, v24.16b, v14.4b[0]\n"
- ".word 0x4f92e31e // sdot v30.4s, v24.16b, v18.4b[0]\n"
- ".word 0x4f96e31f // sdot v31.4s, v24.16b, v22.4b[0]\n"
- "ldr q24, [%[b_ptr0]]\n"
- ".word 0x4fa2e33a // sdot v26.4s, v25.16b, v2.4b[1]\n"
- "add %[b_ptr0], %[b_ptr0], #0x10\n"
- ".word 0x4fa6e33b // sdot v27.4s, v25.16b, v6.4b[1]\n"
- ".word 0x4faae33c // sdot v28.4s, v25.16b, v10.4b[1]\n"
- ".word 0x4faee33d // sdot v29.4s, v25.16b, v14.4b[1]\n"
- ".word 0x4fb2e33e // sdot v30.4s, v25.16b, v18.4b[1]\n"
- ".word 0x4fb6e33f // sdot v31.4s, v25.16b, v22.4b[1]\n"
- ".word 0x4f82eb1a // sdot v26.4s, v24.16b, v2.4b[2]\n"
- ".word 0x4f86eb1b // sdot v27.4s, v24.16b, v6.4b[2]\n"
- ".word 0x4f8aeb1c // sdot v28.4s, v24.16b, v10.4b[2]\n"
- ".word 0x4f8eeb1d // sdot v29.4s, v24.16b, v14.4b[2]\n"
- ".word 0x4f92eb1e // sdot v30.4s, v24.16b, v18.4b[2]\n"
- ".word 0x4f96eb1f // sdot v31.4s, v24.16b, v22.4b[2]\n"
+ "add c_ptr5, c_ptr5, #0x10\n"
+ ".inst 0x4f8ce25e // sdot v30.4s, v18.16b, v12.4b[0]\n"
+ ".inst 0x4fa0e27a // sdot v26.4s, v19.16b, v0.4b[1]\n"
+ ".inst 0x4f8fe25f // sdot v31.4s, v18.16b, v15.4b[0]\n"
+ "ldr q18, [%[b_ptr0]]\n"
+ ".inst 0x4fa3e27b // sdot v27.4s, v19.16b, v3.4b[1]\n"
+ ".inst 0x4fa6e27c // sdot v28.4s, v19.16b, v6.4b[1]\n"
+ ".inst 0x4fa9e27d // sdot v29.4s, v19.16b, v9.4b[1]\n"
+ ".inst 0x4face27e // sdot v30.4s, v19.16b, v12.4b[1]\n"
+ ".inst 0x4fafe27f // sdot v31.4s, v19.16b, v15.4b[1]\n"
+ "ldr q19, [%[b_ptr0], #0x10]\n"
+ ".inst 0x4f80ea9a // sdot v26.4s, v20.16b, v0.4b[2]\n"
+ ".inst 0x4f83ea9b // sdot v27.4s, v20.16b, v3.4b[2]\n"
+ ".inst 0x4f86ea9c // sdot v28.4s, v20.16b, v6.4b[2]\n"
+ ".inst 0x4f89ea9d // sdot v29.4s, v20.16b, v9.4b[2]\n"
+ ".inst 0x4f8cea9e // sdot v30.4s, v20.16b, v12.4b[2]\n"
+ ".inst 0x4f8fea9f // sdot v31.4s, v20.16b, v15.4b[2]\n"
+ "ldr q20, [%[b_ptr0], #0x20]\n"
+ ".inst 0x4fa0eaba // sdot v26.4s, v21.16b, v0.4b[3]\n"
+ "add %[b_ptr0], %[b_ptr0], #0x30\n"
+ ".inst 0x4fa3eabb // sdot v27.4s, v21.16b, v3.4b[3]\n"
+ ".inst 0x4fa6eabc // sdot v28.4s, v21.16b, v6.4b[3]\n"
+ ".inst 0x4fa9eabd // sdot v29.4s, v21.16b, v9.4b[3]\n"
+ ".inst 0x4faceabe // sdot v30.4s, v21.16b, v12.4b[3]\n"
+ ".inst 0x4fafeabf // sdot v31.4s, v21.16b, v15.4b[3]\n"
+ ".inst 0x4f81e2da // sdot v26.4s, v22.16b, v1.4b[0]\n"
+ ".inst 0x4f84e2db // sdot v27.4s, v22.16b, v4.4b[0]\n"
+ ".inst 0x4f87e2dc // sdot v28.4s, v22.16b, v7.4b[0]\n"
+ ".inst 0x4f8ae2dd // sdot v29.4s, v22.16b, v10.4b[0]\n"
+ ".inst 0x4f8de2de // sdot v30.4s, v22.16b, v13.4b[0]\n"
+ ".inst 0x4f90e2df // sdot v31.4s, v22.16b, v16.4b[0]\n"
+ ".inst 0x4fa1e2fa // sdot v26.4s, v23.16b, v1.4b[1]\n"
+ ".inst 0x4fa4e2fb // sdot v27.4s, v23.16b, v4.4b[1]\n"
+ ".inst 0x4fa7e2fc // sdot v28.4s, v23.16b, v7.4b[1]\n"
+ ".inst 0x4faae2fd // sdot v29.4s, v23.16b, v10.4b[1]\n"
+ ".inst 0x4fade2fe // sdot v30.4s, v23.16b, v13.4b[1]\n"
+ ".inst 0x4fb0e2ff // sdot v31.4s, v23.16b, v16.4b[1]\n"
+ ".inst 0x4f81eb1a // sdot v26.4s, v24.16b, v1.4b[2]\n"
+ ".inst 0x4f84eb1b // sdot v27.4s, v24.16b, v4.4b[2]\n"
+ ".inst 0x4f87eb1c // sdot v28.4s, v24.16b, v7.4b[2]\n"
+ ".inst 0x4f8aeb1d // sdot v29.4s, v24.16b, v10.4b[2]\n"
+ ".inst 0x4f8deb1e // sdot v30.4s, v24.16b, v13.4b[2]\n"
+ ".inst 0x4f90eb1f // sdot v31.4s, v24.16b, v16.4b[2]\n"
+ ".inst 0x4fa1eb3a // sdot v26.4s, v25.16b, v1.4b[3]\n"
+ ".inst 0x4fa4eb3b // sdot v27.4s, v25.16b, v4.4b[3]\n"
+ ".inst 0x4fa7eb3c // sdot v28.4s, v25.16b, v7.4b[3]\n"
+ ".inst 0x4faaeb3d // sdot v29.4s, v25.16b, v10.4b[3]\n"
+ ".inst 0x4fadeb3e // sdot v30.4s, v25.16b, v13.4b[3]\n"
+ ".inst 0x4fb0eb3f // sdot v31.4s, v25.16b, v16.4b[3]\n"
+ ".inst 0x4f82e25a // sdot v26.4s, v18.16b, v2.4b[0]\n"
+ ".inst 0x4f85e25b // sdot v27.4s, v18.16b, v5.4b[0]\n"
+ ".inst 0x4f88e25c // sdot v28.4s, v18.16b, v8.4b[0]\n"
+ ".inst 0x4f8be25d // sdot v29.4s, v18.16b, v11.4b[0]\n"
+ ".inst 0x4f8ee25e // sdot v30.4s, v18.16b, v14.4b[0]\n"
+ ".inst 0x4f91e25f // sdot v31.4s, v18.16b, v17.4b[0]\n"
+ ".inst 0x4fa2e27a // sdot v26.4s, v19.16b, v2.4b[1]\n"
+ ".inst 0x4fa5e27b // sdot v27.4s, v19.16b, v5.4b[1]\n"
+ ".inst 0x4fa8e27c // sdot v28.4s, v19.16b, v8.4b[1]\n"
+ ".inst 0x4fabe27d // sdot v29.4s, v19.16b, v11.4b[1]\n"
+ ".inst 0x4faee27e // sdot v30.4s, v19.16b, v14.4b[1]\n"
+ ".inst 0x4fb1e27f // sdot v31.4s, v19.16b, v17.4b[1]\n"
+ ".inst 0x4f82ea9a // sdot v26.4s, v20.16b, v2.4b[2]\n"
+ ".inst 0x4f85ea9b // sdot v27.4s, v20.16b, v5.4b[2]\n"
+ ".inst 0x4f88ea9c // sdot v28.4s, v20.16b, v8.4b[2]\n"
+ ".inst 0x4f8bea9d // sdot v29.4s, v20.16b, v11.4b[2]\n"
+ ".inst 0x4f8eea9e // sdot v30.4s, v20.16b, v14.4b[2]\n"
+ ".inst 0x4f91ea9f // sdot v31.4s, v20.16b, v17.4b[2]\n"
"6:\n"
"str q26, [%[c_ptr0]]\n"
+ "add %[c_ptr0], %[c_ptr0], #0x10\n"
"str q27, [c_ptr1]\n"
+ "add c_ptr1, c_ptr1, #0x10\n"
"str q28, [c_ptr2]\n"
+ "add c_ptr2, c_ptr2, #0x10\n"
"str q29, [c_ptr3]\n"
+ "add c_ptr3, c_ptr3, #0x10\n"
"str q30, [c_ptr4]\n"
+ "add c_ptr4, c_ptr4, #0x10\n"
"str q31, [c_ptr5]\n"
+ "add c_ptr5, c_ptr5, #0x10\n"
".unreq a_ptr1\n"
".unreq a_ptr2\n"
".unreq a_ptr3\n"
@@ -1369,419 +1430,437 @@ void a64_smallK_hybrid_s8s32_dot_4x6_a55(const int8_t *A, int lda, const int8_t
"add a_ptr1, %[a_ptr0], #0x0\n"
"1:\n"
"ldr q0, [%[a_ptr0]], #0x10\n"
- "ldr q4, [a_ptr1], #0x10\n"
- "ldr q8, [a_ptr2], #0x10\n"
- "ldr q12, [a_ptr3], #0x10\n"
- "ldr q16, [a_ptr4], #0x10\n"
- "ldr q20, [a_ptr5], #0x10\n"
+ "ldr q3, [a_ptr1], #0x10\n"
+ "ldr q6, [a_ptr2], #0x10\n"
+ "ldr q9, [a_ptr3], #0x10\n"
+ "ldr q12, [a_ptr4], #0x10\n"
+ "ldr q15, [a_ptr5], #0x10\n"
"ldr q1, [%[a_ptr0]], #0x10\n"
- "ldr q5, [a_ptr1], #0x10\n"
- "ldr q9, [a_ptr2], #0x10\n"
- "ldr q13, [a_ptr3], #0x10\n"
- "ldr q17, [a_ptr4], #0x10\n"
- "ldr q21, [a_ptr5], #0x10\n"
+ "ldr q4, [a_ptr1], #0x10\n"
+ "ldr q7, [a_ptr2], #0x10\n"
+ "ldr q10, [a_ptr3], #0x10\n"
+ "ldr q13, [a_ptr4], #0x10\n"
+ "ldr q16, [a_ptr5], #0x10\n"
"cbnz %[odds], 2f\n"
"ldr q2, [%[a_ptr0]]\n"
- "ldr q6, [a_ptr1]\n"
- "ldr q10, [a_ptr2]\n"
- "ldr q14, [a_ptr3]\n"
- "ldr q18, [a_ptr4]\n"
- "ldr q22, [a_ptr5]\n"
+ "ldr q5, [a_ptr1]\n"
+ "ldr q8, [a_ptr2]\n"
+ "ldr q11, [a_ptr3]\n"
+ "ldr q14, [a_ptr4]\n"
+ "ldr q17, [a_ptr5]\n"
"b 3f\n"
"2:\n"
"ldr d2, [%[a_ptr0]], #0x8\n"
- "ldr d6, [a_ptr1], #0x8\n"
- "ldr d10, [a_ptr2], #0x8\n"
- "ldr d14, [a_ptr3], #0x8\n"
- "ldr d18, [a_ptr4], #0x8\n"
- "ldr d22, [a_ptr5], #0x8\n"
+ "ldr d5, [a_ptr1], #0x8\n"
+ "ldr d8, [a_ptr2], #0x8\n"
+ "ldr d11, [a_ptr3], #0x8\n"
+ "ldr d14, [a_ptr4], #0x8\n"
+ "ldr d17, [a_ptr5], #0x8\n"
"ld1 {v2.s}[2], [%[a_ptr0]], #4\n"
- "ld1 {v6.s}[2], [a_ptr1], #4\n"
- "ld1 {v10.s}[2], [a_ptr2], #4\n"
- "ld1 {v14.s}[2], [a_ptr3], #4\n"
- "ld1 {v18.s}[2], [a_ptr4], #4\n"
- "ld1 {v22.s}[2], [a_ptr5], #4\n"
+ "ld1 {v5.s}[2], [a_ptr1], #4\n"
+ "ld1 {v8.s}[2], [a_ptr2], #4\n"
+ "ld1 {v11.s}[2], [a_ptr3], #4\n"
+ "ld1 {v14.s}[2], [a_ptr4], #4\n"
+ "ld1 {v17.s}[2], [a_ptr5], #4\n"
"subs %[odds], %[odds], #0x1\n"
"b.ne 4f\n"
"ld1 {v2.b}[12], [%[a_ptr0]]\n"
- "ld1 {v6.b}[12], [a_ptr1]\n"
- "ld1 {v10.b}[12], [a_ptr2]\n"
- "ld1 {v14.b}[12], [a_ptr3]\n"
- "ld1 {v18.b}[12], [a_ptr4]\n"
- "ld1 {v22.b}[12], [a_ptr5]\n"
+ "ld1 {v5.b}[12], [a_ptr1]\n"
+ "ld1 {v8.b}[12], [a_ptr2]\n"
+ "ld1 {v11.b}[12], [a_ptr3]\n"
+ "ld1 {v14.b}[12], [a_ptr4]\n"
+ "ld1 {v17.b}[12], [a_ptr5]\n"
"b 3f\n"
"4:\n"
"ld1 {v2.h}[6], [%[a_ptr0]], #2\n"
- "ld1 {v6.h}[6], [a_ptr1], #2\n"
- "ld1 {v10.h}[6], [a_ptr2], #2\n"
- "ld1 {v14.h}[6], [a_ptr3], #2\n"
- "ld1 {v18.h}[6], [a_ptr4], #2\n"
- "ld1 {v22.h}[6], [a_ptr5], #2\n"
+ "ld1 {v5.h}[6], [a_ptr1], #2\n"
+ "ld1 {v8.h}[6], [a_ptr2], #2\n"
+ "ld1 {v11.h}[6], [a_ptr3], #2\n"
+ "ld1 {v14.h}[6], [a_ptr4], #2\n"
+ "ld1 {v17.h}[6], [a_ptr5], #2\n"
"subs %[odds], %[odds], #0x1\n"
"b.ne 5f\n"
"b 3f\n"
"5:\n"
"ld1 {v2.b}[14], [%[a_ptr0]]\n"
- "ld1 {v6.b}[14], [a_ptr1]\n"
- "ld1 {v10.b}[14], [a_ptr2]\n"
- "ld1 {v14.b}[14], [a_ptr3]\n"
- "ld1 {v18.b}[14], [a_ptr4]\n"
- "ld1 {v22.b}[14], [a_ptr5]\n"
+ "ld1 {v5.b}[14], [a_ptr1]\n"
+ "ld1 {v8.b}[14], [a_ptr2]\n"
+ "ld1 {v11.b}[14], [a_ptr3]\n"
+ "ld1 {v14.b}[14], [a_ptr4]\n"
+ "ld1 {v17.b}[14], [a_ptr5]\n"
"3:\n"
"movi v26.4s, #0\n"
- "ldr q24, [%[b_ptr0]]\n"
+ "ldr q18, [%[b_ptr0]]\n"
"movi v27.4s, #0\n"
- "ldr q25, [%[b_ptr0], #0x10]\n"
+ "ldr q19, [%[b_ptr0], #0x10]\n"
"movi v28.4s, #0\n"
- "prfm PLDL1KEEP, [a_ptr5, #0x40]\n"
+ "ldr q20, [%[b_ptr0], #0x20]\n"
"movi v29.4s, #0\n"
- "prfm PLDL1KEEP, [a_ptr5, #0x80]\n"
+ "ldr q21, [%[b_ptr0], #0x30]\n"
"movi v30.4s, #0\n"
- "prfm PLDL1KEEP, [a_ptr5, #0xc0]\n"
+ "ldr q22, [%[b_ptr0], #0x40]\n"
"movi v31.4s, #0\n"
+ "ldr q23, [%[b_ptr0], #0x50]\n"
+ ".inst 0x4f80e25a // sdot v26.4s, v18.16b, v0.4b[0]\n"
+ "ldr q24, [%[b_ptr0], #0x60]\n"
+ ".inst 0x4f83e25b // sdot v27.4s, v18.16b, v3.4b[0]\n"
+ "ldr q25, [%[b_ptr0], #0x70]\n"
+ ".inst 0x4f86e25c // sdot v28.4s, v18.16b, v6.4b[0]\n"
+ "prfm PLDL1KEEP, [a_ptr5, #0x40]\n"
+ ".inst 0x4f89e25d // sdot v29.4s, v18.16b, v9.4b[0]\n"
+ "prfm PLDL1KEEP, [a_ptr5, #0x80]\n"
+ ".inst 0x4f8ce25e // sdot v30.4s, v18.16b, v12.4b[0]\n"
+ "prfm PLDL1KEEP, [a_ptr5, #0xc0]\n"
+ ".inst 0x4f8fe25f // sdot v31.4s, v18.16b, v15.4b[0]\n"
"prfm PLDL1KEEP, [a_ptr5, #0x100]\n"
- ".word 0x4f80e31a // sdot v26.4s, v24.16b, v0.4b[0]\n"
+ ".inst 0x4fa0e27a // sdot v26.4s, v19.16b, v0.4b[1]\n"
"prfm PLDL1KEEP, [a_ptr5, #0x140]\n"
- ".word 0x4f84e31b // sdot v27.4s, v24.16b, v4.4b[0]\n"
+ ".inst 0x4fa3e27b // sdot v27.4s, v19.16b, v3.4b[1]\n"
"prfm PLDL1KEEP, [a_ptr5, #0x180]\n"
- ".word 0x4f88e31c // sdot v28.4s, v24.16b, v8.4b[0]\n"
- "add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f8ce31d // sdot v29.4s, v24.16b, v12.4b[0]\n"
- ".word 0x4f90e31e // sdot v30.4s, v24.16b, v16.4b[0]\n"
- ".word 0x4f94e31f // sdot v31.4s, v24.16b, v20.4b[0]\n"
- "ldr q24, [%[b_ptr0]]\n"
- ".word 0x4fa0e33a // sdot v26.4s, v25.16b, v0.4b[1]\n"
- ".word 0x4fa4e33b // sdot v27.4s, v25.16b, v4.4b[1]\n"
- ".word 0x4fa8e33c // sdot v28.4s, v25.16b, v8.4b[1]\n"
- ".word 0x4face33d // sdot v29.4s, v25.16b, v12.4b[1]\n"
- ".word 0x4fb0e33e // sdot v30.4s, v25.16b, v16.4b[1]\n"
- ".word 0x4fb4e33f // sdot v31.4s, v25.16b, v20.4b[1]\n"
- "ldr q25, [%[b_ptr0], #0x10]\n"
- ".word 0x4f80eb1a // sdot v26.4s, v24.16b, v0.4b[2]\n"
- "add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f84eb1b // sdot v27.4s, v24.16b, v4.4b[2]\n"
- ".word 0x4f88eb1c // sdot v28.4s, v24.16b, v8.4b[2]\n"
- ".word 0x4f8ceb1d // sdot v29.4s, v24.16b, v12.4b[2]\n"
- ".word 0x4f90eb1e // sdot v30.4s, v24.16b, v16.4b[2]\n"
- ".word 0x4f94eb1f // sdot v31.4s, v24.16b, v20.4b[2]\n"
- "ldr q24, [%[b_ptr0]]\n"
- ".word 0x4fa0eb3a // sdot v26.4s, v25.16b, v0.4b[3]\n"
- ".word 0x4fa4eb3b // sdot v27.4s, v25.16b, v4.4b[3]\n"
- ".word 0x4fa8eb3c // sdot v28.4s, v25.16b, v8.4b[3]\n"
- ".word 0x4faceb3d // sdot v29.4s, v25.16b, v12.4b[3]\n"
- ".word 0x4fb0eb3e // sdot v30.4s, v25.16b, v16.4b[3]\n"
- ".word 0x4fb4eb3f // sdot v31.4s, v25.16b, v20.4b[3]\n"
- "ldr q25, [%[b_ptr0], #0x10]\n"
- ".word 0x4f81e31a // sdot v26.4s, v24.16b, v1.4b[0]\n"
- "add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f85e31b // sdot v27.4s, v24.16b, v5.4b[0]\n"
- ".word 0x4f89e31c // sdot v28.4s, v24.16b, v9.4b[0]\n"
- ".word 0x4f8de31d // sdot v29.4s, v24.16b, v13.4b[0]\n"
- ".word 0x4f91e31e // sdot v30.4s, v24.16b, v17.4b[0]\n"
- ".word 0x4f95e31f // sdot v31.4s, v24.16b, v21.4b[0]\n"
- "ldr q24, [%[b_ptr0]]\n"
- ".word 0x4fa1e33a // sdot v26.4s, v25.16b, v1.4b[1]\n"
- ".word 0x4fa5e33b // sdot v27.4s, v25.16b, v5.4b[1]\n"
- ".word 0x4fa9e33c // sdot v28.4s, v25.16b, v9.4b[1]\n"
- ".word 0x4fade33d // sdot v29.4s, v25.16b, v13.4b[1]\n"
- ".word 0x4fb1e33e // sdot v30.4s, v25.16b, v17.4b[1]\n"
- ".word 0x4fb5e33f // sdot v31.4s, v25.16b, v21.4b[1]\n"
- "ldr q25, [%[b_ptr0], #0x10]\n"
- ".word 0x4f81eb1a // sdot v26.4s, v24.16b, v1.4b[2]\n"
- "add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f85eb1b // sdot v27.4s, v24.16b, v5.4b[2]\n"
- ".word 0x4f89eb1c // sdot v28.4s, v24.16b, v9.4b[2]\n"
- ".word 0x4f8deb1d // sdot v29.4s, v24.16b, v13.4b[2]\n"
- ".word 0x4f91eb1e // sdot v30.4s, v24.16b, v17.4b[2]\n"
- ".word 0x4f95eb1f // sdot v31.4s, v24.16b, v21.4b[2]\n"
- "ldr q24, [%[b_ptr0]]\n"
- ".word 0x4fa1eb3a // sdot v26.4s, v25.16b, v1.4b[3]\n"
- ".word 0x4fa5eb3b // sdot v27.4s, v25.16b, v5.4b[3]\n"
- ".word 0x4fa9eb3c // sdot v28.4s, v25.16b, v9.4b[3]\n"
- ".word 0x4fadeb3d // sdot v29.4s, v25.16b, v13.4b[3]\n"
- ".word 0x4fb1eb3e // sdot v30.4s, v25.16b, v17.4b[3]\n"
- ".word 0x4fb5eb3f // sdot v31.4s, v25.16b, v21.4b[3]\n"
- "ldr q25, [%[b_ptr0], #0x10]\n"
- ".word 0x4f82e31a // sdot v26.4s, v24.16b, v2.4b[0]\n"
- "add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f86e31b // sdot v27.4s, v24.16b, v6.4b[0]\n"
- ".word 0x4f8ae31c // sdot v28.4s, v24.16b, v10.4b[0]\n"
- ".word 0x4f8ee31d // sdot v29.4s, v24.16b, v14.4b[0]\n"
- ".word 0x4f92e31e // sdot v30.4s, v24.16b, v18.4b[0]\n"
- ".word 0x4f96e31f // sdot v31.4s, v24.16b, v22.4b[0]\n"
- "ldr q24, [%[b_ptr0]]\n"
- ".word 0x4fa2e33a // sdot v26.4s, v25.16b, v2.4b[1]\n"
- ".word 0x4fa6e33b // sdot v27.4s, v25.16b, v6.4b[1]\n"
- ".word 0x4faae33c // sdot v28.4s, v25.16b, v10.4b[1]\n"
- ".word 0x4faee33d // sdot v29.4s, v25.16b, v14.4b[1]\n"
- ".word 0x4fb2e33e // sdot v30.4s, v25.16b, v18.4b[1]\n"
- ".word 0x4fb6e33f // sdot v31.4s, v25.16b, v22.4b[1]\n"
- "ldr q25, [%[b_ptr0], #0x10]\n"
- ".word 0x4f82eb1a // sdot v26.4s, v24.16b, v2.4b[2]\n"
- "add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f86eb1b // sdot v27.4s, v24.16b, v6.4b[2]\n"
- ".word 0x4f8aeb1c // sdot v28.4s, v24.16b, v10.4b[2]\n"
- ".word 0x4f8eeb1d // sdot v29.4s, v24.16b, v14.4b[2]\n"
- ".word 0x4f92eb1e // sdot v30.4s, v24.16b, v18.4b[2]\n"
- ".word 0x4f96eb1f // sdot v31.4s, v24.16b, v22.4b[2]\n"
- ".word 0x4fa2eb3a // sdot v26.4s, v25.16b, v2.4b[3]\n"
- ".word 0x4fa6eb3b // sdot v27.4s, v25.16b, v6.4b[3]\n"
- ".word 0x4faaeb3c // sdot v28.4s, v25.16b, v10.4b[3]\n"
- ".word 0x4faeeb3d // sdot v29.4s, v25.16b, v14.4b[3]\n"
- ".word 0x4fb2eb3e // sdot v30.4s, v25.16b, v18.4b[3]\n"
- ".word 0x4fb6eb3f // sdot v31.4s, v25.16b, v22.4b[3]\n"
+ ".inst 0x4fa6e27c // sdot v28.4s, v19.16b, v6.4b[1]\n"
+ "add %[b_ptr0], %[b_ptr0], #0x80\n"
+ ".inst 0x4fa9e27d // sdot v29.4s, v19.16b, v9.4b[1]\n"
+ "ldr q18, [%[b_ptr0]]\n"
+ ".inst 0x4face27e // sdot v30.4s, v19.16b, v12.4b[1]\n"
+ ".inst 0x4fafe27f // sdot v31.4s, v19.16b, v15.4b[1]\n"
+ "ldr q19, [%[b_ptr0], #0x10]\n"
+ ".inst 0x4f80ea9a // sdot v26.4s, v20.16b, v0.4b[2]\n"
+ ".inst 0x4f83ea9b // sdot v27.4s, v20.16b, v3.4b[2]\n"
+ ".inst 0x4f86ea9c // sdot v28.4s, v20.16b, v6.4b[2]\n"
+ ".inst 0x4f89ea9d // sdot v29.4s, v20.16b, v9.4b[2]\n"
+ ".inst 0x4f8cea9e // sdot v30.4s, v20.16b, v12.4b[2]\n"
+ ".inst 0x4f8fea9f // sdot v31.4s, v20.16b, v15.4b[2]\n"
+ "ldr q20, [%[b_ptr0], #0x20]\n"
+ ".inst 0x4fa0eaba // sdot v26.4s, v21.16b, v0.4b[3]\n"
+ ".inst 0x4fa3eabb // sdot v27.4s, v21.16b, v3.4b[3]\n"
+ ".inst 0x4fa6eabc // sdot v28.4s, v21.16b, v6.4b[3]\n"
+ ".inst 0x4fa9eabd // sdot v29.4s, v21.16b, v9.4b[3]\n"
+ ".inst 0x4faceabe // sdot v30.4s, v21.16b, v12.4b[3]\n"
+ ".inst 0x4fafeabf // sdot v31.4s, v21.16b, v15.4b[3]\n"
+ "ldr q21, [%[b_ptr0], #0x30]\n"
+ ".inst 0x4f81e2da // sdot v26.4s, v22.16b, v1.4b[0]\n"
+ "add %[b_ptr0], %[b_ptr0], #0x40\n"
+ ".inst 0x4f84e2db // sdot v27.4s, v22.16b, v4.4b[0]\n"
+ ".inst 0x4f87e2dc // sdot v28.4s, v22.16b, v7.4b[0]\n"
+ ".inst 0x4f8ae2dd // sdot v29.4s, v22.16b, v10.4b[0]\n"
+ ".inst 0x4f8de2de // sdot v30.4s, v22.16b, v13.4b[0]\n"
+ ".inst 0x4f90e2df // sdot v31.4s, v22.16b, v16.4b[0]\n"
+ ".inst 0x4fa1e2fa // sdot v26.4s, v23.16b, v1.4b[1]\n"
+ ".inst 0x4fa4e2fb // sdot v27.4s, v23.16b, v4.4b[1]\n"
+ ".inst 0x4fa7e2fc // sdot v28.4s, v23.16b, v7.4b[1]\n"
+ ".inst 0x4faae2fd // sdot v29.4s, v23.16b, v10.4b[1]\n"
+ ".inst 0x4fade2fe // sdot v30.4s, v23.16b, v13.4b[1]\n"
+ ".inst 0x4fb0e2ff // sdot v31.4s, v23.16b, v16.4b[1]\n"
+ ".inst 0x4f81eb1a // sdot v26.4s, v24.16b, v1.4b[2]\n"
+ ".inst 0x4f84eb1b // sdot v27.4s, v24.16b, v4.4b[2]\n"
+ ".inst 0x4f87eb1c // sdot v28.4s, v24.16b, v7.4b[2]\n"
+ ".inst 0x4f8aeb1d // sdot v29.4s, v24.16b, v10.4b[2]\n"
+ ".inst 0x4f8deb1e // sdot v30.4s, v24.16b, v13.4b[2]\n"
+ ".inst 0x4f90eb1f // sdot v31.4s, v24.16b, v16.4b[2]\n"
+ ".inst 0x4fa1eb3a // sdot v26.4s, v25.16b, v1.4b[3]\n"
+ ".inst 0x4fa4eb3b // sdot v27.4s, v25.16b, v4.4b[3]\n"
+ ".inst 0x4fa7eb3c // sdot v28.4s, v25.16b, v7.4b[3]\n"
+ ".inst 0x4faaeb3d // sdot v29.4s, v25.16b, v10.4b[3]\n"
+ ".inst 0x4fadeb3e // sdot v30.4s, v25.16b, v13.4b[3]\n"
+ ".inst 0x4fb0eb3f // sdot v31.4s, v25.16b, v16.4b[3]\n"
+ ".inst 0x4f82e25a // sdot v26.4s, v18.16b, v2.4b[0]\n"
+ ".inst 0x4f85e25b // sdot v27.4s, v18.16b, v5.4b[0]\n"
+ ".inst 0x4f88e25c // sdot v28.4s, v18.16b, v8.4b[0]\n"
+ ".inst 0x4f8be25d // sdot v29.4s, v18.16b, v11.4b[0]\n"
+ ".inst 0x4f8ee25e // sdot v30.4s, v18.16b, v14.4b[0]\n"
+ ".inst 0x4f91e25f // sdot v31.4s, v18.16b, v17.4b[0]\n"
+ ".inst 0x4fa2e27a // sdot v26.4s, v19.16b, v2.4b[1]\n"
+ ".inst 0x4fa5e27b // sdot v27.4s, v19.16b, v5.4b[1]\n"
+ ".inst 0x4fa8e27c // sdot v28.4s, v19.16b, v8.4b[1]\n"
+ ".inst 0x4fabe27d // sdot v29.4s, v19.16b, v11.4b[1]\n"
+ ".inst 0x4faee27e // sdot v30.4s, v19.16b, v14.4b[1]\n"
+ ".inst 0x4fb1e27f // sdot v31.4s, v19.16b, v17.4b[1]\n"
+ ".inst 0x4f82ea9a // sdot v26.4s, v20.16b, v2.4b[2]\n"
+ ".inst 0x4f85ea9b // sdot v27.4s, v20.16b, v5.4b[2]\n"
+ ".inst 0x4f88ea9c // sdot v28.4s, v20.16b, v8.4b[2]\n"
+ ".inst 0x4f8bea9d // sdot v29.4s, v20.16b, v11.4b[2]\n"
+ ".inst 0x4f8eea9e // sdot v30.4s, v20.16b, v14.4b[2]\n"
+ ".inst 0x4f91ea9f // sdot v31.4s, v20.16b, v17.4b[2]\n"
+ ".inst 0x4fa2eaba // sdot v26.4s, v21.16b, v2.4b[3]\n"
+ ".inst 0x4fa5eabb // sdot v27.4s, v21.16b, v5.4b[3]\n"
+ ".inst 0x4fa8eabc // sdot v28.4s, v21.16b, v8.4b[3]\n"
+ ".inst 0x4fabeabd // sdot v29.4s, v21.16b, v11.4b[3]\n"
+ ".inst 0x4faeeabe // sdot v30.4s, v21.16b, v14.4b[3]\n"
+ ".inst 0x4fb1eabf // sdot v31.4s, v21.16b, v17.4b[3]\n"
"cbz %[loops], 6f\n"
- "ldr d24, [%[b_ptr0]]\n"
+ "ldr d18, [%[b_ptr0]]\n"
"subs %[loops], %[loops], #0x1\n"
- "ldr temploadreg0, [%[b_ptr0], #0x8]\n"
- "ldr d25, [%[b_ptr0], #0x10]\n"
- "ldr temploadreg1, [%[b_ptr0], #0x18]\n"
- "add %[b_ptr0], %[b_ptr0], #0x20\n"
+ "ldr temploadreg2, [%[b_ptr0], #0x8]\n"
+ "ldr d19, [%[b_ptr0], #0x10]\n"
+ "ldr temploadreg3, [%[b_ptr0], #0x18]\n"
+ "ldr d20, [%[b_ptr0], #0x20]\n"
+ "ldr temploadreg0, [%[b_ptr0], #0x28]\n"
+ "ldr d21, [%[b_ptr0], #0x30]\n"
+ "ldr temploadreg1, [%[b_ptr0], #0x38]\n"
+ "ldr d22, [%[b_ptr0], #0x40]\n"
+ "ins v18.d[1], temploadreg2\n"
+ "ldr temploadreg2, [%[b_ptr0], #0x48]\n"
+ "ldr d23, [%[b_ptr0], #0x50]\n"
+ "ins v19.d[1], temploadreg3\n"
+ "ldr temploadreg3, [%[b_ptr0], #0x58]\n"
+ "ldr d24, [%[b_ptr0], #0x60]\n"
+ "ins v20.d[1], temploadreg0\n"
+ "ldr temploadreg0, [%[b_ptr0], #0x68]\n"
+ "ldr d25, [%[b_ptr0], #0x70]\n"
+ "ins v21.d[1], temploadreg1\n"
+ "ldr temploadreg1, [%[b_ptr0], #0x78]\n"
+ "add %[b_ptr0], %[b_ptr0], #0x80\n"
+ "ins v22.d[1], temploadreg2\n"
+ "ins v23.d[1], temploadreg3\n"
"ins v24.d[1], temploadreg0\n"
"b.eq 7f\n"
"8:\n"
- "str q26, [%[c_ptr0]], #0x10\n"
+ "str q26, [%[c_ptr0]]\n"
"subs %[loops], %[loops], #0x1\n"
"movi v26.4s, #0\n"
"ins v25.d[1], temploadreg1\n"
- "ldr temploadreg0, [%[b_ptr0], #0x8]\n"
- "prfm PSTL1KEEP, [%[c_ptr0], #0x40]\n"
- "str q27, [c_ptr1], #0x10\n"
+ "ldr temploadreg2, [%[b_ptr0], #0x8]\n"
+ "add %[c_ptr0], %[c_ptr0], #0x10\n"
+ "str q27, [c_ptr1]\n"
+ "add c_ptr1, c_ptr1, #0x10\n"
"movi v27.4s, #0\n"
- "ldr temploadreg1, [%[b_ptr0], #0x18]\n"
- ".word 0x4f80e31a // sdot v26.4s, v24.16b, v0.4b[0]\n"
- "str q28, [c_ptr2], #0x10\n"
- "prfm PSTL1KEEP, [c_ptr1, #0x40]\n"
+ "ldr temploadreg3, [%[b_ptr0], #0x18]\n"
+ ".inst 0x4f80e25a // sdot v26.4s, v18.16b, v0.4b[0]\n"
+ "str q28, [c_ptr2]\n"
"movi v28.4s, #0\n"
- ".word 0x4f84e31b // sdot v27.4s, v24.16b, v4.4b[0]\n"
- "str q29, [c_ptr3], #0x10\n"
+ "ldr temploadreg0, [%[b_ptr0], #0x28]\n"
+ "ldr temploadreg1, [%[b_ptr0], #0x38]\n"
+ "add c_ptr2, c_ptr2, #0x10\n"
+ ".inst 0x4f83e25b // sdot v27.4s, v18.16b, v3.4b[0]\n"
+ "str q29, [c_ptr3]\n"
"movi v29.4s, #0\n"
- "prfm PSTL1KEEP, [c_ptr2, #0x40]\n"
- ".word 0x4f88e31c // sdot v28.4s, v24.16b, v8.4b[0]\n"
- "str q30, [c_ptr4], #0x10\n"
+ "add c_ptr3, c_ptr3, #0x10\n"
+ ".inst 0x4f86e25c // sdot v28.4s, v18.16b, v6.4b[0]\n"
+ "str q30, [c_ptr4]\n"
"movi v30.4s, #0\n"
- "prfm PSTL1KEEP, [c_ptr3, #0x40]\n"
- ".word 0x4f8ce31d // sdot v29.4s, v24.16b, v12.4b[0]\n"
- "str q31, [c_ptr5], #0x10\n"
+ "add c_ptr4, c_ptr4, #0x10\n"
+ ".inst 0x4f89e25d // sdot v29.4s, v18.16b, v9.4b[0]\n"
+ "str q31, [c_ptr5]\n"
"movi v31.4s, #0\n"
+ "add c_ptr5, c_ptr5, #0x10\n"
+ ".inst 0x4f8ce25e // sdot v30.4s, v18.16b, v12.4b[0]\n"
+ "prfm PSTL1KEEP, [%[c_ptr0], #0x40]\n"
+ ".inst 0x4f8fe25f // sdot v31.4s, v18.16b, v15.4b[0]\n"
+ "ldr d18, [%[b_ptr0]]\n"
+ ".inst 0x4fa0e27a // sdot v26.4s, v19.16b, v0.4b[1]\n"
+ "prfm PSTL1KEEP, [c_ptr1, #0x40]\n"
+ ".inst 0x4fa3e27b // sdot v27.4s, v19.16b, v3.4b[1]\n"
+ "ins v18.d[1], temploadreg2\n"
+ ".inst 0x4fa6e27c // sdot v28.4s, v19.16b, v6.4b[1]\n"
+ "prfm PSTL1KEEP, [c_ptr2, #0x40]\n"
+ ".inst 0x4fa9e27d // sdot v29.4s, v19.16b, v9.4b[1]\n"
+ "prfm PSTL1KEEP, [c_ptr3, #0x40]\n"
+ ".inst 0x4face27e // sdot v30.4s, v19.16b, v12.4b[1]\n"
"prfm PSTL1KEEP, [c_ptr4, #0x40]\n"
- ".word 0x4f90e31e // sdot v30.4s, v24.16b, v16.4b[0]\n"
+ ".inst 0x4fafe27f // sdot v31.4s, v19.16b, v15.4b[1]\n"
+ "ldr d19, [%[b_ptr0], #0x10]\n"
+ ".inst 0x4f80ea9a // sdot v26.4s, v20.16b, v0.4b[2]\n"
"prfm PSTL1KEEP, [c_ptr5, #0x40]\n"
- ".word 0x4f94e31f // sdot v31.4s, v24.16b, v20.4b[0]\n"
- "ldr d24, [%[b_ptr0]]\n"
- ".word 0x4fa0e33a // sdot v26.4s, v25.16b, v0.4b[1]\n"
- ".word 0x4fa4e33b // sdot v27.4s, v25.16b, v4.4b[1]\n"
- ".word 0x4fa8e33c // sdot v28.4s, v25.16b, v8.4b[1]\n"
- "ins v24.d[1], temploadreg0\n"
- ".word 0x4face33d // sdot v29.4s, v25.16b, v12.4b[1]\n"
- ".word 0x4fb0e33e // sdot v30.4s, v25.16b, v16.4b[1]\n"
- ".word 0x4fb4e33f // sdot v31.4s, v25.16b, v20.4b[1]\n"
- "ldr d25, [%[b_ptr0], #0x10]\n"
- ".word 0x4f80eb1a // sdot v26.4s, v24.16b, v0.4b[2]\n"
- "add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f84eb1b // sdot v27.4s, v24.16b, v4.4b[2]\n"
- "ins v25.d[1], temploadreg1\n"
- ".word 0x4f88eb1c // sdot v28.4s, v24.16b, v8.4b[2]\n"
- "ldr temploadreg0, [%[b_ptr0], #0x8]\n"
- ".word 0x4f8ceb1d // sdot v29.4s, v24.16b, v12.4b[2]\n"
- "ldr temploadreg1, [%[b_ptr0], #0x18]\n"
- ".word 0x4f90eb1e // sdot v30.4s, v24.16b, v16.4b[2]\n"
- ".word 0x4f94eb1f // sdot v31.4s, v24.16b, v20.4b[2]\n"
- "ldr d24, [%[b_ptr0]]\n"
- ".word 0x4fa0eb3a // sdot v26.4s, v25.16b, v0.4b[3]\n"
- ".word 0x4fa4eb3b // sdot v27.4s, v25.16b, v4.4b[3]\n"
- ".word 0x4fa8eb3c // sdot v28.4s, v25.16b, v8.4b[3]\n"
- "ins v24.d[1], temploadreg0\n"
- ".word 0x4faceb3d // sdot v29.4s, v25.16b, v12.4b[3]\n"
- ".word 0x4fb0eb3e // sdot v30.4s, v25.16b, v16.4b[3]\n"
- ".word 0x4fb4eb3f // sdot v31.4s, v25.16b, v20.4b[3]\n"
- "ldr d25, [%[b_ptr0], #0x10]\n"
- ".word 0x4f81e31a // sdot v26.4s, v24.16b, v1.4b[0]\n"
- "add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f85e31b // sdot v27.4s, v24.16b, v5.4b[0]\n"
