aboutsummaryrefslogtreecommitdiff
path: root/src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_s8s32_dot_4x8/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_4x8/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_4x8/a55.cpp')
-rw-r--r--src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_s8s32_dot_4x8/a55.cpp2611
1 files changed, 1402 insertions, 1209 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_s8s32_dot_4x8/a55.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_s8s32_dot_4x8/a55.cpp
index 4ed2bc823c..c742fcc890 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_s8s32_dot_4x8/a55.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_s8s32_dot_4x8/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_4x8_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_4x8_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;
@@ -114,47 +115,47 @@ void a64_smallK_hybrid_s8s32_dot_4x8_a55(const int8_t *A, int lda, const int8_t
"1:\n"
"cbnz %[odds], 2f\n"
"ldr s0, [%[a_ptr0]]\n"
- "ldr s2, [a_ptr1]\n"
- "ldr s4, [a_ptr2]\n"
- "ldr s6, [a_ptr3]\n"
- "ldr s8, [a_ptr4]\n"
- "ldr s10, [a_ptr5]\n"
- "ldr s12, [a_ptr6]\n"
- "ldr s14, [a_ptr7]\n"
+ "ldr s1, [a_ptr1]\n"
+ "ldr s2, [a_ptr2]\n"
+ "ldr s3, [a_ptr3]\n"
+ "ldr s4, [a_ptr4]\n"
+ "ldr s5, [a_ptr5]\n"
+ "ldr s6, [a_ptr6]\n"
+ "ldr s7, [a_ptr7]\n"
"b 3f\n"
"2:\n"
"subs %[odds], %[odds], #0x1\n"
"b.ne 4f\n"
"ldr b0, [%[a_ptr0]]\n"
- "ldr b2, [a_ptr1]\n"
- "ldr b4, [a_ptr2]\n"
- "ldr b6, [a_ptr3]\n"
- "ldr b8, [a_ptr4]\n"
- "ldr b10, [a_ptr5]\n"
- "ldr b12, [a_ptr6]\n"
- "ldr b14, [a_ptr7]\n"
+ "ldr b1, [a_ptr1]\n"
+ "ldr b2, [a_ptr2]\n"
+ "ldr b3, [a_ptr3]\n"
+ "ldr b4, [a_ptr4]\n"
+ "ldr b5, [a_ptr5]\n"
+ "ldr b6, [a_ptr6]\n"
+ "ldr b7, [a_ptr7]\n"
"b 3f\n"
"4:\n"
"ldr h0, [%[a_ptr0]], #0x2\n"
- "ldr h2, [a_ptr1], #0x2\n"
- "ldr h4, [a_ptr2], #0x2\n"
- "ldr h6, [a_ptr3], #0x2\n"
- "ldr h8, [a_ptr4], #0x2\n"
- "ldr h10, [a_ptr5], #0x2\n"
- "ldr h12, [a_ptr6], #0x2\n"
- "ldr h14, [a_ptr7], #0x2\n"
+ "ldr h1, [a_ptr1], #0x2\n"
+ "ldr h2, [a_ptr2], #0x2\n"
+ "ldr h3, [a_ptr3], #0x2\n"
+ "ldr h4, [a_ptr4], #0x2\n"
+ "ldr h5, [a_ptr5], #0x2\n"
+ "ldr h6, [a_ptr6], #0x2\n"
+ "ldr h7, [a_ptr7], #0x2\n"
"subs %[odds], %[odds], #0x1\n"
"b.ne 5f\n"
"b 3f\n"
"5:\n"
"ld1 {v0.b}[2], [%[a_ptr0]]\n"
- "ld1 {v2.b}[2], [a_ptr1]\n"
- "ld1 {v4.b}[2], [a_ptr2]\n"
- "ld1 {v6.b}[2], [a_ptr3]\n"
- "ld1 {v8.b}[2], [a_ptr4]\n"
- "ld1 {v10.b}[2], [a_ptr5]\n"
- "ld1 {v12.b}[2], [a_ptr6]\n"
- "ld1 {v14.b}[2], [a_ptr7]\n"
+ "ld1 {v1.b}[2], [a_ptr1]\n"
+ "ld1 {v2.b}[2], [a_ptr2]\n"
+ "ld1 {v3.b}[2], [a_ptr3]\n"
+ "ld1 {v4.b}[2], [a_ptr4]\n"
+ "ld1 {v5.b}[2], [a_ptr5]\n"
+ "ld1 {v6.b}[2], [a_ptr6]\n"
+ "ld1 {v7.b}[2], [a_ptr7]\n"
"3:\n"
"movi v24.4s, #0\n"
"ldr q16, [%[b_ptr0]]\n"
@@ -172,94 +173,118 @@ void a64_smallK_hybrid_s8s32_dot_4x8_a55(const int8_t *A, int lda, const int8_t
"prfm PLDL1KEEP, [a_ptr7, #0x180]\n"
"movi v31.4s, #0\n"
"add %[b_ptr0], %[b_ptr0], #0x10\n"
- ".word 0x4f80e218 // sdot v24.4s, v16.16b, v0.4b[0]\n"
- ".word 0x4f82e219 // sdot v25.4s, v16.16b, v2.4b[0]\n"
- ".word 0x4f84e21a // sdot v26.4s, v16.16b, v4.4b[0]\n"
- ".word 0x4f86e21b // sdot v27.4s, v16.16b, v6.4b[0]\n"
- ".word 0x4f88e21c // sdot v28.4s, v16.16b, v8.4b[0]\n"
- ".word 0x4f8ae21d // sdot v29.4s, v16.16b, v10.4b[0]\n"
- ".word 0x4f8ce21e // sdot v30.4s, v16.16b, v12.4b[0]\n"
- ".word 0x4f8ee21f // sdot v31.4s, v16.16b, v14.4b[0]\n"
+ ".inst 0x4f80e218 // sdot v24.4s, v16.16b, v0.4b[0]\n"
+ ".inst 0x4f81e219 // sdot v25.4s, v16.16b, v1.4b[0]\n"
+ ".inst 0x4f82e21a // sdot v26.4s, v16.16b, v2.4b[0]\n"
+ ".inst 0x4f83e21b // sdot v27.4s, v16.16b, v3.4b[0]\n"
+ ".inst 0x4f84e21c // sdot v28.4s, v16.16b, v4.4b[0]\n"
+ ".inst 0x4f85e21d // sdot v29.4s, v16.16b, v5.4b[0]\n"
+ ".inst 0x4f86e21e // sdot v30.4s, v16.16b, v6.4b[0]\n"
+ ".inst 0x4f87e21f // sdot v31.4s, v16.16b, v7.4b[0]\n"
"cbz %[loops], 6f\n"
"ldr d16, [%[b_ptr0]]\n"
"subs %[loops], %[loops], #0x1\n"
"ldr temploadreg0, [%[b_ptr0], #0x8]\n"
+ "add %[b_ptr0], %[b_ptr0], #0x10\n"
+ "ins v16.d[1], temploadreg0\n"
"b.eq 7f\n"
"8:\n"
- "str q24, [%[c_ptr0]], #0x10\n"
- "add %[b_ptr0], %[b_ptr0], #0x10\n"
+ "str q24, [%[c_ptr0]]\n"
+ "subs %[loops], %[loops], #0x1\n"
"movi v24.4s, #0\n"
- "ins v16.d[1], temploadreg0\n"
"ldr temploadreg0, [%[b_ptr0], #0x8]\n"
- "subs %[loops], %[loops], #0x1\n"
- "str q25, [c_ptr1], #0x10\n"
- "prfm PSTL1KEEP, [%[c_ptr0], #0x40]\n"
+ "add %[c_ptr0], %[c_ptr0], #0x10\n"
+ "str q25, [c_ptr1]\n"
+ "add c_ptr1, c_ptr1, #0x10\n"
"movi v25.4s, #0\n"
- ".word 0x4f80e218 // sdot v24.4s, v16.16b, v0.4b[0]\n"
- "str q26, [c_ptr2], #0x10\n"
+ "prfm PSTL1KEEP, [%[c_ptr0], #0x40]\n"
+ ".inst 0x4f80e218 // sdot v24.4s, v16.16b, v0.4b[0]\n"
+ "str q26, [c_ptr2]\n"
"movi v26.4s, #0\n"
- "prfm PSTL1KEEP, [c_ptr1, #0x40]\n"
- ".word 0x4f82e219 // sdot v25.4s, v16.16b, v2.4b[0]\n"
- "str q27, [c_ptr3], #0x10\n"
+ "add c_ptr2, c_ptr2, #0x10\n"
+ ".inst 0x4f81e219 // sdot v25.4s, v16.16b, v1.4b[0]\n"
+ "str q27, [c_ptr3]\n"
"movi v27.4s, #0\n"
- "prfm PSTL1KEEP, [c_ptr2, #0x40]\n"
- ".word 0x4f84e21a // sdot v26.4s, v16.16b, v4.4b[0]\n"
- "str q28, [c_ptr4], #0x10\n"
+ "add c_ptr3, c_ptr3, #0x10\n"
+ ".inst 0x4f82e21a // sdot v26.4s, v16.16b, v2.4b[0]\n"
+ "str q28, [c_ptr4]\n"
"movi v28.4s, #0\n"
- "prfm PSTL1KEEP, [c_ptr3, #0x40]\n"
- ".word 0x4f86e21b // sdot v27.4s, v16.16b, v6.4b[0]\n"
- "str q29, [c_ptr5], #0x10\n"
+ "add c_ptr4, c_ptr4, #0x10\n"
+ ".inst 0x4f83e21b // sdot v27.4s, v16.16b, v3.4b[0]\n"
+ "str q29, [c_ptr5]\n"
"movi v29.4s, #0\n"
- "prfm PSTL1KEEP, [c_ptr4, #0x40]\n"
- ".word 0x4f88e21c // sdot v28.4s, v16.16b, v8.4b[0]\n"
- "str q30, [c_ptr6], #0x10\n"
+ "add c_ptr5, c_ptr5, #0x10\n"
+ ".inst 0x4f84e21c // sdot v28.4s, v16.16b, v4.4b[0]\n"
+ "str q30, [c_ptr6]\n"
"movi v30.4s, #0\n"
- "prfm PSTL1KEEP, [c_ptr5, #0x40]\n"
- ".word 0x4f8ae21d // sdot v29.4s, v16.16b, v10.4b[0]\n"
- "str q31, [c_ptr7], #0x10\n"
+ "add c_ptr6, c_ptr6, #0x10\n"
+ ".inst 0x4f85e21d // sdot v29.4s, v16.16b, v5.4b[0]\n"
+ "str q31, [c_ptr7]\n"
"movi v31.4s, #0\n"
+ "add c_ptr7, c_ptr7, #0x10\n"
+ ".inst 0x4f86e21e // sdot v30.4s, v16.16b, v6.4b[0]\n"
+ "prfm PSTL1KEEP, [c_ptr1, #0x40]\n"
+ ".inst 0x4f87e21f // sdot v31.4s, v16.16b, v7.4b[0]\n"
+ "ldr d16, [%[b_ptr0]]\n"
+ "add %[b_ptr0], %[b_ptr0], #0x10\n"
+ "prfm PSTL1KEEP, [c_ptr2, #0x40]\n"
+ "prfm PSTL1KEEP, [c_ptr3, #0x40]\n"
+ "ins v16.d[1], temploadreg0\n"
+ "prfm PSTL1KEEP, [c_ptr4, #0x40]\n"
+ "prfm PSTL1KEEP, [c_ptr5, #0x40]\n"
"prfm PSTL1KEEP, [c_ptr6, #0x40]\n"
- ".word 0x4f8ce21e // sdot v30.4s, v16.16b, v12.4b[0]\n"
"prfm PSTL1KEEP, [c_ptr7, #0x40]\n"
- ".word 0x4f8ee21f // sdot v31.4s, v16.16b, v14.4b[0]\n"
- "ldr d16, [%[b_ptr0]]\n"
"b.ne 8b\n"
"7:\n"
- "str q24, [%[c_ptr0]], #0x10\n"
- "add %[b_ptr0], %[b_ptr0], #0x10\n"
+ "str q24, [%[c_ptr0]]\n"
+ "add %[c_ptr0], %[c_ptr0], #0x10\n"
"movi v24.4s, #0\n"
- "ins v16.d[1], temploadreg0\n"
- "str q25, [c_ptr1], #0x10\n"
+ "str q25, [c_ptr1]\n"
+ "add c_ptr1, c_ptr1, #0x10\n"
"movi v25.4s, #0\n"
- ".word 0x4f80e218 // sdot v24.4s, v16.16b, v0.4b[0]\n"
- "str q26, [c_ptr2], #0x10\n"
+ ".inst 0x4f80e218 // sdot v24.4s, v16.16b, v0.4b[0]\n"
+ "str q26, [c_ptr2]\n"
"movi v26.4s, #0\n"
- ".word 0x4f82e219 // sdot v25.4s, v16.16b, v2.4b[0]\n"
- "str q27, [c_ptr3], #0x10\n"
+ "add c_ptr2, c_ptr2, #0x10\n"
+ ".inst 0x4f81e219 // sdot v25.4s, v16.16b, v1.4b[0]\n"
+ "str q27, [c_ptr3]\n"
"movi v27.4s, #0\n"
- ".word 0x4f84e21a // sdot v26.4s, v16.16b, v4.4b[0]\n"
- "str q28, [c_ptr4], #0x10\n"
+ "add c_ptr3, c_ptr3, #0x10\n"
+ ".inst 0x4f82e21a // sdot v26.4s, v16.16b, v2.4b[0]\n"
+ "str q28, [c_ptr4]\n"
"movi v28.4s, #0\n"
- ".word 0x4f86e21b // sdot v27.4s, v16.16b, v6.4b[0]\n"
- "str q29, [c_ptr5], #0x10\n"
+ "add c_ptr4, c_ptr4, #0x10\n"
+ ".inst 0x4f83e21b // sdot v27.4s, v16.16b, v3.4b[0]\n"
+ "str q29, [c_ptr5]\n"
"movi v29.4s, #0\n"
- ".word 0x4f88e21c // sdot v28.4s, v16.16b, v8.4b[0]\n"
- "str q30, [c_ptr6], #0x10\n"
+ "add c_ptr5, c_ptr5, #0x10\n"
+ ".inst 0x4f84e21c // sdot v28.4s, v16.16b, v4.4b[0]\n"
+ "str q30, [c_ptr6]\n"
"movi v30.4s, #0\n"
- ".word 0x4f8ae21d // sdot v29.4s, v16.16b, v10.4b[0]\n"
- "str q31, [c_ptr7], #0x10\n"
+ "add c_ptr6, c_ptr6, #0x10\n"
+ ".inst 0x4f85e21d // sdot v29.4s, v16.16b, v5.4b[0]\n"
+ "str q31, [c_ptr7]\n"
"movi v31.4s, #0\n"
- ".word 0x4f8ce21e // sdot v30.4s, v16.16b, v12.4b[0]\n"
- ".word 0x4f8ee21f // sdot v31.4s, v16.16b, v14.4b[0]\n"
+ "add c_ptr7, c_ptr7, #0x10\n"
+ ".inst 0x4f86e21e // sdot v30.4s, v16.16b, v6.4b[0]\n"
+ ".inst 0x4f87e21f // sdot v31.4s, v16.16b, v7.4b[0]\n"
"6:\n"
"str q24, [%[c_ptr0]]\n"
+ "add %[c_ptr0], %[c_ptr0], #0x10\n"
"str q25, [c_ptr1]\n"
+ "add c_ptr1, c_ptr1, #0x10\n"
"str q26, [c_ptr2]\n"
+ "add c_ptr2, c_ptr2, #0x10\n"
"str q27, [c_ptr3]\n"
+ "add c_ptr3, c_ptr3, #0x10\n"
"str q28, [c_ptr4]\n"
+ "add c_ptr4, c_ptr4, #0x10\n"
"str q29, [c_ptr5]\n"
+ "add c_ptr5, c_ptr5, #0x10\n"
"str q30, [c_ptr6]\n"
+ "add c_ptr6, c_ptr6, #0x10\n"
"str q31, [c_ptr7]\n"
+ "add c_ptr7, c_ptr7, #0x10\n"
".unreq a_ptr1\n"
".unreq a_ptr2\n"
".unreq a_ptr3\n"
@@ -348,55 +373,55 @@ void a64_smallK_hybrid_s8s32_dot_4x8_a55(const int8_t *A, int lda, const int8_t
"1:\n"
"cbnz %[odds], 2f\n"
"ldr d0, [%[a_ptr0]]\n"
- "ldr d2, [a_ptr1]\n"
- "ldr d4, [a_ptr2]\n"
- "ldr d6, [a_ptr3]\n"
- "ldr d8, [a_ptr4]\n"
- "ldr d10, [a_ptr5]\n"
- "ldr d12, [a_ptr6]\n"
- "ldr d14, [a_ptr7]\n"
+ "ldr d1, [a_ptr1]\n"
+ "ldr d2, [a_ptr2]\n"
+ "ldr d3, [a_ptr3]\n"
+ "ldr d4, [a_ptr4]\n"
+ "ldr d5, [a_ptr5]\n"
+ "ldr d6, [a_ptr6]\n"
+ "ldr d7, [a_ptr7]\n"
"b 3f\n"
"2:\n"
"ldr s0, [%[a_ptr0]], #0x4\n"
- "ldr s2, [a_ptr1], #0x4\n"
- "ldr s4, [a_ptr2], #0x4\n"
- "ldr s6, [a_ptr3], #0x4\n"
- "ldr s8, [a_ptr4], #0x4\n"
- "ldr s10, [a_ptr5], #0x4\n"
- "ldr s12, [a_ptr6], #0x4\n"
- "ldr s14, [a_ptr7], #0x4\n"
+ "ldr s1, [a_ptr1], #0x4\n"
+ "ldr s2, [a_ptr2], #0x4\n"
+ "ldr s3, [a_ptr3], #0x4\n"
+ "ldr s4, [a_ptr4], #0x4\n"
+ "ldr s5, [a_ptr5], #0x4\n"
+ "ldr s6, [a_ptr6], #0x4\n"
+ "ldr s7, [a_ptr7], #0x4\n"
"subs %[odds], %[odds], #0x1\n"
"b.ne 4f\n"
"ld1 {v0.b}[4], [%[a_ptr0]]\n"
- "ld1 {v2.b}[4], [a_ptr1]\n"
- "ld1 {v4.b}[4], [a_ptr2]\n"
- "ld1 {v6.b}[4], [a_ptr3]\n"
- "ld1 {v8.b}[4], [a_ptr4]\n"
- "ld1 {v10.b}[4], [a_ptr5]\n"
- "ld1 {v12.b}[4], [a_ptr6]\n"
- "ld1 {v14.b}[4], [a_ptr7]\n"
+ "ld1 {v1.b}[4], [a_ptr1]\n"
+ "ld1 {v2.b}[4], [a_ptr2]\n"
+ "ld1 {v3.b}[4], [a_ptr3]\n"
+ "ld1 {v4.b}[4], [a_ptr4]\n"
+ "ld1 {v5.b}[4], [a_ptr5]\n"
+ "ld1 {v6.b}[4], [a_ptr6]\n"
+ "ld1 {v7.b}[4], [a_ptr7]\n"
"b 3f\n"
"4:\n"
"ld1 {v0.h}[2], [%[a_ptr0]], #2\n"
- "ld1 {v2.h}[2], [a_ptr1], #2\n"
- "ld1 {v4.h}[2], [a_ptr2], #2\n"
- "ld1 {v6.h}[2], [a_ptr3], #2\n"
- "ld1 {v8.h}[2], [a_ptr4], #2\n"
- "ld1 {v10.h}[2], [a_ptr5], #2\n"
- "ld1 {v12.h}[2], [a_ptr6], #2\n"
- "ld1 {v14.h}[2], [a_ptr7], #2\n"
+ "ld1 {v1.h}[2], [a_ptr1], #2\n"
+ "ld1 {v2.h}[2], [a_ptr2], #2\n"
+ "ld1 {v3.h}[2], [a_ptr3], #2\n"
+ "ld1 {v4.h}[2], [a_ptr4], #2\n"
+ "ld1 {v5.h}[2], [a_ptr5], #2\n"
+ "ld1 {v6.h}[2], [a_ptr6], #2\n"
+ "ld1 {v7.h}[2], [a_ptr7], #2\n"
"subs %[odds], %[odds], #0x1\n"
"b.ne 5f\n"
"b 3f\n"
"5:\n"
"ld1 {v0.b}[6], [%[a_ptr0]]\n"
- "ld1 {v2.b}[6], [a_ptr1]\n"
- "ld1 {v4.b}[6], [a_ptr2]\n"
- "ld1 {v6.b}[6], [a_ptr3]\n"
- "ld1 {v8.b}[6], [a_ptr4]\n"
- "ld1 {v10.b}[6], [a_ptr5]\n"
- "ld1 {v12.b}[6], [a_ptr6]\n"
- "ld1 {v14.b}[6], [a_ptr7]\n"
+ "ld1 {v1.b}[6], [a_ptr1]\n"
+ "ld1 {v2.b}[6], [a_ptr2]\n"
+ "ld1 {v3.b}[6], [a_ptr3]\n"
+ "ld1 {v4.b}[6], [a_ptr4]\n"
+ "ld1 {v5.b}[6], [a_ptr5]\n"
+ "ld1 {v6.b}[6], [a_ptr6]\n"
+ "ld1 {v7.b}[6], [a_ptr7]\n"
"3:\n"
"movi v24.4s, #0\n"
"ldr q16, [%[b_ptr0]]\n"
@@ -414,23 +439,23 @@ void a64_smallK_hybrid_s8s32_dot_4x8_a55(const int8_t *A, int lda, const int8_t
"prfm PLDL1KEEP, [a_ptr7, #0x140]\n"
"movi v31.4s, #0\n"
"prfm PLDL1KEEP, [a_ptr7, #0x180]\n"
- ".word 0x4f80e218 // sdot v24.4s, v16.16b, v0.4b[0]\n"
+ ".inst 0x4f80e218 // sdot v24.4s, v16.16b, v0.4b[0]\n"
"add %[b_ptr0], %[b_ptr0], #0x20\n"
- ".word 0x4f82e219 // sdot v25.4s, v16.16b, v2.4b[0]\n"
- ".word 0x4f84e21a // sdot v26.4s, v16.16b, v4.4b[0]\n"
- ".word 0x4f86e21b // sdot v27.4s, v16.16b, v6.4b[0]\n"
- ".word 0x4f88e21c // sdot v28.4s, v16.16b, v8.4b[0]\n"
- ".word 0x4f8ae21d // sdot v29.4s, v16.16b, v10.4b[0]\n"
- ".word 0x4f8ce21e // sdot v30.4s, v16.16b, v12.4b[0]\n"
- ".word 0x4f8ee21f // sdot v31.4s, v16.16b, v14.4b[0]\n"
- ".word 0x4fa0e238 // sdot v24.4s, v17.16b, v0.4b[1]\n"
- ".word 0x4fa2e239 // sdot v25.4s, v17.16b, v2.4b[1]\n"
- ".word 0x4fa4e23a // sdot v26.4s, v17.16b, v4.4b[1]\n"
- ".word 0x4fa6e23b // sdot v27.4s, v17.16b, v6.4b[1]\n"
- ".word 0x4fa8e23c // sdot v28.4s, v17.16b, v8.4b[1]\n"
- ".word 0x4faae23d // sdot v29.4s, v17.16b, v10.4b[1]\n"
- ".word 0x4face23e // sdot v30.4s, v17.16b, v12.4b[1]\n"
- ".word 0x4faee23f // sdot v31.4s, v17.16b, v14.4b[1]\n"
+ ".inst 0x4f81e219 // sdot v25.4s, v16.16b, v1.4b[0]\n"
+ ".inst 0x4f82e21a // sdot v26.4s, v16.16b, v2.4b[0]\n"
+ ".inst 0x4f83e21b // sdot v27.4s, v16.16b, v3.4b[0]\n"
+ ".inst 0x4f84e21c // sdot v28.4s, v16.16b, v4.4b[0]\n"
+ ".inst 0x4f85e21d // sdot v29.4s, v16.16b, v5.4b[0]\n"
+ ".inst 0x4f86e21e // sdot v30.4s, v16.16b, v6.4b[0]\n"
+ ".inst 0x4f87e21f // sdot v31.4s, v16.16b, v7.4b[0]\n"
+ ".inst 0x4fa0e238 // sdot v24.4s, v17.16b, v0.4b[1]\n"
+ ".inst 0x4fa1e239 // sdot v25.4s, v17.16b, v1.4b[1]\n"
+ ".inst 0x4fa2e23a // sdot v26.4s, v17.16b, v2.4b[1]\n"
+ ".inst 0x4fa3e23b // sdot v27.4s, v17.16b, v3.4b[1]\n"
+ ".inst 0x4fa4e23c // sdot v28.4s, v17.16b, v4.4b[1]\n"
+ ".inst 0x4fa5e23d // sdot v29.4s, v17.16b, v5.4b[1]\n"
+ ".inst 0x4fa6e23e // sdot v30.4s, v17.16b, v6.4b[1]\n"
+ ".inst 0x4fa7e23f // sdot v31.4s, v17.16b, v7.4b[1]\n"
"cbz %[loops], 6f\n"
"ldr d16, [%[b_ptr0]]\n"
"subs %[loops], %[loops], #0x1\n"
@@ -441,98 +466,122 @@ void a64_smallK_hybrid_s8s32_dot_4x8_a55(const int8_t *A, int lda, const int8_t
"ins v16.d[1], temploadreg0\n"
"b.eq 7f\n"
"8:\n"
- "str q24, [%[c_ptr0]], #0x10\n"
+ "str q24, [%[c_ptr0]]\n"
"subs %[loops], %[loops], #0x1\n"
"movi v24.4s, #0\n"
"ins v17.d[1], temploadreg1\n"
"ldr temploadreg0, [%[b_ptr0], #0x8]\n"
- "prfm PSTL1KEEP, [%[c_ptr0], #0x40]\n"
- "str q25, [c_ptr1], #0x10\n"
+ "add %[c_ptr0], %[c_ptr0], #0x10\n"
+ "str q25, [c_ptr1]\n"
+ "add c_ptr1, c_ptr1, #0x10\n"
"movi v25.4s, #0\n"
"ldr temploadreg1, [%[b_ptr0], #0x18]\n"
- ".word 0x4f80e218 // sdot v24.4s, v16.16b, v0.4b[0]\n"
- "str q26, [c_ptr2], #0x10\n"
- "prfm PSTL1KEEP, [c_ptr1, #0x40]\n"
+ ".inst 0x4f80e218 // sdot v24.4s, v16.16b, v0.4b[0]\n"
+ "str q26, [c_ptr2]\n"
"movi v26.4s, #0\n"
- ".word 0x4f82e219 // sdot v25.4s, v16.16b, v2.4b[0]\n"
- "str q27, [c_ptr3], #0x10\n"
+ "prfm PSTL1KEEP, [%[c_ptr0], #0x40]\n"
+ ".inst 0x4f81e219 // sdot v25.4s, v16.16b, v1.4b[0]\n"
+ "str q27, [c_ptr3]\n"
"movi v27.4s, #0\n"
- "prfm PSTL1KEEP, [c_ptr2, #0x40]\n"
- ".word 0x4f84e21a // sdot v26.4s, v16.16b, v4.4b[0]\n"
- "str q28, [c_ptr4], #0x10\n"
+ "add c_ptr2, c_ptr2, #0x10\n"
+ ".inst 0x4f82e21a // sdot v26.4s, v16.16b, v2.4b[0]\n"
+ "str q28, [c_ptr4]\n"
"movi v28.4s, #0\n"
- "prfm PSTL1KEEP, [c_ptr3, #0x40]\n"
- ".word 0x4f86e21b // sdot v27.4s, v16.16b, v6.4b[0]\n"
- "str q29, [c_ptr5], #0x10\n"
+ "add c_ptr3, c_ptr3, #0x10\n"
+ ".inst 0x4f83e21b // sdot v27.4s, v16.16b, v3.4b[0]\n"
+ "str q29, [c_ptr5]\n"
"movi v29.4s, #0\n"
- "prfm PSTL1KEEP, [c_ptr4, #0x40]\n"
- ".word 0x4f88e21c // sdot v28.4s, v16.16b, v8.4b[0]\n"
- "str q30, [c_ptr6], #0x10\n"
+ "add c_ptr4, c_ptr4, #0x10\n"
+ ".inst 0x4f84e21c // sdot v28.4s, v16.16b, v4.4b[0]\n"
+ "str q30, [c_ptr6]\n"
"movi v30.4s, #0\n"
- "prfm PSTL1KEEP, [c_ptr5, #0x40]\n"
- ".word 0x4f8ae21d // sdot v29.4s, v16.16b, v10.4b[0]\n"
- "str q31, [c_ptr7], #0x10\n"
+ "add c_ptr5, c_ptr5, #0x10\n"
+ ".inst 0x4f85e21d // sdot v29.4s, v16.16b, v5.4b[0]\n"
+ "str q31, [c_ptr7]\n"
"movi v31.4s, #0\n"
- "prfm PSTL1KEEP, [c_ptr6, #0x40]\n"
- ".word 0x4f8ce21e // sdot v30.4s, v16.16b, v12.4b[0]\n"
- "prfm PSTL1KEEP, [c_ptr7, #0x40]\n"
- ".word 0x4f8ee21f // sdot v31.4s, v16.16b, v14.4b[0]\n"
+ "add c_ptr6, c_ptr6, #0x10\n"
+ ".inst 0x4f86e21e // sdot v30.4s, v16.16b, v6.4b[0]\n"
+ "add c_ptr7, c_ptr7, #0x10\n"
+ ".inst 0x4f87e21f // sdot v31.4s, v16.16b, v7.4b[0]\n"
"ldr d16, [%[b_ptr0]]\n"
- ".word 0x4fa0e238 // sdot v24.4s, v17.16b, v0.4b[1]\n"
- ".word 0x4fa2e239 // sdot v25.4s, v17.16b, v2.4b[1]\n"
- ".word 0x4fa4e23a // sdot v26.4s, v17.16b, v4.4b[1]\n"
+ ".inst 0x4fa0e238 // sdot v24.4s, v17.16b, v0.4b[1]\n"
+ "prfm PSTL1KEEP, [c_ptr1, #0x40]\n"
+ ".inst 0x4fa1e239 // sdot v25.4s, v17.16b, v1.4b[1]\n"
"ins v16.d[1], temploadreg0\n"
- ".word 0x4fa6e23b // sdot v27.4s, v17.16b, v6.4b[1]\n"
- ".word 0x4fa8e23c // sdot v28.4s, v17.16b, v8.4b[1]\n"
- ".word 0x4faae23d // sdot v29.4s, v17.16b, v10.4b[1]\n"
- ".word 0x4face23e // sdot v30.4s, v17.16b, v12.4b[1]\n"
- ".word 0x4faee23f // sdot v31.4s, v17.16b, v14.4b[1]\n"
+ ".inst 0x4fa2e23a // sdot v26.4s, v17.16b, v2.4b[1]\n"
+ "prfm PSTL1KEEP, [c_ptr2, #0x40]\n"
+ ".inst 0x4fa3e23b // sdot v27.4s, v17.16b, v3.4b[1]\n"
+ "prfm PSTL1KEEP, [c_ptr3, #0x40]\n"
+ ".inst 0x4fa4e23c // sdot v28.4s, v17.16b, v4.4b[1]\n"
+ "prfm PSTL1KEEP, [c_ptr4, #0x40]\n"
+ ".inst 0x4fa5e23d // sdot v29.4s, v17.16b, v5.4b[1]\n"
+ "prfm PSTL1KEEP, [c_ptr5, #0x40]\n"
+ ".inst 0x4fa6e23e // sdot v30.4s, v17.16b, v6.4b[1]\n"
+ "prfm PSTL1KEEP, [c_ptr6, #0x40]\n"
+ ".inst 0x4fa7e23f // sdot v31.4s, v17.16b, v7.4b[1]\n"
"ldr d17, [%[b_ptr0], #0x10]\n"
+ "prfm PSTL1KEEP, [c_ptr7, #0x40]\n"
"add %[b_ptr0], %[b_ptr0], #0x20\n"
"b.ne 8b\n"
"7:\n"
- "str q24, [%[c_ptr0]], #0x10\n"
+ "str q24, [%[c_ptr0]]\n"
+ "add %[c_ptr0], %[c_ptr0], #0x10\n"
"movi v24.4s, #0\n"
"ins v17.d[1], temploadreg1\n"
- "str q25, [c_ptr1], #0x10\n"
+ "str q25, [c_ptr1]\n"
+ "add c_ptr1, c_ptr1, #0x10\n"
"movi v25.4s, #0\n"
