aboutsummaryrefslogtreecommitdiff
path: root/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8s32_dot_16x4/generic.cpp
diff options
context:
space:
mode:
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8s32_dot_16x4/generic.cpp')
-rw-r--r--src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8s32_dot_16x4/generic.cpp3289
1 files changed, 1763 insertions, 1526 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8s32_dot_16x4/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8s32_dot_16x4/generic.cpp
index 01791391c8..fdd45a03cf 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8s32_dot_16x4/generic.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8s32_dot_16x4/generic.cpp
@@ -37,1569 +37,1806 @@ void a64_hybrid_s8s32_dot_16x4(const int8_t *A, int lda, const int8_t *B, int32_
const long loops_count = ((K + 16) / 32) - 1;
K -= loops_count * 32;
const long regs_count = (K / 16) - 1;
+ K -= (regs_count + 1) * 16;
+ const long blocks_count = K / 4;
+ const long odds_count = K - (blocks_count * 4);
for (int y=0; y<M; y+=4) {
const int8_t * const a_ptr0_base = A + (y * lda);
const unsigned long ldab = lda * sizeof(int8_t);
int32_t *c_ptr0 = C + (y * ldc);
- const unsigned long ldcb = ldc * sizeof(int32_t);
for (int x0=0; x0<N; x0+=16ul) {
const long width = std::min((unsigned long)N-x0, 16ul);
const int32_t *betaptr = &beta;
long loops = loops_count;
long regs = regs_count;
+ long blocks = blocks_count;
+ long odds = odds_count;
const int8_t *a_ptr0 = a_ptr0_base;
const int8_t *b_ptr0 = B + (K_stride * x0);
+ const bool use_result_buffer = (width < 16);
+ int32_t result_buffer[64];
+ const unsigned long ldcb = (use_result_buffer ? 16 : ldc) * sizeof(int32_t);
+ int32_t *c_ptr_real = c_ptr0;
+ if (use_result_buffer && !beta0) {
+ for(int cy=0; cy<std::min(M-y, 4); cy++) {
+ for(unsigned int cx=0; cx<width; cx++) {
+ result_buffer[cy * 16 + cx] = c_ptr_real[cy * ldc + cx];
+ }
+ }
+ }
+ if (use_result_buffer) {
+ c_ptr0 = result_buffer;
+ }
switch(M-y) {
case 1:
__asm __volatile (
- "cbz %[beta0], 1f\n"
- "movi v16.4s, #0\n"
- "ldr q0, [%[a_ptr0]]\n"
- "movi v17.4s, #0\n"
- "ldr q8, [%[b_ptr0]]\n"
- "movi v18.4s, #0\n"
- "ldr q9, [%[b_ptr0], #0x10]\n"
- "movi v19.4s, #0\n"
- "ldr q10, [%[b_ptr0], #0x20]\n"
- "ldr q11, [%[b_ptr0], #0x30]\n"
- "add %[a_ptr0], %[a_ptr0], #0x10\n"
- "ldr q12, [%[b_ptr0], #0x40]\n"
- "ldr q13, [%[b_ptr0], #0x50]\n"
- "ldr q14, [%[b_ptr0], #0x60]\n"
- "add %[b_ptr0], %[b_ptr0], #0x80\n"
- "cbz %[loops], 2f\n"
- "b 3f\n"
- "1:\n"
- "ld1r {v15.4s}, [%[betaptr]]\n"
- "ldr q16, [%[c_ptr0]]\n"
- "ldr q17, [%[c_ptr0], #0x10]\n"
- "ldr q18, [%[c_ptr0], #0x20]\n"
- "ldr q19, [%[c_ptr0], #0x30]\n"
- "mul v16.4s, v16.4s, v15.4s\n"
- "ldr q0, [%[a_ptr0]]\n"
- "mul v17.4s, v17.4s, v15.4s\n"
- "ldr q8, [%[b_ptr0]]\n"
- "mul v18.4s, v18.4s, v15.4s\n"
- "ldr q9, [%[b_ptr0], #0x10]\n"
- "mul v19.4s, v19.4s, v15.4s\n"
- "ldr q10, [%[b_ptr0], #0x20]\n"
- "ldr q11, [%[b_ptr0], #0x30]\n"
- "add %[a_ptr0], %[a_ptr0], #0x10\n"
- "ldr q12, [%[b_ptr0], #0x40]\n"
- "ldr q13, [%[b_ptr0], #0x50]\n"
- "ldr q14, [%[b_ptr0], #0x60]\n"
- "add %[b_ptr0], %[b_ptr0], #0x80\n"
- "cbz %[loops], 2f\n"
- "3:\n"
- ".word 0x4f80e110 // sdot v16.4s, v8.16b, v0.4b[0]\n"
- "ldr q15, [%[b_ptr0], #-0x10]\n"
- ".word 0x4f80e131 // sdot v17.4s, v9.16b, v0.4b[0]\n"
- "ldr q4, [%[a_ptr0]]\n"
- ".word 0x4f80e152 // sdot v18.4s, v10.16b, v0.4b[0]\n"
- "ldr q8, [%[b_ptr0]]\n"
- ".word 0x4f80e173 // sdot v19.4s, v11.16b, v0.4b[0]\n"
- "ldr q9, [%[b_ptr0], #0x10]\n"
- ".word 0x4fa0e190 // sdot v16.4s, v12.16b, v0.4b[1]\n"
- "ldr q10, [%[b_ptr0], #0x20]\n"
- ".word 0x4fa0e1b1 // sdot v17.4s, v13.16b, v0.4b[1]\n"
- "ldr q11, [%[b_ptr0], #0x30]\n"
- ".word 0x4fa0e1d2 // sdot v18.4s, v14.16b, v0.4b[1]\n"
- "ldr q12, [%[b_ptr0], #0x40]\n"
- ".word 0x4fa0e1f3 // sdot v19.4s, v15.16b, v0.4b[1]\n"
- "ldr q13, [%[b_ptr0], #0x50]\n"
- ".word 0x4f80e910 // sdot v16.4s, v8.16b, v0.4b[2]\n"
- "ldr q14, [%[b_ptr0], #0x60]\n"
- ".word 0x4f80e931 // sdot v17.4s, v9.16b, v0.4b[2]\n"
- "ldr q15, [%[b_ptr0], #0x70]\n"
- ".word 0x4f80e952 // sdot v18.4s, v10.16b, v0.4b[2]\n"
- "subs %[loops], %[loops], #0x1\n"
- ".word 0x4f80e973 // sdot v19.4s, v11.16b, v0.4b[2]\n"
- "prfm PLDL1KEEP, [%[a_ptr0], #0x40]\n"
- ".word 0x4fa0e990 // sdot v16.4s, v12.16b, v0.4b[3]\n"
- "add %[b_ptr0], %[b_ptr0], #0x100\n"
- ".word 0x4fa0e9b1 // sdot v17.4s, v13.16b, v0.4b[3]\n"
- "ldr q8, [%[b_ptr0], #-0x80]\n"
- ".word 0x4fa0e9d2 // sdot v18.4s, v14.16b, v0.4b[3]\n"
- "ldr q9, [%[b_ptr0], #-0x70]\n"
- ".word 0x4fa0e9f3 // sdot v19.4s, v15.16b, v0.4b[3]\n"
- "ldr q10, [%[b_ptr0], #-0x60]\n"
- "ldr q11, [%[b_ptr0], #-0x50]\n"
- "add %[a_ptr0], %[a_ptr0], #0x20\n"
- ".word 0x4f84e110 // sdot v16.4s, v8.16b, v4.4b[0]\n"
- "ldr q12, [%[b_ptr0], #-0x40]\n"
- ".word 0x4f84e131 // sdot v17.4s, v9.16b, v4.4b[0]\n"
- "ldr q13, [%[b_ptr0], #-0x30]\n"
- ".word 0x4f84e152 // sdot v18.4s, v10.16b, v4.4b[0]\n"
- "ldr q14, [%[b_ptr0], #-0x20]\n"
- ".word 0x4f84e173 // sdot v19.4s, v11.16b, v4.4b[0]\n"
- "ldr q15, [%[b_ptr0], #-0x10]\n"
- ".word 0x4fa4e190 // sdot v16.4s, v12.16b, v4.4b[1]\n"
- "ldr q0, [%[a_ptr0], #-0x10]\n"
- ".word 0x4fa4e1b1 // sdot v17.4s, v13.16b, v4.4b[1]\n"
- "ldr q8, [%[b_ptr0]]\n"
- ".word 0x4fa4e1d2 // sdot v18.4s, v14.16b, v4.4b[1]\n"
- "ldr q9, [%[b_ptr0], #0x10]\n"
- ".word 0x4fa4e1f3 // sdot v19.4s, v15.16b, v4.4b[1]\n"
- "ldr q10, [%[b_ptr0], #0x20]\n"
- "ldr q11, [%[b_ptr0], #0x30]\n"
- ".word 0x4f84e910 // sdot v16.4s, v8.16b, v4.4b[2]\n"
- "ldr q12, [%[b_ptr0], #0x40]\n"
- ".word 0x4f84e931 // sdot v17.4s, v9.16b, v4.4b[2]\n"
- "ldr q13, [%[b_ptr0], #0x50]\n"
- ".word 0x4f84e952 // sdot v18.4s, v10.16b, v4.4b[2]\n"
- "ldr q14, [%[b_ptr0], #0x60]\n"
- ".word 0x4f84e973 // sdot v19.4s, v11.16b, v4.4b[2]\n"
- "ldr q15, [%[b_ptr0], #0x70]\n"
- ".word 0x4fa4e990 // sdot v16.4s, v12.16b, v4.4b[3]\n"
- "add %[b_ptr0], %[b_ptr0], #0x100\n"
- ".word 0x4fa4e9b1 // sdot v17.4s, v13.16b, v4.4b[3]\n"
- "ldr q8, [%[b_ptr0], #-0x80]\n"
- ".word 0x4fa4e9d2 // sdot v18.4s, v14.16b, v4.4b[3]\n"
- "ldr q9, [%[b_ptr0], #-0x70]\n"
- ".word 0x4fa4e9f3 // sdot v19.4s, v15.16b, v4.4b[3]\n"
- "ldr q10, [%[b_ptr0], #-0x60]\n"
- "ldr q11, [%[b_ptr0], #-0x50]\n"
- "ldr q12, [%[b_ptr0], #-0x40]\n"
- "ldr q13, [%[b_ptr0], #-0x30]\n"
- "ldr q14, [%[b_ptr0], #-0x20]\n"
- "b.ne 3b\n"
- "2:\n"
- "ldr q15, [%[b_ptr0], #-0x10]\n"
- "prfm PSTL1KEEP, [%[c_ptr0]]\n"
- "cbz %[regs], 4f\n"
- ".word 0x4f80e110 // sdot v16.4s, v8.16b, v0.4b[0]\n"
- "ldr q4, [%[a_ptr0]]\n"
- ".word 0x4f80e131 // sdot v17.4s, v9.16b, v0.4b[0]\n"
- "ldr q8, [%[b_ptr0]]\n"
- ".word 0x4f80e152 // sdot v18.4s, v10.16b, v0.4b[0]\n"
- "ldr q9, [%[b_ptr0], #0x10]\n"
- ".word 0x4f80e173 // sdot v19.4s, v11.16b, v0.4b[0]\n"
- "ldr q10, [%[b_ptr0], #0x20]\n"
- ".word 0x4fa0e190 // sdot v16.4s, v12.16b, v0.4b[1]\n"
- "ldr q11, [%[b_ptr0], #0x30]\n"
- ".word 0x4fa0e1b1 // sdot v17.4s, v13.16b, v0.4b[1]\n"
- "ldr q12, [%[b_ptr0], #0x40]\n"
- ".word 0x4fa0e1d2 // sdot v18.4s, v14.16b, v0.4b[1]\n"
- "ldr q13, [%[b_ptr0], #0x50]\n"
- ".word 0x4fa0e1f3 // sdot v19.4s, v15.16b, v0.4b[1]\n"
- "ldr q14, [%[b_ptr0], #0x60]\n"
- ".word 0x4f80e910 // sdot v16.4s, v8.16b, v0.4b[2]\n"
- "ldr q15, [%[b_ptr0], #0x70]\n"
- ".word 0x4f80e931 // sdot v17.4s, v9.16b, v0.4b[2]\n"
- "add %[b_ptr0], %[b_ptr0], #0x100\n"
- ".word 0x4f80e952 // sdot v18.4s, v10.16b, v0.4b[2]\n"
- "ldr q8, [%[b_ptr0], #-0x80]\n"
- ".word 0x4f80e973 // sdot v19.4s, v11.16b, v0.4b[2]\n"
- "ldr q9, [%[b_ptr0], #-0x70]\n"
- ".word 0x4fa0e990 // sdot v16.4s, v12.16b, v0.4b[3]\n"
- "ldr q10, [%[b_ptr0], #-0x60]\n"
- ".word 0x4fa0e9b1 // sdot v17.4s, v13.16b, v0.4b[3]\n"
- "ldr q11, [%[b_ptr0], #-0x50]\n"
- ".word 0x4fa0e9d2 // sdot v18.4s, v14.16b, v0.4b[3]\n"
- "ldr q12, [%[b_ptr0], #-0x40]\n"
- ".word 0x4fa0e9f3 // sdot v19.4s, v15.16b, v0.4b[3]\n"
- "ldr q13, [%[b_ptr0], #-0x30]\n"
- ".word 0x4f84e110 // sdot v16.4s, v8.16b, v4.4b[0]\n"
- "ldr q14, [%[b_ptr0], #-0x20]\n"
- ".word 0x4f84e131 // sdot v17.4s, v9.16b, v4.4b[0]\n"
- "ldr q15, [%[b_ptr0], #-0x10]\n"
- ".word 0x4f84e152 // sdot v18.4s, v10.16b, v4.4b[0]\n"
- "ldr q0, [%[a_ptr0], #0x10]\n"
- ".word 0x4f84e173 // sdot v19.4s, v11.16b, v4.4b[0]\n"
- "ldr q8, [%[b_ptr0]]\n"
- ".word 0x4fa4e190 // sdot v16.4s, v12.16b, v4.4b[1]\n"
- "ldr q9, [%[b_ptr0], #0x10]\n"
- ".word 0x4fa4e1b1 // sdot v17.4s, v13.16b, v4.4b[1]\n"
- "ldr q10, [%[b_ptr0], #0x20]\n"
- ".word 0x4fa4e1d2 // sdot v18.4s, v14.16b, v4.4b[1]\n"
- "ldr q11, [%[b_ptr0], #0x30]\n"
- ".word 0x4fa4e1f3 // sdot v19.4s, v15.16b, v4.4b[1]\n"
- "ldr q12, [%[b_ptr0], #0x40]\n"
- ".word 0x4f84e910 // sdot v16.4s, v8.16b, v4.4b[2]\n"
- "ldr q13, [%[b_ptr0], #0x50]\n"
- ".word 0x4f84e931 // sdot v17.4s, v9.16b, v4.4b[2]\n"
- "ldr q14, [%[b_ptr0], #0x60]\n"
- ".word 0x4f84e952 // sdot v18.4s, v10.16b, v4.4b[2]\n"
- "ldr q15, [%[b_ptr0], #0x70]\n"
- ".word 0x4f84e973 // sdot v19.4s, v11.16b, v4.4b[2]\n"
- ".word 0x4fa4e990 // sdot v16.4s, v12.16b, v4.4b[3]\n"
- ".word 0x4fa4e9b1 // sdot v17.4s, v13.16b, v4.4b[3]\n"
- ".word 0x4fa4e9d2 // sdot v18.4s, v14.16b, v4.4b[3]\n"
- ".word 0x4fa4e9f3 // sdot v19.4s, v15.16b, v4.4b[3]\n"
- "b 5f\n"
- "4:\n"
- ".word 0x4f80e110 // sdot v16.4s, v8.16b, v0.4b[0]\n"
- "ldr q4, [%[a_ptr0]]\n"
- ".word 0x4f80e131 // sdot v17.4s, v9.16b, v0.4b[0]\n"
- "ldr q8, [%[b_ptr0]]\n"
- ".word 0x4f80e152 // sdot v18.4s, v10.16b, v0.4b[0]\n"
- "ldr q9, [%[b_ptr0], #0x10]\n"
- ".word 0x4f80e173 // sdot v19.4s, v11.16b, v0.4b[0]\n"
- "ldr q10, [%[b_ptr0], #0x20]\n"
- ".word 0x4fa0e190 // sdot v16.4s, v12.16b, v0.4b[1]\n"
- "ldr q11, [%[b_ptr0], #0x30]\n"
- ".word 0x4fa0e1b1 // sdot v17.4s, v13.16b, v0.4b[1]\n"
- "ldr q12, [%[b_ptr0], #0x40]\n"
- ".word 0x4fa0e1d2 // sdot v18.4s, v14.16b, v0.4b[1]\n"
- "ldr q13, [%[b_ptr0], #0x50]\n"
- ".word 0x4fa0e1f3 // sdot v19.4s, v15.16b, v0.4b[1]\n"
- "ldr q14, [%[b_ptr0], #0x60]\n"
- ".word 0x4f80e910 // sdot v16.4s, v8.16b, v0.4b[2]\n"
- "ldr q15, [%[b_ptr0], #0x70]\n"
- ".word 0x4f80e931 // sdot v17.4s, v9.16b, v0.4b[2]\n"
- ".word 0x4f80e952 // sdot v18.4s, v10.16b, v0.4b[2]\n"
- ".word 0x4f80e973 // sdot v19.4s, v11.16b, v0.4b[2]\n"
- ".word 0x4fa0e990 // sdot v16.4s, v12.16b, v0.4b[3]\n"