- "ins v25.d[1], temploadreg1\n"
- ".word 0x4f89e31c // sdot v28.4s, v24.16b, v9.4b[0]\n"
- "ldr temploadreg0, [%[b_ptr0], #0x8]\n"
- ".word 0x4f8de31d // sdot v29.4s, v24.16b, v13.4b[0]\n"
- "ldr temploadreg1, [%[b_ptr0], #0x18]\n"
- ".word 0x4f91e31e // sdot v30.4s, v24.16b, v17.4b[0]\n"
- ".word 0x4f95e31f // sdot v31.4s, v24.16b, v21.4b[0]\n"
- "ldr d24, [%[b_ptr0]]\n"
- ".word 0x4fa1e33a // sdot v26.4s, v25.16b, v1.4b[1]\n"
- ".word 0x4fa5e33b // sdot v27.4s, v25.16b, v5.4b[1]\n"
- ".word 0x4fa9e33c // sdot v28.4s, v25.16b, v9.4b[1]\n"
- "ins v24.d[1], temploadreg0\n"
- ".word 0x4fade33d // sdot v29.4s, v25.16b, v13.4b[1]\n"
- ".word 0x4fb1e33e // sdot v30.4s, v25.16b, v17.4b[1]\n"
- ".word 0x4fb5e33f // sdot v31.4s, v25.16b, v21.4b[1]\n"
- "ldr d25, [%[b_ptr0], #0x10]\n"
- ".word 0x4f81eb1a // sdot v26.4s, v24.16b, v1.4b[2]\n"
- "add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f85eb1b // sdot v27.4s, v24.16b, v5.4b[2]\n"
- "ins v25.d[1], temploadreg1\n"
- ".word 0x4f89eb1c // sdot v28.4s, v24.16b, v9.4b[2]\n"
- "ldr temploadreg0, [%[b_ptr0], #0x8]\n"
- ".word 0x4f8deb1d // sdot v29.4s, v24.16b, v13.4b[2]\n"
- "ldr temploadreg1, [%[b_ptr0], #0x18]\n"
- ".word 0x4f91eb1e // sdot v30.4s, v24.16b, v17.4b[2]\n"
- ".word 0x4f95eb1f // sdot v31.4s, v24.16b, v21.4b[2]\n"
- "ldr d24, [%[b_ptr0]]\n"
- ".word 0x4fa1eb3a // sdot v26.4s, v25.16b, v1.4b[3]\n"
- ".word 0x4fa5eb3b // sdot v27.4s, v25.16b, v5.4b[3]\n"
- ".word 0x4fa9eb3c // sdot v28.4s, v25.16b, v9.4b[3]\n"
+ ".inst 0x4f83ea9b // sdot v27.4s, v20.16b, v3.4b[2]\n"
+ "ins v19.d[1], temploadreg3\n"
+ ".inst 0x4f86ea9c // sdot v28.4s, v20.16b, v6.4b[2]\n"
+ ".inst 0x4f89ea9d // sdot v29.4s, v20.16b, v9.4b[2]\n"
+ ".inst 0x4f8cea9e // sdot v30.4s, v20.16b, v12.4b[2]\n"
+ ".inst 0x4f8fea9f // sdot v31.4s, v20.16b, v15.4b[2]\n"
+ "ldr d20, [%[b_ptr0], #0x20]\n"
+ ".inst 0x4fa0eaba // sdot v26.4s, v21.16b, v0.4b[3]\n"
+ ".inst 0x4fa3eabb // sdot v27.4s, v21.16b, v3.4b[3]\n"
+ ".inst 0x4fa6eabc // sdot v28.4s, v21.16b, v6.4b[3]\n"
+ "ins v20.d[1], temploadreg0\n"
+ ".inst 0x4fa9eabd // sdot v29.4s, v21.16b, v9.4b[3]\n"
+ ".inst 0x4faceabe // sdot v30.4s, v21.16b, v12.4b[3]\n"
+ ".inst 0x4fafeabf // sdot v31.4s, v21.16b, v15.4b[3]\n"
+ "ldr d21, [%[b_ptr0], #0x30]\n"
+ ".inst 0x4f81e2da // sdot v26.4s, v22.16b, v1.4b[0]\n"
+ "add %[b_ptr0], %[b_ptr0], #0x40\n"
+ ".inst 0x4f84e2db // sdot v27.4s, v22.16b, v4.4b[0]\n"
+ "ins v21.d[1], temploadreg1\n"
+ ".inst 0x4f87e2dc // sdot v28.4s, v22.16b, v7.4b[0]\n"
+ "ldr temploadreg2, [%[b_ptr0], #0x8]\n"
+ ".inst 0x4f8ae2dd // sdot v29.4s, v22.16b, v10.4b[0]\n"
+ "ldr temploadreg3, [%[b_ptr0], #0x18]\n"
+ ".inst 0x4f8de2de // sdot v30.4s, v22.16b, v13.4b[0]\n"
+ "ldr temploadreg0, [%[b_ptr0], #0x28]\n"
+ ".inst 0x4f90e2df // sdot v31.4s, v22.16b, v16.4b[0]\n"
+ "ldr temploadreg1, [%[b_ptr0], #0x38]\n"
+ ".inst 0x4fa1e2fa // sdot v26.4s, v23.16b, v1.4b[1]\n"
+ "ldr d22, [%[b_ptr0], #0x40]\n"
+ ".inst 0x4fa4e2fb // sdot v27.4s, v23.16b, v4.4b[1]\n"
+ ".inst 0x4fa7e2fc // sdot v28.4s, v23.16b, v7.4b[1]\n"
+ ".inst 0x4faae2fd // sdot v29.4s, v23.16b, v10.4b[1]\n"
+ ".inst 0x4fade2fe // sdot v30.4s, v23.16b, v13.4b[1]\n"
+ ".inst 0x4fb0e2ff // sdot v31.4s, v23.16b, v16.4b[1]\n"
+ "ldr d23, [%[b_ptr0], #0x50]\n"
+ ".inst 0x4f81eb1a // sdot v26.4s, v24.16b, v1.4b[2]\n"
+ ".inst 0x4f84eb1b // sdot v27.4s, v24.16b, v4.4b[2]\n"
+ ".inst 0x4f87eb1c // sdot v28.4s, v24.16b, v7.4b[2]\n"
+ ".inst 0x4f8aeb1d // sdot v29.4s, v24.16b, v10.4b[2]\n"
+ ".inst 0x4f8deb1e // sdot v30.4s, v24.16b, v13.4b[2]\n"
+ ".inst 0x4f90eb1f // sdot v31.4s, v24.16b, v16.4b[2]\n"
+ "ldr d24, [%[b_ptr0], #0x60]\n"
+ ".inst 0x4fa1eb3a // sdot v26.4s, v25.16b, v1.4b[3]\n"
+ ".inst 0x4fa4eb3b // sdot v27.4s, v25.16b, v4.4b[3]\n"
+ ".inst 0x4fa7eb3c // sdot v28.4s, v25.16b, v7.4b[3]\n"
+ ".inst 0x4faaeb3d // sdot v29.4s, v25.16b, v10.4b[3]\n"
+ ".inst 0x4fadeb3e // sdot v30.4s, v25.16b, v13.4b[3]\n"
+ ".inst 0x4fb0eb3f // sdot v31.4s, v25.16b, v16.4b[3]\n"
+ "ldr d25, [%[b_ptr0], #0x70]\n"
+ ".inst 0x4f82e25a // sdot v26.4s, v18.16b, v2.4b[0]\n"
+ ".inst 0x4f85e25b // sdot v27.4s, v18.16b, v5.4b[0]\n"
+ ".inst 0x4f88e25c // sdot v28.4s, v18.16b, v8.4b[0]\n"
+ ".inst 0x4f8be25d // sdot v29.4s, v18.16b, v11.4b[0]\n"
+ ".inst 0x4f8ee25e // sdot v30.4s, v18.16b, v14.4b[0]\n"
+ ".inst 0x4f91e25f // sdot v31.4s, v18.16b, v17.4b[0]\n"
+ "ldr d18, [%[b_ptr0]]\n"
+ ".inst 0x4fa2e27a // sdot v26.4s, v19.16b, v2.4b[1]\n"
+ ".inst 0x4fa5e27b // sdot v27.4s, v19.16b, v5.4b[1]\n"
+ ".inst 0x4fa8e27c // sdot v28.4s, v19.16b, v8.4b[1]\n"
+ "ins v18.d[1], temploadreg2\n"
+ ".inst 0x4fabe27d // sdot v29.4s, v19.16b, v11.4b[1]\n"
+ "ldr temploadreg2, [%[b_ptr0], #0x48]\n"
+ ".inst 0x4faee27e // sdot v30.4s, v19.16b, v14.4b[1]\n"
+ ".inst 0x4fb1e27f // sdot v31.4s, v19.16b, v17.4b[1]\n"
+ "ldr d19, [%[b_ptr0], #0x10]\n"
+ ".inst 0x4f82ea9a // sdot v26.4s, v20.16b, v2.4b[2]\n"
+ "ins v22.d[1], temploadreg2\n"
+ ".inst 0x4f85ea9b // sdot v27.4s, v20.16b, v5.4b[2]\n"
+ ".inst 0x4f88ea9c // sdot v28.4s, v20.16b, v8.4b[2]\n"
+ "ins v19.d[1], temploadreg3\n"
+ ".inst 0x4f8bea9d // sdot v29.4s, v20.16b, v11.4b[2]\n"
+ "ldr temploadreg3, [%[b_ptr0], #0x58]\n"
+ ".inst 0x4f8eea9e // sdot v30.4s, v20.16b, v14.4b[2]\n"
+ ".inst 0x4f91ea9f // sdot v31.4s, v20.16b, v17.4b[2]\n"
+ "ldr d20, [%[b_ptr0], #0x20]\n"
+ ".inst 0x4fa2eaba // sdot v26.4s, v21.16b, v2.4b[3]\n"
+ "ins v23.d[1], temploadreg3\n"
+ ".inst 0x4fa5eabb // sdot v27.4s, v21.16b, v5.4b[3]\n"
+ ".inst 0x4fa8eabc // sdot v28.4s, v21.16b, v8.4b[3]\n"
+ "ins v20.d[1], temploadreg0\n"
+ ".inst 0x4fabeabd // sdot v29.4s, v21.16b, v11.4b[3]\n"
+ "ldr temploadreg0, [%[b_ptr0], #0x68]\n"
+ ".inst 0x4faeeabe // sdot v30.4s, v21.16b, v14.4b[3]\n"
+ ".inst 0x4fb1eabf // sdot v31.4s, v21.16b, v17.4b[3]\n"
+ "ldr d21, [%[b_ptr0], #0x30]\n"
"ins v24.d[1], temploadreg0\n"
- ".word 0x4fadeb3d // sdot v29.4s, v25.16b, v13.4b[3]\n"
- ".word 0x4fb1eb3e // sdot v30.4s, v25.16b, v17.4b[3]\n"
- ".word 0x4fb5eb3f // sdot v31.4s, v25.16b, v21.4b[3]\n"
- "ldr d25, [%[b_ptr0], #0x10]\n"
- ".word 0x4f82e31a // sdot v26.4s, v24.16b, v2.4b[0]\n"
- "add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f86e31b // sdot v27.4s, v24.16b, v6.4b[0]\n"
- "ins v25.d[1], temploadreg1\n"
- ".word 0x4f8ae31c // sdot v28.4s, v24.16b, v10.4b[0]\n"
- "ldr temploadreg0, [%[b_ptr0], #0x8]\n"
- ".word 0x4f8ee31d // sdot v29.4s, v24.16b, v14.4b[0]\n"
- "ldr temploadreg1, [%[b_ptr0], #0x18]\n"
- ".word 0x4f92e31e // sdot v30.4s, v24.16b, v18.4b[0]\n"
- ".word 0x4f96e31f // sdot v31.4s, v24.16b, v22.4b[0]\n"
- "ldr d24, [%[b_ptr0]]\n"
- ".word 0x4fa2e33a // sdot v26.4s, v25.16b, v2.4b[1]\n"
- ".word 0x4fa6e33b // sdot v27.4s, v25.16b, v6.4b[1]\n"
- ".word 0x4faae33c // sdot v28.4s, v25.16b, v10.4b[1]\n"
- "ins v24.d[1], temploadreg0\n"
- ".word 0x4faee33d // sdot v29.4s, v25.16b, v14.4b[1]\n"
- ".word 0x4fb2e33e // sdot v30.4s, v25.16b, v18.4b[1]\n"
- ".word 0x4fb6e33f // sdot v31.4s, v25.16b, v22.4b[1]\n"
- "ldr d25, [%[b_ptr0], #0x10]\n"
- ".word 0x4f82eb1a // sdot v26.4s, v24.16b, v2.4b[2]\n"
- "add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f86eb1b // sdot v27.4s, v24.16b, v6.4b[2]\n"
- "ins v25.d[1], temploadreg1\n"
- ".word 0x4f8aeb1c // sdot v28.4s, v24.16b, v10.4b[2]\n"
- "ldr temploadreg0, [%[b_ptr0], #0x8]\n"
- ".word 0x4f8eeb1d // sdot v29.4s, v24.16b, v14.4b[2]\n"
- "ldr temploadreg1, [%[b_ptr0], #0x18]\n"
- ".word 0x4f92eb1e // sdot v30.4s, v24.16b, v18.4b[2]\n"
- ".word 0x4f96eb1f // sdot v31.4s, v24.16b, v22.4b[2]\n"
- "ldr d24, [%[b_ptr0]]\n"
- ".word 0x4fa2eb3a // sdot v26.4s, v25.16b, v2.4b[3]\n"
- ".word 0x4fa6eb3b // sdot v27.4s, v25.16b, v6.4b[3]\n"
- ".word 0x4faaeb3c // sdot v28.4s, v25.16b, v10.4b[3]\n"
- "ins v24.d[1], temploadreg0\n"
- ".word 0x4faeeb3d // sdot v29.4s, v25.16b, v14.4b[3]\n"
- ".word 0x4fb2eb3e // sdot v30.4s, v25.16b, v18.4b[3]\n"
- ".word 0x4fb6eb3f // sdot v31.4s, v25.16b, v22.4b[3]\n"
- "ldr d25, [%[b_ptr0], #0x10]\n"
- "add %[b_ptr0], %[b_ptr0], #0x20\n"
+ "ins v21.d[1], temploadreg1\n"
+ "ldr temploadreg1, [%[b_ptr0], #0x78]\n"
+ "add %[b_ptr0], %[b_ptr0], #0x80\n"
"b.ne 8b\n"
"7:\n"
- "str q26, [%[c_ptr0]], #0x10\n"
+ "str q26, [%[c_ptr0]]\n"
+ "add %[c_ptr0], %[c_ptr0], #0x10\n"
"movi v26.4s, #0\n"
"ins v25.d[1], temploadreg1\n"
- "str q27, [c_ptr1], #0x10\n"
+ "str q27, [c_ptr1]\n"
+ "add c_ptr1, c_ptr1, #0x10\n"
"movi v27.4s, #0\n"
- ".word 0x4f80e31a // sdot v26.4s, v24.16b, v0.4b[0]\n"
- "str q28, [c_ptr2], #0x10\n"
+ ".inst 0x4f80e25a // sdot v26.4s, v18.16b, v0.4b[0]\n"
+ "str q28, [c_ptr2]\n"
"movi v28.4s, #0\n"
- ".word 0x4f84e31b // sdot v27.4s, v24.16b, v4.4b[0]\n"
- ".word 0x4fa0e33a // sdot v26.4s, v25.16b, v0.4b[1]\n"
- "str q29, [c_ptr3], #0x10\n"
+ "add c_ptr2, c_ptr2, #0x10\n"
+ ".inst 0x4f83e25b // sdot v27.4s, v18.16b, v3.4b[0]\n"
+ "str q29, [c_ptr3]\n"
"movi v29.4s, #0\n"
- ".word 0x4f88e31c // sdot v28.4s, v24.16b, v8.4b[0]\n"
- ".word 0x4fa4e33b // sdot v27.4s, v25.16b, v4.4b[1]\n"
- "str q30, [c_ptr4], #0x10\n"
+ "add c_ptr3, c_ptr3, #0x10\n"
+ ".inst 0x4f86e25c // sdot v28.4s, v18.16b, v6.4b[0]\n"
+ "str q30, [c_ptr4]\n"
"movi v30.4s, #0\n"
- ".word 0x4f8ce31d // sdot v29.4s, v24.16b, v12.4b[0]\n"
- ".word 0x4fa8e33c // sdot v28.4s, v25.16b, v8.4b[1]\n"
- "str q31, [c_ptr5], #0x10\n"
+ "add c_ptr4, c_ptr4, #0x10\n"
+ ".inst 0x4f89e25d // sdot v29.4s, v18.16b, v9.4b[0]\n"
+ "str q31, [c_ptr5]\n"
"movi v31.4s, #0\n"
- ".word 0x4f90e31e // sdot v30.4s, v24.16b, v16.4b[0]\n"
- ".word 0x4face33d // sdot v29.4s, v25.16b, v12.4b[1]\n"
- ".word 0x4f94e31f // sdot v31.4s, v24.16b, v20.4b[0]\n"
- "ldr q24, [%[b_ptr0]]\n"
- ".word 0x4fb0e33e // sdot v30.4s, v25.16b, v16.4b[1]\n"
- ".word 0x4fb4e33f // sdot v31.4s, v25.16b, v20.4b[1]\n"
- "ldr q25, [%[b_ptr0], #0x10]\n"
- ".word 0x4f80eb1a // sdot v26.4s, v24.16b, v0.4b[2]\n"
- "add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f84eb1b // sdot v27.4s, v24.16b, v4.4b[2]\n"
- ".word 0x4f88eb1c // sdot v28.4s, v24.16b, v8.4b[2]\n"
- ".word 0x4f8ceb1d // sdot v29.4s, v24.16b, v12.4b[2]\n"
- ".word 0x4f90eb1e // sdot v30.4s, v24.16b, v16.4b[2]\n"
- ".word 0x4f94eb1f // sdot v31.4s, v24.16b, v20.4b[2]\n"
- "ldr q24, [%[b_ptr0]]\n"
- ".word 0x4fa0eb3a // sdot v26.4s, v25.16b, v0.4b[3]\n"
- ".word 0x4fa4eb3b // sdot v27.4s, v25.16b, v4.4b[3]\n"
- ".word 0x4fa8eb3c // sdot v28.4s, v25.16b, v8.4b[3]\n"
- ".word 0x4faceb3d // sdot v29.4s, v25.16b, v12.4b[3]\n"
- ".word 0x4fb0eb3e // sdot v30.4s, v25.16b, v16.4b[3]\n"
- ".word 0x4fb4eb3f // sdot v31.4s, v25.16b, v20.4b[3]\n"
- "ldr q25, [%[b_ptr0], #0x10]\n"
- ".word 0x4f81e31a // sdot v26.4s, v24.16b, v1.4b[0]\n"
- "add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f85e31b // sdot v27.4s, v24.16b, v5.4b[0]\n"
- ".word 0x4f89e31c // sdot v28.4s, v24.16b, v9.4b[0]\n"
- ".word 0x4f8de31d // sdot v29.4s, v24.16b, v13.4b[0]\n"
- ".word 0x4f91e31e // sdot v30.4s, v24.16b, v17.4b[0]\n"
- ".word 0x4f95e31f // sdot v31.4s, v24.16b, v21.4b[0]\n"
- "ldr q24, [%[b_ptr0]]\n"
- ".word 0x4fa1e33a // sdot v26.4s, v25.16b, v1.4b[1]\n"
- ".word 0x4fa5e33b // sdot v27.4s, v25.16b, v5.4b[1]\n"
- ".word 0x4fa9e33c // sdot v28.4s, v25.16b, v9.4b[1]\n"
- ".word 0x4fade33d // sdot v29.4s, v25.16b, v13.4b[1]\n"
- ".word 0x4fb1e33e // sdot v30.4s, v25.16b, v17.4b[1]\n"
- ".word 0x4fb5e33f // sdot v31.4s, v25.16b, v21.4b[1]\n"
- "ldr q25, [%[b_ptr0], #0x10]\n"
- ".word 0x4f81eb1a // sdot v26.4s, v24.16b, v1.4b[2]\n"
- "add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f85eb1b // sdot v27.4s, v24.16b, v5.4b[2]\n"
- ".word 0x4f89eb1c // sdot v28.4s, v24.16b, v9.4b[2]\n"
- ".word 0x4f8deb1d // sdot v29.4s, v24.16b, v13.4b[2]\n"
- ".word 0x4f91eb1e // sdot v30.4s, v24.16b, v17.4b[2]\n"
- ".word 0x4f95eb1f // sdot v31.4s, v24.16b, v21.4b[2]\n"
- "ldr q24, [%[b_ptr0]]\n"
- ".word 0x4fa1eb3a // sdot v26.4s, v25.16b, v1.4b[3]\n"
- ".word 0x4fa5eb3b // sdot v27.4s, v25.16b, v5.4b[3]\n"
- ".word 0x4fa9eb3c // sdot v28.4s, v25.16b, v9.4b[3]\n"
- ".word 0x4fadeb3d // sdot v29.4s, v25.16b, v13.4b[3]\n"
- ".word 0x4fb1eb3e // sdot v30.4s, v25.16b, v17.4b[3]\n"
- ".word 0x4fb5eb3f // sdot v31.4s, v25.16b, v21.4b[3]\n"
- "ldr q25, [%[b_ptr0], #0x10]\n"
- ".word 0x4f82e31a // sdot v26.4s, v24.16b, v2.4b[0]\n"
- "add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f86e31b // sdot v27.4s, v24.16b, v6.4b[0]\n"
- ".word 0x4f8ae31c // sdot v28.4s, v24.16b, v10.4b[0]\n"
- ".word 0x4f8ee31d // sdot v29.4s, v24.16b, v14.4b[0]\n"
- ".word 0x4f92e31e // sdot v30.4s, v24.16b, v18.4b[0]\n"
- ".word 0x4f96e31f // sdot v31.4s, v24.16b, v22.4b[0]\n"
- "ldr q24, [%[b_ptr0]]\n"
- ".word 0x4fa2e33a // sdot v26.4s, v25.16b, v2.4b[1]\n"
- ".word 0x4fa6e33b // sdot v27.4s, v25.16b, v6.4b[1]\n"
- ".word 0x4faae33c // sdot v28.4s, v25.16b, v10.4b[1]\n"
- ".word 0x4faee33d // sdot v29.4s, v25.16b, v14.4b[1]\n"
- ".word 0x4fb2e33e // sdot v30.4s, v25.16b, v18.4b[1]\n"
- ".word 0x4fb6e33f // sdot v31.4s, v25.16b, v22.4b[1]\n"
- "ldr q25, [%[b_ptr0], #0x10]\n"
- ".word 0x4f82eb1a // sdot v26.4s, v24.16b, v2.4b[2]\n"
- "add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f86eb1b // sdot v27.4s, v24.16b, v6.4b[2]\n"
- ".word 0x4f8aeb1c // sdot v28.4s, v24.16b, v10.4b[2]\n"
- ".word 0x4f8eeb1d // sdot v29.4s, v24.16b, v14.4b[2]\n"
- ".word 0x4f92eb1e // sdot v30.4s, v24.16b, v18.4b[2]\n"
- ".word 0x4f96eb1f // sdot v31.4s, v24.16b, v22.4b[2]\n"
- ".word 0x4fa2eb3a // sdot v26.4s, v25.16b, v2.4b[3]\n"
- ".word 0x4fa6eb3b // sdot v27.4s, v25.16b, v6.4b[3]\n"
- ".word 0x4faaeb3c // sdot v28.4s, v25.16b, v10.4b[3]\n"
- ".word 0x4faeeb3d // sdot v29.4s, v25.16b, v14.4b[3]\n"
- ".word 0x4fb2eb3e // sdot v30.4s, v25.16b, v18.4b[3]\n"
- ".word 0x4fb6eb3f // sdot v31.4s, v25.16b, v22.4b[3]\n"
+ "add c_ptr5, c_ptr5, #0x10\n"
+ ".inst 0x4f8ce25e // sdot v30.4s, v18.16b, v12.4b[0]\n"
+ ".inst 0x4fa0e27a // sdot v26.4s, v19.16b, v0.4b[1]\n"
+ ".inst 0x4f8fe25f // sdot v31.4s, v18.16b, v15.4b[0]\n"
+ "ldr q18, [%[b_ptr0]]\n"
+ ".inst 0x4fa3e27b // sdot v27.4s, v19.16b, v3.4b[1]\n"
+ ".inst 0x4fa6e27c // sdot v28.4s, v19.16b, v6.4b[1]\n"
+ ".inst 0x4fa9e27d // sdot v29.4s, v19.16b, v9.4b[1]\n"
+ ".inst 0x4face27e // sdot v30.4s, v19.16b, v12.4b[1]\n"
+ ".inst 0x4fafe27f // sdot v31.4s, v19.16b, v15.4b[1]\n"
+ "ldr q19, [%[b_ptr0], #0x10]\n"
+ ".inst 0x4f80ea9a // sdot v26.4s, v20.16b, v0.4b[2]\n"
+ ".inst 0x4f83ea9b // sdot v27.4s, v20.16b, v3.4b[2]\n"
+ ".inst 0x4f86ea9c // sdot v28.4s, v20.16b, v6.4b[2]\n"
+ ".inst 0x4f89ea9d // sdot v29.4s, v20.16b, v9.4b[2]\n"
+ ".inst 0x4f8cea9e // sdot v30.4s, v20.16b, v12.4b[2]\n"
+ ".inst 0x4f8fea9f // sdot v31.4s, v20.16b, v15.4b[2]\n"
+ "ldr q20, [%[b_ptr0], #0x20]\n"
+ ".inst 0x4fa0eaba // sdot v26.4s, v21.16b, v0.4b[3]\n"
+ ".inst 0x4fa3eabb // sdot v27.4s, v21.16b, v3.4b[3]\n"
+ ".inst 0x4fa6eabc // sdot v28.4s, v21.16b, v6.4b[3]\n"
+ ".inst 0x4fa9eabd // sdot v29.4s, v21.16b, v9.4b[3]\n"
+ ".inst 0x4faceabe // sdot v30.4s, v21.16b, v12.4b[3]\n"
+ ".inst 0x4fafeabf // sdot v31.4s, v21.16b, v15.4b[3]\n"
+ "ldr q21, [%[b_ptr0], #0x30]\n"
+ ".inst 0x4f81e2da // sdot v26.4s, v22.16b, v1.4b[0]\n"
+ "add %[b_ptr0], %[b_ptr0], #0x40\n"
+ ".inst 0x4f84e2db // sdot v27.4s, v22.16b, v4.4b[0]\n"
+ ".inst 0x4f87e2dc // sdot v28.4s, v22.16b, v7.4b[0]\n"
+ ".inst 0x4f8ae2dd // sdot v29.4s, v22.16b, v10.4b[0]\n"
+ ".inst 0x4f8de2de // sdot v30.4s, v22.16b, v13.4b[0]\n"
+ ".inst 0x4f90e2df // sdot v31.4s, v22.16b, v16.4b[0]\n"
+ ".inst 0x4fa1e2fa // sdot v26.4s, v23.16b, v1.4b[1]\n"
+ ".inst 0x4fa4e2fb // sdot v27.4s, v23.16b, v4.4b[1]\n"
+ ".inst 0x4fa7e2fc // sdot v28.4s, v23.16b, v7.4b[1]\n"
+ ".inst 0x4faae2fd // sdot v29.4s, v23.16b, v10.4b[1]\n"
+ ".inst 0x4fade2fe // sdot v30.4s, v23.16b, v13.4b[1]\n"
+ ".inst 0x4fb0e2ff // sdot v31.4s, v23.16b, v16.4b[1]\n"
+ ".inst 0x4f81eb1a // sdot v26.4s, v24.16b, v1.4b[2]\n"
+ ".inst 0x4f84eb1b // sdot v27.4s, v24.16b, v4.4b[2]\n"
+ ".inst 0x4f87eb1c // sdot v28.4s, v24.16b, v7.4b[2]\n"
+ ".inst 0x4f8aeb1d // sdot v29.4s, v24.16b, v10.4b[2]\n"
+ ".inst 0x4f8deb1e // sdot v30.4s, v24.16b, v13.4b[2]\n"
+ ".inst 0x4f90eb1f // sdot v31.4s, v24.16b, v16.4b[2]\n"
+ ".inst 0x4fa1eb3a // sdot v26.4s, v25.16b, v1.4b[3]\n"
+ ".inst 0x4fa4eb3b // sdot v27.4s, v25.16b, v4.4b[3]\n"
+ ".inst 0x4fa7eb3c // sdot v28.4s, v25.16b, v7.4b[3]\n"
+ ".inst 0x4faaeb3d // sdot v29.4s, v25.16b, v10.4b[3]\n"
+ ".inst 0x4fadeb3e // sdot v30.4s, v25.16b, v13.4b[3]\n"
+ ".inst 0x4fb0eb3f // sdot v31.4s, v25.16b, v16.4b[3]\n"
+ ".inst 0x4f82e25a // sdot v26.4s, v18.16b, v2.4b[0]\n"
+ ".inst 0x4f85e25b // sdot v27.4s, v18.16b, v5.4b[0]\n"
+ ".inst 0x4f88e25c // sdot v28.4s, v18.16b, v8.4b[0]\n"
+ ".inst 0x4f8be25d // sdot v29.4s, v18.16b, v11.4b[0]\n"
+ ".inst 0x4f8ee25e // sdot v30.4s, v18.16b, v14.4b[0]\n"
+ ".inst 0x4f91e25f // sdot v31.4s, v18.16b, v17.4b[0]\n"
+ ".inst 0x4fa2e27a // sdot v26.4s, v19.16b, v2.4b[1]\n"
+ ".inst 0x4fa5e27b // sdot v27.4s, v19.16b, v5.4b[1]\n"
+ ".inst 0x4fa8e27c // sdot v28.4s, v19.16b, v8.4b[1]\n"
+ ".inst 0x4fabe27d // sdot v29.4s, v19.16b, v11.4b[1]\n"
+ ".inst 0x4faee27e // sdot v30.4s, v19.16b, v14.4b[1]\n"
+ ".inst 0x4fb1e27f // sdot v31.4s, v19.16b, v17.4b[1]\n"
+ ".inst 0x4f82ea9a // sdot v26.4s, v20.16b, v2.4b[2]\n"
+ ".inst 0x4f85ea9b // sdot v27.4s, v20.16b, v5.4b[2]\n"
+ ".inst 0x4f88ea9c // sdot v28.4s, v20.16b, v8.4b[2]\n"
+ ".inst 0x4f8bea9d // sdot v29.4s, v20.16b, v11.4b[2]\n"
+ ".inst 0x4f8eea9e // sdot v30.4s, v20.16b, v14.4b[2]\n"
+ ".inst 0x4f91ea9f // sdot v31.4s, v20.16b, v17.4b[2]\n"
+ ".inst 0x4fa2eaba // sdot v26.4s, v21.16b, v2.4b[3]\n"
+ ".inst 0x4fa5eabb // sdot v27.4s, v21.16b, v5.4b[3]\n"
+ ".inst 0x4fa8eabc // sdot v28.4s, v21.16b, v8.4b[3]\n"
+ ".inst 0x4fabeabd // sdot v29.4s, v21.16b, v11.4b[3]\n"
+ ".inst 0x4faeeabe // sdot v30.4s, v21.16b, v14.4b[3]\n"
+ ".inst 0x4fb1eabf // sdot v31.4s, v21.16b, v17.4b[3]\n"
"6:\n"
"str q26, [%[c_ptr0]]\n"
+ "add %[c_ptr0], %[c_ptr0], #0x10\n"
"str q27, [c_ptr1]\n"
+ "add c_ptr1, c_ptr1, #0x10\n"
"str q28, [c_ptr2]\n"
+ "add c_ptr2, c_ptr2, #0x10\n"
"str q29, [c_ptr3]\n"
+ "add c_ptr3, c_ptr3, #0x10\n"
"str q30, [c_ptr4]\n"
+ "add c_ptr4, c_ptr4, #0x10\n"
"str q31, [c_ptr5]\n"
+ "add c_ptr5, c_ptr5, #0x10\n"
".unreq a_ptr1\n"
".unreq a_ptr2\n"
".unreq a_ptr3\n"
@@ -1914,104 +1993,104 @@ void a64_smallK_hybrid_s8s32_dot_4x6_a55(const int8_t *A, int lda, const int8_t
"prfm PLDL1KEEP, [a_ptr5, #0xc0]\n"
"movi v31.4s, #0\n"
"prfm PLDL1KEEP, [a_ptr5, #0x100]\n"
- ".word 0x4f80e31a // sdot v26.4s, v24.16b, v0.4b[0]\n"
+ ".inst 0x4f80e31a // sdot v26.4s, v24.16b, v0.4b[0]\n"
"prfm PLDL1KEEP, [a_ptr5, #0x140]\n"
- ".word 0x4f84e31b // sdot v27.4s, v24.16b, v4.4b[0]\n"
+ ".inst 0x4f84e31b // sdot v27.4s, v24.16b, v4.4b[0]\n"
"prfm PLDL1KEEP, [a_ptr5, #0x180]\n"
- ".word 0x4f88e31c // sdot v28.4s, v24.16b, v8.4b[0]\n"
+ ".inst 0x4f88e31c // sdot v28.4s, v24.16b, v8.4b[0]\n"
"add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f8ce31d // sdot v29.4s, v24.16b, v12.4b[0]\n"
- ".word 0x4f90e31e // sdot v30.4s, v24.16b, v16.4b[0]\n"
- ".word 0x4f94e31f // sdot v31.4s, v24.16b, v20.4b[0]\n"
+ ".inst 0x4f8ce31d // sdot v29.4s, v24.16b, v12.4b[0]\n"
+ ".inst 0x4f90e31e // sdot v30.4s, v24.16b, v16.4b[0]\n"
+ ".inst 0x4f94e31f // sdot v31.4s, v24.16b, v20.4b[0]\n"
"ldr q24, [%[b_ptr0]]\n"
- ".word 0x4fa0e33a // sdot v26.4s, v25.16b, v0.4b[1]\n"
- ".word 0x4fa4e33b // sdot v27.4s, v25.16b, v4.4b[1]\n"
- ".word 0x4fa8e33c // sdot v28.4s, v25.16b, v8.4b[1]\n"
- ".word 0x4face33d // sdot v29.4s, v25.16b, v12.4b[1]\n"
- ".word 0x4fb0e33e // sdot v30.4s, v25.16b, v16.4b[1]\n"
- ".word 0x4fb4e33f // sdot v31.4s, v25.16b, v20.4b[1]\n"
+ ".inst 0x4fa0e33a // sdot v26.4s, v25.16b, v0.4b[1]\n"
+ ".inst 0x4fa4e33b // sdot v27.4s, v25.16b, v4.4b[1]\n"
+ ".inst 0x4fa8e33c // sdot v28.4s, v25.16b, v8.4b[1]\n"
+ ".inst 0x4face33d // sdot v29.4s, v25.16b, v12.4b[1]\n"
+ ".inst 0x4fb0e33e // sdot v30.4s, v25.16b, v16.4b[1]\n"
+ ".inst 0x4fb4e33f // sdot v31.4s, v25.16b, v20.4b[1]\n"
"ldr q25, [%[b_ptr0], #0x10]\n"
- ".word 0x4f80eb1a // sdot v26.4s, v24.16b, v0.4b[2]\n"
+ ".inst 0x4f80eb1a // sdot v26.4s, v24.16b, v0.4b[2]\n"
"add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f84eb1b // sdot v27.4s, v24.16b, v4.4b[2]\n"
- ".word 0x4f88eb1c // sdot v28.4s, v24.16b, v8.4b[2]\n"
- ".word 0x4f8ceb1d // sdot v29.4s, v24.16b, v12.4b[2]\n"
- ".word 0x4f90eb1e // sdot v30.4s, v24.16b, v16.4b[2]\n"
- ".word 0x4f94eb1f // sdot v31.4s, v24.16b, v20.4b[2]\n"
+ ".inst 0x4f84eb1b // sdot v27.4s, v24.16b, v4.4b[2]\n"
+ ".inst 0x4f88eb1c // sdot v28.4s, v24.16b, v8.4b[2]\n"
+ ".inst 0x4f8ceb1d // sdot v29.4s, v24.16b, v12.4b[2]\n"
+ ".inst 0x4f90eb1e // sdot v30.4s, v24.16b, v16.4b[2]\n"
+ ".inst 0x4f94eb1f // sdot v31.4s, v24.16b, v20.4b[2]\n"
"ldr q24, [%[b_ptr0]]\n"
- ".word 0x4fa0eb3a // sdot v26.4s, v25.16b, v0.4b[3]\n"
- ".word 0x4fa4eb3b // sdot v27.4s, v25.16b, v4.4b[3]\n"
- ".word 0x4fa8eb3c // sdot v28.4s, v25.16b, v8.4b[3]\n"
- ".word 0x4faceb3d // sdot v29.4s, v25.16b, v12.4b[3]\n"
- ".word 0x4fb0eb3e // sdot v30.4s, v25.16b, v16.4b[3]\n"
- ".word 0x4fb4eb3f // sdot v31.4s, v25.16b, v20.4b[3]\n"
+ ".inst 0x4fa0eb3a // sdot v26.4s, v25.16b, v0.4b[3]\n"
+ ".inst 0x4fa4eb3b // sdot v27.4s, v25.16b, v4.4b[3]\n"
+ ".inst 0x4fa8eb3c // sdot v28.4s, v25.16b, v8.4b[3]\n"
+ ".inst 0x4faceb3d // sdot v29.4s, v25.16b, v12.4b[3]\n"
+ ".inst 0x4fb0eb3e // sdot v30.4s, v25.16b, v16.4b[3]\n"
+ ".inst 0x4fb4eb3f // sdot v31.4s, v25.16b, v20.4b[3]\n"
"ldr q25, [%[b_ptr0], #0x10]\n"
- ".word 0x4f81e31a // sdot v26.4s, v24.16b, v1.4b[0]\n"
+ ".inst 0x4f81e31a // sdot v26.4s, v24.16b, v1.4b[0]\n"
"add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f85e31b // sdot v27.4s, v24.16b, v5.4b[0]\n"
- ".word 0x4f89e31c // sdot v28.4s, v24.16b, v9.4b[0]\n"
- ".word 0x4f8de31d // sdot v29.4s, v24.16b, v13.4b[0]\n"
- ".word 0x4f91e31e // sdot v30.4s, v24.16b, v17.4b[0]\n"
- ".word 0x4f95e31f // sdot v31.4s, v24.16b, v21.4b[0]\n"
+ ".inst 0x4f85e31b // sdot v27.4s, v24.16b, v5.4b[0]\n"
+ ".inst 0x4f89e31c // sdot v28.4s, v24.16b, v9.4b[0]\n"
+ ".inst 0x4f8de31d // sdot v29.4s, v24.16b, v13.4b[0]\n"
+ ".inst 0x4f91e31e // sdot v30.4s, v24.16b, v17.4b[0]\n"
+ ".inst 0x4f95e31f // sdot v31.4s, v24.16b, v21.4b[0]\n"
"ldr q24, [%[b_ptr0]]\n"
- ".word 0x4fa1e33a // sdot v26.4s, v25.16b, v1.4b[1]\n"
- ".word 0x4fa5e33b // sdot v27.4s, v25.16b, v5.4b[1]\n"
- ".word 0x4fa9e33c // sdot v28.4s, v25.16b, v9.4b[1]\n"
- ".word 0x4fade33d // sdot v29.4s, v25.16b, v13.4b[1]\n"
- ".word 0x4fb1e33e // sdot v30.4s, v25.16b, v17.4b[1]\n"
- ".word 0x4fb5e33f // sdot v31.4s, v25.16b, v21.4b[1]\n"
+ ".inst 0x4fa1e33a // sdot v26.4s, v25.16b, v1.4b[1]\n"
+ ".inst 0x4fa5e33b // sdot v27.4s, v25.16b, v5.4b[1]\n"
+ ".inst 0x4fa9e33c // sdot v28.4s, v25.16b, v9.4b[1]\n"
+ ".inst 0x4fade33d // sdot v29.4s, v25.16b, v13.4b[1]\n"
+ ".inst 0x4fb1e33e // sdot v30.4s, v25.16b, v17.4b[1]\n"
+ ".inst 0x4fb5e33f // sdot v31.4s, v25.16b, v21.4b[1]\n"
"ldr q25, [%[b_ptr0], #0x10]\n"
- ".word 0x4f81eb1a // sdot v26.4s, v24.16b, v1.4b[2]\n"
+ ".inst 0x4f81eb1a // sdot v26.4s, v24.16b, v1.4b[2]\n"
"add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f85eb1b // sdot v27.4s, v24.16b, v5.4b[2]\n"