- ".word 0x4f80e218 // sdot v24.4s, v16.16b, v0.4b[0]\n"
- "str q26, [c_ptr2], #0x10\n"
+ ".inst 0x4f80e218 // sdot v24.4s, v16.16b, v0.4b[0]\n"
+ "str q26, [c_ptr2]\n"
"movi v26.4s, #0\n"
- ".word 0x4f82e219 // sdot v25.4s, v16.16b, v2.4b[0]\n"
- ".word 0x4fa0e238 // sdot v24.4s, v17.16b, v0.4b[1]\n"
- "str q27, [c_ptr3], #0x10\n"
+ "add c_ptr2, c_ptr2, #0x10\n"
+ ".inst 0x4f81e219 // sdot v25.4s, v16.16b, v1.4b[0]\n"
+ "str q27, [c_ptr3]\n"
"movi v27.4s, #0\n"
- ".word 0x4f84e21a // sdot v26.4s, v16.16b, v4.4b[0]\n"
- ".word 0x4fa2e239 // sdot v25.4s, v17.16b, v2.4b[1]\n"
- "str q28, [c_ptr4], #0x10\n"
+ "add c_ptr3, c_ptr3, #0x10\n"
+ ".inst 0x4f82e21a // sdot v26.4s, v16.16b, v2.4b[0]\n"
+ "str q28, [c_ptr4]\n"
"movi v28.4s, #0\n"
- ".word 0x4f86e21b // sdot v27.4s, v16.16b, v6.4b[0]\n"
- ".word 0x4fa4e23a // sdot v26.4s, v17.16b, v4.4b[1]\n"
- "str q29, [c_ptr5], #0x10\n"
+ "add c_ptr4, c_ptr4, #0x10\n"
+ ".inst 0x4f83e21b // sdot v27.4s, v16.16b, v3.4b[0]\n"
+ "str q29, [c_ptr5]\n"
"movi v29.4s, #0\n"
- ".word 0x4f88e21c // sdot v28.4s, v16.16b, v8.4b[0]\n"
- ".word 0x4fa6e23b // sdot v27.4s, v17.16b, v6.4b[1]\n"
- "str q30, [c_ptr6], #0x10\n"
+ "add c_ptr5, c_ptr5, #0x10\n"
+ ".inst 0x4f84e21c // sdot v28.4s, v16.16b, v4.4b[0]\n"
+ "str q30, [c_ptr6]\n"
"movi v30.4s, #0\n"
- ".word 0x4f8ae21d // sdot v29.4s, v16.16b, v10.4b[0]\n"
- ".word 0x4fa8e23c // sdot v28.4s, v17.16b, v8.4b[1]\n"
- "str q31, [c_ptr7], #0x10\n"
+ "add c_ptr6, c_ptr6, #0x10\n"
+ ".inst 0x4f85e21d // sdot v29.4s, v16.16b, v5.4b[0]\n"
+ "str q31, [c_ptr7]\n"
"movi v31.4s, #0\n"
- ".word 0x4f8ce21e // sdot v30.4s, v16.16b, v12.4b[0]\n"
- ".word 0x4faae23d // sdot v29.4s, v17.16b, v10.4b[1]\n"
- ".word 0x4f8ee21f // sdot v31.4s, v16.16b, v14.4b[0]\n"
- ".word 0x4face23e // sdot v30.4s, v17.16b, v12.4b[1]\n"
- ".word 0x4faee23f // sdot v31.4s, v17.16b, v14.4b[1]\n"
+ "add c_ptr7, c_ptr7, #0x10\n"
+ ".inst 0x4f86e21e // sdot v30.4s, v16.16b, v6.4b[0]\n"
+ ".inst 0x4fa0e238 // sdot v24.4s, v17.16b, v0.4b[1]\n"
+ ".inst 0x4f87e21f // sdot v31.4s, v16.16b, v7.4b[0]\n"
+ ".inst 0x4fa1e239 // sdot v25.4s, v17.16b, v1.4b[1]\n"
+ ".inst 0x4fa2e23a // sdot v26.4s, v17.16b, v2.4b[1]\n"
+ ".inst 0x4fa3e23b // sdot v27.4s, v17.16b, v3.4b[1]\n"
+ ".inst 0x4fa4e23c // sdot v28.4s, v17.16b, v4.4b[1]\n"
+ ".inst 0x4fa5e23d // sdot v29.4s, v17.16b, v5.4b[1]\n"
+ ".inst 0x4fa6e23e // sdot v30.4s, v17.16b, v6.4b[1]\n"
+ ".inst 0x4fa7e23f // sdot v31.4s, v17.16b, v7.4b[1]\n"
"6:\n"
"str q24, [%[c_ptr0]]\n"
+ "add %[c_ptr0], %[c_ptr0], #0x10\n"
"str q25, [c_ptr1]\n"
+ "add c_ptr1, c_ptr1, #0x10\n"
"str q26, [c_ptr2]\n"
+ "add c_ptr2, c_ptr2, #0x10\n"
"str q27, [c_ptr3]\n"
+ "add c_ptr3, c_ptr3, #0x10\n"
"str q28, [c_ptr4]\n"
+ "add c_ptr4, c_ptr4, #0x10\n"
"str q29, [c_ptr5]\n"
+ "add c_ptr5, c_ptr5, #0x10\n"
"str q30, [c_ptr6]\n"
+ "add c_ptr6, c_ptr6, #0x10\n"
"str q31, [c_ptr7]\n"
+ "add c_ptr7, c_ptr7, #0x10\n"
".unreq a_ptr1\n"
".unreq a_ptr2\n"
".unreq a_ptr3\n"
@@ -620,56 +669,56 @@ void a64_smallK_hybrid_s8s32_dot_4x8_a55(const int8_t *A, int lda, const int8_t
"add a_ptr1, %[a_ptr0], #0x0\n"
"1:\n"
"ldr d0, [%[a_ptr0]], #0x8\n"
- "ldr d2, [a_ptr1], #0x8\n"
- "ldr d4, [a_ptr2], #0x8\n"
- "ldr d6, [a_ptr3], #0x8\n"
- "ldr d8, [a_ptr4], #0x8\n"
- "ldr d10, [a_ptr5], #0x8\n"
- "ldr d12, [a_ptr6], #0x8\n"
- "ldr d14, [a_ptr7], #0x8\n"
+ "ldr d1, [a_ptr1], #0x8\n"
+ "ldr d2, [a_ptr2], #0x8\n"
+ "ldr d3, [a_ptr3], #0x8\n"
+ "ldr d4, [a_ptr4], #0x8\n"
+ "ldr d5, [a_ptr5], #0x8\n"
+ "ldr d6, [a_ptr6], #0x8\n"
+ "ldr d7, [a_ptr7], #0x8\n"
"cbnz %[odds], 2f\n"
"ld1 {v0.s}[2], [%[a_ptr0]]\n"
- "ld1 {v2.s}[2], [a_ptr1]\n"
- "ld1 {v4.s}[2], [a_ptr2]\n"
- "ld1 {v6.s}[2], [a_ptr3]\n"
- "ld1 {v8.s}[2], [a_ptr4]\n"
- "ld1 {v10.s}[2], [a_ptr5]\n"
- "ld1 {v12.s}[2], [a_ptr6]\n"
- "ld1 {v14.s}[2], [a_ptr7]\n"
+ "ld1 {v1.s}[2], [a_ptr1]\n"
+ "ld1 {v2.s}[2], [a_ptr2]\n"
+ "ld1 {v3.s}[2], [a_ptr3]\n"
+ "ld1 {v4.s}[2], [a_ptr4]\n"
+ "ld1 {v5.s}[2], [a_ptr5]\n"
+ "ld1 {v6.s}[2], [a_ptr6]\n"
+ "ld1 {v7.s}[2], [a_ptr7]\n"
"b 3f\n"
"2:\n"
"subs %[odds], %[odds], #0x1\n"
"b.ne 4f\n"
"ld1 {v0.b}[8], [%[a_ptr0]]\n"
- "ld1 {v2.b}[8], [a_ptr1]\n"
- "ld1 {v4.b}[8], [a_ptr2]\n"
- "ld1 {v6.b}[8], [a_ptr3]\n"
- "ld1 {v8.b}[8], [a_ptr4]\n"
- "ld1 {v10.b}[8], [a_ptr5]\n"
- "ld1 {v12.b}[8], [a_ptr6]\n"
- "ld1 {v14.b}[8], [a_ptr7]\n"
+ "ld1 {v1.b}[8], [a_ptr1]\n"
+ "ld1 {v2.b}[8], [a_ptr2]\n"
+ "ld1 {v3.b}[8], [a_ptr3]\n"
+ "ld1 {v4.b}[8], [a_ptr4]\n"
+ "ld1 {v5.b}[8], [a_ptr5]\n"
+ "ld1 {v6.b}[8], [a_ptr6]\n"
+ "ld1 {v7.b}[8], [a_ptr7]\n"
"b 3f\n"
"4:\n"
"ld1 {v0.h}[4], [%[a_ptr0]], #2\n"
- "ld1 {v2.h}[4], [a_ptr1], #2\n"
- "ld1 {v4.h}[4], [a_ptr2], #2\n"
- "ld1 {v6.h}[4], [a_ptr3], #2\n"
- "ld1 {v8.h}[4], [a_ptr4], #2\n"
- "ld1 {v10.h}[4], [a_ptr5], #2\n"
- "ld1 {v12.h}[4], [a_ptr6], #2\n"
- "ld1 {v14.h}[4], [a_ptr7], #2\n"
+ "ld1 {v1.h}[4], [a_ptr1], #2\n"
+ "ld1 {v2.h}[4], [a_ptr2], #2\n"
+ "ld1 {v3.h}[4], [a_ptr3], #2\n"
+ "ld1 {v4.h}[4], [a_ptr4], #2\n"
+ "ld1 {v5.h}[4], [a_ptr5], #2\n"
+ "ld1 {v6.h}[4], [a_ptr6], #2\n"
+ "ld1 {v7.h}[4], [a_ptr7], #2\n"
"subs %[odds], %[odds], #0x1\n"
"b.ne 5f\n"
"b 3f\n"
"5:\n"
"ld1 {v0.b}[10], [%[a_ptr0]]\n"
- "ld1 {v2.b}[10], [a_ptr1]\n"
- "ld1 {v4.b}[10], [a_ptr2]\n"
- "ld1 {v6.b}[10], [a_ptr3]\n"
- "ld1 {v8.b}[10], [a_ptr4]\n"
- "ld1 {v10.b}[10], [a_ptr5]\n"
- "ld1 {v12.b}[10], [a_ptr6]\n"
- "ld1 {v14.b}[10], [a_ptr7]\n"
+ "ld1 {v1.b}[10], [a_ptr1]\n"
+ "ld1 {v2.b}[10], [a_ptr2]\n"
+ "ld1 {v3.b}[10], [a_ptr3]\n"
+ "ld1 {v4.b}[10], [a_ptr4]\n"
+ "ld1 {v5.b}[10], [a_ptr5]\n"
+ "ld1 {v6.b}[10], [a_ptr6]\n"
+ "ld1 {v7.b}[10], [a_ptr7]\n"
"3:\n"
"movi v24.4s, #0\n"
"ldr q16, [%[b_ptr0]]\n"
@@ -687,32 +736,32 @@ void a64_smallK_hybrid_s8s32_dot_4x8_a55(const int8_t *A, int lda, const int8_t
"prfm PLDL1KEEP, [a_ptr7, #0x100]\n"
"movi v31.4s, #0\n"
"prfm PLDL1KEEP, [a_ptr7, #0x140]\n"
- ".word 0x4f80e218 // sdot v24.4s, v16.16b, v0.4b[0]\n"
+ ".inst 0x4f80e218 // sdot v24.4s, v16.16b, v0.4b[0]\n"
"prfm PLDL1KEEP, [a_ptr7, #0x180]\n"
- ".word 0x4f82e219 // sdot v25.4s, v16.16b, v2.4b[0]\n"
+ ".inst 0x4f81e219 // sdot v25.4s, v16.16b, v1.4b[0]\n"
"add %[b_ptr0], %[b_ptr0], #0x30\n"
- ".word 0x4f84e21a // sdot v26.4s, v16.16b, v4.4b[0]\n"
- ".word 0x4f86e21b // sdot v27.4s, v16.16b, v6.4b[0]\n"
- ".word 0x4f88e21c // sdot v28.4s, v16.16b, v8.4b[0]\n"
- ".word 0x4f8ae21d // sdot v29.4s, v16.16b, v10.4b[0]\n"
- ".word 0x4f8ce21e // sdot v30.4s, v16.16b, v12.4b[0]\n"
- ".word 0x4f8ee21f // sdot v31.4s, v16.16b, v14.4b[0]\n"
- ".word 0x4fa0e238 // sdot v24.4s, v17.16b, v0.4b[1]\n"
- ".word 0x4fa2e239 // sdot v25.4s, v17.16b, v2.4b[1]\n"
- ".word 0x4fa4e23a // sdot v26.4s, v17.16b, v4.4b[1]\n"
- ".word 0x4fa6e23b // sdot v27.4s, v17.16b, v6.4b[1]\n"
- ".word 0x4fa8e23c // sdot v28.4s, v17.16b, v8.4b[1]\n"
- ".word 0x4faae23d // sdot v29.4s, v17.16b, v10.4b[1]\n"
- ".word 0x4face23e // sdot v30.4s, v17.16b, v12.4b[1]\n"
- ".word 0x4faee23f // sdot v31.4s, v17.16b, v14.4b[1]\n"
- ".word 0x4f80ea58 // sdot v24.4s, v18.16b, v0.4b[2]\n"
- ".word 0x4f82ea59 // sdot v25.4s, v18.16b, v2.4b[2]\n"
- ".word 0x4f84ea5a // sdot v26.4s, v18.16b, v4.4b[2]\n"
- ".word 0x4f86ea5b // sdot v27.4s, v18.16b, v6.4b[2]\n"
- ".word 0x4f88ea5c // sdot v28.4s, v18.16b, v8.4b[2]\n"
- ".word 0x4f8aea5d // sdot v29.4s, v18.16b, v10.4b[2]\n"
- ".word 0x4f8cea5e // sdot v30.4s, v18.16b, v12.4b[2]\n"
- ".word 0x4f8eea5f // sdot v31.4s, v18.16b, v14.4b[2]\n"
+ ".inst 0x4f82e21a // sdot v26.4s, v16.16b, v2.4b[0]\n"
+ ".inst 0x4f83e21b // sdot v27.4s, v16.16b, v3.4b[0]\n"
+ ".inst 0x4f84e21c // sdot v28.4s, v16.16b, v4.4b[0]\n"
+ ".inst 0x4f85e21d // sdot v29.4s, v16.16b, v5.4b[0]\n"
+ ".inst 0x4f86e21e // sdot v30.4s, v16.16b, v6.4b[0]\n"
+ ".inst 0x4f87e21f // sdot v31.4s, v16.16b, v7.4b[0]\n"
+ ".inst 0x4fa0e238 // sdot v24.4s, v17.16b, v0.4b[1]\n"
+ ".inst 0x4fa1e239 // sdot v25.4s, v17.16b, v1.4b[1]\n"
+ ".inst 0x4fa2e23a // sdot v26.4s, v17.16b, v2.4b[1]\n"
+ ".inst 0x4fa3e23b // sdot v27.4s, v17.16b, v3.4b[1]\n"
+ ".inst 0x4fa4e23c // sdot v28.4s, v17.16b, v4.4b[1]\n"
+ ".inst 0x4fa5e23d // sdot v29.4s, v17.16b, v5.4b[1]\n"
+ ".inst 0x4fa6e23e // sdot v30.4s, v17.16b, v6.4b[1]\n"
+ ".inst 0x4fa7e23f // sdot v31.4s, v17.16b, v7.4b[1]\n"
+ ".inst 0x4f80ea58 // sdot v24.4s, v18.16b, v0.4b[2]\n"
+ ".inst 0x4f81ea59 // sdot v25.4s, v18.16b, v1.4b[2]\n"
+ ".inst 0x4f82ea5a // sdot v26.4s, v18.16b, v2.4b[2]\n"
+ ".inst 0x4f83ea5b // sdot v27.4s, v18.16b, v3.4b[2]\n"
+ ".inst 0x4f84ea5c // sdot v28.4s, v18.16b, v4.4b[2]\n"
+ ".inst 0x4f85ea5d // sdot v29.4s, v18.16b, v5.4b[2]\n"
+ ".inst 0x4f86ea5e // sdot v30.4s, v18.16b, v6.4b[2]\n"
+ ".inst 0x4f87ea5f // sdot v31.4s, v18.16b, v7.4b[2]\n"
"cbz %[loops], 6f\n"
"ldr d16, [%[b_ptr0]]\n"
"subs %[loops], %[loops], #0x1\n"
@@ -726,117 +775,141 @@ void a64_smallK_hybrid_s8s32_dot_4x8_a55(const int8_t *A, int lda, const int8_t
"ins v17.d[1], temploadreg1\n"
"b.eq 7f\n"
"8:\n"
- "str q24, [%[c_ptr0]], #0x10\n"
+ "str q24, [%[c_ptr0]]\n"
"subs %[loops], %[loops], #0x1\n"
"movi v24.4s, #0\n"
"ins v18.d[1], temploadreg2\n"
"ldr temploadreg0, [%[b_ptr0], #0x8]\n"
- "prfm PSTL1KEEP, [%[c_ptr0], #0x40]\n"
- "str q25, [c_ptr1], #0x10\n"
+ "add %[c_ptr0], %[c_ptr0], #0x10\n"
+ "str q25, [c_ptr1]\n"
+ "add c_ptr1, c_ptr1, #0x10\n"
"movi v25.4s, #0\n"
"ldr temploadreg1, [%[b_ptr0], #0x18]\n"
- ".word 0x4f80e218 // sdot v24.4s, v16.16b, v0.4b[0]\n"
- "ldr temploadreg2, [%[b_ptr0], #0x28]\n"
- "str q26, [c_ptr2], #0x10\n"
- "prfm PSTL1KEEP, [c_ptr1, #0x40]\n"
+ ".inst 0x4f80e218 // sdot v24.4s, v16.16b, v0.4b[0]\n"
+ "str q26, [c_ptr2]\n"
"movi v26.4s, #0\n"
- ".word 0x4f82e219 // sdot v25.4s, v16.16b, v2.4b[0]\n"
- "str q27, [c_ptr3], #0x10\n"
+ "ldr temploadreg2, [%[b_ptr0], #0x28]\n"
+ "add c_ptr2, c_ptr2, #0x10\n"
+ ".inst 0x4f81e219 // sdot v25.4s, v16.16b, v1.4b[0]\n"
+ "str q27, [c_ptr3]\n"
"movi v27.4s, #0\n"
- "prfm PSTL1KEEP, [c_ptr2, #0x40]\n"
- ".word 0x4f84e21a // sdot v26.4s, v16.16b, v4.4b[0]\n"
- "str q28, [c_ptr4], #0x10\n"
+ "add c_ptr3, c_ptr3, #0x10\n"
+ ".inst 0x4f82e21a // sdot v26.4s, v16.16b, v2.4b[0]\n"
+ "str q28, [c_ptr4]\n"
"movi v28.4s, #0\n"
- "prfm PSTL1KEEP, [c_ptr3, #0x40]\n"
- ".word 0x4f86e21b // sdot v27.4s, v16.16b, v6.4b[0]\n"
- "str q29, [c_ptr5], #0x10\n"
+ "add c_ptr4, c_ptr4, #0x10\n"
+ ".inst 0x4f83e21b // sdot v27.4s, v16.16b, v3.4b[0]\n"
+ "str q29, [c_ptr5]\n"
"movi v29.4s, #0\n"
- "prfm PSTL1KEEP, [c_ptr4, #0x40]\n"
- ".word 0x4f88e21c // sdot v28.4s, v16.16b, v8.4b[0]\n"
- "str q30, [c_ptr6], #0x10\n"
+ "add c_ptr5, c_ptr5, #0x10\n"
+ ".inst 0x4f84e21c // sdot v28.4s, v16.16b, v4.4b[0]\n"
+ "str q30, [c_ptr6]\n"
"movi v30.4s, #0\n"
- "prfm PSTL1KEEP, [c_ptr5, #0x40]\n"
- ".word 0x4f8ae21d // sdot v29.4s, v16.16b, v10.4b[0]\n"
- "str q31, [c_ptr7], #0x10\n"
+ "add c_ptr6, c_ptr6, #0x10\n"
+ ".inst 0x4f85e21d // sdot v29.4s, v16.16b, v5.4b[0]\n"
+ "str q31, [c_ptr7]\n"
"movi v31.4s, #0\n"
- "prfm PSTL1KEEP, [c_ptr6, #0x40]\n"
- ".word 0x4f8ce21e // sdot v30.4s, v16.16b, v12.4b[0]\n"
- "prfm PSTL1KEEP, [c_ptr7, #0x40]\n"
- ".word 0x4f8ee21f // sdot v31.4s, v16.16b, v14.4b[0]\n"
+ "add c_ptr7, c_ptr7, #0x10\n"
+ ".inst 0x4f86e21e // sdot v30.4s, v16.16b, v6.4b[0]\n"
+ "prfm PSTL1KEEP, [%[c_ptr0], #0x40]\n"
+ ".inst 0x4f87e21f // sdot v31.4s, v16.16b, v7.4b[0]\n"
"ldr d16, [%[b_ptr0]]\n"
- ".word 0x4fa0e238 // sdot v24.4s, v17.16b, v0.4b[1]\n"
- ".word 0x4fa2e239 // sdot v25.4s, v17.16b, v2.4b[1]\n"
- ".word 0x4fa4e23a // sdot v26.4s, v17.16b, v4.4b[1]\n"
+ ".inst 0x4fa0e238 // sdot v24.4s, v17.16b, v0.4b[1]\n"
+ "prfm PSTL1KEEP, [c_ptr1, #0x40]\n"
+ ".inst 0x4fa1e239 // sdot v25.4s, v17.16b, v1.4b[1]\n"
"ins v16.d[1], temploadreg0\n"
- ".word 0x4fa6e23b // sdot v27.4s, v17.16b, v6.4b[1]\n"
- ".word 0x4fa8e23c // sdot v28.4s, v17.16b, v8.4b[1]\n"
- ".word 0x4faae23d // sdot v29.4s, v17.16b, v10.4b[1]\n"
- ".word 0x4face23e // sdot v30.4s, v17.16b, v12.4b[1]\n"
- ".word 0x4faee23f // sdot v31.4s, v17.16b, v14.4b[1]\n"
+ ".inst 0x4fa2e23a // sdot v26.4s, v17.16b, v2.4b[1]\n"
+ "prfm PSTL1KEEP, [c_ptr2, #0x40]\n"
+ ".inst 0x4fa3e23b // sdot v27.4s, v17.16b, v3.4b[1]\n"
+ "prfm PSTL1KEEP, [c_ptr3, #0x40]\n"
+ ".inst 0x4fa4e23c // sdot v28.4s, v17.16b, v4.4b[1]\n"
+ "prfm PSTL1KEEP, [c_ptr4, #0x40]\n"
+ ".inst 0x4fa5e23d // sdot v29.4s, v17.16b, v5.4b[1]\n"
+ "prfm PSTL1KEEP, [c_ptr5, #0x40]\n"
+ ".inst 0x4fa6e23e // sdot v30.4s, v17.16b, v6.4b[1]\n"
+ "prfm PSTL1KEEP, [c_ptr6, #0x40]\n"
+ ".inst 0x4fa7e23f // sdot v31.4s, v17.16b, v7.4b[1]\n"
"ldr d17, [%[b_ptr0], #0x10]\n"
- ".word 0x4f80ea58 // sdot v24.4s, v18.16b, v0.4b[2]\n"
- ".word 0x4f82ea59 // sdot v25.4s, v18.16b, v2.4b[2]\n"
- ".word 0x4f84ea5a // sdot v26.4s, v18.16b, v4.4b[2]\n"
+ ".inst 0x4f80ea58 // sdot v24.4s, v18.16b, v0.4b[2]\n"
+ "prfm PSTL1KEEP, [c_ptr7, #0x40]\n"
+ ".inst 0x4f81ea59 // sdot v25.4s, v18.16b, v1.4b[2]\n"
"ins v17.d[1], temploadreg1\n"
- ".word 0x4f86ea5b // sdot v27.4s, v18.16b, v6.4b[2]\n"
- ".word 0x4f88ea5c // sdot v28.4s, v18.16b, v8.4b[2]\n"
- ".word 0x4f8aea5d // sdot v29.4s, v18.16b, v10.4b[2]\n"
- ".word 0x4f8cea5e // sdot v30.4s, v18.16b, v12.4b[2]\n"
- ".word 0x4f8eea5f // sdot v31.4s, v18.16b, v14.4b[2]\n"
+ ".inst 0x4f82ea5a // sdot v26.4s, v18.16b, v2.4b[2]\n"
+ ".inst 0x4f83ea5b // sdot v27.4s, v18.16b, v3.4b[2]\n"
+ ".inst 0x4f84ea5c // sdot v28.4s, v18.16b, v4.4b[2]\n"
+ ".inst 0x4f85ea5d // sdot v29.4s, v18.16b, v5.4b[2]\n"
+ ".inst 0x4f86ea5e // sdot v30.4s, v18.16b, v6.4b[2]\n"
+ ".inst 0x4f87ea5f // sdot v31.4s, v18.16b, v7.4b[2]\n"
"ldr d18, [%[b_ptr0], #0x20]\n"
"add %[b_ptr0], %[b_ptr0], #0x30\n"
"b.ne 8b\n"
"7:\n"
- "str q24, [%[c_ptr0]], #0x10\n"
+ "str q24, [%[c_ptr0]]\n"
+ "add %[c_ptr0], %[c_ptr0], #0x10\n"
"movi v24.4s, #0\n"
"ins v18.d[1], temploadreg2\n"
- "str q25, [c_ptr1], #0x10\n"
+ "str q25, [c_ptr1]\n"
+ "add c_ptr1, c_ptr1, #0x10\n"
"movi v25.4s, #0\n"
- ".word 0x4f80e218 // sdot v24.4s, v16.16b, v0.4b[0]\n"
- "str q26, [c_ptr2], #0x10\n"
+ ".inst 0x4f80e218 // sdot v24.4s, v16.16b, v0.4b[0]\n"
+ "str q26, [c_ptr2]\n"
"movi v26.4s, #0\n"
- ".word 0x4f82e219 // sdot v25.4s, v16.16b, v2.4b[0]\n"
- ".word 0x4fa0e238 // sdot v24.4s, v17.16b, v0.4b[1]\n"
- "str q27, [c_ptr3], #0x10\n"
+ "add c_ptr2, c_ptr2, #0x10\n"
+ ".inst 0x4f81e219 // sdot v25.4s, v16.16b, v1.4b[0]\n"
+ "str q27, [c_ptr3]\n"
"movi v27.4s, #0\n"
- ".word 0x4f84e21a // sdot v26.4s, v16.16b, v4.4b[0]\n"
- ".word 0x4fa2e239 // sdot v25.4s, v17.16b, v2.4b[1]\n"
- "str q28, [c_ptr4], #0x10\n"
+ "add c_ptr3, c_ptr3, #0x10\n"
+ ".inst 0x4f82e21a // sdot v26.4s, v16.16b, v2.4b[0]\n"
+ "str q28, [c_ptr4]\n"
"movi v28.4s, #0\n"
- ".word 0x4f86e21b // sdot v27.4s, v16.16b, v6.4b[0]\n"
- ".word 0x4fa4e23a // sdot v26.4s, v17.16b, v4.4b[1]\n"
- "str q29, [c_ptr5], #0x10\n"
+ "add c_ptr4, c_ptr4, #0x10\n"
+ ".inst 0x4f83e21b // sdot v27.4s, v16.16b, v3.4b[0]\n"
+ "str q29, [c_ptr5]\n"
"movi v29.4s, #0\n"
- ".word 0x4f88e21c // sdot v28.4s, v16.16b, v8.4b[0]\n"
- ".word 0x4fa6e23b // sdot v27.4s, v17.16b, v6.4b[1]\n"
- "str q30, [c_ptr6], #0x10\n"
+ "add c_ptr5, c_ptr5, #0x10\n"
+ ".inst 0x4f84e21c // sdot v28.4s, v16.16b, v4.4b[0]\n"
+ "str q30, [c_ptr6]\n"
"movi v30.4s, #0\n"
- ".word 0x4f8ae21d // sdot v29.4s, v16.16b, v10.4b[0]\n"
- ".word 0x4fa8e23c // sdot v28.4s, v17.16b, v8.4b[1]\n"
- "str q31, [c_ptr7], #0x10\n"
+ "add c_ptr6, c_ptr6, #0x10\n"
+ ".inst 0x4f85e21d // sdot v29.4s, v16.16b, v5.4b[0]\n"
+ "str q31, [c_ptr7]\n"
"movi v31.4s, #0\n"
- ".word 0x4f8ce21e // sdot v30.4s, v16.16b, v12.4b[0]\n"
- ".word 0x4faae23d // sdot v29.4s, v17.16b, v10.4b[1]\n"
- ".word 0x4f80ea58 // sdot v24.4s, v18.16b, v0.4b[2]\n"
- ".word 0x4f8ee21f // sdot v31.4s, v16.16b, v14.4b[0]\n"
- ".word 0x4face23e // sdot v30.4s, v17.16b, v12.4b[1]\n"
- ".word 0x4f82ea59 // sdot v25.4s, v18.16b, v2.4b[2]\n"
- ".word 0x4f84ea5a // sdot v26.4s, v18.16b, v4.4b[2]\n"
- ".word 0x4faee23f // sdot v31.4s, v17.16b, v14.4b[1]\n"
- ".word 0x4f86ea5b // sdot v27.4s, v18.16b, v6.4b[2]\n"
- ".word 0x4f88ea5c // sdot v28.4s, v18.16b, v8.4b[2]\n"
- ".word 0x4f8aea5d // sdot v29.4s, v18.16b, v10.4b[2]\n"
- ".word 0x4f8cea5e // sdot v30.4s, v18.16b, v12.4b[2]\n"
- ".word 0x4f8eea5f // sdot v31.4s, v18.16b, v14.4b[2]\n"
+ "add c_ptr7, c_ptr7, #0x10\n"
+ ".inst 0x4f86e21e // sdot v30.4s, v16.16b, v6.4b[0]\n"
+ ".inst 0x4fa0e238 // sdot v24.4s, v17.16b, v0.4b[1]\n"
+ ".inst 0x4f87e21f // sdot v31.4s, v16.16b, v7.4b[0]\n"
+ ".inst 0x4fa1e239 // sdot v25.4s, v17.16b, v1.4b[1]\n"
+ ".inst 0x4fa2e23a // sdot v26.4s, v17.16b, v2.4b[1]\n"
+ ".inst 0x4fa3e23b // sdot v27.4s, v17.16b, v3.4b[1]\n"
+ ".inst 0x4fa4e23c // sdot v28.4s, v17.16b, v4.4b[1]\n"
+ ".inst 0x4fa5e23d // sdot v29.4s, v17.16b, v5.4b[1]\n"
+ ".inst 0x4fa6e23e // sdot v30.4s, v17.16b, v6.4b[1]\n"
+ ".inst 0x4fa7e23f // sdot v31.4s, v17.16b, v7.4b[1]\n"
+ ".inst 0x4f80ea58 // sdot v24.4s, v18.16b, v0.4b[2]\n"
+ ".inst 0x4f81ea59 // sdot v25.4s, v18.16b, v1.4b[2]\n"
+ ".inst 0x4f82ea5a // sdot v26.4s, v18.16b, v2.4b[2]\n"
+ ".inst 0x4f83ea5b // sdot v27.4s, v18.16b, v3.4b[2]\n"
+ ".inst 0x4f84ea5c // sdot v28.4s, v18.16b, v4.4b[2]\n"
+ ".inst 0x4f85ea5d // sdot v29.4s, v18.16b, v5.4b[2]\n"
+ ".inst 0x4f86ea5e // sdot v30.4s, v18.16b, v6.4b[2]\n"
+ ".inst 0x4f87ea5f // sdot v31.4s, v18.16b, v7.4b[2]\n"
"6:\n"
"str q24, [%[c_ptr0]]\n"
+ "add %[c_ptr0], %[c_ptr0], #0x10\n"
"str q25, [c_ptr1]\n"
+ "add c_ptr1, c_ptr1, #0x10\n"
"str q26, [c_ptr2]\n"
+ "add c_ptr2, c_ptr2, #0x10\n"
"str q27, [c_ptr3]\n"
+ "add c_ptr3, c_ptr3, #0x10\n"
"str q28, [c_ptr4]\n"
+ "add c_ptr4, c_ptr4, #0x10\n"
"str q29, [c_ptr5]\n"
+ "add c_ptr5, c_ptr5, #0x10\n"
"str q30, [c_ptr6]\n"
+ "add c_ptr6, c_ptr6, #0x10\n"
"str q31, [c_ptr7]\n"
+ "add c_ptr7, c_ptr7, #0x10\n"
".unreq a_ptr1\n"
".unreq a_ptr2\n"
".unreq a_ptr3\n"
@@ -925,63 +998,63 @@ void a64_smallK_hybrid_s8s32_dot_4x8_a55(const int8_t *A, int lda, const int8_t