- ".word 0x4fa0e9b1 // sdot v17.4s, v13.16b, v0.4b[3]\n"
- ".word 0x4fa0e9d2 // sdot v18.4s, v14.16b, v0.4b[3]\n"
- ".word 0x4fa0e9f3 // sdot v19.4s, v15.16b, v0.4b[3]\n"
- "5:\n"
- "str q16, [%[c_ptr0]]\n"
- "str q17, [%[c_ptr0], #0x10]\n"
- "str q18, [%[c_ptr0], #0x20]\n"
- "str q19, [%[c_ptr0], #0x30]\n"
- "add %[c_ptr0], %[c_ptr0], #0x40\n"
- : [a_ptr0] "+r" (a_ptr0), [b_ptr0] "+r" (b_ptr0), [c_ptr0] "+r" (c_ptr0), [loops] "+r" (loops), [regs] "+r" (regs)
- : [betaptr] "r" (betaptr), [width] "r" (width), [beta0] "r" (beta0), [lda] "r" (ldab), [ldc] "r" (ldcb)
- : "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "cc", "memory"
+ "cbz %[beta0], 1f\n"
+ "movi v16.4s, #0\n"
+ "ldr q0, [%[a_ptr0]]\n"
+ "movi v17.4s, #0\n"
+ "ldr q8, [%[b_ptr0]]\n"
+ "movi v18.4s, #0\n"
+ "ldr q9, [%[b_ptr0], #0x10]\n"
+ "movi v19.4s, #0\n"
+ "ldr q10, [%[b_ptr0], #0x20]\n"
+ "ldr q11, [%[b_ptr0], #0x30]\n"
+ "add %[a_ptr0], %[a_ptr0], #0x10\n"
+ "ldr q12, [%[b_ptr0], #0x40]\n"
+ "ldr q13, [%[b_ptr0], #0x50]\n"
+ "ldr q14, [%[b_ptr0], #0x60]\n"
+ "add %[b_ptr0], %[b_ptr0], #0x80\n"
+ "cbz %[loops], 2f\n"
+ "b 3f\n"
+ "1:\n"
+ "ld1r {v15.4s}, [%[betaptr]]\n"
+ "ldr q16, [%[c_ptr0]]\n"
+ "ldr q17, [%[c_ptr0], #0x10]\n"
+ "ldr q18, [%[c_ptr0], #0x20]\n"
+ "ldr q19, [%[c_ptr0], #0x30]\n"
+ "mul v16.4s, v16.4s, v15.4s\n"
+ "ldr q0, [%[a_ptr0]]\n"
+ "mul v17.4s, v17.4s, v15.4s\n"
+ "ldr q8, [%[b_ptr0]]\n"
+ "mul v18.4s, v18.4s, v15.4s\n"
+ "ldr q9, [%[b_ptr0], #0x10]\n"
+ "mul v19.4s, v19.4s, v15.4s\n"
+ "ldr q10, [%[b_ptr0], #0x20]\n"
+ "ldr q11, [%[b_ptr0], #0x30]\n"
+ "add %[a_ptr0], %[a_ptr0], #0x10\n"
+ "ldr q12, [%[b_ptr0], #0x40]\n"
+ "ldr q13, [%[b_ptr0], #0x50]\n"
+ "ldr q14, [%[b_ptr0], #0x60]\n"
+ "add %[b_ptr0], %[b_ptr0], #0x80\n"
+ "cbz %[loops], 2f\n"
+ "3:\n"
+ ".word 0x4f80e110 // sdot v16.4s, v8.16b, v0.4b[0]\n"
+ "ldr q15, [%[b_ptr0], #-0x10]\n"
+ ".word 0x4f80e131 // sdot v17.4s, v9.16b, v0.4b[0]\n"
+ "ldr q4, [%[a_ptr0]]\n"
+ ".word 0x4f80e152 // sdot v18.4s, v10.16b, v0.4b[0]\n"
+ "ldr q8, [%[b_ptr0]]\n"
+ ".word 0x4f80e173 // sdot v19.4s, v11.16b, v0.4b[0]\n"
+ "ldr q9, [%[b_ptr0], #0x10]\n"
+ ".word 0x4fa0e190 // sdot v16.4s, v12.16b, v0.4b[1]\n"
+ "ldr q10, [%[b_ptr0], #0x20]\n"
+ ".word 0x4fa0e1b1 // sdot v17.4s, v13.16b, v0.4b[1]\n"
+ "ldr q11, [%[b_ptr0], #0x30]\n"
+ ".word 0x4fa0e1d2 // sdot v18.4s, v14.16b, v0.4b[1]\n"
+ "ldr q12, [%[b_ptr0], #0x40]\n"
+ ".word 0x4fa0e1f3 // sdot v19.4s, v15.16b, v0.4b[1]\n"
+ "ldr q13, [%[b_ptr0], #0x50]\n"
+ ".word 0x4f80e910 // sdot v16.4s, v8.16b, v0.4b[2]\n"
+ "ldr q14, [%[b_ptr0], #0x60]\n"
+ ".word 0x4f80e931 // sdot v17.4s, v9.16b, v0.4b[2]\n"
+ "ldr q15, [%[b_ptr0], #0x70]\n"
+ ".word 0x4f80e952 // sdot v18.4s, v10.16b, v0.4b[2]\n"
+ "subs %[loops], %[loops], #0x1\n"
+ ".word 0x4f80e973 // sdot v19.4s, v11.16b, v0.4b[2]\n"
+ "prfm PLDL1KEEP, [%[a_ptr0], #0x40]\n"
+ ".word 0x4fa0e990 // sdot v16.4s, v12.16b, v0.4b[3]\n"
+ "add %[b_ptr0], %[b_ptr0], #0x100\n"
+ ".word 0x4fa0e9b1 // sdot v17.4s, v13.16b, v0.4b[3]\n"
+ "ldr q8, [%[b_ptr0], #-0x80]\n"
+ ".word 0x4fa0e9d2 // sdot v18.4s, v14.16b, v0.4b[3]\n"
+ "ldr q9, [%[b_ptr0], #-0x70]\n"
+ ".word 0x4fa0e9f3 // sdot v19.4s, v15.16b, v0.4b[3]\n"
+ "ldr q10, [%[b_ptr0], #-0x60]\n"
+ "ldr q11, [%[b_ptr0], #-0x50]\n"
+ "add %[a_ptr0], %[a_ptr0], #0x20\n"
+ ".word 0x4f84e110 // sdot v16.4s, v8.16b, v4.4b[0]\n"
+ "ldr q12, [%[b_ptr0], #-0x40]\n"
+ ".word 0x4f84e131 // sdot v17.4s, v9.16b, v4.4b[0]\n"
+ "ldr q13, [%[b_ptr0], #-0x30]\n"
+ ".word 0x4f84e152 // sdot v18.4s, v10.16b, v4.4b[0]\n"
+ "ldr q14, [%[b_ptr0], #-0x20]\n"
+ ".word 0x4f84e173 // sdot v19.4s, v11.16b, v4.4b[0]\n"
+ "ldr q15, [%[b_ptr0], #-0x10]\n"
+ ".word 0x4fa4e190 // sdot v16.4s, v12.16b, v4.4b[1]\n"
+ "ldr q0, [%[a_ptr0], #-0x10]\n"
+ ".word 0x4fa4e1b1 // sdot v17.4s, v13.16b, v4.4b[1]\n"
+ "ldr q8, [%[b_ptr0]]\n"
+ ".word 0x4fa4e1d2 // sdot v18.4s, v14.16b, v4.4b[1]\n"
+ "ldr q9, [%[b_ptr0], #0x10]\n"
+ ".word 0x4fa4e1f3 // sdot v19.4s, v15.16b, v4.4b[1]\n"
+ "ldr q10, [%[b_ptr0], #0x20]\n"
+ "ldr q11, [%[b_ptr0], #0x30]\n"
+ ".word 0x4f84e910 // sdot v16.4s, v8.16b, v4.4b[2]\n"
+ "ldr q12, [%[b_ptr0], #0x40]\n"
+ ".word 0x4f84e931 // sdot v17.4s, v9.16b, v4.4b[2]\n"
+ "ldr q13, [%[b_ptr0], #0x50]\n"
+ ".word 0x4f84e952 // sdot v18.4s, v10.16b, v4.4b[2]\n"
+ "ldr q14, [%[b_ptr0], #0x60]\n"
+ ".word 0x4f84e973 // sdot v19.4s, v11.16b, v4.4b[2]\n"
+ "ldr q15, [%[b_ptr0], #0x70]\n"
+ ".word 0x4fa4e990 // sdot v16.4s, v12.16b, v4.4b[3]\n"
+ "add %[b_ptr0], %[b_ptr0], #0x100\n"
+ ".word 0x4fa4e9b1 // sdot v17.4s, v13.16b, v4.4b[3]\n"
+ "ldr q8, [%[b_ptr0], #-0x80]\n"
+ ".word 0x4fa4e9d2 // sdot v18.4s, v14.16b, v4.4b[3]\n"
+ "ldr q9, [%[b_ptr0], #-0x70]\n"
+ ".word 0x4fa4e9f3 // sdot v19.4s, v15.16b, v4.4b[3]\n"
+ "ldr q10, [%[b_ptr0], #-0x60]\n"
+ "ldr q11, [%[b_ptr0], #-0x50]\n"
+ "ldr q12, [%[b_ptr0], #-0x40]\n"
+ "ldr q13, [%[b_ptr0], #-0x30]\n"
+ "ldr q14, [%[b_ptr0], #-0x20]\n"
+ "b.ne 3b\n"
+ "2:\n"
+ "ldr q15, [%[b_ptr0], #-0x10]\n"
+ "prfm PSTL1KEEP, [%[c_ptr0]]\n"
+ "cbz %[regs], 4f\n"
+ ".word 0x4f80e110 // sdot v16.4s, v8.16b, v0.4b[0]\n"
+ "ldr q4, [%[a_ptr0]]\n"
+ ".word 0x4f80e131 // sdot v17.4s, v9.16b, v0.4b[0]\n"
+ "ldr q8, [%[b_ptr0]]\n"
+ ".word 0x4f80e152 // sdot v18.4s, v10.16b, v0.4b[0]\n"
+ "ldr q9, [%[b_ptr0], #0x10]\n"
+ ".word 0x4f80e173 // sdot v19.4s, v11.16b, v0.4b[0]\n"
+ "ldr q10, [%[b_ptr0], #0x20]\n"
+ ".word 0x4fa0e190 // sdot v16.4s, v12.16b, v0.4b[1]\n"
+ "ldr q11, [%[b_ptr0], #0x30]\n"
+ ".word 0x4fa0e1b1 // sdot v17.4s, v13.16b, v0.4b[1]\n"
+ "ldr q12, [%[b_ptr0], #0x40]\n"
+ ".word 0x4fa0e1d2 // sdot v18.4s, v14.16b, v0.4b[1]\n"
+ "ldr q13, [%[b_ptr0], #0x50]\n"
+ ".word 0x4fa0e1f3 // sdot v19.4s, v15.16b, v0.4b[1]\n"
+ "ldr q14, [%[b_ptr0], #0x60]\n"
+ ".word 0x4f80e910 // sdot v16.4s, v8.16b, v0.4b[2]\n"
+ "ldr q15, [%[b_ptr0], #0x70]\n"
+ ".word 0x4f80e931 // sdot v17.4s, v9.16b, v0.4b[2]\n"
+ "add %[b_ptr0], %[b_ptr0], #0x100\n"
+ ".word 0x4f80e952 // sdot v18.4s, v10.16b, v0.4b[2]\n"
+ "ldr q8, [%[b_ptr0], #-0x80]\n"
+ ".word 0x4f80e973 // sdot v19.4s, v11.16b, v0.4b[2]\n"
+ "ldr q9, [%[b_ptr0], #-0x70]\n"
+ ".word 0x4fa0e990 // sdot v16.4s, v12.16b, v0.4b[3]\n"
+ "ldr q10, [%[b_ptr0], #-0x60]\n"
+ ".word 0x4fa0e9b1 // sdot v17.4s, v13.16b, v0.4b[3]\n"
+ "ldr q11, [%[b_ptr0], #-0x50]\n"
+ ".word 0x4fa0e9d2 // sdot v18.4s, v14.16b, v0.4b[3]\n"
+ "ldr q12, [%[b_ptr0], #-0x40]\n"
+ ".word 0x4fa0e9f3 // sdot v19.4s, v15.16b, v0.4b[3]\n"
+ "ldr q13, [%[b_ptr0], #-0x30]\n"
+ ".word 0x4f84e110 // sdot v16.4s, v8.16b, v4.4b[0]\n"
+ "ldr q14, [%[b_ptr0], #-0x20]\n"
+ ".word 0x4f84e131 // sdot v17.4s, v9.16b, v4.4b[0]\n"
+ "ldr q15, [%[b_ptr0], #-0x10]\n"
+ ".word 0x4f84e152 // sdot v18.4s, v10.16b, v4.4b[0]\n"
+ "ldr q8, [%[b_ptr0]]\n"
+ ".word 0x4f84e173 // sdot v19.4s, v11.16b, v4.4b[0]\n"
+ "ldr q9, [%[b_ptr0], #0x10]\n"
+ ".word 0x4fa4e190 // sdot v16.4s, v12.16b, v4.4b[1]\n"
+ "ldr q10, [%[b_ptr0], #0x20]\n"
+ ".word 0x4fa4e1b1 // sdot v17.4s, v13.16b, v4.4b[1]\n"
+ "ldr q11, [%[b_ptr0], #0x30]\n"
+ ".word 0x4fa4e1d2 // sdot v18.4s, v14.16b, v4.4b[1]\n"
+ "ldr q12, [%[b_ptr0], #0x40]\n"
+ ".word 0x4fa4e1f3 // sdot v19.4s, v15.16b, v4.4b[1]\n"
+ "ldr q13, [%[b_ptr0], #0x50]\n"
+ ".word 0x4f84e910 // sdot v16.4s, v8.16b, v4.4b[2]\n"
+ "ldr q14, [%[b_ptr0], #0x60]\n"
+ ".word 0x4f84e931 // sdot v17.4s, v9.16b, v4.4b[2]\n"
+ "ldr q15, [%[b_ptr0], #0x70]\n"
+ ".word 0x4f84e952 // sdot v18.4s, v10.16b, v4.4b[2]\n"
+ "add %[b_ptr0], %[b_ptr0], #0x80\n"
+ ".word 0x4f84e973 // sdot v19.4s, v11.16b, v4.4b[2]\n"
+ "add %[a_ptr0], %[a_ptr0], #0x10\n"
+ ".word 0x4fa4e990 // sdot v16.4s, v12.16b, v4.4b[3]\n"
+ ".word 0x4fa4e9b1 // sdot v17.4s, v13.16b, v4.4b[3]\n"
+ ".word 0x4fa4e9d2 // sdot v18.4s, v14.16b, v4.4b[3]\n"
+ ".word 0x4fa4e9f3 // sdot v19.4s, v15.16b, v4.4b[3]\n"
+ "b 5f\n"
+ "4:\n"
+ ".word 0x4f80e110 // sdot v16.4s, v8.16b, v0.4b[0]\n"
+ "ldr q8, [%[b_ptr0]]\n"
+ ".word 0x4f80e131 // sdot v17.4s, v9.16b, v0.4b[0]\n"
+ "ldr q9, [%[b_ptr0], #0x10]\n"
+ ".word 0x4f80e152 // sdot v18.4s, v10.16b, v0.4b[0]\n"
+ "ldr q10, [%[b_ptr0], #0x20]\n"
+ ".word 0x4f80e173 // sdot v19.4s, v11.16b, v0.4b[0]\n"
+ "ldr q11, [%[b_ptr0], #0x30]\n"
+ ".word 0x4fa0e190 // sdot v16.4s, v12.16b, v0.4b[1]\n"
+ "ldr q12, [%[b_ptr0], #0x40]\n"
+ ".word 0x4fa0e1b1 // sdot v17.4s, v13.16b, v0.4b[1]\n"
+ "ldr q13, [%[b_ptr0], #0x50]\n"
+ ".word 0x4fa0e1d2 // sdot v18.4s, v14.16b, v0.4b[1]\n"
+ "ldr q14, [%[b_ptr0], #0x60]\n"
+ ".word 0x4fa0e1f3 // sdot v19.4s, v15.16b, v0.4b[1]\n"
+ "ldr q15, [%[b_ptr0], #0x70]\n"
+ ".word 0x4f80e910 // sdot v16.4s, v8.16b, v0.4b[2]\n"
+ "add %[b_ptr0], %[b_ptr0], #0x80\n"
+ ".word 0x4f80e931 // sdot v17.4s, v9.16b, v0.4b[2]\n"
+ ".word 0x4f80e952 // sdot v18.4s, v10.16b, v0.4b[2]\n"
+ ".word 0x4f80e973 // sdot v19.4s, v11.16b, v0.4b[2]\n"
+ ".word 0x4fa0e990 // sdot v16.4s, v12.16b, v0.4b[3]\n"
+ ".word 0x4fa0e9b1 // sdot v17.4s, v13.16b, v0.4b[3]\n"
+ ".word 0x4fa0e9d2 // sdot v18.4s, v14.16b, v0.4b[3]\n"
+ ".word 0x4fa0e9f3 // sdot v19.4s, v15.16b, v0.4b[3]\n"
+ "5:\n"
+ "cbz %[blocks], 6f\n"
+ "7:\n"
+ "ldr q8, [%[b_ptr0]]\n"
+ "subs %[blocks], %[blocks], #0x1\n"
+ "ldr q9, [%[b_ptr0], #0x10]\n"
+ "ldr s0, [%[a_ptr0]]\n"
+ "ldr q10, [%[b_ptr0], #0x20]\n"
+ "add %[a_ptr0], %[a_ptr0], #0x4\n"
+ "ldr q11, [%[b_ptr0], #0x30]\n"
+ "add %[b_ptr0], %[b_ptr0], #0x40\n"
+ ".word 0x4f80e110 // sdot v16.4s, v8.16b, v0.4b[0]\n"
+ ".word 0x4f80e131 // sdot v17.4s, v9.16b, v0.4b[0]\n"
+ ".word 0x4f80e152 // sdot v18.4s, v10.16b, v0.4b[0]\n"
+ ".word 0x4f80e173 // sdot v19.4s, v11.16b, v0.4b[0]\n"
+ "b.ne 7b\n"
+ "6:\n"
+ "cbz %[odds], 8f\n"
+ "ld1 {v0.b}[0], [%[a_ptr0]], #1\n"
+ "subs %[odds], %[odds], #0x1\n"
+ "b.eq 9f\n"
+ "ld1 {v0.b}[1], [%[a_ptr0]], #1\n"
+ "subs %[odds], %[odds], #0x1\n"
+ "b.eq 9f\n"
+ "ld1 {v0.b}[2], [%[a_ptr0]]\n"
+ "9:\n"
+ "ldr q8, [%[b_ptr0]]\n"
+ "ldr q9, [%[b_ptr0], #0x10]\n"
+ "ldr q10, [%[b_ptr0], #0x20]\n"
+ "ldr q11, [%[b_ptr0], #0x30]\n"
+ ".word 0x4f80e110 // sdot v16.4s, v8.16b, v0.4b[0]\n"
+ ".word 0x4f80e131 // sdot v17.4s, v9.16b, v0.4b[0]\n"
+ ".word 0x4f80e152 // sdot v18.4s, v10.16b, v0.4b[0]\n"
+ ".word 0x4f80e173 // sdot v19.4s, v11.16b, v0.4b[0]\n"