- ".word 0x4f89eb1c // sdot v28.4s, v24.16b, v9.4b[2]\n"
- ".word 0x4f8deb1d // sdot v29.4s, v24.16b, v13.4b[2]\n"
- ".word 0x4f91eb1e // sdot v30.4s, v24.16b, v17.4b[2]\n"
- ".word 0x4f95eb1f // sdot v31.4s, v24.16b, v21.4b[2]\n"
+ ".inst 0x4f85eb1b // sdot v27.4s, v24.16b, v5.4b[2]\n"
+ ".inst 0x4f89eb1c // sdot v28.4s, v24.16b, v9.4b[2]\n"
+ ".inst 0x4f8deb1d // sdot v29.4s, v24.16b, v13.4b[2]\n"
+ ".inst 0x4f91eb1e // sdot v30.4s, v24.16b, v17.4b[2]\n"
+ ".inst 0x4f95eb1f // sdot v31.4s, v24.16b, v21.4b[2]\n"
"ldr q24, [%[b_ptr0]]\n"
- ".word 0x4fa1eb3a // sdot v26.4s, v25.16b, v1.4b[3]\n"
- ".word 0x4fa5eb3b // sdot v27.4s, v25.16b, v5.4b[3]\n"
- ".word 0x4fa9eb3c // sdot v28.4s, v25.16b, v9.4b[3]\n"
- ".word 0x4fadeb3d // sdot v29.4s, v25.16b, v13.4b[3]\n"
- ".word 0x4fb1eb3e // sdot v30.4s, v25.16b, v17.4b[3]\n"
- ".word 0x4fb5eb3f // sdot v31.4s, v25.16b, v21.4b[3]\n"
+ ".inst 0x4fa1eb3a // sdot v26.4s, v25.16b, v1.4b[3]\n"
+ ".inst 0x4fa5eb3b // sdot v27.4s, v25.16b, v5.4b[3]\n"
+ ".inst 0x4fa9eb3c // sdot v28.4s, v25.16b, v9.4b[3]\n"
+ ".inst 0x4fadeb3d // sdot v29.4s, v25.16b, v13.4b[3]\n"
+ ".inst 0x4fb1eb3e // sdot v30.4s, v25.16b, v17.4b[3]\n"
+ ".inst 0x4fb5eb3f // sdot v31.4s, v25.16b, v21.4b[3]\n"
"ldr q25, [%[b_ptr0], #0x10]\n"
- ".word 0x4f82e31a // sdot v26.4s, v24.16b, v2.4b[0]\n"
+ ".inst 0x4f82e31a // sdot v26.4s, v24.16b, v2.4b[0]\n"
"add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f86e31b // sdot v27.4s, v24.16b, v6.4b[0]\n"
- ".word 0x4f8ae31c // sdot v28.4s, v24.16b, v10.4b[0]\n"
- ".word 0x4f8ee31d // sdot v29.4s, v24.16b, v14.4b[0]\n"
- ".word 0x4f92e31e // sdot v30.4s, v24.16b, v18.4b[0]\n"
- ".word 0x4f96e31f // sdot v31.4s, v24.16b, v22.4b[0]\n"
+ ".inst 0x4f86e31b // sdot v27.4s, v24.16b, v6.4b[0]\n"
+ ".inst 0x4f8ae31c // sdot v28.4s, v24.16b, v10.4b[0]\n"
+ ".inst 0x4f8ee31d // sdot v29.4s, v24.16b, v14.4b[0]\n"
+ ".inst 0x4f92e31e // sdot v30.4s, v24.16b, v18.4b[0]\n"
+ ".inst 0x4f96e31f // sdot v31.4s, v24.16b, v22.4b[0]\n"
"ldr q24, [%[b_ptr0]]\n"
- ".word 0x4fa2e33a // sdot v26.4s, v25.16b, v2.4b[1]\n"
- ".word 0x4fa6e33b // sdot v27.4s, v25.16b, v6.4b[1]\n"
- ".word 0x4faae33c // sdot v28.4s, v25.16b, v10.4b[1]\n"
- ".word 0x4faee33d // sdot v29.4s, v25.16b, v14.4b[1]\n"
- ".word 0x4fb2e33e // sdot v30.4s, v25.16b, v18.4b[1]\n"
- ".word 0x4fb6e33f // sdot v31.4s, v25.16b, v22.4b[1]\n"
+ ".inst 0x4fa2e33a // sdot v26.4s, v25.16b, v2.4b[1]\n"
+ ".inst 0x4fa6e33b // sdot v27.4s, v25.16b, v6.4b[1]\n"
+ ".inst 0x4faae33c // sdot v28.4s, v25.16b, v10.4b[1]\n"
+ ".inst 0x4faee33d // sdot v29.4s, v25.16b, v14.4b[1]\n"
+ ".inst 0x4fb2e33e // sdot v30.4s, v25.16b, v18.4b[1]\n"
+ ".inst 0x4fb6e33f // sdot v31.4s, v25.16b, v22.4b[1]\n"
"ldr q25, [%[b_ptr0], #0x10]\n"
- ".word 0x4f82eb1a // sdot v26.4s, v24.16b, v2.4b[2]\n"
+ ".inst 0x4f82eb1a // sdot v26.4s, v24.16b, v2.4b[2]\n"
"add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f86eb1b // sdot v27.4s, v24.16b, v6.4b[2]\n"
- ".word 0x4f8aeb1c // sdot v28.4s, v24.16b, v10.4b[2]\n"
- ".word 0x4f8eeb1d // sdot v29.4s, v24.16b, v14.4b[2]\n"
- ".word 0x4f92eb1e // sdot v30.4s, v24.16b, v18.4b[2]\n"
- ".word 0x4f96eb1f // sdot v31.4s, v24.16b, v22.4b[2]\n"
+ ".inst 0x4f86eb1b // sdot v27.4s, v24.16b, v6.4b[2]\n"
+ ".inst 0x4f8aeb1c // sdot v28.4s, v24.16b, v10.4b[2]\n"
+ ".inst 0x4f8eeb1d // sdot v29.4s, v24.16b, v14.4b[2]\n"
+ ".inst 0x4f92eb1e // sdot v30.4s, v24.16b, v18.4b[2]\n"
+ ".inst 0x4f96eb1f // sdot v31.4s, v24.16b, v22.4b[2]\n"
"ldr q24, [%[b_ptr0]]\n"
- ".word 0x4fa2eb3a // sdot v26.4s, v25.16b, v2.4b[3]\n"
+ ".inst 0x4fa2eb3a // sdot v26.4s, v25.16b, v2.4b[3]\n"
"add %[b_ptr0], %[b_ptr0], #0x10\n"
- ".word 0x4fa6eb3b // sdot v27.4s, v25.16b, v6.4b[3]\n"
- ".word 0x4faaeb3c // sdot v28.4s, v25.16b, v10.4b[3]\n"
- ".word 0x4faeeb3d // sdot v29.4s, v25.16b, v14.4b[3]\n"
- ".word 0x4fb2eb3e // sdot v30.4s, v25.16b, v18.4b[3]\n"
- ".word 0x4fb6eb3f // sdot v31.4s, v25.16b, v22.4b[3]\n"
- ".word 0x4f83e31a // sdot v26.4s, v24.16b, v3.4b[0]\n"
- ".word 0x4f87e31b // sdot v27.4s, v24.16b, v7.4b[0]\n"
- ".word 0x4f8be31c // sdot v28.4s, v24.16b, v11.4b[0]\n"
- ".word 0x4f8fe31d // sdot v29.4s, v24.16b, v15.4b[0]\n"
- ".word 0x4f93e31e // sdot v30.4s, v24.16b, v19.4b[0]\n"
- ".word 0x4f97e31f // sdot v31.4s, v24.16b, v23.4b[0]\n"
+ ".inst 0x4fa6eb3b // sdot v27.4s, v25.16b, v6.4b[3]\n"
+ ".inst 0x4faaeb3c // sdot v28.4s, v25.16b, v10.4b[3]\n"
+ ".inst 0x4faeeb3d // sdot v29.4s, v25.16b, v14.4b[3]\n"
+ ".inst 0x4fb2eb3e // sdot v30.4s, v25.16b, v18.4b[3]\n"
+ ".inst 0x4fb6eb3f // sdot v31.4s, v25.16b, v22.4b[3]\n"
+ ".inst 0x4f83e31a // sdot v26.4s, v24.16b, v3.4b[0]\n"
+ ".inst 0x4f87e31b // sdot v27.4s, v24.16b, v7.4b[0]\n"
+ ".inst 0x4f8be31c // sdot v28.4s, v24.16b, v11.4b[0]\n"
+ ".inst 0x4f8fe31d // sdot v29.4s, v24.16b, v15.4b[0]\n"
+ ".inst 0x4f93e31e // sdot v30.4s, v24.16b, v19.4b[0]\n"
+ ".inst 0x4f97e31f // sdot v31.4s, v24.16b, v23.4b[0]\n"
"cbz %[loops], 6f\n"
"ldr d24, [%[b_ptr0]]\n"
"subs %[loops], %[loops], #0x1\n"
@@ -2023,265 +2102,283 @@ void a64_smallK_hybrid_s8s32_dot_4x6_a55(const int8_t *A, int lda, const int8_t
"ins v25.d[1], temploadreg1\n"
"b.eq 7f\n"
"8:\n"
- "str q26, [%[c_ptr0]], #0x10\n"
+ "str q26, [%[c_ptr0]]\n"
"subs %[loops], %[loops], #0x1\n"
"movi v26.4s, #0\n"
"ldr temploadreg0, [%[b_ptr0], #0x8]\n"
"ldr temploadreg1, [%[b_ptr0], #0x18]\n"
- "prfm PSTL1KEEP, [%[c_ptr0], #0x40]\n"
- "str q27, [c_ptr1], #0x10\n"
+ "add %[c_ptr0], %[c_ptr0], #0x10\n"
+ "str q27, [c_ptr1]\n"
+ "add c_ptr1, c_ptr1, #0x10\n"
"movi v27.4s, #0\n"
- ".word 0x4f80e31a // sdot v26.4s, v24.16b, v0.4b[0]\n"
- "str q28, [c_ptr2], #0x10\n"
- "prfm PSTL1KEEP, [c_ptr1, #0x40]\n"
+ "prfm PSTL1KEEP, [%[c_ptr0], #0x40]\n"
+ ".inst 0x4f80e31a // sdot v26.4s, v24.16b, v0.4b[0]\n"
+ "str q28, [c_ptr2]\n"
"movi v28.4s, #0\n"
- ".word 0x4f84e31b // sdot v27.4s, v24.16b, v4.4b[0]\n"
- "str q29, [c_ptr3], #0x10\n"
+ "add c_ptr2, c_ptr2, #0x10\n"
+ ".inst 0x4f84e31b // sdot v27.4s, v24.16b, v4.4b[0]\n"
+ "str q29, [c_ptr3]\n"
"movi v29.4s, #0\n"
- "prfm PSTL1KEEP, [c_ptr2, #0x40]\n"
- ".word 0x4f88e31c // sdot v28.4s, v24.16b, v8.4b[0]\n"
- "str q30, [c_ptr4], #0x10\n"
+ "add c_ptr3, c_ptr3, #0x10\n"
+ ".inst 0x4f88e31c // sdot v28.4s, v24.16b, v8.4b[0]\n"
+ "str q30, [c_ptr4]\n"
"movi v30.4s, #0\n"
- "prfm PSTL1KEEP, [c_ptr3, #0x40]\n"
- ".word 0x4f8ce31d // sdot v29.4s, v24.16b, v12.4b[0]\n"
- "str q31, [c_ptr5], #0x10\n"
+ "add c_ptr4, c_ptr4, #0x10\n"
+ ".inst 0x4f8ce31d // sdot v29.4s, v24.16b, v12.4b[0]\n"
+ "str q31, [c_ptr5]\n"
"movi v31.4s, #0\n"
- "prfm PSTL1KEEP, [c_ptr4, #0x40]\n"
- ".word 0x4f90e31e // sdot v30.4s, v24.16b, v16.4b[0]\n"
- "prfm PSTL1KEEP, [c_ptr5, #0x40]\n"
- ".word 0x4f94e31f // sdot v31.4s, v24.16b, v20.4b[0]\n"
+ "add c_ptr5, c_ptr5, #0x10\n"
+ ".inst 0x4f90e31e // sdot v30.4s, v24.16b, v16.4b[0]\n"
+ "prfm PSTL1KEEP, [c_ptr1, #0x40]\n"
+ ".inst 0x4f94e31f // sdot v31.4s, v24.16b, v20.4b[0]\n"
"ldr d24, [%[b_ptr0]]\n"
- ".word 0x4fa0e33a // sdot v26.4s, v25.16b, v0.4b[1]\n"
- ".word 0x4fa4e33b // sdot v27.4s, v25.16b, v4.4b[1]\n"
- ".word 0x4fa8e33c // sdot v28.4s, v25.16b, v8.4b[1]\n"
+ ".inst 0x4fa0e33a // sdot v26.4s, v25.16b, v0.4b[1]\n"
+ "prfm PSTL1KEEP, [c_ptr2, #0x40]\n"
+ ".inst 0x4fa4e33b // sdot v27.4s, v25.16b, v4.4b[1]\n"
"ins v24.d[1], temploadreg0\n"
- ".word 0x4face33d // sdot v29.4s, v25.16b, v12.4b[1]\n"
- ".word 0x4fb0e33e // sdot v30.4s, v25.16b, v16.4b[1]\n"
- ".word 0x4fb4e33f // sdot v31.4s, v25.16b, v20.4b[1]\n"
+ ".inst 0x4fa8e33c // sdot v28.4s, v25.16b, v8.4b[1]\n"
+ "prfm PSTL1KEEP, [c_ptr3, #0x40]\n"
+ ".inst 0x4face33d // sdot v29.4s, v25.16b, v12.4b[1]\n"
+ "prfm PSTL1KEEP, [c_ptr4, #0x40]\n"
+ ".inst 0x4fb0e33e // sdot v30.4s, v25.16b, v16.4b[1]\n"
+ "prfm PSTL1KEEP, [c_ptr5, #0x40]\n"
+ ".inst 0x4fb4e33f // sdot v31.4s, v25.16b, v20.4b[1]\n"
"ldr d25, [%[b_ptr0], #0x10]\n"
- ".word 0x4f80eb1a // sdot v26.4s, v24.16b, v0.4b[2]\n"
+ ".inst 0x4f80eb1a // sdot v26.4s, v24.16b, v0.4b[2]\n"
"add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f84eb1b // sdot v27.4s, v24.16b, v4.4b[2]\n"
+ ".inst 0x4f84eb1b // sdot v27.4s, v24.16b, v4.4b[2]\n"
"ins v25.d[1], temploadreg1\n"
- ".word 0x4f88eb1c // sdot v28.4s, v24.16b, v8.4b[2]\n"
+ ".inst 0x4f88eb1c // sdot v28.4s, v24.16b, v8.4b[2]\n"
"ldr temploadreg0, [%[b_ptr0], #0x8]\n"
- ".word 0x4f8ceb1d // sdot v29.4s, v24.16b, v12.4b[2]\n"
+ ".inst 0x4f8ceb1d // sdot v29.4s, v24.16b, v12.4b[2]\n"
"ldr temploadreg1, [%[b_ptr0], #0x18]\n"
- ".word 0x4f90eb1e // sdot v30.4s, v24.16b, v16.4b[2]\n"
- ".word 0x4f94eb1f // sdot v31.4s, v24.16b, v20.4b[2]\n"
+ ".inst 0x4f90eb1e // sdot v30.4s, v24.16b, v16.4b[2]\n"
+ ".inst 0x4f94eb1f // sdot v31.4s, v24.16b, v20.4b[2]\n"
"ldr d24, [%[b_ptr0]]\n"
- ".word 0x4fa0eb3a // sdot v26.4s, v25.16b, v0.4b[3]\n"
- ".word 0x4fa4eb3b // sdot v27.4s, v25.16b, v4.4b[3]\n"
- ".word 0x4fa8eb3c // sdot v28.4s, v25.16b, v8.4b[3]\n"
+ ".inst 0x4fa0eb3a // sdot v26.4s, v25.16b, v0.4b[3]\n"
+ ".inst 0x4fa4eb3b // sdot v27.4s, v25.16b, v4.4b[3]\n"
+ ".inst 0x4fa8eb3c // sdot v28.4s, v25.16b, v8.4b[3]\n"
"ins v24.d[1], temploadreg0\n"
- ".word 0x4faceb3d // sdot v29.4s, v25.16b, v12.4b[3]\n"
- ".word 0x4fb0eb3e // sdot v30.4s, v25.16b, v16.4b[3]\n"
- ".word 0x4fb4eb3f // sdot v31.4s, v25.16b, v20.4b[3]\n"
+ ".inst 0x4faceb3d // sdot v29.4s, v25.16b, v12.4b[3]\n"
+ ".inst 0x4fb0eb3e // sdot v30.4s, v25.16b, v16.4b[3]\n"
+ ".inst 0x4fb4eb3f // sdot v31.4s, v25.16b, v20.4b[3]\n"
"ldr d25, [%[b_ptr0], #0x10]\n"
- ".word 0x4f81e31a // sdot v26.4s, v24.16b, v1.4b[0]\n"
+ ".inst 0x4f81e31a // sdot v26.4s, v24.16b, v1.4b[0]\n"
"add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f85e31b // sdot v27.4s, v24.16b, v5.4b[0]\n"
+ ".inst 0x4f85e31b // sdot v27.4s, v24.16b, v5.4b[0]\n"
"ins v25.d[1], temploadreg1\n"
- ".word 0x4f89e31c // sdot v28.4s, v24.16b, v9.4b[0]\n"
+ ".inst 0x4f89e31c // sdot v28.4s, v24.16b, v9.4b[0]\n"
"ldr temploadreg0, [%[b_ptr0], #0x8]\n"
- ".word 0x4f8de31d // sdot v29.4s, v24.16b, v13.4b[0]\n"
+ ".inst 0x4f8de31d // sdot v29.4s, v24.16b, v13.4b[0]\n"
"ldr temploadreg1, [%[b_ptr0], #0x18]\n"
- ".word 0x4f91e31e // sdot v30.4s, v24.16b, v17.4b[0]\n"
- ".word 0x4f95e31f // sdot v31.4s, v24.16b, v21.4b[0]\n"
+ ".inst 0x4f91e31e // sdot v30.4s, v24.16b, v17.4b[0]\n"
+ ".inst 0x4f95e31f // sdot v31.4s, v24.16b, v21.4b[0]\n"
"ldr d24, [%[b_ptr0]]\n"
- ".word 0x4fa1e33a // sdot v26.4s, v25.16b, v1.4b[1]\n"
- ".word 0x4fa5e33b // sdot v27.4s, v25.16b, v5.4b[1]\n"
- ".word 0x4fa9e33c // sdot v28.4s, v25.16b, v9.4b[1]\n"
+ ".inst 0x4fa1e33a // sdot v26.4s, v25.16b, v1.4b[1]\n"
+ ".inst 0x4fa5e33b // sdot v27.4s, v25.16b, v5.4b[1]\n"
+ ".inst 0x4fa9e33c // sdot v28.4s, v25.16b, v9.4b[1]\n"
"ins v24.d[1], temploadreg0\n"
- ".word 0x4fade33d // sdot v29.4s, v25.16b, v13.4b[1]\n"
- ".word 0x4fb1e33e // sdot v30.4s, v25.16b, v17.4b[1]\n"
- ".word 0x4fb5e33f // sdot v31.4s, v25.16b, v21.4b[1]\n"
+ ".inst 0x4fade33d // sdot v29.4s, v25.16b, v13.4b[1]\n"
+ ".inst 0x4fb1e33e // sdot v30.4s, v25.16b, v17.4b[1]\n"
+ ".inst 0x4fb5e33f // sdot v31.4s, v25.16b, v21.4b[1]\n"
"ldr d25, [%[b_ptr0], #0x10]\n"
- ".word 0x4f81eb1a // sdot v26.4s, v24.16b, v1.4b[2]\n"
+ ".inst 0x4f81eb1a // sdot v26.4s, v24.16b, v1.4b[2]\n"
"add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f85eb1b // sdot v27.4s, v24.16b, v5.4b[2]\n"
+ ".inst 0x4f85eb1b // sdot v27.4s, v24.16b, v5.4b[2]\n"
"ins v25.d[1], temploadreg1\n"
- ".word 0x4f89eb1c // sdot v28.4s, v24.16b, v9.4b[2]\n"
+ ".inst 0x4f89eb1c // sdot v28.4s, v24.16b, v9.4b[2]\n"
"ldr temploadreg0, [%[b_ptr0], #0x8]\n"
- ".word 0x4f8deb1d // sdot v29.4s, v24.16b, v13.4b[2]\n"
+ ".inst 0x4f8deb1d // sdot v29.4s, v24.16b, v13.4b[2]\n"
"ldr temploadreg1, [%[b_ptr0], #0x18]\n"
- ".word 0x4f91eb1e // sdot v30.4s, v24.16b, v17.4b[2]\n"
- ".word 0x4f95eb1f // sdot v31.4s, v24.16b, v21.4b[2]\n"
+ ".inst 0x4f91eb1e // sdot v30.4s, v24.16b, v17.4b[2]\n"
+ ".inst 0x4f95eb1f // sdot v31.4s, v24.16b, v21.4b[2]\n"
"ldr d24, [%[b_ptr0]]\n"
- ".word 0x4fa1eb3a // sdot v26.4s, v25.16b, v1.4b[3]\n"
- ".word 0x4fa5eb3b // sdot v27.4s, v25.16b, v5.4b[3]\n"
- ".word 0x4fa9eb3c // sdot v28.4s, v25.16b, v9.4b[3]\n"
+ ".inst 0x4fa1eb3a // sdot v26.4s, v25.16b, v1.4b[3]\n"
+ ".inst 0x4fa5eb3b // sdot v27.4s, v25.16b, v5.4b[3]\n"
+ ".inst 0x4fa9eb3c // sdot v28.4s, v25.16b, v9.4b[3]\n"
"ins v24.d[1], temploadreg0\n"
- ".word 0x4fadeb3d // sdot v29.4s, v25.16b, v13.4b[3]\n"
- ".word 0x4fb1eb3e // sdot v30.4s, v25.16b, v17.4b[3]\n"
- ".word 0x4fb5eb3f // sdot v31.4s, v25.16b, v21.4b[3]\n"
+ ".inst 0x4fadeb3d // sdot v29.4s, v25.16b, v13.4b[3]\n"
+ ".inst 0x4fb1eb3e // sdot v30.4s, v25.16b, v17.4b[3]\n"
+ ".inst 0x4fb5eb3f // sdot v31.4s, v25.16b, v21.4b[3]\n"
"ldr d25, [%[b_ptr0], #0x10]\n"
- ".word 0x4f82e31a // sdot v26.4s, v24.16b, v2.4b[0]\n"
+ ".inst 0x4f82e31a // sdot v26.4s, v24.16b, v2.4b[0]\n"
"add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f86e31b // sdot v27.4s, v24.16b, v6.4b[0]\n"
+ ".inst 0x4f86e31b // sdot v27.4s, v24.16b, v6.4b[0]\n"
"ins v25.d[1], temploadreg1\n"
- ".word 0x4f8ae31c // sdot v28.4s, v24.16b, v10.4b[0]\n"
+ ".inst 0x4f8ae31c // sdot v28.4s, v24.16b, v10.4b[0]\n"
"ldr temploadreg0, [%[b_ptr0], #0x8]\n"
- ".word 0x4f8ee31d // sdot v29.4s, v24.16b, v14.4b[0]\n"
+ ".inst 0x4f8ee31d // sdot v29.4s, v24.16b, v14.4b[0]\n"
"ldr temploadreg1, [%[b_ptr0], #0x18]\n"
- ".word 0x4f92e31e // sdot v30.4s, v24.16b, v18.4b[0]\n"
- ".word 0x4f96e31f // sdot v31.4s, v24.16b, v22.4b[0]\n"
+ ".inst 0x4f92e31e // sdot v30.4s, v24.16b, v18.4b[0]\n"
+ ".inst 0x4f96e31f // sdot v31.4s, v24.16b, v22.4b[0]\n"
"ldr d24, [%[b_ptr0]]\n"
- ".word 0x4fa2e33a // sdot v26.4s, v25.16b, v2.4b[1]\n"
- ".word 0x4fa6e33b // sdot v27.4s, v25.16b, v6.4b[1]\n"
- ".word 0x4faae33c // sdot v28.4s, v25.16b, v10.4b[1]\n"
+ ".inst 0x4fa2e33a // sdot v26.4s, v25.16b, v2.4b[1]\n"
+ ".inst 0x4fa6e33b // sdot v27.4s, v25.16b, v6.4b[1]\n"
+ ".inst 0x4faae33c // sdot v28.4s, v25.16b, v10.4b[1]\n"
"ins v24.d[1], temploadreg0\n"
- ".word 0x4faee33d // sdot v29.4s, v25.16b, v14.4b[1]\n"
- ".word 0x4fb2e33e // sdot v30.4s, v25.16b, v18.4b[1]\n"
- ".word 0x4fb6e33f // sdot v31.4s, v25.16b, v22.4b[1]\n"
+ ".inst 0x4faee33d // sdot v29.4s, v25.16b, v14.4b[1]\n"
+ ".inst 0x4fb2e33e // sdot v30.4s, v25.16b, v18.4b[1]\n"
+ ".inst 0x4fb6e33f // sdot v31.4s, v25.16b, v22.4b[1]\n"
"ldr d25, [%[b_ptr0], #0x10]\n"
- ".word 0x4f82eb1a // sdot v26.4s, v24.16b, v2.4b[2]\n"
+ ".inst 0x4f82eb1a // sdot v26.4s, v24.16b, v2.4b[2]\n"
"add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f86eb1b // sdot v27.4s, v24.16b, v6.4b[2]\n"
+ ".inst 0x4f86eb1b // sdot v27.4s, v24.16b, v6.4b[2]\n"
"ins v25.d[1], temploadreg1\n"
- ".word 0x4f8aeb1c // sdot v28.4s, v24.16b, v10.4b[2]\n"
+ ".inst 0x4f8aeb1c // sdot v28.4s, v24.16b, v10.4b[2]\n"
"ldr temploadreg0, [%[b_ptr0], #0x8]\n"
- ".word 0x4f8eeb1d // sdot v29.4s, v24.16b, v14.4b[2]\n"
- ".word 0x4f92eb1e // sdot v30.4s, v24.16b, v18.4b[2]\n"
- ".word 0x4f96eb1f // sdot v31.4s, v24.16b, v22.4b[2]\n"
+ ".inst 0x4f8eeb1d // sdot v29.4s, v24.16b, v14.4b[2]\n"
+ ".inst 0x4f92eb1e // sdot v30.4s, v24.16b, v18.4b[2]\n"
+ ".inst 0x4f96eb1f // sdot v31.4s, v24.16b, v22.4b[2]\n"
"ldr d24, [%[b_ptr0]]\n"
- ".word 0x4fa2eb3a // sdot v26.4s, v25.16b, v2.4b[3]\n"
+ ".inst 0x4fa2eb3a // sdot v26.4s, v25.16b, v2.4b[3]\n"
"add %[b_ptr0], %[b_ptr0], #0x10\n"
- ".word 0x4fa6eb3b // sdot v27.4s, v25.16b, v6.4b[3]\n"
+ ".inst 0x4fa6eb3b // sdot v27.4s, v25.16b, v6.4b[3]\n"
"ins v24.d[1], temploadreg0\n"
- ".word 0x4faaeb3c // sdot v28.4s, v25.16b, v10.4b[3]\n"
+ ".inst 0x4faaeb3c // sdot v28.4s, v25.16b, v10.4b[3]\n"
"ldr temploadreg0, [%[b_ptr0], #0x8]\n"
- ".word 0x4faeeb3d // sdot v29.4s, v25.16b, v14.4b[3]\n"
+ ".inst 0x4faeeb3d // sdot v29.4s, v25.16b, v14.4b[3]\n"
"ldr temploadreg1, [%[b_ptr0], #0x18]\n"
- ".word 0x4fb2eb3e // sdot v30.4s, v25.16b, v18.4b[3]\n"
- ".word 0x4fb6eb3f // sdot v31.4s, v25.16b, v22.4b[3]\n"
+ ".inst 0x4fb2eb3e // sdot v30.4s, v25.16b, v18.4b[3]\n"
+ ".inst 0x4fb6eb3f // sdot v31.4s, v25.16b, v22.4b[3]\n"
"ldr d25, [%[b_ptr0], #0x10]\n"
- ".word 0x4f83e31a // sdot v26.4s, v24.16b, v3.4b[0]\n"
- ".word 0x4f87e31b // sdot v27.4s, v24.16b, v7.4b[0]\n"
- ".word 0x4f8be31c // sdot v28.4s, v24.16b, v11.4b[0]\n"
+ ".inst 0x4f83e31a // sdot v26.4s, v24.16b, v3.4b[0]\n"
+ ".inst 0x4f87e31b // sdot v27.4s, v24.16b, v7.4b[0]\n"
+ ".inst 0x4f8be31c // sdot v28.4s, v24.16b, v11.4b[0]\n"
"ins v25.d[1], temploadreg1\n"
- ".word 0x4f8fe31d // sdot v29.4s, v24.16b, v15.4b[0]\n"
- ".word 0x4f93e31e // sdot v30.4s, v24.16b, v19.4b[0]\n"
- ".word 0x4f97e31f // sdot v31.4s, v24.16b, v23.4b[0]\n"
+ ".inst 0x4f8fe31d // sdot v29.4s, v24.16b, v15.4b[0]\n"
+ ".inst 0x4f93e31e // sdot v30.4s, v24.16b, v19.4b[0]\n"
+ ".inst 0x4f97e31f // sdot v31.4s, v24.16b, v23.4b[0]\n"
"ldr d24, [%[b_ptr0]]\n"
"add %[b_ptr0], %[b_ptr0], #0x20\n"
"ins v24.d[1], temploadreg0\n"
"b.ne 8b\n"
"7:\n"
- "str q26, [%[c_ptr0]], #0x10\n"
+ "str q26, [%[c_ptr0]]\n"
+ "add %[c_ptr0], %[c_ptr0], #0x10\n"
"movi v26.4s, #0\n"
- "str q27, [c_ptr1], #0x10\n"
+ "str q27, [c_ptr1]\n"
+ "add c_ptr1, c_ptr1, #0x10\n"
"movi v27.4s, #0\n"
- ".word 0x4f80e31a // sdot v26.4s, v24.16b, v0.4b[0]\n"
- "str q28, [c_ptr2], #0x10\n"
+ ".inst 0x4f80e31a // sdot v26.4s, v24.16b, v0.4b[0]\n"
+ "str q28, [c_ptr2]\n"
"movi v28.4s, #0\n"
- ".word 0x4f84e31b // sdot v27.4s, v24.16b, v4.4b[0]\n"
- ".word 0x4fa0e33a // sdot v26.4s, v25.16b, v0.4b[1]\n"
- "str q29, [c_ptr3], #0x10\n"
+ "add c_ptr2, c_ptr2, #0x10\n"
+ ".inst 0x4f84e31b // sdot v27.4s, v24.16b, v4.4b[0]\n"
+ "str q29, [c_ptr3]\n"
"movi v29.4s, #0\n"
- ".word 0x4f88e31c // sdot v28.4s, v24.16b, v8.4b[0]\n"
- ".word 0x4fa4e33b // sdot v27.4s, v25.16b, v4.4b[1]\n"
- "str q30, [c_ptr4], #0x10\n"
+ "add c_ptr3, c_ptr3, #0x10\n"
+ ".inst 0x4f88e31c // sdot v28.4s, v24.16b, v8.4b[0]\n"
+ "str q30, [c_ptr4]\n"
"movi v30.4s, #0\n"
- ".word 0x4f8ce31d // sdot v29.4s, v24.16b, v12.4b[0]\n"
- ".word 0x4fa8e33c // sdot v28.4s, v25.16b, v8.4b[1]\n"
- "str q31, [c_ptr5], #0x10\n"
+ "add c_ptr4, c_ptr4, #0x10\n"
+ ".inst 0x4f8ce31d // sdot v29.4s, v24.16b, v12.4b[0]\n"
+ "str q31, [c_ptr5]\n"
"movi v31.4s, #0\n"
- ".word 0x4f90e31e // sdot v30.4s, v24.16b, v16.4b[0]\n"
- ".word 0x4face33d // sdot v29.4s, v25.16b, v12.4b[1]\n"
- ".word 0x4f94e31f // sdot v31.4s, v24.16b, v20.4b[0]\n"
+ "add c_ptr5, c_ptr5, #0x10\n"
+ ".inst 0x4f90e31e // sdot v30.4s, v24.16b, v16.4b[0]\n"
+ ".inst 0x4fa0e33a // sdot v26.4s, v25.16b, v0.4b[1]\n"
+ ".inst 0x4f94e31f // sdot v31.4s, v24.16b, v20.4b[0]\n"
"ldr q24, [%[b_ptr0]]\n"
- ".word 0x4fb0e33e // sdot v30.4s, v25.16b, v16.4b[1]\n"
- ".word 0x4fb4e33f // sdot v31.4s, v25.16b, v20.4b[1]\n"
+ ".inst 0x4fa4e33b // sdot v27.4s, v25.16b, v4.4b[1]\n"
+ ".inst 0x4fa8e33c // sdot v28.4s, v25.16b, v8.4b[1]\n"
+ ".inst 0x4face33d // sdot v29.4s, v25.16b, v12.4b[1]\n"
+ ".inst 0x4fb0e33e // sdot v30.4s, v25.16b, v16.4b[1]\n"
+ ".inst 0x4fb4e33f // sdot v31.4s, v25.16b, v20.4b[1]\n"
"ldr q25, [%[b_ptr0], #0x10]\n"
- ".word 0x4f80eb1a // sdot v26.4s, v24.16b, v0.4b[2]\n"
+ ".inst 0x4f80eb1a // sdot v26.4s, v24.16b, v0.4b[2]\n"
"add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f84eb1b // sdot v27.4s, v24.16b, v4.4b[2]\n"
- ".word 0x4f88eb1c // sdot v28.4s, v24.16b, v8.4b[2]\n"
- ".word 0x4f8ceb1d // sdot v29.4s, v24.16b, v12.4b[2]\n"
- ".word 0x4f90eb1e // sdot v30.4s, v24.16b, v16.4b[2]\n"
- ".word 0x4f94eb1f // sdot v31.4s, v24.16b, v20.4b[2]\n"
+ ".inst 0x4f84eb1b // sdot v27.4s, v24.16b, v4.4b[2]\n"
+ ".inst 0x4f88eb1c // sdot v28.4s, v24.16b, v8.4b[2]\n"
+ ".inst 0x4f8ceb1d // sdot v29.4s, v24.16b, v12.4b[2]\n"
+ ".inst 0x4f90eb1e // sdot v30.4s, v24.16b, v16.4b[2]\n"
+ ".inst 0x4f94eb1f // sdot v31.4s, v24.16b, v20.4b[2]\n"
"ldr q24, [%[b_ptr0]]\n"
- ".word 0x4fa0eb3a // sdot v26.4s, v25.16b, v0.4b[3]\n"
- ".word 0x4fa4eb3b // sdot v27.4s, v25.16b, v4.4b[3]\n"
- ".word 0x4fa8eb3c // sdot v28.4s, v25.16b, v8.4b[3]\n"
- ".word 0x4faceb3d // sdot v29.4s, v25.16b, v12.4b[3]\n"
- ".word 0x4fb0eb3e // sdot v30.4s, v25.16b, v16.4b[3]\n"
- ".word 0x4fb4eb3f // sdot v31.4s, v25.16b, v20.4b[3]\n"
+ ".inst 0x4fa0eb3a // sdot v26.4s, v25.16b, v0.4b[3]\n"
+ ".inst 0x4fa4eb3b // sdot v27.4s, v25.16b, v4.4b[3]\n"
+ ".inst 0x4fa8eb3c // sdot v28.4s, v25.16b, v8.4b[3]\n"
+ ".inst 0x4faceb3d // sdot v29.4s, v25.16b, v12.4b[3]\n"
+ ".inst 0x4fb0eb3e // sdot v30.4s, v25.16b, v16.4b[3]\n"
+ ".inst 0x4fb4eb3f // sdot v31.4s, v25.16b, v20.4b[3]\n"
"ldr q25, [%[b_ptr0], #0x10]\n"
- ".word 0x4f81e31a // sdot v26.4s, v24.16b, v1.4b[0]\n"
+ ".inst 0x4f81e31a // sdot v26.4s, v24.16b, v1.4b[0]\n"
"add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f85e31b // sdot v27.4s, v24.16b, v5.4b[0]\n"
- ".word 0x4f89e31c // sdot v28.4s, v24.16b, v9.4b[0]\n"
- ".word 0x4f8de31d // sdot v29.4s, v24.16b, v13.4b[0]\n"
- ".word 0x4f91e31e // sdot v30.4s, v24.16b, v17.4b[0]\n"
- ".word 0x4f95e31f // sdot v31.4s, v24.16b, v21.4b[0]\n"
+ ".inst 0x4f85e31b // sdot v27.4s, v24.16b, v5.4b[0]\n"
+ ".inst 0x4f89e31c // sdot v28.4s, v24.16b, v9.4b[0]\n"
+ ".inst 0x4f8de31d // sdot v29.4s, v24.16b, v13.4b[0]\n"
+ ".inst 0x4f91e31e // sdot v30.4s, v24.16b, v17.4b[0]\n"
+ ".inst 0x4f95e31f // sdot v31.4s, v24.16b, v21.4b[0]\n"
"ldr q24, [%[b_ptr0]]\n"
- ".word 0x4fa1e33a // sdot v26.4s, v25.16b, v1.4b[1]\n"
- ".word 0x4fa5e33b // sdot v27.4s, v25.16b, v5.4b[1]\n"
- ".word 0x4fa9e33c // sdot v28.4s, v25.16b, v9.4b[1]\n"
- ".word 0x4fade33d // sdot v29.4s, v25.16b, v13.4b[1]\n"
- ".word 0x4fb1e33e // sdot v30.4s, v25.16b, v17.4b[1]\n"
- ".word 0x4fb5e33f // sdot v31.4s, v25.16b, v21.4b[1]\n"
+ ".inst 0x4fa1e33a // sdot v26.4s, v25.16b, v1.4b[1]\n"
+ ".inst 0x4fa5e33b // sdot v27.4s, v25.16b, v5.4b[1]\n"
+ ".inst 0x4fa9e33c // sdot v28.4s, v25.16b, v9.4b[1]\n"
+ ".inst 0x4fade33d // sdot v29.4s, v25.16b, v13.4b[1]\n"
+ ".inst 0x4fb1e33e // sdot v30.4s, v25.16b, v17.4b[1]\n"
+ ".inst 0x4fb5e33f // sdot v31.4s, v25.16b, v21.4b[1]\n"
"ldr q25, [%[b_ptr0], #0x10]\n"
- ".word 0x4f81eb1a // sdot v26.4s, v24.16b, v1.4b[2]\n"
+ ".inst 0x4f81eb1a // sdot v26.4s, v24.16b, v1.4b[2]\n"
"add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f85eb1b // sdot v27.4s, v24.16b, v5.4b[2]\n"
- ".word 0x4f89eb1c // sdot v28.4s, v24.16b, v9.4b[2]\n"
- ".word 0x4f8deb1d // sdot v29.4s, v24.16b, v13.4b[2]\n"
- ".word 0x4f91eb1e // sdot v30.4s, v24.16b, v17.4b[2]\n"
- ".word 0x4f95eb1f // sdot v31.4s, v24.16b, v21.4b[2]\n"
+ ".inst 0x4f85eb1b // sdot v27.4s, v24.16b, v5.4b[2]\n"
+ ".inst 0x4f89eb1c // sdot v28.4s, v24.16b, v9.4b[2]\n"
+ ".inst 0x4f8deb1d // sdot v29.4s, v24.16b, v13.4b[2]\n"
+ ".inst 0x4f91eb1e // sdot v30.4s, v24.16b, v17.4b[2]\n"
+ ".inst 0x4f95eb1f // sdot v31.4s, v24.16b, v21.4b[2]\n"
"ldr q24, [%[b_ptr0]]\n"
- ".word 0x4fa1eb3a // sdot v26.4s, v25.16b, v1.4b[3]\n"
- ".word 0x4fa5eb3b // sdot v27.4s, v25.16b, v5.4b[3]\n"
- ".word 0x4fa9eb3c // sdot v28.4s, v25.16b, v9.4b[3]\n"
- ".word 0x4fadeb3d // sdot v29.4s, v25.16b, v13.4b[3]\n"
- ".word 0x4fb1eb3e // sdot v30.4s, v25.16b, v17.4b[3]\n"
- ".word 0x4fb5eb3f // sdot v31.4s, v25.16b, v21.4b[3]\n"
+ ".inst 0x4fa1eb3a // sdot v26.4s, v25.16b, v1.4b[3]\n"
+ ".inst 0x4fa5eb3b // sdot v27.4s, v25.16b, v5.4b[3]\n"