"1:\n"
"cbnz %[odds], 2f\n"
"ldr q0, [%[a_ptr0]]\n"
- "ldr q2, [a_ptr1]\n"
- "ldr q4, [a_ptr2]\n"
- "ldr q6, [a_ptr3]\n"
- "ldr q8, [a_ptr4]\n"
- "ldr q10, [a_ptr5]\n"
- "ldr q12, [a_ptr6]\n"
- "ldr q14, [a_ptr7]\n"
+ "ldr q1, [a_ptr1]\n"
+ "ldr q2, [a_ptr2]\n"
+ "ldr q3, [a_ptr3]\n"
+ "ldr q4, [a_ptr4]\n"
+ "ldr q5, [a_ptr5]\n"
+ "ldr q6, [a_ptr6]\n"
+ "ldr q7, [a_ptr7]\n"
"b 3f\n"
"2:\n"
"ldr d0, [%[a_ptr0]], #0x8\n"
- "ldr d2, [a_ptr1], #0x8\n"
- "ldr d4, [a_ptr2], #0x8\n"
- "ldr d6, [a_ptr3], #0x8\n"
- "ldr d8, [a_ptr4], #0x8\n"
- "ldr d10, [a_ptr5], #0x8\n"
- "ldr d12, [a_ptr6], #0x8\n"
- "ldr d14, [a_ptr7], #0x8\n"
+ "ldr d1, [a_ptr1], #0x8\n"
+ "ldr d2, [a_ptr2], #0x8\n"
+ "ldr d3, [a_ptr3], #0x8\n"
+ "ldr d4, [a_ptr4], #0x8\n"
+ "ldr d5, [a_ptr5], #0x8\n"
+ "ldr d6, [a_ptr6], #0x8\n"
+ "ldr d7, [a_ptr7], #0x8\n"
"ld1 {v0.s}[2], [%[a_ptr0]], #4\n"
- "ld1 {v2.s}[2], [a_ptr1], #4\n"
- "ld1 {v4.s}[2], [a_ptr2], #4\n"
- "ld1 {v6.s}[2], [a_ptr3], #4\n"
- "ld1 {v8.s}[2], [a_ptr4], #4\n"
- "ld1 {v10.s}[2], [a_ptr5], #4\n"
- "ld1 {v12.s}[2], [a_ptr6], #4\n"
- "ld1 {v14.s}[2], [a_ptr7], #4\n"
+ "ld1 {v1.s}[2], [a_ptr1], #4\n"
+ "ld1 {v2.s}[2], [a_ptr2], #4\n"
+ "ld1 {v3.s}[2], [a_ptr3], #4\n"
+ "ld1 {v4.s}[2], [a_ptr4], #4\n"
+ "ld1 {v5.s}[2], [a_ptr5], #4\n"
+ "ld1 {v6.s}[2], [a_ptr6], #4\n"
+ "ld1 {v7.s}[2], [a_ptr7], #4\n"
"subs %[odds], %[odds], #0x1\n"
"b.ne 4f\n"
"ld1 {v0.b}[12], [%[a_ptr0]]\n"
- "ld1 {v2.b}[12], [a_ptr1]\n"
- "ld1 {v4.b}[12], [a_ptr2]\n"
- "ld1 {v6.b}[12], [a_ptr3]\n"
- "ld1 {v8.b}[12], [a_ptr4]\n"
- "ld1 {v10.b}[12], [a_ptr5]\n"
- "ld1 {v12.b}[12], [a_ptr6]\n"
- "ld1 {v14.b}[12], [a_ptr7]\n"
+ "ld1 {v1.b}[12], [a_ptr1]\n"
+ "ld1 {v2.b}[12], [a_ptr2]\n"
+ "ld1 {v3.b}[12], [a_ptr3]\n"
+ "ld1 {v4.b}[12], [a_ptr4]\n"
+ "ld1 {v5.b}[12], [a_ptr5]\n"
+ "ld1 {v6.b}[12], [a_ptr6]\n"
+ "ld1 {v7.b}[12], [a_ptr7]\n"
"b 3f\n"
"4:\n"
"ld1 {v0.h}[6], [%[a_ptr0]], #2\n"
- "ld1 {v2.h}[6], [a_ptr1], #2\n"
- "ld1 {v4.h}[6], [a_ptr2], #2\n"
- "ld1 {v6.h}[6], [a_ptr3], #2\n"
- "ld1 {v8.h}[6], [a_ptr4], #2\n"
- "ld1 {v10.h}[6], [a_ptr5], #2\n"
- "ld1 {v12.h}[6], [a_ptr6], #2\n"
- "ld1 {v14.h}[6], [a_ptr7], #2\n"
+ "ld1 {v1.h}[6], [a_ptr1], #2\n"
+ "ld1 {v2.h}[6], [a_ptr2], #2\n"
+ "ld1 {v3.h}[6], [a_ptr3], #2\n"
+ "ld1 {v4.h}[6], [a_ptr4], #2\n"
+ "ld1 {v5.h}[6], [a_ptr5], #2\n"
+ "ld1 {v6.h}[6], [a_ptr6], #2\n"
+ "ld1 {v7.h}[6], [a_ptr7], #2\n"
"subs %[odds], %[odds], #0x1\n"
"b.ne 5f\n"
"b 3f\n"
"5:\n"
"ld1 {v0.b}[14], [%[a_ptr0]]\n"
- "ld1 {v2.b}[14], [a_ptr1]\n"
- "ld1 {v4.b}[14], [a_ptr2]\n"
- "ld1 {v6.b}[14], [a_ptr3]\n"
- "ld1 {v8.b}[14], [a_ptr4]\n"
- "ld1 {v10.b}[14], [a_ptr5]\n"
- "ld1 {v12.b}[14], [a_ptr6]\n"
- "ld1 {v14.b}[14], [a_ptr7]\n"
+ "ld1 {v1.b}[14], [a_ptr1]\n"
+ "ld1 {v2.b}[14], [a_ptr2]\n"
+ "ld1 {v3.b}[14], [a_ptr3]\n"
+ "ld1 {v4.b}[14], [a_ptr4]\n"
+ "ld1 {v5.b}[14], [a_ptr5]\n"
+ "ld1 {v6.b}[14], [a_ptr6]\n"
+ "ld1 {v7.b}[14], [a_ptr7]\n"
"3:\n"
"movi v24.4s, #0\n"
"ldr q16, [%[b_ptr0]]\n"
@@ -999,41 +1072,41 @@ void a64_smallK_hybrid_s8s32_dot_4x8_a55(const int8_t *A, int lda, const int8_t
"prfm PLDL1KEEP, [a_ptr7, #0xc0]\n"
"movi v31.4s, #0\n"
"prfm PLDL1KEEP, [a_ptr7, #0x100]\n"
- ".word 0x4f80e218 // sdot v24.4s, v16.16b, v0.4b[0]\n"
+ ".inst 0x4f80e218 // sdot v24.4s, v16.16b, v0.4b[0]\n"
"prfm PLDL1KEEP, [a_ptr7, #0x140]\n"
- ".word 0x4f82e219 // sdot v25.4s, v16.16b, v2.4b[0]\n"
+ ".inst 0x4f81e219 // sdot v25.4s, v16.16b, v1.4b[0]\n"
"prfm PLDL1KEEP, [a_ptr7, #0x180]\n"
- ".word 0x4f84e21a // sdot v26.4s, v16.16b, v4.4b[0]\n"
+ ".inst 0x4f82e21a // sdot v26.4s, v16.16b, v2.4b[0]\n"
"add %[b_ptr0], %[b_ptr0], #0x40\n"
- ".word 0x4f86e21b // sdot v27.4s, v16.16b, v6.4b[0]\n"
- ".word 0x4f88e21c // sdot v28.4s, v16.16b, v8.4b[0]\n"
- ".word 0x4f8ae21d // sdot v29.4s, v16.16b, v10.4b[0]\n"
- ".word 0x4f8ce21e // sdot v30.4s, v16.16b, v12.4b[0]\n"
- ".word 0x4f8ee21f // sdot v31.4s, v16.16b, v14.4b[0]\n"
- ".word 0x4fa0e238 // sdot v24.4s, v17.16b, v0.4b[1]\n"
- ".word 0x4fa2e239 // sdot v25.4s, v17.16b, v2.4b[1]\n"
- ".word 0x4fa4e23a // sdot v26.4s, v17.16b, v4.4b[1]\n"
- ".word 0x4fa6e23b // sdot v27.4s, v17.16b, v6.4b[1]\n"
- ".word 0x4fa8e23c // sdot v28.4s, v17.16b, v8.4b[1]\n"
- ".word 0x4faae23d // sdot v29.4s, v17.16b, v10.4b[1]\n"
- ".word 0x4face23e // sdot v30.4s, v17.16b, v12.4b[1]\n"
- ".word 0x4faee23f // sdot v31.4s, v17.16b, v14.4b[1]\n"
- ".word 0x4f80ea58 // sdot v24.4s, v18.16b, v0.4b[2]\n"
- ".word 0x4f82ea59 // sdot v25.4s, v18.16b, v2.4b[2]\n"
- ".word 0x4f84ea5a // sdot v26.4s, v18.16b, v4.4b[2]\n"
- ".word 0x4f86ea5b // sdot v27.4s, v18.16b, v6.4b[2]\n"
- ".word 0x4f88ea5c // sdot v28.4s, v18.16b, v8.4b[2]\n"
- ".word 0x4f8aea5d // sdot v29.4s, v18.16b, v10.4b[2]\n"
- ".word 0x4f8cea5e // sdot v30.4s, v18.16b, v12.4b[2]\n"
- ".word 0x4f8eea5f // sdot v31.4s, v18.16b, v14.4b[2]\n"
- ".word 0x4fa0ea78 // sdot v24.4s, v19.16b, v0.4b[3]\n"
- ".word 0x4fa2ea79 // sdot v25.4s, v19.16b, v2.4b[3]\n"
- ".word 0x4fa4ea7a // sdot v26.4s, v19.16b, v4.4b[3]\n"
- ".word 0x4fa6ea7b // sdot v27.4s, v19.16b, v6.4b[3]\n"
- ".word 0x4fa8ea7c // sdot v28.4s, v19.16b, v8.4b[3]\n"
- ".word 0x4faaea7d // sdot v29.4s, v19.16b, v10.4b[3]\n"
- ".word 0x4facea7e // sdot v30.4s, v19.16b, v12.4b[3]\n"
- ".word 0x4faeea7f // sdot v31.4s, v19.16b, v14.4b[3]\n"
+ ".inst 0x4f83e21b // sdot v27.4s, v16.16b, v3.4b[0]\n"
+ ".inst 0x4f84e21c // sdot v28.4s, v16.16b, v4.4b[0]\n"
+ ".inst 0x4f85e21d // sdot v29.4s, v16.16b, v5.4b[0]\n"
+ ".inst 0x4f86e21e // sdot v30.4s, v16.16b, v6.4b[0]\n"
+ ".inst 0x4f87e21f // sdot v31.4s, v16.16b, v7.4b[0]\n"
+ ".inst 0x4fa0e238 // sdot v24.4s, v17.16b, v0.4b[1]\n"
+ ".inst 0x4fa1e239 // sdot v25.4s, v17.16b, v1.4b[1]\n"
+ ".inst 0x4fa2e23a // sdot v26.4s, v17.16b, v2.4b[1]\n"
+ ".inst 0x4fa3e23b // sdot v27.4s, v17.16b, v3.4b[1]\n"
+ ".inst 0x4fa4e23c // sdot v28.4s, v17.16b, v4.4b[1]\n"
+ ".inst 0x4fa5e23d // sdot v29.4s, v17.16b, v5.4b[1]\n"
+ ".inst 0x4fa6e23e // sdot v30.4s, v17.16b, v6.4b[1]\n"
+ ".inst 0x4fa7e23f // sdot v31.4s, v17.16b, v7.4b[1]\n"
+ ".inst 0x4f80ea58 // sdot v24.4s, v18.16b, v0.4b[2]\n"
+ ".inst 0x4f81ea59 // sdot v25.4s, v18.16b, v1.4b[2]\n"
+ ".inst 0x4f82ea5a // sdot v26.4s, v18.16b, v2.4b[2]\n"
+ ".inst 0x4f83ea5b // sdot v27.4s, v18.16b, v3.4b[2]\n"
+ ".inst 0x4f84ea5c // sdot v28.4s, v18.16b, v4.4b[2]\n"
+ ".inst 0x4f85ea5d // sdot v29.4s, v18.16b, v5.4b[2]\n"
+ ".inst 0x4f86ea5e // sdot v30.4s, v18.16b, v6.4b[2]\n"
+ ".inst 0x4f87ea5f // sdot v31.4s, v18.16b, v7.4b[2]\n"
+ ".inst 0x4fa0ea78 // sdot v24.4s, v19.16b, v0.4b[3]\n"
+ ".inst 0x4fa1ea79 // sdot v25.4s, v19.16b, v1.4b[3]\n"
+ ".inst 0x4fa2ea7a // sdot v26.4s, v19.16b, v2.4b[3]\n"
+ ".inst 0x4fa3ea7b // sdot v27.4s, v19.16b, v3.4b[3]\n"
+ ".inst 0x4fa4ea7c // sdot v28.4s, v19.16b, v4.4b[3]\n"
+ ".inst 0x4fa5ea7d // sdot v29.4s, v19.16b, v5.4b[3]\n"
+ ".inst 0x4fa6ea7e // sdot v30.4s, v19.16b, v6.4b[3]\n"
+ ".inst 0x4fa7ea7f // sdot v31.4s, v19.16b, v7.4b[3]\n"
"cbz %[loops], 6f\n"
"ldr d16, [%[b_ptr0]]\n"
"subs %[loops], %[loops], #0x1\n"
@@ -1050,136 +1123,160 @@ void a64_smallK_hybrid_s8s32_dot_4x8_a55(const int8_t *A, int lda, const int8_t
"ins v18.d[1], temploadreg2\n"
"b.eq 7f\n"
"8:\n"
- "str q24, [%[c_ptr0]], #0x10\n"
+ "str q24, [%[c_ptr0]]\n"
"subs %[loops], %[loops], #0x1\n"
"movi v24.4s, #0\n"
"ins v19.d[1], temploadreg3\n"
"ldr temploadreg0, [%[b_ptr0], #0x8]\n"
- "prfm PSTL1KEEP, [%[c_ptr0], #0x40]\n"
- "str q25, [c_ptr1], #0x10\n"
+ "add %[c_ptr0], %[c_ptr0], #0x10\n"
+ "str q25, [c_ptr1]\n"
+ "add c_ptr1, c_ptr1, #0x10\n"
"movi v25.4s, #0\n"
"ldr temploadreg1, [%[b_ptr0], #0x18]\n"
- ".word 0x4f80e218 // sdot v24.4s, v16.16b, v0.4b[0]\n"
+ ".inst 0x4f80e218 // sdot v24.4s, v16.16b, v0.4b[0]\n"
+ "str q26, [c_ptr2]\n"
+ "movi v26.4s, #0\n"
"ldr temploadreg2, [%[b_ptr0], #0x28]\n"
"ldr temploadreg3, [%[b_ptr0], #0x38]\n"
- "prfm PSTL1KEEP, [c_ptr1, #0x40]\n"
- ".word 0x4f82e219 // sdot v25.4s, v16.16b, v2.4b[0]\n"
- "str q26, [c_ptr2], #0x10\n"
- "movi v26.4s, #0\n"
- ".word 0x4fa0e238 // sdot v24.4s, v17.16b, v0.4b[1]\n"
- ".word 0x4fa2e239 // sdot v25.4s, v17.16b, v2.4b[1]\n"
- "str q27, [c_ptr3], #0x10\n"
+ "add c_ptr2, c_ptr2, #0x10\n"
+ ".inst 0x4f81e219 // sdot v25.4s, v16.16b, v1.4b[0]\n"
+ "str q27, [c_ptr3]\n"
"movi v27.4s, #0\n"
- "prfm PSTL1KEEP, [c_ptr2, #0x40]\n"
- ".word 0x4f84e21a // sdot v26.4s, v16.16b, v4.4b[0]\n"
- "str q28, [c_ptr4], #0x10\n"
+ "add c_ptr3, c_ptr3, #0x10\n"
+ ".inst 0x4f82e21a // sdot v26.4s, v16.16b, v2.4b[0]\n"
+ "str q28, [c_ptr4]\n"
"movi v28.4s, #0\n"
- "prfm PSTL1KEEP, [c_ptr3, #0x40]\n"
- ".word 0x4f86e21b // sdot v27.4s, v16.16b, v6.4b[0]\n"
- "str q29, [c_ptr5], #0x10\n"
+ "add c_ptr4, c_ptr4, #0x10\n"
+ ".inst 0x4f83e21b // sdot v27.4s, v16.16b, v3.4b[0]\n"
+ "str q29, [c_ptr5]\n"
"movi v29.4s, #0\n"
- "prfm PSTL1KEEP, [c_ptr4, #0x40]\n"
- ".word 0x4f88e21c // sdot v28.4s, v16.16b, v8.4b[0]\n"
- "str q30, [c_ptr6], #0x10\n"
+ "add c_ptr5, c_ptr5, #0x10\n"
+ ".inst 0x4f84e21c // sdot v28.4s, v16.16b, v4.4b[0]\n"
+ "str q30, [c_ptr6]\n"
"movi v30.4s, #0\n"
- "prfm PSTL1KEEP, [c_ptr5, #0x40]\n"
- ".word 0x4f8ae21d // sdot v29.4s, v16.16b, v10.4b[0]\n"
- "str q31, [c_ptr7], #0x10\n"
+ "add c_ptr6, c_ptr6, #0x10\n"
+ ".inst 0x4f85e21d // sdot v29.4s, v16.16b, v5.4b[0]\n"
+ "str q31, [c_ptr7]\n"
"movi v31.4s, #0\n"
- "prfm PSTL1KEEP, [c_ptr6, #0x40]\n"
- ".word 0x4f8ce21e // sdot v30.4s, v16.16b, v12.4b[0]\n"
- "prfm PSTL1KEEP, [c_ptr7, #0x40]\n"
- ".word 0x4f8ee21f // sdot v31.4s, v16.16b, v14.4b[0]\n"
+ "add c_ptr7, c_ptr7, #0x10\n"
+ ".inst 0x4f86e21e // sdot v30.4s, v16.16b, v6.4b[0]\n"
+ "prfm PSTL1KEEP, [%[c_ptr0], #0x40]\n"
+ ".inst 0x4f87e21f // sdot v31.4s, v16.16b, v7.4b[0]\n"
"ldr d16, [%[b_ptr0]]\n"
- ".word 0x4fa4e23a // sdot v26.4s, v17.16b, v4.4b[1]\n"
- ".word 0x4fa6e23b // sdot v27.4s, v17.16b, v6.4b[1]\n"
- ".word 0x4fa8e23c // sdot v28.4s, v17.16b, v8.4b[1]\n"
+ ".inst 0x4fa0e238 // sdot v24.4s, v17.16b, v0.4b[1]\n"
+ "prfm PSTL1KEEP, [c_ptr1, #0x40]\n"
+ ".inst 0x4fa1e239 // sdot v25.4s, v17.16b, v1.4b[1]\n"
"ins v16.d[1], temploadreg0\n"
- ".word 0x4faae23d // sdot v29.4s, v17.16b, v10.4b[1]\n"
- ".word 0x4face23e // sdot v30.4s, v17.16b, v12.4b[1]\n"
- ".word 0x4faee23f // sdot v31.4s, v17.16b, v14.4b[1]\n"
+ ".inst 0x4fa2e23a // sdot v26.4s, v17.16b, v2.4b[1]\n"
+ "prfm PSTL1KEEP, [c_ptr2, #0x40]\n"
+ ".inst 0x4fa3e23b // sdot v27.4s, v17.16b, v3.4b[1]\n"
+ "prfm PSTL1KEEP, [c_ptr3, #0x40]\n"
+ ".inst 0x4fa4e23c // sdot v28.4s, v17.16b, v4.4b[1]\n"
+ "prfm PSTL1KEEP, [c_ptr4, #0x40]\n"
+ ".inst 0x4fa5e23d // sdot v29.4s, v17.16b, v5.4b[1]\n"
+ "prfm PSTL1KEEP, [c_ptr5, #0x40]\n"
+ ".inst 0x4fa6e23e // sdot v30.4s, v17.16b, v6.4b[1]\n"
+ "prfm PSTL1KEEP, [c_ptr6, #0x40]\n"
+ ".inst 0x4fa7e23f // sdot v31.4s, v17.16b, v7.4b[1]\n"
"ldr d17, [%[b_ptr0], #0x10]\n"
- ".word 0x4f80ea58 // sdot v24.4s, v18.16b, v0.4b[2]\n"
- ".word 0x4f82ea59 // sdot v25.4s, v18.16b, v2.4b[2]\n"
- ".word 0x4f84ea5a // sdot v26.4s, v18.16b, v4.4b[2]\n"
+ ".inst 0x4f80ea58 // sdot v24.4s, v18.16b, v0.4b[2]\n"
+ "prfm PSTL1KEEP, [c_ptr7, #0x40]\n"
+ ".inst 0x4f81ea59 // sdot v25.4s, v18.16b, v1.4b[2]\n"
"ins v17.d[1], temploadreg1\n"
- ".word 0x4f86ea5b // sdot v27.4s, v18.16b, v6.4b[2]\n"
- ".word 0x4f88ea5c // sdot v28.4s, v18.16b, v8.4b[2]\n"
- ".word 0x4f8aea5d // sdot v29.4s, v18.16b, v10.4b[2]\n"
- ".word 0x4f8cea5e // sdot v30.4s, v18.16b, v12.4b[2]\n"
- ".word 0x4f8eea5f // sdot v31.4s, v18.16b, v14.4b[2]\n"
+ ".inst 0x4f82ea5a // sdot v26.4s, v18.16b, v2.4b[2]\n"
+ ".inst 0x4f83ea5b // sdot v27.4s, v18.16b, v3.4b[2]\n"
+ ".inst 0x4f84ea5c // sdot v28.4s, v18.16b, v4.4b[2]\n"
+ ".inst 0x4f85ea5d // sdot v29.4s, v18.16b, v5.4b[2]\n"
+ ".inst 0x4f86ea5e // sdot v30.4s, v18.16b, v6.4b[2]\n"
+ ".inst 0x4f87ea5f // sdot v31.4s, v18.16b, v7.4b[2]\n"
"ldr d18, [%[b_ptr0], #0x20]\n"
- ".word 0x4fa0ea78 // sdot v24.4s, v19.16b, v0.4b[3]\n"
- ".word 0x4fa2ea79 // sdot v25.4s, v19.16b, v2.4b[3]\n"
- ".word 0x4fa4ea7a // sdot v26.4s, v19.16b, v4.4b[3]\n"
+ ".inst 0x4fa0ea78 // sdot v24.4s, v19.16b, v0.4b[3]\n"
+ ".inst 0x4fa1ea79 // sdot v25.4s, v19.16b, v1.4b[3]\n"
+ ".inst 0x4fa2ea7a // sdot v26.4s, v19.16b, v2.4b[3]\n"
"ins v18.d[1], temploadreg2\n"
- ".word 0x4fa6ea7b // sdot v27.4s, v19.16b, v6.4b[3]\n"
- ".word 0x4fa8ea7c // sdot v28.4s, v19.16b, v8.4b[3]\n"
- ".word 0x4faaea7d // sdot v29.4s, v19.16b, v10.4b[3]\n"
- ".word 0x4facea7e // sdot v30.4s, v19.16b, v12.4b[3]\n"
- ".word 0x4faeea7f // sdot v31.4s, v19.16b, v14.4b[3]\n"
+ ".inst 0x4fa3ea7b // sdot v27.4s, v19.16b, v3.4b[3]\n"
+ ".inst 0x4fa4ea7c // sdot v28.4s, v19.16b, v4.4b[3]\n"
+ ".inst 0x4fa5ea7d // sdot v29.4s, v19.16b, v5.4b[3]\n"
+ ".inst 0x4fa6ea7e // sdot v30.4s, v19.16b, v6.4b[3]\n"
+ ".inst 0x4fa7ea7f // sdot v31.4s, v19.16b, v7.4b[3]\n"
"ldr d19, [%[b_ptr0], #0x30]\n"
"add %[b_ptr0], %[b_ptr0], #0x40\n"
"b.ne 8b\n"
"7:\n"
- "str q24, [%[c_ptr0]], #0x10\n"
+ "str q24, [%[c_ptr0]]\n"
+ "add %[c_ptr0], %[c_ptr0], #0x10\n"
"movi v24.4s, #0\n"
"ins v19.d[1], temploadreg3\n"
- "str q25, [c_ptr1], #0x10\n"
+ "str q25, [c_ptr1]\n"
+ "add c_ptr1, c_ptr1, #0x10\n"
"movi v25.4s, #0\n"
- ".word 0x4f80e218 // sdot v24.4s, v16.16b, v0.4b[0]\n"
- "str q26, [c_ptr2], #0x10\n"
+ ".inst 0x4f80e218 // sdot v24.4s, v16.16b, v0.4b[0]\n"
+ "str q26, [c_ptr2]\n"
"movi v26.4s, #0\n"
- ".word 0x4f82e219 // sdot v25.4s, v16.16b, v2.4b[0]\n"
- ".word 0x4fa0e238 // sdot v24.4s, v17.16b, v0.4b[1]\n"
- "str q27, [c_ptr3], #0x10\n"
+ "add c_ptr2, c_ptr2, #0x10\n"
+ ".inst 0x4f81e219 // sdot v25.4s, v16.16b, v1.4b[0]\n"
+ "str q27, [c_ptr3]\n"
"movi v27.4s, #0\n"
- ".word 0x4f84e21a // sdot v26.4s, v16.16b, v4.4b[0]\n"
- ".word 0x4fa2e239 // sdot v25.4s, v17.16b, v2.4b[1]\n"
- "str q28, [c_ptr4], #0x10\n"
+ "add c_ptr3, c_ptr3, #0x10\n"
+ ".inst 0x4f82e21a // sdot v26.4s, v16.16b, v2.4b[0]\n"
+ "str q28, [c_ptr4]\n"
"movi v28.4s, #0\n"
- ".word 0x4f86e21b // sdot v27.4s, v16.16b, v6.4b[0]\n"
- ".word 0x4fa4e23a // sdot v26.4s, v17.16b, v4.4b[1]\n"
- "str q29, [c_ptr5], #0x10\n"
+ "add c_ptr4, c_ptr4, #0x10\n"
+ ".inst 0x4f83e21b // sdot v27.4s, v16.16b, v3.4b[0]\n"
+ "str q29, [c_ptr5]\n"
"movi v29.4s, #0\n"
- ".word 0x4f88e21c // sdot v28.4s, v16.16b, v8.4b[0]\n"
- ".word 0x4fa6e23b // sdot v27.4s, v17.16b, v6.4b[1]\n"
- "str q30, [c_ptr6], #0x10\n"
+ "add c_ptr5, c_ptr5, #0x10\n"
+ ".inst 0x4f84e21c // sdot v28.4s, v16.16b, v4.4b[0]\n"
+ "str q30, [c_ptr6]\n"
"movi v30.4s, #0\n"
- ".word 0x4f8ae21d // sdot v29.4s, v16.16b, v10.4b[0]\n"
- ".word 0x4fa8e23c // sdot v28.4s, v17.16b, v8.4b[1]\n"
- "str q31, [c_ptr7], #0x10\n"
+ "add c_ptr6, c_ptr6, #0x10\n"
+ ".inst 0x4f85e21d // sdot v29.4s, v16.16b, v5.4b[0]\n"
+ "str q31, [c_ptr7]\n"
"movi v31.4s, #0\n"
- ".word 0x4f8ce21e // sdot v30.4s, v16.16b, v12.4b[0]\n"
- ".word 0x4faae23d // sdot v29.4s, v17.16b, v10.4b[1]\n"
- ".word 0x4f80ea58 // sdot v24.4s, v18.16b, v0.4b[2]\n"
- ".word 0x4f8ee21f // sdot v31.4s, v16.16b, v14.4b[0]\n"
- ".word 0x4face23e // sdot v30.4s, v17.16b, v12.4b[1]\n"
- ".word 0x4f82ea59 // sdot v25.4s, v18.16b, v2.4b[2]\n"
- ".word 0x4f84ea5a // sdot v26.4s, v18.16b, v4.4b[2]\n"
- ".word 0x4faee23f // sdot v31.4s, v17.16b, v14.4b[1]\n"
- ".word 0x4f86ea5b // sdot v27.4s, v18.16b, v6.4b[2]\n"
- ".word 0x4f88ea5c // sdot v28.4s, v18.16b, v8.4b[2]\n"
- ".word 0x4f8aea5d // sdot v29.4s, v18.16b, v10.4b[2]\n"
- ".word 0x4f8cea5e // sdot v30.4s, v18.16b, v12.4b[2]\n"
- ".word 0x4f8eea5f // sdot v31.4s, v18.16b, v14.4b[2]\n"
- ".word 0x4fa0ea78 // sdot v24.4s, v19.16b, v0.4b[3]\n"
- ".word 0x4fa2ea79 // sdot v25.4s, v19.16b, v2.4b[3]\n"
- ".word 0x4fa4ea7a // sdot v26.4s, v19.16b, v4.4b[3]\n"
- ".word 0x4fa6ea7b // sdot v27.4s, v19.16b, v6.4b[3]\n"
- ".word 0x4fa8ea7c // sdot v28.4s, v19.16b, v8.4b[3]\n"
- ".word 0x4faaea7d // sdot v29.4s, v19.16b, v10.4b[3]\n"
- ".word 0x4facea7e // sdot v30.4s, v19.16b, v12.4b[3]\n"
- ".word 0x4faeea7f // sdot v31.4s, v19.16b, v14.4b[3]\n"
+ "add c_ptr7, c_ptr7, #0x10\n"
+ ".inst 0x4f86e21e // sdot v30.4s, v16.16b, v6.4b[0]\n"
+ ".inst 0x4fa0e238 // sdot v24.4s, v17.16b, v0.4b[1]\n"
+ ".inst 0x4f87e21f // sdot v31.4s, v16.16b, v7.4b[0]\n"
+ ".inst 0x4fa1e239 // sdot v25.4s, v17.16b, v1.4b[1]\n"
+ ".inst 0x4fa2e23a // sdot v26.4s, v17.16b, v2.4b[1]\n"
+ ".inst 0x4fa3e23b // sdot v27.4s, v17.16b, v3.4b[1]\n"
+ ".inst 0x4fa4e23c // sdot v28.4s, v17.16b, v4.4b[1]\n"
+ ".inst 0x4fa5e23d // sdot v29.4s, v17.16b, v5.4b[1]\n"
+ ".inst 0x4fa6e23e // sdot v30.4s, v17.16b, v6.4b[1]\n"
+ ".inst 0x4fa7e23f // sdot v31.4s, v17.16b, v7.4b[1]\n"
+ ".inst 0x4f80ea58 // sdot v24.4s, v18.16b, v0.4b[2]\n"
+ ".inst 0x4f81ea59 // sdot v25.4s, v18.16b, v1.4b[2]\n"
+ ".inst 0x4f82ea5a // sdot v26.4s, v18.16b, v2.4b[2]\n"
+ ".inst 0x4f83ea5b // sdot v27.4s, v18.16b, v3.4b[2]\n"
+ ".inst 0x4f84ea5c // sdot v28.4s, v18.16b, v4.4b[2]\n"
+ ".inst 0x4f85ea5d // sdot v29.4s, v18.16b, v5.4b[2]\n"
+ ".inst 0x4f86ea5e // sdot v30.4s, v18.16b, v6.4b[2]\n"
+ ".inst 0x4f87ea5f // sdot v31.4s, v18.16b, v7.4b[2]\n"
+ ".inst 0x4fa0ea78 // sdot v24.4s, v19.16b, v0.4b[3]\n"
+ ".inst 0x4fa1ea79 // sdot v25.4s, v19.16b, v1.4b[3]\n"
+ ".inst 0x4fa2ea7a // sdot v26.4s, v19.16b, v2.4b[3]\n"
+ ".inst 0x4fa3ea7b // sdot v27.4s, v19.16b, v3.4b[3]\n"
+ ".inst 0x4fa4ea7c // sdot v28.4s, v19.16b, v4.4b[3]\n"
+ ".inst 0x4fa5ea7d // sdot v29.4s, v19.16b, v5.4b[3]\n"
+ ".inst 0x4fa6ea7e // sdot v30.4s, v19.16b, v6.4b[3]\n"
+ ".inst 0x4fa7ea7f // sdot v31.4s, v19.16b, v7.4b[3]\n"
"6:\n"
"str q24, [%[c_ptr0]]\n"
+ "add %[c_ptr0], %[c_ptr0], #0x10\n"
"str q25, [c_ptr1]\n"
+ "add c_ptr1, c_ptr1, #0x10\n"
"str q26, [c_ptr2]\n"
+ "add c_ptr2, c_ptr2, #0x10\n"
"str q27, [c_ptr3]\n"
+ "add c_ptr3, c_ptr3, #0x10\n"
"str q28, [c_ptr4]\n"
+ "add c_ptr4, c_ptr4, #0x10\n"
"str q29, [c_ptr5]\n"
+ "add c_ptr5, c_ptr5, #0x10\n"
"str q30, [c_ptr6]\n"