+ "8:\n"
+ "str q16, [%[c_ptr0]]\n"
+ "str q17, [%[c_ptr0], #0x10]\n"
+ "str q18, [%[c_ptr0], #0x20]\n"
+ "str q19, [%[c_ptr0], #0x30]\n"
+ "add %[c_ptr0], %[c_ptr0], #0x40\n"
+ : [a_ptr0] "+r" (a_ptr0), [b_ptr0] "+r" (b_ptr0), [c_ptr0] "+r" (c_ptr0), [loops] "+r" (loops), [regs] "+r" (regs), [blocks] "+r" (blocks), [odds] "+r" (odds)
+ : [betaptr] "r" (betaptr), [width] "r" (width), [beta0] "r" (beta0), [lda] "r" (ldab), [ldc] "r" (ldcb)
+ : "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "cc", "memory"
);
break;
case 2:
__asm __volatile (
- "a_ptr1 .req X0\n"
- "c_ptr1 .req X1\n"
- "add a_ptr1, %[a_ptr0], %[lda]\n"
- "add c_ptr1, %[c_ptr0], %[ldc]\n"
- "cbz %[beta0], 1f\n"
- "movi v16.4s, #0\n"
- "ldr q0, [%[a_ptr0]]\n"
- "movi v17.4s, #0\n"
- "ldr q1, [a_ptr1]\n"
- "movi v18.4s, #0\n"
- "ldr q8, [%[b_ptr0]]\n"
- "movi v19.4s, #0\n"
- "ldr q9, [%[b_ptr0], #0x10]\n"
- "movi v20.4s, #0\n"
- "ldr q10, [%[b_ptr0], #0x20]\n"
- "movi v21.4s, #0\n"
- "ldr q11, [%[b_ptr0], #0x30]\n"
- "movi v22.4s, #0\n"
- "ldr q12, [%[b_ptr0], #0x40]\n"
- "movi v23.4s, #0\n"
- "ldr q13, [%[b_ptr0], #0x50]\n"
- "ldr q14, [%[b_ptr0], #0x60]\n"
- "add %[a_ptr0], %[a_ptr0], #0x10\n"
- "add a_ptr1, a_ptr1, #0x10\n"
- "add %[b_ptr0], %[b_ptr0], #0x80\n"
- "cbz %[loops], 2f\n"
- "b 3f\n"
- "1:\n"
- "ld1r {v15.4s}, [%[betaptr]]\n"
- "ldr q16, [%[c_ptr0]]\n"
- "ldr q17, [%[c_ptr0], #0x10]\n"
- "ldr q18, [%[c_ptr0], #0x20]\n"
- "ldr q19, [%[c_ptr0], #0x30]\n"
- "mul v16.4s, v16.4s, v15.4s\n"
- "ldr q20, [c_ptr1]\n"
- "mul v17.4s, v17.4s, v15.4s\n"
- "ldr q21, [c_ptr1, #0x10]\n"
- "mul v18.4s, v18.4s, v15.4s\n"
- "ldr q22, [c_ptr1, #0x20]\n"
- "mul v19.4s, v19.4s, v15.4s\n"
- "ldr q23, [c_ptr1, #0x30]\n"
- "mul v20.4s, v20.4s, v15.4s\n"
- "ldr q0, [%[a_ptr0]]\n"
- "mul v21.4s, v21.4s, v15.4s\n"
- "ldr q1, [a_ptr1]\n"
- "mul v22.4s, v22.4s, v15.4s\n"
- "ldr q8, [%[b_ptr0]]\n"
- "mul v23.4s, v23.4s, v15.4s\n"
- "ldr q9, [%[b_ptr0], #0x10]\n"
- "ldr q10, [%[b_ptr0], #0x20]\n"
- "add %[a_ptr0], %[a_ptr0], #0x10\n"
- "ldr q11, [%[b_ptr0], #0x30]\n"
- "add a_ptr1, a_ptr1, #0x10\n"
- "ldr q12, [%[b_ptr0], #0x40]\n"
- "ldr q13, [%[b_ptr0], #0x50]\n"
- "ldr q14, [%[b_ptr0], #0x60]\n"
- "add %[b_ptr0], %[b_ptr0], #0x80\n"
- "cbz %[loops], 2f\n"
- "3:\n"
- ".word 0x4f80e110 // sdot v16.4s, v8.16b, v0.4b[0]\n"
- "ldr q15, [%[b_ptr0], #-0x10]\n"
- ".word 0x4f81e114 // sdot v20.4s, v8.16b, v1.4b[0]\n"
- "ldr q4, [%[a_ptr0]]\n"
- ".word 0x4f80e131 // sdot v17.4s, v9.16b, v0.4b[0]\n"
- "ldr q5, [a_ptr1]\n"
- ".word 0x4f81e135 // sdot v21.4s, v9.16b, v1.4b[0]\n"
- "ldr q8, [%[b_ptr0]]\n"
- ".word 0x4f80e152 // sdot v18.4s, v10.16b, v0.4b[0]\n"
- "ldr q9, [%[b_ptr0], #0x10]\n"
- ".word 0x4f81e156 // sdot v22.4s, v10.16b, v1.4b[0]\n"
- "ldr q10, [%[b_ptr0], #0x20]\n"
- ".word 0x4f80e173 // sdot v19.4s, v11.16b, v0.4b[0]\n"
- "subs %[loops], %[loops], #0x1\n"
- ".word 0x4f81e177 // sdot v23.4s, v11.16b, v1.4b[0]\n"
- "ldr q11, [%[b_ptr0], #0x30]\n"
- ".word 0x4fa0e190 // sdot v16.4s, v12.16b, v0.4b[1]\n"
- "prfm PLDL1KEEP, [%[a_ptr0], #0x40]\n"
- ".word 0x4fa1e194 // sdot v20.4s, v12.16b, v1.4b[1]\n"
- "ldr q12, [%[b_ptr0], #0x40]\n"
- ".word 0x4fa0e1b1 // sdot v17.4s, v13.16b, v0.4b[1]\n"
- "add %[a_ptr0], %[a_ptr0], #0x20\n"
- ".word 0x4fa1e1b5 // sdot v21.4s, v13.16b, v1.4b[1]\n"
- "ldr q13, [%[b_ptr0], #0x50]\n"
- ".word 0x4fa0e1d2 // sdot v18.4s, v14.16b, v0.4b[1]\n"
- "add a_ptr1, a_ptr1, #0x20\n"
- ".word 0x4fa1e1d6 // sdot v22.4s, v14.16b, v1.4b[1]\n"
- "ldr q14, [%[b_ptr0], #0x60]\n"
- ".word 0x4fa0e1f3 // sdot v19.4s, v15.16b, v0.4b[1]\n"
- "prfm PLDL1KEEP, [a_ptr1, #0x40]\n"
- ".word 0x4fa1e1f7 // sdot v23.4s, v15.16b, v1.4b[1]\n"
- "ldr q15, [%[b_ptr0], #0x70]\n"
- ".word 0x4f80e910 // sdot v16.4s, v8.16b, v0.4b[2]\n"
- "add %[b_ptr0], %[b_ptr0], #0x100\n"
- ".word 0x4f81e914 // sdot v20.4s, v8.16b, v1.4b[2]\n"
- "ldr q8, [%[b_ptr0], #-0x80]\n"
- ".word 0x4f80e931 // sdot v17.4s, v9.16b, v0.4b[2]\n"
- ".word 0x4f81e935 // sdot v21.4s, v9.16b, v1.4b[2]\n"
- "ldr q9, [%[b_ptr0], #-0x70]\n"
- ".word 0x4f80e952 // sdot v18.4s, v10.16b, v0.4b[2]\n"
- ".word 0x4f81e956 // sdot v22.4s, v10.16b, v1.4b[2]\n"
- "ldr q10, [%[b_ptr0], #-0x60]\n"
- ".word 0x4f80e973 // sdot v19.4s, v11.16b, v0.4b[2]\n"
- ".word 0x4f81e977 // sdot v23.4s, v11.16b, v1.4b[2]\n"
- "ldr q11, [%[b_ptr0], #-0x50]\n"
- ".word 0x4fa0e990 // sdot v16.4s, v12.16b, v0.4b[3]\n"
- ".word 0x4fa1e994 // sdot v20.4s, v12.16b, v1.4b[3]\n"
- "ldr q12, [%[b_ptr0], #-0x40]\n"
- ".word 0x4fa0e9b1 // sdot v17.4s, v13.16b, v0.4b[3]\n"
- ".word 0x4fa1e9b5 // sdot v21.4s, v13.16b, v1.4b[3]\n"
- "ldr q13, [%[b_ptr0], #-0x30]\n"
- ".word 0x4fa0e9d2 // sdot v18.4s, v14.16b, v0.4b[3]\n"
- ".word 0x4fa1e9d6 // sdot v22.4s, v14.16b, v1.4b[3]\n"
- "ldr q14, [%[b_ptr0], #-0x20]\n"
- ".word 0x4fa0e9f3 // sdot v19.4s, v15.16b, v0.4b[3]\n"
- "ldr q0, [%[a_ptr0], #-0x10]\n"
- ".word 0x4fa1e9f7 // sdot v23.4s, v15.16b, v1.4b[3]\n"
- "ldr q15, [%[b_ptr0], #-0x10]\n"
- ".word 0x4f84e110 // sdot v16.4s, v8.16b, v4.4b[0]\n"
- "ldr q1, [a_ptr1, #-0x10]\n"
- ".word 0x4f85e114 // sdot v20.4s, v8.16b, v5.4b[0]\n"
- "ldr q8, [%[b_ptr0]]\n"
- ".word 0x4f84e131 // sdot v17.4s, v9.16b, v4.4b[0]\n"
- ".word 0x4f85e135 // sdot v21.4s, v9.16b, v5.4b[0]\n"
- "ldr q9, [%[b_ptr0], #0x10]\n"
- ".word 0x4f84e152 // sdot v18.4s, v10.16b, v4.4b[0]\n"
- ".word 0x4f85e156 // sdot v22.4s, v10.16b, v5.4b[0]\n"
- "ldr q10, [%[b_ptr0], #0x20]\n"
- ".word 0x4f84e173 // sdot v19.4s, v11.16b, v4.4b[0]\n"
- ".word 0x4f85e177 // sdot v23.4s, v11.16b, v5.4b[0]\n"
- "ldr q11, [%[b_ptr0], #0x30]\n"
- ".word 0x4fa4e190 // sdot v16.4s, v12.16b, v4.4b[1]\n"
- ".word 0x4fa5e194 // sdot v20.4s, v12.16b, v5.4b[1]\n"
- "ldr q12, [%[b_ptr0], #0x40]\n"
- ".word 0x4fa4e1b1 // sdot v17.4s, v13.16b, v4.4b[1]\n"
- ".word 0x4fa5e1b5 // sdot v21.4s, v13.16b, v5.4b[1]\n"
- "ldr q13, [%[b_ptr0], #0x50]\n"
- ".word 0x4fa4e1d2 // sdot v18.4s, v14.16b, v4.4b[1]\n"
- ".word 0x4fa5e1d6 // sdot v22.4s, v14.16b, v5.4b[1]\n"
- "ldr q14, [%[b_ptr0], #0x60]\n"
- ".word 0x4fa4e1f3 // sdot v19.4s, v15.16b, v4.4b[1]\n"
- ".word 0x4fa5e1f7 // sdot v23.4s, v15.16b, v5.4b[1]\n"
- "ldr q15, [%[b_ptr0], #0x70]\n"
- ".word 0x4f84e910 // sdot v16.4s, v8.16b, v4.4b[2]\n"
- "add %[b_ptr0], %[b_ptr0], #0x100\n"
- ".word 0x4f85e914 // sdot v20.4s, v8.16b, v5.4b[2]\n"
- "ldr q8, [%[b_ptr0], #-0x80]\n"
- ".word 0x4f84e931 // sdot v17.4s, v9.16b, v4.4b[2]\n"
- ".word 0x4f85e935 // sdot v21.4s, v9.16b, v5.4b[2]\n"
- "ldr q9, [%[b_ptr0], #-0x70]\n"
- ".word 0x4f84e952 // sdot v18.4s, v10.16b, v4.4b[2]\n"
- ".word 0x4f85e956 // sdot v22.4s, v10.16b, v5.4b[2]\n"
- "ldr q10, [%[b_ptr0], #-0x60]\n"
- ".word 0x4f84e973 // sdot v19.4s, v11.16b, v4.4b[2]\n"
- ".word 0x4f85e977 // sdot v23.4s, v11.16b, v5.4b[2]\n"
- "ldr q11, [%[b_ptr0], #-0x50]\n"
- ".word 0x4fa4e990 // sdot v16.4s, v12.16b, v4.4b[3]\n"
- ".word 0x4fa5e994 // sdot v20.4s, v12.16b, v5.4b[3]\n"
- "ldr q12, [%[b_ptr0], #-0x40]\n"
- ".word 0x4fa4e9b1 // sdot v17.4s, v13.16b, v4.4b[3]\n"
- ".word 0x4fa5e9b5 // sdot v21.4s, v13.16b, v5.4b[3]\n"
- "ldr q13, [%[b_ptr0], #-0x30]\n"
- ".word 0x4fa4e9d2 // sdot v18.4s, v14.16b, v4.4b[3]\n"
- ".word 0x4fa5e9d6 // sdot v22.4s, v14.16b, v5.4b[3]\n"
- "ldr q14, [%[b_ptr0], #-0x20]\n"
- ".word 0x4fa4e9f3 // sdot v19.4s, v15.16b, v4.4b[3]\n"
- ".word 0x4fa5e9f7 // sdot v23.4s, v15.16b, v5.4b[3]\n"
- "b.ne 3b\n"
- "2:\n"
- "ldr q15, [%[b_ptr0], #-0x10]\n"
- "prfm PSTL1KEEP, [%[c_ptr0]]\n"
- "prfm PSTL1KEEP, [c_ptr1]\n"
- "cbz %[regs], 4f\n"
- ".word 0x4f80e110 // sdot v16.4s, v8.16b, v0.4b[0]\n"
- "ldr q4, [%[a_ptr0]]\n"
- ".word 0x4f81e114 // sdot v20.4s, v8.16b, v1.4b[0]\n"
- "ldr q5, [a_ptr1]\n"
- ".word 0x4f80e131 // sdot v17.4s, v9.16b, v0.4b[0]\n"
- "ldr q8, [%[b_ptr0]]\n"
- ".word 0x4f81e135 // sdot v21.4s, v9.16b, v1.4b[0]\n"
- "ldr q9, [%[b_ptr0], #0x10]\n"
- ".word 0x4f80e152 // sdot v18.4s, v10.16b, v0.4b[0]\n"
- ".word 0x4f81e156 // sdot v22.4s, v10.16b, v1.4b[0]\n"
- "ldr q10, [%[b_ptr0], #0x20]\n"
- ".word 0x4f80e173 // sdot v19.4s, v11.16b, v0.4b[0]\n"
- ".word 0x4f81e177 // sdot v23.4s, v11.16b, v1.4b[0]\n"
- "ldr q11, [%[b_ptr0], #0x30]\n"
- ".word 0x4fa0e190 // sdot v16.4s, v12.16b, v0.4b[1]\n"
- ".word 0x4fa1e194 // sdot v20.4s, v12.16b, v1.4b[1]\n"
- "ldr q12, [%[b_ptr0], #0x40]\n"
- ".word 0x4fa0e1b1 // sdot v17.4s, v13.16b, v0.4b[1]\n"
- ".word 0x4fa1e1b5 // sdot v21.4s, v13.16b, v1.4b[1]\n"
- "ldr q13, [%[b_ptr0], #0x50]\n"
- ".word 0x4fa0e1d2 // sdot v18.4s, v14.16b, v0.4b[1]\n"
- ".word 0x4fa1e1d6 // sdot v22.4s, v14.16b, v1.4b[1]\n"
- "ldr q14, [%[b_ptr0], #0x60]\n"
- ".word 0x4fa0e1f3 // sdot v19.4s, v15.16b, v0.4b[1]\n"
- ".word 0x4fa1e1f7 // sdot v23.4s, v15.16b, v1.4b[1]\n"
- "ldr q15, [%[b_ptr0], #0x70]\n"
- ".word 0x4f80e910 // sdot v16.4s, v8.16b, v0.4b[2]\n"
- "add %[b_ptr0], %[b_ptr0], #0x100\n"
- ".word 0x4f81e914 // sdot v20.4s, v8.16b, v1.4b[2]\n"
- "ldr q8, [%[b_ptr0], #-0x80]\n"
- ".word 0x4f80e931 // sdot v17.4s, v9.16b, v0.4b[2]\n"
- ".word 0x4f81e935 // sdot v21.4s, v9.16b, v1.4b[2]\n"
- "ldr q9, [%[b_ptr0], #-0x70]\n"
- ".word 0x4f80e952 // sdot v18.4s, v10.16b, v0.4b[2]\n"
- ".word 0x4f81e956 // sdot v22.4s, v10.16b, v1.4b[2]\n"
- "ldr q10, [%[b_ptr0], #-0x60]\n"
- ".word 0x4f80e973 // sdot v19.4s, v11.16b, v0.4b[2]\n"
- ".word 0x4f81e977 // sdot v23.4s, v11.16b, v1.4b[2]\n"
- "ldr q11, [%[b_ptr0], #-0x50]\n"
- ".word 0x4fa0e990 // sdot v16.4s, v12.16b, v0.4b[3]\n"
- ".word 0x4fa1e994 // sdot v20.4s, v12.16b, v1.4b[3]\n"
- "ldr q12, [%[b_ptr0], #-0x40]\n"
- ".word 0x4fa0e9b1 // sdot v17.4s, v13.16b, v0.4b[3]\n"
- ".word 0x4fa1e9b5 // sdot v21.4s, v13.16b, v1.4b[3]\n"
- "ldr q13, [%[b_ptr0], #-0x30]\n"
- ".word 0x4fa0e9d2 // sdot v18.4s, v14.16b, v0.4b[3]\n"
- ".word 0x4fa1e9d6 // sdot v22.4s, v14.16b, v1.4b[3]\n"
- "ldr q14, [%[b_ptr0], #-0x20]\n"
- ".word 0x4fa0e9f3 // sdot v19.4s, v15.16b, v0.4b[3]\n"
- "ldr q0, [%[a_ptr0], #0x10]\n"
- ".word 0x4fa1e9f7 // sdot v23.4s, v15.16b, v1.4b[3]\n"
- "ldr q15, [%[b_ptr0], #-0x10]\n"
- ".word 0x4f84e110 // sdot v16.4s, v8.16b, v4.4b[0]\n"
- "ldr q1, [a_ptr1, #0x10]\n"
- ".word 0x4f85e114 // sdot v20.4s, v8.16b, v5.4b[0]\n"
- "ldr q8, [%[b_ptr0]]\n"