+ ".inst 0x4fa9eb3c // sdot v28.4s, v25.16b, v9.4b[3]\n"
+ ".inst 0x4fadeb3d // sdot v29.4s, v25.16b, v13.4b[3]\n"
+ ".inst 0x4fb1eb3e // sdot v30.4s, v25.16b, v17.4b[3]\n"
+ ".inst 0x4fb5eb3f // sdot v31.4s, v25.16b, v21.4b[3]\n"
"ldr q25, [%[b_ptr0], #0x10]\n"
- ".word 0x4f82e31a // sdot v26.4s, v24.16b, v2.4b[0]\n"
+ ".inst 0x4f82e31a // sdot v26.4s, v24.16b, v2.4b[0]\n"
"add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f86e31b // sdot v27.4s, v24.16b, v6.4b[0]\n"
- ".word 0x4f8ae31c // sdot v28.4s, v24.16b, v10.4b[0]\n"
- ".word 0x4f8ee31d // sdot v29.4s, v24.16b, v14.4b[0]\n"
- ".word 0x4f92e31e // sdot v30.4s, v24.16b, v18.4b[0]\n"
- ".word 0x4f96e31f // sdot v31.4s, v24.16b, v22.4b[0]\n"
+ ".inst 0x4f86e31b // sdot v27.4s, v24.16b, v6.4b[0]\n"
+ ".inst 0x4f8ae31c // sdot v28.4s, v24.16b, v10.4b[0]\n"
+ ".inst 0x4f8ee31d // sdot v29.4s, v24.16b, v14.4b[0]\n"
+ ".inst 0x4f92e31e // sdot v30.4s, v24.16b, v18.4b[0]\n"
+ ".inst 0x4f96e31f // sdot v31.4s, v24.16b, v22.4b[0]\n"
"ldr q24, [%[b_ptr0]]\n"
- ".word 0x4fa2e33a // sdot v26.4s, v25.16b, v2.4b[1]\n"
- ".word 0x4fa6e33b // sdot v27.4s, v25.16b, v6.4b[1]\n"
- ".word 0x4faae33c // sdot v28.4s, v25.16b, v10.4b[1]\n"
- ".word 0x4faee33d // sdot v29.4s, v25.16b, v14.4b[1]\n"
- ".word 0x4fb2e33e // sdot v30.4s, v25.16b, v18.4b[1]\n"
- ".word 0x4fb6e33f // sdot v31.4s, v25.16b, v22.4b[1]\n"
+ ".inst 0x4fa2e33a // sdot v26.4s, v25.16b, v2.4b[1]\n"
+ ".inst 0x4fa6e33b // sdot v27.4s, v25.16b, v6.4b[1]\n"
+ ".inst 0x4faae33c // sdot v28.4s, v25.16b, v10.4b[1]\n"
+ ".inst 0x4faee33d // sdot v29.4s, v25.16b, v14.4b[1]\n"
+ ".inst 0x4fb2e33e // sdot v30.4s, v25.16b, v18.4b[1]\n"
+ ".inst 0x4fb6e33f // sdot v31.4s, v25.16b, v22.4b[1]\n"
"ldr q25, [%[b_ptr0], #0x10]\n"
- ".word 0x4f82eb1a // sdot v26.4s, v24.16b, v2.4b[2]\n"
+ ".inst 0x4f82eb1a // sdot v26.4s, v24.16b, v2.4b[2]\n"
"add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f86eb1b // sdot v27.4s, v24.16b, v6.4b[2]\n"
- ".word 0x4f8aeb1c // sdot v28.4s, v24.16b, v10.4b[2]\n"
- ".word 0x4f8eeb1d // sdot v29.4s, v24.16b, v14.4b[2]\n"
- ".word 0x4f92eb1e // sdot v30.4s, v24.16b, v18.4b[2]\n"
- ".word 0x4f96eb1f // sdot v31.4s, v24.16b, v22.4b[2]\n"
+ ".inst 0x4f86eb1b // sdot v27.4s, v24.16b, v6.4b[2]\n"
+ ".inst 0x4f8aeb1c // sdot v28.4s, v24.16b, v10.4b[2]\n"
+ ".inst 0x4f8eeb1d // sdot v29.4s, v24.16b, v14.4b[2]\n"
+ ".inst 0x4f92eb1e // sdot v30.4s, v24.16b, v18.4b[2]\n"
+ ".inst 0x4f96eb1f // sdot v31.4s, v24.16b, v22.4b[2]\n"
"ldr q24, [%[b_ptr0]]\n"
- ".word 0x4fa2eb3a // sdot v26.4s, v25.16b, v2.4b[3]\n"
+ ".inst 0x4fa2eb3a // sdot v26.4s, v25.16b, v2.4b[3]\n"
"add %[b_ptr0], %[b_ptr0], #0x10\n"
- ".word 0x4fa6eb3b // sdot v27.4s, v25.16b, v6.4b[3]\n"
- ".word 0x4faaeb3c // sdot v28.4s, v25.16b, v10.4b[3]\n"
- ".word 0x4faeeb3d // sdot v29.4s, v25.16b, v14.4b[3]\n"
- ".word 0x4fb2eb3e // sdot v30.4s, v25.16b, v18.4b[3]\n"
- ".word 0x4fb6eb3f // sdot v31.4s, v25.16b, v22.4b[3]\n"
- ".word 0x4f83e31a // sdot v26.4s, v24.16b, v3.4b[0]\n"
- ".word 0x4f87e31b // sdot v27.4s, v24.16b, v7.4b[0]\n"
- ".word 0x4f8be31c // sdot v28.4s, v24.16b, v11.4b[0]\n"
- ".word 0x4f8fe31d // sdot v29.4s, v24.16b, v15.4b[0]\n"
- ".word 0x4f93e31e // sdot v30.4s, v24.16b, v19.4b[0]\n"
- ".word 0x4f97e31f // sdot v31.4s, v24.16b, v23.4b[0]\n"
+ ".inst 0x4fa6eb3b // sdot v27.4s, v25.16b, v6.4b[3]\n"
+ ".inst 0x4faaeb3c // sdot v28.4s, v25.16b, v10.4b[3]\n"
+ ".inst 0x4faeeb3d // sdot v29.4s, v25.16b, v14.4b[3]\n"
+ ".inst 0x4fb2eb3e // sdot v30.4s, v25.16b, v18.4b[3]\n"
+ ".inst 0x4fb6eb3f // sdot v31.4s, v25.16b, v22.4b[3]\n"
+ ".inst 0x4f83e31a // sdot v26.4s, v24.16b, v3.4b[0]\n"
+ ".inst 0x4f87e31b // sdot v27.4s, v24.16b, v7.4b[0]\n"
+ ".inst 0x4f8be31c // sdot v28.4s, v24.16b, v11.4b[0]\n"
+ ".inst 0x4f8fe31d // sdot v29.4s, v24.16b, v15.4b[0]\n"
+ ".inst 0x4f93e31e // sdot v30.4s, v24.16b, v19.4b[0]\n"
+ ".inst 0x4f97e31f // sdot v31.4s, v24.16b, v23.4b[0]\n"
"6:\n"
"str q26, [%[c_ptr0]]\n"
+ "add %[c_ptr0], %[c_ptr0], #0x10\n"
"str q27, [c_ptr1]\n"
+ "add c_ptr1, c_ptr1, #0x10\n"
"str q28, [c_ptr2]\n"
+ "add c_ptr2, c_ptr2, #0x10\n"
"str q29, [c_ptr3]\n"
+ "add c_ptr3, c_ptr3, #0x10\n"
"str q30, [c_ptr4]\n"
+ "add c_ptr4, c_ptr4, #0x10\n"
"str q31, [c_ptr5]\n"
+ "add c_ptr5, c_ptr5, #0x10\n"
".unreq a_ptr1\n"
".unreq a_ptr2\n"
".unreq a_ptr3\n"
@@ -2420,111 +2517,111 @@ void a64_smallK_hybrid_s8s32_dot_4x6_a55(const int8_t *A, int lda, const int8_t
"prfm PLDL1KEEP, [a_ptr5, #0xc0]\n"
"movi v31.4s, #0\n"
"prfm PLDL1KEEP, [a_ptr5, #0x100]\n"
- ".word 0x4f80e31a // sdot v26.4s, v24.16b, v0.4b[0]\n"
+ ".inst 0x4f80e31a // sdot v26.4s, v24.16b, v0.4b[0]\n"
"prfm PLDL1KEEP, [a_ptr5, #0x140]\n"
- ".word 0x4f84e31b // sdot v27.4s, v24.16b, v4.4b[0]\n"
+ ".inst 0x4f84e31b // sdot v27.4s, v24.16b, v4.4b[0]\n"
"prfm PLDL1KEEP, [a_ptr5, #0x180]\n"
- ".word 0x4f88e31c // sdot v28.4s, v24.16b, v8.4b[0]\n"
+ ".inst 0x4f88e31c // sdot v28.4s, v24.16b, v8.4b[0]\n"
"add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f8ce31d // sdot v29.4s, v24.16b, v12.4b[0]\n"
- ".word 0x4f90e31e // sdot v30.4s, v24.16b, v16.4b[0]\n"
- ".word 0x4f94e31f // sdot v31.4s, v24.16b, v20.4b[0]\n"
+ ".inst 0x4f8ce31d // sdot v29.4s, v24.16b, v12.4b[0]\n"
+ ".inst 0x4f90e31e // sdot v30.4s, v24.16b, v16.4b[0]\n"
+ ".inst 0x4f94e31f // sdot v31.4s, v24.16b, v20.4b[0]\n"
"ldr q24, [%[b_ptr0]]\n"
- ".word 0x4fa0e33a // sdot v26.4s, v25.16b, v0.4b[1]\n"
- ".word 0x4fa4e33b // sdot v27.4s, v25.16b, v4.4b[1]\n"
- ".word 0x4fa8e33c // sdot v28.4s, v25.16b, v8.4b[1]\n"
- ".word 0x4face33d // sdot v29.4s, v25.16b, v12.4b[1]\n"
- ".word 0x4fb0e33e // sdot v30.4s, v25.16b, v16.4b[1]\n"
- ".word 0x4fb4e33f // sdot v31.4s, v25.16b, v20.4b[1]\n"
+ ".inst 0x4fa0e33a // sdot v26.4s, v25.16b, v0.4b[1]\n"
+ ".inst 0x4fa4e33b // sdot v27.4s, v25.16b, v4.4b[1]\n"
+ ".inst 0x4fa8e33c // sdot v28.4s, v25.16b, v8.4b[1]\n"
+ ".inst 0x4face33d // sdot v29.4s, v25.16b, v12.4b[1]\n"
+ ".inst 0x4fb0e33e // sdot v30.4s, v25.16b, v16.4b[1]\n"
+ ".inst 0x4fb4e33f // sdot v31.4s, v25.16b, v20.4b[1]\n"
"ldr q25, [%[b_ptr0], #0x10]\n"
- ".word 0x4f80eb1a // sdot v26.4s, v24.16b, v0.4b[2]\n"
+ ".inst 0x4f80eb1a // sdot v26.4s, v24.16b, v0.4b[2]\n"
"add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f84eb1b // sdot v27.4s, v24.16b, v4.4b[2]\n"
- ".word 0x4f88eb1c // sdot v28.4s, v24.16b, v8.4b[2]\n"
- ".word 0x4f8ceb1d // sdot v29.4s, v24.16b, v12.4b[2]\n"
- ".word 0x4f90eb1e // sdot v30.4s, v24.16b, v16.4b[2]\n"
- ".word 0x4f94eb1f // sdot v31.4s, v24.16b, v20.4b[2]\n"
+ ".inst 0x4f84eb1b // sdot v27.4s, v24.16b, v4.4b[2]\n"
+ ".inst 0x4f88eb1c // sdot v28.4s, v24.16b, v8.4b[2]\n"
+ ".inst 0x4f8ceb1d // sdot v29.4s, v24.16b, v12.4b[2]\n"
+ ".inst 0x4f90eb1e // sdot v30.4s, v24.16b, v16.4b[2]\n"
+ ".inst 0x4f94eb1f // sdot v31.4s, v24.16b, v20.4b[2]\n"
"ldr q24, [%[b_ptr0]]\n"
- ".word 0x4fa0eb3a // sdot v26.4s, v25.16b, v0.4b[3]\n"
- ".word 0x4fa4eb3b // sdot v27.4s, v25.16b, v4.4b[3]\n"
- ".word 0x4fa8eb3c // sdot v28.4s, v25.16b, v8.4b[3]\n"
- ".word 0x4faceb3d // sdot v29.4s, v25.16b, v12.4b[3]\n"
- ".word 0x4fb0eb3e // sdot v30.4s, v25.16b, v16.4b[3]\n"
- ".word 0x4fb4eb3f // sdot v31.4s, v25.16b, v20.4b[3]\n"
+ ".inst 0x4fa0eb3a // sdot v26.4s, v25.16b, v0.4b[3]\n"
+ ".inst 0x4fa4eb3b // sdot v27.4s, v25.16b, v4.4b[3]\n"
+ ".inst 0x4fa8eb3c // sdot v28.4s, v25.16b, v8.4b[3]\n"
+ ".inst 0x4faceb3d // sdot v29.4s, v25.16b, v12.4b[3]\n"
+ ".inst 0x4fb0eb3e // sdot v30.4s, v25.16b, v16.4b[3]\n"
+ ".inst 0x4fb4eb3f // sdot v31.4s, v25.16b, v20.4b[3]\n"
"ldr q25, [%[b_ptr0], #0x10]\n"
- ".word 0x4f81e31a // sdot v26.4s, v24.16b, v1.4b[0]\n"
+ ".inst 0x4f81e31a // sdot v26.4s, v24.16b, v1.4b[0]\n"
"add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f85e31b // sdot v27.4s, v24.16b, v5.4b[0]\n"
- ".word 0x4f89e31c // sdot v28.4s, v24.16b, v9.4b[0]\n"
- ".word 0x4f8de31d // sdot v29.4s, v24.16b, v13.4b[0]\n"
- ".word 0x4f91e31e // sdot v30.4s, v24.16b, v17.4b[0]\n"
- ".word 0x4f95e31f // sdot v31.4s, v24.16b, v21.4b[0]\n"
+ ".inst 0x4f85e31b // sdot v27.4s, v24.16b, v5.4b[0]\n"
+ ".inst 0x4f89e31c // sdot v28.4s, v24.16b, v9.4b[0]\n"
+ ".inst 0x4f8de31d // sdot v29.4s, v24.16b, v13.4b[0]\n"
+ ".inst 0x4f91e31e // sdot v30.4s, v24.16b, v17.4b[0]\n"
+ ".inst 0x4f95e31f // sdot v31.4s, v24.16b, v21.4b[0]\n"
"ldr q24, [%[b_ptr0]]\n"
- ".word 0x4fa1e33a // sdot v26.4s, v25.16b, v1.4b[1]\n"
- ".word 0x4fa5e33b // sdot v27.4s, v25.16b, v5.4b[1]\n"
- ".word 0x4fa9e33c // sdot v28.4s, v25.16b, v9.4b[1]\n"
- ".word 0x4fade33d // sdot v29.4s, v25.16b, v13.4b[1]\n"
- ".word 0x4fb1e33e // sdot v30.4s, v25.16b, v17.4b[1]\n"
- ".word 0x4fb5e33f // sdot v31.4s, v25.16b, v21.4b[1]\n"
+ ".inst 0x4fa1e33a // sdot v26.4s, v25.16b, v1.4b[1]\n"
+ ".inst 0x4fa5e33b // sdot v27.4s, v25.16b, v5.4b[1]\n"
+ ".inst 0x4fa9e33c // sdot v28.4s, v25.16b, v9.4b[1]\n"
+ ".inst 0x4fade33d // sdot v29.4s, v25.16b, v13.4b[1]\n"
+ ".inst 0x4fb1e33e // sdot v30.4s, v25.16b, v17.4b[1]\n"
+ ".inst 0x4fb5e33f // sdot v31.4s, v25.16b, v21.4b[1]\n"
"ldr q25, [%[b_ptr0], #0x10]\n"
- ".word 0x4f81eb1a // sdot v26.4s, v24.16b, v1.4b[2]\n"
+ ".inst 0x4f81eb1a // sdot v26.4s, v24.16b, v1.4b[2]\n"
"add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f85eb1b // sdot v27.4s, v24.16b, v5.4b[2]\n"
- ".word 0x4f89eb1c // sdot v28.4s, v24.16b, v9.4b[2]\n"
- ".word 0x4f8deb1d // sdot v29.4s, v24.16b, v13.4b[2]\n"
- ".word 0x4f91eb1e // sdot v30.4s, v24.16b, v17.4b[2]\n"
- ".word 0x4f95eb1f // sdot v31.4s, v24.16b, v21.4b[2]\n"
+ ".inst 0x4f85eb1b // sdot v27.4s, v24.16b, v5.4b[2]\n"
+ ".inst 0x4f89eb1c // sdot v28.4s, v24.16b, v9.4b[2]\n"
+ ".inst 0x4f8deb1d // sdot v29.4s, v24.16b, v13.4b[2]\n"
+ ".inst 0x4f91eb1e // sdot v30.4s, v24.16b, v17.4b[2]\n"
+ ".inst 0x4f95eb1f // sdot v31.4s, v24.16b, v21.4b[2]\n"
"ldr q24, [%[b_ptr0]]\n"
- ".word 0x4fa1eb3a // sdot v26.4s, v25.16b, v1.4b[3]\n"
- ".word 0x4fa5eb3b // sdot v27.4s, v25.16b, v5.4b[3]\n"
- ".word 0x4fa9eb3c // sdot v28.4s, v25.16b, v9.4b[3]\n"
- ".word 0x4fadeb3d // sdot v29.4s, v25.16b, v13.4b[3]\n"
- ".word 0x4fb1eb3e // sdot v30.4s, v25.16b, v17.4b[3]\n"
- ".word 0x4fb5eb3f // sdot v31.4s, v25.16b, v21.4b[3]\n"
+ ".inst 0x4fa1eb3a // sdot v26.4s, v25.16b, v1.4b[3]\n"
+ ".inst 0x4fa5eb3b // sdot v27.4s, v25.16b, v5.4b[3]\n"
+ ".inst 0x4fa9eb3c // sdot v28.4s, v25.16b, v9.4b[3]\n"
+ ".inst 0x4fadeb3d // sdot v29.4s, v25.16b, v13.4b[3]\n"
+ ".inst 0x4fb1eb3e // sdot v30.4s, v25.16b, v17.4b[3]\n"
+ ".inst 0x4fb5eb3f // sdot v31.4s, v25.16b, v21.4b[3]\n"
"ldr q25, [%[b_ptr0], #0x10]\n"
- ".word 0x4f82e31a // sdot v26.4s, v24.16b, v2.4b[0]\n"
+ ".inst 0x4f82e31a // sdot v26.4s, v24.16b, v2.4b[0]\n"
"add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f86e31b // sdot v27.4s, v24.16b, v6.4b[0]\n"
- ".word 0x4f8ae31c // sdot v28.4s, v24.16b, v10.4b[0]\n"
- ".word 0x4f8ee31d // sdot v29.4s, v24.16b, v14.4b[0]\n"
- ".word 0x4f92e31e // sdot v30.4s, v24.16b, v18.4b[0]\n"
- ".word 0x4f96e31f // sdot v31.4s, v24.16b, v22.4b[0]\n"
+ ".inst 0x4f86e31b // sdot v27.4s, v24.16b, v6.4b[0]\n"
+ ".inst 0x4f8ae31c // sdot v28.4s, v24.16b, v10.4b[0]\n"
+ ".inst 0x4f8ee31d // sdot v29.4s, v24.16b, v14.4b[0]\n"
+ ".inst 0x4f92e31e // sdot v30.4s, v24.16b, v18.4b[0]\n"
+ ".inst 0x4f96e31f // sdot v31.4s, v24.16b, v22.4b[0]\n"
"ldr q24, [%[b_ptr0]]\n"
- ".word 0x4fa2e33a // sdot v26.4s, v25.16b, v2.4b[1]\n"
- ".word 0x4fa6e33b // sdot v27.4s, v25.16b, v6.4b[1]\n"
- ".word 0x4faae33c // sdot v28.4s, v25.16b, v10.4b[1]\n"
- ".word 0x4faee33d // sdot v29.4s, v25.16b, v14.4b[1]\n"
- ".word 0x4fb2e33e // sdot v30.4s, v25.16b, v18.4b[1]\n"
- ".word 0x4fb6e33f // sdot v31.4s, v25.16b, v22.4b[1]\n"
+ ".inst 0x4fa2e33a // sdot v26.4s, v25.16b, v2.4b[1]\n"
+ ".inst 0x4fa6e33b // sdot v27.4s, v25.16b, v6.4b[1]\n"
+ ".inst 0x4faae33c // sdot v28.4s, v25.16b, v10.4b[1]\n"
+ ".inst 0x4faee33d // sdot v29.4s, v25.16b, v14.4b[1]\n"
+ ".inst 0x4fb2e33e // sdot v30.4s, v25.16b, v18.4b[1]\n"
+ ".inst 0x4fb6e33f // sdot v31.4s, v25.16b, v22.4b[1]\n"
"ldr q25, [%[b_ptr0], #0x10]\n"
- ".word 0x4f82eb1a // sdot v26.4s, v24.16b, v2.4b[2]\n"
+ ".inst 0x4f82eb1a // sdot v26.4s, v24.16b, v2.4b[2]\n"
"add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f86eb1b // sdot v27.4s, v24.16b, v6.4b[2]\n"
- ".word 0x4f8aeb1c // sdot v28.4s, v24.16b, v10.4b[2]\n"
- ".word 0x4f8eeb1d // sdot v29.4s, v24.16b, v14.4b[2]\n"
- ".word 0x4f92eb1e // sdot v30.4s, v24.16b, v18.4b[2]\n"
- ".word 0x4f96eb1f // sdot v31.4s, v24.16b, v22.4b[2]\n"
+ ".inst 0x4f86eb1b // sdot v27.4s, v24.16b, v6.4b[2]\n"
+ ".inst 0x4f8aeb1c // sdot v28.4s, v24.16b, v10.4b[2]\n"
+ ".inst 0x4f8eeb1d // sdot v29.4s, v24.16b, v14.4b[2]\n"
+ ".inst 0x4f92eb1e // sdot v30.4s, v24.16b, v18.4b[2]\n"
+ ".inst 0x4f96eb1f // sdot v31.4s, v24.16b, v22.4b[2]\n"
"ldr q24, [%[b_ptr0]]\n"
- ".word 0x4fa2eb3a // sdot v26.4s, v25.16b, v2.4b[3]\n"
- ".word 0x4fa6eb3b // sdot v27.4s, v25.16b, v6.4b[3]\n"
- ".word 0x4faaeb3c // sdot v28.4s, v25.16b, v10.4b[3]\n"
- ".word 0x4faeeb3d // sdot v29.4s, v25.16b, v14.4b[3]\n"
- ".word 0x4fb2eb3e // sdot v30.4s, v25.16b, v18.4b[3]\n"
- ".word 0x4fb6eb3f // sdot v31.4s, v25.16b, v22.4b[3]\n"
+ ".inst 0x4fa2eb3a // sdot v26.4s, v25.16b, v2.4b[3]\n"
+ ".inst 0x4fa6eb3b // sdot v27.4s, v25.16b, v6.4b[3]\n"
+ ".inst 0x4faaeb3c // sdot v28.4s, v25.16b, v10.4b[3]\n"
+ ".inst 0x4faeeb3d // sdot v29.4s, v25.16b, v14.4b[3]\n"
+ ".inst 0x4fb2eb3e // sdot v30.4s, v25.16b, v18.4b[3]\n"
+ ".inst 0x4fb6eb3f // sdot v31.4s, v25.16b, v22.4b[3]\n"
"ldr q25, [%[b_ptr0], #0x10]\n"
- ".word 0x4f83e31a // sdot v26.4s, v24.16b, v3.4b[0]\n"
- "add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f87e31b // sdot v27.4s, v24.16b, v7.4b[0]\n"
- ".word 0x4f8be31c // sdot v28.4s, v24.16b, v11.4b[0]\n"
- ".word 0x4f8fe31d // sdot v29.4s, v24.16b, v15.4b[0]\n"
- ".word 0x4f93e31e // sdot v30.4s, v24.16b, v19.4b[0]\n"
- ".word 0x4f97e31f // sdot v31.4s, v24.16b, v23.4b[0]\n"
- ".word 0x4fa3e33a // sdot v26.4s, v25.16b, v3.4b[1]\n"
- ".word 0x4fa7e33b // sdot v27.4s, v25.16b, v7.4b[1]\n"
- ".word 0x4fabe33c // sdot v28.4s, v25.16b, v11.4b[1]\n"
- ".word 0x4fafe33d // sdot v29.4s, v25.16b, v15.4b[1]\n"
- ".word 0x4fb3e33e // sdot v30.4s, v25.16b, v19.4b[1]\n"
- ".word 0x4fb7e33f // sdot v31.4s, v25.16b, v23.4b[1]\n"
+ ".inst 0x4f83e31a // sdot v26.4s, v24.16b, v3.4b[0]\n"
+ "add %[b_ptr0], %[b_ptr0], #0x20\n"
+ ".inst 0x4f87e31b // sdot v27.4s, v24.16b, v7.4b[0]\n"
+ ".inst 0x4f8be31c // sdot v28.4s, v24.16b, v11.4b[0]\n"
+ ".inst 0x4f8fe31d // sdot v29.4s, v24.16b, v15.4b[0]\n"
+ ".inst 0x4f93e31e // sdot v30.4s, v24.16b, v19.4b[0]\n"
+ ".inst 0x4f97e31f // sdot v31.4s, v24.16b, v23.4b[0]\n"
+ ".inst 0x4fa3e33a // sdot v26.4s, v25.16b, v3.4b[1]\n"
+ ".inst 0x4fa7e33b // sdot v27.4s, v25.16b, v7.4b[1]\n"
+ ".inst 0x4fabe33c // sdot v28.4s, v25.16b, v11.4b[1]\n"
+ ".inst 0x4fafe33d // sdot v29.4s, v25.16b, v15.4b[1]\n"
+ ".inst 0x4fb3e33e // sdot v30.4s, v25.16b, v19.4b[1]\n"
+ ".inst 0x4fb7e33f // sdot v31.4s, v25.16b, v23.4b[1]\n"
"cbz %[loops], 6f\n"
"ldr d24, [%[b_ptr0]]\n"
"subs %[loops], %[loops], #0x1\n"
@@ -2535,282 +2632,300 @@ void a64_smallK_hybrid_s8s32_dot_4x6_a55(const int8_t *A, int lda, const int8_t
"ins v24.d[1], temploadreg0\n"
"b.eq 7f\n"
"8:\n"
- "str q26, [%[c_ptr0]], #0x10\n"
+ "str q26, [%[c_ptr0]]\n"
"subs %[loops], %[loops], #0x1\n"
"movi v26.4s, #0\n"
"ins v25.d[1], temploadreg1\n"
"ldr temploadreg0, [%[b_ptr0], #0x8]\n"
- "prfm PSTL1KEEP, [%[c_ptr0], #0x40]\n"
- "str q27, [c_ptr1], #0x10\n"
+ "add %[c_ptr0], %[c_ptr0], #0x10\n"
+ "str q27, [c_ptr1]\n"
+ "add c_ptr1, c_ptr1, #0x10\n"
"movi v27.4s, #0\n"
"ldr temploadreg1, [%[b_ptr0], #0x18]\n"
- ".word 0x4f80e31a // sdot v26.4s, v24.16b, v0.4b[0]\n"
- "str q28, [c_ptr2], #0x10\n"
- "prfm PSTL1KEEP, [c_ptr1, #0x40]\n"
+ ".inst 0x4f80e31a // sdot v26.4s, v24.16b, v0.4b[0]\n"
+ "str q28, [c_ptr2]\n"
"movi v28.4s, #0\n"
- ".word 0x4f84e31b // sdot v27.4s, v24.16b, v4.4b[0]\n"
- "str q29, [c_ptr3], #0x10\n"
+ "prfm PSTL1KEEP, [%[c_ptr0], #0x40]\n"
+ ".inst 0x4f84e31b // sdot v27.4s, v24.16b, v4.4b[0]\n"
+ "str q29, [c_ptr3]\n"
"movi v29.4s, #0\n"
- "prfm PSTL1KEEP, [c_ptr2, #0x40]\n"
- ".word 0x4f88e31c // sdot v28.4s, v24.16b, v8.4b[0]\n"
- "str q30, [c_ptr4], #0x10\n"
+ "add c_ptr2, c_ptr2, #0x10\n"
+ ".inst 0x4f88e31c // sdot v28.4s, v24.16b, v8.4b[0]\n"
+ "str q30, [c_ptr4]\n"
"movi v30.4s, #0\n"
- "prfm PSTL1KEEP, [c_ptr3, #0x40]\n"
- ".word 0x4f8ce31d // sdot v29.4s, v24.16b, v12.4b[0]\n"
- "str q31, [c_ptr5], #0x10\n"
+ "add c_ptr3, c_ptr3, #0x10\n"
+ ".inst 0x4f8ce31d // sdot v29.4s, v24.16b, v12.4b[0]\n"
+ "str q31, [c_ptr5]\n"
"movi v31.4s, #0\n"
- "prfm PSTL1KEEP, [c_ptr4, #0x40]\n"
- ".word 0x4f90e31e // sdot v30.4s, v24.16b, v16.4b[0]\n"
- "prfm PSTL1KEEP, [c_ptr5, #0x40]\n"
- ".word 0x4f94e31f // sdot v31.4s, v24.16b, v20.4b[0]\n"
+ "add c_ptr4, c_ptr4, #0x10\n"
+ ".inst 0x4f90e31e // sdot v30.4s, v24.16b, v16.4b[0]\n"
+ "add c_ptr5, c_ptr5, #0x10\n"
+ ".inst 0x4f94e31f // sdot v31.4s, v24.16b, v20.4b[0]\n"
"ldr d24, [%[b_ptr0]]\n"
- ".word 0x4fa0e33a // sdot v26.4s, v25.16b, v0.4b[1]\n"
- ".word 0x4fa4e33b // sdot v27.4s, v25.16b, v4.4b[1]\n"
- ".word 0x4fa8e33c // sdot v28.4s, v25.16b, v8.4b[1]\n"
+ ".inst 0x4fa0e33a // sdot v26.4s, v25.16b, v0.4b[1]\n"
+ "prfm PSTL1KEEP, [c_ptr1, #0x40]\n"
+ ".inst 0x4fa4e33b // sdot v27.4s, v25.16b, v4.4b[1]\n"
"ins v24.d[1], temploadreg0\n"
- ".word 0x4face33d // sdot v29.4s, v25.16b, v12.4b[1]\n"
- ".word 0x4fb0e33e // sdot v30.4s, v25.16b, v16.4b[1]\n"
- ".word 0x4fb4e33f // sdot v31.4s, v25.16b, v20.4b[1]\n"
+ ".inst 0x4fa8e33c // sdot v28.4s, v25.16b, v8.4b[1]\n"
+ "prfm PSTL1KEEP, [c_ptr2, #0x40]\n"
+ ".inst 0x4face33d // sdot v29.4s, v25.16b, v12.4b[1]\n"
+ "prfm PSTL1KEEP, [c_ptr3, #0x40]\n"
+ ".inst 0x4fb0e33e // sdot v30.4s, v25.16b, v16.4b[1]\n"
+ "prfm PSTL1KEEP, [c_ptr4, #0x40]\n"
+ ".inst 0x4fb4e33f // sdot v31.4s, v25.16b, v20.4b[1]\n"
"ldr d25, [%[b_ptr0], #0x10]\n"
- ".word 0x4f80eb1a // sdot v26.4s, v24.16b, v0.4b[2]\n"
- "add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f84eb1b // sdot v27.4s, v24.16b, v4.4b[2]\n"
+ ".inst 0x4f80eb1a // sdot v26.4s, v24.16b, v0.4b[2]\n"
+ "prfm PSTL1KEEP, [c_ptr5, #0x40]\n"
+ ".inst 0x4f84eb1b // sdot v27.4s, v24.16b, v4.4b[2]\n"
"ins v25.d[1], temploadreg1\n"
- ".word 0x4f88eb1c // sdot v28.4s, v24.16b, v8.4b[2]\n"
+ ".inst 0x4f88eb1c // sdot v28.4s, v24.16b, v8.4b[2]\n"
+ "add %[b_ptr0], %[b_ptr0], #0x20\n"
+ ".inst 0x4f8ceb1d // sdot v29.4s, v24.16b, v12.4b[2]\n"
"ldr temploadreg0, [%[b_ptr0], #0x8]\n"
- ".word 0x4f8ceb1d // sdot v29.4s, v24.16b, v12.4b[2]\n"
+ ".inst 0x4f90eb1e // sdot v30.4s, v24.16b, v16.4b[2]\n"
"ldr temploadreg1, [%[b_ptr0], #0x18]\n"
- ".word 0x4f90eb1e // sdot v30.4s, v24.16b, v16.4b[2]\n"
- ".word 0x4f94eb1f // sdot v31.4s, v24.16b, v20.4b[2]\n"
+ ".inst 0x4f94eb1f // sdot v31.4s, v24.16b, v20.4b[2]\n"
"ldr d24, [%[b_ptr0]]\n"
- ".word 0x4fa0eb3a // sdot v26.4s, v25.16b, v0.4b[3]\n"
- ".word 0x4fa4eb3b // sdot v27.4s, v25.16b, v4.4b[3]\n"
- ".word 0x4fa8eb3c // sdot v28.4s, v25.16b, v8.4b[3]\n"
+ ".inst 0x4fa0eb3a // sdot v26.4s, v25.16b, v0.4b[3]\n"
+ ".inst 0x4fa4eb3b // sdot v27.4s, v25.16b, v4.4b[3]\n"
+ ".inst 0x4fa8eb3c // sdot v28.4s, v25.16b, v8.4b[3]\n"
"ins v24.d[1], temploadreg0\n"
- ".word 0x4faceb3d // sdot v29.4s, v25.16b, v12.4b[3]\n"
- ".word 0x4fb0eb3e // sdot v30.4s, v25.16b, v16.4b[3]\n"
- ".word 0x4fb4eb3f // sdot v31.4s, v25.16b, v20.4b[3]\n"
+ ".inst 0x4faceb3d // sdot v29.4s, v25.16b, v12.4b[3]\n"
+ ".inst 0x4fb0eb3e // sdot v30.4s, v25.16b, v16.4b[3]\n"
+ ".inst 0x4fb4eb3f // sdot v31.4s, v25.16b, v20.4b[3]\n"
"ldr d25, [%[b_ptr0], #0x10]\n"
- ".word 0x4f81e31a // sdot v26.4s, v24.16b, v1.4b[0]\n"
+ ".inst 0x4f81e31a // sdot v26.4s, v24.16b, v1.4b[0]\n"
"add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f85e31b // sdot v27.4s, v24.16b, v5.4b[0]\n"
+ ".inst 0x4f85e31b // sdot v27.4s, v24.16b, v5.4b[0]\n"
"ins v25.d[1], temploadreg1\n"
- ".word 0x4f89e31c // sdot v28.4s, v24.16b, v9.4b[0]\n"
+ ".inst 0x4f89e31c // sdot v28.4s, v24.16b, v9.4b[0]\n"
"ldr temploadreg0, [%[b_ptr0], #0x8]\n"
- ".word 0x4f8de31d // sdot v29.4s, v24.16b, v13.4b[0]\n"
+ ".inst 0x4f8de31d // sdot v29.4s, v24.16b, v13.4b[0]\n"
"ldr temploadreg1, [%[b_ptr0], #0x18]\n"
- ".word 0x4f91e31e // sdot v30.4s, v24.16b, v17.4b[0]\n"
- ".word 0x4f95e31f // sdot v31.4s, v24.16b, v21.4b[0]\n"
+ ".inst 0x4f91e31e // sdot v30.4s, v24.16b, v17.4b[0]\n"
+ ".inst 0x4f95e31f // sdot v31.4s, v24.16b, v21.4b[0]\n"
"ldr d24, [%[b_ptr0]]\n"
- ".word 0x4fa1e33a // sdot v26.4s, v25.16b, v1.4b[1]\n"
- ".word 0x4fa5e33b // sdot v27.4s, v25.16b, v5.4b[1]\n"
- ".word 0x4fa9e33c // sdot v28.4s, v25.16b, v9.4b[1]\n"
+ ".inst 0x4fa1e33a // sdot v26.4s, v25.16b, v1.4b[1]\n"
+ ".inst 0x4fa5e33b // sdot v27.4s, v25.16b, v5.4b[1]\n"
+ ".inst 0x4fa9e33c // sdot v28.4s, v25.16b, v9.4b[1]\n"
"ins v24.d[1], temploadreg0\n"
- ".word 0x4fade33d // sdot v29.4s, v25.16b, v13.4b[1]\n"
- ".word 0x4fb1e33e // sdot v30.4s, v25.16b, v17.4b[1]\n"
- ".word 0x4fb5e33f // sdot v31.4s, v25.16b, v21.4b[1]\n"
+ ".inst 0x4fade33d // sdot v29.4s, v25.16b, v13.4b[1]\n"
+ ".inst 0x4fb1e33e // sdot v30.4s, v25.16b, v17.4b[1]\n"
+ ".inst 0x4fb5e33f // sdot v31.4s, v25.16b, v21.4b[1]\n"
"ldr d25, [%[b_ptr0], #0x10]\n"
- ".word 0x4f81eb1a // sdot v26.4s, v24.16b, v1.4b[2]\n"
+ ".inst 0x4f81eb1a // sdot v26.4s, v24.16b, v1.4b[2]\n"
"add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f85eb1b // sdot v27.4s, v24.16b, v5.4b[2]\n"
+ ".inst 0x4f85eb1b // sdot v27.4s, v24.16b, v5.4b[2]\n"
"ins v25.d[1], temploadreg1\n"
- ".word 0x4f89eb1c // sdot v28.4s, v24.16b, v9.4b[2]\n"
+ ".inst 0x4f89eb1c // sdot v28.4s, v24.16b, v9.4b[2]\n"
"ldr temploadreg0, [%[b_ptr0], #0x8]\n"
- ".word 0x4f8deb1d // sdot v29.4s, v24.16b, v13.4b[2]\n"
+ ".inst 0x4f8deb1d // sdot v29.4s, v24.16b, v13.4b[2]\n"
"ldr temploadreg1, [%[b_ptr0], #0x18]\n"
- ".word 0x4f91eb1e // sdot v30.4s, v24.16b, v17.4b[2]\n"
- ".word 0x4f95eb1f // sdot v31.4s, v24.16b, v21.4b[2]\n"
+ ".inst 0x4f91eb1e // sdot v30.4s, v24.16b, v17.4b[2]\n"
+ ".inst 0x4f95eb1f // sdot v31.4s, v24.16b, v21.4b[2]\n"
"ldr d24, [%[b_ptr0]]\n"
- ".word 0x4fa1eb3a // sdot v26.4s, v25.16b, v1.4b[3]\n"
- ".word 0x4fa5eb3b // sdot v27.4s, v25.16b, v5.4b[3]\n"
- ".word 0x4fa9eb3c // sdot v28.4s, v25.16b, v9.4b[3]\n"
+ ".inst 0x4fa1eb3a // sdot v26.4s, v25.16b, v1.4b[3]\n"
+ ".inst 0x4fa5eb3b // sdot v27.4s, v25.16b, v5.4b[3]\n"
+ ".inst 0x4fa9eb3c // sdot v28.4s, v25.16b, v9.4b[3]\n"
"ins v24.d[1], temploadreg0\n"
- ".word 0x4fadeb3d // sdot v29.4s, v25.16b, v13.4b[3]\n"
- ".word 0x4fb1eb3e // sdot v30.4s, v25.16b, v17.4b[3]\n"
- ".word 0x4fb5eb3f // sdot v31.4s, v25.16b, v21.4b[3]\n"
+ ".inst 0x4fadeb3d // sdot v29.4s, v25.16b, v13.4b[3]\n"
+ ".inst 0x4fb1eb3e // sdot v30.4s, v25.16b, v17.4b[3]\n"
+ ".inst 0x4fb5eb3f // sdot v31.4s, v25.16b, v21.4b[3]\n"
"ldr d25, [%[b_ptr0], #0x10]\n"
- ".word 0x4f82e31a // sdot v26.4s, v24.16b, v2.4b[0]\n"
+ ".inst 0x4f82e31a // sdot v26.4s, v24.16b, v2.4b[0]\n"
"add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f86e31b // sdot v27.4s, v24.16b, v6.4b[0]\n"
+ ".inst 0x4f86e31b // sdot v27.4s, v24.16b, v6.4b[0]\n"