+ "add c_ptr6, c_ptr6, #0x10\n"
"str q31, [c_ptr7]\n"
+ "add c_ptr7, c_ptr7, #0x10\n"
".unreq a_ptr1\n"
".unreq a_ptr2\n"
".unreq a_ptr3\n"
@@ -1334,50 +1431,50 @@ void a64_smallK_hybrid_s8s32_dot_4x8_a55(const int8_t *A, int lda, const int8_t
"prfm PLDL1KEEP, [a_ptr7, #0x80]\n"
"movi v31.4s, #0\n"
"prfm PLDL1KEEP, [a_ptr7, #0xc0]\n"
- ".word 0x4f80e218 // sdot v24.4s, v16.16b, v0.4b[0]\n"
+ ".inst 0x4f80e218 // sdot v24.4s, v16.16b, v0.4b[0]\n"
"prfm PLDL1KEEP, [a_ptr7, #0x100]\n"
- ".word 0x4f82e219 // sdot v25.4s, v16.16b, v2.4b[0]\n"
+ ".inst 0x4f82e219 // sdot v25.4s, v16.16b, v2.4b[0]\n"
"prfm PLDL1KEEP, [a_ptr7, #0x140]\n"
- ".word 0x4f84e21a // sdot v26.4s, v16.16b, v4.4b[0]\n"
+ ".inst 0x4f84e21a // sdot v26.4s, v16.16b, v4.4b[0]\n"
"prfm PLDL1KEEP, [a_ptr7, #0x180]\n"
- ".word 0x4f86e21b // sdot v27.4s, v16.16b, v6.4b[0]\n"
+ ".inst 0x4f86e21b // sdot v27.4s, v16.16b, v6.4b[0]\n"
"add %[b_ptr0], %[b_ptr0], #0x50\n"
- ".word 0x4f88e21c // sdot v28.4s, v16.16b, v8.4b[0]\n"
- ".word 0x4f8ae21d // sdot v29.4s, v16.16b, v10.4b[0]\n"
- ".word 0x4f8ce21e // sdot v30.4s, v16.16b, v12.4b[0]\n"
- ".word 0x4f8ee21f // sdot v31.4s, v16.16b, v14.4b[0]\n"
- ".word 0x4fa0e238 // sdot v24.4s, v17.16b, v0.4b[1]\n"
- ".word 0x4fa2e239 // sdot v25.4s, v17.16b, v2.4b[1]\n"
- ".word 0x4fa4e23a // sdot v26.4s, v17.16b, v4.4b[1]\n"
- ".word 0x4fa6e23b // sdot v27.4s, v17.16b, v6.4b[1]\n"
- ".word 0x4fa8e23c // sdot v28.4s, v17.16b, v8.4b[1]\n"
- ".word 0x4faae23d // sdot v29.4s, v17.16b, v10.4b[1]\n"
- ".word 0x4face23e // sdot v30.4s, v17.16b, v12.4b[1]\n"
- ".word 0x4faee23f // sdot v31.4s, v17.16b, v14.4b[1]\n"
- ".word 0x4f80ea58 // sdot v24.4s, v18.16b, v0.4b[2]\n"
- ".word 0x4f82ea59 // sdot v25.4s, v18.16b, v2.4b[2]\n"
- ".word 0x4f84ea5a // sdot v26.4s, v18.16b, v4.4b[2]\n"
- ".word 0x4f86ea5b // sdot v27.4s, v18.16b, v6.4b[2]\n"
- ".word 0x4f88ea5c // sdot v28.4s, v18.16b, v8.4b[2]\n"
- ".word 0x4f8aea5d // sdot v29.4s, v18.16b, v10.4b[2]\n"
- ".word 0x4f8cea5e // sdot v30.4s, v18.16b, v12.4b[2]\n"
- ".word 0x4f8eea5f // sdot v31.4s, v18.16b, v14.4b[2]\n"
- ".word 0x4fa0ea78 // sdot v24.4s, v19.16b, v0.4b[3]\n"
- ".word 0x4fa2ea79 // sdot v25.4s, v19.16b, v2.4b[3]\n"
- ".word 0x4fa4ea7a // sdot v26.4s, v19.16b, v4.4b[3]\n"
- ".word 0x4fa6ea7b // sdot v27.4s, v19.16b, v6.4b[3]\n"
- ".word 0x4fa8ea7c // sdot v28.4s, v19.16b, v8.4b[3]\n"
- ".word 0x4faaea7d // sdot v29.4s, v19.16b, v10.4b[3]\n"
- ".word 0x4facea7e // sdot v30.4s, v19.16b, v12.4b[3]\n"
- ".word 0x4faeea7f // sdot v31.4s, v19.16b, v14.4b[3]\n"
- ".word 0x4f81e298 // sdot v24.4s, v20.16b, v1.4b[0]\n"
- ".word 0x4f83e299 // sdot v25.4s, v20.16b, v3.4b[0]\n"
- ".word 0x4f85e29a // sdot v26.4s, v20.16b, v5.4b[0]\n"
- ".word 0x4f87e29b // sdot v27.4s, v20.16b, v7.4b[0]\n"
- ".word 0x4f89e29c // sdot v28.4s, v20.16b, v9.4b[0]\n"
- ".word 0x4f8be29d // sdot v29.4s, v20.16b, v11.4b[0]\n"
- ".word 0x4f8de29e // sdot v30.4s, v20.16b, v13.4b[0]\n"
- ".word 0x4f8fe29f // sdot v31.4s, v20.16b, v15.4b[0]\n"
+ ".inst 0x4f88e21c // sdot v28.4s, v16.16b, v8.4b[0]\n"
+ ".inst 0x4f8ae21d // sdot v29.4s, v16.16b, v10.4b[0]\n"
+ ".inst 0x4f8ce21e // sdot v30.4s, v16.16b, v12.4b[0]\n"
+ ".inst 0x4f8ee21f // sdot v31.4s, v16.16b, v14.4b[0]\n"
+ ".inst 0x4fa0e238 // sdot v24.4s, v17.16b, v0.4b[1]\n"
+ ".inst 0x4fa2e239 // sdot v25.4s, v17.16b, v2.4b[1]\n"
+ ".inst 0x4fa4e23a // sdot v26.4s, v17.16b, v4.4b[1]\n"
+ ".inst 0x4fa6e23b // sdot v27.4s, v17.16b, v6.4b[1]\n"
+ ".inst 0x4fa8e23c // sdot v28.4s, v17.16b, v8.4b[1]\n"
+ ".inst 0x4faae23d // sdot v29.4s, v17.16b, v10.4b[1]\n"
+ ".inst 0x4face23e // sdot v30.4s, v17.16b, v12.4b[1]\n"
+ ".inst 0x4faee23f // sdot v31.4s, v17.16b, v14.4b[1]\n"
+ ".inst 0x4f80ea58 // sdot v24.4s, v18.16b, v0.4b[2]\n"
+ ".inst 0x4f82ea59 // sdot v25.4s, v18.16b, v2.4b[2]\n"
+ ".inst 0x4f84ea5a // sdot v26.4s, v18.16b, v4.4b[2]\n"
+ ".inst 0x4f86ea5b // sdot v27.4s, v18.16b, v6.4b[2]\n"
+ ".inst 0x4f88ea5c // sdot v28.4s, v18.16b, v8.4b[2]\n"
+ ".inst 0x4f8aea5d // sdot v29.4s, v18.16b, v10.4b[2]\n"
+ ".inst 0x4f8cea5e // sdot v30.4s, v18.16b, v12.4b[2]\n"
+ ".inst 0x4f8eea5f // sdot v31.4s, v18.16b, v14.4b[2]\n"
+ ".inst 0x4fa0ea78 // sdot v24.4s, v19.16b, v0.4b[3]\n"
+ ".inst 0x4fa2ea79 // sdot v25.4s, v19.16b, v2.4b[3]\n"
+ ".inst 0x4fa4ea7a // sdot v26.4s, v19.16b, v4.4b[3]\n"
+ ".inst 0x4fa6ea7b // sdot v27.4s, v19.16b, v6.4b[3]\n"
+ ".inst 0x4fa8ea7c // sdot v28.4s, v19.16b, v8.4b[3]\n"
+ ".inst 0x4faaea7d // sdot v29.4s, v19.16b, v10.4b[3]\n"
+ ".inst 0x4facea7e // sdot v30.4s, v19.16b, v12.4b[3]\n"
+ ".inst 0x4faeea7f // sdot v31.4s, v19.16b, v14.4b[3]\n"
+ ".inst 0x4f81e298 // sdot v24.4s, v20.16b, v1.4b[0]\n"
+ ".inst 0x4f83e299 // sdot v25.4s, v20.16b, v3.4b[0]\n"
+ ".inst 0x4f85e29a // sdot v26.4s, v20.16b, v5.4b[0]\n"
+ ".inst 0x4f87e29b // sdot v27.4s, v20.16b, v7.4b[0]\n"
+ ".inst 0x4f89e29c // sdot v28.4s, v20.16b, v9.4b[0]\n"
+ ".inst 0x4f8be29d // sdot v29.4s, v20.16b, v11.4b[0]\n"
+ ".inst 0x4f8de29e // sdot v30.4s, v20.16b, v13.4b[0]\n"
+ ".inst 0x4f8fe29f // sdot v31.4s, v20.16b, v15.4b[0]\n"
"cbz %[loops], 6f\n"
"ldr d16, [%[b_ptr0]]\n"
"subs %[loops], %[loops], #0x1\n"
@@ -1397,155 +1494,179 @@ void a64_smallK_hybrid_s8s32_dot_4x8_a55(const int8_t *A, int lda, const int8_t
"ins v19.d[1], temploadreg3\n"
"b.eq 7f\n"
"8:\n"
- "str q24, [%[c_ptr0]], #0x10\n"
+ "str q24, [%[c_ptr0]]\n"
"subs %[loops], %[loops], #0x1\n"
"movi v24.4s, #0\n"
"ins v20.d[1], temploadreg0\n"
"ldr temploadreg0, [%[b_ptr0], #0x8]\n"
- "prfm PSTL1KEEP, [%[c_ptr0], #0x40]\n"
- "str q25, [c_ptr1], #0x10\n"
+ "add %[c_ptr0], %[c_ptr0], #0x10\n"
+ "str q25, [c_ptr1]\n"
+ "add c_ptr1, c_ptr1, #0x10\n"
"movi v25.4s, #0\n"
"ldr temploadreg1, [%[b_ptr0], #0x18]\n"
- ".word 0x4f80e218 // sdot v24.4s, v16.16b, v0.4b[0]\n"
+ ".inst 0x4f80e218 // sdot v24.4s, v16.16b, v0.4b[0]\n"
+ "str q26, [c_ptr2]\n"
+ "movi v26.4s, #0\n"
"ldr temploadreg2, [%[b_ptr0], #0x28]\n"
"ldr temploadreg3, [%[b_ptr0], #0x38]\n"
- "prfm PSTL1KEEP, [c_ptr1, #0x40]\n"
- ".word 0x4f82e219 // sdot v25.4s, v16.16b, v2.4b[0]\n"
- "str q26, [c_ptr2], #0x10\n"
- "movi v26.4s, #0\n"
- ".word 0x4fa0e238 // sdot v24.4s, v17.16b, v0.4b[1]\n"
- ".word 0x4fa2e239 // sdot v25.4s, v17.16b, v2.4b[1]\n"
- "str q27, [c_ptr3], #0x10\n"
+ "add c_ptr2, c_ptr2, #0x10\n"
+ ".inst 0x4f82e219 // sdot v25.4s, v16.16b, v2.4b[0]\n"
+ "str q27, [c_ptr3]\n"
"movi v27.4s, #0\n"
- "prfm PSTL1KEEP, [c_ptr2, #0x40]\n"
- ".word 0x4f84e21a // sdot v26.4s, v16.16b, v4.4b[0]\n"
- "str q28, [c_ptr4], #0x10\n"
+ "add c_ptr3, c_ptr3, #0x10\n"
+ ".inst 0x4f84e21a // sdot v26.4s, v16.16b, v4.4b[0]\n"
+ "str q28, [c_ptr4]\n"
"movi v28.4s, #0\n"
- "prfm PSTL1KEEP, [c_ptr3, #0x40]\n"
- ".word 0x4f86e21b // sdot v27.4s, v16.16b, v6.4b[0]\n"
- "str q29, [c_ptr5], #0x10\n"
+ "add c_ptr4, c_ptr4, #0x10\n"
+ ".inst 0x4f86e21b // sdot v27.4s, v16.16b, v6.4b[0]\n"
+ "str q29, [c_ptr5]\n"
"movi v29.4s, #0\n"
- "prfm PSTL1KEEP, [c_ptr4, #0x40]\n"
- ".word 0x4f88e21c // sdot v28.4s, v16.16b, v8.4b[0]\n"
- "str q30, [c_ptr6], #0x10\n"
+ "add c_ptr5, c_ptr5, #0x10\n"
+ ".inst 0x4f88e21c // sdot v28.4s, v16.16b, v8.4b[0]\n"
+ "str q30, [c_ptr6]\n"
"movi v30.4s, #0\n"
- "prfm PSTL1KEEP, [c_ptr5, #0x40]\n"
- ".word 0x4f8ae21d // sdot v29.4s, v16.16b, v10.4b[0]\n"
- "str q31, [c_ptr7], #0x10\n"
+ "add c_ptr6, c_ptr6, #0x10\n"
+ ".inst 0x4f8ae21d // sdot v29.4s, v16.16b, v10.4b[0]\n"
+ "str q31, [c_ptr7]\n"
"movi v31.4s, #0\n"
- "prfm PSTL1KEEP, [c_ptr6, #0x40]\n"
- ".word 0x4f8ce21e // sdot v30.4s, v16.16b, v12.4b[0]\n"
- "prfm PSTL1KEEP, [c_ptr7, #0x40]\n"
- ".word 0x4f8ee21f // sdot v31.4s, v16.16b, v14.4b[0]\n"
+ "add c_ptr7, c_ptr7, #0x10\n"
+ ".inst 0x4f8ce21e // sdot v30.4s, v16.16b, v12.4b[0]\n"
+ "prfm PSTL1KEEP, [%[c_ptr0], #0x40]\n"
+ ".inst 0x4f8ee21f // sdot v31.4s, v16.16b, v14.4b[0]\n"
"ldr d16, [%[b_ptr0]]\n"
- ".word 0x4fa4e23a // sdot v26.4s, v17.16b, v4.4b[1]\n"
- ".word 0x4fa6e23b // sdot v27.4s, v17.16b, v6.4b[1]\n"
- ".word 0x4fa8e23c // sdot v28.4s, v17.16b, v8.4b[1]\n"
+ ".inst 0x4fa0e238 // sdot v24.4s, v17.16b, v0.4b[1]\n"
+ "prfm PSTL1KEEP, [c_ptr1, #0x40]\n"
+ ".inst 0x4fa2e239 // sdot v25.4s, v17.16b, v2.4b[1]\n"
"ins v16.d[1], temploadreg0\n"
- ".word 0x4faae23d // sdot v29.4s, v17.16b, v10.4b[1]\n"
+ ".inst 0x4fa4e23a // sdot v26.4s, v17.16b, v4.4b[1]\n"
"ldr temploadreg0, [%[b_ptr0], #0x48]\n"
- ".word 0x4face23e // sdot v30.4s, v17.16b, v12.4b[1]\n"
- ".word 0x4faee23f // sdot v31.4s, v17.16b, v14.4b[1]\n"
+ ".inst 0x4fa6e23b // sdot v27.4s, v17.16b, v6.4b[1]\n"
+ "prfm PSTL1KEEP, [c_ptr2, #0x40]\n"
+ ".inst 0x4fa8e23c // sdot v28.4s, v17.16b, v8.4b[1]\n"
+ "prfm PSTL1KEEP, [c_ptr3, #0x40]\n"
+ ".inst 0x4faae23d // sdot v29.4s, v17.16b, v10.4b[1]\n"
+ "prfm PSTL1KEEP, [c_ptr4, #0x40]\n"
+ ".inst 0x4face23e // sdot v30.4s, v17.16b, v12.4b[1]\n"
+ "prfm PSTL1KEEP, [c_ptr5, #0x40]\n"
+ ".inst 0x4faee23f // sdot v31.4s, v17.16b, v14.4b[1]\n"
"ldr d17, [%[b_ptr0], #0x10]\n"
- ".word 0x4f80ea58 // sdot v24.4s, v18.16b, v0.4b[2]\n"
- ".word 0x4f82ea59 // sdot v25.4s, v18.16b, v2.4b[2]\n"
- ".word 0x4f84ea5a // sdot v26.4s, v18.16b, v4.4b[2]\n"
+ ".inst 0x4f80ea58 // sdot v24.4s, v18.16b, v0.4b[2]\n"
+ "prfm PSTL1KEEP, [c_ptr6, #0x40]\n"
+ ".inst 0x4f82ea59 // sdot v25.4s, v18.16b, v2.4b[2]\n"
"ins v17.d[1], temploadreg1\n"
- ".word 0x4f86ea5b // sdot v27.4s, v18.16b, v6.4b[2]\n"
- ".word 0x4f88ea5c // sdot v28.4s, v18.16b, v8.4b[2]\n"
- ".word 0x4f8aea5d // sdot v29.4s, v18.16b, v10.4b[2]\n"
- ".word 0x4f8cea5e // sdot v30.4s, v18.16b, v12.4b[2]\n"
- ".word 0x4f8eea5f // sdot v31.4s, v18.16b, v14.4b[2]\n"
+ ".inst 0x4f84ea5a // sdot v26.4s, v18.16b, v4.4b[2]\n"
+ "prfm PSTL1KEEP, [c_ptr7, #0x40]\n"
+ ".inst 0x4f86ea5b // sdot v27.4s, v18.16b, v6.4b[2]\n"
+ ".inst 0x4f88ea5c // sdot v28.4s, v18.16b, v8.4b[2]\n"
+ ".inst 0x4f8aea5d // sdot v29.4s, v18.16b, v10.4b[2]\n"
+ ".inst 0x4f8cea5e // sdot v30.4s, v18.16b, v12.4b[2]\n"
+ ".inst 0x4f8eea5f // sdot v31.4s, v18.16b, v14.4b[2]\n"
"ldr d18, [%[b_ptr0], #0x20]\n"
- ".word 0x4fa0ea78 // sdot v24.4s, v19.16b, v0.4b[3]\n"
- ".word 0x4fa2ea79 // sdot v25.4s, v19.16b, v2.4b[3]\n"
- ".word 0x4fa4ea7a // sdot v26.4s, v19.16b, v4.4b[3]\n"
+ ".inst 0x4fa0ea78 // sdot v24.4s, v19.16b, v0.4b[3]\n"
+ ".inst 0x4fa2ea79 // sdot v25.4s, v19.16b, v2.4b[3]\n"
+ ".inst 0x4fa4ea7a // sdot v26.4s, v19.16b, v4.4b[3]\n"
"ins v18.d[1], temploadreg2\n"
- ".word 0x4fa6ea7b // sdot v27.4s, v19.16b, v6.4b[3]\n"
- ".word 0x4fa8ea7c // sdot v28.4s, v19.16b, v8.4b[3]\n"
- ".word 0x4faaea7d // sdot v29.4s, v19.16b, v10.4b[3]\n"
- ".word 0x4facea7e // sdot v30.4s, v19.16b, v12.4b[3]\n"
- ".word 0x4faeea7f // sdot v31.4s, v19.16b, v14.4b[3]\n"
+ ".inst 0x4fa6ea7b // sdot v27.4s, v19.16b, v6.4b[3]\n"
+ ".inst 0x4fa8ea7c // sdot v28.4s, v19.16b, v8.4b[3]\n"
+ ".inst 0x4faaea7d // sdot v29.4s, v19.16b, v10.4b[3]\n"
+ ".inst 0x4facea7e // sdot v30.4s, v19.16b, v12.4b[3]\n"
+ ".inst 0x4faeea7f // sdot v31.4s, v19.16b, v14.4b[3]\n"
"ldr d19, [%[b_ptr0], #0x30]\n"
- ".word 0x4f81e298 // sdot v24.4s, v20.16b, v1.4b[0]\n"
- ".word 0x4f83e299 // sdot v25.4s, v20.16b, v3.4b[0]\n"
- ".word 0x4f85e29a // sdot v26.4s, v20.16b, v5.4b[0]\n"
+ ".inst 0x4f81e298 // sdot v24.4s, v20.16b, v1.4b[0]\n"
+ ".inst 0x4f83e299 // sdot v25.4s, v20.16b, v3.4b[0]\n"
+ ".inst 0x4f85e29a // sdot v26.4s, v20.16b, v5.4b[0]\n"
"ins v19.d[1], temploadreg3\n"
- ".word 0x4f87e29b // sdot v27.4s, v20.16b, v7.4b[0]\n"
- ".word 0x4f89e29c // sdot v28.4s, v20.16b, v9.4b[0]\n"
- ".word 0x4f8be29d // sdot v29.4s, v20.16b, v11.4b[0]\n"
- ".word 0x4f8de29e // sdot v30.4s, v20.16b, v13.4b[0]\n"
- ".word 0x4f8fe29f // sdot v31.4s, v20.16b, v15.4b[0]\n"
+ ".inst 0x4f87e29b // sdot v27.4s, v20.16b, v7.4b[0]\n"
+ ".inst 0x4f89e29c // sdot v28.4s, v20.16b, v9.4b[0]\n"
+ ".inst 0x4f8be29d // sdot v29.4s, v20.16b, v11.4b[0]\n"
+ ".inst 0x4f8de29e // sdot v30.4s, v20.16b, v13.4b[0]\n"
+ ".inst 0x4f8fe29f // sdot v31.4s, v20.16b, v15.4b[0]\n"
"ldr d20, [%[b_ptr0], #0x40]\n"
"add %[b_ptr0], %[b_ptr0], #0x50\n"
"b.ne 8b\n"
"7:\n"
- "str q24, [%[c_ptr0]], #0x10\n"
+ "str q24, [%[c_ptr0]]\n"
+ "add %[c_ptr0], %[c_ptr0], #0x10\n"
"movi v24.4s, #0\n"
"ins v20.d[1], temploadreg0\n"
- "str q25, [c_ptr1], #0x10\n"
+ "str q25, [c_ptr1]\n"
+ "add c_ptr1, c_ptr1, #0x10\n"
"movi v25.4s, #0\n"
- ".word 0x4f80e218 // sdot v24.4s, v16.16b, v0.4b[0]\n"
- "str q26, [c_ptr2], #0x10\n"
+ ".inst 0x4f80e218 // sdot v24.4s, v16.16b, v0.4b[0]\n"
+ "str q26, [c_ptr2]\n"
"movi v26.4s, #0\n"
- ".word 0x4f82e219 // sdot v25.4s, v16.16b, v2.4b[0]\n"
- ".word 0x4fa0e238 // sdot v24.4s, v17.16b, v0.4b[1]\n"
- "str q27, [c_ptr3], #0x10\n"
+ "add c_ptr2, c_ptr2, #0x10\n"
+ ".inst 0x4f82e219 // sdot v25.4s, v16.16b, v2.4b[0]\n"
+ "str q27, [c_ptr3]\n"
"movi v27.4s, #0\n"
- ".word 0x4f84e21a // sdot v26.4s, v16.16b, v4.4b[0]\n"
- ".word 0x4fa2e239 // sdot v25.4s, v17.16b, v2.4b[1]\n"
- "str q28, [c_ptr4], #0x10\n"
+ "add c_ptr3, c_ptr3, #0x10\n"
+ ".inst 0x4f84e21a // sdot v26.4s, v16.16b, v4.4b[0]\n"
+ "str q28, [c_ptr4]\n"
"movi v28.4s, #0\n"
- ".word 0x4f86e21b // sdot v27.4s, v16.16b, v6.4b[0]\n"
- ".word 0x4fa4e23a // sdot v26.4s, v17.16b, v4.4b[1]\n"
- "str q29, [c_ptr5], #0x10\n"
+ "add c_ptr4, c_ptr4, #0x10\n"
+ ".inst 0x4f86e21b // sdot v27.4s, v16.16b, v6.4b[0]\n"
+ "str q29, [c_ptr5]\n"
"movi v29.4s, #0\n"
- ".word 0x4f88e21c // sdot v28.4s, v16.16b, v8.4b[0]\n"
- ".word 0x4fa6e23b // sdot v27.4s, v17.16b, v6.4b[1]\n"
- "str q30, [c_ptr6], #0x10\n"
+ "add c_ptr5, c_ptr5, #0x10\n"
+ ".inst 0x4f88e21c // sdot v28.4s, v16.16b, v8.4b[0]\n"
+ "str q30, [c_ptr6]\n"
"movi v30.4s, #0\n"
- ".word 0x4f8ae21d // sdot v29.4s, v16.16b, v10.4b[0]\n"
- ".word 0x4fa8e23c // sdot v28.4s, v17.16b, v8.4b[1]\n"
- "str q31, [c_ptr7], #0x10\n"
+ "add c_ptr6, c_ptr6, #0x10\n"
+ ".inst 0x4f8ae21d // sdot v29.4s, v16.16b, v10.4b[0]\n"
+ "str q31, [c_ptr7]\n"
"movi v31.4s, #0\n"
- ".word 0x4f8ce21e // sdot v30.4s, v16.16b, v12.4b[0]\n"
- ".word 0x4faae23d // sdot v29.4s, v17.16b, v10.4b[1]\n"
- ".word 0x4f80ea58 // sdot v24.4s, v18.16b, v0.4b[2]\n"
- ".word 0x4f8ee21f // sdot v31.4s, v16.16b, v14.4b[0]\n"
- ".word 0x4face23e // sdot v30.4s, v17.16b, v12.4b[1]\n"
- ".word 0x4f82ea59 // sdot v25.4s, v18.16b, v2.4b[2]\n"
- ".word 0x4f84ea5a // sdot v26.4s, v18.16b, v4.4b[2]\n"
- ".word 0x4faee23f // sdot v31.4s, v17.16b, v14.4b[1]\n"
- ".word 0x4f86ea5b // sdot v27.4s, v18.16b, v6.4b[2]\n"
- ".word 0x4f88ea5c // sdot v28.4s, v18.16b, v8.4b[2]\n"
- ".word 0x4f8aea5d // sdot v29.4s, v18.16b, v10.4b[2]\n"
- ".word 0x4f8cea5e // sdot v30.4s, v18.16b, v12.4b[2]\n"
- ".word 0x4f8eea5f // sdot v31.4s, v18.16b, v14.4b[2]\n"
- ".word 0x4fa0ea78 // sdot v24.4s, v19.16b, v0.4b[3]\n"
- ".word 0x4fa2ea79 // sdot v25.4s, v19.16b, v2.4b[3]\n"
- ".word 0x4fa4ea7a // sdot v26.4s, v19.16b, v4.4b[3]\n"
- ".word 0x4fa6ea7b // sdot v27.4s, v19.16b, v6.4b[3]\n"
- ".word 0x4fa8ea7c // sdot v28.4s, v19.16b, v8.4b[3]\n"
- ".word 0x4faaea7d // sdot v29.4s, v19.16b, v10.4b[3]\n"
- ".word 0x4facea7e // sdot v30.4s, v19.16b, v12.4b[3]\n"
- ".word 0x4faeea7f // sdot v31.4s, v19.16b, v14.4b[3]\n"
- ".word 0x4f81e298 // sdot v24.4s, v20.16b, v1.4b[0]\n"
- ".word 0x4f83e299 // sdot v25.4s, v20.16b, v3.4b[0]\n"
- ".word 0x4f85e29a // sdot v26.4s, v20.16b, v5.4b[0]\n"
- ".word 0x4f87e29b // sdot v27.4s, v20.16b, v7.4b[0]\n"
- ".word 0x4f89e29c // sdot v28.4s, v20.16b, v9.4b[0]\n"
- ".word 0x4f8be29d // sdot v29.4s, v20.16b, v11.4b[0]\n"
- ".word 0x4f8de29e // sdot v30.4s, v20.16b, v13.4b[0]\n"
- ".word 0x4f8fe29f // sdot v31.4s, v20.16b, v15.4b[0]\n"
+ "add c_ptr7, c_ptr7, #0x10\n"
+ ".inst 0x4f8ce21e // sdot v30.4s, v16.16b, v12.4b[0]\n"
+ ".inst 0x4fa0e238 // sdot v24.4s, v17.16b, v0.4b[1]\n"
+ ".inst 0x4f8ee21f // sdot v31.4s, v16.16b, v14.4b[0]\n"
+ ".inst 0x4fa2e239 // sdot v25.4s, v17.16b, v2.4b[1]\n"
+ ".inst 0x4fa4e23a // sdot v26.4s, v17.16b, v4.4b[1]\n"
+ ".inst 0x4fa6e23b // sdot v27.4s, v17.16b, v6.4b[1]\n"
+ ".inst 0x4fa8e23c // sdot v28.4s, v17.16b, v8.4b[1]\n"
+ ".inst 0x4faae23d // sdot v29.4s, v17.16b, v10.4b[1]\n"
+ ".inst 0x4face23e // sdot v30.4s, v17.16b, v12.4b[1]\n"
+ ".inst 0x4faee23f // sdot v31.4s, v17.16b, v14.4b[1]\n"
+ ".inst 0x4f80ea58 // sdot v24.4s, v18.16b, v0.4b[2]\n"
+ ".inst 0x4f82ea59 // sdot v25.4s, v18.16b, v2.4b[2]\n"
+ ".inst 0x4f84ea5a // sdot v26.4s, v18.16b, v4.4b[2]\n"
+ ".inst 0x4f86ea5b // sdot v27.4s, v18.16b, v6.4b[2]\n"
+ ".inst 0x4f88ea5c // sdot v28.4s, v18.16b, v8.4b[2]\n"
+ ".inst 0x4f8aea5d // sdot v29.4s, v18.16b, v10.4b[2]\n"
+ ".inst 0x4f8cea5e // sdot v30.4s, v18.16b, v12.4b[2]\n"
+ ".inst 0x4f8eea5f // sdot v31.4s, v18.16b, v14.4b[2]\n"
+ ".inst 0x4fa0ea78 // sdot v24.4s, v19.16b, v0.4b[3]\n"
+ ".inst 0x4fa2ea79 // sdot v25.4s, v19.16b, v2.4b[3]\n"
+ ".inst 0x4fa4ea7a // sdot v26.4s, v19.16b, v4.4b[3]\n"
+ ".inst 0x4fa6ea7b // sdot v27.4s, v19.16b, v6.4b[3]\n"
+ ".inst 0x4fa8ea7c // sdot v28.4s, v19.16b, v8.4b[3]\n"
+ ".inst 0x4faaea7d // sdot v29.4s, v19.16b, v10.4b[3]\n"
+ ".inst 0x4facea7e // sdot v30.4s, v19.16b, v12.4b[3]\n"
+ ".inst 0x4faeea7f // sdot v31.4s, v19.16b, v14.4b[3]\n"
+ ".inst 0x4f81e298 // sdot v24.4s, v20.16b, v1.4b[0]\n"
+ ".inst 0x4f83e299 // sdot v25.4s, v20.16b, v3.4b[0]\n"
+ ".inst 0x4f85e29a // sdot v26.4s, v20.16b, v5.4b[0]\n"
+ ".inst 0x4f87e29b // sdot v27.4s, v20.16b, v7.4b[0]\n"
+ ".inst 0x4f89e29c // sdot v28.4s, v20.16b, v9.4b[0]\n"
+ ".inst 0x4f8be29d // sdot v29.4s, v20.16b, v11.4b[0]\n"
+ ".inst 0x4f8de29e // sdot v30.4s, v20.16b, v13.4b[0]\n"
+ ".inst 0x4f8fe29f // sdot v31.4s, v20.16b, v15.4b[0]\n"
"6:\n"
"str q24, [%[c_ptr0]]\n"
+ "add %[c_ptr0], %[c_ptr0], #0x10\n"
"str q25, [c_ptr1]\n"
+ "add c_ptr1, c_ptr1, #0x10\n"
"str q26, [c_ptr2]\n"
+ "add c_ptr2, c_ptr2, #0x10\n"
"str q27, [c_ptr3]\n"
+ "add c_ptr3, c_ptr3, #0x10\n"
"str q28, [c_ptr4]\n"
+ "add c_ptr4, c_ptr4, #0x10\n"