- ".word 0x4f84e131 // sdot v17.4s, v9.16b, v4.4b[0]\n"
- ".word 0x4f85e135 // sdot v21.4s, v9.16b, v5.4b[0]\n"
- "ldr q9, [%[b_ptr0], #0x10]\n"
- ".word 0x4f84e152 // sdot v18.4s, v10.16b, v4.4b[0]\n"
- ".word 0x4f85e156 // sdot v22.4s, v10.16b, v5.4b[0]\n"
- "ldr q10, [%[b_ptr0], #0x20]\n"
- ".word 0x4f84e173 // sdot v19.4s, v11.16b, v4.4b[0]\n"
- ".word 0x4f85e177 // sdot v23.4s, v11.16b, v5.4b[0]\n"
- "ldr q11, [%[b_ptr0], #0x30]\n"
- ".word 0x4fa4e190 // sdot v16.4s, v12.16b, v4.4b[1]\n"
- ".word 0x4fa5e194 // sdot v20.4s, v12.16b, v5.4b[1]\n"
- "ldr q12, [%[b_ptr0], #0x40]\n"
- ".word 0x4fa4e1b1 // sdot v17.4s, v13.16b, v4.4b[1]\n"
- ".word 0x4fa5e1b5 // sdot v21.4s, v13.16b, v5.4b[1]\n"
- "ldr q13, [%[b_ptr0], #0x50]\n"
- ".word 0x4fa4e1d2 // sdot v18.4s, v14.16b, v4.4b[1]\n"
- ".word 0x4fa5e1d6 // sdot v22.4s, v14.16b, v5.4b[1]\n"
- "ldr q14, [%[b_ptr0], #0x60]\n"
- ".word 0x4fa4e1f3 // sdot v19.4s, v15.16b, v4.4b[1]\n"
- ".word 0x4fa5e1f7 // sdot v23.4s, v15.16b, v5.4b[1]\n"
- "ldr q15, [%[b_ptr0], #0x70]\n"
- ".word 0x4f84e910 // sdot v16.4s, v8.16b, v4.4b[2]\n"
- ".word 0x4f85e914 // sdot v20.4s, v8.16b, v5.4b[2]\n"
- ".word 0x4f84e931 // sdot v17.4s, v9.16b, v4.4b[2]\n"
- ".word 0x4f85e935 // sdot v21.4s, v9.16b, v5.4b[2]\n"
- ".word 0x4f84e952 // sdot v18.4s, v10.16b, v4.4b[2]\n"
- ".word 0x4f85e956 // sdot v22.4s, v10.16b, v5.4b[2]\n"
- ".word 0x4f84e973 // sdot v19.4s, v11.16b, v4.4b[2]\n"
- ".word 0x4f85e977 // sdot v23.4s, v11.16b, v5.4b[2]\n"
- ".word 0x4fa4e990 // sdot v16.4s, v12.16b, v4.4b[3]\n"
- ".word 0x4fa5e994 // sdot v20.4s, v12.16b, v5.4b[3]\n"
- ".word 0x4fa4e9b1 // sdot v17.4s, v13.16b, v4.4b[3]\n"
- ".word 0x4fa5e9b5 // sdot v21.4s, v13.16b, v5.4b[3]\n"
- ".word 0x4fa4e9d2 // sdot v18.4s, v14.16b, v4.4b[3]\n"
- ".word 0x4fa5e9d6 // sdot v22.4s, v14.16b, v5.4b[3]\n"
- ".word 0x4fa4e9f3 // sdot v19.4s, v15.16b, v4.4b[3]\n"
- ".word 0x4fa5e9f7 // sdot v23.4s, v15.16b, v5.4b[3]\n"
- "b 5f\n"
- "4:\n"
- ".word 0x4f80e110 // sdot v16.4s, v8.16b, v0.4b[0]\n"
- "ldr q4, [%[a_ptr0]]\n"
- ".word 0x4f81e114 // sdot v20.4s, v8.16b, v1.4b[0]\n"
- "ldr q5, [a_ptr1]\n"
- ".word 0x4f80e131 // sdot v17.4s, v9.16b, v0.4b[0]\n"
- "ldr q8, [%[b_ptr0]]\n"
- ".word 0x4f81e135 // sdot v21.4s, v9.16b, v1.4b[0]\n"
- "ldr q9, [%[b_ptr0], #0x10]\n"
- ".word 0x4f80e152 // sdot v18.4s, v10.16b, v0.4b[0]\n"
- ".word 0x4f81e156 // sdot v22.4s, v10.16b, v1.4b[0]\n"
- "ldr q10, [%[b_ptr0], #0x20]\n"
- ".word 0x4f80e173 // sdot v19.4s, v11.16b, v0.4b[0]\n"
- ".word 0x4f81e177 // sdot v23.4s, v11.16b, v1.4b[0]\n"
- "ldr q11, [%[b_ptr0], #0x30]\n"
- ".word 0x4fa0e190 // sdot v16.4s, v12.16b, v0.4b[1]\n"
- ".word 0x4fa1e194 // sdot v20.4s, v12.16b, v1.4b[1]\n"
- "ldr q12, [%[b_ptr0], #0x40]\n"
- ".word 0x4fa0e1b1 // sdot v17.4s, v13.16b, v0.4b[1]\n"
- ".word 0x4fa1e1b5 // sdot v21.4s, v13.16b, v1.4b[1]\n"
- "ldr q13, [%[b_ptr0], #0x50]\n"
- ".word 0x4fa0e1d2 // sdot v18.4s, v14.16b, v0.4b[1]\n"
- ".word 0x4fa1e1d6 // sdot v22.4s, v14.16b, v1.4b[1]\n"
- "ldr q14, [%[b_ptr0], #0x60]\n"
- ".word 0x4fa0e1f3 // sdot v19.4s, v15.16b, v0.4b[1]\n"
- ".word 0x4fa1e1f7 // sdot v23.4s, v15.16b, v1.4b[1]\n"
- "ldr q15, [%[b_ptr0], #0x70]\n"
- ".word 0x4f80e910 // sdot v16.4s, v8.16b, v0.4b[2]\n"
- ".word 0x4f81e914 // sdot v20.4s, v8.16b, v1.4b[2]\n"
- ".word 0x4f80e931 // sdot v17.4s, v9.16b, v0.4b[2]\n"
- ".word 0x4f81e935 // sdot v21.4s, v9.16b, v1.4b[2]\n"
- ".word 0x4f80e952 // sdot v18.4s, v10.16b, v0.4b[2]\n"
- ".word 0x4f81e956 // sdot v22.4s, v10.16b, v1.4b[2]\n"
- ".word 0x4f80e973 // sdot v19.4s, v11.16b, v0.4b[2]\n"
- ".word 0x4f81e977 // sdot v23.4s, v11.16b, v1.4b[2]\n"
- ".word 0x4fa0e990 // sdot v16.4s, v12.16b, v0.4b[3]\n"
- ".word 0x4fa1e994 // sdot v20.4s, v12.16b, v1.4b[3]\n"
- ".word 0x4fa0e9b1 // sdot v17.4s, v13.16b, v0.4b[3]\n"
- ".word 0x4fa1e9b5 // sdot v21.4s, v13.16b, v1.4b[3]\n"
- ".word 0x4fa0e9d2 // sdot v18.4s, v14.16b, v0.4b[3]\n"
- ".word 0x4fa1e9d6 // sdot v22.4s, v14.16b, v1.4b[3]\n"
- ".word 0x4fa0e9f3 // sdot v19.4s, v15.16b, v0.4b[3]\n"
- ".word 0x4fa1e9f7 // sdot v23.4s, v15.16b, v1.4b[3]\n"
- "5:\n"
- "str q16, [%[c_ptr0]]\n"
- "str q17, [%[c_ptr0], #0x10]\n"
- "str q18, [%[c_ptr0], #0x20]\n"
- "str q19, [%[c_ptr0], #0x30]\n"
- "add %[c_ptr0], %[c_ptr0], #0x40\n"
- "str q20, [c_ptr1]\n"
- "str q21, [c_ptr1, #0x10]\n"
- "str q22, [c_ptr1, #0x20]\n"
- "str q23, [c_ptr1, #0x30]\n"
- ".unreq a_ptr1\n"
- ".unreq c_ptr1\n"
- : [a_ptr0] "+r" (a_ptr0), [b_ptr0] "+r" (b_ptr0), [c_ptr0] "+r" (c_ptr0), [loops] "+r" (loops), [regs] "+r" (regs)
- : [betaptr] "r" (betaptr), [width] "r" (width), [beta0] "r" (beta0), [lda] "r" (ldab), [ldc] "r" (ldcb)
- : "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x0", "x1", "cc", "memory"
+ "a_ptr1 .req X0\n"
+ "c_ptr1 .req X1\n"
+ "add a_ptr1, %[a_ptr0], %[lda]\n"
+ "add c_ptr1, %[c_ptr0], %[ldc]\n"
+ "cbz %[beta0], 1f\n"
+ "movi v16.4s, #0\n"
+ "ldr q0, [%[a_ptr0]]\n"
+ "movi v17.4s, #0\n"
+ "ldr q1, [a_ptr1]\n"
+ "movi v18.4s, #0\n"
+ "ldr q8, [%[b_ptr0]]\n"
+ "movi v19.4s, #0\n"
+ "ldr q9, [%[b_ptr0], #0x10]\n"
+ "movi v20.4s, #0\n"
+ "ldr q10, [%[b_ptr0], #0x20]\n"
+ "movi v21.4s, #0\n"
+ "ldr q11, [%[b_ptr0], #0x30]\n"
+ "movi v22.4s, #0\n"
+ "ldr q12, [%[b_ptr0], #0x40]\n"
+ "movi v23.4s, #0\n"
+ "ldr q13, [%[b_ptr0], #0x50]\n"
+ "ldr q14, [%[b_ptr0], #0x60]\n"
+ "add %[a_ptr0], %[a_ptr0], #0x10\n"
+ "add a_ptr1, a_ptr1, #0x10\n"
+ "add %[b_ptr0], %[b_ptr0], #0x80\n"
+ "cbz %[loops], 2f\n"
+ "b 3f\n"
+ "1:\n"
+ "ld1r {v15.4s}, [%[betaptr]]\n"
+ "ldr q16, [%[c_ptr0]]\n"
+ "ldr q17, [%[c_ptr0], #0x10]\n"
+ "ldr q18, [%[c_ptr0], #0x20]\n"
+ "ldr q19, [%[c_ptr0], #0x30]\n"
+ "mul v16.4s, v16.4s, v15.4s\n"
+ "ldr q20, [c_ptr1]\n"
+ "mul v17.4s, v17.4s, v15.4s\n"
+ "ldr q21, [c_ptr1, #0x10]\n"
+ "mul v18.4s, v18.4s, v15.4s\n"
+ "ldr q22, [c_ptr1, #0x20]\n"
+ "mul v19.4s, v19.4s, v15.4s\n"
+ "ldr q23, [c_ptr1, #0x30]\n"
+ "mul v20.4s, v20.4s, v15.4s\n"
+ "ldr q0, [%[a_ptr0]]\n"
+ "mul v21.4s, v21.4s, v15.4s\n"
+ "ldr q1, [a_ptr1]\n"
+ "mul v22.4s, v22.4s, v15.4s\n"
+ "ldr q8, [%[b_ptr0]]\n"
+ "mul v23.4s, v23.4s, v15.4s\n"
+ "ldr q9, [%[b_ptr0], #0x10]\n"
+ "ldr q10, [%[b_ptr0], #0x20]\n"
+ "add %[a_ptr0], %[a_ptr0], #0x10\n"
+ "ldr q11, [%[b_ptr0], #0x30]\n"
+ "add a_ptr1, a_ptr1, #0x10\n"
+ "ldr q12, [%[b_ptr0], #0x40]\n"
+ "ldr q13, [%[b_ptr0], #0x50]\n"
+ "ldr q14, [%[b_ptr0], #0x60]\n"
+ "add %[b_ptr0], %[b_ptr0], #0x80\n"
+ "cbz %[loops], 2f\n"
+ "3:\n"
+ ".word 0x4f80e110 // sdot v16.4s, v8.16b, v0.4b[0]\n"
+ "ldr q15, [%[b_ptr0], #-0x10]\n"
+ ".word 0x4f81e114 // sdot v20.4s, v8.16b, v1.4b[0]\n"
+ "ldr q4, [%[a_ptr0]]\n"
+ ".word 0x4f80e131 // sdot v17.4s, v9.16b, v0.4b[0]\n"
+ "ldr q5, [a_ptr1]\n"
+ ".word 0x4f81e135 // sdot v21.4s, v9.16b, v1.4b[0]\n"
+ "ldr q8, [%[b_ptr0]]\n"
+ ".word 0x4f80e152 // sdot v18.4s, v10.16b, v0.4b[0]\n"
+ "ldr q9, [%[b_ptr0], #0x10]\n"
+ ".word 0x4f81e156 // sdot v22.4s, v10.16b, v1.4b[0]\n"
+ "ldr q10, [%[b_ptr0], #0x20]\n"
+ ".word 0x4f80e173 // sdot v19.4s, v11.16b, v0.4b[0]\n"
+ "subs %[loops], %[loops], #0x1\n"
+ ".word 0x4f81e177 // sdot v23.4s, v11.16b, v1.4b[0]\n"
+ "ldr q11, [%[b_ptr0], #0x30]\n"
+ ".word 0x4fa0e190 // sdot v16.4s, v12.16b, v0.4b[1]\n"
+ "prfm PLDL1KEEP, [%[a_ptr0], #0x40]\n"
+ ".word 0x4fa1e194 // sdot v20.4s, v12.16b, v1.4b[1]\n"
+ "ldr q12, [%[b_ptr0], #0x40]\n"
+ ".word 0x4fa0e1b1 // sdot v17.4s, v13.16b, v0.4b[1]\n"
+ "add %[a_ptr0], %[a_ptr0], #0x20\n"
+ ".word 0x4fa1e1b5 // sdot v21.4s, v13.16b, v1.4b[1]\n"
+ "ldr q13, [%[b_ptr0], #0x50]\n"
+ ".word 0x4fa0e1d2 // sdot v18.4s, v14.16b, v0.4b[1]\n"
+ "add a_ptr1, a_ptr1, #0x20\n"
+ ".word 0x4fa1e1d6 // sdot v22.4s, v14.16b, v1.4b[1]\n"
+ "ldr q14, [%[b_ptr0], #0x60]\n"
+ ".word 0x4fa0e1f3 // sdot v19.4s, v15.16b, v0.4b[1]\n"
+ "prfm PLDL1KEEP, [a_ptr1, #0x40]\n"
+ ".word 0x4fa1e1f7 // sdot v23.4s, v15.16b, v1.4b[1]\n"
+ "ldr q15, [%[b_ptr0], #0x70]\n"
+ ".word 0x4f80e910 // sdot v16.4s, v8.16b, v0.4b[2]\n"
+ "add %[b_ptr0], %[b_ptr0], #0x100\n"
+ ".word 0x4f81e914 // sdot v20.4s, v8.16b, v1.4b[2]\n"
+ "ldr q8, [%[b_ptr0], #-0x80]\n"
+ ".word 0x4f80e931 // sdot v17.4s, v9.16b, v0.4b[2]\n"
+ ".word 0x4f81e935 // sdot v21.4s, v9.16b, v1.4b[2]\n"
+ "ldr q9, [%[b_ptr0], #-0x70]\n"
+ ".word 0x4f80e952 // sdot v18.4s, v10.16b, v0.4b[2]\n"
+ ".word 0x4f81e956 // sdot v22.4s, v10.16b, v1.4b[2]\n"
+ "ldr q10, [%[b_ptr0], #-0x60]\n"
+ ".word 0x4f80e973 // sdot v19.4s, v11.16b, v0.4b[2]\n"
+ ".word 0x4f81e977 // sdot v23.4s, v11.16b, v1.4b[2]\n"
+ "ldr q11, [%[b_ptr0], #-0x50]\n"
+ ".word 0x4fa0e990 // sdot v16.4s, v12.16b, v0.4b[3]\n"
+ ".word 0x4fa1e994 // sdot v20.4s, v12.16b, v1.4b[3]\n"
+ "ldr q12, [%[b_ptr0], #-0x40]\n"
+ ".word 0x4fa0e9b1 // sdot v17.4s, v13.16b, v0.4b[3]\n"
+ ".word 0x4fa1e9b5 // sdot v21.4s, v13.16b, v1.4b[3]\n"
+ "ldr q13, [%[b_ptr0], #-0x30]\n"
+ ".word 0x4fa0e9d2 // sdot v18.4s, v14.16b, v0.4b[3]\n"
+ ".word 0x4fa1e9d6 // sdot v22.4s, v14.16b, v1.4b[3]\n"
+ "ldr q14, [%[b_ptr0], #-0x20]\n"
+ ".word 0x4fa0e9f3 // sdot v19.4s, v15.16b, v0.4b[3]\n"
+ "ldr q0, [%[a_ptr0], #-0x10]\n"
+ ".word 0x4fa1e9f7 // sdot v23.4s, v15.16b, v1.4b[3]\n"
+ "ldr q15, [%[b_ptr0], #-0x10]\n"
+ ".word 0x4f84e110 // sdot v16.4s, v8.16b, v4.4b[0]\n"
+ "ldr q1, [a_ptr1, #-0x10]\n"
+ ".word 0x4f85e114 // sdot v20.4s, v8.16b, v5.4b[0]\n"
+ "ldr q8, [%[b_ptr0]]\n"
+ ".word 0x4f84e131 // sdot v17.4s, v9.16b, v4.4b[0]\n"
+ ".word 0x4f85e135 // sdot v21.4s, v9.16b, v5.4b[0]\n"
+ "ldr q9, [%[b_ptr0], #0x10]\n"
+ ".word 0x4f84e152 // sdot v18.4s, v10.16b, v4.4b[0]\n"
+ ".word 0x4f85e156 // sdot v22.4s, v10.16b, v5.4b[0]\n"
+ "ldr q10, [%[b_ptr0], #0x20]\n"
+ ".word 0x4f84e173 // sdot v19.4s, v11.16b, v4.4b[0]\n"
+ ".word 0x4f85e177 // sdot v23.4s, v11.16b, v5.4b[0]\n"
+ "ldr q11, [%[b_ptr0], #0x30]\n"
+ ".word 0x4fa4e190 // sdot v16.4s, v12.16b, v4.4b[1]\n"
+ ".word 0x4fa5e194 // sdot v20.4s, v12.16b, v5.4b[1]\n"
+ "ldr q12, [%[b_ptr0], #0x40]\n"
+ ".word 0x4fa4e1b1 // sdot v17.4s, v13.16b, v4.4b[1]\n"
+ ".word 0x4fa5e1b5 // sdot v21.4s, v13.16b, v5.4b[1]\n"
+ "ldr q13, [%[b_ptr0], #0x50]\n"
+ ".word 0x4fa4e1d2 // sdot v18.4s, v14.16b, v4.4b[1]\n"
+ ".word 0x4fa5e1d6 // sdot v22.4s, v14.16b, v5.4b[1]\n"
+ "ldr q14, [%[b_ptr0], #0x60]\n"