"ins v25.d[1], temploadreg1\n"
- ".word 0x4f8ae31c // sdot v28.4s, v24.16b, v10.4b[0]\n"
+ ".inst 0x4f8ae31c // sdot v28.4s, v24.16b, v10.4b[0]\n"
"ldr temploadreg0, [%[b_ptr0], #0x8]\n"
- ".word 0x4f8ee31d // sdot v29.4s, v24.16b, v14.4b[0]\n"
+ ".inst 0x4f8ee31d // sdot v29.4s, v24.16b, v14.4b[0]\n"
"ldr temploadreg1, [%[b_ptr0], #0x18]\n"
- ".word 0x4f92e31e // sdot v30.4s, v24.16b, v18.4b[0]\n"
- ".word 0x4f96e31f // sdot v31.4s, v24.16b, v22.4b[0]\n"
+ ".inst 0x4f92e31e // sdot v30.4s, v24.16b, v18.4b[0]\n"
+ ".inst 0x4f96e31f // sdot v31.4s, v24.16b, v22.4b[0]\n"
"ldr d24, [%[b_ptr0]]\n"
- ".word 0x4fa2e33a // sdot v26.4s, v25.16b, v2.4b[1]\n"
- ".word 0x4fa6e33b // sdot v27.4s, v25.16b, v6.4b[1]\n"
- ".word 0x4faae33c // sdot v28.4s, v25.16b, v10.4b[1]\n"
+ ".inst 0x4fa2e33a // sdot v26.4s, v25.16b, v2.4b[1]\n"
+ ".inst 0x4fa6e33b // sdot v27.4s, v25.16b, v6.4b[1]\n"
+ ".inst 0x4faae33c // sdot v28.4s, v25.16b, v10.4b[1]\n"
"ins v24.d[1], temploadreg0\n"
- ".word 0x4faee33d // sdot v29.4s, v25.16b, v14.4b[1]\n"
- ".word 0x4fb2e33e // sdot v30.4s, v25.16b, v18.4b[1]\n"
- ".word 0x4fb6e33f // sdot v31.4s, v25.16b, v22.4b[1]\n"
+ ".inst 0x4faee33d // sdot v29.4s, v25.16b, v14.4b[1]\n"
+ ".inst 0x4fb2e33e // sdot v30.4s, v25.16b, v18.4b[1]\n"
+ ".inst 0x4fb6e33f // sdot v31.4s, v25.16b, v22.4b[1]\n"
"ldr d25, [%[b_ptr0], #0x10]\n"
- ".word 0x4f82eb1a // sdot v26.4s, v24.16b, v2.4b[2]\n"
+ ".inst 0x4f82eb1a // sdot v26.4s, v24.16b, v2.4b[2]\n"
"add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f86eb1b // sdot v27.4s, v24.16b, v6.4b[2]\n"
+ ".inst 0x4f86eb1b // sdot v27.4s, v24.16b, v6.4b[2]\n"
"ins v25.d[1], temploadreg1\n"
- ".word 0x4f8aeb1c // sdot v28.4s, v24.16b, v10.4b[2]\n"
+ ".inst 0x4f8aeb1c // sdot v28.4s, v24.16b, v10.4b[2]\n"
"ldr temploadreg0, [%[b_ptr0], #0x8]\n"
- ".word 0x4f8eeb1d // sdot v29.4s, v24.16b, v14.4b[2]\n"
+ ".inst 0x4f8eeb1d // sdot v29.4s, v24.16b, v14.4b[2]\n"
"ldr temploadreg1, [%[b_ptr0], #0x18]\n"
- ".word 0x4f92eb1e // sdot v30.4s, v24.16b, v18.4b[2]\n"
- ".word 0x4f96eb1f // sdot v31.4s, v24.16b, v22.4b[2]\n"
+ ".inst 0x4f92eb1e // sdot v30.4s, v24.16b, v18.4b[2]\n"
+ ".inst 0x4f96eb1f // sdot v31.4s, v24.16b, v22.4b[2]\n"
"ldr d24, [%[b_ptr0]]\n"
- ".word 0x4fa2eb3a // sdot v26.4s, v25.16b, v2.4b[3]\n"
- ".word 0x4fa6eb3b // sdot v27.4s, v25.16b, v6.4b[3]\n"
- ".word 0x4faaeb3c // sdot v28.4s, v25.16b, v10.4b[3]\n"
+ ".inst 0x4fa2eb3a // sdot v26.4s, v25.16b, v2.4b[3]\n"
+ ".inst 0x4fa6eb3b // sdot v27.4s, v25.16b, v6.4b[3]\n"
+ ".inst 0x4faaeb3c // sdot v28.4s, v25.16b, v10.4b[3]\n"
"ins v24.d[1], temploadreg0\n"
- ".word 0x4faeeb3d // sdot v29.4s, v25.16b, v14.4b[3]\n"
- ".word 0x4fb2eb3e // sdot v30.4s, v25.16b, v18.4b[3]\n"
- ".word 0x4fb6eb3f // sdot v31.4s, v25.16b, v22.4b[3]\n"
+ ".inst 0x4faeeb3d // sdot v29.4s, v25.16b, v14.4b[3]\n"
+ ".inst 0x4fb2eb3e // sdot v30.4s, v25.16b, v18.4b[3]\n"
+ ".inst 0x4fb6eb3f // sdot v31.4s, v25.16b, v22.4b[3]\n"
"ldr d25, [%[b_ptr0], #0x10]\n"
- ".word 0x4f83e31a // sdot v26.4s, v24.16b, v3.4b[0]\n"
+ ".inst 0x4f83e31a // sdot v26.4s, v24.16b, v3.4b[0]\n"
"add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f87e31b // sdot v27.4s, v24.16b, v7.4b[0]\n"
+ ".inst 0x4f87e31b // sdot v27.4s, v24.16b, v7.4b[0]\n"
"ins v25.d[1], temploadreg1\n"
- ".word 0x4f8be31c // sdot v28.4s, v24.16b, v11.4b[0]\n"
+ ".inst 0x4f8be31c // sdot v28.4s, v24.16b, v11.4b[0]\n"
"ldr temploadreg0, [%[b_ptr0], #0x8]\n"
- ".word 0x4f8fe31d // sdot v29.4s, v24.16b, v15.4b[0]\n"
+ ".inst 0x4f8fe31d // sdot v29.4s, v24.16b, v15.4b[0]\n"
"ldr temploadreg1, [%[b_ptr0], #0x18]\n"
- ".word 0x4f93e31e // sdot v30.4s, v24.16b, v19.4b[0]\n"
- ".word 0x4f97e31f // sdot v31.4s, v24.16b, v23.4b[0]\n"
+ ".inst 0x4f93e31e // sdot v30.4s, v24.16b, v19.4b[0]\n"
+ ".inst 0x4f97e31f // sdot v31.4s, v24.16b, v23.4b[0]\n"
"ldr d24, [%[b_ptr0]]\n"
- ".word 0x4fa3e33a // sdot v26.4s, v25.16b, v3.4b[1]\n"
- ".word 0x4fa7e33b // sdot v27.4s, v25.16b, v7.4b[1]\n"
- ".word 0x4fabe33c // sdot v28.4s, v25.16b, v11.4b[1]\n"
+ ".inst 0x4fa3e33a // sdot v26.4s, v25.16b, v3.4b[1]\n"
+ ".inst 0x4fa7e33b // sdot v27.4s, v25.16b, v7.4b[1]\n"
+ ".inst 0x4fabe33c // sdot v28.4s, v25.16b, v11.4b[1]\n"
"ins v24.d[1], temploadreg0\n"
- ".word 0x4fafe33d // sdot v29.4s, v25.16b, v15.4b[1]\n"
- ".word 0x4fb3e33e // sdot v30.4s, v25.16b, v19.4b[1]\n"
- ".word 0x4fb7e33f // sdot v31.4s, v25.16b, v23.4b[1]\n"
+ ".inst 0x4fafe33d // sdot v29.4s, v25.16b, v15.4b[1]\n"
+ ".inst 0x4fb3e33e // sdot v30.4s, v25.16b, v19.4b[1]\n"
+ ".inst 0x4fb7e33f // sdot v31.4s, v25.16b, v23.4b[1]\n"
"ldr d25, [%[b_ptr0], #0x10]\n"
"add %[b_ptr0], %[b_ptr0], #0x20\n"
"b.ne 8b\n"
"7:\n"
- "str q26, [%[c_ptr0]], #0x10\n"
+ "str q26, [%[c_ptr0]]\n"
+ "add %[c_ptr0], %[c_ptr0], #0x10\n"
"movi v26.4s, #0\n"
"ins v25.d[1], temploadreg1\n"
- "str q27, [c_ptr1], #0x10\n"
+ "str q27, [c_ptr1]\n"
+ "add c_ptr1, c_ptr1, #0x10\n"
"movi v27.4s, #0\n"
- ".word 0x4f80e31a // sdot v26.4s, v24.16b, v0.4b[0]\n"
- "str q28, [c_ptr2], #0x10\n"
+ ".inst 0x4f80e31a // sdot v26.4s, v24.16b, v0.4b[0]\n"
+ "str q28, [c_ptr2]\n"
"movi v28.4s, #0\n"
- ".word 0x4f84e31b // sdot v27.4s, v24.16b, v4.4b[0]\n"
- ".word 0x4fa0e33a // sdot v26.4s, v25.16b, v0.4b[1]\n"
- "str q29, [c_ptr3], #0x10\n"
+ "add c_ptr2, c_ptr2, #0x10\n"
+ ".inst 0x4f84e31b // sdot v27.4s, v24.16b, v4.4b[0]\n"
+ "str q29, [c_ptr3]\n"
"movi v29.4s, #0\n"
- ".word 0x4f88e31c // sdot v28.4s, v24.16b, v8.4b[0]\n"
- ".word 0x4fa4e33b // sdot v27.4s, v25.16b, v4.4b[1]\n"
- "str q30, [c_ptr4], #0x10\n"
+ "add c_ptr3, c_ptr3, #0x10\n"
+ ".inst 0x4f88e31c // sdot v28.4s, v24.16b, v8.4b[0]\n"
+ "str q30, [c_ptr4]\n"
"movi v30.4s, #0\n"
- ".word 0x4f8ce31d // sdot v29.4s, v24.16b, v12.4b[0]\n"
- ".word 0x4fa8e33c // sdot v28.4s, v25.16b, v8.4b[1]\n"
- "str q31, [c_ptr5], #0x10\n"
+ "add c_ptr4, c_ptr4, #0x10\n"
+ ".inst 0x4f8ce31d // sdot v29.4s, v24.16b, v12.4b[0]\n"
+ "str q31, [c_ptr5]\n"
"movi v31.4s, #0\n"
- ".word 0x4f90e31e // sdot v30.4s, v24.16b, v16.4b[0]\n"
- ".word 0x4face33d // sdot v29.4s, v25.16b, v12.4b[1]\n"
- ".word 0x4f94e31f // sdot v31.4s, v24.16b, v20.4b[0]\n"
+ "add c_ptr5, c_ptr5, #0x10\n"
+ ".inst 0x4f90e31e // sdot v30.4s, v24.16b, v16.4b[0]\n"
+ ".inst 0x4fa0e33a // sdot v26.4s, v25.16b, v0.4b[1]\n"
+ ".inst 0x4f94e31f // sdot v31.4s, v24.16b, v20.4b[0]\n"
"ldr q24, [%[b_ptr0]]\n"
- ".word 0x4fb0e33e // sdot v30.4s, v25.16b, v16.4b[1]\n"
- ".word 0x4fb4e33f // sdot v31.4s, v25.16b, v20.4b[1]\n"
+ ".inst 0x4fa4e33b // sdot v27.4s, v25.16b, v4.4b[1]\n"
+ ".inst 0x4fa8e33c // sdot v28.4s, v25.16b, v8.4b[1]\n"
+ ".inst 0x4face33d // sdot v29.4s, v25.16b, v12.4b[1]\n"
+ ".inst 0x4fb0e33e // sdot v30.4s, v25.16b, v16.4b[1]\n"
+ ".inst 0x4fb4e33f // sdot v31.4s, v25.16b, v20.4b[1]\n"
"ldr q25, [%[b_ptr0], #0x10]\n"
- ".word 0x4f80eb1a // sdot v26.4s, v24.16b, v0.4b[2]\n"
+ ".inst 0x4f80eb1a // sdot v26.4s, v24.16b, v0.4b[2]\n"
"add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f84eb1b // sdot v27.4s, v24.16b, v4.4b[2]\n"
- ".word 0x4f88eb1c // sdot v28.4s, v24.16b, v8.4b[2]\n"
- ".word 0x4f8ceb1d // sdot v29.4s, v24.16b, v12.4b[2]\n"
- ".word 0x4f90eb1e // sdot v30.4s, v24.16b, v16.4b[2]\n"
- ".word 0x4f94eb1f // sdot v31.4s, v24.16b, v20.4b[2]\n"
+ ".inst 0x4f84eb1b // sdot v27.4s, v24.16b, v4.4b[2]\n"
+ ".inst 0x4f88eb1c // sdot v28.4s, v24.16b, v8.4b[2]\n"
+ ".inst 0x4f8ceb1d // sdot v29.4s, v24.16b, v12.4b[2]\n"
+ ".inst 0x4f90eb1e // sdot v30.4s, v24.16b, v16.4b[2]\n"
+ ".inst 0x4f94eb1f // sdot v31.4s, v24.16b, v20.4b[2]\n"
"ldr q24, [%[b_ptr0]]\n"
- ".word 0x4fa0eb3a // sdot v26.4s, v25.16b, v0.4b[3]\n"
- ".word 0x4fa4eb3b // sdot v27.4s, v25.16b, v4.4b[3]\n"
- ".word 0x4fa8eb3c // sdot v28.4s, v25.16b, v8.4b[3]\n"
- ".word 0x4faceb3d // sdot v29.4s, v25.16b, v12.4b[3]\n"
- ".word 0x4fb0eb3e // sdot v30.4s, v25.16b, v16.4b[3]\n"
- ".word 0x4fb4eb3f // sdot v31.4s, v25.16b, v20.4b[3]\n"
+ ".inst 0x4fa0eb3a // sdot v26.4s, v25.16b, v0.4b[3]\n"
+ ".inst 0x4fa4eb3b // sdot v27.4s, v25.16b, v4.4b[3]\n"
+ ".inst 0x4fa8eb3c // sdot v28.4s, v25.16b, v8.4b[3]\n"
+ ".inst 0x4faceb3d // sdot v29.4s, v25.16b, v12.4b[3]\n"
+ ".inst 0x4fb0eb3e // sdot v30.4s, v25.16b, v16.4b[3]\n"
+ ".inst 0x4fb4eb3f // sdot v31.4s, v25.16b, v20.4b[3]\n"
"ldr q25, [%[b_ptr0], #0x10]\n"
- ".word 0x4f81e31a // sdot v26.4s, v24.16b, v1.4b[0]\n"
+ ".inst 0x4f81e31a // sdot v26.4s, v24.16b, v1.4b[0]\n"
"add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f85e31b // sdot v27.4s, v24.16b, v5.4b[0]\n"
- ".word 0x4f89e31c // sdot v28.4s, v24.16b, v9.4b[0]\n"
- ".word 0x4f8de31d // sdot v29.4s, v24.16b, v13.4b[0]\n"
- ".word 0x4f91e31e // sdot v30.4s, v24.16b, v17.4b[0]\n"
- ".word 0x4f95e31f // sdot v31.4s, v24.16b, v21.4b[0]\n"
+ ".inst 0x4f85e31b // sdot v27.4s, v24.16b, v5.4b[0]\n"
+ ".inst 0x4f89e31c // sdot v28.4s, v24.16b, v9.4b[0]\n"
+ ".inst 0x4f8de31d // sdot v29.4s, v24.16b, v13.4b[0]\n"
+ ".inst 0x4f91e31e // sdot v30.4s, v24.16b, v17.4b[0]\n"
+ ".inst 0x4f95e31f // sdot v31.4s, v24.16b, v21.4b[0]\n"
"ldr q24, [%[b_ptr0]]\n"
- ".word 0x4fa1e33a // sdot v26.4s, v25.16b, v1.4b[1]\n"
- ".word 0x4fa5e33b // sdot v27.4s, v25.16b, v5.4b[1]\n"
- ".word 0x4fa9e33c // sdot v28.4s, v25.16b, v9.4b[1]\n"
- ".word 0x4fade33d // sdot v29.4s, v25.16b, v13.4b[1]\n"
- ".word 0x4fb1e33e // sdot v30.4s, v25.16b, v17.4b[1]\n"
- ".word 0x4fb5e33f // sdot v31.4s, v25.16b, v21.4b[1]\n"
+ ".inst 0x4fa1e33a // sdot v26.4s, v25.16b, v1.4b[1]\n"
+ ".inst 0x4fa5e33b // sdot v27.4s, v25.16b, v5.4b[1]\n"
+ ".inst 0x4fa9e33c // sdot v28.4s, v25.16b, v9.4b[1]\n"
+ ".inst 0x4fade33d // sdot v29.4s, v25.16b, v13.4b[1]\n"
+ ".inst 0x4fb1e33e // sdot v30.4s, v25.16b, v17.4b[1]\n"
+ ".inst 0x4fb5e33f // sdot v31.4s, v25.16b, v21.4b[1]\n"
"ldr q25, [%[b_ptr0], #0x10]\n"
- ".word 0x4f81eb1a // sdot v26.4s, v24.16b, v1.4b[2]\n"
+ ".inst 0x4f81eb1a // sdot v26.4s, v24.16b, v1.4b[2]\n"
"add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f85eb1b // sdot v27.4s, v24.16b, v5.4b[2]\n"
- ".word 0x4f89eb1c // sdot v28.4s, v24.16b, v9.4b[2]\n"
- ".word 0x4f8deb1d // sdot v29.4s, v24.16b, v13.4b[2]\n"
- ".word 0x4f91eb1e // sdot v30.4s, v24.16b, v17.4b[2]\n"
- ".word 0x4f95eb1f // sdot v31.4s, v24.16b, v21.4b[2]\n"
+ ".inst 0x4f85eb1b // sdot v27.4s, v24.16b, v5.4b[2]\n"
+ ".inst 0x4f89eb1c // sdot v28.4s, v24.16b, v9.4b[2]\n"
+ ".inst 0x4f8deb1d // sdot v29.4s, v24.16b, v13.4b[2]\n"
+ ".inst 0x4f91eb1e // sdot v30.4s, v24.16b, v17.4b[2]\n"
+ ".inst 0x4f95eb1f // sdot v31.4s, v24.16b, v21.4b[2]\n"
"ldr q24, [%[b_ptr0]]\n"
- ".word 0x4fa1eb3a // sdot v26.4s, v25.16b, v1.4b[3]\n"
- ".word 0x4fa5eb3b // sdot v27.4s, v25.16b, v5.4b[3]\n"
- ".word 0x4fa9eb3c // sdot v28.4s, v25.16b, v9.4b[3]\n"
- ".word 0x4fadeb3d // sdot v29.4s, v25.16b, v13.4b[3]\n"
- ".word 0x4fb1eb3e // sdot v30.4s, v25.16b, v17.4b[3]\n"
- ".word 0x4fb5eb3f // sdot v31.4s, v25.16b, v21.4b[3]\n"
+ ".inst 0x4fa1eb3a // sdot v26.4s, v25.16b, v1.4b[3]\n"
+ ".inst 0x4fa5eb3b // sdot v27.4s, v25.16b, v5.4b[3]\n"
+ ".inst 0x4fa9eb3c // sdot v28.4s, v25.16b, v9.4b[3]\n"
+ ".inst 0x4fadeb3d // sdot v29.4s, v25.16b, v13.4b[3]\n"
+ ".inst 0x4fb1eb3e // sdot v30.4s, v25.16b, v17.4b[3]\n"
+ ".inst 0x4fb5eb3f // sdot v31.4s, v25.16b, v21.4b[3]\n"
"ldr q25, [%[b_ptr0], #0x10]\n"
- ".word 0x4f82e31a // sdot v26.4s, v24.16b, v2.4b[0]\n"
+ ".inst 0x4f82e31a // sdot v26.4s, v24.16b, v2.4b[0]\n"
"add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f86e31b // sdot v27.4s, v24.16b, v6.4b[0]\n"
- ".word 0x4f8ae31c // sdot v28.4s, v24.16b, v10.4b[0]\n"
- ".word 0x4f8ee31d // sdot v29.4s, v24.16b, v14.4b[0]\n"
- ".word 0x4f92e31e // sdot v30.4s, v24.16b, v18.4b[0]\n"
- ".word 0x4f96e31f // sdot v31.4s, v24.16b, v22.4b[0]\n"
+ ".inst 0x4f86e31b // sdot v27.4s, v24.16b, v6.4b[0]\n"
+ ".inst 0x4f8ae31c // sdot v28.4s, v24.16b, v10.4b[0]\n"
+ ".inst 0x4f8ee31d // sdot v29.4s, v24.16b, v14.4b[0]\n"
+ ".inst 0x4f92e31e // sdot v30.4s, v24.16b, v18.4b[0]\n"
+ ".inst 0x4f96e31f // sdot v31.4s, v24.16b, v22.4b[0]\n"
"ldr q24, [%[b_ptr0]]\n"
- ".word 0x4fa2e33a // sdot v26.4s, v25.16b, v2.4b[1]\n"
- ".word 0x4fa6e33b // sdot v27.4s, v25.16b, v6.4b[1]\n"
- ".word 0x4faae33c // sdot v28.4s, v25.16b, v10.4b[1]\n"
- ".word 0x4faee33d // sdot v29.4s, v25.16b, v14.4b[1]\n"
- ".word 0x4fb2e33e // sdot v30.4s, v25.16b, v18.4b[1]\n"
- ".word 0x4fb6e33f // sdot v31.4s, v25.16b, v22.4b[1]\n"
+ ".inst 0x4fa2e33a // sdot v26.4s, v25.16b, v2.4b[1]\n"
+ ".inst 0x4fa6e33b // sdot v27.4s, v25.16b, v6.4b[1]\n"
+ ".inst 0x4faae33c // sdot v28.4s, v25.16b, v10.4b[1]\n"
+ ".inst 0x4faee33d // sdot v29.4s, v25.16b, v14.4b[1]\n"
+ ".inst 0x4fb2e33e // sdot v30.4s, v25.16b, v18.4b[1]\n"
+ ".inst 0x4fb6e33f // sdot v31.4s, v25.16b, v22.4b[1]\n"
"ldr q25, [%[b_ptr0], #0x10]\n"
- ".word 0x4f82eb1a // sdot v26.4s, v24.16b, v2.4b[2]\n"
+ ".inst 0x4f82eb1a // sdot v26.4s, v24.16b, v2.4b[2]\n"
"add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f86eb1b // sdot v27.4s, v24.16b, v6.4b[2]\n"
- ".word 0x4f8aeb1c // sdot v28.4s, v24.16b, v10.4b[2]\n"
- ".word 0x4f8eeb1d // sdot v29.4s, v24.16b, v14.4b[2]\n"
- ".word 0x4f92eb1e // sdot v30.4s, v24.16b, v18.4b[2]\n"
- ".word 0x4f96eb1f // sdot v31.4s, v24.16b, v22.4b[2]\n"
+ ".inst 0x4f86eb1b // sdot v27.4s, v24.16b, v6.4b[2]\n"
+ ".inst 0x4f8aeb1c // sdot v28.4s, v24.16b, v10.4b[2]\n"
+ ".inst 0x4f8eeb1d // sdot v29.4s, v24.16b, v14.4b[2]\n"
+ ".inst 0x4f92eb1e // sdot v30.4s, v24.16b, v18.4b[2]\n"
+ ".inst 0x4f96eb1f // sdot v31.4s, v24.16b, v22.4b[2]\n"
"ldr q24, [%[b_ptr0]]\n"
- ".word 0x4fa2eb3a // sdot v26.4s, v25.16b, v2.4b[3]\n"
- ".word 0x4fa6eb3b // sdot v27.4s, v25.16b, v6.4b[3]\n"
- ".word 0x4faaeb3c // sdot v28.4s, v25.16b, v10.4b[3]\n"
- ".word 0x4faeeb3d // sdot v29.4s, v25.16b, v14.4b[3]\n"
- ".word 0x4fb2eb3e // sdot v30.4s, v25.16b, v18.4b[3]\n"
- ".word 0x4fb6eb3f // sdot v31.4s, v25.16b, v22.4b[3]\n"
+ ".inst 0x4fa2eb3a // sdot v26.4s, v25.16b, v2.4b[3]\n"
+ ".inst 0x4fa6eb3b // sdot v27.4s, v25.16b, v6.4b[3]\n"
+ ".inst 0x4faaeb3c // sdot v28.4s, v25.16b, v10.4b[3]\n"
+ ".inst 0x4faeeb3d // sdot v29.4s, v25.16b, v14.4b[3]\n"
+ ".inst 0x4fb2eb3e // sdot v30.4s, v25.16b, v18.4b[3]\n"
+ ".inst 0x4fb6eb3f // sdot v31.4s, v25.16b, v22.4b[3]\n"
"ldr q25, [%[b_ptr0], #0x10]\n"
- ".word 0x4f83e31a // sdot v26.4s, v24.16b, v3.4b[0]\n"
- "add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f87e31b // sdot v27.4s, v24.16b, v7.4b[0]\n"
- ".word 0x4f8be31c // sdot v28.4s, v24.16b, v11.4b[0]\n"
- ".word 0x4f8fe31d // sdot v29.4s, v24.16b, v15.4b[0]\n"
- ".word 0x4f93e31e // sdot v30.4s, v24.16b, v19.4b[0]\n"
- ".word 0x4f97e31f // sdot v31.4s, v24.16b, v23.4b[0]\n"
- ".word 0x4fa3e33a // sdot v26.4s, v25.16b, v3.4b[1]\n"
- ".word 0x4fa7e33b // sdot v27.4s, v25.16b, v7.4b[1]\n"
- ".word 0x4fabe33c // sdot v28.4s, v25.16b, v11.4b[1]\n"
- ".word 0x4fafe33d // sdot v29.4s, v25.16b, v15.4b[1]\n"
- ".word 0x4fb3e33e // sdot v30.4s, v25.16b, v19.4b[1]\n"
- ".word 0x4fb7e33f // sdot v31.4s, v25.16b, v23.4b[1]\n"
+ ".inst 0x4f83e31a // sdot v26.4s, v24.16b, v3.4b[0]\n"
+ "add %[b_ptr0], %[b_ptr0], #0x20\n"
+ ".inst 0x4f87e31b // sdot v27.4s, v24.16b, v7.4b[0]\n"
+ ".inst 0x4f8be31c // sdot v28.4s, v24.16b, v11.4b[0]\n"
+ ".inst 0x4f8fe31d // sdot v29.4s, v24.16b, v15.4b[0]\n"
+ ".inst 0x4f93e31e // sdot v30.4s, v24.16b, v19.4b[0]\n"
+ ".inst 0x4f97e31f // sdot v31.4s, v24.16b, v23.4b[0]\n"
+ ".inst 0x4fa3e33a // sdot v26.4s, v25.16b, v3.4b[1]\n"
+ ".inst 0x4fa7e33b // sdot v27.4s, v25.16b, v7.4b[1]\n"
+ ".inst 0x4fabe33c // sdot v28.4s, v25.16b, v11.4b[1]\n"
+ ".inst 0x4fafe33d // sdot v29.4s, v25.16b, v15.4b[1]\n"
+ ".inst 0x4fb3e33e // sdot v30.4s, v25.16b, v19.4b[1]\n"
+ ".inst 0x4fb7e33f // sdot v31.4s, v25.16b, v23.4b[1]\n"
"6:\n"
"str q26, [%[c_ptr0]]\n"
+ "add %[c_ptr0], %[c_ptr0], #0x10\n"
"str q27, [c_ptr1]\n"
+ "add c_ptr1, c_ptr1, #0x10\n"
"str q28, [c_ptr2]\n"
+ "add c_ptr2, c_ptr2, #0x10\n"
"str q29, [c_ptr3]\n"
+ "add c_ptr3, c_ptr3, #0x10\n"
"str q30, [c_ptr4]\n"
+ "add c_ptr4, c_ptr4, #0x10\n"
"str q31, [c_ptr5]\n"
+ "add c_ptr5, c_ptr5, #0x10\n"
".unreq a_ptr1\n"
".unreq a_ptr2\n"
".unreq a_ptr3\n"
@@ -2949,119 +3064,119 @@ void a64_smallK_hybrid_s8s32_dot_4x6_a55(const int8_t *A, int lda, const int8_t
"prfm PLDL1KEEP, [a_ptr5, #0xc0]\n"
"movi v31.4s, #0\n"
"prfm PLDL1KEEP, [a_ptr5, #0x100]\n"
- ".word 0x4f80e31a // sdot v26.4s, v24.16b, v0.4b[0]\n"
+ ".inst 0x4f80e31a // sdot v26.4s, v24.16b, v0.4b[0]\n"
"prfm PLDL1KEEP, [a_ptr5, #0x140]\n"
- ".word 0x4f84e31b // sdot v27.4s, v24.16b, v4.4b[0]\n"
+ ".inst 0x4f84e31b // sdot v27.4s, v24.16b, v4.4b[0]\n"
"prfm PLDL1KEEP, [a_ptr5, #0x180]\n"
- ".word 0x4f88e31c // sdot v28.4s, v24.16b, v8.4b[0]\n"
+ ".inst 0x4f88e31c // sdot v28.4s, v24.16b, v8.4b[0]\n"
"add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f8ce31d // sdot v29.4s, v24.16b, v12.4b[0]\n"
- ".word 0x4f90e31e // sdot v30.4s, v24.16b, v16.4b[0]\n"
- ".word 0x4f94e31f // sdot v31.4s, v24.16b, v20.4b[0]\n"
+ ".inst 0x4f8ce31d // sdot v29.4s, v24.16b, v12.4b[0]\n"
+ ".inst 0x4f90e31e // sdot v30.4s, v24.16b, v16.4b[0]\n"
+ ".inst 0x4f94e31f // sdot v31.4s, v24.16b, v20.4b[0]\n"
"ldr q24, [%[b_ptr0]]\n"
- ".word 0x4fa0e33a // sdot v26.4s, v25.16b, v0.4b[1]\n"
- ".word 0x4fa4e33b // sdot v27.4s, v25.16b, v4.4b[1]\n"
- ".word 0x4fa8e33c // sdot v28.4s, v25.16b, v8.4b[1]\n"
- ".word 0x4face33d // sdot v29.4s, v25.16b, v12.4b[1]\n"
- ".word 0x4fb0e33e // sdot v30.4s, v25.16b, v16.4b[1]\n"
- ".word 0x4fb4e33f // sdot v31.4s, v25.16b, v20.4b[1]\n"
+ ".inst 0x4fa0e33a // sdot v26.4s, v25.16b, v0.4b[1]\n"
+ ".inst 0x4fa4e33b // sdot v27.4s, v25.16b, v4.4b[1]\n"
+ ".inst 0x4fa8e33c // sdot v28.4s, v25.16b, v8.4b[1]\n"
+ ".inst 0x4face33d // sdot v29.4s, v25.16b, v12.4b[1]\n"
+ ".inst 0x4fb0e33e // sdot v30.4s, v25.16b, v16.4b[1]\n"
+ ".inst 0x4fb4e33f // sdot v31.4s, v25.16b, v20.4b[1]\n"
"ldr q25, [%[b_ptr0], #0x10]\n"
- ".word 0x4f80eb1a // sdot v26.4s, v24.16b, v0.4b[2]\n"
+ ".inst 0x4f80eb1a // sdot v26.4s, v24.16b, v0.4b[2]\n"
"add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f84eb1b // sdot v27.4s, v24.16b, v4.4b[2]\n"
- ".word 0x4f88eb1c // sdot v28.4s, v24.16b, v8.4b[2]\n"
- ".word 0x4f8ceb1d // sdot v29.4s, v24.16b, v12.4b[2]\n"
- ".word 0x4f90eb1e // sdot v30.4s, v24.16b, v16.4b[2]\n"
- ".word 0x4f94eb1f // sdot v31.4s, v24.16b, v20.4b[2]\n"
+ ".inst 0x4f84eb1b // sdot v27.4s, v24.16b, v4.4b[2]\n"
+ ".inst 0x4f88eb1c // sdot v28.4s, v24.16b, v8.4b[2]\n"
+ ".inst 0x4f8ceb1d // sdot v29.4s, v24.16b, v12.4b[2]\n"
+ ".inst 0x4f90eb1e // sdot v30.4s, v24.16b, v16.4b[2]\n"
+ ".inst 0x4f94eb1f // sdot v31.4s, v24.16b, v20.4b[2]\n"
"ldr q24, [%[b_ptr0]]\n"
- ".word 0x4fa0eb3a // sdot v26.4s, v25.16b, v0.4b[3]\n"
- ".word 0x4fa4eb3b // sdot v27.4s, v25.16b, v4.4b[3]\n"
- ".word 0x4fa8eb3c // sdot v28.4s, v25.16b, v8.4b[3]\n"
- ".word 0x4faceb3d // sdot v29.4s, v25.16b, v12.4b[3]\n"
- ".word 0x4fb0eb3e // sdot v30.4s, v25.16b, v16.4b[3]\n"
- ".word 0x4fb4eb3f // sdot v31.4s, v25.16b, v20.4b[3]\n"
+ ".inst 0x4fa0eb3a // sdot v26.4s, v25.16b, v0.4b[3]\n"
+ ".inst 0x4fa4eb3b // sdot v27.4s, v25.16b, v4.4b[3]\n"
+ ".inst 0x4fa8eb3c // sdot v28.4s, v25.16b, v8.4b[3]\n"
+ ".inst 0x4faceb3d // sdot v29.4s, v25.16b, v12.4b[3]\n"
+ ".inst 0x4fb0eb3e // sdot v30.4s, v25.16b, v16.4b[3]\n"
+ ".inst 0x4fb4eb3f // sdot v31.4s, v25.16b, v20.4b[3]\n"
"ldr q25, [%[b_ptr0], #0x10]\n"
- ".word 0x4f81e31a // sdot v26.4s, v24.16b, v1.4b[0]\n"
+ ".inst 0x4f81e31a // sdot v26.4s, v24.16b, v1.4b[0]\n"
"add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f85e31b // sdot v27.4s, v24.16b, v5.4b[0]\n"
- ".word 0x4f89e31c // sdot v28.4s, v24.16b, v9.4b[0]\n"
- ".word 0x4f8de31d // sdot v29.4s, v24.16b, v13.4b[0]\n"
- ".word 0x4f91e31e // sdot v30.4s, v24.16b, v17.4b[0]\n"
- ".word 0x4f95e31f // sdot v31.4s, v24.16b, v21.4b[0]\n"
+ ".inst 0x4f85e31b // sdot v27.4s, v24.16b, v5.4b[0]\n"
+ ".inst 0x4f89e31c // sdot v28.4s, v24.16b, v9.4b[0]\n"
+ ".inst 0x4f8de31d // sdot v29.4s, v24.16b, v13.4b[0]\n"
+ ".inst 0x4f91e31e // sdot v30.4s, v24.16b, v17.4b[0]\n"
+ ".inst 0x4f95e31f // sdot v31.4s, v24.16b, v21.4b[0]\n"
"ldr q24, [%[b_ptr0]]\n"
- ".word 0x4fa1e33a // sdot v26.4s, v25.16b, v1.4b[1]\n"
- ".word 0x4fa5e33b // sdot v27.4s, v25.16b, v5.4b[1]\n"
- ".word 0x4fa9e33c // sdot v28.4s, v25.16b, v9.4b[1]\n"
- ".word 0x4fade33d // sdot v29.4s, v25.16b, v13.4b[1]\n"
- ".word 0x4fb1e33e // sdot v30.4s, v25.16b, v17.4b[1]\n"
- ".word 0x4fb5e33f // sdot v31.4s, v25.16b, v21.4b[1]\n"
+ ".inst 0x4fa1e33a // sdot v26.4s, v25.16b, v1.4b[1]\n"
+ ".inst 0x4fa5e33b // sdot v27.4s, v25.16b, v5.4b[1]\n"
+ ".inst 0x4fa9e33c // sdot v28.4s, v25.16b, v9.4b[1]\n"
+ ".inst 0x4fade33d // sdot v29.4s, v25.16b, v13.4b[1]\n"
+ ".inst 0x4fb1e33e // sdot v30.4s, v25.16b, v17.4b[1]\n"
+ ".inst 0x4fb5e33f // sdot v31.4s, v25.16b, v21.4b[1]\n"
"ldr q25, [%[b_ptr0], #0x10]\n"
- ".word 0x4f81eb1a // sdot v26.4s, v24.16b, v1.4b[2]\n"
+ ".inst 0x4f81eb1a // sdot v26.4s, v24.16b, v1.4b[2]\n"
"add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f85eb1b // sdot v27.4s, v24.16b, v5.4b[2]\n"
- ".word 0x4f89eb1c // sdot v28.4s, v24.16b, v9.4b[2]\n"
- ".word 0x4f8deb1d // sdot v29.4s, v24.16b, v13.4b[2]\n"
- ".word 0x4f91eb1e // sdot v30.4s, v24.16b, v17.4b[2]\n"
- ".word 0x4f95eb1f // sdot v31.4s, v24.16b, v21.4b[2]\n"
+ ".inst 0x4f85eb1b // sdot v27.4s, v24.16b, v5.4b[2]\n"
+ ".inst 0x4f89eb1c // sdot v28.4s, v24.16b, v9.4b[2]\n"
+ ".inst 0x4f8deb1d // sdot v29.4s, v24.16b, v13.4b[2]\n"
+ ".inst 0x4f91eb1e // sdot v30.4s, v24.16b, v17.4b[2]\n"
+ ".inst 0x4f95eb1f // sdot v31.4s, v24.16b, v21.4b[2]\n"
"ldr q24, [%[b_ptr0]]\n"
- ".word 0x4fa1eb3a // sdot v26.4s, v25.16b, v1.4b[3]\n"
- ".word 0x4fa5eb3b // sdot v27.4s, v25.16b, v5.4b[3]\n"
- ".word 0x4fa9eb3c // sdot v28.4s, v25.16b, v9.4b[3]\n"
- ".word 0x4fadeb3d // sdot v29.4s, v25.16b, v13.4b[3]\n"
- ".word 0x4fb1eb3e // sdot v30.4s, v25.16b, v17.4b[3]\n"
- ".word 0x4fb5eb3f // sdot v31.4s, v25.16b, v21.4b[3]\n"
+ ".inst 0x4fa1eb3a // sdot v26.4s, v25.16b, v1.4b[3]\n"
+ ".inst 0x4fa5eb3b // sdot v27.4s, v25.16b, v5.4b[3]\n"
+ ".inst 0x4fa9eb3c // sdot v28.4s, v25.16b, v9.4b[3]\n"
+ ".inst 0x4fadeb3d // sdot v29.4s, v25.16b, v13.4b[3]\n"
+ ".inst 0x4fb1eb3e // sdot v30.4s, v25.16b, v17.4b[3]\n"
+ ".inst 0x4fb5eb3f // sdot v31.4s, v25.16b, v21.4b[3]\n"
"ldr q25, [%[b_ptr0], #0x10]\n"
- ".word 0x4f82e31a // sdot v26.4s, v24.16b, v2.4b[0]\n"
+ ".inst 0x4f82e31a // sdot v26.4s, v24.16b, v2.4b[0]\n"
"add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f86e31b // sdot v27.4s, v24.16b, v6.4b[0]\n"
- ".word 0x4f8ae31c // sdot v28.4s, v24.16b, v10.4b[0]\n"
- ".word 0x4f8ee31d // sdot v29.4s, v24.16b, v14.4b[0]\n"
- ".word 0x4f92e31e // sdot v30.4s, v24.16b, v18.4b[0]\n"
- ".word 0x4f96e31f // sdot v31.4s, v24.16b, v22.4b[0]\n"
+ ".inst 0x4f86e31b // sdot v27.4s, v24.16b, v6.4b[0]\n"
+ ".inst 0x4f8ae31c // sdot v28.4s, v24.16b, v10.4b[0]\n"
+ ".inst 0x4f8ee31d // sdot v29.4s, v24.16b, v14.4b[0]\n"
+ ".inst 0x4f92e31e // sdot v30.4s, v24.16b, v18.4b[0]\n"
+ ".inst 0x4f96e31f // sdot v31.4s, v24.16b, v22.4b[0]\n"
"ldr q24, [%[b_ptr0]]\n"
- ".word 0x4fa2e33a // sdot v26.4s, v25.16b, v2.4b[1]\n"
- ".word 0x4fa6e33b // sdot v27.4s, v25.16b, v6.4b[1]\n"