"str q29, [c_ptr5]\n"
+ "add c_ptr5, c_ptr5, #0x10\n"
"str q30, [c_ptr6]\n"
+ "add c_ptr6, c_ptr6, #0x10\n"
"str q31, [c_ptr7]\n"
+ "add c_ptr7, c_ptr7, #0x10\n"
".unreq a_ptr1\n"
".unreq a_ptr2\n"
".unreq a_ptr3\n"
@@ -1708,59 +1829,59 @@ void a64_smallK_hybrid_s8s32_dot_4x8_a55(const int8_t *A, int lda, const int8_t
"prfm PLDL1KEEP, [a_ptr7, #0x40]\n"
"movi v31.4s, #0\n"
"prfm PLDL1KEEP, [a_ptr7, #0x80]\n"
- ".word 0x4f80e218 // sdot v24.4s, v16.16b, v0.4b[0]\n"
+ ".inst 0x4f80e218 // sdot v24.4s, v16.16b, v0.4b[0]\n"
"prfm PLDL1KEEP, [a_ptr7, #0xc0]\n"
- ".word 0x4f82e219 // sdot v25.4s, v16.16b, v2.4b[0]\n"
+ ".inst 0x4f82e219 // sdot v25.4s, v16.16b, v2.4b[0]\n"
"prfm PLDL1KEEP, [a_ptr7, #0x100]\n"
- ".word 0x4f84e21a // sdot v26.4s, v16.16b, v4.4b[0]\n"
+ ".inst 0x4f84e21a // sdot v26.4s, v16.16b, v4.4b[0]\n"
"prfm PLDL1KEEP, [a_ptr7, #0x140]\n"
- ".word 0x4f86e21b // sdot v27.4s, v16.16b, v6.4b[0]\n"
+ ".inst 0x4f86e21b // sdot v27.4s, v16.16b, v6.4b[0]\n"
"prfm PLDL1KEEP, [a_ptr7, #0x180]\n"
- ".word 0x4f88e21c // sdot v28.4s, v16.16b, v8.4b[0]\n"
+ ".inst 0x4f88e21c // sdot v28.4s, v16.16b, v8.4b[0]\n"
"add %[b_ptr0], %[b_ptr0], #0x60\n"
- ".word 0x4f8ae21d // sdot v29.4s, v16.16b, v10.4b[0]\n"
- ".word 0x4f8ce21e // sdot v30.4s, v16.16b, v12.4b[0]\n"
- ".word 0x4f8ee21f // sdot v31.4s, v16.16b, v14.4b[0]\n"
- ".word 0x4fa0e238 // sdot v24.4s, v17.16b, v0.4b[1]\n"
- ".word 0x4fa2e239 // sdot v25.4s, v17.16b, v2.4b[1]\n"
- ".word 0x4fa4e23a // sdot v26.4s, v17.16b, v4.4b[1]\n"
- ".word 0x4fa6e23b // sdot v27.4s, v17.16b, v6.4b[1]\n"
- ".word 0x4fa8e23c // sdot v28.4s, v17.16b, v8.4b[1]\n"
- ".word 0x4faae23d // sdot v29.4s, v17.16b, v10.4b[1]\n"
- ".word 0x4face23e // sdot v30.4s, v17.16b, v12.4b[1]\n"
- ".word 0x4faee23f // sdot v31.4s, v17.16b, v14.4b[1]\n"
- ".word 0x4f80ea58 // sdot v24.4s, v18.16b, v0.4b[2]\n"
- ".word 0x4f82ea59 // sdot v25.4s, v18.16b, v2.4b[2]\n"
- ".word 0x4f84ea5a // sdot v26.4s, v18.16b, v4.4b[2]\n"
- ".word 0x4f86ea5b // sdot v27.4s, v18.16b, v6.4b[2]\n"
- ".word 0x4f88ea5c // sdot v28.4s, v18.16b, v8.4b[2]\n"
- ".word 0x4f8aea5d // sdot v29.4s, v18.16b, v10.4b[2]\n"
- ".word 0x4f8cea5e // sdot v30.4s, v18.16b, v12.4b[2]\n"
- ".word 0x4f8eea5f // sdot v31.4s, v18.16b, v14.4b[2]\n"
- ".word 0x4fa0ea78 // sdot v24.4s, v19.16b, v0.4b[3]\n"
- ".word 0x4fa2ea79 // sdot v25.4s, v19.16b, v2.4b[3]\n"
- ".word 0x4fa4ea7a // sdot v26.4s, v19.16b, v4.4b[3]\n"
- ".word 0x4fa6ea7b // sdot v27.4s, v19.16b, v6.4b[3]\n"
- ".word 0x4fa8ea7c // sdot v28.4s, v19.16b, v8.4b[3]\n"
- ".word 0x4faaea7d // sdot v29.4s, v19.16b, v10.4b[3]\n"
- ".word 0x4facea7e // sdot v30.4s, v19.16b, v12.4b[3]\n"
- ".word 0x4faeea7f // sdot v31.4s, v19.16b, v14.4b[3]\n"
- ".word 0x4f81e298 // sdot v24.4s, v20.16b, v1.4b[0]\n"
- ".word 0x4f83e299 // sdot v25.4s, v20.16b, v3.4b[0]\n"
- ".word 0x4f85e29a // sdot v26.4s, v20.16b, v5.4b[0]\n"
- ".word 0x4f87e29b // sdot v27.4s, v20.16b, v7.4b[0]\n"
- ".word 0x4f89e29c // sdot v28.4s, v20.16b, v9.4b[0]\n"
- ".word 0x4f8be29d // sdot v29.4s, v20.16b, v11.4b[0]\n"
- ".word 0x4f8de29e // sdot v30.4s, v20.16b, v13.4b[0]\n"
- ".word 0x4f8fe29f // sdot v31.4s, v20.16b, v15.4b[0]\n"
- ".word 0x4fa1e2b8 // sdot v24.4s, v21.16b, v1.4b[1]\n"
- ".word 0x4fa3e2b9 // sdot v25.4s, v21.16b, v3.4b[1]\n"
- ".word 0x4fa5e2ba // sdot v26.4s, v21.16b, v5.4b[1]\n"
- ".word 0x4fa7e2bb // sdot v27.4s, v21.16b, v7.4b[1]\n"
- ".word 0x4fa9e2bc // sdot v28.4s, v21.16b, v9.4b[1]\n"
- ".word 0x4fabe2bd // sdot v29.4s, v21.16b, v11.4b[1]\n"
- ".word 0x4fade2be // sdot v30.4s, v21.16b, v13.4b[1]\n"
- ".word 0x4fafe2bf // sdot v31.4s, v21.16b, v15.4b[1]\n"
+ ".inst 0x4f8ae21d // sdot v29.4s, v16.16b, v10.4b[0]\n"
+ ".inst 0x4f8ce21e // sdot v30.4s, v16.16b, v12.4b[0]\n"
+ ".inst 0x4f8ee21f // sdot v31.4s, v16.16b, v14.4b[0]\n"
+ ".inst 0x4fa0e238 // sdot v24.4s, v17.16b, v0.4b[1]\n"
+ ".inst 0x4fa2e239 // sdot v25.4s, v17.16b, v2.4b[1]\n"
+ ".inst 0x4fa4e23a // sdot v26.4s, v17.16b, v4.4b[1]\n"
+ ".inst 0x4fa6e23b // sdot v27.4s, v17.16b, v6.4b[1]\n"
+ ".inst 0x4fa8e23c // sdot v28.4s, v17.16b, v8.4b[1]\n"
+ ".inst 0x4faae23d // sdot v29.4s, v17.16b, v10.4b[1]\n"
+ ".inst 0x4face23e // sdot v30.4s, v17.16b, v12.4b[1]\n"
+ ".inst 0x4faee23f // sdot v31.4s, v17.16b, v14.4b[1]\n"
+ ".inst 0x4f80ea58 // sdot v24.4s, v18.16b, v0.4b[2]\n"
+ ".inst 0x4f82ea59 // sdot v25.4s, v18.16b, v2.4b[2]\n"
+ ".inst 0x4f84ea5a // sdot v26.4s, v18.16b, v4.4b[2]\n"
+ ".inst 0x4f86ea5b // sdot v27.4s, v18.16b, v6.4b[2]\n"
+ ".inst 0x4f88ea5c // sdot v28.4s, v18.16b, v8.4b[2]\n"
+ ".inst 0x4f8aea5d // sdot v29.4s, v18.16b, v10.4b[2]\n"
+ ".inst 0x4f8cea5e // sdot v30.4s, v18.16b, v12.4b[2]\n"
+ ".inst 0x4f8eea5f // sdot v31.4s, v18.16b, v14.4b[2]\n"
+ ".inst 0x4fa0ea78 // sdot v24.4s, v19.16b, v0.4b[3]\n"
+ ".inst 0x4fa2ea79 // sdot v25.4s, v19.16b, v2.4b[3]\n"
+ ".inst 0x4fa4ea7a // sdot v26.4s, v19.16b, v4.4b[3]\n"
+ ".inst 0x4fa6ea7b // sdot v27.4s, v19.16b, v6.4b[3]\n"
+ ".inst 0x4fa8ea7c // sdot v28.4s, v19.16b, v8.4b[3]\n"
+ ".inst 0x4faaea7d // sdot v29.4s, v19.16b, v10.4b[3]\n"
+ ".inst 0x4facea7e // sdot v30.4s, v19.16b, v12.4b[3]\n"
+ ".inst 0x4faeea7f // sdot v31.4s, v19.16b, v14.4b[3]\n"
+ ".inst 0x4f81e298 // sdot v24.4s, v20.16b, v1.4b[0]\n"
+ ".inst 0x4f83e299 // sdot v25.4s, v20.16b, v3.4b[0]\n"
+ ".inst 0x4f85e29a // sdot v26.4s, v20.16b, v5.4b[0]\n"
+ ".inst 0x4f87e29b // sdot v27.4s, v20.16b, v7.4b[0]\n"
+ ".inst 0x4f89e29c // sdot v28.4s, v20.16b, v9.4b[0]\n"
+ ".inst 0x4f8be29d // sdot v29.4s, v20.16b, v11.4b[0]\n"
+ ".inst 0x4f8de29e // sdot v30.4s, v20.16b, v13.4b[0]\n"
+ ".inst 0x4f8fe29f // sdot v31.4s, v20.16b, v15.4b[0]\n"
+ ".inst 0x4fa1e2b8 // sdot v24.4s, v21.16b, v1.4b[1]\n"
+ ".inst 0x4fa3e2b9 // sdot v25.4s, v21.16b, v3.4b[1]\n"
+ ".inst 0x4fa5e2ba // sdot v26.4s, v21.16b, v5.4b[1]\n"
+ ".inst 0x4fa7e2bb // sdot v27.4s, v21.16b, v7.4b[1]\n"
+ ".inst 0x4fa9e2bc // sdot v28.4s, v21.16b, v9.4b[1]\n"
+ ".inst 0x4fabe2bd // sdot v29.4s, v21.16b, v11.4b[1]\n"
+ ".inst 0x4fade2be // sdot v30.4s, v21.16b, v13.4b[1]\n"
+ ".inst 0x4fafe2bf // sdot v31.4s, v21.16b, v15.4b[1]\n"
"cbz %[loops], 6f\n"
"ldr d16, [%[b_ptr0]]\n"
"subs %[loops], %[loops], #0x1\n"
@@ -1783,174 +1904,198 @@ void a64_smallK_hybrid_s8s32_dot_4x8_a55(const int8_t *A, int lda, const int8_t
"ins v20.d[1], temploadreg0\n"
"b.eq 7f\n"
"8:\n"
- "str q24, [%[c_ptr0]], #0x10\n"
+ "str q24, [%[c_ptr0]]\n"
"subs %[loops], %[loops], #0x1\n"
"movi v24.4s, #0\n"
"ins v21.d[1], temploadreg1\n"
"ldr temploadreg0, [%[b_ptr0], #0x8]\n"
- "prfm PSTL1KEEP, [%[c_ptr0], #0x40]\n"
- "str q25, [c_ptr1], #0x10\n"
+ "add %[c_ptr0], %[c_ptr0], #0x10\n"
+ "str q25, [c_ptr1]\n"
+ "add c_ptr1, c_ptr1, #0x10\n"
"movi v25.4s, #0\n"
"ldr temploadreg1, [%[b_ptr0], #0x18]\n"
- ".word 0x4f80e218 // sdot v24.4s, v16.16b, v0.4b[0]\n"
+ ".inst 0x4f80e218 // sdot v24.4s, v16.16b, v0.4b[0]\n"
+ "str q26, [c_ptr2]\n"
+ "movi v26.4s, #0\n"
"ldr temploadreg2, [%[b_ptr0], #0x28]\n"
"ldr temploadreg3, [%[b_ptr0], #0x38]\n"
- "prfm PSTL1KEEP, [c_ptr1, #0x40]\n"
- ".word 0x4f82e219 // sdot v25.4s, v16.16b, v2.4b[0]\n"
- "str q26, [c_ptr2], #0x10\n"
- "movi v26.4s, #0\n"
- ".word 0x4fa0e238 // sdot v24.4s, v17.16b, v0.4b[1]\n"
- ".word 0x4fa2e239 // sdot v25.4s, v17.16b, v2.4b[1]\n"
- "str q27, [c_ptr3], #0x10\n"
+ "add c_ptr2, c_ptr2, #0x10\n"
+ ".inst 0x4f82e219 // sdot v25.4s, v16.16b, v2.4b[0]\n"
+ "str q27, [c_ptr3]\n"
"movi v27.4s, #0\n"
- "prfm PSTL1KEEP, [c_ptr2, #0x40]\n"
- ".word 0x4f84e21a // sdot v26.4s, v16.16b, v4.4b[0]\n"
- "str q28, [c_ptr4], #0x10\n"
+ "add c_ptr3, c_ptr3, #0x10\n"
+ ".inst 0x4f84e21a // sdot v26.4s, v16.16b, v4.4b[0]\n"
+ "str q28, [c_ptr4]\n"
"movi v28.4s, #0\n"
- "prfm PSTL1KEEP, [c_ptr3, #0x40]\n"
- ".word 0x4f86e21b // sdot v27.4s, v16.16b, v6.4b[0]\n"
- "str q29, [c_ptr5], #0x10\n"
+ "add c_ptr4, c_ptr4, #0x10\n"
+ ".inst 0x4f86e21b // sdot v27.4s, v16.16b, v6.4b[0]\n"
+ "str q29, [c_ptr5]\n"
"movi v29.4s, #0\n"
- "prfm PSTL1KEEP, [c_ptr4, #0x40]\n"
- ".word 0x4f88e21c // sdot v28.4s, v16.16b, v8.4b[0]\n"
- "str q30, [c_ptr6], #0x10\n"
+ "add c_ptr5, c_ptr5, #0x10\n"
+ ".inst 0x4f88e21c // sdot v28.4s, v16.16b, v8.4b[0]\n"
+ "str q30, [c_ptr6]\n"
"movi v30.4s, #0\n"
- "prfm PSTL1KEEP, [c_ptr5, #0x40]\n"
- ".word 0x4f8ae21d // sdot v29.4s, v16.16b, v10.4b[0]\n"
- "str q31, [c_ptr7], #0x10\n"
+ "add c_ptr6, c_ptr6, #0x10\n"
+ ".inst 0x4f8ae21d // sdot v29.4s, v16.16b, v10.4b[0]\n"
+ "str q31, [c_ptr7]\n"
"movi v31.4s, #0\n"
- "prfm PSTL1KEEP, [c_ptr6, #0x40]\n"
- ".word 0x4f8ce21e // sdot v30.4s, v16.16b, v12.4b[0]\n"
- "prfm PSTL1KEEP, [c_ptr7, #0x40]\n"
- ".word 0x4f8ee21f // sdot v31.4s, v16.16b, v14.4b[0]\n"
+ "add c_ptr7, c_ptr7, #0x10\n"
+ ".inst 0x4f8ce21e // sdot v30.4s, v16.16b, v12.4b[0]\n"
+ "prfm PSTL1KEEP, [%[c_ptr0], #0x40]\n"
+ ".inst 0x4f8ee21f // sdot v31.4s, v16.16b, v14.4b[0]\n"
"ldr d16, [%[b_ptr0]]\n"
- ".word 0x4fa4e23a // sdot v26.4s, v17.16b, v4.4b[1]\n"
- ".word 0x4fa6e23b // sdot v27.4s, v17.16b, v6.4b[1]\n"
- ".word 0x4fa8e23c // sdot v28.4s, v17.16b, v8.4b[1]\n"
+ ".inst 0x4fa0e238 // sdot v24.4s, v17.16b, v0.4b[1]\n"
+ "prfm PSTL1KEEP, [c_ptr1, #0x40]\n"
+ ".inst 0x4fa2e239 // sdot v25.4s, v17.16b, v2.4b[1]\n"
"ins v16.d[1], temploadreg0\n"
- ".word 0x4faae23d // sdot v29.4s, v17.16b, v10.4b[1]\n"
+ ".inst 0x4fa4e23a // sdot v26.4s, v17.16b, v4.4b[1]\n"
"ldr temploadreg0, [%[b_ptr0], #0x48]\n"
- ".word 0x4face23e // sdot v30.4s, v17.16b, v12.4b[1]\n"
- ".word 0x4faee23f // sdot v31.4s, v17.16b, v14.4b[1]\n"
+ ".inst 0x4fa6e23b // sdot v27.4s, v17.16b, v6.4b[1]\n"
+ "prfm PSTL1KEEP, [c_ptr2, #0x40]\n"
+ ".inst 0x4fa8e23c // sdot v28.4s, v17.16b, v8.4b[1]\n"
+ "prfm PSTL1KEEP, [c_ptr3, #0x40]\n"
+ ".inst 0x4faae23d // sdot v29.4s, v17.16b, v10.4b[1]\n"
+ "prfm PSTL1KEEP, [c_ptr4, #0x40]\n"
+ ".inst 0x4face23e // sdot v30.4s, v17.16b, v12.4b[1]\n"
+ "prfm PSTL1KEEP, [c_ptr5, #0x40]\n"
+ ".inst 0x4faee23f // sdot v31.4s, v17.16b, v14.4b[1]\n"
"ldr d17, [%[b_ptr0], #0x10]\n"
- ".word 0x4f80ea58 // sdot v24.4s, v18.16b, v0.4b[2]\n"
- ".word 0x4f82ea59 // sdot v25.4s, v18.16b, v2.4b[2]\n"
- ".word 0x4f84ea5a // sdot v26.4s, v18.16b, v4.4b[2]\n"
+ ".inst 0x4f80ea58 // sdot v24.4s, v18.16b, v0.4b[2]\n"
+ "prfm PSTL1KEEP, [c_ptr6, #0x40]\n"
+ ".inst 0x4f82ea59 // sdot v25.4s, v18.16b, v2.4b[2]\n"
"ins v17.d[1], temploadreg1\n"
- ".word 0x4f86ea5b // sdot v27.4s, v18.16b, v6.4b[2]\n"
+ ".inst 0x4f84ea5a // sdot v26.4s, v18.16b, v4.4b[2]\n"
"ldr temploadreg1, [%[b_ptr0], #0x58]\n"
- ".word 0x4f88ea5c // sdot v28.4s, v18.16b, v8.4b[2]\n"
- ".word 0x4f8aea5d // sdot v29.4s, v18.16b, v10.4b[2]\n"
- ".word 0x4f8cea5e // sdot v30.4s, v18.16b, v12.4b[2]\n"
- ".word 0x4f8eea5f // sdot v31.4s, v18.16b, v14.4b[2]\n"
+ ".inst 0x4f86ea5b // sdot v27.4s, v18.16b, v6.4b[2]\n"
+ "prfm PSTL1KEEP, [c_ptr7, #0x40]\n"
+ ".inst 0x4f88ea5c // sdot v28.4s, v18.16b, v8.4b[2]\n"
+ ".inst 0x4f8aea5d // sdot v29.4s, v18.16b, v10.4b[2]\n"
+ ".inst 0x4f8cea5e // sdot v30.4s, v18.16b, v12.4b[2]\n"
+ ".inst 0x4f8eea5f // sdot v31.4s, v18.16b, v14.4b[2]\n"
"ldr d18, [%[b_ptr0], #0x20]\n"
- ".word 0x4fa0ea78 // sdot v24.4s, v19.16b, v0.4b[3]\n"
- ".word 0x4fa2ea79 // sdot v25.4s, v19.16b, v2.4b[3]\n"
- ".word 0x4fa4ea7a // sdot v26.4s, v19.16b, v4.4b[3]\n"
+ ".inst 0x4fa0ea78 // sdot v24.4s, v19.16b, v0.4b[3]\n"
+ ".inst 0x4fa2ea79 // sdot v25.4s, v19.16b, v2.4b[3]\n"
+ ".inst 0x4fa4ea7a // sdot v26.4s, v19.16b, v4.4b[3]\n"
"ins v18.d[1], temploadreg2\n"
- ".word 0x4fa6ea7b // sdot v27.4s, v19.16b, v6.4b[3]\n"
- ".word 0x4fa8ea7c // sdot v28.4s, v19.16b, v8.4b[3]\n"
- ".word 0x4faaea7d // sdot v29.4s, v19.16b, v10.4b[3]\n"
- ".word 0x4facea7e // sdot v30.4s, v19.16b, v12.4b[3]\n"
- ".word 0x4faeea7f // sdot v31.4s, v19.16b, v14.4b[3]\n"
+ ".inst 0x4fa6ea7b // sdot v27.4s, v19.16b, v6.4b[3]\n"
+ ".inst 0x4fa8ea7c // sdot v28.4s, v19.16b, v8.4b[3]\n"
+ ".inst 0x4faaea7d // sdot v29.4s, v19.16b, v10.4b[3]\n"
+ ".inst 0x4facea7e // sdot v30.4s, v19.16b, v12.4b[3]\n"
+ ".inst 0x4faeea7f // sdot v31.4s, v19.16b, v14.4b[3]\n"
"ldr d19, [%[b_ptr0], #0x30]\n"
- ".word 0x4f81e298 // sdot v24.4s, v20.16b, v1.4b[0]\n"
- ".word 0x4f83e299 // sdot v25.4s, v20.16b, v3.4b[0]\n"
- ".word 0x4f85e29a // sdot v26.4s, v20.16b, v5.4b[0]\n"
+ ".inst 0x4f81e298 // sdot v24.4s, v20.16b, v1.4b[0]\n"
+ ".inst 0x4f83e299 // sdot v25.4s, v20.16b, v3.4b[0]\n"
+ ".inst 0x4f85e29a // sdot v26.4s, v20.16b, v5.4b[0]\n"
"ins v19.d[1], temploadreg3\n"
- ".word 0x4f87e29b // sdot v27.4s, v20.16b, v7.4b[0]\n"
- ".word 0x4f89e29c // sdot v28.4s, v20.16b, v9.4b[0]\n"
- ".word 0x4f8be29d // sdot v29.4s, v20.16b, v11.4b[0]\n"
- ".word 0x4f8de29e // sdot v30.4s, v20.16b, v13.4b[0]\n"
- ".word 0x4f8fe29f // sdot v31.4s, v20.16b, v15.4b[0]\n"
+ ".inst 0x4f87e29b // sdot v27.4s, v20.16b, v7.4b[0]\n"
+ ".inst 0x4f89e29c // sdot v28.4s, v20.16b, v9.4b[0]\n"
+ ".inst 0x4f8be29d // sdot v29.4s, v20.16b, v11.4b[0]\n"
+ ".inst 0x4f8de29e // sdot v30.4s, v20.16b, v13.4b[0]\n"
+ ".inst 0x4f8fe29f // sdot v31.4s, v20.16b, v15.4b[0]\n"
"ldr d20, [%[b_ptr0], #0x40]\n"
- ".word 0x4fa1e2b8 // sdot v24.4s, v21.16b, v1.4b[1]\n"
- ".word 0x4fa3e2b9 // sdot v25.4s, v21.16b, v3.4b[1]\n"
- ".word 0x4fa5e2ba // sdot v26.4s, v21.16b, v5.4b[1]\n"
+ ".inst 0x4fa1e2b8 // sdot v24.4s, v21.16b, v1.4b[1]\n"
+ ".inst 0x4fa3e2b9 // sdot v25.4s, v21.16b, v3.4b[1]\n"
+ ".inst 0x4fa5e2ba // sdot v26.4s, v21.16b, v5.4b[1]\n"
"ins v20.d[1], temploadreg0\n"
- ".word 0x4fa7e2bb // sdot v27.4s, v21.16b, v7.4b[1]\n"
- ".word 0x4fa9e2bc // sdot v28.4s, v21.16b, v9.4b[1]\n"
- ".word 0x4fabe2bd // sdot v29.4s, v21.16b, v11.4b[1]\n"
- ".word 0x4fade2be // sdot v30.4s, v21.16b, v13.4b[1]\n"
- ".word 0x4fafe2bf // sdot v31.4s, v21.16b, v15.4b[1]\n"
+ ".inst 0x4fa7e2bb // sdot v27.4s, v21.16b, v7.4b[1]\n"
+ ".inst 0x4fa9e2bc // sdot v28.4s, v21.16b, v9.4b[1]\n"
+ ".inst 0x4fabe2bd // sdot v29.4s, v21.16b, v11.4b[1]\n"
+ ".inst 0x4fade2be // sdot v30.4s, v21.16b, v13.4b[1]\n"
+ ".inst 0x4fafe2bf // sdot v31.4s, v21.16b, v15.4b[1]\n"
"ldr d21, [%[b_ptr0], #0x50]\n"
"add %[b_ptr0], %[b_ptr0], #0x60\n"
"b.ne 8b\n"
"7:\n"
- "str q24, [%[c_ptr0]], #0x10\n"
+ "str q24, [%[c_ptr0]]\n"
+ "add %[c_ptr0], %[c_ptr0], #0x10\n"
"movi v24.4s, #0\n"
"ins v21.d[1], temploadreg1\n"
- "str q25, [c_ptr1], #0x10\n"
+ "str q25, [c_ptr1]\n"
+ "add c_ptr1, c_ptr1, #0x10\n"
"movi v25.4s, #0\n"
- ".word 0x4f80e218 // sdot v24.4s, v16.16b, v0.4b[0]\n"
- "str q26, [c_ptr2], #0x10\n"
+ ".inst 0x4f80e218 // sdot v24.4s, v16.16b, v0.4b[0]\n"
+ "str q26, [c_ptr2]\n"
"movi v26.4s, #0\n"
- ".word 0x4f82e219 // sdot v25.4s, v16.16b, v2.4b[0]\n"
- ".word 0x4fa0e238 // sdot v24.4s, v17.16b, v0.4b[1]\n"
- "str q27, [c_ptr3], #0x10\n"
+ "add c_ptr2, c_ptr2, #0x10\n"
+ ".inst 0x4f82e219 // sdot v25.4s, v16.16b, v2.4b[0]\n"
+ "str q27, [c_ptr3]\n"
"movi v27.4s, #0\n"
- ".word 0x4f84e21a // sdot v26.4s, v16.16b, v4.4b[0]\n"
- ".word 0x4fa2e239 // sdot v25.4s, v17.16b, v2.4b[1]\n"
- "str q28, [c_ptr4], #0x10\n"
+ "add c_ptr3, c_ptr3, #0x10\n"
+ ".inst 0x4f84e21a // sdot v26.4s, v16.16b, v4.4b[0]\n"
+ "str q28, [c_ptr4]\n"
"movi v28.4s, #0\n"
- ".word 0x4f86e21b // sdot v27.4s, v16.16b, v6.4b[0]\n"
- ".word 0x4fa4e23a // sdot v26.4s, v17.16b, v4.4b[1]\n"
- "str q29, [c_ptr5], #0x10\n"
+ "add c_ptr4, c_ptr4, #0x10\n"
+ ".inst 0x4f86e21b // sdot v27.4s, v16.16b, v6.4b[0]\n"
+ "str q29, [c_ptr5]\n"
"movi v29.4s, #0\n"
- ".word 0x4f88e21c // sdot v28.4s, v16.16b, v8.4b[0]\n"
- ".word 0x4fa6e23b // sdot v27.4s, v17.16b, v6.4b[1]\n"
- "str q30, [c_ptr6], #0x10\n"
+ "add c_ptr5, c_ptr5, #0x10\n"
+ ".inst 0x4f88e21c // sdot v28.4s, v16.16b, v8.4b[0]\n"
+ "str q30, [c_ptr6]\n"
"movi v30.4s, #0\n"
- ".word 0x4f8ae21d // sdot v29.4s, v16.16b, v10.4b[0]\n"
- ".word 0x4fa8e23c // sdot v28.4s, v17.16b, v8.4b[1]\n"
- "str q31, [c_ptr7], #0x10\n"
+ "add c_ptr6, c_ptr6, #0x10\n"
+ ".inst 0x4f8ae21d // sdot v29.4s, v16.16b, v10.4b[0]\n"
+ "str q31, [c_ptr7]\n"
"movi v31.4s, #0\n"
- ".word 0x4f8ce21e // sdot v30.4s, v16.16b, v12.4b[0]\n"
- ".word 0x4faae23d // sdot v29.4s, v17.16b, v10.4b[1]\n"
- ".word 0x4f80ea58 // sdot v24.4s, v18.16b, v0.4b[2]\n"
- ".word 0x4f8ee21f // sdot v31.4s, v16.16b, v14.4b[0]\n"
- ".word 0x4face23e // sdot v30.4s, v17.16b, v12.4b[1]\n"
- ".word 0x4f82ea59 // sdot v25.4s, v18.16b, v2.4b[2]\n"
- ".word 0x4f84ea5a // sdot v26.4s, v18.16b, v4.4b[2]\n"
- ".word 0x4faee23f // sdot v31.4s, v17.16b, v14.4b[1]\n"
- ".word 0x4f86ea5b // sdot v27.4s, v18.16b, v6.4b[2]\n"
- ".word 0x4f88ea5c // sdot v28.4s, v18.16b, v8.4b[2]\n"
- ".word 0x4f8aea5d // sdot v29.4s, v18.16b, v10.4b[2]\n"
- ".word 0x4f8cea5e // sdot v30.4s, v18.16b, v12.4b[2]\n"
- ".word 0x4f8eea5f // sdot v31.4s, v18.16b, v14.4b[2]\n"
- ".word 0x4fa0ea78 // sdot v24.4s, v19.16b, v0.4b[3]\n"
- ".word 0x4fa2ea79 // sdot v25.4s, v19.16b, v2.4b[3]\n"
- ".word 0x4fa4ea7a // sdot v26.4s, v19.16b, v4.4b[3]\n"
- ".word 0x4fa6ea7b // sdot v27.4s, v19.16b, v6.4b[3]\n"
- ".word 0x4fa8ea7c // sdot v28.4s, v19.16b, v8.4b[3]\n"
- ".word 0x4faaea7d // sdot v29.4s, v19.16b, v10.4b[3]\n"
- ".word 0x4facea7e // sdot v30.4s, v19.16b, v12.4b[3]\n"
- ".word 0x4faeea7f // sdot v31.4s, v19.16b, v14.4b[3]\n"
- ".word 0x4f81e298 // sdot v24.4s, v20.16b, v1.4b[0]\n"
- ".word 0x4f83e299 // sdot v25.4s, v20.16b, v3.4b[0]\n"
- ".word 0x4f85e29a // sdot v26.4s, v20.16b, v5.4b[0]\n"
- ".word 0x4f87e29b // sdot v27.4s, v20.16b, v7.4b[0]\n"
- ".word 0x4f89e29c // sdot v28.4s, v20.16b, v9.4b[0]\n"
- ".word 0x4f8be29d // sdot v29.4s, v20.16b, v11.4b[0]\n"
- ".word 0x4f8de29e // sdot v30.4s, v20.16b, v13.4b[0]\n"
- ".word 0x4f8fe29f // sdot v31.4s, v20.16b, v15.4b[0]\n"
- ".word 0x4fa1e2b8 // sdot v24.4s, v21.16b, v1.4b[1]\n"
- ".word 0x4fa3e2b9 // sdot v25.4s, v21.16b, v3.4b[1]\n"
- ".word 0x4fa5e2ba // sdot v26.4s, v21.16b, v5.4b[1]\n"
- ".word 0x4fa7e2bb // sdot v27.4s, v21.16b, v7.4b[1]\n"