+ ".word 0x4fa4e1f3 // sdot v19.4s, v15.16b, v4.4b[1]\n"
+ ".word 0x4fa5e1f7 // sdot v23.4s, v15.16b, v5.4b[1]\n"
+ "ldr q15, [%[b_ptr0], #0x70]\n"
+ ".word 0x4f84e910 // sdot v16.4s, v8.16b, v4.4b[2]\n"
+ "add %[b_ptr0], %[b_ptr0], #0x100\n"
+ ".word 0x4f85e914 // sdot v20.4s, v8.16b, v5.4b[2]\n"
+ "ldr q8, [%[b_ptr0], #-0x80]\n"
+ ".word 0x4f84e931 // sdot v17.4s, v9.16b, v4.4b[2]\n"
+ ".word 0x4f85e935 // sdot v21.4s, v9.16b, v5.4b[2]\n"
+ "ldr q9, [%[b_ptr0], #-0x70]\n"
+ ".word 0x4f84e952 // sdot v18.4s, v10.16b, v4.4b[2]\n"
+ ".word 0x4f85e956 // sdot v22.4s, v10.16b, v5.4b[2]\n"
+ "ldr q10, [%[b_ptr0], #-0x60]\n"
+ ".word 0x4f84e973 // sdot v19.4s, v11.16b, v4.4b[2]\n"
+ ".word 0x4f85e977 // sdot v23.4s, v11.16b, v5.4b[2]\n"
+ "ldr q11, [%[b_ptr0], #-0x50]\n"
+ ".word 0x4fa4e990 // sdot v16.4s, v12.16b, v4.4b[3]\n"
+ ".word 0x4fa5e994 // sdot v20.4s, v12.16b, v5.4b[3]\n"
+ "ldr q12, [%[b_ptr0], #-0x40]\n"
+ ".word 0x4fa4e9b1 // sdot v17.4s, v13.16b, v4.4b[3]\n"
+ ".word 0x4fa5e9b5 // sdot v21.4s, v13.16b, v5.4b[3]\n"
+ "ldr q13, [%[b_ptr0], #-0x30]\n"
+ ".word 0x4fa4e9d2 // sdot v18.4s, v14.16b, v4.4b[3]\n"
+ ".word 0x4fa5e9d6 // sdot v22.4s, v14.16b, v5.4b[3]\n"
+ "ldr q14, [%[b_ptr0], #-0x20]\n"
+ ".word 0x4fa4e9f3 // sdot v19.4s, v15.16b, v4.4b[3]\n"
+ ".word 0x4fa5e9f7 // sdot v23.4s, v15.16b, v5.4b[3]\n"
+ "b.ne 3b\n"
+ "2:\n"
+ "ldr q15, [%[b_ptr0], #-0x10]\n"
+ "prfm PSTL1KEEP, [%[c_ptr0]]\n"
+ "prfm PSTL1KEEP, [c_ptr1]\n"
+ "cbz %[regs], 4f\n"
+ ".word 0x4f80e110 // sdot v16.4s, v8.16b, v0.4b[0]\n"
+ "ldr q4, [%[a_ptr0]]\n"
+ ".word 0x4f81e114 // sdot v20.4s, v8.16b, v1.4b[0]\n"
+ "ldr q5, [a_ptr1]\n"
+ ".word 0x4f80e131 // sdot v17.4s, v9.16b, v0.4b[0]\n"
+ "ldr q8, [%[b_ptr0]]\n"
+ ".word 0x4f81e135 // sdot v21.4s, v9.16b, v1.4b[0]\n"
+ "ldr q9, [%[b_ptr0], #0x10]\n"
+ ".word 0x4f80e152 // sdot v18.4s, v10.16b, v0.4b[0]\n"
+ "add %[a_ptr0], %[a_ptr0], #0x10\n"
+ ".word 0x4f81e156 // sdot v22.4s, v10.16b, v1.4b[0]\n"
+ "ldr q10, [%[b_ptr0], #0x20]\n"
+ ".word 0x4f80e173 // sdot v19.4s, v11.16b, v0.4b[0]\n"
+ "add a_ptr1, a_ptr1, #0x10\n"
+ ".word 0x4f81e177 // sdot v23.4s, v11.16b, v1.4b[0]\n"
+ "ldr q11, [%[b_ptr0], #0x30]\n"
+ ".word 0x4fa0e190 // sdot v16.4s, v12.16b, v0.4b[1]\n"
+ ".word 0x4fa1e194 // sdot v20.4s, v12.16b, v1.4b[1]\n"
+ "ldr q12, [%[b_ptr0], #0x40]\n"
+ ".word 0x4fa0e1b1 // sdot v17.4s, v13.16b, v0.4b[1]\n"
+ ".word 0x4fa1e1b5 // sdot v21.4s, v13.16b, v1.4b[1]\n"
+ "ldr q13, [%[b_ptr0], #0x50]\n"
+ ".word 0x4fa0e1d2 // sdot v18.4s, v14.16b, v0.4b[1]\n"
+ ".word 0x4fa1e1d6 // sdot v22.4s, v14.16b, v1.4b[1]\n"
+ "ldr q14, [%[b_ptr0], #0x60]\n"
+ ".word 0x4fa0e1f3 // sdot v19.4s, v15.16b, v0.4b[1]\n"
+ ".word 0x4fa1e1f7 // sdot v23.4s, v15.16b, v1.4b[1]\n"
+ "ldr q15, [%[b_ptr0], #0x70]\n"
+ ".word 0x4f80e910 // sdot v16.4s, v8.16b, v0.4b[2]\n"
+ "add %[b_ptr0], %[b_ptr0], #0x100\n"
+ ".word 0x4f81e914 // sdot v20.4s, v8.16b, v1.4b[2]\n"
+ "ldr q8, [%[b_ptr0], #-0x80]\n"
+ ".word 0x4f80e931 // sdot v17.4s, v9.16b, v0.4b[2]\n"
+ ".word 0x4f81e935 // sdot v21.4s, v9.16b, v1.4b[2]\n"
+ "ldr q9, [%[b_ptr0], #-0x70]\n"
+ ".word 0x4f80e952 // sdot v18.4s, v10.16b, v0.4b[2]\n"
+ ".word 0x4f81e956 // sdot v22.4s, v10.16b, v1.4b[2]\n"
+ "ldr q10, [%[b_ptr0], #-0x60]\n"
+ ".word 0x4f80e973 // sdot v19.4s, v11.16b, v0.4b[2]\n"
+ ".word 0x4f81e977 // sdot v23.4s, v11.16b, v1.4b[2]\n"
+ "ldr q11, [%[b_ptr0], #-0x50]\n"
+ ".word 0x4fa0e990 // sdot v16.4s, v12.16b, v0.4b[3]\n"
+ ".word 0x4fa1e994 // sdot v20.4s, v12.16b, v1.4b[3]\n"
+ "ldr q12, [%[b_ptr0], #-0x40]\n"
+ ".word 0x4fa0e9b1 // sdot v17.4s, v13.16b, v0.4b[3]\n"
+ ".word 0x4fa1e9b5 // sdot v21.4s, v13.16b, v1.4b[3]\n"
+ "ldr q13, [%[b_ptr0], #-0x30]\n"
+ ".word 0x4fa0e9d2 // sdot v18.4s, v14.16b, v0.4b[3]\n"
+ ".word 0x4fa1e9d6 // sdot v22.4s, v14.16b, v1.4b[3]\n"
+ "ldr q14, [%[b_ptr0], #-0x20]\n"
+ ".word 0x4fa0e9f3 // sdot v19.4s, v15.16b, v0.4b[3]\n"
+ ".word 0x4fa1e9f7 // sdot v23.4s, v15.16b, v1.4b[3]\n"
+ "ldr q15, [%[b_ptr0], #-0x10]\n"
+ ".word 0x4f84e110 // sdot v16.4s, v8.16b, v4.4b[0]\n"
+ ".word 0x4f85e114 // sdot v20.4s, v8.16b, v5.4b[0]\n"
+ "ldr q8, [%[b_ptr0]]\n"
+ ".word 0x4f84e131 // sdot v17.4s, v9.16b, v4.4b[0]\n"
+ ".word 0x4f85e135 // sdot v21.4s, v9.16b, v5.4b[0]\n"
+ "ldr q9, [%[b_ptr0], #0x10]\n"
+ ".word 0x4f84e152 // sdot v18.4s, v10.16b, v4.4b[0]\n"
+ ".word 0x4f85e156 // sdot v22.4s, v10.16b, v5.4b[0]\n"
+ "ldr q10, [%[b_ptr0], #0x20]\n"
+ ".word 0x4f84e173 // sdot v19.4s, v11.16b, v4.4b[0]\n"
+ ".word 0x4f85e177 // sdot v23.4s, v11.16b, v5.4b[0]\n"
+ "ldr q11, [%[b_ptr0], #0x30]\n"
+ ".word 0x4fa4e190 // sdot v16.4s, v12.16b, v4.4b[1]\n"
+ ".word 0x4fa5e194 // sdot v20.4s, v12.16b, v5.4b[1]\n"
+ "ldr q12, [%[b_ptr0], #0x40]\n"
+ ".word 0x4fa4e1b1 // sdot v17.4s, v13.16b, v4.4b[1]\n"
+ ".word 0x4fa5e1b5 // sdot v21.4s, v13.16b, v5.4b[1]\n"
+ "ldr q13, [%[b_ptr0], #0x50]\n"
+ ".word 0x4fa4e1d2 // sdot v18.4s, v14.16b, v4.4b[1]\n"
+ ".word 0x4fa5e1d6 // sdot v22.4s, v14.16b, v5.4b[1]\n"
+ "ldr q14, [%[b_ptr0], #0x60]\n"
+ ".word 0x4fa4e1f3 // sdot v19.4s, v15.16b, v4.4b[1]\n"
+ ".word 0x4fa5e1f7 // sdot v23.4s, v15.16b, v5.4b[1]\n"
+ "ldr q15, [%[b_ptr0], #0x70]\n"
+ ".word 0x4f84e910 // sdot v16.4s, v8.16b, v4.4b[2]\n"
+ "add %[b_ptr0], %[b_ptr0], #0x80\n"
+ ".word 0x4f85e914 // sdot v20.4s, v8.16b, v5.4b[2]\n"
+ ".word 0x4f84e931 // sdot v17.4s, v9.16b, v4.4b[2]\n"
+ ".word 0x4f85e935 // sdot v21.4s, v9.16b, v5.4b[2]\n"
+ ".word 0x4f84e952 // sdot v18.4s, v10.16b, v4.4b[2]\n"
+ ".word 0x4f85e956 // sdot v22.4s, v10.16b, v5.4b[2]\n"
+ ".word 0x4f84e973 // sdot v19.4s, v11.16b, v4.4b[2]\n"
+ ".word 0x4f85e977 // sdot v23.4s, v11.16b, v5.4b[2]\n"
+ ".word 0x4fa4e990 // sdot v16.4s, v12.16b, v4.4b[3]\n"
+ ".word 0x4fa5e994 // sdot v20.4s, v12.16b, v5.4b[3]\n"
+ ".word 0x4fa4e9b1 // sdot v17.4s, v13.16b, v4.4b[3]\n"
+ ".word 0x4fa5e9b5 // sdot v21.4s, v13.16b, v5.4b[3]\n"
+ ".word 0x4fa4e9d2 // sdot v18.4s, v14.16b, v4.4b[3]\n"
+ ".word 0x4fa5e9d6 // sdot v22.4s, v14.16b, v5.4b[3]\n"
+ ".word 0x4fa4e9f3 // sdot v19.4s, v15.16b, v4.4b[3]\n"
+ ".word 0x4fa5e9f7 // sdot v23.4s, v15.16b, v5.4b[3]\n"
+ "b 5f\n"
+ "4:\n"
+ ".word 0x4f80e110 // sdot v16.4s, v8.16b, v0.4b[0]\n"
+ ".word 0x4f81e114 // sdot v20.4s, v8.16b, v1.4b[0]\n"
+ "ldr q8, [%[b_ptr0]]\n"
+ ".word 0x4f80e131 // sdot v17.4s, v9.16b, v0.4b[0]\n"
+ ".word 0x4f81e135 // sdot v21.4s, v9.16b, v1.4b[0]\n"
+ "ldr q9, [%[b_ptr0], #0x10]\n"
+ ".word 0x4f80e152 // sdot v18.4s, v10.16b, v0.4b[0]\n"
+ ".word 0x4f81e156 // sdot v22.4s, v10.16b, v1.4b[0]\n"
+ "ldr q10, [%[b_ptr0], #0x20]\n"
+ ".word 0x4f80e173 // sdot v19.4s, v11.16b, v0.4b[0]\n"
+ ".word 0x4f81e177 // sdot v23.4s, v11.16b, v1.4b[0]\n"
+ "ldr q11, [%[b_ptr0], #0x30]\n"
+ ".word 0x4fa0e190 // sdot v16.4s, v12.16b, v0.4b[1]\n"
+ ".word 0x4fa1e194 // sdot v20.4s, v12.16b, v1.4b[1]\n"
+ "ldr q12, [%[b_ptr0], #0x40]\n"
+ ".word 0x4fa0e1b1 // sdot v17.4s, v13.16b, v0.4b[1]\n"
+ ".word 0x4fa1e1b5 // sdot v21.4s, v13.16b, v1.4b[1]\n"
+ "ldr q13, [%[b_ptr0], #0x50]\n"
+ ".word 0x4fa0e1d2 // sdot v18.4s, v14.16b, v0.4b[1]\n"
+ ".word 0x4fa1e1d6 // sdot v22.4s, v14.16b, v1.4b[1]\n"
+ "ldr q14, [%[b_ptr0], #0x60]\n"
+ ".word 0x4fa0e1f3 // sdot v19.4s, v15.16b, v0.4b[1]\n"
+ ".word 0x4fa1e1f7 // sdot v23.4s, v15.16b, v1.4b[1]\n"
+ "ldr q15, [%[b_ptr0], #0x70]\n"
+ ".word 0x4f80e910 // sdot v16.4s, v8.16b, v0.4b[2]\n"
+ "add %[b_ptr0], %[b_ptr0], #0x80\n"
+ ".word 0x4f81e914 // sdot v20.4s, v8.16b, v1.4b[2]\n"
+ ".word 0x4f80e931 // sdot v17.4s, v9.16b, v0.4b[2]\n"
+ ".word 0x4f81e935 // sdot v21.4s, v9.16b, v1.4b[2]\n"
+ ".word 0x4f80e952 // sdot v18.4s, v10.16b, v0.4b[2]\n"
+ ".word 0x4f81e956 // sdot v22.4s, v10.16b, v1.4b[2]\n"
+ ".word 0x4f80e973 // sdot v19.4s, v11.16b, v0.4b[2]\n"
+ ".word 0x4f81e977 // sdot v23.4s, v11.16b, v1.4b[2]\n"
+ ".word 0x4fa0e990 // sdot v16.4s, v12.16b, v0.4b[3]\n"
+ ".word 0x4fa1e994 // sdot v20.4s, v12.16b, v1.4b[3]\n"
+ ".word 0x4fa0e9b1 // sdot v17.4s, v13.16b, v0.4b[3]\n"
+ ".word 0x4fa1e9b5 // sdot v21.4s, v13.16b, v1.4b[3]\n"
+ ".word 0x4fa0e9d2 // sdot v18.4s, v14.16b, v0.4b[3]\n"
+ ".word 0x4fa1e9d6 // sdot v22.4s, v14.16b, v1.4b[3]\n"
+ ".word 0x4fa0e9f3 // sdot v19.4s, v15.16b, v0.4b[3]\n"
+ ".word 0x4fa1e9f7 // sdot v23.4s, v15.16b, v1.4b[3]\n"
+ "5:\n"
+ "cbz %[blocks], 6f\n"
+ "7:\n"
+ "ldr q8, [%[b_ptr0]]\n"
+ "subs %[blocks], %[blocks], #0x1\n"
+ "ldr q9, [%[b_ptr0], #0x10]\n"
+ "ldr s0, [%[a_ptr0]]\n"
+ "ldr q10, [%[b_ptr0], #0x20]\n"
+ "add %[a_ptr0], %[a_ptr0], #0x4\n"
+ "ldr q11, [%[b_ptr0], #0x30]\n"
+ "add %[b_ptr0], %[b_ptr0], #0x40\n"
+ ".word 0x4f80e110 // sdot v16.4s, v8.16b, v0.4b[0]\n"
+ "ldr s1, [a_ptr1]\n"
+ ".word 0x4f80e131 // sdot v17.4s, v9.16b, v0.4b[0]\n"
+ "add a_ptr1, a_ptr1, #0x4\n"
+ ".word 0x4f80e152 // sdot v18.4s, v10.16b, v0.4b[0]\n"
+ ".word 0x4f81e114 // sdot v20.4s, v8.16b, v1.4b[0]\n"
+ ".word 0x4f81e135 // sdot v21.4s, v9.16b, v1.4b[0]\n"
+ ".word 0x4f81e156 // sdot v22.4s, v10.16b, v1.4b[0]\n"
+ ".word 0x4f80e173 // sdot v19.4s, v11.16b, v0.4b[0]\n"
+ ".word 0x4f81e177 // sdot v23.4s, v11.16b, v1.4b[0]\n"
+ "b.ne 7b\n"
+ "6:\n"
+ "cbz %[odds], 8f\n"
+ "ld1 {v0.b}[0], [%[a_ptr0]], #1\n"
+ "ld1 {v1.b}[0], [a_ptr1], #1\n"
+ "subs %[odds], %[odds], #0x1\n"
+ "b.eq 9f\n"
+ "ld1 {v0.b}[1], [%[a_ptr0]], #1\n"
+ "ld1 {v1.b}[1], [a_ptr1], #1\n"
+ "subs %[odds], %[odds], #0x1\n"
+ "b.eq 9f\n"
+ "ld1 {v0.b}[2], [%[a_ptr0]]\n"
+ "ld1 {v1.b}[2], [a_ptr1]\n"
+ "9:\n"
+ "ldr q8, [%[b_ptr0]]\n"
+ "ldr q9, [%[b_ptr0], #0x10]\n"
+ "ldr q10, [%[b_ptr0], #0x20]\n"
+ "ldr q11, [%[b_ptr0], #0x30]\n"
+ ".word 0x4f80e110 // sdot v16.4s, v8.16b, v0.4b[0]\n"
+ ".word 0x4f81e114 // sdot v20.4s, v8.16b, v1.4b[0]\n"
+ ".word 0x4f80e131 // sdot v17.4s, v9.16b, v0.4b[0]\n"
+ ".word 0x4f81e135 // sdot v21.4s, v9.16b, v1.4b[0]\n"
+ ".word 0x4f80e152 // sdot v18.4s, v10.16b, v0.4b[0]\n"
+ ".word 0x4f81e156 // sdot v22.4s, v10.16b, v1.4b[0]\n"
+ ".word 0x4f80e173 // sdot v19.4s, v11.16b, v0.4b[0]\n"
+ ".word 0x4f81e177 // sdot v23.4s, v11.16b, v1.4b[0]\n"
+ "8:\n"
+ "str q16, [%[c_ptr0]]\n"
+ "str q17, [%[c_ptr0], #0x10]\n"
+ "str q18, [%[c_ptr0], #0x20]\n"
+ "str q19, [%[c_ptr0], #0x30]\n"
+ "add %[c_ptr0], %[c_ptr0], #0x40\n"
+ "str q20, [c_ptr1]\n"