- ".word 0x4faae33c // sdot v28.4s, v25.16b, v10.4b[1]\n"
- ".word 0x4faee33d // sdot v29.4s, v25.16b, v14.4b[1]\n"
- ".word 0x4fb2e33e // sdot v30.4s, v25.16b, v18.4b[1]\n"
- ".word 0x4fb6e33f // sdot v31.4s, v25.16b, v22.4b[1]\n"
+ ".inst 0x4fa2e33a // sdot v26.4s, v25.16b, v2.4b[1]\n"
+ ".inst 0x4fa6e33b // sdot v27.4s, v25.16b, v6.4b[1]\n"
+ ".inst 0x4faae33c // sdot v28.4s, v25.16b, v10.4b[1]\n"
+ ".inst 0x4faee33d // sdot v29.4s, v25.16b, v14.4b[1]\n"
+ ".inst 0x4fb2e33e // sdot v30.4s, v25.16b, v18.4b[1]\n"
+ ".inst 0x4fb6e33f // sdot v31.4s, v25.16b, v22.4b[1]\n"
"ldr q25, [%[b_ptr0], #0x10]\n"
- ".word 0x4f82eb1a // sdot v26.4s, v24.16b, v2.4b[2]\n"
+ ".inst 0x4f82eb1a // sdot v26.4s, v24.16b, v2.4b[2]\n"
"add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f86eb1b // sdot v27.4s, v24.16b, v6.4b[2]\n"
- ".word 0x4f8aeb1c // sdot v28.4s, v24.16b, v10.4b[2]\n"
- ".word 0x4f8eeb1d // sdot v29.4s, v24.16b, v14.4b[2]\n"
- ".word 0x4f92eb1e // sdot v30.4s, v24.16b, v18.4b[2]\n"
- ".word 0x4f96eb1f // sdot v31.4s, v24.16b, v22.4b[2]\n"
+ ".inst 0x4f86eb1b // sdot v27.4s, v24.16b, v6.4b[2]\n"
+ ".inst 0x4f8aeb1c // sdot v28.4s, v24.16b, v10.4b[2]\n"
+ ".inst 0x4f8eeb1d // sdot v29.4s, v24.16b, v14.4b[2]\n"
+ ".inst 0x4f92eb1e // sdot v30.4s, v24.16b, v18.4b[2]\n"
+ ".inst 0x4f96eb1f // sdot v31.4s, v24.16b, v22.4b[2]\n"
"ldr q24, [%[b_ptr0]]\n"
- ".word 0x4fa2eb3a // sdot v26.4s, v25.16b, v2.4b[3]\n"
- ".word 0x4fa6eb3b // sdot v27.4s, v25.16b, v6.4b[3]\n"
- ".word 0x4faaeb3c // sdot v28.4s, v25.16b, v10.4b[3]\n"
- ".word 0x4faeeb3d // sdot v29.4s, v25.16b, v14.4b[3]\n"
- ".word 0x4fb2eb3e // sdot v30.4s, v25.16b, v18.4b[3]\n"
- ".word 0x4fb6eb3f // sdot v31.4s, v25.16b, v22.4b[3]\n"
+ ".inst 0x4fa2eb3a // sdot v26.4s, v25.16b, v2.4b[3]\n"
+ ".inst 0x4fa6eb3b // sdot v27.4s, v25.16b, v6.4b[3]\n"
+ ".inst 0x4faaeb3c // sdot v28.4s, v25.16b, v10.4b[3]\n"
+ ".inst 0x4faeeb3d // sdot v29.4s, v25.16b, v14.4b[3]\n"
+ ".inst 0x4fb2eb3e // sdot v30.4s, v25.16b, v18.4b[3]\n"
+ ".inst 0x4fb6eb3f // sdot v31.4s, v25.16b, v22.4b[3]\n"
"ldr q25, [%[b_ptr0], #0x10]\n"
- ".word 0x4f83e31a // sdot v26.4s, v24.16b, v3.4b[0]\n"
+ ".inst 0x4f83e31a // sdot v26.4s, v24.16b, v3.4b[0]\n"
"add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f87e31b // sdot v27.4s, v24.16b, v7.4b[0]\n"
- ".word 0x4f8be31c // sdot v28.4s, v24.16b, v11.4b[0]\n"
- ".word 0x4f8fe31d // sdot v29.4s, v24.16b, v15.4b[0]\n"
- ".word 0x4f93e31e // sdot v30.4s, v24.16b, v19.4b[0]\n"
- ".word 0x4f97e31f // sdot v31.4s, v24.16b, v23.4b[0]\n"
+ ".inst 0x4f87e31b // sdot v27.4s, v24.16b, v7.4b[0]\n"
+ ".inst 0x4f8be31c // sdot v28.4s, v24.16b, v11.4b[0]\n"
+ ".inst 0x4f8fe31d // sdot v29.4s, v24.16b, v15.4b[0]\n"
+ ".inst 0x4f93e31e // sdot v30.4s, v24.16b, v19.4b[0]\n"
+ ".inst 0x4f97e31f // sdot v31.4s, v24.16b, v23.4b[0]\n"
"ldr q24, [%[b_ptr0]]\n"
- ".word 0x4fa3e33a // sdot v26.4s, v25.16b, v3.4b[1]\n"
+ ".inst 0x4fa3e33a // sdot v26.4s, v25.16b, v3.4b[1]\n"
"add %[b_ptr0], %[b_ptr0], #0x10\n"
- ".word 0x4fa7e33b // sdot v27.4s, v25.16b, v7.4b[1]\n"
- ".word 0x4fabe33c // sdot v28.4s, v25.16b, v11.4b[1]\n"
- ".word 0x4fafe33d // sdot v29.4s, v25.16b, v15.4b[1]\n"
- ".word 0x4fb3e33e // sdot v30.4s, v25.16b, v19.4b[1]\n"
- ".word 0x4fb7e33f // sdot v31.4s, v25.16b, v23.4b[1]\n"
- ".word 0x4f83eb1a // sdot v26.4s, v24.16b, v3.4b[2]\n"
- ".word 0x4f87eb1b // sdot v27.4s, v24.16b, v7.4b[2]\n"
- ".word 0x4f8beb1c // sdot v28.4s, v24.16b, v11.4b[2]\n"
- ".word 0x4f8feb1d // sdot v29.4s, v24.16b, v15.4b[2]\n"
- ".word 0x4f93eb1e // sdot v30.4s, v24.16b, v19.4b[2]\n"
- ".word 0x4f97eb1f // sdot v31.4s, v24.16b, v23.4b[2]\n"
+ ".inst 0x4fa7e33b // sdot v27.4s, v25.16b, v7.4b[1]\n"
+ ".inst 0x4fabe33c // sdot v28.4s, v25.16b, v11.4b[1]\n"
+ ".inst 0x4fafe33d // sdot v29.4s, v25.16b, v15.4b[1]\n"
+ ".inst 0x4fb3e33e // sdot v30.4s, v25.16b, v19.4b[1]\n"
+ ".inst 0x4fb7e33f // sdot v31.4s, v25.16b, v23.4b[1]\n"
+ ".inst 0x4f83eb1a // sdot v26.4s, v24.16b, v3.4b[2]\n"
+ ".inst 0x4f87eb1b // sdot v27.4s, v24.16b, v7.4b[2]\n"
+ ".inst 0x4f8beb1c // sdot v28.4s, v24.16b, v11.4b[2]\n"
+ ".inst 0x4f8feb1d // sdot v29.4s, v24.16b, v15.4b[2]\n"
+ ".inst 0x4f93eb1e // sdot v30.4s, v24.16b, v19.4b[2]\n"
+ ".inst 0x4f97eb1f // sdot v31.4s, v24.16b, v23.4b[2]\n"
"cbz %[loops], 6f\n"
"ldr d24, [%[b_ptr0]]\n"
"subs %[loops], %[loops], #0x1\n"
@@ -3073,299 +3188,317 @@ void a64_smallK_hybrid_s8s32_dot_4x6_a55(const int8_t *A, int lda, const int8_t
"ins v25.d[1], temploadreg1\n"
"b.eq 7f\n"
"8:\n"
- "str q26, [%[c_ptr0]], #0x10\n"
+ "str q26, [%[c_ptr0]]\n"
"subs %[loops], %[loops], #0x1\n"
"movi v26.4s, #0\n"
"ldr temploadreg0, [%[b_ptr0], #0x8]\n"
"ldr temploadreg1, [%[b_ptr0], #0x18]\n"
- "prfm PSTL1KEEP, [%[c_ptr0], #0x40]\n"
- "str q27, [c_ptr1], #0x10\n"
+ "add %[c_ptr0], %[c_ptr0], #0x10\n"
+ "str q27, [c_ptr1]\n"
+ "add c_ptr1, c_ptr1, #0x10\n"
"movi v27.4s, #0\n"
- ".word 0x4f80e31a // sdot v26.4s, v24.16b, v0.4b[0]\n"
- "str q28, [c_ptr2], #0x10\n"
- "prfm PSTL1KEEP, [c_ptr1, #0x40]\n"
+ "prfm PSTL1KEEP, [%[c_ptr0], #0x40]\n"
+ ".inst 0x4f80e31a // sdot v26.4s, v24.16b, v0.4b[0]\n"
+ "str q28, [c_ptr2]\n"
"movi v28.4s, #0\n"
- ".word 0x4f84e31b // sdot v27.4s, v24.16b, v4.4b[0]\n"
- "str q29, [c_ptr3], #0x10\n"
+ "add c_ptr2, c_ptr2, #0x10\n"
+ ".inst 0x4f84e31b // sdot v27.4s, v24.16b, v4.4b[0]\n"
+ "str q29, [c_ptr3]\n"
"movi v29.4s, #0\n"
- "prfm PSTL1KEEP, [c_ptr2, #0x40]\n"
- ".word 0x4f88e31c // sdot v28.4s, v24.16b, v8.4b[0]\n"
- "str q30, [c_ptr4], #0x10\n"
+ "add c_ptr3, c_ptr3, #0x10\n"
+ ".inst 0x4f88e31c // sdot v28.4s, v24.16b, v8.4b[0]\n"
+ "str q30, [c_ptr4]\n"
"movi v30.4s, #0\n"
- "prfm PSTL1KEEP, [c_ptr3, #0x40]\n"
- ".word 0x4f8ce31d // sdot v29.4s, v24.16b, v12.4b[0]\n"
- "str q31, [c_ptr5], #0x10\n"
+ "add c_ptr4, c_ptr4, #0x10\n"
+ ".inst 0x4f8ce31d // sdot v29.4s, v24.16b, v12.4b[0]\n"
+ "str q31, [c_ptr5]\n"
"movi v31.4s, #0\n"
- "prfm PSTL1KEEP, [c_ptr4, #0x40]\n"
- ".word 0x4f90e31e // sdot v30.4s, v24.16b, v16.4b[0]\n"
- "prfm PSTL1KEEP, [c_ptr5, #0x40]\n"
- ".word 0x4f94e31f // sdot v31.4s, v24.16b, v20.4b[0]\n"
+ "add c_ptr5, c_ptr5, #0x10\n"
+ ".inst 0x4f90e31e // sdot v30.4s, v24.16b, v16.4b[0]\n"
+ "prfm PSTL1KEEP, [c_ptr1, #0x40]\n"
+ ".inst 0x4f94e31f // sdot v31.4s, v24.16b, v20.4b[0]\n"
"ldr d24, [%[b_ptr0]]\n"
- ".word 0x4fa0e33a // sdot v26.4s, v25.16b, v0.4b[1]\n"
- ".word 0x4fa4e33b // sdot v27.4s, v25.16b, v4.4b[1]\n"
- ".word 0x4fa8e33c // sdot v28.4s, v25.16b, v8.4b[1]\n"
+ ".inst 0x4fa0e33a // sdot v26.4s, v25.16b, v0.4b[1]\n"
+ "prfm PSTL1KEEP, [c_ptr2, #0x40]\n"
+ ".inst 0x4fa4e33b // sdot v27.4s, v25.16b, v4.4b[1]\n"
"ins v24.d[1], temploadreg0\n"
- ".word 0x4face33d // sdot v29.4s, v25.16b, v12.4b[1]\n"
- ".word 0x4fb0e33e // sdot v30.4s, v25.16b, v16.4b[1]\n"
- ".word 0x4fb4e33f // sdot v31.4s, v25.16b, v20.4b[1]\n"
+ ".inst 0x4fa8e33c // sdot v28.4s, v25.16b, v8.4b[1]\n"
+ "prfm PSTL1KEEP, [c_ptr3, #0x40]\n"
+ ".inst 0x4face33d // sdot v29.4s, v25.16b, v12.4b[1]\n"
+ "prfm PSTL1KEEP, [c_ptr4, #0x40]\n"
+ ".inst 0x4fb0e33e // sdot v30.4s, v25.16b, v16.4b[1]\n"
+ "prfm PSTL1KEEP, [c_ptr5, #0x40]\n"
+ ".inst 0x4fb4e33f // sdot v31.4s, v25.16b, v20.4b[1]\n"
"ldr d25, [%[b_ptr0], #0x10]\n"
- ".word 0x4f80eb1a // sdot v26.4s, v24.16b, v0.4b[2]\n"
+ ".inst 0x4f80eb1a // sdot v26.4s, v24.16b, v0.4b[2]\n"
"add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f84eb1b // sdot v27.4s, v24.16b, v4.4b[2]\n"
+ ".inst 0x4f84eb1b // sdot v27.4s, v24.16b, v4.4b[2]\n"
"ins v25.d[1], temploadreg1\n"
- ".word 0x4f88eb1c // sdot v28.4s, v24.16b, v8.4b[2]\n"
+ ".inst 0x4f88eb1c // sdot v28.4s, v24.16b, v8.4b[2]\n"
"ldr temploadreg0, [%[b_ptr0], #0x8]\n"
- ".word 0x4f8ceb1d // sdot v29.4s, v24.16b, v12.4b[2]\n"
+ ".inst 0x4f8ceb1d // sdot v29.4s, v24.16b, v12.4b[2]\n"
"ldr temploadreg1, [%[b_ptr0], #0x18]\n"
- ".word 0x4f90eb1e // sdot v30.4s, v24.16b, v16.4b[2]\n"
- ".word 0x4f94eb1f // sdot v31.4s, v24.16b, v20.4b[2]\n"
+ ".inst 0x4f90eb1e // sdot v30.4s, v24.16b, v16.4b[2]\n"
+ ".inst 0x4f94eb1f // sdot v31.4s, v24.16b, v20.4b[2]\n"
"ldr d24, [%[b_ptr0]]\n"
- ".word 0x4fa0eb3a // sdot v26.4s, v25.16b, v0.4b[3]\n"
- ".word 0x4fa4eb3b // sdot v27.4s, v25.16b, v4.4b[3]\n"
- ".word 0x4fa8eb3c // sdot v28.4s, v25.16b, v8.4b[3]\n"
+ ".inst 0x4fa0eb3a // sdot v26.4s, v25.16b, v0.4b[3]\n"
+ ".inst 0x4fa4eb3b // sdot v27.4s, v25.16b, v4.4b[3]\n"
+ ".inst 0x4fa8eb3c // sdot v28.4s, v25.16b, v8.4b[3]\n"
"ins v24.d[1], temploadreg0\n"
- ".word 0x4faceb3d // sdot v29.4s, v25.16b, v12.4b[3]\n"
- ".word 0x4fb0eb3e // sdot v30.4s, v25.16b, v16.4b[3]\n"
- ".word 0x4fb4eb3f // sdot v31.4s, v25.16b, v20.4b[3]\n"
+ ".inst 0x4faceb3d // sdot v29.4s, v25.16b, v12.4b[3]\n"
+ ".inst 0x4fb0eb3e // sdot v30.4s, v25.16b, v16.4b[3]\n"
+ ".inst 0x4fb4eb3f // sdot v31.4s, v25.16b, v20.4b[3]\n"
"ldr d25, [%[b_ptr0], #0x10]\n"
- ".word 0x4f81e31a // sdot v26.4s, v24.16b, v1.4b[0]\n"
+ ".inst 0x4f81e31a // sdot v26.4s, v24.16b, v1.4b[0]\n"
"add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f85e31b // sdot v27.4s, v24.16b, v5.4b[0]\n"
+ ".inst 0x4f85e31b // sdot v27.4s, v24.16b, v5.4b[0]\n"
"ins v25.d[1], temploadreg1\n"
- ".word 0x4f89e31c // sdot v28.4s, v24.16b, v9.4b[0]\n"
+ ".inst 0x4f89e31c // sdot v28.4s, v24.16b, v9.4b[0]\n"
"ldr temploadreg0, [%[b_ptr0], #0x8]\n"
- ".word 0x4f8de31d // sdot v29.4s, v24.16b, v13.4b[0]\n"
+ ".inst 0x4f8de31d // sdot v29.4s, v24.16b, v13.4b[0]\n"
"ldr temploadreg1, [%[b_ptr0], #0x18]\n"
- ".word 0x4f91e31e // sdot v30.4s, v24.16b, v17.4b[0]\n"
- ".word 0x4f95e31f // sdot v31.4s, v24.16b, v21.4b[0]\n"
+ ".inst 0x4f91e31e // sdot v30.4s, v24.16b, v17.4b[0]\n"
+ ".inst 0x4f95e31f // sdot v31.4s, v24.16b, v21.4b[0]\n"
"ldr d24, [%[b_ptr0]]\n"
- ".word 0x4fa1e33a // sdot v26.4s, v25.16b, v1.4b[1]\n"
- ".word 0x4fa5e33b // sdot v27.4s, v25.16b, v5.4b[1]\n"
- ".word 0x4fa9e33c // sdot v28.4s, v25.16b, v9.4b[1]\n"
+ ".inst 0x4fa1e33a // sdot v26.4s, v25.16b, v1.4b[1]\n"
+ ".inst 0x4fa5e33b // sdot v27.4s, v25.16b, v5.4b[1]\n"
+ ".inst 0x4fa9e33c // sdot v28.4s, v25.16b, v9.4b[1]\n"
"ins v24.d[1], temploadreg0\n"
- ".word 0x4fade33d // sdot v29.4s, v25.16b, v13.4b[1]\n"
- ".word 0x4fb1e33e // sdot v30.4s, v25.16b, v17.4b[1]\n"
- ".word 0x4fb5e33f // sdot v31.4s, v25.16b, v21.4b[1]\n"
+ ".inst 0x4fade33d // sdot v29.4s, v25.16b, v13.4b[1]\n"
+ ".inst 0x4fb1e33e // sdot v30.4s, v25.16b, v17.4b[1]\n"
+ ".inst 0x4fb5e33f // sdot v31.4s, v25.16b, v21.4b[1]\n"
"ldr d25, [%[b_ptr0], #0x10]\n"
- ".word 0x4f81eb1a // sdot v26.4s, v24.16b, v1.4b[2]\n"
+ ".inst 0x4f81eb1a // sdot v26.4s, v24.16b, v1.4b[2]\n"
"add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f85eb1b // sdot v27.4s, v24.16b, v5.4b[2]\n"
+ ".inst 0x4f85eb1b // sdot v27.4s, v24.16b, v5.4b[2]\n"
"ins v25.d[1], temploadreg1\n"
- ".word 0x4f89eb1c // sdot v28.4s, v24.16b, v9.4b[2]\n"
+ ".inst 0x4f89eb1c // sdot v28.4s, v24.16b, v9.4b[2]\n"
"ldr temploadreg0, [%[b_ptr0], #0x8]\n"
- ".word 0x4f8deb1d // sdot v29.4s, v24.16b, v13.4b[2]\n"
+ ".inst 0x4f8deb1d // sdot v29.4s, v24.16b, v13.4b[2]\n"
"ldr temploadreg1, [%[b_ptr0], #0x18]\n"
- ".word 0x4f91eb1e // sdot v30.4s, v24.16b, v17.4b[2]\n"
- ".word 0x4f95eb1f // sdot v31.4s, v24.16b, v21.4b[2]\n"
+ ".inst 0x4f91eb1e // sdot v30.4s, v24.16b, v17.4b[2]\n"
+ ".inst 0x4f95eb1f // sdot v31.4s, v24.16b, v21.4b[2]\n"
"ldr d24, [%[b_ptr0]]\n"
- ".word 0x4fa1eb3a // sdot v26.4s, v25.16b, v1.4b[3]\n"
- ".word 0x4fa5eb3b // sdot v27.4s, v25.16b, v5.4b[3]\n"
- ".word 0x4fa9eb3c // sdot v28.4s, v25.16b, v9.4b[3]\n"
+ ".inst 0x4fa1eb3a // sdot v26.4s, v25.16b, v1.4b[3]\n"
+ ".inst 0x4fa5eb3b // sdot v27.4s, v25.16b, v5.4b[3]\n"
+ ".inst 0x4fa9eb3c // sdot v28.4s, v25.16b, v9.4b[3]\n"
"ins v24.d[1], temploadreg0\n"
- ".word 0x4fadeb3d // sdot v29.4s, v25.16b, v13.4b[3]\n"
- ".word 0x4fb1eb3e // sdot v30.4s, v25.16b, v17.4b[3]\n"
- ".word 0x4fb5eb3f // sdot v31.4s, v25.16b, v21.4b[3]\n"
+ ".inst 0x4fadeb3d // sdot v29.4s, v25.16b, v13.4b[3]\n"
+ ".inst 0x4fb1eb3e // sdot v30.4s, v25.16b, v17.4b[3]\n"
+ ".inst 0x4fb5eb3f // sdot v31.4s, v25.16b, v21.4b[3]\n"
"ldr d25, [%[b_ptr0], #0x10]\n"
- ".word 0x4f82e31a // sdot v26.4s, v24.16b, v2.4b[0]\n"
+ ".inst 0x4f82e31a // sdot v26.4s, v24.16b, v2.4b[0]\n"
"add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f86e31b // sdot v27.4s, v24.16b, v6.4b[0]\n"
+ ".inst 0x4f86e31b // sdot v27.4s, v24.16b, v6.4b[0]\n"
"ins v25.d[1], temploadreg1\n"
- ".word 0x4f8ae31c // sdot v28.4s, v24.16b, v10.4b[0]\n"
+ ".inst 0x4f8ae31c // sdot v28.4s, v24.16b, v10.4b[0]\n"
"ldr temploadreg0, [%[b_ptr0], #0x8]\n"
- ".word 0x4f8ee31d // sdot v29.4s, v24.16b, v14.4b[0]\n"
+ ".inst 0x4f8ee31d // sdot v29.4s, v24.16b, v14.4b[0]\n"
"ldr temploadreg1, [%[b_ptr0], #0x18]\n"
- ".word 0x4f92e31e // sdot v30.4s, v24.16b, v18.4b[0]\n"
- ".word 0x4f96e31f // sdot v31.4s, v24.16b, v22.4b[0]\n"
+ ".inst 0x4f92e31e // sdot v30.4s, v24.16b, v18.4b[0]\n"
+ ".inst 0x4f96e31f // sdot v31.4s, v24.16b, v22.4b[0]\n"
"ldr d24, [%[b_ptr0]]\n"
- ".word 0x4fa2e33a // sdot v26.4s, v25.16b, v2.4b[1]\n"
- ".word 0x4fa6e33b // sdot v27.4s, v25.16b, v6.4b[1]\n"
- ".word 0x4faae33c // sdot v28.4s, v25.16b, v10.4b[1]\n"
+ ".inst 0x4fa2e33a // sdot v26.4s, v25.16b, v2.4b[1]\n"
+ ".inst 0x4fa6e33b // sdot v27.4s, v25.16b, v6.4b[1]\n"
+ ".inst 0x4faae33c // sdot v28.4s, v25.16b, v10.4b[1]\n"
"ins v24.d[1], temploadreg0\n"
- ".word 0x4faee33d // sdot v29.4s, v25.16b, v14.4b[1]\n"
- ".word 0x4fb2e33e // sdot v30.4s, v25.16b, v18.4b[1]\n"
- ".word 0x4fb6e33f // sdot v31.4s, v25.16b, v22.4b[1]\n"
+ ".inst 0x4faee33d // sdot v29.4s, v25.16b, v14.4b[1]\n"
+ ".inst 0x4fb2e33e // sdot v30.4s, v25.16b, v18.4b[1]\n"
+ ".inst 0x4fb6e33f // sdot v31.4s, v25.16b, v22.4b[1]\n"
"ldr d25, [%[b_ptr0], #0x10]\n"
- ".word 0x4f82eb1a // sdot v26.4s, v24.16b, v2.4b[2]\n"
+ ".inst 0x4f82eb1a // sdot v26.4s, v24.16b, v2.4b[2]\n"
"add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f86eb1b // sdot v27.4s, v24.16b, v6.4b[2]\n"
+ ".inst 0x4f86eb1b // sdot v27.4s, v24.16b, v6.4b[2]\n"
"ins v25.d[1], temploadreg1\n"
- ".word 0x4f8aeb1c // sdot v28.4s, v24.16b, v10.4b[2]\n"
+ ".inst 0x4f8aeb1c // sdot v28.4s, v24.16b, v10.4b[2]\n"
"ldr temploadreg0, [%[b_ptr0], #0x8]\n"
- ".word 0x4f8eeb1d // sdot v29.4s, v24.16b, v14.4b[2]\n"
+ ".inst 0x4f8eeb1d // sdot v29.4s, v24.16b, v14.4b[2]\n"
"ldr temploadreg1, [%[b_ptr0], #0x18]\n"
- ".word 0x4f92eb1e // sdot v30.4s, v24.16b, v18.4b[2]\n"
- ".word 0x4f96eb1f // sdot v31.4s, v24.16b, v22.4b[2]\n"
+ ".inst 0x4f92eb1e // sdot v30.4s, v24.16b, v18.4b[2]\n"
+ ".inst 0x4f96eb1f // sdot v31.4s, v24.16b, v22.4b[2]\n"
"ldr d24, [%[b_ptr0]]\n"
- ".word 0x4fa2eb3a // sdot v26.4s, v25.16b, v2.4b[3]\n"
- ".word 0x4fa6eb3b // sdot v27.4s, v25.16b, v6.4b[3]\n"
- ".word 0x4faaeb3c // sdot v28.4s, v25.16b, v10.4b[3]\n"
+ ".inst 0x4fa2eb3a // sdot v26.4s, v25.16b, v2.4b[3]\n"
+ ".inst 0x4fa6eb3b // sdot v27.4s, v25.16b, v6.4b[3]\n"
+ ".inst 0x4faaeb3c // sdot v28.4s, v25.16b, v10.4b[3]\n"
"ins v24.d[1], temploadreg0\n"
- ".word 0x4faeeb3d // sdot v29.4s, v25.16b, v14.4b[3]\n"
- ".word 0x4fb2eb3e // sdot v30.4s, v25.16b, v18.4b[3]\n"
- ".word 0x4fb6eb3f // sdot v31.4s, v25.16b, v22.4b[3]\n"
+ ".inst 0x4faeeb3d // sdot v29.4s, v25.16b, v14.4b[3]\n"
+ ".inst 0x4fb2eb3e // sdot v30.4s, v25.16b, v18.4b[3]\n"
+ ".inst 0x4fb6eb3f // sdot v31.4s, v25.16b, v22.4b[3]\n"
"ldr d25, [%[b_ptr0], #0x10]\n"
- ".word 0x4f83e31a // sdot v26.4s, v24.16b, v3.4b[0]\n"
+ ".inst 0x4f83e31a // sdot v26.4s, v24.16b, v3.4b[0]\n"
"add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f87e31b // sdot v27.4s, v24.16b, v7.4b[0]\n"
+ ".inst 0x4f87e31b // sdot v27.4s, v24.16b, v7.4b[0]\n"
"ins v25.d[1], temploadreg1\n"
- ".word 0x4f8be31c // sdot v28.4s, v24.16b, v11.4b[0]\n"
+ ".inst 0x4f8be31c // sdot v28.4s, v24.16b, v11.4b[0]\n"
"ldr temploadreg0, [%[b_ptr0], #0x8]\n"
- ".word 0x4f8fe31d // sdot v29.4s, v24.16b, v15.4b[0]\n"
- ".word 0x4f93e31e // sdot v30.4s, v24.16b, v19.4b[0]\n"
- ".word 0x4f97e31f // sdot v31.4s, v24.16b, v23.4b[0]\n"
+ ".inst 0x4f8fe31d // sdot v29.4s, v24.16b, v15.4b[0]\n"
+ ".inst 0x4f93e31e // sdot v30.4s, v24.16b, v19.4b[0]\n"
+ ".inst 0x4f97e31f // sdot v31.4s, v24.16b, v23.4b[0]\n"
"ldr d24, [%[b_ptr0]]\n"
- ".word 0x4fa3e33a // sdot v26.4s, v25.16b, v3.4b[1]\n"
+ ".inst 0x4fa3e33a // sdot v26.4s, v25.16b, v3.4b[1]\n"
"add %[b_ptr0], %[b_ptr0], #0x10\n"
- ".word 0x4fa7e33b // sdot v27.4s, v25.16b, v7.4b[1]\n"
+ ".inst 0x4fa7e33b // sdot v27.4s, v25.16b, v7.4b[1]\n"
"ins v24.d[1], temploadreg0\n"
- ".word 0x4fabe33c // sdot v28.4s, v25.16b, v11.4b[1]\n"
+ ".inst 0x4fabe33c // sdot v28.4s, v25.16b, v11.4b[1]\n"
"ldr temploadreg0, [%[b_ptr0], #0x8]\n"
- ".word 0x4fafe33d // sdot v29.4s, v25.16b, v15.4b[1]\n"
+ ".inst 0x4fafe33d // sdot v29.4s, v25.16b, v15.4b[1]\n"
"ldr temploadreg1, [%[b_ptr0], #0x18]\n"
- ".word 0x4fb3e33e // sdot v30.4s, v25.16b, v19.4b[1]\n"
- ".word 0x4fb7e33f // sdot v31.4s, v25.16b, v23.4b[1]\n"
+ ".inst 0x4fb3e33e // sdot v30.4s, v25.16b, v19.4b[1]\n"
+ ".inst 0x4fb7e33f // sdot v31.4s, v25.16b, v23.4b[1]\n"
"ldr d25, [%[b_ptr0], #0x10]\n"
- ".word 0x4f83eb1a // sdot v26.4s, v24.16b, v3.4b[2]\n"
- ".word 0x4f87eb1b // sdot v27.4s, v24.16b, v7.4b[2]\n"
- ".word 0x4f8beb1c // sdot v28.4s, v24.16b, v11.4b[2]\n"
+ ".inst 0x4f83eb1a // sdot v26.4s, v24.16b, v3.4b[2]\n"
+ ".inst 0x4f87eb1b // sdot v27.4s, v24.16b, v7.4b[2]\n"
+ ".inst 0x4f8beb1c // sdot v28.4s, v24.16b, v11.4b[2]\n"
"ins v25.d[1], temploadreg1\n"
- ".word 0x4f8feb1d // sdot v29.4s, v24.16b, v15.4b[2]\n"
- ".word 0x4f93eb1e // sdot v30.4s, v24.16b, v19.4b[2]\n"
- ".word 0x4f97eb1f // sdot v31.4s, v24.16b, v23.4b[2]\n"
+ ".inst 0x4f8feb1d // sdot v29.4s, v24.16b, v15.4b[2]\n"
+ ".inst 0x4f93eb1e // sdot v30.4s, v24.16b, v19.4b[2]\n"
+ ".inst 0x4f97eb1f // sdot v31.4s, v24.16b, v23.4b[2]\n"
"ldr d24, [%[b_ptr0]]\n"
"add %[b_ptr0], %[b_ptr0], #0x20\n"
"ins v24.d[1], temploadreg0\n"
"b.ne 8b\n"
"7:\n"
- "str q26, [%[c_ptr0]], #0x10\n"
+ "str q26, [%[c_ptr0]]\n"
+ "add %[c_ptr0], %[c_ptr0], #0x10\n"
"movi v26.4s, #0\n"
- "str q27, [c_ptr1], #0x10\n"
+ "str q27, [c_ptr1]\n"
+ "add c_ptr1, c_ptr1, #0x10\n"
"movi v27.4s, #0\n"
- ".word 0x4f80e31a // sdot v26.4s, v24.16b, v0.4b[0]\n"
- "str q28, [c_ptr2], #0x10\n"
+ ".inst 0x4f80e31a // sdot v26.4s, v24.16b, v0.4b[0]\n"
+ "str q28, [c_ptr2]\n"
"movi v28.4s, #0\n"
- ".word 0x4f84e31b // sdot v27.4s, v24.16b, v4.4b[0]\n"
- ".word 0x4fa0e33a // sdot v26.4s, v25.16b, v0.4b[1]\n"
- "str q29, [c_ptr3], #0x10\n"
+ "add c_ptr2, c_ptr2, #0x10\n"
+ ".inst 0x4f84e31b // sdot v27.4s, v24.16b, v4.4b[0]\n"
+ "str q29, [c_ptr3]\n"
"movi v29.4s, #0\n"
- ".word 0x4f88e31c // sdot v28.4s, v24.16b, v8.4b[0]\n"
- ".word 0x4fa4e33b // sdot v27.4s, v25.16b, v4.4b[1]\n"
- "str q30, [c_ptr4], #0x10\n"
+ "add c_ptr3, c_ptr3, #0x10\n"
+ ".inst 0x4f88e31c // sdot v28.4s, v24.16b, v8.4b[0]\n"
+ "str q30, [c_ptr4]\n"
"movi v30.4s, #0\n"
- ".word 0x4f8ce31d // sdot v29.4s, v24.16b, v12.4b[0]\n"
- ".word 0x4fa8e33c // sdot v28.4s, v25.16b, v8.4b[1]\n"
- "str q31, [c_ptr5], #0x10\n"
+ "add c_ptr4, c_ptr4, #0x10\n"
+ ".inst 0x4f8ce31d // sdot v29.4s, v24.16b, v12.4b[0]\n"
+ "str q31, [c_ptr5]\n"
"movi v31.4s, #0\n"
- ".word 0x4f90e31e // sdot v30.4s, v24.16b, v16.4b[0]\n"
- ".word 0x4face33d // sdot v29.4s, v25.16b, v12.4b[1]\n"
- ".word 0x4f94e31f // sdot v31.4s, v24.16b, v20.4b[0]\n"
+ "add c_ptr5, c_ptr5, #0x10\n"
+ ".inst 0x4f90e31e // sdot v30.4s, v24.16b, v16.4b[0]\n"
+ ".inst 0x4fa0e33a // sdot v26.4s, v25.16b, v0.4b[1]\n"
+ ".inst 0x4f94e31f // sdot v31.4s, v24.16b, v20.4b[0]\n"
"ldr q24, [%[b_ptr0]]\n"
- ".word 0x4fb0e33e // sdot v30.4s, v25.16b, v16.4b[1]\n"
- ".word 0x4fb4e33f // sdot v31.4s, v25.16b, v20.4b[1]\n"
+ ".inst 0x4fa4e33b // sdot v27.4s, v25.16b, v4.4b[1]\n"
+ ".inst 0x4fa8e33c // sdot v28.4s, v25.16b, v8.4b[1]\n"
+ ".inst 0x4face33d // sdot v29.4s, v25.16b, v12.4b[1]\n"
+ ".inst 0x4fb0e33e // sdot v30.4s, v25.16b, v16.4b[1]\n"
+ ".inst 0x4fb4e33f // sdot v31.4s, v25.16b, v20.4b[1]\n"
"ldr q25, [%[b_ptr0], #0x10]\n"
- ".word 0x4f80eb1a // sdot v26.4s, v24.16b, v0.4b[2]\n"
+ ".inst 0x4f80eb1a // sdot v26.4s, v24.16b, v0.4b[2]\n"
"add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f84eb1b // sdot v27.4s, v24.16b, v4.4b[2]\n"
- ".word 0x4f88eb1c // sdot v28.4s, v24.16b, v8.4b[2]\n"
- ".word 0x4f8ceb1d // sdot v29.4s, v24.16b, v12.4b[2]\n"
- ".word 0x4f90eb1e // sdot v30.4s, v24.16b, v16.4b[2]\n"
- ".word 0x4f94eb1f // sdot v31.4s, v24.16b, v20.4b[2]\n"
+ ".inst 0x4f84eb1b // sdot v27.4s, v24.16b, v4.4b[2]\n"
+ ".inst 0x4f88eb1c // sdot v28.4s, v24.16b, v8.4b[2]\n"
+ ".inst 0x4f8ceb1d // sdot v29.4s, v24.16b, v12.4b[2]\n"
+ ".inst 0x4f90eb1e // sdot v30.4s, v24.16b, v16.4b[2]\n"
+ ".inst 0x4f94eb1f // sdot v31.4s, v24.16b, v20.4b[2]\n"
"ldr q24, [%[b_ptr0]]\n"
- ".word 0x4fa0eb3a // sdot v26.4s, v25.16b, v0.4b[3]\n"
- ".word 0x4fa4eb3b // sdot v27.4s, v25.16b, v4.4b[3]\n"
- ".word 0x4fa8eb3c // sdot v28.4s, v25.16b, v8.4b[3]\n"
- ".word 0x4faceb3d // sdot v29.4s, v25.16b, v12.4b[3]\n"
- ".word 0x4fb0eb3e // sdot v30.4s, v25.16b, v16.4b[3]\n"
- ".word 0x4fb4eb3f // sdot v31.4s, v25.16b, v20.4b[3]\n"
+ ".inst 0x4fa0eb3a // sdot v26.4s, v25.16b, v0.4b[3]\n"
+ ".inst 0x4fa4eb3b // sdot v27.4s, v25.16b, v4.4b[3]\n"
+ ".inst 0x4fa8eb3c // sdot v28.4s, v25.16b, v8.4b[3]\n"
+ ".inst 0x4faceb3d // sdot v29.4s, v25.16b, v12.4b[3]\n"
+ ".inst 0x4fb0eb3e // sdot v30.4s, v25.16b, v16.4b[3]\n"
+ ".inst 0x4fb4eb3f // sdot v31.4s, v25.16b, v20.4b[3]\n"
"ldr q25, [%[b_ptr0], #0x10]\n"
- ".word 0x4f81e31a // sdot v26.4s, v24.16b, v1.4b[0]\n"
+ ".inst 0x4f81e31a // sdot v26.4s, v24.16b, v1.4b[0]\n"
"add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f85e31b // sdot v27.4s, v24.16b, v5.4b[0]\n"
- ".word 0x4f89e31c // sdot v28.4s, v24.16b, v9.4b[0]\n"
- ".word 0x4f8de31d // sdot v29.4s, v24.16b, v13.4b[0]\n"
- ".word 0x4f91e31e // sdot v30.4s, v24.16b, v17.4b[0]\n"
- ".word 0x4f95e31f // sdot v31.4s, v24.16b, v21.4b[0]\n"
+ ".inst 0x4f85e31b // sdot v27.4s, v24.16b, v5.4b[0]\n"
+ ".inst 0x4f89e31c // sdot v28.4s, v24.16b, v9.4b[0]\n"
+ ".inst 0x4f8de31d // sdot v29.4s, v24.16b, v13.4b[0]\n"
+ ".inst 0x4f91e31e // sdot v30.4s, v24.16b, v17.4b[0]\n"
+ ".inst 0x4f95e31f // sdot v31.4s, v24.16b, v21.4b[0]\n"
"ldr q24, [%[b_ptr0]]\n"
- ".word 0x4fa1e33a // sdot v26.4s, v25.16b, v1.4b[1]\n"
- ".word 0x4fa5e33b // sdot v27.4s, v25.16b, v5.4b[1]\n"
- ".word 0x4fa9e33c // sdot v28.4s, v25.16b, v9.4b[1]\n"
- ".word 0x4fade33d // sdot v29.4s, v25.16b, v13.4b[1]\n"
- ".word 0x4fb1e33e // sdot v30.4s, v25.16b, v17.4b[1]\n"
- ".word 0x4fb5e33f // sdot v31.4s, v25.16b, v21.4b[1]\n"
+ ".inst 0x4fa1e33a // sdot v26.4s, v25.16b, v1.4b[1]\n"
+ ".inst 0x4fa5e33b // sdot v27.4s, v25.16b, v5.4b[1]\n"
+ ".inst 0x4fa9e33c // sdot v28.4s, v25.16b, v9.4b[1]\n"
+ ".inst 0x4fade33d // sdot v29.4s, v25.16b, v13.4b[1]\n"
+ ".inst 0x4fb1e33e // sdot v30.4s, v25.16b, v17.4b[1]\n"
+ ".inst 0x4fb5e33f // sdot v31.4s, v25.16b, v21.4b[1]\n"
"ldr q25, [%[b_ptr0], #0x10]\n"
- ".word 0x4f81eb1a // sdot v26.4s, v24.16b, v1.4b[2]\n"
+ ".inst 0x4f81eb1a // sdot v26.4s, v24.16b, v1.4b[2]\n"
"add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f85eb1b // sdot v27.4s, v24.16b, v5.4b[2]\n"