- ".word 0x4fa9e2bc // sdot v28.4s, v21.16b, v9.4b[1]\n"
- ".word 0x4fabe2bd // sdot v29.4s, v21.16b, v11.4b[1]\n"
- ".word 0x4fade2be // sdot v30.4s, v21.16b, v13.4b[1]\n"
- ".word 0x4fafe2bf // sdot v31.4s, v21.16b, v15.4b[1]\n"
+ "add c_ptr7, c_ptr7, #0x10\n"
+ ".inst 0x4f8ce21e // sdot v30.4s, v16.16b, v12.4b[0]\n"
+ ".inst 0x4fa0e238 // sdot v24.4s, v17.16b, v0.4b[1]\n"
+ ".inst 0x4f8ee21f // sdot v31.4s, v16.16b, v14.4b[0]\n"
+ ".inst 0x4fa2e239 // sdot v25.4s, v17.16b, v2.4b[1]\n"
+ ".inst 0x4fa4e23a // sdot v26.4s, v17.16b, v4.4b[1]\n"
+ ".inst 0x4fa6e23b // sdot v27.4s, v17.16b, v6.4b[1]\n"
+ ".inst 0x4fa8e23c // sdot v28.4s, v17.16b, v8.4b[1]\n"
+ ".inst 0x4faae23d // sdot v29.4s, v17.16b, v10.4b[1]\n"
+ ".inst 0x4face23e // sdot v30.4s, v17.16b, v12.4b[1]\n"
+ ".inst 0x4faee23f // sdot v31.4s, v17.16b, v14.4b[1]\n"
+ ".inst 0x4f80ea58 // sdot v24.4s, v18.16b, v0.4b[2]\n"
+ ".inst 0x4f82ea59 // sdot v25.4s, v18.16b, v2.4b[2]\n"
+ ".inst 0x4f84ea5a // sdot v26.4s, v18.16b, v4.4b[2]\n"
+ ".inst 0x4f86ea5b // sdot v27.4s, v18.16b, v6.4b[2]\n"
+ ".inst 0x4f88ea5c // sdot v28.4s, v18.16b, v8.4b[2]\n"
+ ".inst 0x4f8aea5d // sdot v29.4s, v18.16b, v10.4b[2]\n"
+ ".inst 0x4f8cea5e // sdot v30.4s, v18.16b, v12.4b[2]\n"
+ ".inst 0x4f8eea5f // sdot v31.4s, v18.16b, v14.4b[2]\n"
+ ".inst 0x4fa0ea78 // sdot v24.4s, v19.16b, v0.4b[3]\n"
+ ".inst 0x4fa2ea79 // sdot v25.4s, v19.16b, v2.4b[3]\n"
+ ".inst 0x4fa4ea7a // sdot v26.4s, v19.16b, v4.4b[3]\n"
+ ".inst 0x4fa6ea7b // sdot v27.4s, v19.16b, v6.4b[3]\n"
+ ".inst 0x4fa8ea7c // sdot v28.4s, v19.16b, v8.4b[3]\n"
+ ".inst 0x4faaea7d // sdot v29.4s, v19.16b, v10.4b[3]\n"
+ ".inst 0x4facea7e // sdot v30.4s, v19.16b, v12.4b[3]\n"
+ ".inst 0x4faeea7f // sdot v31.4s, v19.16b, v14.4b[3]\n"
+ ".inst 0x4f81e298 // sdot v24.4s, v20.16b, v1.4b[0]\n"
+ ".inst 0x4f83e299 // sdot v25.4s, v20.16b, v3.4b[0]\n"
+ ".inst 0x4f85e29a // sdot v26.4s, v20.16b, v5.4b[0]\n"
+ ".inst 0x4f87e29b // sdot v27.4s, v20.16b, v7.4b[0]\n"
+ ".inst 0x4f89e29c // sdot v28.4s, v20.16b, v9.4b[0]\n"
+ ".inst 0x4f8be29d // sdot v29.4s, v20.16b, v11.4b[0]\n"
+ ".inst 0x4f8de29e // sdot v30.4s, v20.16b, v13.4b[0]\n"
+ ".inst 0x4f8fe29f // sdot v31.4s, v20.16b, v15.4b[0]\n"
+ ".inst 0x4fa1e2b8 // sdot v24.4s, v21.16b, v1.4b[1]\n"
+ ".inst 0x4fa3e2b9 // sdot v25.4s, v21.16b, v3.4b[1]\n"
+ ".inst 0x4fa5e2ba // sdot v26.4s, v21.16b, v5.4b[1]\n"
+ ".inst 0x4fa7e2bb // sdot v27.4s, v21.16b, v7.4b[1]\n"
+ ".inst 0x4fa9e2bc // sdot v28.4s, v21.16b, v9.4b[1]\n"
+ ".inst 0x4fabe2bd // sdot v29.4s, v21.16b, v11.4b[1]\n"
+ ".inst 0x4fade2be // sdot v30.4s, v21.16b, v13.4b[1]\n"
+ ".inst 0x4fafe2bf // sdot v31.4s, v21.16b, v15.4b[1]\n"
"6:\n"
"str q24, [%[c_ptr0]]\n"
+ "add %[c_ptr0], %[c_ptr0], #0x10\n"
"str q25, [c_ptr1]\n"
+ "add c_ptr1, c_ptr1, #0x10\n"
"str q26, [c_ptr2]\n"
+ "add c_ptr2, c_ptr2, #0x10\n"
"str q27, [c_ptr3]\n"
+ "add c_ptr3, c_ptr3, #0x10\n"
"str q28, [c_ptr4]\n"
+ "add c_ptr4, c_ptr4, #0x10\n"
"str q29, [c_ptr5]\n"
+ "add c_ptr5, c_ptr5, #0x10\n"
"str q30, [c_ptr6]\n"
+ "add c_ptr6, c_ptr6, #0x10\n"
"str q31, [c_ptr7]\n"
+ "add c_ptr7, c_ptr7, #0x10\n"
".unreq a_ptr1\n"
".unreq a_ptr2\n"
".unreq a_ptr3\n"
@@ -2113,68 +2258,68 @@ void a64_smallK_hybrid_s8s32_dot_4x8_a55(const int8_t *A, int lda, const int8_t
"ldr q22, [%[b_ptr0], #0x60]\n"
"movi v31.4s, #0\n"
"prfm PLDL1KEEP, [a_ptr7, #0x40]\n"
- ".word 0x4f80e218 // sdot v24.4s, v16.16b, v0.4b[0]\n"
+ ".inst 0x4f80e218 // sdot v24.4s, v16.16b, v0.4b[0]\n"
"prfm PLDL1KEEP, [a_ptr7, #0x80]\n"
- ".word 0x4f82e219 // sdot v25.4s, v16.16b, v2.4b[0]\n"
+ ".inst 0x4f82e219 // sdot v25.4s, v16.16b, v2.4b[0]\n"
"prfm PLDL1KEEP, [a_ptr7, #0xc0]\n"
- ".word 0x4f84e21a // sdot v26.4s, v16.16b, v4.4b[0]\n"
+ ".inst 0x4f84e21a // sdot v26.4s, v16.16b, v4.4b[0]\n"
"prfm PLDL1KEEP, [a_ptr7, #0x100]\n"
- ".word 0x4f86e21b // sdot v27.4s, v16.16b, v6.4b[0]\n"
+ ".inst 0x4f86e21b // sdot v27.4s, v16.16b, v6.4b[0]\n"
"prfm PLDL1KEEP, [a_ptr7, #0x140]\n"
- ".word 0x4f88e21c // sdot v28.4s, v16.16b, v8.4b[0]\n"
+ ".inst 0x4f88e21c // sdot v28.4s, v16.16b, v8.4b[0]\n"
"prfm PLDL1KEEP, [a_ptr7, #0x180]\n"
- ".word 0x4f8ae21d // sdot v29.4s, v16.16b, v10.4b[0]\n"
+ ".inst 0x4f8ae21d // sdot v29.4s, v16.16b, v10.4b[0]\n"
"add %[b_ptr0], %[b_ptr0], #0x70\n"
- ".word 0x4f8ce21e // sdot v30.4s, v16.16b, v12.4b[0]\n"
- ".word 0x4f8ee21f // sdot v31.4s, v16.16b, v14.4b[0]\n"
- ".word 0x4fa0e238 // sdot v24.4s, v17.16b, v0.4b[1]\n"
- ".word 0x4fa2e239 // sdot v25.4s, v17.16b, v2.4b[1]\n"
- ".word 0x4fa4e23a // sdot v26.4s, v17.16b, v4.4b[1]\n"
- ".word 0x4fa6e23b // sdot v27.4s, v17.16b, v6.4b[1]\n"
- ".word 0x4fa8e23c // sdot v28.4s, v17.16b, v8.4b[1]\n"
- ".word 0x4faae23d // sdot v29.4s, v17.16b, v10.4b[1]\n"
- ".word 0x4face23e // sdot v30.4s, v17.16b, v12.4b[1]\n"
- ".word 0x4faee23f // sdot v31.4s, v17.16b, v14.4b[1]\n"
- ".word 0x4f80ea58 // sdot v24.4s, v18.16b, v0.4b[2]\n"
- ".word 0x4f82ea59 // sdot v25.4s, v18.16b, v2.4b[2]\n"
- ".word 0x4f84ea5a // sdot v26.4s, v18.16b, v4.4b[2]\n"
- ".word 0x4f86ea5b // sdot v27.4s, v18.16b, v6.4b[2]\n"
- ".word 0x4f88ea5c // sdot v28.4s, v18.16b, v8.4b[2]\n"
- ".word 0x4f8aea5d // sdot v29.4s, v18.16b, v10.4b[2]\n"
- ".word 0x4f8cea5e // sdot v30.4s, v18.16b, v12.4b[2]\n"
- ".word 0x4f8eea5f // sdot v31.4s, v18.16b, v14.4b[2]\n"
- ".word 0x4fa0ea78 // sdot v24.4s, v19.16b, v0.4b[3]\n"
- ".word 0x4fa2ea79 // sdot v25.4s, v19.16b, v2.4b[3]\n"
- ".word 0x4fa4ea7a // sdot v26.4s, v19.16b, v4.4b[3]\n"
- ".word 0x4fa6ea7b // sdot v27.4s, v19.16b, v6.4b[3]\n"
- ".word 0x4fa8ea7c // sdot v28.4s, v19.16b, v8.4b[3]\n"
- ".word 0x4faaea7d // sdot v29.4s, v19.16b, v10.4b[3]\n"
- ".word 0x4facea7e // sdot v30.4s, v19.16b, v12.4b[3]\n"
- ".word 0x4faeea7f // sdot v31.4s, v19.16b, v14.4b[3]\n"
- ".word 0x4f81e298 // sdot v24.4s, v20.16b, v1.4b[0]\n"
- ".word 0x4f83e299 // sdot v25.4s, v20.16b, v3.4b[0]\n"
- ".word 0x4f85e29a // sdot v26.4s, v20.16b, v5.4b[0]\n"
- ".word 0x4f87e29b // sdot v27.4s, v20.16b, v7.4b[0]\n"
- ".word 0x4f89e29c // sdot v28.4s, v20.16b, v9.4b[0]\n"
- ".word 0x4f8be29d // sdot v29.4s, v20.16b, v11.4b[0]\n"
- ".word 0x4f8de29e // sdot v30.4s, v20.16b, v13.4b[0]\n"
- ".word 0x4f8fe29f // sdot v31.4s, v20.16b, v15.4b[0]\n"
- ".word 0x4fa1e2b8 // sdot v24.4s, v21.16b, v1.4b[1]\n"
- ".word 0x4fa3e2b9 // sdot v25.4s, v21.16b, v3.4b[1]\n"
- ".word 0x4fa5e2ba // sdot v26.4s, v21.16b, v5.4b[1]\n"
- ".word 0x4fa7e2bb // sdot v27.4s, v21.16b, v7.4b[1]\n"
- ".word 0x4fa9e2bc // sdot v28.4s, v21.16b, v9.4b[1]\n"
- ".word 0x4fabe2bd // sdot v29.4s, v21.16b, v11.4b[1]\n"
- ".word 0x4fade2be // sdot v30.4s, v21.16b, v13.4b[1]\n"
- ".word 0x4fafe2bf // sdot v31.4s, v21.16b, v15.4b[1]\n"
- ".word 0x4f81ead8 // sdot v24.4s, v22.16b, v1.4b[2]\n"
- ".word 0x4f83ead9 // sdot v25.4s, v22.16b, v3.4b[2]\n"
- ".word 0x4f85eada // sdot v26.4s, v22.16b, v5.4b[2]\n"
- ".word 0x4f87eadb // sdot v27.4s, v22.16b, v7.4b[2]\n"
- ".word 0x4f89eadc // sdot v28.4s, v22.16b, v9.4b[2]\n"
- ".word 0x4f8beadd // sdot v29.4s, v22.16b, v11.4b[2]\n"
- ".word 0x4f8deade // sdot v30.4s, v22.16b, v13.4b[2]\n"
- ".word 0x4f8feadf // sdot v31.4s, v22.16b, v15.4b[2]\n"
+ ".inst 0x4f8ce21e // sdot v30.4s, v16.16b, v12.4b[0]\n"
+ ".inst 0x4f8ee21f // sdot v31.4s, v16.16b, v14.4b[0]\n"
+ ".inst 0x4fa0e238 // sdot v24.4s, v17.16b, v0.4b[1]\n"
+ ".inst 0x4fa2e239 // sdot v25.4s, v17.16b, v2.4b[1]\n"
+ ".inst 0x4fa4e23a // sdot v26.4s, v17.16b, v4.4b[1]\n"
+ ".inst 0x4fa6e23b // sdot v27.4s, v17.16b, v6.4b[1]\n"
+ ".inst 0x4fa8e23c // sdot v28.4s, v17.16b, v8.4b[1]\n"
+ ".inst 0x4faae23d // sdot v29.4s, v17.16b, v10.4b[1]\n"
+ ".inst 0x4face23e // sdot v30.4s, v17.16b, v12.4b[1]\n"
+ ".inst 0x4faee23f // sdot v31.4s, v17.16b, v14.4b[1]\n"
+ ".inst 0x4f80ea58 // sdot v24.4s, v18.16b, v0.4b[2]\n"
+ ".inst 0x4f82ea59 // sdot v25.4s, v18.16b, v2.4b[2]\n"
+ ".inst 0x4f84ea5a // sdot v26.4s, v18.16b, v4.4b[2]\n"
+ ".inst 0x4f86ea5b // sdot v27.4s, v18.16b, v6.4b[2]\n"
+ ".inst 0x4f88ea5c // sdot v28.4s, v18.16b, v8.4b[2]\n"
+ ".inst 0x4f8aea5d // sdot v29.4s, v18.16b, v10.4b[2]\n"
+ ".inst 0x4f8cea5e // sdot v30.4s, v18.16b, v12.4b[2]\n"
+ ".inst 0x4f8eea5f // sdot v31.4s, v18.16b, v14.4b[2]\n"
+ ".inst 0x4fa0ea78 // sdot v24.4s, v19.16b, v0.4b[3]\n"
+ ".inst 0x4fa2ea79 // sdot v25.4s, v19.16b, v2.4b[3]\n"
+ ".inst 0x4fa4ea7a // sdot v26.4s, v19.16b, v4.4b[3]\n"
+ ".inst 0x4fa6ea7b // sdot v27.4s, v19.16b, v6.4b[3]\n"
+ ".inst 0x4fa8ea7c // sdot v28.4s, v19.16b, v8.4b[3]\n"
+ ".inst 0x4faaea7d // sdot v29.4s, v19.16b, v10.4b[3]\n"
+ ".inst 0x4facea7e // sdot v30.4s, v19.16b, v12.4b[3]\n"
+ ".inst 0x4faeea7f // sdot v31.4s, v19.16b, v14.4b[3]\n"
+ ".inst 0x4f81e298 // sdot v24.4s, v20.16b, v1.4b[0]\n"
+ ".inst 0x4f83e299 // sdot v25.4s, v20.16b, v3.4b[0]\n"
+ ".inst 0x4f85e29a // sdot v26.4s, v20.16b, v5.4b[0]\n"
+ ".inst 0x4f87e29b // sdot v27.4s, v20.16b, v7.4b[0]\n"
+ ".inst 0x4f89e29c // sdot v28.4s, v20.16b, v9.4b[0]\n"
+ ".inst 0x4f8be29d // sdot v29.4s, v20.16b, v11.4b[0]\n"
+ ".inst 0x4f8de29e // sdot v30.4s, v20.16b, v13.4b[0]\n"
+ ".inst 0x4f8fe29f // sdot v31.4s, v20.16b, v15.4b[0]\n"
+ ".inst 0x4fa1e2b8 // sdot v24.4s, v21.16b, v1.4b[1]\n"
+ ".inst 0x4fa3e2b9 // sdot v25.4s, v21.16b, v3.4b[1]\n"
+ ".inst 0x4fa5e2ba // sdot v26.4s, v21.16b, v5.4b[1]\n"
+ ".inst 0x4fa7e2bb // sdot v27.4s, v21.16b, v7.4b[1]\n"
+ ".inst 0x4fa9e2bc // sdot v28.4s, v21.16b, v9.4b[1]\n"
+ ".inst 0x4fabe2bd // sdot v29.4s, v21.16b, v11.4b[1]\n"
+ ".inst 0x4fade2be // sdot v30.4s, v21.16b, v13.4b[1]\n"
+ ".inst 0x4fafe2bf // sdot v31.4s, v21.16b, v15.4b[1]\n"
+ ".inst 0x4f81ead8 // sdot v24.4s, v22.16b, v1.4b[2]\n"
+ ".inst 0x4f83ead9 // sdot v25.4s, v22.16b, v3.4b[2]\n"
+ ".inst 0x4f85eada // sdot v26.4s, v22.16b, v5.4b[2]\n"
+ ".inst 0x4f87eadb // sdot v27.4s, v22.16b, v7.4b[2]\n"
+ ".inst 0x4f89eadc // sdot v28.4s, v22.16b, v9.4b[2]\n"
+ ".inst 0x4f8beadd // sdot v29.4s, v22.16b, v11.4b[2]\n"
+ ".inst 0x4f8deade // sdot v30.4s, v22.16b, v13.4b[2]\n"
+ ".inst 0x4f8feadf // sdot v31.4s, v22.16b, v15.4b[2]\n"
"cbz %[loops], 6f\n"
"ldr d16, [%[b_ptr0]]\n"
"subs %[loops], %[loops], #0x1\n"
@@ -2200,193 +2345,217 @@ void a64_smallK_hybrid_s8s32_dot_4x8_a55(const int8_t *A, int lda, const int8_t
"ins v21.d[1], temploadreg1\n"
"b.eq 7f\n"
"8:\n"
- "str q24, [%[c_ptr0]], #0x10\n"
+ "str q24, [%[c_ptr0]]\n"
"subs %[loops], %[loops], #0x1\n"
"movi v24.4s, #0\n"
"ins v22.d[1], temploadreg2\n"
"ldr temploadreg0, [%[b_ptr0], #0x8]\n"
- "prfm PSTL1KEEP, [%[c_ptr0], #0x40]\n"
- "str q25, [c_ptr1], #0x10\n"
+ "add %[c_ptr0], %[c_ptr0], #0x10\n"
+ "str q25, [c_ptr1]\n"
+ "add c_ptr1, c_ptr1, #0x10\n"
"movi v25.4s, #0\n"
"ldr temploadreg1, [%[b_ptr0], #0x18]\n"
- ".word 0x4f80e218 // sdot v24.4s, v16.16b, v0.4b[0]\n"
+ ".inst 0x4f80e218 // sdot v24.4s, v16.16b, v0.4b[0]\n"
+ "str q26, [c_ptr2]\n"
+ "movi v26.4s, #0\n"
"ldr temploadreg2, [%[b_ptr0], #0x28]\n"
"ldr temploadreg3, [%[b_ptr0], #0x38]\n"
- "prfm PSTL1KEEP, [c_ptr1, #0x40]\n"
- ".word 0x4f82e219 // sdot v25.4s, v16.16b, v2.4b[0]\n"
- "str q26, [c_ptr2], #0x10\n"
- "movi v26.4s, #0\n"
- ".word 0x4fa0e238 // sdot v24.4s, v17.16b, v0.4b[1]\n"
- ".word 0x4fa2e239 // sdot v25.4s, v17.16b, v2.4b[1]\n"
- "str q27, [c_ptr3], #0x10\n"
+ "add c_ptr2, c_ptr2, #0x10\n"
+ ".inst 0x4f82e219 // sdot v25.4s, v16.16b, v2.4b[0]\n"
+ "str q27, [c_ptr3]\n"
"movi v27.4s, #0\n"
- "prfm PSTL1KEEP, [c_ptr2, #0x40]\n"
- ".word 0x4f84e21a // sdot v26.4s, v16.16b, v4.4b[0]\n"
- "str q28, [c_ptr4], #0x10\n"
+ "add c_ptr3, c_ptr3, #0x10\n"
+ ".inst 0x4f84e21a // sdot v26.4s, v16.16b, v4.4b[0]\n"
+ "str q28, [c_ptr4]\n"
"movi v28.4s, #0\n"
- "prfm PSTL1KEEP, [c_ptr3, #0x40]\n"
- ".word 0x4f86e21b // sdot v27.4s, v16.16b, v6.4b[0]\n"
- "str q29, [c_ptr5], #0x10\n"
+ "add c_ptr4, c_ptr4, #0x10\n"
+ ".inst 0x4f86e21b // sdot v27.4s, v16.16b, v6.4b[0]\n"
+ "str q29, [c_ptr5]\n"
"movi v29.4s, #0\n"
- "prfm PSTL1KEEP, [c_ptr4, #0x40]\n"
- ".word 0x4f88e21c // sdot v28.4s, v16.16b, v8.4b[0]\n"
- "str q30, [c_ptr6], #0x10\n"
+ "add c_ptr5, c_ptr5, #0x10\n"
+ ".inst 0x4f88e21c // sdot v28.4s, v16.16b, v8.4b[0]\n"
+ "str q30, [c_ptr6]\n"
"movi v30.4s, #0\n"
- "prfm PSTL1KEEP, [c_ptr5, #0x40]\n"
- ".word 0x4f8ae21d // sdot v29.4s, v16.16b, v10.4b[0]\n"
- "str q31, [c_ptr7], #0x10\n"
+ "add c_ptr6, c_ptr6, #0x10\n"
+ ".inst 0x4f8ae21d // sdot v29.4s, v16.16b, v10.4b[0]\n"
+ "str q31, [c_ptr7]\n"
"movi v31.4s, #0\n"
- "prfm PSTL1KEEP, [c_ptr6, #0x40]\n"
- ".word 0x4f8ce21e // sdot v30.4s, v16.16b, v12.4b[0]\n"
- "prfm PSTL1KEEP, [c_ptr7, #0x40]\n"
- ".word 0x4f8ee21f // sdot v31.4s, v16.16b, v14.4b[0]\n"
+ "add c_ptr7, c_ptr7, #0x10\n"
+ ".inst 0x4f8ce21e // sdot v30.4s, v16.16b, v12.4b[0]\n"
+ "prfm PSTL1KEEP, [%[c_ptr0], #0x40]\n"
+ ".inst 0x4f8ee21f // sdot v31.4s, v16.16b, v14.4b[0]\n"
"ldr d16, [%[b_ptr0]]\n"
- ".word 0x4fa4e23a // sdot v26.4s, v17.16b, v4.4b[1]\n"
- ".word 0x4fa6e23b // sdot v27.4s, v17.16b, v6.4b[1]\n"
- ".word 0x4fa8e23c // sdot v28.4s, v17.16b, v8.4b[1]\n"
+ ".inst 0x4fa0e238 // sdot v24.4s, v17.16b, v0.4b[1]\n"
+ "prfm PSTL1KEEP, [c_ptr1, #0x40]\n"
+ ".inst 0x4fa2e239 // sdot v25.4s, v17.16b, v2.4b[1]\n"
"ins v16.d[1], temploadreg0\n"
- ".word 0x4faae23d // sdot v29.4s, v17.16b, v10.4b[1]\n"
+ ".inst 0x4fa4e23a // sdot v26.4s, v17.16b, v4.4b[1]\n"
"ldr temploadreg0, [%[b_ptr0], #0x48]\n"
- ".word 0x4face23e // sdot v30.4s, v17.16b, v12.4b[1]\n"
- ".word 0x4faee23f // sdot v31.4s, v17.16b, v14.4b[1]\n"
+ ".inst 0x4fa6e23b // sdot v27.4s, v17.16b, v6.4b[1]\n"
+ "prfm PSTL1KEEP, [c_ptr2, #0x40]\n"
+ ".inst 0x4fa8e23c // sdot v28.4s, v17.16b, v8.4b[1]\n"
+ "prfm PSTL1KEEP, [c_ptr3, #0x40]\n"
+ ".inst 0x4faae23d // sdot v29.4s, v17.16b, v10.4b[1]\n"
+ "prfm PSTL1KEEP, [c_ptr4, #0x40]\n"
+ ".inst 0x4face23e // sdot v30.4s, v17.16b, v12.4b[1]\n"
+ "prfm PSTL1KEEP, [c_ptr5, #0x40]\n"
+ ".inst 0x4faee23f // sdot v31.4s, v17.16b, v14.4b[1]\n"
"ldr d17, [%[b_ptr0], #0x10]\n"
- ".word 0x4f80ea58 // sdot v24.4s, v18.16b, v0.4b[2]\n"
- ".word 0x4f82ea59 // sdot v25.4s, v18.16b, v2.4b[2]\n"
- ".word 0x4f84ea5a // sdot v26.4s, v18.16b, v4.4b[2]\n"
+ ".inst 0x4f80ea58 // sdot v24.4s, v18.16b, v0.4b[2]\n"
+ "prfm PSTL1KEEP, [c_ptr6, #0x40]\n"
+ ".inst 0x4f82ea59 // sdot v25.4s, v18.16b, v2.4b[2]\n"
"ins v17.d[1], temploadreg1\n"
- ".word 0x4f86ea5b // sdot v27.4s, v18.16b, v6.4b[2]\n"
+ ".inst 0x4f84ea5a // sdot v26.4s, v18.16b, v4.4b[2]\n"
"ldr temploadreg1, [%[b_ptr0], #0x58]\n"
- ".word 0x4f88ea5c // sdot v28.4s, v18.16b, v8.4b[2]\n"
- ".word 0x4f8aea5d // sdot v29.4s, v18.16b, v10.4b[2]\n"
- ".word 0x4f8cea5e // sdot v30.4s, v18.16b, v12.4b[2]\n"
- ".word 0x4f8eea5f // sdot v31.4s, v18.16b, v14.4b[2]\n"
+ ".inst 0x4f86ea5b // sdot v27.4s, v18.16b, v6.4b[2]\n"
+ "prfm PSTL1KEEP, [c_ptr7, #0x40]\n"
+ ".inst 0x4f88ea5c // sdot v28.4s, v18.16b, v8.4b[2]\n"
+ ".inst 0x4f8aea5d // sdot v29.4s, v18.16b, v10.4b[2]\n"
+ ".inst 0x4f8cea5e // sdot v30.4s, v18.16b, v12.4b[2]\n"
+ ".inst 0x4f8eea5f // sdot v31.4s, v18.16b, v14.4b[2]\n"
"ldr d18, [%[b_ptr0], #0x20]\n"
- ".word 0x4fa0ea78 // sdot v24.4s, v19.16b, v0.4b[3]\n"
- ".word 0x4fa2ea79 // sdot v25.4s, v19.16b, v2.4b[3]\n"
- ".word 0x4fa4ea7a // sdot v26.4s, v19.16b, v4.4b[3]\n"
+ ".inst 0x4fa0ea78 // sdot v24.4s, v19.16b, v0.4b[3]\n"
+ ".inst 0x4fa2ea79 // sdot v25.4s, v19.16b, v2.4b[3]\n"
+ ".inst 0x4fa4ea7a // sdot v26.4s, v19.16b, v4.4b[3]\n"
"ins v18.d[1], temploadreg2\n"
- ".word 0x4fa6ea7b // sdot v27.4s, v19.16b, v6.4b[3]\n"
+ ".inst 0x4fa6ea7b // sdot v27.4s, v19.16b, v6.4b[3]\n"
"ldr temploadreg2, [%[b_ptr0], #0x68]\n"
- ".word 0x4fa8ea7c // sdot v28.4s, v19.16b, v8.4b[3]\n"
- ".word 0x4faaea7d // sdot v29.4s, v19.16b, v10.4b[3]\n"
- ".word 0x4facea7e // sdot v30.4s, v19.16b, v12.4b[3]\n"
- ".word 0x4faeea7f // sdot v31.4s, v19.16b, v14.4b[3]\n"
+ ".inst 0x4fa8ea7c // sdot v28.4s, v19.16b, v8.4b[3]\n"
+ ".inst 0x4faaea7d // sdot v29.4s, v19.16b, v10.4b[3]\n"
+ ".inst 0x4facea7e // sdot v30.4s, v19.16b, v12.4b[3]\n"
+ ".inst 0x4faeea7f // sdot v31.4s, v19.16b, v14.4b[3]\n"
"ldr d19, [%[b_ptr0], #0x30]\n"
- ".word 0x4f81e298 // sdot v24.4s, v20.16b, v1.4b[0]\n"
- ".word 0x4f83e299 // sdot v25.4s, v20.16b, v3.4b[0]\n"
- ".word 0x4f85e29a // sdot v26.4s, v20.16b, v5.4b[0]\n"
+ ".inst 0x4f81e298 // sdot v24.4s, v20.16b, v1.4b[0]\n"
+ ".inst 0x4f83e299 // sdot v25.4s, v20.16b, v3.4b[0]\n"
+ ".inst 0x4f85e29a // sdot v26.4s, v20.16b, v5.4b[0]\n"
"ins v19.d[1], temploadreg3\n"
- ".word 0x4f87e29b // sdot v27.4s, v20.16b, v7.4b[0]\n"
- ".word 0x4f89e29c // sdot v28.4s, v20.16b, v9.4b[0]\n"
- ".word 0x4f8be29d // sdot v29.4s, v20.16b, v11.4b[0]\n"
- ".word 0x4f8de29e // sdot v30.4s, v20.16b, v13.4b[0]\n"
- ".word 0x4f8fe29f // sdot v31.4s, v20.16b, v15.4b[0]\n"
+ ".inst 0x4f87e29b // sdot v27.4s, v20.16b, v7.4b[0]\n"
+ ".inst 0x4f89e29c // sdot v28.4s, v20.16b, v9.4b[0]\n"
+ ".inst 0x4f8be29d // sdot v29.4s, v20.16b, v11.4b[0]\n"
+ ".inst 0x4f8de29e // sdot v30.4s, v20.16b, v13.4b[0]\n"
+ ".inst 0x4f8fe29f // sdot v31.4s, v20.16b, v15.4b[0]\n"
"ldr d20, [%[b_ptr0], #0x40]\n"
- ".word 0x4fa1e2b8 // sdot v24.4s, v21.16b, v1.4b[1]\n"
- ".word 0x4fa3e2b9 // sdot v25.4s, v21.16b, v3.4b[1]\n"
- ".word 0x4fa5e2ba // sdot v26.4s, v21.16b, v5.4b[1]\n"
+ ".inst 0x4fa1e2b8 // sdot v24.4s, v21.16b, v1.4b[1]\n"
+ ".inst 0x4fa3e2b9 // sdot v25.4s, v21.16b, v3.4b[1]\n"
+ ".inst 0x4fa5e2ba // sdot v26.4s, v21.16b, v5.4b[1]\n"
"ins v20.d[1], temploadreg0\n"
- ".word 0x4fa7e2bb // sdot v27.4s, v21.16b, v7.4b[1]\n"
- ".word 0x4fa9e2bc // sdot v28.4s, v21.16b, v9.4b[1]\n"
- ".word 0x4fabe2bd // sdot v29.4s, v21.16b, v11.4b[1]\n"
- ".word 0x4fade2be // sdot v30.4s, v21.16b, v13.4b[1]\n"
- ".word 0x4fafe2bf // sdot v31.4s, v21.16b, v15.4b[1]\n"
+ ".inst 0x4fa7e2bb // sdot v27.4s, v21.16b, v7.4b[1]\n"
+ ".inst 0x4fa9e2bc // sdot v28.4s, v21.16b, v9.4b[1]\n"
+ ".inst 0x4fabe2bd // sdot v29.4s, v21.16b, v11.4b[1]\n"