+ "str q21, [c_ptr1, #0x10]\n"
+ "str q22, [c_ptr1, #0x20]\n"
+ "str q23, [c_ptr1, #0x30]\n"
+ ".unreq a_ptr1\n"
+ ".unreq c_ptr1\n"
+ : [a_ptr0] "+r" (a_ptr0), [b_ptr0] "+r" (b_ptr0), [c_ptr0] "+r" (c_ptr0), [loops] "+r" (loops), [regs] "+r" (regs), [blocks] "+r" (blocks), [odds] "+r" (odds)
+ : [betaptr] "r" (betaptr), [width] "r" (width), [beta0] "r" (beta0), [lda] "r" (ldab), [ldc] "r" (ldcb)
+ : "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x0", "x1", "cc", "memory"
);
break;
case 3:
__asm __volatile (
- "a_ptr1 .req X0\n"
- "a_ptr2 .req X1\n"
- "c_ptr1 .req X2\n"
- "c_ptr2 .req X3\n"
- "add a_ptr1, %[a_ptr0], %[lda]\n"
- "add c_ptr1, %[c_ptr0], %[ldc]\n"
- "add a_ptr2, a_ptr1, %[lda]\n"
- "add c_ptr2, c_ptr1, %[ldc]\n"
- "cbz %[beta0], 1f\n"
- "movi v16.4s, #0\n"
- "ldr q0, [%[a_ptr0]]\n"
- "movi v17.4s, #0\n"
- "ldr q1, [a_ptr1]\n"
- "movi v18.4s, #0\n"
- "ldr q2, [a_ptr2]\n"
- "movi v19.4s, #0\n"
- "ldr q8, [%[b_ptr0]]\n"
- "movi v20.4s, #0\n"
- "ldr q9, [%[b_ptr0], #0x10]\n"
- "movi v21.4s, #0\n"
- "ldr q10, [%[b_ptr0], #0x20]\n"
- "movi v22.4s, #0\n"
- "ldr q11, [%[b_ptr0], #0x30]\n"
- "movi v23.4s, #0\n"
- "ldr q12, [%[b_ptr0], #0x40]\n"
- "movi v24.4s, #0\n"
- "ldr q13, [%[b_ptr0], #0x50]\n"
- "movi v25.4s, #0\n"
- "ldr q14, [%[b_ptr0], #0x60]\n"
- "movi v26.4s, #0\n"
- "add %[a_ptr0], %[a_ptr0], #0x10\n"
- "movi v27.4s, #0\n"
- "add a_ptr1, a_ptr1, #0x10\n"
- "add a_ptr2, a_ptr2, #0x10\n"
- "add %[b_ptr0], %[b_ptr0], #0x80\n"
- "cbz %[loops], 2f\n"
- "b 3f\n"
- "1:\n"
- "ld1r {v15.4s}, [%[betaptr]]\n"
- "ldr q16, [%[c_ptr0]]\n"
- "ldr q17, [%[c_ptr0], #0x10]\n"
- "ldr q18, [%[c_ptr0], #0x20]\n"
- "ldr q19, [%[c_ptr0], #0x30]\n"
- "mul v16.4s, v16.4s, v15.4s\n"
- "ldr q20, [c_ptr1]\n"
- "mul v17.4s, v17.4s, v15.4s\n"
- "ldr q21, [c_ptr1, #0x10]\n"
- "mul v18.4s, v18.4s, v15.4s\n"
- "ldr q22, [c_ptr1, #0x20]\n"
- "mul v19.4s, v19.4s, v15.4s\n"
- "ldr q23, [c_ptr1, #0x30]\n"
- "mul v20.4s, v20.4s, v15.4s\n"
- "ldr q24, [c_ptr2]\n"
- "mul v21.4s, v21.4s, v15.4s\n"
- "ldr q25, [c_ptr2, #0x10]\n"
- "mul v22.4s, v22.4s, v15.4s\n"
- "ldr q26, [c_ptr2, #0x20]\n"
- "mul v23.4s, v23.4s, v15.4s\n"
- "ldr q27, [c_ptr2, #0x30]\n"
- "mul v24.4s, v24.4s, v15.4s\n"
- "ldr q0, [%[a_ptr0]]\n"
- "mul v25.4s, v25.4s, v15.4s\n"
- "ldr q1, [a_ptr1]\n"
- "mul v26.4s, v26.4s, v15.4s\n"
- "ldr q2, [a_ptr2]\n"
- "mul v27.4s, v27.4s, v15.4s\n"
- "ldr q8, [%[b_ptr0]]\n"
- "ldr q9, [%[b_ptr0], #0x10]\n"
- "add %[a_ptr0], %[a_ptr0], #0x10\n"
- "ldr q10, [%[b_ptr0], #0x20]\n"
- "add a_ptr1, a_ptr1, #0x10\n"
- "ldr q11, [%[b_ptr0], #0x30]\n"
- "add a_ptr2, a_ptr2, #0x10\n"
- "ldr q12, [%[b_ptr0], #0x40]\n"
- "ldr q13, [%[b_ptr0], #0x50]\n"
- "ldr q14, [%[b_ptr0], #0x60]\n"
- "add %[b_ptr0], %[b_ptr0], #0x80\n"
- "cbz %[loops], 2f\n"
- "3:\n"
- ".word 0x4f80e110 // sdot v16.4s, v8.16b, v0.4b[0]\n"
- "ldr q15, [%[b_ptr0], #-0x10]\n"
- ".word 0x4f81e114 // sdot v20.4s, v8.16b, v1.4b[0]\n"
- "ldr q4, [%[a_ptr0]]\n"
- ".word 0x4f82e118 // sdot v24.4s, v8.16b, v2.4b[0]\n"
- "ldr q5, [a_ptr1]\n"
- ".word 0x4f80e131 // sdot v17.4s, v9.16b, v0.4b[0]\n"
- "ldr q6, [a_ptr2]\n"
- ".word 0x4f81e135 // sdot v21.4s, v9.16b, v1.4b[0]\n"
- "ldr q8, [%[b_ptr0]]\n"
- ".word 0x4f82e139 // sdot v25.4s, v9.16b, v2.4b[0]\n"
- "ldr q9, [%[b_ptr0], #0x10]\n"
- ".word 0x4f80e152 // sdot v18.4s, v10.16b, v0.4b[0]\n"
- "subs %[loops], %[loops], #0x1\n"
- ".word 0x4f81e156 // sdot v22.4s, v10.16b, v1.4b[0]\n"
- "prfm PLDL1KEEP, [%[a_ptr0], #0x40]\n"
- ".word 0x4f82e15a // sdot v26.4s, v10.16b, v2.4b[0]\n"
- "ldr q10, [%[b_ptr0], #0x20]\n"
- ".word 0x4f80e173 // sdot v19.4s, v11.16b, v0.4b[0]\n"
- "add %[a_ptr0], %[a_ptr0], #0x20\n"
- ".word 0x4f81e177 // sdot v23.4s, v11.16b, v1.4b[0]\n"
- "add a_ptr1, a_ptr1, #0x20\n"
- ".word 0x4f82e17b // sdot v27.4s, v11.16b, v2.4b[0]\n"
- "ldr q11, [%[b_ptr0], #0x30]\n"
- ".word 0x4fa0e190 // sdot v16.4s, v12.16b, v0.4b[1]\n"
- "add a_ptr2, a_ptr2, #0x20\n"
- ".word 0x4fa1e194 // sdot v20.4s, v12.16b, v1.4b[1]\n"
- "prfm PLDL1KEEP, [a_ptr1, #0x40]\n"
- ".word 0x4fa2e198 // sdot v24.4s, v12.16b, v2.4b[1]\n"
- "ldr q12, [%[b_ptr0], #0x40]\n"
- ".word 0x4fa0e1b1 // sdot v17.4s, v13.16b, v0.4b[1]\n"
- "prfm PLDL1KEEP, [a_ptr2, #0x40]\n"
- ".word 0x4fa1e1b5 // sdot v21.4s, v13.16b, v1.4b[1]\n"
- ".word 0x4fa2e1b9 // sdot v25.4s, v13.16b, v2.4b[1]\n"
- "ldr q13, [%[b_ptr0], #0x50]\n"
- ".word 0x4fa0e1d2 // sdot v18.4s, v14.16b, v0.4b[1]\n"
- ".word 0x4fa1e1d6 // sdot v22.4s, v14.16b, v1.4b[1]\n"
- ".word 0x4fa2e1da // sdot v26.4s, v14.16b, v2.4b[1]\n"
- "ldr q14, [%[b_ptr0], #0x60]\n"
- ".word 0x4fa0e1f3 // sdot v19.4s, v15.16b, v0.4b[1]\n"
- ".word 0x4fa1e1f7 // sdot v23.4s, v15.16b, v1.4b[1]\n"
- ".word 0x4fa2e1fb // sdot v27.4s, v15.16b, v2.4b[1]\n"
- "ldr q15, [%[b_ptr0], #0x70]\n"
- ".word 0x4f80e910 // sdot v16.4s, v8.16b, v0.4b[2]\n"
- "add %[b_ptr0], %[b_ptr0], #0x100\n"
- ".word 0x4f81e914 // sdot v20.4s, v8.16b, v1.4b[2]\n"
- ".word 0x4f82e918 // sdot v24.4s, v8.16b, v2.4b[2]\n"
- "ldr q8, [%[b_ptr0], #-0x80]\n"
- ".word 0x4f80e931 // sdot v17.4s, v9.16b, v0.4b[2]\n"
- ".word 0x4f81e935 // sdot v21.4s, v9.16b, v1.4b[2]\n"
- ".word 0x4f82e939 // sdot v25.4s, v9.16b, v2.4b[2]\n"
- "ldr q9, [%[b_ptr0], #-0x70]\n"
- ".word 0x4f80e952 // sdot v18.4s, v10.16b, v0.4b[2]\n"
- ".word 0x4f81e956 // sdot v22.4s, v10.16b, v1.4b[2]\n"
- ".word 0x4f82e95a // sdot v26.4s, v10.16b, v2.4b[2]\n"
- "ldr q10, [%[b_ptr0], #-0x60]\n"
- ".word 0x4f80e973 // sdot v19.4s, v11.16b, v0.4b[2]\n"
- ".word 0x4f81e977 // sdot v23.4s, v11.16b, v1.4b[2]\n"
- ".word 0x4f82e97b // sdot v27.4s, v11.16b, v2.4b[2]\n"
- "ldr q11, [%[b_ptr0], #-0x50]\n"
- ".word 0x4fa0e990 // sdot v16.4s, v12.16b, v0.4b[3]\n"
- ".word 0x4fa1e994 // sdot v20.4s, v12.16b, v1.4b[3]\n"
- ".word 0x4fa2e998 // sdot v24.4s, v12.16b, v2.4b[3]\n"
- "ldr q12, [%[b_ptr0], #-0x40]\n"
- ".word 0x4fa0e9b1 // sdot v17.4s, v13.16b, v0.4b[3]\n"
- ".word 0x4fa1e9b5 // sdot v21.4s, v13.16b, v1.4b[3]\n"
- ".word 0x4fa2e9b9 // sdot v25.4s, v13.16b, v2.4b[3]\n"
- "ldr q13, [%[b_ptr0], #-0x30]\n"
- ".word 0x4fa0e9d2 // sdot v18.4s, v14.16b, v0.4b[3]\n"
- ".word 0x4fa1e9d6 // sdot v22.4s, v14.16b, v1.4b[3]\n"
- ".word 0x4fa2e9da // sdot v26.4s, v14.16b, v2.4b[3]\n"
- "ldr q14, [%[b_ptr0], #-0x20]\n"
- ".word 0x4fa0e9f3 // sdot v19.4s, v15.16b, v0.4b[3]\n"
- "ldr q0, [%[a_ptr0], #-0x10]\n"
- ".word 0x4fa1e9f7 // sdot v23.4s, v15.16b, v1.4b[3]\n"
- "ldr q1, [a_ptr1, #-0x10]\n"
- ".word 0x4fa2e9fb // sdot v27.4s, v15.16b, v2.4b[3]\n"
- "ldr q15, [%[b_ptr0], #-0x10]\n"
- ".word 0x4f84e110 // sdot v16.4s, v8.16b, v4.4b[0]\n"
- "ldr q2, [a_ptr2, #-0x10]\n"
- ".word 0x4f85e114 // sdot v20.4s, v8.16b, v5.4b[0]\n"
- ".word 0x4f86e118 // sdot v24.4s, v8.16b, v6.4b[0]\n"
- "ldr q8, [%[b_ptr0]]\n"
- ".word 0x4f84e131 // sdot v17.4s, v9.16b, v4.4b[0]\n"
- ".word 0x4f85e135 // sdot v21.4s, v9.16b, v5.4b[0]\n"
- ".word 0x4f86e139 // sdot v25.4s, v9.16b, v6.4b[0]\n"
- "ldr q9, [%[b_ptr0], #0x10]\n"
- ".word 0x4f84e152 // sdot v18.4s, v10.16b, v4.4b[0]\n"
- ".word 0x4f85e156 // sdot v22.4s, v10.16b, v5.4b[0]\n"
- ".word 0x4f86e15a // sdot v26.4s, v10.16b, v6.4b[0]\n"
- "ldr q10, [%[b_ptr0], #0x20]\n"
- ".word 0x4f84e173 // sdot v19.4s, v11.16b, v4.4b[0]\n"
- ".word 0x4f85e177 // sdot v23.4s, v11.16b, v5.4b[0]\n"
- ".word 0x4f86e17b // sdot v27.4s, v11.16b, v6.4b[0]\n"
- "ldr q11, [%[b_ptr0], #0x30]\n"
- ".word 0x4fa4e190 // sdot v16.4s, v12.16b, v4.4b[1]\n"
- ".word 0x4fa5e194 // sdot v20.4s, v12.16b, v5.4b[1]\n"
- ".word 0x4fa6e198 // sdot v24.4s, v12.16b, v6.4b[1]\n"
- "ldr q12, [%[b_ptr0], #0x40]\n"
- ".word 0x4fa4e1b1 // sdot v17.4s, v13.16b, v4.4b[1]\n"
- ".word 0x4fa5e1b5 // sdot v21.4s, v13.16b, v5.4b[1]\n"
- ".word 0x4fa6e1b9 // sdot v25.4s, v13.16b, v6.4b[1]\n"
- "ldr q13, [%[b_ptr0], #0x50]\n"
- ".word 0x4fa4e1d2 // sdot v18.4s, v14.16b, v4.4b[1]\n"
- ".word 0x4fa5e1d6 // sdot v22.4s, v14.16b, v5.4b[1]\n"
- ".word 0x4fa6e1da // sdot v26.4s, v14.16b, v6.4b[1]\n"
- "ldr q14, [%[b_ptr0], #0x60]\n"
- ".word 0x4fa4e1f3 // sdot v19.4s, v15.16b, v4.4b[1]\n"
- ".word 0x4fa5e1f7 // sdot v23.4s, v15.16b, v5.4b[1]\n"
- ".word 0x4fa6e1fb // sdot v27.4s, v15.16b, v6.4b[1]\n"
- "ldr q15, [%[b_ptr0], #0x70]\n"
- ".word 0x4f84e910 // sdot v16.4s, v8.16b, v4.4b[2]\n"
- "add %[b_ptr0], %[b_ptr0], #0x100\n"
- ".word 0x4f85e914 // sdot v20.4s, v8.16b, v5.4b[2]\n"
- ".word 0x4f86e918 // sdot v24.4s, v8.16b, v6.4b[2]\n"
- "ldr q8, [%[b_ptr0], #-0x80]\n"
- ".word 0x4f84e931 // sdot v17.4s, v9.16b, v4.4b[2]\n"
- ".word 0x4f85e935 // sdot v21.4s, v9.16b, v5.4b[2]\n"
- ".word 0x4f86e939 // sdot v25.4s, v9.16b, v6.4b[2]\n"
- "ldr q9, [%[b_ptr0], #-0x70]\n"
- ".word 0x4f84e952 // sdot v18.4s, v10.16b, v4.4b[2]\n"
- ".word 0x4f85e956 // sdot v22.4s, v10.16b, v5.4b[2]\n"
- ".word 0x4f86e95a // sdot v26.4s, v10.16b, v6.4b[2]\n"
- "ldr q10, [%[b_ptr0], #-0x60]\n"
- ".word 0x4f84e973 // sdot v19.4s, v11.16b, v4.4b[2]\n"
- ".word 0x4f85e977 // sdot v23.4s, v11.16b, v5.4b[2]\n"
- ".word 0x4f86e97b // sdot v27.4s, v11.16b, v6.4b[2]\n"
- "ldr q11, [%[b_ptr0], #-0x50]\n"
- ".word 0x4fa4e990 // sdot v16.4s, v12.16b, v4.4b[3]\n"
- ".word 0x4fa5e994 // sdot v20.4s, v12.16b, v5.4b[3]\n"
- ".word 0x4fa6e998 // sdot v24.4s, v12.16b, v6.4b[3]\n"
- "ldr q12, [%[b_ptr0], #-0x40]\n"
- ".word 0x4fa4e9b1 // sdot v17.4s, v13.16b, v4.4b[3]\n"
- ".word 0x4fa5e9b5 // sdot v21.4s, v13.16b, v5.4b[3]\n"
- ".word 0x4fa6e9b9 // sdot v25.4s, v13.16b, v6.4b[3]\n"
- "ldr q13, [%[b_ptr0], #-0x30]\n"
- ".word 0x4fa4e9d2 // sdot v18.4s, v14.16b, v4.4b[3]\n"
- ".word 0x4fa5e9d6 // sdot v22.4s, v14.16b, v5.4b[3]\n"
- ".word 0x4fa6e9da // sdot v26.4s, v14.16b, v6.4b[3]\n"
- "ldr q14, [%[b_ptr0], #-0x20]\n"
- ".word 0x4fa4e9f3 // sdot v19.4s, v15.16b, v4.4b[3]\n"
- ".word 0x4fa5e9f7 // sdot v23.4s, v15.16b, v5.4b[3]\n"
- ".word 0x4fa6e9fb // sdot v27.4s, v15.16b, v6.4b[3]\n"
- "b.ne 3b\n"
- "2:\n"
- "ldr q15, [%[b_ptr0], #-0x10]\n"
- "prfm PSTL1KEEP, [%[c_ptr0]]\n"
- "prfm PSTL1KEEP, [c_ptr1]\n"
- "prfm PSTL1KEEP, [c_ptr2]\n"
- "cbz %[regs], 4f\n"
- ".word 0x4f80e110 // sdot v16.4s, v8.16b, v0.4b[0]\n"
- "ldr q4, [%[a_ptr0]]\n"
- ".word 0x4f81e114 // sdot v20.4s, v8.16b, v1.4b[0]\n"
- "ldr q5, [a_ptr1]\n"
- ".word 0x4f82e118 // sdot v24.4s, v8.16b, v2.4b[0]\n"