- ".word 0x4f89eb1c // sdot v28.4s, v24.16b, v9.4b[2]\n"
- ".word 0x4f8deb1d // sdot v29.4s, v24.16b, v13.4b[2]\n"
- ".word 0x4f91eb1e // sdot v30.4s, v24.16b, v17.4b[2]\n"
- ".word 0x4f95eb1f // sdot v31.4s, v24.16b, v21.4b[2]\n"
+ ".inst 0x4f85eb1b // sdot v27.4s, v24.16b, v5.4b[2]\n"
+ ".inst 0x4f89eb1c // sdot v28.4s, v24.16b, v9.4b[2]\n"
+ ".inst 0x4f8deb1d // sdot v29.4s, v24.16b, v13.4b[2]\n"
+ ".inst 0x4f91eb1e // sdot v30.4s, v24.16b, v17.4b[2]\n"
+ ".inst 0x4f95eb1f // sdot v31.4s, v24.16b, v21.4b[2]\n"
"ldr q24, [%[b_ptr0]]\n"
- ".word 0x4fa1eb3a // sdot v26.4s, v25.16b, v1.4b[3]\n"
- ".word 0x4fa5eb3b // sdot v27.4s, v25.16b, v5.4b[3]\n"
- ".word 0x4fa9eb3c // sdot v28.4s, v25.16b, v9.4b[3]\n"
- ".word 0x4fadeb3d // sdot v29.4s, v25.16b, v13.4b[3]\n"
- ".word 0x4fb1eb3e // sdot v30.4s, v25.16b, v17.4b[3]\n"
- ".word 0x4fb5eb3f // sdot v31.4s, v25.16b, v21.4b[3]\n"
+ ".inst 0x4fa1eb3a // sdot v26.4s, v25.16b, v1.4b[3]\n"
+ ".inst 0x4fa5eb3b // sdot v27.4s, v25.16b, v5.4b[3]\n"
+ ".inst 0x4fa9eb3c // sdot v28.4s, v25.16b, v9.4b[3]\n"
+ ".inst 0x4fadeb3d // sdot v29.4s, v25.16b, v13.4b[3]\n"
+ ".inst 0x4fb1eb3e // sdot v30.4s, v25.16b, v17.4b[3]\n"
+ ".inst 0x4fb5eb3f // sdot v31.4s, v25.16b, v21.4b[3]\n"
"ldr q25, [%[b_ptr0], #0x10]\n"
- ".word 0x4f82e31a // sdot v26.4s, v24.16b, v2.4b[0]\n"
+ ".inst 0x4f82e31a // sdot v26.4s, v24.16b, v2.4b[0]\n"
"add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f86e31b // sdot v27.4s, v24.16b, v6.4b[0]\n"
- ".word 0x4f8ae31c // sdot v28.4s, v24.16b, v10.4b[0]\n"
- ".word 0x4f8ee31d // sdot v29.4s, v24.16b, v14.4b[0]\n"
- ".word 0x4f92e31e // sdot v30.4s, v24.16b, v18.4b[0]\n"
- ".word 0x4f96e31f // sdot v31.4s, v24.16b, v22.4b[0]\n"
+ ".inst 0x4f86e31b // sdot v27.4s, v24.16b, v6.4b[0]\n"
+ ".inst 0x4f8ae31c // sdot v28.4s, v24.16b, v10.4b[0]\n"
+ ".inst 0x4f8ee31d // sdot v29.4s, v24.16b, v14.4b[0]\n"
+ ".inst 0x4f92e31e // sdot v30.4s, v24.16b, v18.4b[0]\n"
+ ".inst 0x4f96e31f // sdot v31.4s, v24.16b, v22.4b[0]\n"
"ldr q24, [%[b_ptr0]]\n"
- ".word 0x4fa2e33a // sdot v26.4s, v25.16b, v2.4b[1]\n"
- ".word 0x4fa6e33b // sdot v27.4s, v25.16b, v6.4b[1]\n"
- ".word 0x4faae33c // sdot v28.4s, v25.16b, v10.4b[1]\n"
- ".word 0x4faee33d // sdot v29.4s, v25.16b, v14.4b[1]\n"
- ".word 0x4fb2e33e // sdot v30.4s, v25.16b, v18.4b[1]\n"
- ".word 0x4fb6e33f // sdot v31.4s, v25.16b, v22.4b[1]\n"
+ ".inst 0x4fa2e33a // sdot v26.4s, v25.16b, v2.4b[1]\n"
+ ".inst 0x4fa6e33b // sdot v27.4s, v25.16b, v6.4b[1]\n"
+ ".inst 0x4faae33c // sdot v28.4s, v25.16b, v10.4b[1]\n"
+ ".inst 0x4faee33d // sdot v29.4s, v25.16b, v14.4b[1]\n"
+ ".inst 0x4fb2e33e // sdot v30.4s, v25.16b, v18.4b[1]\n"
+ ".inst 0x4fb6e33f // sdot v31.4s, v25.16b, v22.4b[1]\n"
"ldr q25, [%[b_ptr0], #0x10]\n"
- ".word 0x4f82eb1a // sdot v26.4s, v24.16b, v2.4b[2]\n"
+ ".inst 0x4f82eb1a // sdot v26.4s, v24.16b, v2.4b[2]\n"
"add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f86eb1b // sdot v27.4s, v24.16b, v6.4b[2]\n"
- ".word 0x4f8aeb1c // sdot v28.4s, v24.16b, v10.4b[2]\n"
- ".word 0x4f8eeb1d // sdot v29.4s, v24.16b, v14.4b[2]\n"
- ".word 0x4f92eb1e // sdot v30.4s, v24.16b, v18.4b[2]\n"
- ".word 0x4f96eb1f // sdot v31.4s, v24.16b, v22.4b[2]\n"
+ ".inst 0x4f86eb1b // sdot v27.4s, v24.16b, v6.4b[2]\n"
+ ".inst 0x4f8aeb1c // sdot v28.4s, v24.16b, v10.4b[2]\n"
+ ".inst 0x4f8eeb1d // sdot v29.4s, v24.16b, v14.4b[2]\n"
+ ".inst 0x4f92eb1e // sdot v30.4s, v24.16b, v18.4b[2]\n"
+ ".inst 0x4f96eb1f // sdot v31.4s, v24.16b, v22.4b[2]\n"
"ldr q24, [%[b_ptr0]]\n"
- ".word 0x4fa2eb3a // sdot v26.4s, v25.16b, v2.4b[3]\n"
- ".word 0x4fa6eb3b // sdot v27.4s, v25.16b, v6.4b[3]\n"
- ".word 0x4faaeb3c // sdot v28.4s, v25.16b, v10.4b[3]\n"
- ".word 0x4faeeb3d // sdot v29.4s, v25.16b, v14.4b[3]\n"
- ".word 0x4fb2eb3e // sdot v30.4s, v25.16b, v18.4b[3]\n"
- ".word 0x4fb6eb3f // sdot v31.4s, v25.16b, v22.4b[3]\n"
+ ".inst 0x4fa2eb3a // sdot v26.4s, v25.16b, v2.4b[3]\n"
+ ".inst 0x4fa6eb3b // sdot v27.4s, v25.16b, v6.4b[3]\n"
+ ".inst 0x4faaeb3c // sdot v28.4s, v25.16b, v10.4b[3]\n"
+ ".inst 0x4faeeb3d // sdot v29.4s, v25.16b, v14.4b[3]\n"
+ ".inst 0x4fb2eb3e // sdot v30.4s, v25.16b, v18.4b[3]\n"
+ ".inst 0x4fb6eb3f // sdot v31.4s, v25.16b, v22.4b[3]\n"
"ldr q25, [%[b_ptr0], #0x10]\n"
- ".word 0x4f83e31a // sdot v26.4s, v24.16b, v3.4b[0]\n"
+ ".inst 0x4f83e31a // sdot v26.4s, v24.16b, v3.4b[0]\n"
"add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f87e31b // sdot v27.4s, v24.16b, v7.4b[0]\n"
- ".word 0x4f8be31c // sdot v28.4s, v24.16b, v11.4b[0]\n"
- ".word 0x4f8fe31d // sdot v29.4s, v24.16b, v15.4b[0]\n"
- ".word 0x4f93e31e // sdot v30.4s, v24.16b, v19.4b[0]\n"
- ".word 0x4f97e31f // sdot v31.4s, v24.16b, v23.4b[0]\n"
+ ".inst 0x4f87e31b // sdot v27.4s, v24.16b, v7.4b[0]\n"
+ ".inst 0x4f8be31c // sdot v28.4s, v24.16b, v11.4b[0]\n"
+ ".inst 0x4f8fe31d // sdot v29.4s, v24.16b, v15.4b[0]\n"
+ ".inst 0x4f93e31e // sdot v30.4s, v24.16b, v19.4b[0]\n"
+ ".inst 0x4f97e31f // sdot v31.4s, v24.16b, v23.4b[0]\n"
"ldr q24, [%[b_ptr0]]\n"
- ".word 0x4fa3e33a // sdot v26.4s, v25.16b, v3.4b[1]\n"
+ ".inst 0x4fa3e33a // sdot v26.4s, v25.16b, v3.4b[1]\n"
"add %[b_ptr0], %[b_ptr0], #0x10\n"
- ".word 0x4fa7e33b // sdot v27.4s, v25.16b, v7.4b[1]\n"
- ".word 0x4fabe33c // sdot v28.4s, v25.16b, v11.4b[1]\n"
- ".word 0x4fafe33d // sdot v29.4s, v25.16b, v15.4b[1]\n"
- ".word 0x4fb3e33e // sdot v30.4s, v25.16b, v19.4b[1]\n"
- ".word 0x4fb7e33f // sdot v31.4s, v25.16b, v23.4b[1]\n"
- ".word 0x4f83eb1a // sdot v26.4s, v24.16b, v3.4b[2]\n"
- ".word 0x4f87eb1b // sdot v27.4s, v24.16b, v7.4b[2]\n"
- ".word 0x4f8beb1c // sdot v28.4s, v24.16b, v11.4b[2]\n"
- ".word 0x4f8feb1d // sdot v29.4s, v24.16b, v15.4b[2]\n"
- ".word 0x4f93eb1e // sdot v30.4s, v24.16b, v19.4b[2]\n"
- ".word 0x4f97eb1f // sdot v31.4s, v24.16b, v23.4b[2]\n"
+ ".inst 0x4fa7e33b // sdot v27.4s, v25.16b, v7.4b[1]\n"
+ ".inst 0x4fabe33c // sdot v28.4s, v25.16b, v11.4b[1]\n"
+ ".inst 0x4fafe33d // sdot v29.4s, v25.16b, v15.4b[1]\n"
+ ".inst 0x4fb3e33e // sdot v30.4s, v25.16b, v19.4b[1]\n"
+ ".inst 0x4fb7e33f // sdot v31.4s, v25.16b, v23.4b[1]\n"
+ ".inst 0x4f83eb1a // sdot v26.4s, v24.16b, v3.4b[2]\n"
+ ".inst 0x4f87eb1b // sdot v27.4s, v24.16b, v7.4b[2]\n"
+ ".inst 0x4f8beb1c // sdot v28.4s, v24.16b, v11.4b[2]\n"
+ ".inst 0x4f8feb1d // sdot v29.4s, v24.16b, v15.4b[2]\n"
+ ".inst 0x4f93eb1e // sdot v30.4s, v24.16b, v19.4b[2]\n"
+ ".inst 0x4f97eb1f // sdot v31.4s, v24.16b, v23.4b[2]\n"
"6:\n"
"str q26, [%[c_ptr0]]\n"
+ "add %[c_ptr0], %[c_ptr0], #0x10\n"
"str q27, [c_ptr1]\n"
+ "add c_ptr1, c_ptr1, #0x10\n"
"str q28, [c_ptr2]\n"
+ "add c_ptr2, c_ptr2, #0x10\n"
"str q29, [c_ptr3]\n"
+ "add c_ptr3, c_ptr3, #0x10\n"
"str q30, [c_ptr4]\n"
+ "add c_ptr4, c_ptr4, #0x10\n"
"str q31, [c_ptr5]\n"
+ "add c_ptr5, c_ptr5, #0x10\n"
".unreq a_ptr1\n"
".unreq a_ptr2\n"
".unreq a_ptr3\n"
@@ -3511,126 +3644,126 @@ void a64_smallK_hybrid_s8s32_dot_4x6_a55(const int8_t *A, int lda, const int8_t
"prfm PLDL1KEEP, [a_ptr5, #0xc0]\n"
"movi v31.4s, #0\n"
"prfm PLDL1KEEP, [a_ptr5, #0x100]\n"
- ".word 0x4f80e31a // sdot v26.4s, v24.16b, v0.4b[0]\n"
+ ".inst 0x4f80e31a // sdot v26.4s, v24.16b, v0.4b[0]\n"
"prfm PLDL1KEEP, [a_ptr5, #0x140]\n"
- ".word 0x4f84e31b // sdot v27.4s, v24.16b, v4.4b[0]\n"
+ ".inst 0x4f84e31b // sdot v27.4s, v24.16b, v4.4b[0]\n"
"prfm PLDL1KEEP, [a_ptr5, #0x180]\n"
- ".word 0x4f88e31c // sdot v28.4s, v24.16b, v8.4b[0]\n"
+ ".inst 0x4f88e31c // sdot v28.4s, v24.16b, v8.4b[0]\n"
"add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f8ce31d // sdot v29.4s, v24.16b, v12.4b[0]\n"
- ".word 0x4f90e31e // sdot v30.4s, v24.16b, v16.4b[0]\n"
- ".word 0x4f94e31f // sdot v31.4s, v24.16b, v20.4b[0]\n"
+ ".inst 0x4f8ce31d // sdot v29.4s, v24.16b, v12.4b[0]\n"
+ ".inst 0x4f90e31e // sdot v30.4s, v24.16b, v16.4b[0]\n"
+ ".inst 0x4f94e31f // sdot v31.4s, v24.16b, v20.4b[0]\n"
"ldr q24, [%[b_ptr0]]\n"
- ".word 0x4fa0e33a // sdot v26.4s, v25.16b, v0.4b[1]\n"
- ".word 0x4fa4e33b // sdot v27.4s, v25.16b, v4.4b[1]\n"
- ".word 0x4fa8e33c // sdot v28.4s, v25.16b, v8.4b[1]\n"
- ".word 0x4face33d // sdot v29.4s, v25.16b, v12.4b[1]\n"
- ".word 0x4fb0e33e // sdot v30.4s, v25.16b, v16.4b[1]\n"
- ".word 0x4fb4e33f // sdot v31.4s, v25.16b, v20.4b[1]\n"
+ ".inst 0x4fa0e33a // sdot v26.4s, v25.16b, v0.4b[1]\n"
+ ".inst 0x4fa4e33b // sdot v27.4s, v25.16b, v4.4b[1]\n"
+ ".inst 0x4fa8e33c // sdot v28.4s, v25.16b, v8.4b[1]\n"
+ ".inst 0x4face33d // sdot v29.4s, v25.16b, v12.4b[1]\n"
+ ".inst 0x4fb0e33e // sdot v30.4s, v25.16b, v16.4b[1]\n"
+ ".inst 0x4fb4e33f // sdot v31.4s, v25.16b, v20.4b[1]\n"
"ldr q25, [%[b_ptr0], #0x10]\n"
- ".word 0x4f80eb1a // sdot v26.4s, v24.16b, v0.4b[2]\n"
+ ".inst 0x4f80eb1a // sdot v26.4s, v24.16b, v0.4b[2]\n"
"add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f84eb1b // sdot v27.4s, v24.16b, v4.4b[2]\n"
- ".word 0x4f88eb1c // sdot v28.4s, v24.16b, v8.4b[2]\n"
- ".word 0x4f8ceb1d // sdot v29.4s, v24.16b, v12.4b[2]\n"
- ".word 0x4f90eb1e // sdot v30.4s, v24.16b, v16.4b[2]\n"
- ".word 0x4f94eb1f // sdot v31.4s, v24.16b, v20.4b[2]\n"
+ ".inst 0x4f84eb1b // sdot v27.4s, v24.16b, v4.4b[2]\n"
+ ".inst 0x4f88eb1c // sdot v28.4s, v24.16b, v8.4b[2]\n"
+ ".inst 0x4f8ceb1d // sdot v29.4s, v24.16b, v12.4b[2]\n"
+ ".inst 0x4f90eb1e // sdot v30.4s, v24.16b, v16.4b[2]\n"
+ ".inst 0x4f94eb1f // sdot v31.4s, v24.16b, v20.4b[2]\n"
"ldr q24, [%[b_ptr0]]\n"
- ".word 0x4fa0eb3a // sdot v26.4s, v25.16b, v0.4b[3]\n"
- ".word 0x4fa4eb3b // sdot v27.4s, v25.16b, v4.4b[3]\n"
- ".word 0x4fa8eb3c // sdot v28.4s, v25.16b, v8.4b[3]\n"
- ".word 0x4faceb3d // sdot v29.4s, v25.16b, v12.4b[3]\n"
- ".word 0x4fb0eb3e // sdot v30.4s, v25.16b, v16.4b[3]\n"
- ".word 0x4fb4eb3f // sdot v31.4s, v25.16b, v20.4b[3]\n"
+ ".inst 0x4fa0eb3a // sdot v26.4s, v25.16b, v0.4b[3]\n"
+ ".inst 0x4fa4eb3b // sdot v27.4s, v25.16b, v4.4b[3]\n"
+ ".inst 0x4fa8eb3c // sdot v28.4s, v25.16b, v8.4b[3]\n"
+ ".inst 0x4faceb3d // sdot v29.4s, v25.16b, v12.4b[3]\n"
+ ".inst 0x4fb0eb3e // sdot v30.4s, v25.16b, v16.4b[3]\n"
+ ".inst 0x4fb4eb3f // sdot v31.4s, v25.16b, v20.4b[3]\n"
"ldr q25, [%[b_ptr0], #0x10]\n"
- ".word 0x4f81e31a // sdot v26.4s, v24.16b, v1.4b[0]\n"
+ ".inst 0x4f81e31a // sdot v26.4s, v24.16b, v1.4b[0]\n"
"add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f85e31b // sdot v27.4s, v24.16b, v5.4b[0]\n"
- ".word 0x4f89e31c // sdot v28.4s, v24.16b, v9.4b[0]\n"
- ".word 0x4f8de31d // sdot v29.4s, v24.16b, v13.4b[0]\n"
- ".word 0x4f91e31e // sdot v30.4s, v24.16b, v17.4b[0]\n"
- ".word 0x4f95e31f // sdot v31.4s, v24.16b, v21.4b[0]\n"
+ ".inst 0x4f85e31b // sdot v27.4s, v24.16b, v5.4b[0]\n"
+ ".inst 0x4f89e31c // sdot v28.4s, v24.16b, v9.4b[0]\n"
+ ".inst 0x4f8de31d // sdot v29.4s, v24.16b, v13.4b[0]\n"
+ ".inst 0x4f91e31e // sdot v30.4s, v24.16b, v17.4b[0]\n"
+ ".inst 0x4f95e31f // sdot v31.4s, v24.16b, v21.4b[0]\n"
"ldr q24, [%[b_ptr0]]\n"
- ".word 0x4fa1e33a // sdot v26.4s, v25.16b, v1.4b[1]\n"
- ".word 0x4fa5e33b // sdot v27.4s, v25.16b, v5.4b[1]\n"
- ".word 0x4fa9e33c // sdot v28.4s, v25.16b, v9.4b[1]\n"
- ".word 0x4fade33d // sdot v29.4s, v25.16b, v13.4b[1]\n"
- ".word 0x4fb1e33e // sdot v30.4s, v25.16b, v17.4b[1]\n"
- ".word 0x4fb5e33f // sdot v31.4s, v25.16b, v21.4b[1]\n"
+ ".inst 0x4fa1e33a // sdot v26.4s, v25.16b, v1.4b[1]\n"
+ ".inst 0x4fa5e33b // sdot v27.4s, v25.16b, v5.4b[1]\n"
+ ".inst 0x4fa9e33c // sdot v28.4s, v25.16b, v9.4b[1]\n"
+ ".inst 0x4fade33d // sdot v29.4s, v25.16b, v13.4b[1]\n"
+ ".inst 0x4fb1e33e // sdot v30.4s, v25.16b, v17.4b[1]\n"
+ ".inst 0x4fb5e33f // sdot v31.4s, v25.16b, v21.4b[1]\n"
"ldr q25, [%[b_ptr0], #0x10]\n"
- ".word 0x4f81eb1a // sdot v26.4s, v24.16b, v1.4b[2]\n"
+ ".inst 0x4f81eb1a // sdot v26.4s, v24.16b, v1.4b[2]\n"
"add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f85eb1b // sdot v27.4s, v24.16b, v5.4b[2]\n"
- ".word 0x4f89eb1c // sdot v28.4s, v24.16b, v9.4b[2]\n"
- ".word 0x4f8deb1d // sdot v29.4s, v24.16b, v13.4b[2]\n"
- ".word 0x4f91eb1e // sdot v30.4s, v24.16b, v17.4b[2]\n"
- ".word 0x4f95eb1f // sdot v31.4s, v24.16b, v21.4b[2]\n"
+ ".inst 0x4f85eb1b // sdot v27.4s, v24.16b, v5.4b[2]\n"
+ ".inst 0x4f89eb1c // sdot v28.4s, v24.16b, v9.4b[2]\n"
+ ".inst 0x4f8deb1d // sdot v29.4s, v24.16b, v13.4b[2]\n"
+ ".inst 0x4f91eb1e // sdot v30.4s, v24.16b, v17.4b[2]\n"
+ ".inst 0x4f95eb1f // sdot v31.4s, v24.16b, v21.4b[2]\n"
"ldr q24, [%[b_ptr0]]\n"
- ".word 0x4fa1eb3a // sdot v26.4s, v25.16b, v1.4b[3]\n"
- ".word 0x4fa5eb3b // sdot v27.4s, v25.16b, v5.4b[3]\n"
- ".word 0x4fa9eb3c // sdot v28.4s, v25.16b, v9.4b[3]\n"
- ".word 0x4fadeb3d // sdot v29.4s, v25.16b, v13.4b[3]\n"
- ".word 0x4fb1eb3e // sdot v30.4s, v25.16b, v17.4b[3]\n"
- ".word 0x4fb5eb3f // sdot v31.4s, v25.16b, v21.4b[3]\n"
+ ".inst 0x4fa1eb3a // sdot v26.4s, v25.16b, v1.4b[3]\n"
+ ".inst 0x4fa5eb3b // sdot v27.4s, v25.16b, v5.4b[3]\n"
+ ".inst 0x4fa9eb3c // sdot v28.4s, v25.16b, v9.4b[3]\n"
+ ".inst 0x4fadeb3d // sdot v29.4s, v25.16b, v13.4b[3]\n"
+ ".inst 0x4fb1eb3e // sdot v30.4s, v25.16b, v17.4b[3]\n"
+ ".inst 0x4fb5eb3f // sdot v31.4s, v25.16b, v21.4b[3]\n"
"ldr q25, [%[b_ptr0], #0x10]\n"
- ".word 0x4f82e31a // sdot v26.4s, v24.16b, v2.4b[0]\n"
+ ".inst 0x4f82e31a // sdot v26.4s, v24.16b, v2.4b[0]\n"
"add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f86e31b // sdot v27.4s, v24.16b, v6.4b[0]\n"
- ".word 0x4f8ae31c // sdot v28.4s, v24.16b, v10.4b[0]\n"
- ".word 0x4f8ee31d // sdot v29.4s, v24.16b, v14.4b[0]\n"
- ".word 0x4f92e31e // sdot v30.4s, v24.16b, v18.4b[0]\n"
- ".word 0x4f96e31f // sdot v31.4s, v24.16b, v22.4b[0]\n"
+ ".inst 0x4f86e31b // sdot v27.4s, v24.16b, v6.4b[0]\n"
+ ".inst 0x4f8ae31c // sdot v28.4s, v24.16b, v10.4b[0]\n"
+ ".inst 0x4f8ee31d // sdot v29.4s, v24.16b, v14.4b[0]\n"
+ ".inst 0x4f92e31e // sdot v30.4s, v24.16b, v18.4b[0]\n"
+ ".inst 0x4f96e31f // sdot v31.4s, v24.16b, v22.4b[0]\n"
"ldr q24, [%[b_ptr0]]\n"
- ".word 0x4fa2e33a // sdot v26.4s, v25.16b, v2.4b[1]\n"
- ".word 0x4fa6e33b // sdot v27.4s, v25.16b, v6.4b[1]\n"
- ".word 0x4faae33c // sdot v28.4s, v25.16b, v10.4b[1]\n"
- ".word 0x4faee33d // sdot v29.4s, v25.16b, v14.4b[1]\n"
- ".word 0x4fb2e33e // sdot v30.4s, v25.16b, v18.4b[1]\n"
- ".word 0x4fb6e33f // sdot v31.4s, v25.16b, v22.4b[1]\n"
+ ".inst 0x4fa2e33a // sdot v26.4s, v25.16b, v2.4b[1]\n"
+ ".inst 0x4fa6e33b // sdot v27.4s, v25.16b, v6.4b[1]\n"
+ ".inst 0x4faae33c // sdot v28.4s, v25.16b, v10.4b[1]\n"
+ ".inst 0x4faee33d // sdot v29.4s, v25.16b, v14.4b[1]\n"
+ ".inst 0x4fb2e33e // sdot v30.4s, v25.16b, v18.4b[1]\n"
+ ".inst 0x4fb6e33f // sdot v31.4s, v25.16b, v22.4b[1]\n"
"ldr q25, [%[b_ptr0], #0x10]\n"
- ".word 0x4f82eb1a // sdot v26.4s, v24.16b, v2.4b[2]\n"
+ ".inst 0x4f82eb1a // sdot v26.4s, v24.16b, v2.4b[2]\n"
"add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f86eb1b // sdot v27.4s, v24.16b, v6.4b[2]\n"
- ".word 0x4f8aeb1c // sdot v28.4s, v24.16b, v10.4b[2]\n"
- ".word 0x4f8eeb1d // sdot v29.4s, v24.16b, v14.4b[2]\n"
- ".word 0x4f92eb1e // sdot v30.4s, v24.16b, v18.4b[2]\n"
- ".word 0x4f96eb1f // sdot v31.4s, v24.16b, v22.4b[2]\n"
+ ".inst 0x4f86eb1b // sdot v27.4s, v24.16b, v6.4b[2]\n"
+ ".inst 0x4f8aeb1c // sdot v28.4s, v24.16b, v10.4b[2]\n"
+ ".inst 0x4f8eeb1d // sdot v29.4s, v24.16b, v14.4b[2]\n"
+ ".inst 0x4f92eb1e // sdot v30.4s, v24.16b, v18.4b[2]\n"
+ ".inst 0x4f96eb1f // sdot v31.4s, v24.16b, v22.4b[2]\n"
"ldr q24, [%[b_ptr0]]\n"
- ".word 0x4fa2eb3a // sdot v26.4s, v25.16b, v2.4b[3]\n"
- ".word 0x4fa6eb3b // sdot v27.4s, v25.16b, v6.4b[3]\n"
- ".word 0x4faaeb3c // sdot v28.4s, v25.16b, v10.4b[3]\n"
- ".word 0x4faeeb3d // sdot v29.4s, v25.16b, v14.4b[3]\n"
- ".word 0x4fb2eb3e // sdot v30.4s, v25.16b, v18.4b[3]\n"
- ".word 0x4fb6eb3f // sdot v31.4s, v25.16b, v22.4b[3]\n"
+ ".inst 0x4fa2eb3a // sdot v26.4s, v25.16b, v2.4b[3]\n"
+ ".inst 0x4fa6eb3b // sdot v27.4s, v25.16b, v6.4b[3]\n"
+ ".inst 0x4faaeb3c // sdot v28.4s, v25.16b, v10.4b[3]\n"
+ ".inst 0x4faeeb3d // sdot v29.4s, v25.16b, v14.4b[3]\n"
+ ".inst 0x4fb2eb3e // sdot v30.4s, v25.16b, v18.4b[3]\n"
+ ".inst 0x4fb6eb3f // sdot v31.4s, v25.16b, v22.4b[3]\n"
"ldr q25, [%[b_ptr0], #0x10]\n"
- ".word 0x4f83e31a // sdot v26.4s, v24.16b, v3.4b[0]\n"
+ ".inst 0x4f83e31a // sdot v26.4s, v24.16b, v3.4b[0]\n"
"add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f87e31b // sdot v27.4s, v24.16b, v7.4b[0]\n"
- ".word 0x4f8be31c // sdot v28.4s, v24.16b, v11.4b[0]\n"
- ".word 0x4f8fe31d // sdot v29.4s, v24.16b, v15.4b[0]\n"
- ".word 0x4f93e31e // sdot v30.4s, v24.16b, v19.4b[0]\n"
- ".word 0x4f97e31f // sdot v31.4s, v24.16b, v23.4b[0]\n"
+ ".inst 0x4f87e31b // sdot v27.4s, v24.16b, v7.4b[0]\n"
+ ".inst 0x4f8be31c // sdot v28.4s, v24.16b, v11.4b[0]\n"
+ ".inst 0x4f8fe31d // sdot v29.4s, v24.16b, v15.4b[0]\n"
+ ".inst 0x4f93e31e // sdot v30.4s, v24.16b, v19.4b[0]\n"
+ ".inst 0x4f97e31f // sdot v31.4s, v24.16b, v23.4b[0]\n"
"ldr q24, [%[b_ptr0]]\n"
- ".word 0x4fa3e33a // sdot v26.4s, v25.16b, v3.4b[1]\n"
- ".word 0x4fa7e33b // sdot v27.4s, v25.16b, v7.4b[1]\n"
- ".word 0x4fabe33c // sdot v28.4s, v25.16b, v11.4b[1]\n"
- ".word 0x4fafe33d // sdot v29.4s, v25.16b, v15.4b[1]\n"
- ".word 0x4fb3e33e // sdot v30.4s, v25.16b, v19.4b[1]\n"
- ".word 0x4fb7e33f // sdot v31.4s, v25.16b, v23.4b[1]\n"
+ ".inst 0x4fa3e33a // sdot v26.4s, v25.16b, v3.4b[1]\n"
+ ".inst 0x4fa7e33b // sdot v27.4s, v25.16b, v7.4b[1]\n"
+ ".inst 0x4fabe33c // sdot v28.4s, v25.16b, v11.4b[1]\n"
+ ".inst 0x4fafe33d // sdot v29.4s, v25.16b, v15.4b[1]\n"
+ ".inst 0x4fb3e33e // sdot v30.4s, v25.16b, v19.4b[1]\n"
+ ".inst 0x4fb7e33f // sdot v31.4s, v25.16b, v23.4b[1]\n"
"ldr q25, [%[b_ptr0], #0x10]\n"
- ".word 0x4f83eb1a // sdot v26.4s, v24.16b, v3.4b[2]\n"
- "add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f87eb1b // sdot v27.4s, v24.16b, v7.4b[2]\n"
- ".word 0x4f8beb1c // sdot v28.4s, v24.16b, v11.4b[2]\n"
- ".word 0x4f8feb1d // sdot v29.4s, v24.16b, v15.4b[2]\n"
- ".word 0x4f93eb1e // sdot v30.4s, v24.16b, v19.4b[2]\n"
- ".word 0x4f97eb1f // sdot v31.4s, v24.16b, v23.4b[2]\n"
- ".word 0x4fa3eb3a // sdot v26.4s, v25.16b, v3.4b[3]\n"
- ".word 0x4fa7eb3b // sdot v27.4s, v25.16b, v7.4b[3]\n"
- ".word 0x4fabeb3c // sdot v28.4s, v25.16b, v11.4b[3]\n"
- ".word 0x4fafeb3d // sdot v29.4s, v25.16b, v15.4b[3]\n"
- ".word 0x4fb3eb3e // sdot v30.4s, v25.16b, v19.4b[3]\n"
- ".word 0x4fb7eb3f // sdot v31.4s, v25.16b, v23.4b[3]\n"
+ ".inst 0x4f83eb1a // sdot v26.4s, v24.16b, v3.4b[2]\n"
+ "add %[b_ptr0], %[b_ptr0], #0x20\n"
+ ".inst 0x4f87eb1b // sdot v27.4s, v24.16b, v7.4b[2]\n"
+ ".inst 0x4f8beb1c // sdot v28.4s, v24.16b, v11.4b[2]\n"
+ ".inst 0x4f8feb1d // sdot v29.4s, v24.16b, v15.4b[2]\n"
+ ".inst 0x4f93eb1e // sdot v30.4s, v24.16b, v19.4b[2]\n"
+ ".inst 0x4f97eb1f // sdot v31.4s, v24.16b, v23.4b[2]\n"
+ ".inst 0x4fa3eb3a // sdot v26.4s, v25.16b, v3.4b[3]\n"
+ ".inst 0x4fa7eb3b // sdot v27.4s, v25.16b, v7.4b[3]\n"
+ ".inst 0x4fabeb3c // sdot v28.4s, v25.16b, v11.4b[3]\n"
+ ".inst 0x4fafeb3d // sdot v29.4s, v25.16b, v15.4b[3]\n"
+ ".inst 0x4fb3eb3e // sdot v30.4s, v25.16b, v19.4b[3]\n"
+ ".inst 0x4fb7eb3f // sdot v31.4s, v25.16b, v23.4b[3]\n"
"cbz %[loops], 6f\n"
"ldr d24, [%[b_ptr0]]\n"
"subs %[loops], %[loops], #0x1\n"
@@ -3641,316 +3774,334 @@ void a64_smallK_hybrid_s8s32_dot_4x6_a55(const int8_t *A, int lda, const int8_t
"ins v24.d[1], temploadreg0\n"
"b.eq 7f\n"
"8:\n"
- "str q26, [%[c_ptr0]], #0x10\n"
+ "str q26, [%[c_ptr0]]\n"
"subs %[loops], %[loops], #0x1\n"
"movi v26.4s, #0\n"
"ins v25.d[1], temploadreg1\n"
"ldr temploadreg0, [%[b_ptr0], #0x8]\n"
- "prfm PSTL1KEEP, [%[c_ptr0], #0x40]\n"
- "str q27, [c_ptr1], #0x10\n"
+ "add %[c_ptr0], %[c_ptr0], #0x10\n"
+ "str q27, [c_ptr1]\n"
+ "add c_ptr1, c_ptr1, #0x10\n"
"movi v27.4s, #0\n"
"ldr temploadreg1, [%[b_ptr0], #0x18]\n"
- ".word 0x4f80e31a // sdot v26.4s, v24.16b, v0.4b[0]\n"
- "str q28, [c_ptr2], #0x10\n"
- "prfm PSTL1KEEP, [c_ptr1, #0x40]\n"
+ ".inst 0x4f80e31a // sdot v26.4s, v24.16b, v0.4b[0]\n"
+ "str q28, [c_ptr2]\n"
"movi v28.4s, #0\n"
- ".word 0x4f84e31b // sdot v27.4s, v24.16b, v4.4b[0]\n"
- "str q29, [c_ptr3], #0x10\n"
+ "prfm PSTL1KEEP, [%[c_ptr0], #0x40]\n"
+ ".inst 0x4f84e31b // sdot v27.4s, v24.16b, v4.4b[0]\n"
+ "str q29, [c_ptr3]\n"
"movi v29.4s, #0\n"
- "prfm PSTL1KEEP, [c_ptr2, #0x40]\n"
- ".word 0x4f88e31c // sdot v28.4s, v24.16b, v8.4b[0]\n"
- "str q30, [c_ptr4], #0x10\n"
+ "add c_ptr2, c_ptr2, #0x10\n"
+ ".inst 0x4f88e31c // sdot v28.4s, v24.16b, v8.4b[0]\n"
+ "str q30, [c_ptr4]\n"
"movi v30.4s, #0\n"
- "prfm PSTL1KEEP, [c_ptr3, #0x40]\n"
- ".word 0x4f8ce31d // sdot v29.4s, v24.16b, v12.4b[0]\n"
- "str q31, [c_ptr5], #0x10\n"
+ "add c_ptr3, c_ptr3, #0x10\n"
+ ".inst 0x4f8ce31d // sdot v29.4s, v24.16b, v12.4b[0]\n"
+ "str q31, [c_ptr5]\n"
"movi v31.4s, #0\n"
- "prfm PSTL1KEEP, [c_ptr4, #0x40]\n"
- ".word 0x4f90e31e // sdot v30.4s, v24.16b, v16.4b[0]\n"
- "prfm PSTL1KEEP, [c_ptr5, #0x40]\n"
- ".word 0x4f94e31f // sdot v31.4s, v24.16b, v20.4b[0]\n"
+ "add c_ptr4, c_ptr4, #0x10\n"
+ ".inst 0x4f90e31e // sdot v30.4s, v24.16b, v16.4b[0]\n"
+ "add c_ptr5, c_ptr5, #0x10\n"
+ ".inst 0x4f94e31f // sdot v31.4s, v24.16b, v20.4b[0]\n"
"ldr d24, [%[b_ptr0]]\n"
- ".word 0x4fa0e33a // sdot v26.4s, v25.16b, v0.4b[1]\n"
- ".word 0x4fa4e33b // sdot v27.4s, v25.16b, v4.4b[1]\n"
- ".word 0x4fa8e33c // sdot v28.4s, v25.16b, v8.4b[1]\n"
+ ".inst 0x4fa0e33a // sdot v26.4s, v25.16b, v0.4b[1]\n"
+ "prfm PSTL1KEEP, [c_ptr1, #0x40]\n"
+ ".inst 0x4fa4e33b // sdot v27.4s, v25.16b, v4.4b[1]\n"
"ins v24.d[1], temploadreg0\n"
- ".word 0x4face33d // sdot v29.4s, v25.16b, v12.4b[1]\n"
- ".word 0x4fb0e33e // sdot v30.4s, v25.16b, v16.4b[1]\n"
- ".word 0x4fb4e33f // sdot v31.4s, v25.16b, v20.4b[1]\n"
+ ".inst 0x4fa8e33c // sdot v28.4s, v25.16b, v8.4b[1]\n"
+ "prfm PSTL1KEEP, [c_ptr2, #0x40]\n"
+ ".inst 0x4face33d // sdot v29.4s, v25.16b, v12.4b[1]\n"
+ "prfm PSTL1KEEP, [c_ptr3, #0x40]\n"
+ ".inst 0x4fb0e33e // sdot v30.4s, v25.16b, v16.4b[1]\n"
+ "prfm PSTL1KEEP, [c_ptr4, #0x40]\n"
+ ".inst 0x4fb4e33f // sdot v31.4s, v25.16b, v20.4b[1]\n"
"ldr d25, [%[b_ptr0], #0x10]\n"
- ".word 0x4f80eb1a // sdot v26.4s, v24.16b, v0.4b[2]\n"
- "add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f84eb1b // sdot v27.4s, v24.16b, v4.4b[2]\n"
+ ".inst 0x4f80eb1a // sdot v26.4s, v24.16b, v0.4b[2]\n"
+ "prfm PSTL1KEEP, [c_ptr5, #0x40]\n"
+ ".inst 0x4f84eb1b // sdot v27.4s, v24.16b, v4.4b[2]\n"
"ins v25.d[1], temploadreg1\n"
- ".word 0x4f88eb1c // sdot v28.4s, v24.16b, v8.4b[2]\n"
+ ".inst 0x4f88eb1c // sdot v28.4s, v24.16b, v8.4b[2]\n"
+ "add %[b_ptr0], %[b_ptr0], #0x20\n"
+ ".inst 0x4f8ceb1d // sdot v29.4s, v24.16b, v12.4b[2]\n"
"ldr temploadreg0, [%[b_ptr0], #0x8]\n"
- ".word 0x4f8ceb1d // sdot v29.4s, v24.16b, v12.4b[2]\n"
+ ".inst 0x4f90eb1e // sdot v30.4s, v24.16b, v16.4b[2]\n"
"ldr temploadreg1, [%[b_ptr0], #0x18]\n"
- ".word 0x4f90eb1e // sdot v30.4s, v24.16b, v16.4b[2]\n"
- ".word 0x4f94eb1f // sdot v31.4s, v24.16b, v20.4b[2]\n"
+ ".inst 0x4f94eb1f // sdot v31.4s, v24.16b, v20.4b[2]\n"
"ldr d24, [%[b_ptr0]]\n"
- ".word 0x4fa0eb3a // sdot v26.4s, v25.16b, v0.4b[3]\n"
- ".word 0x4fa4eb3b // sdot v27.4s, v25.16b, v4.4b[3]\n"
- ".word 0x4fa8eb3c // sdot v28.4s, v25.16b, v8.4b[3]\n"
+ ".inst 0x4fa0eb3a // sdot v26.4s, v25.16b, v0.4b[3]\n"
+ ".inst 0x4fa4eb3b // sdot v27.4s, v25.16b, v4.4b[3]\n"
+ ".inst 0x4fa8eb3c // sdot v28.4s, v25.16b, v8.4b[3]\n"