+ ".inst 0x4fade2be // sdot v30.4s, v21.16b, v13.4b[1]\n"
+ ".inst 0x4fafe2bf // sdot v31.4s, v21.16b, v15.4b[1]\n"
"ldr d21, [%[b_ptr0], #0x50]\n"
- ".word 0x4f81ead8 // sdot v24.4s, v22.16b, v1.4b[2]\n"
- ".word 0x4f83ead9 // sdot v25.4s, v22.16b, v3.4b[2]\n"
- ".word 0x4f85eada // sdot v26.4s, v22.16b, v5.4b[2]\n"
+ ".inst 0x4f81ead8 // sdot v24.4s, v22.16b, v1.4b[2]\n"
+ ".inst 0x4f83ead9 // sdot v25.4s, v22.16b, v3.4b[2]\n"
+ ".inst 0x4f85eada // sdot v26.4s, v22.16b, v5.4b[2]\n"
"ins v21.d[1], temploadreg1\n"
- ".word 0x4f87eadb // sdot v27.4s, v22.16b, v7.4b[2]\n"
- ".word 0x4f89eadc // sdot v28.4s, v22.16b, v9.4b[2]\n"
- ".word 0x4f8beadd // sdot v29.4s, v22.16b, v11.4b[2]\n"
- ".word 0x4f8deade // sdot v30.4s, v22.16b, v13.4b[2]\n"
- ".word 0x4f8feadf // sdot v31.4s, v22.16b, v15.4b[2]\n"
+ ".inst 0x4f87eadb // sdot v27.4s, v22.16b, v7.4b[2]\n"
+ ".inst 0x4f89eadc // sdot v28.4s, v22.16b, v9.4b[2]\n"
+ ".inst 0x4f8beadd // sdot v29.4s, v22.16b, v11.4b[2]\n"
+ ".inst 0x4f8deade // sdot v30.4s, v22.16b, v13.4b[2]\n"
+ ".inst 0x4f8feadf // sdot v31.4s, v22.16b, v15.4b[2]\n"
"ldr d22, [%[b_ptr0], #0x60]\n"
"add %[b_ptr0], %[b_ptr0], #0x70\n"
"b.ne 8b\n"
"7:\n"
- "str q24, [%[c_ptr0]], #0x10\n"
+ "str q24, [%[c_ptr0]]\n"
+ "add %[c_ptr0], %[c_ptr0], #0x10\n"
"movi v24.4s, #0\n"
"ins v22.d[1], temploadreg2\n"
- "str q25, [c_ptr1], #0x10\n"
+ "str q25, [c_ptr1]\n"
+ "add c_ptr1, c_ptr1, #0x10\n"
"movi v25.4s, #0\n"
- ".word 0x4f80e218 // sdot v24.4s, v16.16b, v0.4b[0]\n"
- "str q26, [c_ptr2], #0x10\n"
+ ".inst 0x4f80e218 // sdot v24.4s, v16.16b, v0.4b[0]\n"
+ "str q26, [c_ptr2]\n"
"movi v26.4s, #0\n"
- ".word 0x4f82e219 // sdot v25.4s, v16.16b, v2.4b[0]\n"
- ".word 0x4fa0e238 // sdot v24.4s, v17.16b, v0.4b[1]\n"
- "str q27, [c_ptr3], #0x10\n"
+ "add c_ptr2, c_ptr2, #0x10\n"
+ ".inst 0x4f82e219 // sdot v25.4s, v16.16b, v2.4b[0]\n"
+ "str q27, [c_ptr3]\n"
"movi v27.4s, #0\n"
- ".word 0x4f84e21a // sdot v26.4s, v16.16b, v4.4b[0]\n"
- ".word 0x4fa2e239 // sdot v25.4s, v17.16b, v2.4b[1]\n"
- "str q28, [c_ptr4], #0x10\n"
+ "add c_ptr3, c_ptr3, #0x10\n"
+ ".inst 0x4f84e21a // sdot v26.4s, v16.16b, v4.4b[0]\n"
+ "str q28, [c_ptr4]\n"
"movi v28.4s, #0\n"
- ".word 0x4f86e21b // sdot v27.4s, v16.16b, v6.4b[0]\n"
- ".word 0x4fa4e23a // sdot v26.4s, v17.16b, v4.4b[1]\n"
- "str q29, [c_ptr5], #0x10\n"
+ "add c_ptr4, c_ptr4, #0x10\n"
+ ".inst 0x4f86e21b // sdot v27.4s, v16.16b, v6.4b[0]\n"
+ "str q29, [c_ptr5]\n"
"movi v29.4s, #0\n"
- ".word 0x4f88e21c // sdot v28.4s, v16.16b, v8.4b[0]\n"
- ".word 0x4fa6e23b // sdot v27.4s, v17.16b, v6.4b[1]\n"
- "str q30, [c_ptr6], #0x10\n"
+ "add c_ptr5, c_ptr5, #0x10\n"
+ ".inst 0x4f88e21c // sdot v28.4s, v16.16b, v8.4b[0]\n"
+ "str q30, [c_ptr6]\n"
"movi v30.4s, #0\n"
- ".word 0x4f8ae21d // sdot v29.4s, v16.16b, v10.4b[0]\n"
- ".word 0x4fa8e23c // sdot v28.4s, v17.16b, v8.4b[1]\n"
- "str q31, [c_ptr7], #0x10\n"
+ "add c_ptr6, c_ptr6, #0x10\n"
+ ".inst 0x4f8ae21d // sdot v29.4s, v16.16b, v10.4b[0]\n"
+ "str q31, [c_ptr7]\n"
"movi v31.4s, #0\n"
- ".word 0x4f8ce21e // sdot v30.4s, v16.16b, v12.4b[0]\n"
- ".word 0x4faae23d // sdot v29.4s, v17.16b, v10.4b[1]\n"
- ".word 0x4f80ea58 // sdot v24.4s, v18.16b, v0.4b[2]\n"
- ".word 0x4f8ee21f // sdot v31.4s, v16.16b, v14.4b[0]\n"
- ".word 0x4face23e // sdot v30.4s, v17.16b, v12.4b[1]\n"
- ".word 0x4f82ea59 // sdot v25.4s, v18.16b, v2.4b[2]\n"
- ".word 0x4f84ea5a // sdot v26.4s, v18.16b, v4.4b[2]\n"
- ".word 0x4faee23f // sdot v31.4s, v17.16b, v14.4b[1]\n"
- ".word 0x4f86ea5b // sdot v27.4s, v18.16b, v6.4b[2]\n"
- ".word 0x4f88ea5c // sdot v28.4s, v18.16b, v8.4b[2]\n"
- ".word 0x4f8aea5d // sdot v29.4s, v18.16b, v10.4b[2]\n"
- ".word 0x4f8cea5e // sdot v30.4s, v18.16b, v12.4b[2]\n"
- ".word 0x4f8eea5f // sdot v31.4s, v18.16b, v14.4b[2]\n"
- ".word 0x4fa0ea78 // sdot v24.4s, v19.16b, v0.4b[3]\n"
- ".word 0x4fa2ea79 // sdot v25.4s, v19.16b, v2.4b[3]\n"
- ".word 0x4fa4ea7a // sdot v26.4s, v19.16b, v4.4b[3]\n"
- ".word 0x4fa6ea7b // sdot v27.4s, v19.16b, v6.4b[3]\n"
- ".word 0x4fa8ea7c // sdot v28.4s, v19.16b, v8.4b[3]\n"
- ".word 0x4faaea7d // sdot v29.4s, v19.16b, v10.4b[3]\n"
- ".word 0x4facea7e // sdot v30.4s, v19.16b, v12.4b[3]\n"
- ".word 0x4faeea7f // sdot v31.4s, v19.16b, v14.4b[3]\n"
- ".word 0x4f81e298 // sdot v24.4s, v20.16b, v1.4b[0]\n"
- ".word 0x4f83e299 // sdot v25.4s, v20.16b, v3.4b[0]\n"
- ".word 0x4f85e29a // sdot v26.4s, v20.16b, v5.4b[0]\n"
- ".word 0x4f87e29b // sdot v27.4s, v20.16b, v7.4b[0]\n"
- ".word 0x4f89e29c // sdot v28.4s, v20.16b, v9.4b[0]\n"
- ".word 0x4f8be29d // sdot v29.4s, v20.16b, v11.4b[0]\n"
- ".word 0x4f8de29e // sdot v30.4s, v20.16b, v13.4b[0]\n"
- ".word 0x4f8fe29f // sdot v31.4s, v20.16b, v15.4b[0]\n"
- ".word 0x4fa1e2b8 // sdot v24.4s, v21.16b, v1.4b[1]\n"
- ".word 0x4fa3e2b9 // sdot v25.4s, v21.16b, v3.4b[1]\n"
- ".word 0x4fa5e2ba // sdot v26.4s, v21.16b, v5.4b[1]\n"
- ".word 0x4fa7e2bb // sdot v27.4s, v21.16b, v7.4b[1]\n"
- ".word 0x4fa9e2bc // sdot v28.4s, v21.16b, v9.4b[1]\n"
- ".word 0x4fabe2bd // sdot v29.4s, v21.16b, v11.4b[1]\n"
- ".word 0x4fade2be // sdot v30.4s, v21.16b, v13.4b[1]\n"
- ".word 0x4fafe2bf // sdot v31.4s, v21.16b, v15.4b[1]\n"
- ".word 0x4f81ead8 // sdot v24.4s, v22.16b, v1.4b[2]\n"
- ".word 0x4f83ead9 // sdot v25.4s, v22.16b, v3.4b[2]\n"
- ".word 0x4f85eada // sdot v26.4s, v22.16b, v5.4b[2]\n"
- ".word 0x4f87eadb // sdot v27.4s, v22.16b, v7.4b[2]\n"
- ".word 0x4f89eadc // sdot v28.4s, v22.16b, v9.4b[2]\n"
- ".word 0x4f8beadd // sdot v29.4s, v22.16b, v11.4b[2]\n"
- ".word 0x4f8deade // sdot v30.4s, v22.16b, v13.4b[2]\n"
- ".word 0x4f8feadf // sdot v31.4s, v22.16b, v15.4b[2]\n"
+ "add c_ptr7, c_ptr7, #0x10\n"
+ ".inst 0x4f8ce21e // sdot v30.4s, v16.16b, v12.4b[0]\n"
+ ".inst 0x4fa0e238 // sdot v24.4s, v17.16b, v0.4b[1]\n"
+ ".inst 0x4f8ee21f // sdot v31.4s, v16.16b, v14.4b[0]\n"
+ ".inst 0x4fa2e239 // sdot v25.4s, v17.16b, v2.4b[1]\n"
+ ".inst 0x4fa4e23a // sdot v26.4s, v17.16b, v4.4b[1]\n"
+ ".inst 0x4fa6e23b // sdot v27.4s, v17.16b, v6.4b[1]\n"
+ ".inst 0x4fa8e23c // sdot v28.4s, v17.16b, v8.4b[1]\n"
+ ".inst 0x4faae23d // sdot v29.4s, v17.16b, v10.4b[1]\n"
+ ".inst 0x4face23e // sdot v30.4s, v17.16b, v12.4b[1]\n"
+ ".inst 0x4faee23f // sdot v31.4s, v17.16b, v14.4b[1]\n"
+ ".inst 0x4f80ea58 // sdot v24.4s, v18.16b, v0.4b[2]\n"
+ ".inst 0x4f82ea59 // sdot v25.4s, v18.16b, v2.4b[2]\n"
+ ".inst 0x4f84ea5a // sdot v26.4s, v18.16b, v4.4b[2]\n"
+ ".inst 0x4f86ea5b // sdot v27.4s, v18.16b, v6.4b[2]\n"
+ ".inst 0x4f88ea5c // sdot v28.4s, v18.16b, v8.4b[2]\n"
+ ".inst 0x4f8aea5d // sdot v29.4s, v18.16b, v10.4b[2]\n"
+ ".inst 0x4f8cea5e // sdot v30.4s, v18.16b, v12.4b[2]\n"
+ ".inst 0x4f8eea5f // sdot v31.4s, v18.16b, v14.4b[2]\n"
+ ".inst 0x4fa0ea78 // sdot v24.4s, v19.16b, v0.4b[3]\n"
+ ".inst 0x4fa2ea79 // sdot v25.4s, v19.16b, v2.4b[3]\n"
+ ".inst 0x4fa4ea7a // sdot v26.4s, v19.16b, v4.4b[3]\n"
+ ".inst 0x4fa6ea7b // sdot v27.4s, v19.16b, v6.4b[3]\n"
+ ".inst 0x4fa8ea7c // sdot v28.4s, v19.16b, v8.4b[3]\n"
+ ".inst 0x4faaea7d // sdot v29.4s, v19.16b, v10.4b[3]\n"
+ ".inst 0x4facea7e // sdot v30.4s, v19.16b, v12.4b[3]\n"
+ ".inst 0x4faeea7f // sdot v31.4s, v19.16b, v14.4b[3]\n"
+ ".inst 0x4f81e298 // sdot v24.4s, v20.16b, v1.4b[0]\n"
+ ".inst 0x4f83e299 // sdot v25.4s, v20.16b, v3.4b[0]\n"
+ ".inst 0x4f85e29a // sdot v26.4s, v20.16b, v5.4b[0]\n"
+ ".inst 0x4f87e29b // sdot v27.4s, v20.16b, v7.4b[0]\n"
+ ".inst 0x4f89e29c // sdot v28.4s, v20.16b, v9.4b[0]\n"
+ ".inst 0x4f8be29d // sdot v29.4s, v20.16b, v11.4b[0]\n"
+ ".inst 0x4f8de29e // sdot v30.4s, v20.16b, v13.4b[0]\n"
+ ".inst 0x4f8fe29f // sdot v31.4s, v20.16b, v15.4b[0]\n"
+ ".inst 0x4fa1e2b8 // sdot v24.4s, v21.16b, v1.4b[1]\n"
+ ".inst 0x4fa3e2b9 // sdot v25.4s, v21.16b, v3.4b[1]\n"
+ ".inst 0x4fa5e2ba // sdot v26.4s, v21.16b, v5.4b[1]\n"
+ ".inst 0x4fa7e2bb // sdot v27.4s, v21.16b, v7.4b[1]\n"
+ ".inst 0x4fa9e2bc // sdot v28.4s, v21.16b, v9.4b[1]\n"
+ ".inst 0x4fabe2bd // sdot v29.4s, v21.16b, v11.4b[1]\n"
+ ".inst 0x4fade2be // sdot v30.4s, v21.16b, v13.4b[1]\n"
+ ".inst 0x4fafe2bf // sdot v31.4s, v21.16b, v15.4b[1]\n"
+ ".inst 0x4f81ead8 // sdot v24.4s, v22.16b, v1.4b[2]\n"
+ ".inst 0x4f83ead9 // sdot v25.4s, v22.16b, v3.4b[2]\n"
+ ".inst 0x4f85eada // sdot v26.4s, v22.16b, v5.4b[2]\n"
+ ".inst 0x4f87eadb // sdot v27.4s, v22.16b, v7.4b[2]\n"
+ ".inst 0x4f89eadc // sdot v28.4s, v22.16b, v9.4b[2]\n"
+ ".inst 0x4f8beadd // sdot v29.4s, v22.16b, v11.4b[2]\n"
+ ".inst 0x4f8deade // sdot v30.4s, v22.16b, v13.4b[2]\n"
+ ".inst 0x4f8feadf // sdot v31.4s, v22.16b, v15.4b[2]\n"
"6:\n"
"str q24, [%[c_ptr0]]\n"
+ "add %[c_ptr0], %[c_ptr0], #0x10\n"
"str q25, [c_ptr1]\n"
+ "add c_ptr1, c_ptr1, #0x10\n"
"str q26, [c_ptr2]\n"
+ "add c_ptr2, c_ptr2, #0x10\n"
"str q27, [c_ptr3]\n"
+ "add c_ptr3, c_ptr3, #0x10\n"
"str q28, [c_ptr4]\n"
+ "add c_ptr4, c_ptr4, #0x10\n"
"str q29, [c_ptr5]\n"
+ "add c_ptr5, c_ptr5, #0x10\n"
"str q30, [c_ptr6]\n"
+ "add c_ptr6, c_ptr6, #0x10\n"
"str q31, [c_ptr7]\n"
+ "add c_ptr7, c_ptr7, #0x10\n"
".unreq a_ptr1\n"
".unreq a_ptr2\n"
".unreq a_ptr3\n"
@@ -2558,77 +2727,77 @@ void a64_smallK_hybrid_s8s32_dot_4x8_a55(const int8_t *A, int lda, const int8_t
"ldr q22, [%[b_ptr0], #0x60]\n"
"movi v31.4s, #0\n"
"ldr q23, [%[b_ptr0], #0x70]\n"
- ".word 0x4f80e218 // sdot v24.4s, v16.16b, v0.4b[0]\n"
+ ".inst 0x4f80e218 // sdot v24.4s, v16.16b, v0.4b[0]\n"
"prfm PLDL1KEEP, [a_ptr7, #0x40]\n"
- ".word 0x4f82e219 // sdot v25.4s, v16.16b, v2.4b[0]\n"
+ ".inst 0x4f82e219 // sdot v25.4s, v16.16b, v2.4b[0]\n"
"prfm PLDL1KEEP, [a_ptr7, #0x80]\n"
- ".word 0x4f84e21a // sdot v26.4s, v16.16b, v4.4b[0]\n"
+ ".inst 0x4f84e21a // sdot v26.4s, v16.16b, v4.4b[0]\n"
"prfm PLDL1KEEP, [a_ptr7, #0xc0]\n"
- ".word 0x4f86e21b // sdot v27.4s, v16.16b, v6.4b[0]\n"
+ ".inst 0x4f86e21b // sdot v27.4s, v16.16b, v6.4b[0]\n"
"prfm PLDL1KEEP, [a_ptr7, #0x100]\n"
- ".word 0x4f88e21c // sdot v28.4s, v16.16b, v8.4b[0]\n"
+ ".inst 0x4f88e21c // sdot v28.4s, v16.16b, v8.4b[0]\n"
"prfm PLDL1KEEP, [a_ptr7, #0x140]\n"
- ".word 0x4f8ae21d // sdot v29.4s, v16.16b, v10.4b[0]\n"
+ ".inst 0x4f8ae21d // sdot v29.4s, v16.16b, v10.4b[0]\n"
"prfm PLDL1KEEP, [a_ptr7, #0x180]\n"
- ".word 0x4f8ce21e // sdot v30.4s, v16.16b, v12.4b[0]\n"
+ ".inst 0x4f8ce21e // sdot v30.4s, v16.16b, v12.4b[0]\n"
"add %[b_ptr0], %[b_ptr0], #0x80\n"
- ".word 0x4f8ee21f // sdot v31.4s, v16.16b, v14.4b[0]\n"
- ".word 0x4fa0e238 // sdot v24.4s, v17.16b, v0.4b[1]\n"
- ".word 0x4fa2e239 // sdot v25.4s, v17.16b, v2.4b[1]\n"
- ".word 0x4fa4e23a // sdot v26.4s, v17.16b, v4.4b[1]\n"
- ".word 0x4fa6e23b // sdot v27.4s, v17.16b, v6.4b[1]\n"
- ".word 0x4fa8e23c // sdot v28.4s, v17.16b, v8.4b[1]\n"
- ".word 0x4faae23d // sdot v29.4s, v17.16b, v10.4b[1]\n"
- ".word 0x4face23e // sdot v30.4s, v17.16b, v12.4b[1]\n"
- ".word 0x4faee23f // sdot v31.4s, v17.16b, v14.4b[1]\n"
- ".word 0x4f80ea58 // sdot v24.4s, v18.16b, v0.4b[2]\n"
- ".word 0x4f82ea59 // sdot v25.4s, v18.16b, v2.4b[2]\n"
- ".word 0x4f84ea5a // sdot v26.4s, v18.16b, v4.4b[2]\n"
- ".word 0x4f86ea5b // sdot v27.4s, v18.16b, v6.4b[2]\n"
- ".word 0x4f88ea5c // sdot v28.4s, v18.16b, v8.4b[2]\n"
- ".word 0x4f8aea5d // sdot v29.4s, v18.16b, v10.4b[2]\n"
- ".word 0x4f8cea5e // sdot v30.4s, v18.16b, v12.4b[2]\n"
- ".word 0x4f8eea5f // sdot v31.4s, v18.16b, v14.4b[2]\n"
- ".word 0x4fa0ea78 // sdot v24.4s, v19.16b, v0.4b[3]\n"
- ".word 0x4fa2ea79 // sdot v25.4s, v19.16b, v2.4b[3]\n"
- ".word 0x4fa4ea7a // sdot v26.4s, v19.16b, v4.4b[3]\n"
- ".word 0x4fa6ea7b // sdot v27.4s, v19.16b, v6.4b[3]\n"
- ".word 0x4fa8ea7c // sdot v28.4s, v19.16b, v8.4b[3]\n"
- ".word 0x4faaea7d // sdot v29.4s, v19.16b, v10.4b[3]\n"
- ".word 0x4facea7e // sdot v30.4s, v19.16b, v12.4b[3]\n"
- ".word 0x4faeea7f // sdot v31.4s, v19.16b, v14.4b[3]\n"
- ".word 0x4f81e298 // sdot v24.4s, v20.16b, v1.4b[0]\n"
- ".word 0x4f83e299 // sdot v25.4s, v20.16b, v3.4b[0]\n"
- ".word 0x4f85e29a // sdot v26.4s, v20.16b, v5.4b[0]\n"
- ".word 0x4f87e29b // sdot v27.4s, v20.16b, v7.4b[0]\n"
- ".word 0x4f89e29c // sdot v28.4s, v20.16b, v9.4b[0]\n"
- ".word 0x4f8be29d // sdot v29.4s, v20.16b, v11.4b[0]\n"
- ".word 0x4f8de29e // sdot v30.4s, v20.16b, v13.4b[0]\n"
- ".word 0x4f8fe29f // sdot v31.4s, v20.16b, v15.4b[0]\n"
- ".word 0x4fa1e2b8 // sdot v24.4s, v21.16b, v1.4b[1]\n"
- ".word 0x4fa3e2b9 // sdot v25.4s, v21.16b, v3.4b[1]\n"
- ".word 0x4fa5e2ba // sdot v26.4s, v21.16b, v5.4b[1]\n"
- ".word 0x4fa7e2bb // sdot v27.4s, v21.16b, v7.4b[1]\n"
- ".word 0x4fa9e2bc // sdot v28.4s, v21.16b, v9.4b[1]\n"
- ".word 0x4fabe2bd // sdot v29.4s, v21.16b, v11.4b[1]\n"
- ".word 0x4fade2be // sdot v30.4s, v21.16b, v13.4b[1]\n"
- ".word 0x4fafe2bf // sdot v31.4s, v21.16b, v15.4b[1]\n"
- ".word 0x4f81ead8 // sdot v24.4s, v22.16b, v1.4b[2]\n"
- ".word 0x4f83ead9 // sdot v25.4s, v22.16b, v3.4b[2]\n"
- ".word 0x4f85eada // sdot v26.4s, v22.16b, v5.4b[2]\n"
- ".word 0x4f87eadb // sdot v27.4s, v22.16b, v7.4b[2]\n"
- ".word 0x4f89eadc // sdot v28.4s, v22.16b, v9.4b[2]\n"
- ".word 0x4f8beadd // sdot v29.4s, v22.16b, v11.4b[2]\n"
- ".word 0x4f8deade // sdot v30.4s, v22.16b, v13.4b[2]\n"
- ".word 0x4f8feadf // sdot v31.4s, v22.16b, v15.4b[2]\n"
- ".word 0x4fa1eaf8 // sdot v24.4s, v23.16b, v1.4b[3]\n"
- ".word 0x4fa3eaf9 // sdot v25.4s, v23.16b, v3.4b[3]\n"
- ".word 0x4fa5eafa // sdot v26.4s, v23.16b, v5.4b[3]\n"
- ".word 0x4fa7eafb // sdot v27.4s, v23.16b, v7.4b[3]\n"
- ".word 0x4fa9eafc // sdot v28.4s, v23.16b, v9.4b[3]\n"
- ".word 0x4fabeafd // sdot v29.4s, v23.16b, v11.4b[3]\n"
- ".word 0x4fadeafe // sdot v30.4s, v23.16b, v13.4b[3]\n"
- ".word 0x4fafeaff // sdot v31.4s, v23.16b, v15.4b[3]\n"
+ ".inst 0x4f8ee21f // sdot v31.4s, v16.16b, v14.4b[0]\n"
+ ".inst 0x4fa0e238 // sdot v24.4s, v17.16b, v0.4b[1]\n"
+ ".inst 0x4fa2e239 // sdot v25.4s, v17.16b, v2.4b[1]\n"
+ ".inst 0x4fa4e23a // sdot v26.4s, v17.16b, v4.4b[1]\n"
+ ".inst 0x4fa6e23b // sdot v27.4s, v17.16b, v6.4b[1]\n"
+ ".inst 0x4fa8e23c // sdot v28.4s, v17.16b, v8.4b[1]\n"
+ ".inst 0x4faae23d // sdot v29.4s, v17.16b, v10.4b[1]\n"
+ ".inst 0x4face23e // sdot v30.4s, v17.16b, v12.4b[1]\n"
+ ".inst 0x4faee23f // sdot v31.4s, v17.16b, v14.4b[1]\n"
+ ".inst 0x4f80ea58 // sdot v24.4s, v18.16b, v0.4b[2]\n"
+ ".inst 0x4f82ea59 // sdot v25.4s, v18.16b, v2.4b[2]\n"
+ ".inst 0x4f84ea5a // sdot v26.4s, v18.16b, v4.4b[2]\n"
+ ".inst 0x4f86ea5b // sdot v27.4s, v18.16b, v6.4b[2]\n"
+ ".inst 0x4f88ea5c // sdot v28.4s, v18.16b, v8.4b[2]\n"
+ ".inst 0x4f8aea5d // sdot v29.4s, v18.16b, v10.4b[2]\n"
+ ".inst 0x4f8cea5e // sdot v30.4s, v18.16b, v12.4b[2]\n"
+ ".inst 0x4f8eea5f // sdot v31.4s, v18.16b, v14.4b[2]\n"
+ ".inst 0x4fa0ea78 // sdot v24.4s, v19.16b, v0.4b[3]\n"
+ ".inst 0x4fa2ea79 // sdot v25.4s, v19.16b, v2.4b[3]\n"
+ ".inst 0x4fa4ea7a // sdot v26.4s, v19.16b, v4.4b[3]\n"
+ ".inst 0x4fa6ea7b // sdot v27.4s, v19.16b, v6.4b[3]\n"
+ ".inst 0x4fa8ea7c // sdot v28.4s, v19.16b, v8.4b[3]\n"
+ ".inst 0x4faaea7d // sdot v29.4s, v19.16b, v10.4b[3]\n"
+ ".inst 0x4facea7e // sdot v30.4s, v19.16b, v12.4b[3]\n"
+ ".inst 0x4faeea7f // sdot v31.4s, v19.16b, v14.4b[3]\n"
+ ".inst 0x4f81e298 // sdot v24.4s, v20.16b, v1.4b[0]\n"
+ ".inst 0x4f83e299 // sdot v25.4s, v20.16b, v3.4b[0]\n"
+ ".inst 0x4f85e29a // sdot v26.4s, v20.16b, v5.4b[0]\n"
+ ".inst 0x4f87e29b // sdot v27.4s, v20.16b, v7.4b[0]\n"
+ ".inst 0x4f89e29c // sdot v28.4s, v20.16b, v9.4b[0]\n"
+ ".inst 0x4f8be29d // sdot v29.4s, v20.16b, v11.4b[0]\n"
+ ".inst 0x4f8de29e // sdot v30.4s, v20.16b, v13.4b[0]\n"
+ ".inst 0x4f8fe29f // sdot v31.4s, v20.16b, v15.4b[0]\n"
+ ".inst 0x4fa1e2b8 // sdot v24.4s, v21.16b, v1.4b[1]\n"
+ ".inst 0x4fa3e2b9 // sdot v25.4s, v21.16b, v3.4b[1]\n"
+ ".inst 0x4fa5e2ba // sdot v26.4s, v21.16b, v5.4b[1]\n"
+ ".inst 0x4fa7e2bb // sdot v27.4s, v21.16b, v7.4b[1]\n"
+ ".inst 0x4fa9e2bc // sdot v28.4s, v21.16b, v9.4b[1]\n"
+ ".inst 0x4fabe2bd // sdot v29.4s, v21.16b, v11.4b[1]\n"
+ ".inst 0x4fade2be // sdot v30.4s, v21.16b, v13.4b[1]\n"
+ ".inst 0x4fafe2bf // sdot v31.4s, v21.16b, v15.4b[1]\n"
+ ".inst 0x4f81ead8 // sdot v24.4s, v22.16b, v1.4b[2]\n"
+ ".inst 0x4f83ead9 // sdot v25.4s, v22.16b, v3.4b[2]\n"
+ ".inst 0x4f85eada // sdot v26.4s, v22.16b, v5.4b[2]\n"
+ ".inst 0x4f87eadb // sdot v27.4s, v22.16b, v7.4b[2]\n"
+ ".inst 0x4f89eadc // sdot v28.4s, v22.16b, v9.4b[2]\n"
+ ".inst 0x4f8beadd // sdot v29.4s, v22.16b, v11.4b[2]\n"
+ ".inst 0x4f8deade // sdot v30.4s, v22.16b, v13.4b[2]\n"
+ ".inst 0x4f8feadf // sdot v31.4s, v22.16b, v15.4b[2]\n"
+ ".inst 0x4fa1eaf8 // sdot v24.4s, v23.16b, v1.4b[3]\n"
+ ".inst 0x4fa3eaf9 // sdot v25.4s, v23.16b, v3.4b[3]\n"
+ ".inst 0x4fa5eafa // sdot v26.4s, v23.16b, v5.4b[3]\n"
+ ".inst 0x4fa7eafb // sdot v27.4s, v23.16b, v7.4b[3]\n"
+ ".inst 0x4fa9eafc // sdot v28.4s, v23.16b, v9.4b[3]\n"
+ ".inst 0x4fabeafd // sdot v29.4s, v23.16b, v11.4b[3]\n"
+ ".inst 0x4fadeafe // sdot v30.4s, v23.16b, v13.4b[3]\n"
+ ".inst 0x4fafeaff // sdot v31.4s, v23.16b, v15.4b[3]\n"
"cbz %[loops], 6f\n"
"ldr d16, [%[b_ptr0]]\n"
"subs %[loops], %[loops], #0x1\n"
@@ -2657,212 +2826,236 @@ void a64_smallK_hybrid_s8s32_dot_4x8_a55(const int8_t *A, int lda, const int8_t
"ins v22.d[1], temploadreg2\n"
"b.eq 7f\n"
"8:\n"
- "str q24, [%[c_ptr0]], #0x10\n"
+ "str q24, [%[c_ptr0]]\n"
"subs %[loops], %[loops], #0x1\n"
"movi v24.4s, #0\n"
"ins v23.d[1], temploadreg3\n"
"ldr temploadreg0, [%[b_ptr0], #0x8]\n"
- "prfm PSTL1KEEP, [%[c_ptr0], #0x40]\n"
- "str q25, [c_ptr1], #0x10\n"
+ "add %[c_ptr0], %[c_ptr0], #0x10\n"
+ "str q25, [c_ptr1]\n"
+ "add c_ptr1, c_ptr1, #0x10\n"
"movi v25.4s, #0\n"
"ldr temploadreg1, [%[b_ptr0], #0x18]\n"
- ".word 0x4f80e218 // sdot v24.4s, v16.16b, v0.4b[0]\n"
+ ".inst 0x4f80e218 // sdot v24.4s, v16.16b, v0.4b[0]\n"
+ "str q26, [c_ptr2]\n"
+ "movi v26.4s, #0\n"
"ldr temploadreg2, [%[b_ptr0], #0x28]\n"
"ldr temploadreg3, [%[b_ptr0], #0x38]\n"
- "prfm PSTL1KEEP, [c_ptr1, #0x40]\n"
- ".word 0x4f82e219 // sdot v25.4s, v16.16b, v2.4b[0]\n"