- "ldr q6, [a_ptr2]\n"
- ".word 0x4f80e131 // sdot v17.4s, v9.16b, v0.4b[0]\n"
- "ldr q8, [%[b_ptr0]]\n"
- ".word 0x4f81e135 // sdot v21.4s, v9.16b, v1.4b[0]\n"
- ".word 0x4f82e139 // sdot v25.4s, v9.16b, v2.4b[0]\n"
- "ldr q9, [%[b_ptr0], #0x10]\n"
- ".word 0x4f80e152 // sdot v18.4s, v10.16b, v0.4b[0]\n"
- ".word 0x4f81e156 // sdot v22.4s, v10.16b, v1.4b[0]\n"
- ".word 0x4f82e15a // sdot v26.4s, v10.16b, v2.4b[0]\n"
- "ldr q10, [%[b_ptr0], #0x20]\n"
- ".word 0x4f80e173 // sdot v19.4s, v11.16b, v0.4b[0]\n"
- ".word 0x4f81e177 // sdot v23.4s, v11.16b, v1.4b[0]\n"
- ".word 0x4f82e17b // sdot v27.4s, v11.16b, v2.4b[0]\n"
- "ldr q11, [%[b_ptr0], #0x30]\n"
- ".word 0x4fa0e190 // sdot v16.4s, v12.16b, v0.4b[1]\n"
- ".word 0x4fa1e194 // sdot v20.4s, v12.16b, v1.4b[1]\n"
- ".word 0x4fa2e198 // sdot v24.4s, v12.16b, v2.4b[1]\n"
- "ldr q12, [%[b_ptr0], #0x40]\n"
- ".word 0x4fa0e1b1 // sdot v17.4s, v13.16b, v0.4b[1]\n"
- ".word 0x4fa1e1b5 // sdot v21.4s, v13.16b, v1.4b[1]\n"
- ".word 0x4fa2e1b9 // sdot v25.4s, v13.16b, v2.4b[1]\n"
- "ldr q13, [%[b_ptr0], #0x50]\n"
- ".word 0x4fa0e1d2 // sdot v18.4s, v14.16b, v0.4b[1]\n"
- ".word 0x4fa1e1d6 // sdot v22.4s, v14.16b, v1.4b[1]\n"
- ".word 0x4fa2e1da // sdot v26.4s, v14.16b, v2.4b[1]\n"
- "ldr q14, [%[b_ptr0], #0x60]\n"
- ".word 0x4fa0e1f3 // sdot v19.4s, v15.16b, v0.4b[1]\n"
- ".word 0x4fa1e1f7 // sdot v23.4s, v15.16b, v1.4b[1]\n"
- ".word 0x4fa2e1fb // sdot v27.4s, v15.16b, v2.4b[1]\n"
- "ldr q15, [%[b_ptr0], #0x70]\n"
- ".word 0x4f80e910 // sdot v16.4s, v8.16b, v0.4b[2]\n"
- "add %[b_ptr0], %[b_ptr0], #0x100\n"
- ".word 0x4f81e914 // sdot v20.4s, v8.16b, v1.4b[2]\n"
- ".word 0x4f82e918 // sdot v24.4s, v8.16b, v2.4b[2]\n"
- "ldr q8, [%[b_ptr0], #-0x80]\n"
- ".word 0x4f80e931 // sdot v17.4s, v9.16b, v0.4b[2]\n"
- ".word 0x4f81e935 // sdot v21.4s, v9.16b, v1.4b[2]\n"
- ".word 0x4f82e939 // sdot v25.4s, v9.16b, v2.4b[2]\n"
- "ldr q9, [%[b_ptr0], #-0x70]\n"
- ".word 0x4f80e952 // sdot v18.4s, v10.16b, v0.4b[2]\n"
- ".word 0x4f81e956 // sdot v22.4s, v10.16b, v1.4b[2]\n"
- ".word 0x4f82e95a // sdot v26.4s, v10.16b, v2.4b[2]\n"
- "ldr q10, [%[b_ptr0], #-0x60]\n"
- ".word 0x4f80e973 // sdot v19.4s, v11.16b, v0.4b[2]\n"
- ".word 0x4f81e977 // sdot v23.4s, v11.16b, v1.4b[2]\n"
- ".word 0x4f82e97b // sdot v27.4s, v11.16b, v2.4b[2]\n"
- "ldr q11, [%[b_ptr0], #-0x50]\n"
- ".word 0x4fa0e990 // sdot v16.4s, v12.16b, v0.4b[3]\n"
- ".word 0x4fa1e994 // sdot v20.4s, v12.16b, v1.4b[3]\n"
- ".word 0x4fa2e998 // sdot v24.4s, v12.16b, v2.4b[3]\n"
- "ldr q12, [%[b_ptr0], #-0x40]\n"
- ".word 0x4fa0e9b1 // sdot v17.4s, v13.16b, v0.4b[3]\n"
- ".word 0x4fa1e9b5 // sdot v21.4s, v13.16b, v1.4b[3]\n"
- ".word 0x4fa2e9b9 // sdot v25.4s, v13.16b, v2.4b[3]\n"
- "ldr q13, [%[b_ptr0], #-0x30]\n"
- ".word 0x4fa0e9d2 // sdot v18.4s, v14.16b, v0.4b[3]\n"
- ".word 0x4fa1e9d6 // sdot v22.4s, v14.16b, v1.4b[3]\n"
- ".word 0x4fa2e9da // sdot v26.4s, v14.16b, v2.4b[3]\n"
- "ldr q14, [%[b_ptr0], #-0x20]\n"
- ".word 0x4fa0e9f3 // sdot v19.4s, v15.16b, v0.4b[3]\n"
- "ldr q0, [%[a_ptr0], #0x10]\n"
- ".word 0x4fa1e9f7 // sdot v23.4s, v15.16b, v1.4b[3]\n"
- "ldr q1, [a_ptr1, #0x10]\n"
- ".word 0x4fa2e9fb // sdot v27.4s, v15.16b, v2.4b[3]\n"
- "ldr q15, [%[b_ptr0], #-0x10]\n"
- ".word 0x4f84e110 // sdot v16.4s, v8.16b, v4.4b[0]\n"
- "ldr q2, [a_ptr2, #0x10]\n"
- ".word 0x4f85e114 // sdot v20.4s, v8.16b, v5.4b[0]\n"
- ".word 0x4f86e118 // sdot v24.4s, v8.16b, v6.4b[0]\n"
- "ldr q8, [%[b_ptr0]]\n"
- ".word 0x4f84e131 // sdot v17.4s, v9.16b, v4.4b[0]\n"
- ".word 0x4f85e135 // sdot v21.4s, v9.16b, v5.4b[0]\n"
- ".word 0x4f86e139 // sdot v25.4s, v9.16b, v6.4b[0]\n"
- "ldr q9, [%[b_ptr0], #0x10]\n"
- ".word 0x4f84e152 // sdot v18.4s, v10.16b, v4.4b[0]\n"
- ".word 0x4f85e156 // sdot v22.4s, v10.16b, v5.4b[0]\n"
- ".word 0x4f86e15a // sdot v26.4s, v10.16b, v6.4b[0]\n"
- "ldr q10, [%[b_ptr0], #0x20]\n"
- ".word 0x4f84e173 // sdot v19.4s, v11.16b, v4.4b[0]\n"
- ".word 0x4f85e177 // sdot v23.4s, v11.16b, v5.4b[0]\n"
- ".word 0x4f86e17b // sdot v27.4s, v11.16b, v6.4b[0]\n"
- "ldr q11, [%[b_ptr0], #0x30]\n"
- ".word 0x4fa4e190 // sdot v16.4s, v12.16b, v4.4b[1]\n"
- ".word 0x4fa5e194 // sdot v20.4s, v12.16b, v5.4b[1]\n"
- ".word 0x4fa6e198 // sdot v24.4s, v12.16b, v6.4b[1]\n"
- "ldr q12, [%[b_ptr0], #0x40]\n"
- ".word 0x4fa4e1b1 // sdot v17.4s, v13.16b, v4.4b[1]\n"
- ".word 0x4fa5e1b5 // sdot v21.4s, v13.16b, v5.4b[1]\n"
- ".word 0x4fa6e1b9 // sdot v25.4s, v13.16b, v6.4b[1]\n"
- "ldr q13, [%[b_ptr0], #0x50]\n"
- ".word 0x4fa4e1d2 // sdot v18.4s, v14.16b, v4.4b[1]\n"
- ".word 0x4fa5e1d6 // sdot v22.4s, v14.16b, v5.4b[1]\n"
- ".word 0x4fa6e1da // sdot v26.4s, v14.16b, v6.4b[1]\n"
- "ldr q14, [%[b_ptr0], #0x60]\n"
- ".word 0x4fa4e1f3 // sdot v19.4s, v15.16b, v4.4b[1]\n"
- ".word 0x4fa5e1f7 // sdot v23.4s, v15.16b, v5.4b[1]\n"
- ".word 0x4fa6e1fb // sdot v27.4s, v15.16b, v6.4b[1]\n"
- "ldr q15, [%[b_ptr0], #0x70]\n"
- ".word 0x4f84e910 // sdot v16.4s, v8.16b, v4.4b[2]\n"
- ".word 0x4f85e914 // sdot v20.4s, v8.16b, v5.4b[2]\n"
- ".word 0x4f86e918 // sdot v24.4s, v8.16b, v6.4b[2]\n"
- ".word 0x4f84e931 // sdot v17.4s, v9.16b, v4.4b[2]\n"
- ".word 0x4f85e935 // sdot v21.4s, v9.16b, v5.4b[2]\n"
- ".word 0x4f86e939 // sdot v25.4s, v9.16b, v6.4b[2]\n"
- ".word 0x4f84e952 // sdot v18.4s, v10.16b, v4.4b[2]\n"
- ".word 0x4f85e956 // sdot v22.4s, v10.16b, v5.4b[2]\n"
- ".word 0x4f86e95a // sdot v26.4s, v10.16b, v6.4b[2]\n"
- ".word 0x4f84e973 // sdot v19.4s, v11.16b, v4.4b[2]\n"
- ".word 0x4f85e977 // sdot v23.4s, v11.16b, v5.4b[2]\n"
- ".word 0x4f86e97b // sdot v27.4s, v11.16b, v6.4b[2]\n"
- ".word 0x4fa4e990 // sdot v16.4s, v12.16b, v4.4b[3]\n"
- ".word 0x4fa5e994 // sdot v20.4s, v12.16b, v5.4b[3]\n"
- ".word 0x4fa6e998 // sdot v24.4s, v12.16b, v6.4b[3]\n"
- ".word 0x4fa4e9b1 // sdot v17.4s, v13.16b, v4.4b[3]\n"
- ".word 0x4fa5e9b5 // sdot v21.4s, v13.16b, v5.4b[3]\n"
- ".word 0x4fa6e9b9 // sdot v25.4s, v13.16b, v6.4b[3]\n"
- ".word 0x4fa4e9d2 // sdot v18.4s, v14.16b, v4.4b[3]\n"
- ".word 0x4fa5e9d6 // sdot v22.4s, v14.16b, v5.4b[3]\n"
- ".word 0x4fa6e9da // sdot v26.4s, v14.16b, v6.4b[3]\n"
- ".word 0x4fa4e9f3 // sdot v19.4s, v15.16b, v4.4b[3]\n"
- ".word 0x4fa5e9f7 // sdot v23.4s, v15.16b, v5.4b[3]\n"
- ".word 0x4fa6e9fb // sdot v27.4s, v15.16b, v6.4b[3]\n"
- "b 5f\n"
- "4:\n"
- ".word 0x4f80e110 // sdot v16.4s, v8.16b, v0.4b[0]\n"
- "ldr q4, [%[a_ptr0]]\n"
- ".word 0x4f81e114 // sdot v20.4s, v8.16b, v1.4b[0]\n"
- "ldr q5, [a_ptr1]\n"
- ".word 0x4f82e118 // sdot v24.4s, v8.16b, v2.4b[0]\n"
- "ldr q6, [a_ptr2]\n"
- ".word 0x4f80e131 // sdot v17.4s, v9.16b, v0.4b[0]\n"
- "ldr q8, [%[b_ptr0]]\n"
- ".word 0x4f81e135 // sdot v21.4s, v9.16b, v1.4b[0]\n"
- ".word 0x4f82e139 // sdot v25.4s, v9.16b, v2.4b[0]\n"
- "ldr q9, [%[b_ptr0], #0x10]\n"
- ".word 0x4f80e152 // sdot v18.4s, v10.16b, v0.4b[0]\n"
- ".word 0x4f81e156 // sdot v22.4s, v10.16b, v1.4b[0]\n"
- ".word 0x4f82e15a // sdot v26.4s, v10.16b, v2.4b[0]\n"
- "ldr q10, [%[b_ptr0], #0x20]\n"
- ".word 0x4f80e173 // sdot v19.4s, v11.16b, v0.4b[0]\n"
- ".word 0x4f81e177 // sdot v23.4s, v11.16b, v1.4b[0]\n"
- ".word 0x4f82e17b // sdot v27.4s, v11.16b, v2.4b[0]\n"
- "ldr q11, [%[b_ptr0], #0x30]\n"
- ".word 0x4fa0e190 // sdot v16.4s, v12.16b, v0.4b[1]\n"
- ".word 0x4fa1e194 // sdot v20.4s, v12.16b, v1.4b[1]\n"
- ".word 0x4fa2e198 // sdot v24.4s, v12.16b, v2.4b[1]\n"
- "ldr q12, [%[b_ptr0], #0x40]\n"
- ".word 0x4fa0e1b1 // sdot v17.4s, v13.16b, v0.4b[1]\n"
- ".word 0x4fa1e1b5 // sdot v21.4s, v13.16b, v1.4b[1]\n"
- ".word 0x4fa2e1b9 // sdot v25.4s, v13.16b, v2.4b[1]\n"
- "ldr q13, [%[b_ptr0], #0x50]\n"
- ".word 0x4fa0e1d2 // sdot v18.4s, v14.16b, v0.4b[1]\n"
- ".word 0x4fa1e1d6 // sdot v22.4s, v14.16b, v1.4b[1]\n"
- ".word 0x4fa2e1da // sdot v26.4s, v14.16b, v2.4b[1]\n"
- "ldr q14, [%[b_ptr0], #0x60]\n"
- ".word 0x4fa0e1f3 // sdot v19.4s, v15.16b, v0.4b[1]\n"
- ".word 0x4fa1e1f7 // sdot v23.4s, v15.16b, v1.4b[1]\n"
- ".word 0x4fa2e1fb // sdot v27.4s, v15.16b, v2.4b[1]\n"
- "ldr q15, [%[b_ptr0], #0x70]\n"
- ".word 0x4f80e910 // sdot v16.4s, v8.16b, v0.4b[2]\n"
- ".word 0x4f81e914 // sdot v20.4s, v8.16b, v1.4b[2]\n"
- ".word 0x4f82e918 // sdot v24.4s, v8.16b, v2.4b[2]\n"
- ".word 0x4f80e931 // sdot v17.4s, v9.16b, v0.4b[2]\n"
- ".word 0x4f81e935 // sdot v21.4s, v9.16b, v1.4b[2]\n"
- ".word 0x4f82e939 // sdot v25.4s, v9.16b, v2.4b[2]\n"
- ".word 0x4f80e952 // sdot v18.4s, v10.16b, v0.4b[2]\n"
- ".word 0x4f81e956 // sdot v22.4s, v10.16b, v1.4b[2]\n"
- ".word 0x4f82e95a // sdot v26.4s, v10.16b, v2.4b[2]\n"
- ".word 0x4f80e973 // sdot v19.4s, v11.16b, v0.4b[2]\n"
- ".word 0x4f81e977 // sdot v23.4s, v11.16b, v1.4b[2]\n"
- ".word 0x4f82e97b // sdot v27.4s, v11.16b, v2.4b[2]\n"
- ".word 0x4fa0e990 // sdot v16.4s, v12.16b, v0.4b[3]\n"
- ".word 0x4fa1e994 // sdot v20.4s, v12.16b, v1.4b[3]\n"
- ".word 0x4fa2e998 // sdot v24.4s, v12.16b, v2.4b[3]\n"
- ".word 0x4fa0e9b1 // sdot v17.4s, v13.16b, v0.4b[3]\n"
- ".word 0x4fa1e9b5 // sdot v21.4s, v13.16b, v1.4b[3]\n"
- ".word 0x4fa2e9b9 // sdot v25.4s, v13.16b, v2.4b[3]\n"
- ".word 0x4fa0e9d2 // sdot v18.4s, v14.16b, v0.4b[3]\n"
- ".word 0x4fa1e9d6 // sdot v22.4s, v14.16b, v1.4b[3]\n"
- ".word 0x4fa2e9da // sdot v26.4s, v14.16b, v2.4b[3]\n"
- ".word 0x4fa0e9f3 // sdot v19.4s, v15.16b, v0.4b[3]\n"
- ".word 0x4fa1e9f7 // sdot v23.4s, v15.16b, v1.4b[3]\n"
- ".word 0x4fa2e9fb // sdot v27.4s, v15.16b, v2.4b[3]\n"
- "5:\n"
- "str q16, [%[c_ptr0]]\n"
- "str q17, [%[c_ptr0], #0x10]\n"
- "str q18, [%[c_ptr0], #0x20]\n"
- "str q19, [%[c_ptr0], #0x30]\n"
- "add %[c_ptr0], %[c_ptr0], #0x40\n"
- "str q20, [c_ptr1]\n"
- "str q21, [c_ptr1, #0x10]\n"
- "str q22, [c_ptr1, #0x20]\n"
- "str q23, [c_ptr1, #0x30]\n"
- "str q24, [c_ptr2]\n"
- "str q25, [c_ptr2, #0x10]\n"
- "str q26, [c_ptr2, #0x20]\n"
- "str q27, [c_ptr2, #0x30]\n"
- ".unreq a_ptr1\n"
- ".unreq a_ptr2\n"
- ".unreq c_ptr1\n"
- ".unreq c_ptr2\n"
- : [a_ptr0] "+r" (a_ptr0), [b_ptr0] "+r" (b_ptr0), [c_ptr0] "+r" (c_ptr0), [loops] "+r" (loops), [regs] "+r" (regs)
- : [betaptr] "r" (betaptr), [width] "r" (width), [beta0] "r" (beta0), [lda] "r" (ldab), [ldc] "r" (ldcb)
- : "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x0", "x1", "x2", "x3", "cc", "memory"
+ "a_ptr1 .req X0\n"