"ins v24.d[1], temploadreg0\n"
- ".word 0x4faceb3d // sdot v29.4s, v25.16b, v12.4b[3]\n"
- ".word 0x4fb0eb3e // sdot v30.4s, v25.16b, v16.4b[3]\n"
- ".word 0x4fb4eb3f // sdot v31.4s, v25.16b, v20.4b[3]\n"
+ ".inst 0x4faceb3d // sdot v29.4s, v25.16b, v12.4b[3]\n"
+ ".inst 0x4fb0eb3e // sdot v30.4s, v25.16b, v16.4b[3]\n"
+ ".inst 0x4fb4eb3f // sdot v31.4s, v25.16b, v20.4b[3]\n"
"ldr d25, [%[b_ptr0], #0x10]\n"
- ".word 0x4f81e31a // sdot v26.4s, v24.16b, v1.4b[0]\n"
+ ".inst 0x4f81e31a // sdot v26.4s, v24.16b, v1.4b[0]\n"
"add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f85e31b // sdot v27.4s, v24.16b, v5.4b[0]\n"
+ ".inst 0x4f85e31b // sdot v27.4s, v24.16b, v5.4b[0]\n"
"ins v25.d[1], temploadreg1\n"
- ".word 0x4f89e31c // sdot v28.4s, v24.16b, v9.4b[0]\n"
+ ".inst 0x4f89e31c // sdot v28.4s, v24.16b, v9.4b[0]\n"
"ldr temploadreg0, [%[b_ptr0], #0x8]\n"
- ".word 0x4f8de31d // sdot v29.4s, v24.16b, v13.4b[0]\n"
+ ".inst 0x4f8de31d // sdot v29.4s, v24.16b, v13.4b[0]\n"
"ldr temploadreg1, [%[b_ptr0], #0x18]\n"
- ".word 0x4f91e31e // sdot v30.4s, v24.16b, v17.4b[0]\n"
- ".word 0x4f95e31f // sdot v31.4s, v24.16b, v21.4b[0]\n"
+ ".inst 0x4f91e31e // sdot v30.4s, v24.16b, v17.4b[0]\n"
+ ".inst 0x4f95e31f // sdot v31.4s, v24.16b, v21.4b[0]\n"
"ldr d24, [%[b_ptr0]]\n"
- ".word 0x4fa1e33a // sdot v26.4s, v25.16b, v1.4b[1]\n"
- ".word 0x4fa5e33b // sdot v27.4s, v25.16b, v5.4b[1]\n"
- ".word 0x4fa9e33c // sdot v28.4s, v25.16b, v9.4b[1]\n"
+ ".inst 0x4fa1e33a // sdot v26.4s, v25.16b, v1.4b[1]\n"
+ ".inst 0x4fa5e33b // sdot v27.4s, v25.16b, v5.4b[1]\n"
+ ".inst 0x4fa9e33c // sdot v28.4s, v25.16b, v9.4b[1]\n"
"ins v24.d[1], temploadreg0\n"
- ".word 0x4fade33d // sdot v29.4s, v25.16b, v13.4b[1]\n"
- ".word 0x4fb1e33e // sdot v30.4s, v25.16b, v17.4b[1]\n"
- ".word 0x4fb5e33f // sdot v31.4s, v25.16b, v21.4b[1]\n"
+ ".inst 0x4fade33d // sdot v29.4s, v25.16b, v13.4b[1]\n"
+ ".inst 0x4fb1e33e // sdot v30.4s, v25.16b, v17.4b[1]\n"
+ ".inst 0x4fb5e33f // sdot v31.4s, v25.16b, v21.4b[1]\n"
"ldr d25, [%[b_ptr0], #0x10]\n"
- ".word 0x4f81eb1a // sdot v26.4s, v24.16b, v1.4b[2]\n"
+ ".inst 0x4f81eb1a // sdot v26.4s, v24.16b, v1.4b[2]\n"
"add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f85eb1b // sdot v27.4s, v24.16b, v5.4b[2]\n"
+ ".inst 0x4f85eb1b // sdot v27.4s, v24.16b, v5.4b[2]\n"
"ins v25.d[1], temploadreg1\n"
- ".word 0x4f89eb1c // sdot v28.4s, v24.16b, v9.4b[2]\n"
+ ".inst 0x4f89eb1c // sdot v28.4s, v24.16b, v9.4b[2]\n"
"ldr temploadreg0, [%[b_ptr0], #0x8]\n"
- ".word 0x4f8deb1d // sdot v29.4s, v24.16b, v13.4b[2]\n"
+ ".inst 0x4f8deb1d // sdot v29.4s, v24.16b, v13.4b[2]\n"
"ldr temploadreg1, [%[b_ptr0], #0x18]\n"
- ".word 0x4f91eb1e // sdot v30.4s, v24.16b, v17.4b[2]\n"
- ".word 0x4f95eb1f // sdot v31.4s, v24.16b, v21.4b[2]\n"
+ ".inst 0x4f91eb1e // sdot v30.4s, v24.16b, v17.4b[2]\n"
+ ".inst 0x4f95eb1f // sdot v31.4s, v24.16b, v21.4b[2]\n"
"ldr d24, [%[b_ptr0]]\n"
- ".word 0x4fa1eb3a // sdot v26.4s, v25.16b, v1.4b[3]\n"
- ".word 0x4fa5eb3b // sdot v27.4s, v25.16b, v5.4b[3]\n"
- ".word 0x4fa9eb3c // sdot v28.4s, v25.16b, v9.4b[3]\n"
+ ".inst 0x4fa1eb3a // sdot v26.4s, v25.16b, v1.4b[3]\n"
+ ".inst 0x4fa5eb3b // sdot v27.4s, v25.16b, v5.4b[3]\n"
+ ".inst 0x4fa9eb3c // sdot v28.4s, v25.16b, v9.4b[3]\n"
"ins v24.d[1], temploadreg0\n"
- ".word 0x4fadeb3d // sdot v29.4s, v25.16b, v13.4b[3]\n"
- ".word 0x4fb1eb3e // sdot v30.4s, v25.16b, v17.4b[3]\n"
- ".word 0x4fb5eb3f // sdot v31.4s, v25.16b, v21.4b[3]\n"
+ ".inst 0x4fadeb3d // sdot v29.4s, v25.16b, v13.4b[3]\n"
+ ".inst 0x4fb1eb3e // sdot v30.4s, v25.16b, v17.4b[3]\n"
+ ".inst 0x4fb5eb3f // sdot v31.4s, v25.16b, v21.4b[3]\n"
"ldr d25, [%[b_ptr0], #0x10]\n"
- ".word 0x4f82e31a // sdot v26.4s, v24.16b, v2.4b[0]\n"
+ ".inst 0x4f82e31a // sdot v26.4s, v24.16b, v2.4b[0]\n"
"add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f86e31b // sdot v27.4s, v24.16b, v6.4b[0]\n"
+ ".inst 0x4f86e31b // sdot v27.4s, v24.16b, v6.4b[0]\n"
"ins v25.d[1], temploadreg1\n"
- ".word 0x4f8ae31c // sdot v28.4s, v24.16b, v10.4b[0]\n"
+ ".inst 0x4f8ae31c // sdot v28.4s, v24.16b, v10.4b[0]\n"
"ldr temploadreg0, [%[b_ptr0], #0x8]\n"
- ".word 0x4f8ee31d // sdot v29.4s, v24.16b, v14.4b[0]\n"
+ ".inst 0x4f8ee31d // sdot v29.4s, v24.16b, v14.4b[0]\n"
"ldr temploadreg1, [%[b_ptr0], #0x18]\n"
- ".word 0x4f92e31e // sdot v30.4s, v24.16b, v18.4b[0]\n"
- ".word 0x4f96e31f // sdot v31.4s, v24.16b, v22.4b[0]\n"
+ ".inst 0x4f92e31e // sdot v30.4s, v24.16b, v18.4b[0]\n"
+ ".inst 0x4f96e31f // sdot v31.4s, v24.16b, v22.4b[0]\n"
"ldr d24, [%[b_ptr0]]\n"
- ".word 0x4fa2e33a // sdot v26.4s, v25.16b, v2.4b[1]\n"
- ".word 0x4fa6e33b // sdot v27.4s, v25.16b, v6.4b[1]\n"
- ".word 0x4faae33c // sdot v28.4s, v25.16b, v10.4b[1]\n"
+ ".inst 0x4fa2e33a // sdot v26.4s, v25.16b, v2.4b[1]\n"
+ ".inst 0x4fa6e33b // sdot v27.4s, v25.16b, v6.4b[1]\n"
+ ".inst 0x4faae33c // sdot v28.4s, v25.16b, v10.4b[1]\n"
"ins v24.d[1], temploadreg0\n"
- ".word 0x4faee33d // sdot v29.4s, v25.16b, v14.4b[1]\n"
- ".word 0x4fb2e33e // sdot v30.4s, v25.16b, v18.4b[1]\n"
- ".word 0x4fb6e33f // sdot v31.4s, v25.16b, v22.4b[1]\n"
+ ".inst 0x4faee33d // sdot v29.4s, v25.16b, v14.4b[1]\n"
+ ".inst 0x4fb2e33e // sdot v30.4s, v25.16b, v18.4b[1]\n"
+ ".inst 0x4fb6e33f // sdot v31.4s, v25.16b, v22.4b[1]\n"
"ldr d25, [%[b_ptr0], #0x10]\n"
- ".word 0x4f82eb1a // sdot v26.4s, v24.16b, v2.4b[2]\n"
+ ".inst 0x4f82eb1a // sdot v26.4s, v24.16b, v2.4b[2]\n"
"add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f86eb1b // sdot v27.4s, v24.16b, v6.4b[2]\n"
+ ".inst 0x4f86eb1b // sdot v27.4s, v24.16b, v6.4b[2]\n"
"ins v25.d[1], temploadreg1\n"
- ".word 0x4f8aeb1c // sdot v28.4s, v24.16b, v10.4b[2]\n"
+ ".inst 0x4f8aeb1c // sdot v28.4s, v24.16b, v10.4b[2]\n"
"ldr temploadreg0, [%[b_ptr0], #0x8]\n"
- ".word 0x4f8eeb1d // sdot v29.4s, v24.16b, v14.4b[2]\n"
+ ".inst 0x4f8eeb1d // sdot v29.4s, v24.16b, v14.4b[2]\n"
"ldr temploadreg1, [%[b_ptr0], #0x18]\n"
- ".word 0x4f92eb1e // sdot v30.4s, v24.16b, v18.4b[2]\n"
- ".word 0x4f96eb1f // sdot v31.4s, v24.16b, v22.4b[2]\n"
+ ".inst 0x4f92eb1e // sdot v30.4s, v24.16b, v18.4b[2]\n"
+ ".inst 0x4f96eb1f // sdot v31.4s, v24.16b, v22.4b[2]\n"
"ldr d24, [%[b_ptr0]]\n"
- ".word 0x4fa2eb3a // sdot v26.4s, v25.16b, v2.4b[3]\n"
- ".word 0x4fa6eb3b // sdot v27.4s, v25.16b, v6.4b[3]\n"
- ".word 0x4faaeb3c // sdot v28.4s, v25.16b, v10.4b[3]\n"
+ ".inst 0x4fa2eb3a // sdot v26.4s, v25.16b, v2.4b[3]\n"
+ ".inst 0x4fa6eb3b // sdot v27.4s, v25.16b, v6.4b[3]\n"
+ ".inst 0x4faaeb3c // sdot v28.4s, v25.16b, v10.4b[3]\n"
"ins v24.d[1], temploadreg0\n"
- ".word 0x4faeeb3d // sdot v29.4s, v25.16b, v14.4b[3]\n"
- ".word 0x4fb2eb3e // sdot v30.4s, v25.16b, v18.4b[3]\n"
- ".word 0x4fb6eb3f // sdot v31.4s, v25.16b, v22.4b[3]\n"
+ ".inst 0x4faeeb3d // sdot v29.4s, v25.16b, v14.4b[3]\n"
+ ".inst 0x4fb2eb3e // sdot v30.4s, v25.16b, v18.4b[3]\n"
+ ".inst 0x4fb6eb3f // sdot v31.4s, v25.16b, v22.4b[3]\n"
"ldr d25, [%[b_ptr0], #0x10]\n"
- ".word 0x4f83e31a // sdot v26.4s, v24.16b, v3.4b[0]\n"
+ ".inst 0x4f83e31a // sdot v26.4s, v24.16b, v3.4b[0]\n"
"add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f87e31b // sdot v27.4s, v24.16b, v7.4b[0]\n"
+ ".inst 0x4f87e31b // sdot v27.4s, v24.16b, v7.4b[0]\n"
"ins v25.d[1], temploadreg1\n"
- ".word 0x4f8be31c // sdot v28.4s, v24.16b, v11.4b[0]\n"
+ ".inst 0x4f8be31c // sdot v28.4s, v24.16b, v11.4b[0]\n"
"ldr temploadreg0, [%[b_ptr0], #0x8]\n"
- ".word 0x4f8fe31d // sdot v29.4s, v24.16b, v15.4b[0]\n"
+ ".inst 0x4f8fe31d // sdot v29.4s, v24.16b, v15.4b[0]\n"
"ldr temploadreg1, [%[b_ptr0], #0x18]\n"
- ".word 0x4f93e31e // sdot v30.4s, v24.16b, v19.4b[0]\n"
- ".word 0x4f97e31f // sdot v31.4s, v24.16b, v23.4b[0]\n"
+ ".inst 0x4f93e31e // sdot v30.4s, v24.16b, v19.4b[0]\n"
+ ".inst 0x4f97e31f // sdot v31.4s, v24.16b, v23.4b[0]\n"
"ldr d24, [%[b_ptr0]]\n"
- ".word 0x4fa3e33a // sdot v26.4s, v25.16b, v3.4b[1]\n"
- ".word 0x4fa7e33b // sdot v27.4s, v25.16b, v7.4b[1]\n"
- ".word 0x4fabe33c // sdot v28.4s, v25.16b, v11.4b[1]\n"
+ ".inst 0x4fa3e33a // sdot v26.4s, v25.16b, v3.4b[1]\n"
+ ".inst 0x4fa7e33b // sdot v27.4s, v25.16b, v7.4b[1]\n"
+ ".inst 0x4fabe33c // sdot v28.4s, v25.16b, v11.4b[1]\n"
"ins v24.d[1], temploadreg0\n"
- ".word 0x4fafe33d // sdot v29.4s, v25.16b, v15.4b[1]\n"
- ".word 0x4fb3e33e // sdot v30.4s, v25.16b, v19.4b[1]\n"
- ".word 0x4fb7e33f // sdot v31.4s, v25.16b, v23.4b[1]\n"
+ ".inst 0x4fafe33d // sdot v29.4s, v25.16b, v15.4b[1]\n"
+ ".inst 0x4fb3e33e // sdot v30.4s, v25.16b, v19.4b[1]\n"
+ ".inst 0x4fb7e33f // sdot v31.4s, v25.16b, v23.4b[1]\n"
"ldr d25, [%[b_ptr0], #0x10]\n"
- ".word 0x4f83eb1a // sdot v26.4s, v24.16b, v3.4b[2]\n"
+ ".inst 0x4f83eb1a // sdot v26.4s, v24.16b, v3.4b[2]\n"
"add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f87eb1b // sdot v27.4s, v24.16b, v7.4b[2]\n"
+ ".inst 0x4f87eb1b // sdot v27.4s, v24.16b, v7.4b[2]\n"
"ins v25.d[1], temploadreg1\n"
- ".word 0x4f8beb1c // sdot v28.4s, v24.16b, v11.4b[2]\n"
+ ".inst 0x4f8beb1c // sdot v28.4s, v24.16b, v11.4b[2]\n"
"ldr temploadreg0, [%[b_ptr0], #0x8]\n"
- ".word 0x4f8feb1d // sdot v29.4s, v24.16b, v15.4b[2]\n"
+ ".inst 0x4f8feb1d // sdot v29.4s, v24.16b, v15.4b[2]\n"
"ldr temploadreg1, [%[b_ptr0], #0x18]\n"
- ".word 0x4f93eb1e // sdot v30.4s, v24.16b, v19.4b[2]\n"
- ".word 0x4f97eb1f // sdot v31.4s, v24.16b, v23.4b[2]\n"
+ ".inst 0x4f93eb1e // sdot v30.4s, v24.16b, v19.4b[2]\n"
+ ".inst 0x4f97eb1f // sdot v31.4s, v24.16b, v23.4b[2]\n"
"ldr d24, [%[b_ptr0]]\n"
- ".word 0x4fa3eb3a // sdot v26.4s, v25.16b, v3.4b[3]\n"
- ".word 0x4fa7eb3b // sdot v27.4s, v25.16b, v7.4b[3]\n"
- ".word 0x4fabeb3c // sdot v28.4s, v25.16b, v11.4b[3]\n"
+ ".inst 0x4fa3eb3a // sdot v26.4s, v25.16b, v3.4b[3]\n"
+ ".inst 0x4fa7eb3b // sdot v27.4s, v25.16b, v7.4b[3]\n"
+ ".inst 0x4fabeb3c // sdot v28.4s, v25.16b, v11.4b[3]\n"
"ins v24.d[1], temploadreg0\n"
- ".word 0x4fafeb3d // sdot v29.4s, v25.16b, v15.4b[3]\n"
- ".word 0x4fb3eb3e // sdot v30.4s, v25.16b, v19.4b[3]\n"
- ".word 0x4fb7eb3f // sdot v31.4s, v25.16b, v23.4b[3]\n"
+ ".inst 0x4fafeb3d // sdot v29.4s, v25.16b, v15.4b[3]\n"
+ ".inst 0x4fb3eb3e // sdot v30.4s, v25.16b, v19.4b[3]\n"
+ ".inst 0x4fb7eb3f // sdot v31.4s, v25.16b, v23.4b[3]\n"
"ldr d25, [%[b_ptr0], #0x10]\n"
"add %[b_ptr0], %[b_ptr0], #0x20\n"
"b.ne 8b\n"
"7:\n"
- "str q26, [%[c_ptr0]], #0x10\n"
+ "str q26, [%[c_ptr0]]\n"
+ "add %[c_ptr0], %[c_ptr0], #0x10\n"
"movi v26.4s, #0\n"
"ins v25.d[1], temploadreg1\n"
- "str q27, [c_ptr1], #0x10\n"
+ "str q27, [c_ptr1]\n"
+ "add c_ptr1, c_ptr1, #0x10\n"
"movi v27.4s, #0\n"
- ".word 0x4f80e31a // sdot v26.4s, v24.16b, v0.4b[0]\n"
- "str q28, [c_ptr2], #0x10\n"
+ ".inst 0x4f80e31a // sdot v26.4s, v24.16b, v0.4b[0]\n"
+ "str q28, [c_ptr2]\n"
"movi v28.4s, #0\n"
- ".word 0x4f84e31b // sdot v27.4s, v24.16b, v4.4b[0]\n"
- ".word 0x4fa0e33a // sdot v26.4s, v25.16b, v0.4b[1]\n"
- "str q29, [c_ptr3], #0x10\n"
+ "add c_ptr2, c_ptr2, #0x10\n"
+ ".inst 0x4f84e31b // sdot v27.4s, v24.16b, v4.4b[0]\n"
+ "str q29, [c_ptr3]\n"
"movi v29.4s, #0\n"
- ".word 0x4f88e31c // sdot v28.4s, v24.16b, v8.4b[0]\n"
- ".word 0x4fa4e33b // sdot v27.4s, v25.16b, v4.4b[1]\n"
- "str q30, [c_ptr4], #0x10\n"
+ "add c_ptr3, c_ptr3, #0x10\n"
+ ".inst 0x4f88e31c // sdot v28.4s, v24.16b, v8.4b[0]\n"
+ "str q30, [c_ptr4]\n"
"movi v30.4s, #0\n"
- ".word 0x4f8ce31d // sdot v29.4s, v24.16b, v12.4b[0]\n"
- ".word 0x4fa8e33c // sdot v28.4s, v25.16b, v8.4b[1]\n"
- "str q31, [c_ptr5], #0x10\n"
+ "add c_ptr4, c_ptr4, #0x10\n"
+ ".inst 0x4f8ce31d // sdot v29.4s, v24.16b, v12.4b[0]\n"
+ "str q31, [c_ptr5]\n"
"movi v31.4s, #0\n"
- ".word 0x4f90e31e // sdot v30.4s, v24.16b, v16.4b[0]\n"
- ".word 0x4face33d // sdot v29.4s, v25.16b, v12.4b[1]\n"
- ".word 0x4f94e31f // sdot v31.4s, v24.16b, v20.4b[0]\n"
+ "add c_ptr5, c_ptr5, #0x10\n"
+ ".inst 0x4f90e31e // sdot v30.4s, v24.16b, v16.4b[0]\n"
+ ".inst 0x4fa0e33a // sdot v26.4s, v25.16b, v0.4b[1]\n"
+ ".inst 0x4f94e31f // sdot v31.4s, v24.16b, v20.4b[0]\n"
"ldr q24, [%[b_ptr0]]\n"
- ".word 0x4fb0e33e // sdot v30.4s, v25.16b, v16.4b[1]\n"
- ".word 0x4fb4e33f // sdot v31.4s, v25.16b, v20.4b[1]\n"
+ ".inst 0x4fa4e33b // sdot v27.4s, v25.16b, v4.4b[1]\n"
+ ".inst 0x4fa8e33c // sdot v28.4s, v25.16b, v8.4b[1]\n"
+ ".inst 0x4face33d // sdot v29.4s, v25.16b, v12.4b[1]\n"
+ ".inst 0x4fb0e33e // sdot v30.4s, v25.16b, v16.4b[1]\n"
+ ".inst 0x4fb4e33f // sdot v31.4s, v25.16b, v20.4b[1]\n"
"ldr q25, [%[b_ptr0], #0x10]\n"
- ".word 0x4f80eb1a // sdot v26.4s, v24.16b, v0.4b[2]\n"
+ ".inst 0x4f80eb1a // sdot v26.4s, v24.16b, v0.4b[2]\n"
"add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f84eb1b // sdot v27.4s, v24.16b, v4.4b[2]\n"
- ".word 0x4f88eb1c // sdot v28.4s, v24.16b, v8.4b[2]\n"
- ".word 0x4f8ceb1d // sdot v29.4s, v24.16b, v12.4b[2]\n"
- ".word 0x4f90eb1e // sdot v30.4s, v24.16b, v16.4b[2]\n"
- ".word 0x4f94eb1f // sdot v31.4s, v24.16b, v20.4b[2]\n"
+ ".inst 0x4f84eb1b // sdot v27.4s, v24.16b, v4.4b[2]\n"
+ ".inst 0x4f88eb1c // sdot v28.4s, v24.16b, v8.4b[2]\n"
+ ".inst 0x4f8ceb1d // sdot v29.4s, v24.16b, v12.4b[2]\n"
+ ".inst 0x4f90eb1e // sdot v30.4s, v24.16b, v16.4b[2]\n"
+ ".inst 0x4f94eb1f // sdot v31.4s, v24.16b, v20.4b[2]\n"
"ldr q24, [%[b_ptr0]]\n"
- ".word 0x4fa0eb3a // sdot v26.4s, v25.16b, v0.4b[3]\n"
- ".word 0x4fa4eb3b // sdot v27.4s, v25.16b, v4.4b[3]\n"
- ".word 0x4fa8eb3c // sdot v28.4s, v25.16b, v8.4b[3]\n"
- ".word 0x4faceb3d // sdot v29.4s, v25.16b, v12.4b[3]\n"
- ".word 0x4fb0eb3e // sdot v30.4s, v25.16b, v16.4b[3]\n"
- ".word 0x4fb4eb3f // sdot v31.4s, v25.16b, v20.4b[3]\n"
+ ".inst 0x4fa0eb3a // sdot v26.4s, v25.16b, v0.4b[3]\n"
+ ".inst 0x4fa4eb3b // sdot v27.4s, v25.16b, v4.4b[3]\n"
+ ".inst 0x4fa8eb3c // sdot v28.4s, v25.16b, v8.4b[3]\n"
+ ".inst 0x4faceb3d // sdot v29.4s, v25.16b, v12.4b[3]\n"
+ ".inst 0x4fb0eb3e // sdot v30.4s, v25.16b, v16.4b[3]\n"
+ ".inst 0x4fb4eb3f // sdot v31.4s, v25.16b, v20.4b[3]\n"
"ldr q25, [%[b_ptr0], #0x10]\n"
- ".word 0x4f81e31a // sdot v26.4s, v24.16b, v1.4b[0]\n"
+ ".inst 0x4f81e31a // sdot v26.4s, v24.16b, v1.4b[0]\n"
"add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f85e31b // sdot v27.4s, v24.16b, v5.4b[0]\n"
- ".word 0x4f89e31c // sdot v28.4s, v24.16b, v9.4b[0]\n"
- ".word 0x4f8de31d // sdot v29.4s, v24.16b, v13.4b[0]\n"
- ".word 0x4f91e31e // sdot v30.4s, v24.16b, v17.4b[0]\n"
- ".word 0x4f95e31f // sdot v31.4s, v24.16b, v21.4b[0]\n"
+ ".inst 0x4f85e31b // sdot v27.4s, v24.16b, v5.4b[0]\n"
+ ".inst 0x4f89e31c // sdot v28.4s, v24.16b, v9.4b[0]\n"
+ ".inst 0x4f8de31d // sdot v29.4s, v24.16b, v13.4b[0]\n"
+ ".inst 0x4f91e31e // sdot v30.4s, v24.16b, v17.4b[0]\n"
+ ".inst 0x4f95e31f // sdot v31.4s, v24.16b, v21.4b[0]\n"
"ldr q24, [%[b_ptr0]]\n"
- ".word 0x4fa1e33a // sdot v26.4s, v25.16b, v1.4b[1]\n"
- ".word 0x4fa5e33b // sdot v27.4s, v25.16b, v5.4b[1]\n"
- ".word 0x4fa9e33c // sdot v28.4s, v25.16b, v9.4b[1]\n"
- ".word 0x4fade33d // sdot v29.4s, v25.16b, v13.4b[1]\n"
- ".word 0x4fb1e33e // sdot v30.4s, v25.16b, v17.4b[1]\n"
- ".word 0x4fb5e33f // sdot v31.4s, v25.16b, v21.4b[1]\n"
+ ".inst 0x4fa1e33a // sdot v26.4s, v25.16b, v1.4b[1]\n"
+ ".inst 0x4fa5e33b // sdot v27.4s, v25.16b, v5.4b[1]\n"
+ ".inst 0x4fa9e33c // sdot v28.4s, v25.16b, v9.4b[1]\n"
+ ".inst 0x4fade33d // sdot v29.4s, v25.16b, v13.4b[1]\n"
+ ".inst 0x4fb1e33e // sdot v30.4s, v25.16b, v17.4b[1]\n"
+ ".inst 0x4fb5e33f // sdot v31.4s, v25.16b, v21.4b[1]\n"
"ldr q25, [%[b_ptr0], #0x10]\n"
- ".word 0x4f81eb1a // sdot v26.4s, v24.16b, v1.4b[2]\n"
+ ".inst 0x4f81eb1a // sdot v26.4s, v24.16b, v1.4b[2]\n"
"add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f85eb1b // sdot v27.4s, v24.16b, v5.4b[2]\n"
- ".word 0x4f89eb1c // sdot v28.4s, v24.16b, v9.4b[2]\n"
- ".word 0x4f8deb1d // sdot v29.4s, v24.16b, v13.4b[2]\n"
- ".word 0x4f91eb1e // sdot v30.4s, v24.16b, v17.4b[2]\n"
- ".word 0x4f95eb1f // sdot v31.4s, v24.16b, v21.4b[2]\n"
+ ".inst 0x4f85eb1b // sdot v27.4s, v24.16b, v5.4b[2]\n"
+ ".inst 0x4f89eb1c // sdot v28.4s, v24.16b, v9.4b[2]\n"
+ ".inst 0x4f8deb1d // sdot v29.4s, v24.16b, v13.4b[2]\n"
+ ".inst 0x4f91eb1e // sdot v30.4s, v24.16b, v17.4b[2]\n"
+ ".inst 0x4f95eb1f // sdot v31.4s, v24.16b, v21.4b[2]\n"
"ldr q24, [%[b_ptr0]]\n"
- ".word 0x4fa1eb3a // sdot v26.4s, v25.16b, v1.4b[3]\n"
- ".word 0x4fa5eb3b // sdot v27.4s, v25.16b, v5.4b[3]\n"
- ".word 0x4fa9eb3c // sdot v28.4s, v25.16b, v9.4b[3]\n"
- ".word 0x4fadeb3d // sdot v29.4s, v25.16b, v13.4b[3]\n"
- ".word 0x4fb1eb3e // sdot v30.4s, v25.16b, v17.4b[3]\n"
- ".word 0x4fb5eb3f // sdot v31.4s, v25.16b, v21.4b[3]\n"
+ ".inst 0x4fa1eb3a // sdot v26.4s, v25.16b, v1.4b[3]\n"
+ ".inst 0x4fa5eb3b // sdot v27.4s, v25.16b, v5.4b[3]\n"
+ ".inst 0x4fa9eb3c // sdot v28.4s, v25.16b, v9.4b[3]\n"
+ ".inst 0x4fadeb3d // sdot v29.4s, v25.16b, v13.4b[3]\n"
+ ".inst 0x4fb1eb3e // sdot v30.4s, v25.16b, v17.4b[3]\n"
+ ".inst 0x4fb5eb3f // sdot v31.4s, v25.16b, v21.4b[3]\n"
"ldr q25, [%[b_ptr0], #0x10]\n"
- ".word 0x4f82e31a // sdot v26.4s, v24.16b, v2.4b[0]\n"
+ ".inst 0x4f82e31a // sdot v26.4s, v24.16b, v2.4b[0]\n"
"add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f86e31b // sdot v27.4s, v24.16b, v6.4b[0]\n"
- ".word 0x4f8ae31c // sdot v28.4s, v24.16b, v10.4b[0]\n"
- ".word 0x4f8ee31d // sdot v29.4s, v24.16b, v14.4b[0]\n"
- ".word 0x4f92e31e // sdot v30.4s, v24.16b, v18.4b[0]\n"
- ".word 0x4f96e31f // sdot v31.4s, v24.16b, v22.4b[0]\n"
+ ".inst 0x4f86e31b // sdot v27.4s, v24.16b, v6.4b[0]\n"
+ ".inst 0x4f8ae31c // sdot v28.4s, v24.16b, v10.4b[0]\n"
+ ".inst 0x4f8ee31d // sdot v29.4s, v24.16b, v14.4b[0]\n"
+ ".inst 0x4f92e31e // sdot v30.4s, v24.16b, v18.4b[0]\n"
+ ".inst 0x4f96e31f // sdot v31.4s, v24.16b, v22.4b[0]\n"
"ldr q24, [%[b_ptr0]]\n"
- ".word 0x4fa2e33a // sdot v26.4s, v25.16b, v2.4b[1]\n"
- ".word 0x4fa6e33b // sdot v27.4s, v25.16b, v6.4b[1]\n"
- ".word 0x4faae33c // sdot v28.4s, v25.16b, v10.4b[1]\n"
- ".word 0x4faee33d // sdot v29.4s, v25.16b, v14.4b[1]\n"
- ".word 0x4fb2e33e // sdot v30.4s, v25.16b, v18.4b[1]\n"
- ".word 0x4fb6e33f // sdot v31.4s, v25.16b, v22.4b[1]\n"
+ ".inst 0x4fa2e33a // sdot v26.4s, v25.16b, v2.4b[1]\n"
+ ".inst 0x4fa6e33b // sdot v27.4s, v25.16b, v6.4b[1]\n"
+ ".inst 0x4faae33c // sdot v28.4s, v25.16b, v10.4b[1]\n"
+ ".inst 0x4faee33d // sdot v29.4s, v25.16b, v14.4b[1]\n"
+ ".inst 0x4fb2e33e // sdot v30.4s, v25.16b, v18.4b[1]\n"
+ ".inst 0x4fb6e33f // sdot v31.4s, v25.16b, v22.4b[1]\n"
"ldr q25, [%[b_ptr0], #0x10]\n"
- ".word 0x4f82eb1a // sdot v26.4s, v24.16b, v2.4b[2]\n"
+ ".inst 0x4f82eb1a // sdot v26.4s, v24.16b, v2.4b[2]\n"
"add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f86eb1b // sdot v27.4s, v24.16b, v6.4b[2]\n"
- ".word 0x4f8aeb1c // sdot v28.4s, v24.16b, v10.4b[2]\n"
- ".word 0x4f8eeb1d // sdot v29.4s, v24.16b, v14.4b[2]\n"
- ".word 0x4f92eb1e // sdot v30.4s, v24.16b, v18.4b[2]\n"
- ".word 0x4f96eb1f // sdot v31.4s, v24.16b, v22.4b[2]\n"
+ ".inst 0x4f86eb1b // sdot v27.4s, v24.16b, v6.4b[2]\n"
+ ".inst 0x4f8aeb1c // sdot v28.4s, v24.16b, v10.4b[2]\n"
+ ".inst 0x4f8eeb1d // sdot v29.4s, v24.16b, v14.4b[2]\n"
+ ".inst 0x4f92eb1e // sdot v30.4s, v24.16b, v18.4b[2]\n"
+ ".inst 0x4f96eb1f // sdot v31.4s, v24.16b, v22.4b[2]\n"
"ldr q24, [%[b_ptr0]]\n"
- ".word 0x4fa2eb3a // sdot v26.4s, v25.16b, v2.4b[3]\n"
- ".word 0x4fa6eb3b // sdot v27.4s, v25.16b, v6.4b[3]\n"
- ".word 0x4faaeb3c // sdot v28.4s, v25.16b, v10.4b[3]\n"
- ".word 0x4faeeb3d // sdot v29.4s, v25.16b, v14.4b[3]\n"
- ".word 0x4fb2eb3e // sdot v30.4s, v25.16b, v18.4b[3]\n"
- ".word 0x4fb6eb3f // sdot v31.4s, v25.16b, v22.4b[3]\n"
+ ".inst 0x4fa2eb3a // sdot v26.4s, v25.16b, v2.4b[3]\n"
+ ".inst 0x4fa6eb3b // sdot v27.4s, v25.16b, v6.4b[3]\n"
+ ".inst 0x4faaeb3c // sdot v28.4s, v25.16b, v10.4b[3]\n"
+ ".inst 0x4faeeb3d // sdot v29.4s, v25.16b, v14.4b[3]\n"
+ ".inst 0x4fb2eb3e // sdot v30.4s, v25.16b, v18.4b[3]\n"
+ ".inst 0x4fb6eb3f // sdot v31.4s, v25.16b, v22.4b[3]\n"
"ldr q25, [%[b_ptr0], #0x10]\n"
- ".word 0x4f83e31a // sdot v26.4s, v24.16b, v3.4b[0]\n"
+ ".inst 0x4f83e31a // sdot v26.4s, v24.16b, v3.4b[0]\n"
"add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f87e31b // sdot v27.4s, v24.16b, v7.4b[0]\n"
- ".word 0x4f8be31c // sdot v28.4s, v24.16b, v11.4b[0]\n"
- ".word 0x4f8fe31d // sdot v29.4s, v24.16b, v15.4b[0]\n"
- ".word 0x4f93e31e // sdot v30.4s, v24.16b, v19.4b[0]\n"
- ".word 0x4f97e31f // sdot v31.4s, v24.16b, v23.4b[0]\n"
+ ".inst 0x4f87e31b // sdot v27.4s, v24.16b, v7.4b[0]\n"
+ ".inst 0x4f8be31c // sdot v28.4s, v24.16b, v11.4b[0]\n"
+ ".inst 0x4f8fe31d // sdot v29.4s, v24.16b, v15.4b[0]\n"
+ ".inst 0x4f93e31e // sdot v30.4s, v24.16b, v19.4b[0]\n"
+ ".inst 0x4f97e31f // sdot v31.4s, v24.16b, v23.4b[0]\n"
"ldr q24, [%[b_ptr0]]\n"
- ".word 0x4fa3e33a // sdot v26.4s, v25.16b, v3.4b[1]\n"
- ".word 0x4fa7e33b // sdot v27.4s, v25.16b, v7.4b[1]\n"
- ".word 0x4fabe33c // sdot v28.4s, v25.16b, v11.4b[1]\n"
- ".word 0x4fafe33d // sdot v29.4s, v25.16b, v15.4b[1]\n"
- ".word 0x4fb3e33e // sdot v30.4s, v25.16b, v19.4b[1]\n"
- ".word 0x4fb7e33f // sdot v31.4s, v25.16b, v23.4b[1]\n"
+ ".inst 0x4fa3e33a // sdot v26.4s, v25.16b, v3.4b[1]\n"
+ ".inst 0x4fa7e33b // sdot v27.4s, v25.16b, v7.4b[1]\n"
+ ".inst 0x4fabe33c // sdot v28.4s, v25.16b, v11.4b[1]\n"
+ ".inst 0x4fafe33d // sdot v29.4s, v25.16b, v15.4b[1]\n"
+ ".inst 0x4fb3e33e // sdot v30.4s, v25.16b, v19.4b[1]\n"
+ ".inst 0x4fb7e33f // sdot v31.4s, v25.16b, v23.4b[1]\n"
"ldr q25, [%[b_ptr0], #0x10]\n"
- ".word 0x4f83eb1a // sdot v26.4s, v24.16b, v3.4b[2]\n"
- "add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f87eb1b // sdot v27.4s, v24.16b, v7.4b[2]\n"
- ".word 0x4f8beb1c // sdot v28.4s, v24.16b, v11.4b[2]\n"
- ".word 0x4f8feb1d // sdot v29.4s, v24.16b, v15.4b[2]\n"
- ".word 0x4f93eb1e // sdot v30.4s, v24.16b, v19.4b[2]\n"
- ".word 0x4f97eb1f // sdot v31.4s, v24.16b, v23.4b[2]\n"
- ".word 0x4fa3eb3a // sdot v26.4s, v25.16b, v3.4b[3]\n"
- ".word 0x4fa7eb3b // sdot v27.4s, v25.16b, v7.4b[3]\n"
- ".word 0x4fabeb3c // sdot v28.4s, v25.16b, v11.4b[3]\n"
- ".word 0x4fafeb3d // sdot v29.4s, v25.16b, v15.4b[3]\n"
- ".word 0x4fb3eb3e // sdot v30.4s, v25.16b, v19.4b[3]\n"
- ".word 0x4fb7eb3f // sdot v31.4s, v25.16b, v23.4b[3]\n"
+ ".inst 0x4f83eb1a // sdot v26.4s, v24.16b, v3.4b[2]\n"
+ "add %[b_ptr0], %[b_ptr0], #0x20\n"
+ ".inst 0x4f87eb1b // sdot v27.4s, v24.16b, v7.4b[2]\n"
+ ".inst 0x4f8beb1c // sdot v28.4s, v24.16b, v11.4b[2]\n"
+ ".inst 0x4f8feb1d // sdot v29.4s, v24.16b, v15.4b[2]\n"
+ ".inst 0x4f93eb1e // sdot v30.4s, v24.16b, v19.4b[2]\n"
+ ".inst 0x4f97eb1f // sdot v31.4s, v24.16b, v23.4b[2]\n"
+ ".inst 0x4fa3eb3a // sdot v26.4s, v25.16b, v3.4b[3]\n"
+ ".inst 0x4fa7eb3b // sdot v27.4s, v25.16b, v7.4b[3]\n"
+ ".inst 0x4fabeb3c // sdot v28.4s, v25.16b, v11.4b[3]\n"
+ ".inst 0x4fafeb3d // sdot v29.4s, v25.16b, v15.4b[3]\n"
+ ".inst 0x4fb3eb3e // sdot v30.4s, v25.16b, v19.4b[3]\n"
+ ".inst 0x4fb7eb3f // sdot v31.4s, v25.16b, v23.4b[3]\n"
"6:\n"
"str q26, [%[c_ptr0]]\n"
+ "add %[c_ptr0], %[c_ptr0], #0x10\n"
"str q27, [c_ptr1]\n"
+ "add c_ptr1, c_ptr1, #0x10\n"
"str q28, [c_ptr2]\n"
+ "add c_ptr2, c_ptr2, #0x10\n"
"str q29, [c_ptr3]\n"
+ "add c_ptr3, c_ptr3, #0x10\n"
"str q30, [c_ptr4]\n"
+ "add c_ptr4, c_ptr4, #0x10\n"
"str q31, [c_ptr5]\n"
+ "add c_ptr5, c_ptr5, #0x10\n"
".unreq a_ptr1\n"
".unreq a_ptr2\n"
".unreq a_ptr3\n"