- "str q26, [c_ptr2], #0x10\n"
- "movi v26.4s, #0\n"
- ".word 0x4fa0e238 // sdot v24.4s, v17.16b, v0.4b[1]\n"
- ".word 0x4fa2e239 // sdot v25.4s, v17.16b, v2.4b[1]\n"
- "str q27, [c_ptr3], #0x10\n"
+ "add c_ptr2, c_ptr2, #0x10\n"
+ ".inst 0x4f82e219 // sdot v25.4s, v16.16b, v2.4b[0]\n"
+ "str q27, [c_ptr3]\n"
"movi v27.4s, #0\n"
- "prfm PSTL1KEEP, [c_ptr2, #0x40]\n"
- ".word 0x4f84e21a // sdot v26.4s, v16.16b, v4.4b[0]\n"
- "str q28, [c_ptr4], #0x10\n"
+ "add c_ptr3, c_ptr3, #0x10\n"
+ ".inst 0x4f84e21a // sdot v26.4s, v16.16b, v4.4b[0]\n"
+ "str q28, [c_ptr4]\n"
"movi v28.4s, #0\n"
- "prfm PSTL1KEEP, [c_ptr3, #0x40]\n"
- ".word 0x4f86e21b // sdot v27.4s, v16.16b, v6.4b[0]\n"
- "str q29, [c_ptr5], #0x10\n"
+ "add c_ptr4, c_ptr4, #0x10\n"
+ ".inst 0x4f86e21b // sdot v27.4s, v16.16b, v6.4b[0]\n"
+ "str q29, [c_ptr5]\n"
"movi v29.4s, #0\n"
- "prfm PSTL1KEEP, [c_ptr4, #0x40]\n"
- ".word 0x4f88e21c // sdot v28.4s, v16.16b, v8.4b[0]\n"
- "str q30, [c_ptr6], #0x10\n"
+ "add c_ptr5, c_ptr5, #0x10\n"
+ ".inst 0x4f88e21c // sdot v28.4s, v16.16b, v8.4b[0]\n"
+ "str q30, [c_ptr6]\n"
"movi v30.4s, #0\n"
- "prfm PSTL1KEEP, [c_ptr5, #0x40]\n"
- ".word 0x4f8ae21d // sdot v29.4s, v16.16b, v10.4b[0]\n"
- "str q31, [c_ptr7], #0x10\n"
+ "add c_ptr6, c_ptr6, #0x10\n"
+ ".inst 0x4f8ae21d // sdot v29.4s, v16.16b, v10.4b[0]\n"
+ "str q31, [c_ptr7]\n"
"movi v31.4s, #0\n"
- "prfm PSTL1KEEP, [c_ptr6, #0x40]\n"
- ".word 0x4f8ce21e // sdot v30.4s, v16.16b, v12.4b[0]\n"
- "prfm PSTL1KEEP, [c_ptr7, #0x40]\n"
- ".word 0x4f8ee21f // sdot v31.4s, v16.16b, v14.4b[0]\n"
+ "add c_ptr7, c_ptr7, #0x10\n"
+ ".inst 0x4f8ce21e // sdot v30.4s, v16.16b, v12.4b[0]\n"
+ "prfm PSTL1KEEP, [%[c_ptr0], #0x40]\n"
+ ".inst 0x4f8ee21f // sdot v31.4s, v16.16b, v14.4b[0]\n"
"ldr d16, [%[b_ptr0]]\n"
- ".word 0x4fa4e23a // sdot v26.4s, v17.16b, v4.4b[1]\n"
- ".word 0x4fa6e23b // sdot v27.4s, v17.16b, v6.4b[1]\n"
- ".word 0x4fa8e23c // sdot v28.4s, v17.16b, v8.4b[1]\n"
+ ".inst 0x4fa0e238 // sdot v24.4s, v17.16b, v0.4b[1]\n"
+ "prfm PSTL1KEEP, [c_ptr1, #0x40]\n"
+ ".inst 0x4fa2e239 // sdot v25.4s, v17.16b, v2.4b[1]\n"
"ins v16.d[1], temploadreg0\n"
- ".word 0x4faae23d // sdot v29.4s, v17.16b, v10.4b[1]\n"
+ ".inst 0x4fa4e23a // sdot v26.4s, v17.16b, v4.4b[1]\n"
"ldr temploadreg0, [%[b_ptr0], #0x48]\n"
- ".word 0x4face23e // sdot v30.4s, v17.16b, v12.4b[1]\n"
- ".word 0x4faee23f // sdot v31.4s, v17.16b, v14.4b[1]\n"
+ ".inst 0x4fa6e23b // sdot v27.4s, v17.16b, v6.4b[1]\n"
+ "prfm PSTL1KEEP, [c_ptr2, #0x40]\n"
+ ".inst 0x4fa8e23c // sdot v28.4s, v17.16b, v8.4b[1]\n"
+ "prfm PSTL1KEEP, [c_ptr3, #0x40]\n"
+ ".inst 0x4faae23d // sdot v29.4s, v17.16b, v10.4b[1]\n"
+ "prfm PSTL1KEEP, [c_ptr4, #0x40]\n"
+ ".inst 0x4face23e // sdot v30.4s, v17.16b, v12.4b[1]\n"
+ "prfm PSTL1KEEP, [c_ptr5, #0x40]\n"
+ ".inst 0x4faee23f // sdot v31.4s, v17.16b, v14.4b[1]\n"
"ldr d17, [%[b_ptr0], #0x10]\n"
- ".word 0x4f80ea58 // sdot v24.4s, v18.16b, v0.4b[2]\n"
- ".word 0x4f82ea59 // sdot v25.4s, v18.16b, v2.4b[2]\n"
- ".word 0x4f84ea5a // sdot v26.4s, v18.16b, v4.4b[2]\n"
+ ".inst 0x4f80ea58 // sdot v24.4s, v18.16b, v0.4b[2]\n"
+ "prfm PSTL1KEEP, [c_ptr6, #0x40]\n"
+ ".inst 0x4f82ea59 // sdot v25.4s, v18.16b, v2.4b[2]\n"
"ins v17.d[1], temploadreg1\n"
- ".word 0x4f86ea5b // sdot v27.4s, v18.16b, v6.4b[2]\n"
+ ".inst 0x4f84ea5a // sdot v26.4s, v18.16b, v4.4b[2]\n"
"ldr temploadreg1, [%[b_ptr0], #0x58]\n"
- ".word 0x4f88ea5c // sdot v28.4s, v18.16b, v8.4b[2]\n"
- ".word 0x4f8aea5d // sdot v29.4s, v18.16b, v10.4b[2]\n"
- ".word 0x4f8cea5e // sdot v30.4s, v18.16b, v12.4b[2]\n"
- ".word 0x4f8eea5f // sdot v31.4s, v18.16b, v14.4b[2]\n"
+ ".inst 0x4f86ea5b // sdot v27.4s, v18.16b, v6.4b[2]\n"
+ "prfm PSTL1KEEP, [c_ptr7, #0x40]\n"
+ ".inst 0x4f88ea5c // sdot v28.4s, v18.16b, v8.4b[2]\n"
+ ".inst 0x4f8aea5d // sdot v29.4s, v18.16b, v10.4b[2]\n"
+ ".inst 0x4f8cea5e // sdot v30.4s, v18.16b, v12.4b[2]\n"
+ ".inst 0x4f8eea5f // sdot v31.4s, v18.16b, v14.4b[2]\n"
"ldr d18, [%[b_ptr0], #0x20]\n"
- ".word 0x4fa0ea78 // sdot v24.4s, v19.16b, v0.4b[3]\n"
- ".word 0x4fa2ea79 // sdot v25.4s, v19.16b, v2.4b[3]\n"
- ".word 0x4fa4ea7a // sdot v26.4s, v19.16b, v4.4b[3]\n"
+ ".inst 0x4fa0ea78 // sdot v24.4s, v19.16b, v0.4b[3]\n"
+ ".inst 0x4fa2ea79 // sdot v25.4s, v19.16b, v2.4b[3]\n"
+ ".inst 0x4fa4ea7a // sdot v26.4s, v19.16b, v4.4b[3]\n"
"ins v18.d[1], temploadreg2\n"
- ".word 0x4fa6ea7b // sdot v27.4s, v19.16b, v6.4b[3]\n"
+ ".inst 0x4fa6ea7b // sdot v27.4s, v19.16b, v6.4b[3]\n"
"ldr temploadreg2, [%[b_ptr0], #0x68]\n"
- ".word 0x4fa8ea7c // sdot v28.4s, v19.16b, v8.4b[3]\n"
- ".word 0x4faaea7d // sdot v29.4s, v19.16b, v10.4b[3]\n"
- ".word 0x4facea7e // sdot v30.4s, v19.16b, v12.4b[3]\n"
- ".word 0x4faeea7f // sdot v31.4s, v19.16b, v14.4b[3]\n"
+ ".inst 0x4fa8ea7c // sdot v28.4s, v19.16b, v8.4b[3]\n"
+ ".inst 0x4faaea7d // sdot v29.4s, v19.16b, v10.4b[3]\n"
+ ".inst 0x4facea7e // sdot v30.4s, v19.16b, v12.4b[3]\n"
+ ".inst 0x4faeea7f // sdot v31.4s, v19.16b, v14.4b[3]\n"
"ldr d19, [%[b_ptr0], #0x30]\n"
- ".word 0x4f81e298 // sdot v24.4s, v20.16b, v1.4b[0]\n"
- ".word 0x4f83e299 // sdot v25.4s, v20.16b, v3.4b[0]\n"
- ".word 0x4f85e29a // sdot v26.4s, v20.16b, v5.4b[0]\n"
+ ".inst 0x4f81e298 // sdot v24.4s, v20.16b, v1.4b[0]\n"
+ ".inst 0x4f83e299 // sdot v25.4s, v20.16b, v3.4b[0]\n"
+ ".inst 0x4f85e29a // sdot v26.4s, v20.16b, v5.4b[0]\n"
"ins v19.d[1], temploadreg3\n"
- ".word 0x4f87e29b // sdot v27.4s, v20.16b, v7.4b[0]\n"
+ ".inst 0x4f87e29b // sdot v27.4s, v20.16b, v7.4b[0]\n"
"ldr temploadreg3, [%[b_ptr0], #0x78]\n"
- ".word 0x4f89e29c // sdot v28.4s, v20.16b, v9.4b[0]\n"
- ".word 0x4f8be29d // sdot v29.4s, v20.16b, v11.4b[0]\n"
- ".word 0x4f8de29e // sdot v30.4s, v20.16b, v13.4b[0]\n"
- ".word 0x4f8fe29f // sdot v31.4s, v20.16b, v15.4b[0]\n"
+ ".inst 0x4f89e29c // sdot v28.4s, v20.16b, v9.4b[0]\n"
+ ".inst 0x4f8be29d // sdot v29.4s, v20.16b, v11.4b[0]\n"
+ ".inst 0x4f8de29e // sdot v30.4s, v20.16b, v13.4b[0]\n"
+ ".inst 0x4f8fe29f // sdot v31.4s, v20.16b, v15.4b[0]\n"
"ldr d20, [%[b_ptr0], #0x40]\n"
- ".word 0x4fa1e2b8 // sdot v24.4s, v21.16b, v1.4b[1]\n"
- ".word 0x4fa3e2b9 // sdot v25.4s, v21.16b, v3.4b[1]\n"
- ".word 0x4fa5e2ba // sdot v26.4s, v21.16b, v5.4b[1]\n"
+ ".inst 0x4fa1e2b8 // sdot v24.4s, v21.16b, v1.4b[1]\n"
+ ".inst 0x4fa3e2b9 // sdot v25.4s, v21.16b, v3.4b[1]\n"
+ ".inst 0x4fa5e2ba // sdot v26.4s, v21.16b, v5.4b[1]\n"
"ins v20.d[1], temploadreg0\n"
- ".word 0x4fa7e2bb // sdot v27.4s, v21.16b, v7.4b[1]\n"
- ".word 0x4fa9e2bc // sdot v28.4s, v21.16b, v9.4b[1]\n"
- ".word 0x4fabe2bd // sdot v29.4s, v21.16b, v11.4b[1]\n"
- ".word 0x4fade2be // sdot v30.4s, v21.16b, v13.4b[1]\n"
- ".word 0x4fafe2bf // sdot v31.4s, v21.16b, v15.4b[1]\n"
+ ".inst 0x4fa7e2bb // sdot v27.4s, v21.16b, v7.4b[1]\n"
+ ".inst 0x4fa9e2bc // sdot v28.4s, v21.16b, v9.4b[1]\n"
+ ".inst 0x4fabe2bd // sdot v29.4s, v21.16b, v11.4b[1]\n"
+ ".inst 0x4fade2be // sdot v30.4s, v21.16b, v13.4b[1]\n"
+ ".inst 0x4fafe2bf // sdot v31.4s, v21.16b, v15.4b[1]\n"
"ldr d21, [%[b_ptr0], #0x50]\n"
- ".word 0x4f81ead8 // sdot v24.4s, v22.16b, v1.4b[2]\n"
- ".word 0x4f83ead9 // sdot v25.4s, v22.16b, v3.4b[2]\n"
- ".word 0x4f85eada // sdot v26.4s, v22.16b, v5.4b[2]\n"
+ ".inst 0x4f81ead8 // sdot v24.4s, v22.16b, v1.4b[2]\n"
+ ".inst 0x4f83ead9 // sdot v25.4s, v22.16b, v3.4b[2]\n"
+ ".inst 0x4f85eada // sdot v26.4s, v22.16b, v5.4b[2]\n"
"ins v21.d[1], temploadreg1\n"
- ".word 0x4f87eadb // sdot v27.4s, v22.16b, v7.4b[2]\n"
- ".word 0x4f89eadc // sdot v28.4s, v22.16b, v9.4b[2]\n"
- ".word 0x4f8beadd // sdot v29.4s, v22.16b, v11.4b[2]\n"
- ".word 0x4f8deade // sdot v30.4s, v22.16b, v13.4b[2]\n"
- ".word 0x4f8feadf // sdot v31.4s, v22.16b, v15.4b[2]\n"
+ ".inst 0x4f87eadb // sdot v27.4s, v22.16b, v7.4b[2]\n"
+ ".inst 0x4f89eadc // sdot v28.4s, v22.16b, v9.4b[2]\n"
+ ".inst 0x4f8beadd // sdot v29.4s, v22.16b, v11.4b[2]\n"
+ ".inst 0x4f8deade // sdot v30.4s, v22.16b, v13.4b[2]\n"
+ ".inst 0x4f8feadf // sdot v31.4s, v22.16b, v15.4b[2]\n"
"ldr d22, [%[b_ptr0], #0x60]\n"
- ".word 0x4fa1eaf8 // sdot v24.4s, v23.16b, v1.4b[3]\n"
- ".word 0x4fa3eaf9 // sdot v25.4s, v23.16b, v3.4b[3]\n"
- ".word 0x4fa5eafa // sdot v26.4s, v23.16b, v5.4b[3]\n"
+ ".inst 0x4fa1eaf8 // sdot v24.4s, v23.16b, v1.4b[3]\n"
+ ".inst 0x4fa3eaf9 // sdot v25.4s, v23.16b, v3.4b[3]\n"
+ ".inst 0x4fa5eafa // sdot v26.4s, v23.16b, v5.4b[3]\n"
"ins v22.d[1], temploadreg2\n"
- ".word 0x4fa7eafb // sdot v27.4s, v23.16b, v7.4b[3]\n"
- ".word 0x4fa9eafc // sdot v28.4s, v23.16b, v9.4b[3]\n"
- ".word 0x4fabeafd // sdot v29.4s, v23.16b, v11.4b[3]\n"
- ".word 0x4fadeafe // sdot v30.4s, v23.16b, v13.4b[3]\n"
- ".word 0x4fafeaff // sdot v31.4s, v23.16b, v15.4b[3]\n"
+ ".inst 0x4fa7eafb // sdot v27.4s, v23.16b, v7.4b[3]\n"
+ ".inst 0x4fa9eafc // sdot v28.4s, v23.16b, v9.4b[3]\n"
+ ".inst 0x4fabeafd // sdot v29.4s, v23.16b, v11.4b[3]\n"
+ ".inst 0x4fadeafe // sdot v30.4s, v23.16b, v13.4b[3]\n"
+ ".inst 0x4fafeaff // sdot v31.4s, v23.16b, v15.4b[3]\n"
"ldr d23, [%[b_ptr0], #0x70]\n"
"add %[b_ptr0], %[b_ptr0], #0x80\n"
"b.ne 8b\n"
"7:\n"
- "str q24, [%[c_ptr0]], #0x10\n"
+ "str q24, [%[c_ptr0]]\n"
+ "add %[c_ptr0], %[c_ptr0], #0x10\n"
"movi v24.4s, #0\n"
"ins v23.d[1], temploadreg3\n"
- "str q25, [c_ptr1], #0x10\n"
+ "str q25, [c_ptr1]\n"
+ "add c_ptr1, c_ptr1, #0x10\n"
"movi v25.4s, #0\n"
- ".word 0x4f80e218 // sdot v24.4s, v16.16b, v0.4b[0]\n"
- "str q26, [c_ptr2], #0x10\n"
+ ".inst 0x4f80e218 // sdot v24.4s, v16.16b, v0.4b[0]\n"
+ "str q26, [c_ptr2]\n"
"movi v26.4s, #0\n"
- ".word 0x4f82e219 // sdot v25.4s, v16.16b, v2.4b[0]\n"
- ".word 0x4fa0e238 // sdot v24.4s, v17.16b, v0.4b[1]\n"
- "str q27, [c_ptr3], #0x10\n"
+ "add c_ptr2, c_ptr2, #0x10\n"
+ ".inst 0x4f82e219 // sdot v25.4s, v16.16b, v2.4b[0]\n"
+ "str q27, [c_ptr3]\n"
"movi v27.4s, #0\n"
- ".word 0x4f84e21a // sdot v26.4s, v16.16b, v4.4b[0]\n"
- ".word 0x4fa2e239 // sdot v25.4s, v17.16b, v2.4b[1]\n"
- "str q28, [c_ptr4], #0x10\n"
+ "add c_ptr3, c_ptr3, #0x10\n"
+ ".inst 0x4f84e21a // sdot v26.4s, v16.16b, v4.4b[0]\n"
+ "str q28, [c_ptr4]\n"
"movi v28.4s, #0\n"
- ".word 0x4f86e21b // sdot v27.4s, v16.16b, v6.4b[0]\n"
- ".word 0x4fa4e23a // sdot v26.4s, v17.16b, v4.4b[1]\n"
- "str q29, [c_ptr5], #0x10\n"
+ "add c_ptr4, c_ptr4, #0x10\n"
+ ".inst 0x4f86e21b // sdot v27.4s, v16.16b, v6.4b[0]\n"
+ "str q29, [c_ptr5]\n"
"movi v29.4s, #0\n"
- ".word 0x4f88e21c // sdot v28.4s, v16.16b, v8.4b[0]\n"
- ".word 0x4fa6e23b // sdot v27.4s, v17.16b, v6.4b[1]\n"
- "str q30, [c_ptr6], #0x10\n"
+ "add c_ptr5, c_ptr5, #0x10\n"
+ ".inst 0x4f88e21c // sdot v28.4s, v16.16b, v8.4b[0]\n"
+ "str q30, [c_ptr6]\n"
"movi v30.4s, #0\n"
- ".word 0x4f8ae21d // sdot v29.4s, v16.16b, v10.4b[0]\n"
- ".word 0x4fa8e23c // sdot v28.4s, v17.16b, v8.4b[1]\n"
- "str q31, [c_ptr7], #0x10\n"
+ "add c_ptr6, c_ptr6, #0x10\n"
+ ".inst 0x4f8ae21d // sdot v29.4s, v16.16b, v10.4b[0]\n"
+ "str q31, [c_ptr7]\n"
"movi v31.4s, #0\n"
- ".word 0x4f8ce21e // sdot v30.4s, v16.16b, v12.4b[0]\n"
- ".word 0x4faae23d // sdot v29.4s, v17.16b, v10.4b[1]\n"
- ".word 0x4f80ea58 // sdot v24.4s, v18.16b, v0.4b[2]\n"
- ".word 0x4f8ee21f // sdot v31.4s, v16.16b, v14.4b[0]\n"
- ".word 0x4face23e // sdot v30.4s, v17.16b, v12.4b[1]\n"
- ".word 0x4f82ea59 // sdot v25.4s, v18.16b, v2.4b[2]\n"
- ".word 0x4f84ea5a // sdot v26.4s, v18.16b, v4.4b[2]\n"
- ".word 0x4faee23f // sdot v31.4s, v17.16b, v14.4b[1]\n"
- ".word 0x4f86ea5b // sdot v27.4s, v18.16b, v6.4b[2]\n"
- ".word 0x4f88ea5c // sdot v28.4s, v18.16b, v8.4b[2]\n"
- ".word 0x4f8aea5d // sdot v29.4s, v18.16b, v10.4b[2]\n"
- ".word 0x4f8cea5e // sdot v30.4s, v18.16b, v12.4b[2]\n"
- ".word 0x4f8eea5f // sdot v31.4s, v18.16b, v14.4b[2]\n"
- ".word 0x4fa0ea78 // sdot v24.4s, v19.16b, v0.4b[3]\n"
- ".word 0x4fa2ea79 // sdot v25.4s, v19.16b, v2.4b[3]\n"
- ".word 0x4fa4ea7a // sdot v26.4s, v19.16b, v4.4b[3]\n"
- ".word 0x4fa6ea7b // sdot v27.4s, v19.16b, v6.4b[3]\n"
- ".word 0x4fa8ea7c // sdot v28.4s, v19.16b, v8.4b[3]\n"
- ".word 0x4faaea7d // sdot v29.4s, v19.16b, v10.4b[3]\n"
- ".word 0x4facea7e // sdot v30.4s, v19.16b, v12.4b[3]\n"
- ".word 0x4faeea7f // sdot v31.4s, v19.16b, v14.4b[3]\n"
- ".word 0x4f81e298 // sdot v24.4s, v20.16b, v1.4b[0]\n"
- ".word 0x4f83e299 // sdot v25.4s, v20.16b, v3.4b[0]\n"
- ".word 0x4f85e29a // sdot v26.4s, v20.16b, v5.4b[0]\n"
- ".word 0x4f87e29b // sdot v27.4s, v20.16b, v7.4b[0]\n"
- ".word 0x4f89e29c // sdot v28.4s, v20.16b, v9.4b[0]\n"
- ".word 0x4f8be29d // sdot v29.4s, v20.16b, v11.4b[0]\n"
- ".word 0x4f8de29e // sdot v30.4s, v20.16b, v13.4b[0]\n"
- ".word 0x4f8fe29f // sdot v31.4s, v20.16b, v15.4b[0]\n"
- ".word 0x4fa1e2b8 // sdot v24.4s, v21.16b, v1.4b[1]\n"
- ".word 0x4fa3e2b9 // sdot v25.4s, v21.16b, v3.4b[1]\n"
- ".word 0x4fa5e2ba // sdot v26.4s, v21.16b, v5.4b[1]\n"
- ".word 0x4fa7e2bb // sdot v27.4s, v21.16b, v7.4b[1]\n"
- ".word 0x4fa9e2bc // sdot v28.4s, v21.16b, v9.4b[1]\n"
- ".word 0x4fabe2bd // sdot v29.4s, v21.16b, v11.4b[1]\n"
- ".word 0x4fade2be // sdot v30.4s, v21.16b, v13.4b[1]\n"
- ".word 0x4fafe2bf // sdot v31.4s, v21.16b, v15.4b[1]\n"
- ".word 0x4f81ead8 // sdot v24.4s, v22.16b, v1.4b[2]\n"
- ".word 0x4f83ead9 // sdot v25.4s, v22.16b, v3.4b[2]\n"
- ".word 0x4f85eada // sdot v26.4s, v22.16b, v5.4b[2]\n"
- ".word 0x4f87eadb // sdot v27.4s, v22.16b, v7.4b[2]\n"
- ".word 0x4f89eadc // sdot v28.4s, v22.16b, v9.4b[2]\n"
- ".word 0x4f8beadd // sdot v29.4s, v22.16b, v11.4b[2]\n"
- ".word 0x4f8deade // sdot v30.4s, v22.16b, v13.4b[2]\n"
- ".word 0x4f8feadf // sdot v31.4s, v22.16b, v15.4b[2]\n"
- ".word 0x4fa1eaf8 // sdot v24.4s, v23.16b, v1.4b[3]\n"
- ".word 0x4fa3eaf9 // sdot v25.4s, v23.16b, v3.4b[3]\n"
- ".word 0x4fa5eafa // sdot v26.4s, v23.16b, v5.4b[3]\n"
- ".word 0x4fa7eafb // sdot v27.4s, v23.16b, v7.4b[3]\n"
- ".word 0x4fa9eafc // sdot v28.4s, v23.16b, v9.4b[3]\n"
- ".word 0x4fabeafd // sdot v29.4s, v23.16b, v11.4b[3]\n"
- ".word 0x4fadeafe // sdot v30.4s, v23.16b, v13.4b[3]\n"
- ".word 0x4fafeaff // sdot v31.4s, v23.16b, v15.4b[3]\n"
+ "add c_ptr7, c_ptr7, #0x10\n"
+ ".inst 0x4f8ce21e // sdot v30.4s, v16.16b, v12.4b[0]\n"
+ ".inst 0x4fa0e238 // sdot v24.4s, v17.16b, v0.4b[1]\n"
+ ".inst 0x4f8ee21f // sdot v31.4s, v16.16b, v14.4b[0]\n"
+ ".inst 0x4fa2e239 // sdot v25.4s, v17.16b, v2.4b[1]\n"
+ ".inst 0x4fa4e23a // sdot v26.4s, v17.16b, v4.4b[1]\n"
+ ".inst 0x4fa6e23b // sdot v27.4s, v17.16b, v6.4b[1]\n"
+ ".inst 0x4fa8e23c // sdot v28.4s, v17.16b, v8.4b[1]\n"
+ ".inst 0x4faae23d // sdot v29.4s, v17.16b, v10.4b[1]\n"
+ ".inst 0x4face23e // sdot v30.4s, v17.16b, v12.4b[1]\n"
+ ".inst 0x4faee23f // sdot v31.4s, v17.16b, v14.4b[1]\n"
+ ".inst 0x4f80ea58 // sdot v24.4s, v18.16b, v0.4b[2]\n"
+ ".inst 0x4f82ea59 // sdot v25.4s, v18.16b, v2.4b[2]\n"
+ ".inst 0x4f84ea5a // sdot v26.4s, v18.16b, v4.4b[2]\n"
+ ".inst 0x4f86ea5b // sdot v27.4s, v18.16b, v6.4b[2]\n"
+ ".inst 0x4f88ea5c // sdot v28.4s, v18.16b, v8.4b[2]\n"
+ ".inst 0x4f8aea5d // sdot v29.4s, v18.16b, v10.4b[2]\n"
+ ".inst 0x4f8cea5e // sdot v30.4s, v18.16b, v12.4b[2]\n"
+ ".inst 0x4f8eea5f // sdot v31.4s, v18.16b, v14.4b[2]\n"
+ ".inst 0x4fa0ea78 // sdot v24.4s, v19.16b, v0.4b[3]\n"
+ ".inst 0x4fa2ea79 // sdot v25.4s, v19.16b, v2.4b[3]\n"
+ ".inst 0x4fa4ea7a // sdot v26.4s, v19.16b, v4.4b[3]\n"
+ ".inst 0x4fa6ea7b // sdot v27.4s, v19.16b, v6.4b[3]\n"
+ ".inst 0x4fa8ea7c // sdot v28.4s, v19.16b, v8.4b[3]\n"
+ ".inst 0x4faaea7d // sdot v29.4s, v19.16b, v10.4b[3]\n"
+ ".inst 0x4facea7e // sdot v30.4s, v19.16b, v12.4b[3]\n"
+ ".inst 0x4faeea7f // sdot v31.4s, v19.16b, v14.4b[3]\n"
+ ".inst 0x4f81e298 // sdot v24.4s, v20.16b, v1.4b[0]\n"
+ ".inst 0x4f83e299 // sdot v25.4s, v20.16b, v3.4b[0]\n"
+ ".inst 0x4f85e29a // sdot v26.4s, v20.16b, v5.4b[0]\n"
+ ".inst 0x4f87e29b // sdot v27.4s, v20.16b, v7.4b[0]\n"
+ ".inst 0x4f89e29c // sdot v28.4s, v20.16b, v9.4b[0]\n"
+ ".inst 0x4f8be29d // sdot v29.4s, v20.16b, v11.4b[0]\n"
+ ".inst 0x4f8de29e // sdot v30.4s, v20.16b, v13.4b[0]\n"
+ ".inst 0x4f8fe29f // sdot v31.4s, v20.16b, v15.4b[0]\n"
+ ".inst 0x4fa1e2b8 // sdot v24.4s, v21.16b, v1.4b[1]\n"
+ ".inst 0x4fa3e2b9 // sdot v25.4s, v21.16b, v3.4b[1]\n"
+ ".inst 0x4fa5e2ba // sdot v26.4s, v21.16b, v5.4b[1]\n"
+ ".inst 0x4fa7e2bb // sdot v27.4s, v21.16b, v7.4b[1]\n"
+ ".inst 0x4fa9e2bc // sdot v28.4s, v21.16b, v9.4b[1]\n"
+ ".inst 0x4fabe2bd // sdot v29.4s, v21.16b, v11.4b[1]\n"
+ ".inst 0x4fade2be // sdot v30.4s, v21.16b, v13.4b[1]\n"
+ ".inst 0x4fafe2bf // sdot v31.4s, v21.16b, v15.4b[1]\n"
+ ".inst 0x4f81ead8 // sdot v24.4s, v22.16b, v1.4b[2]\n"
+ ".inst 0x4f83ead9 // sdot v25.4s, v22.16b, v3.4b[2]\n"
+ ".inst 0x4f85eada // sdot v26.4s, v22.16b, v5.4b[2]\n"
+ ".inst 0x4f87eadb // sdot v27.4s, v22.16b, v7.4b[2]\n"
+ ".inst 0x4f89eadc // sdot v28.4s, v22.16b, v9.4b[2]\n"
+ ".inst 0x4f8beadd // sdot v29.4s, v22.16b, v11.4b[2]\n"
+ ".inst 0x4f8deade // sdot v30.4s, v22.16b, v13.4b[2]\n"
+ ".inst 0x4f8feadf // sdot v31.4s, v22.16b, v15.4b[2]\n"
+ ".inst 0x4fa1eaf8 // sdot v24.4s, v23.16b, v1.4b[3]\n"
+ ".inst 0x4fa3eaf9 // sdot v25.4s, v23.16b, v3.4b[3]\n"
+ ".inst 0x4fa5eafa // sdot v26.4s, v23.16b, v5.4b[3]\n"
+ ".inst 0x4fa7eafb // sdot v27.4s, v23.16b, v7.4b[3]\n"
+ ".inst 0x4fa9eafc // sdot v28.4s, v23.16b, v9.4b[3]\n"
+ ".inst 0x4fabeafd // sdot v29.4s, v23.16b, v11.4b[3]\n"
+ ".inst 0x4fadeafe // sdot v30.4s, v23.16b, v13.4b[3]\n"
+ ".inst 0x4fafeaff // sdot v31.4s, v23.16b, v15.4b[3]\n"
"6:\n"
"str q24, [%[c_ptr0]]\n"
+ "add %[c_ptr0], %[c_ptr0], #0x10\n"
"str q25, [c_ptr1]\n"
+ "add c_ptr1, c_ptr1, #0x10\n"
"str q26, [c_ptr2]\n"
+ "add c_ptr2, c_ptr2, #0x10\n"
"str q27, [c_ptr3]\n"
+ "add c_ptr3, c_ptr3, #0x10\n"
"str q28, [c_ptr4]\n"
+ "add c_ptr4, c_ptr4, #0x10\n"
"str q29, [c_ptr5]\n"
+ "add c_ptr5, c_ptr5, #0x10\n"
"str q30, [c_ptr6]\n"
+ "add c_ptr6, c_ptr6, #0x10\n"
"str q31, [c_ptr7]\n"
+ "add c_ptr7, c_ptr7, #0x10\n"
".unreq a_ptr1\n"
".unreq a_ptr2\n"
".unreq a_ptr3\n"