+ "a_ptr2 .req X1\n"
+ "c_ptr1 .req X2\n"
+ "c_ptr2 .req X3\n"
+ "add a_ptr1, %[a_ptr0], %[lda]\n"
+ "add c_ptr1, %[c_ptr0], %[ldc]\n"
+ "add a_ptr2, a_ptr1, %[lda]\n"
+ "add c_ptr2, c_ptr1, %[ldc]\n"
+ "cbz %[beta0], 1f\n"
+ "movi v16.4s, #0\n"
+ "ldr q0, [%[a_ptr0]]\n"
+ "movi v17.4s, #0\n"
+ "ldr q1, [a_ptr1]\n"
+ "movi v18.4s, #0\n"
+ "ldr q2, [a_ptr2]\n"
+ "movi v19.4s, #0\n"
+ "ldr q8, [%[b_ptr0]]\n"
+ "movi v20.4s, #0\n"
+ "ldr q9, [%[b_ptr0], #0x10]\n"
+ "movi v21.4s, #0\n"
+ "ldr q10, [%[b_ptr0], #0x20]\n"
+ "movi v22.4s, #0\n"
+ "ldr q11, [%[b_ptr0], #0x30]\n"
+ "movi v23.4s, #0\n"
+ "ldr q12, [%[b_ptr0], #0x40]\n"
+ "movi v24.4s, #0\n"
+ "ldr q13, [%[b_ptr0], #0x50]\n"
+ "movi v25.4s, #0\n"
+ "ldr q14, [%[b_ptr0], #0x60]\n"
+ "movi v26.4s, #0\n"
+ "add %[a_ptr0], %[a_ptr0], #0x10\n"
+ "movi v27.4s, #0\n"
+ "add a_ptr1, a_ptr1, #0x10\n"
+ "add a_ptr2, a_ptr2, #0x10\n"
+ "add %[b_ptr0], %[b_ptr0], #0x80\n"
+ "cbz %[loops], 2f\n"
+ "b 3f\n"
+ "1:\n"
+ "ld1r {v15.4s}, [%[betaptr]]\n"
+ "ldr q16, [%[c_ptr0]]\n"
+ "ldr q17, [%[c_ptr0], #0x10]\n"
+ "ldr q18, [%[c_ptr0], #0x20]\n"
+ "ldr q19, [%[c_ptr0], #0x30]\n"
+ "mul v16.4s, v16.4s, v15.4s\n"
+ "ldr q20, [c_ptr1]\n"
+ "mul v17.4s, v17.4s, v15.4s\n"
+ "ldr q21, [c_ptr1, #0x10]\n"
+ "mul v18.4s, v18.4s, v15.4s\n"
+ "ldr q22, [c_ptr1, #0x20]\n"
+ "mul v19.4s, v19.4s, v15.4s\n"
+ "ldr q23, [c_ptr1, #0x30]\n"
+ "mul v20.4s, v20.4s, v15.4s\n"
+ "ldr q24, [c_ptr2]\n"
+ "mul v21.4s, v21.4s, v15.4s\n"
+ "ldr q25, [c_ptr2, #0x10]\n"
+ "mul v22.4s, v22.4s, v15.4s\n"
+ "ldr q26, [c_ptr2, #0x20]\n"
+ "mul v23.4s, v23.4s, v15.4s\n"
+ "ldr q27, [c_ptr2, #0x30]\n"
+ "mul v24.4s, v24.4s, v15.4s\n"
+ "ldr q0, [%[a_ptr0]]\n"
+ "mul v25.4s, v25.4s, v15.4s\n"
+ "ldr q1, [a_ptr1]\n"
+ "mul v26.4s, v26.4s, v15.4s\n"
+ "ldr q2, [a_ptr2]\n"
+ "mul v27.4s, v27.4s, v15.4s\n"
+ "ldr q8, [%[b_ptr0]]\n"
+ "ldr q9, [%[b_ptr0], #0x10]\n"
+ "add %[a_ptr0], %[a_ptr0], #0x10\n"
+ "ldr q10, [%[b_ptr0], #0x20]\n"
+ "add a_ptr1, a_ptr1, #0x10\n"
+ "ldr q11, [%[b_ptr0], #0x30]\n"
+ "add a_ptr2, a_ptr2, #0x10\n"
+ "ldr q12, [%[b_ptr0], #0x40]\n"
+ "ldr q13, [%[b_ptr0], #0x50]\n"
+ "ldr q14, [%[b_ptr0], #0x60]\n"
+ "add %[b_ptr0], %[b_ptr0], #0x80\n"
+ "cbz %[loops], 2f\n"
+ "3:\n"
+ ".word 0x4f80e110 // sdot v16.4s, v8.16b, v0.4b[0]\n"
+ "ldr q15, [%[b_ptr0], #-0x10]\n"
+ ".word 0x4f81e114 // sdot v20.4s, v8.16b, v1.4b[0]\n"
+ "ldr q4, [%[a_ptr0]]\n"
+ ".word 0x4f82e118 // sdot v24.4s, v8.16b, v2.4b[0]\n"
+ "ldr q5, [a_ptr1]\n"
+ ".word 0x4f80e131 // sdot v17.4s, v9.16b, v0.4b[0]\n"
+ "ldr q6, [a_ptr2]\n"
+ ".word 0x4f81e135 // sdot v21.4s, v9.16b, v1.4b[0]\n"
+ "ldr q8, [%[b_ptr0]]\n"
+ ".word 0x4f82e139 // sdot v25.4s, v9.16b, v2.4b[0]\n"
+ "ldr q9, [%[b_ptr0], #0x10]\n"
+ ".word 0x4f80e152 // sdot v18.4s, v10.16b, v0.4b[0]\n"
+ "subs %[loops], %[loops], #0x1\n"
+ ".word 0x4f81e156 // sdot v22.4s, v10.16b, v1.4b[0]\n"
+ "prfm PLDL1KEEP, [%[a_ptr0], #0x40]\n"
+ ".word 0x4f82e15a // sdot v26.4s, v10.16b, v2.4b[0]\n"
+ "ldr q10, [%[b_ptr0], #0x20]\n"
+ ".word 0x4f80e173 // sdot v19.4s, v11.16b, v0.4b[0]\n"
+ "add %[a_ptr0], %[a_ptr0], #0x20\n"
+ ".word 0x4f81e177 // sdot v23.4s, v11.16b, v1.4b[0]\n"
+ "add a_ptr1, a_ptr1, #0x20\n"
+ ".word 0x4f82e17b // sdot v27.4s, v11.16b, v2.4b[0]\n"
+ "ldr q11, [%[b_ptr0], #0x30]\n"
+ ".word 0x4fa0e190 // sdot v16.4s, v12.16b, v0.4b[1]\n"
+ "add a_ptr2, a_ptr2, #0x20\n"
+ ".word 0x4fa1e194 // sdot v20.4s, v12.16b, v1.4b[1]\n"
+ "prfm PLDL1KEEP, [a_ptr1, #0x40]\n"
+ ".word 0x4fa2e198 // sdot v24.4s, v12.16b, v2.4b[1]\n"
+ "ldr q12, [%[b_ptr0], #0x40]\n"
+ ".word 0x4fa0e1b1 // sdot v17.4s, v13.16b, v0.4b[1]\n"
+ "prfm PLDL1KEEP, [a_ptr2, #0x40]\n"
+ ".word 0x4fa1e1b5 // sdot v21.4s, v13.16b, v1.4b[1]\n"
+ ".word 0x4fa2e1b9 // sdot v25.4s, v13.16b, v2.4b[1]\n"
+ "ldr q13, [%[b_ptr0], #0x50]\n"
+ ".word 0x4fa0e1d2 // sdot v18.4s, v14.16b, v0.4b[1]\n"
+ ".word 0x4fa1e1d6 // sdot v22.4s, v14.16b, v1.4b[1]\n"
+ ".word 0x4fa2e1da // sdot v26.4s, v14.16b, v2.4b[1]\n"
+ "ldr q14, [%[b_ptr0], #0x60]\n"
+ ".word 0x4fa0e1f3 // sdot v19.4s, v15.16b, v0.4b[1]\n"
+ ".word 0x4fa1e1f7 // sdot v23.4s, v15.16b, v1.4b[1]\n"
+ ".word 0x4fa2e1fb // sdot v27.4s, v15.16b, v2.4b[1]\n"
+ "ldr q15, [%[b_ptr0], #0x70]\n"
+ ".word 0x4f80e910 // sdot v16.4s, v8.16b, v0.4b[2]\n"
+ "add %[b_ptr0], %[b_ptr0], #0x100\n"
+ ".word 0x4f81e914 // sdot v20.4s, v8.16b, v1.4b[2]\n"
+ ".word 0x4f82e918 // sdot v24.4s, v8.16b, v2.4b[2]\n"
+ "ldr q8, [%[b_ptr0], #-0x80]\n"
+ ".word 0x4f80e931 // sdot v17.4s, v9.16b, v0.4b[2]\n"
+ ".word 0x4f81e935 // sdot v21.4s, v9.16b, v1.4b[2]\n"
+ ".word 0x4f82e939 // sdot v25.4s, v9.16b, v2.4b[2]\n"
+ "ldr q9, [%[b_ptr0], #-0x70]\n"
+ ".word 0x4f80e952 // sdot v18.4s, v10.16b, v0.4b[2]\n"
+ ".word 0x4f81e956 // sdot v22.4s, v10.16b, v1.4b[2]\n"
+ ".word 0x4f82e95a // sdot v26.4s, v10.16b, v2.4b[2]\n"
+ "ldr q10, [%[b_ptr0], #-0x60]\n"
+ ".word 0x4f80e973 // sdot v19.4s, v11.16b, v0.4b[2]\n"
+ ".word 0x4f81e977 // sdot v23.4s, v11.16b, v1.4b[2]\n"
+ ".word 0x4f82e97b // sdot v27.4s, v11.16b, v2.4b[2]\n"
+ "ldr q11, [%[b_ptr0], #-0x50]\n"
+ ".word 0x4fa0e990 // sdot v16.4s, v12.16b, v0.4b[3]\n"
+ ".word 0x4fa1e994 // sdot v20.4s, v12.16b, v1.4b[3]\n"
+ ".word 0x4fa2e998 // sdot v24.4s, v12.16b, v2.4b[3]\n"
+ "ldr q12, [%[b_ptr0], #-0x40]\n"
+ ".word 0x4fa0e9b1 // sdot v17.4s, v13.16b, v0.4b[3]\n"
+ ".word 0x4fa1e9b5 // sdot v21.4s, v13.16b, v1.4b[3]\n"
+ ".word 0x4fa2e9b9 // sdot v25.4s, v13.16b, v2.4b[3]\n"
+ "ldr q13, [%[b_ptr0], #-0x30]\n"
+ ".word 0x4fa0e9d2 // sdot v18.4s, v14.16b, v0.4b[3]\n"
+ ".word 0x4fa1e9d6 // sdot v22.4s, v14.16b, v1.4b[3]\n"
+ ".word 0x4fa2e9da // sdot v26.4s, v14.16b, v2.4b[3]\n"
+ "ldr q14, [%[b_ptr0], #-0x20]\n"
+ ".word 0x4fa0e9f3 // sdot v19.4s, v15.16b, v0.4b[3]\n"
+ "ldr q0, [%[a_ptr0], #-0x10]\n"
+ ".word 0x4fa1e9f7 // sdot v23.4s, v15.16b, v1.4b[3]\n"
+ "ldr q1, [a_ptr1, #-0x10]\n"
+ ".word 0x4fa2e9fb // sdot v27.4s, v15.16b, v2.4b[3]\n"
+ "ldr q15, [%[b_ptr0], #-0x10]\n"
+ ".word 0x4f84e110 // sdot v16.4s, v8.16b, v4.4b[0]\n"
+ "ldr q2, [a_ptr2, #-0x10]\n"
+ ".word 0x4f85e114 // sdot v20.4s, v8.16b, v5.4b[0]\n"
+ ".word 0x4f86e118 // sdot v24.4s, v8.16b, v6.4b[0]\n"
+ "ldr q8, [%[b_ptr0]]\n"
+ ".word 0x4f84e131 // sdot v17.4s, v9.16b, v4.4b[0]\n"
+ ".word 0x4f85e135 // sdot v21.4s, v9.16b, v5.4b[0]\n"
+ ".word 0x4f86e139 // sdot v25.4s, v9.16b, v6.4b[0]\n"
+ "ldr q9, [%[b_ptr0], #0x10]\n"
+ ".word 0x4f84e152 // sdot v18.4s, v10.16b, v4.4b[0]\n"
+ ".word 0x4f85e156 // sdot v22.4s, v10.16b, v5.4b[0]\n"
+ ".word 0x4f86e15a // sdot v26.4s, v10.16b, v6.4b[0]\n"
+ "ldr q10, [%[b_ptr0], #0x20]\n"
+ ".word 0x4f84e173 // sdot v19.4s, v11.16b, v4.4b[0]\n"
+ ".word 0x4f85e177 // sdot v23.4s, v11.16b, v5.4b[0]\n"
+ ".word 0x4f86e17b // sdot v27.4s, v11.16b, v6.4b[0]\n"
+ "ldr q11, [%[b_ptr0], #0x30]\n"
+ ".word 0x4fa4e190 // sdot v16.4s, v12.16b, v4.4b[1]\n"
+ ".word 0x4fa5e194 // sdot v20.4s, v12.16b, v5.4b[1]\n"
+ ".word 0x4fa6e198 // sdot v24.4s, v12.16b, v6.4b[1]\n"
+ "ldr q12, [%[b_ptr0], #0x40]\n"
+ ".word 0x4fa4e1b1 // sdot v17.4s, v13.16b, v4.4b[1]\n"
+ ".word 0x4fa5e1b5 // sdot v21.4s, v13.16b, v5.4b[1]\n"
+ ".word 0x4fa6e1b9 // sdot v25.4s, v13.16b, v6.4b[1]\n"
+ "ldr q13, [%[b_ptr0], #0x50]\n"
+ ".word 0x4fa4e1d2 // sdot v18.4s, v14.16b, v4.4b[1]\n"
+ ".word 0x4fa5e1d6 // sdot v22.4s, v14.16b, v5.4b[1]\n"
+ ".word 0x4fa6e1da // sdot v26.4s, v14.16b, v6.4b[1]\n"
+ "ldr q14, [%[b_ptr0], #0x60]\n"
+ ".word 0x4fa4e1f3 // sdot v19.4s, v15.16b, v4.4b[1]\n"
+ ".word 0x4fa5e1f7 // sdot v23.4s, v15.16b, v5.4b[1]\n"
+ ".word 0x4fa6e1fb // sdot v27.4s, v15.16b, v6.4b[1]\n"
+ "ldr q15, [%[b_ptr0], #0x70]\n"
+ ".word 0x4f84e910 // sdot v16.4s, v8.16b, v4.4b[2]\n"
+ "add %[b_ptr0], %[b_ptr0], #0x100\n"
+ ".word 0x4f85e914 // sdot v20.4s, v8.16b, v5.4b[2]\n"
+ ".word 0x4f86e918 // sdot v24.4s, v8.16b, v6.4b[2]\n"
+ "ldr q8, [%[b_ptr0], #-0x80]\n"
+ ".word 0x4f84e931 // sdot v17.4s, v9.16b, v4.4b[2]\n"
+ ".word 0x4f85e935 // sdot v21.4s, v9.16b, v5.4b[2]\n"
+ ".word 0x4f86e939 // sdot v25.4s, v9.16b, v6.4b[2]\n"
+ "ldr q9, [%[b_ptr0], #-0x70]\n"
+ ".word 0x4f84e952 // sdot v18.4s, v10.16b, v4.4b[2]\n"
+ ".word 0x4f85e956 // sdot v22.4s, v10.16b, v5.4b[2]\n"
+ ".word 0x4f86e95a // sdot v26.4s, v10.16b, v6.4b[2]\n"
+ "ldr q10, [%[b_ptr0], #-0x60]\n"
+ ".word 0x4f84e973 // sdot v19.4s, v11.16b, v4.4b[2]\n"
+ ".word 0x4f85e977 // sdot v23.4s, v11.16b, v5.4b[2]\n"
+ ".word 0x4f86e97b // sdot v27.4s, v11.16b, v6.4b[2]\n"
+ "ldr q11, [%[b_ptr0], #-0x50]\n"
+ ".word 0x4fa4e990 // sdot v16.4s, v12.16b, v4.4b[3]\n"
+ ".word 0x4fa5e994 // sdot v20.4s, v12.16b, v5.4b[3]\n"
+ ".word 0x4fa6e998 // sdot v24.4s, v12.16b, v6.4b[3]\n"
+ "ldr q12, [%[b_ptr0], #-0x40]\n"
+ ".word 0x4fa4e9b1 // sdot v17.4s, v13.16b, v4.4b[3]\n"
+ ".word 0x4fa5e9b5 // sdot v21.4s, v13.16b, v5.4b[3]\n"
+ ".word 0x4fa6e9b9 // sdot v25.4s, v13.16b, v6.4b[3]\n"
+ "ldr q13, [%[b_ptr0], #-0x30]\n"
+ ".word 0x4fa4e9d2 // sdot v18.4s, v14.16b, v4.4b[3]\n"
+ ".word 0x4fa5e9d6 // sdot v22.4s, v14.16b, v5.4b[3]\n"
+ ".word 0x4fa6e9da // sdot v26.4s, v14.16b, v6.4b[3]\n"
+ "ldr q14, [%[b_ptr0], #-0x20]\n"
+ ".word 0x4fa4e9f3 // sdot v19.4s, v15.16b, v4.4b[3]\n"
+ ".word 0x4fa5e9f7 // sdot v23.4s, v15.16b, v5.4b[3]\n"
+ ".word 0x4fa6e9fb // sdot v27.4s, v15.16b, v6.4b[3]\n"
+ "b.ne 3b\n"
+ "2:\n"
+ "ldr q15, [%[b_ptr0], #-0x10]\n"
+ "prfm PSTL1KEEP, [%[c_ptr0]]\n"
+ "prfm PSTL1KEEP, [c_ptr1]\n"
+ "prfm PSTL1KEEP, [c_ptr2]\n"
+ "cbz %[regs], 4f\n"
+ ".word 0x4f80e110 // sdot v16.4s, v8.16b, v0.4b[0]\n"
+ "ldr q4, [%[a_ptr0]]\n"
+ ".word 0x4f81e114 // sdot v20.4s, v8.16b, v1.4b[0]\n"
+ "ldr q5, [a_ptr1]\n"
+ ".word 0x4f82e118 // sdot v24.4s, v8.16b, v2.4b[0]\n"
+ "ldr q6, [a_ptr2]\n"
+ ".word 0x4f80e131 // sdot v17.4s, v9.16b, v0.4b[0]\n"
+ "ldr q8, [%[b_ptr0]]\n"
+ ".word 0x4f81e135 // sdot v21.4s, v9.16b, v1.4b[0]\n"
+ "add %[a_ptr0], %[a_ptr0], #0x10\n"
+ ".word 0x4f82e139 // sdot v25.4s, v9.16b, v2.4b[0]\n"
+ "ldr q9, [%[b_ptr0], #0x10]\n"
+ ".word 0x4f80e152 // sdot v18.4s, v10.16b, v0.4b[0]\n"
+ "add a_ptr1, a_ptr1, #0x10\n"
+ ".word 0x4f81e156 // sdot v22.4s, v10.16b, v1.4b[0]\n"
+ "add a_ptr2, a_ptr2, #0x10\n"
+ ".word 0x4f82e15a // sdot v26.4s, v10.16b, v2.4b[0]\n"
+ "ldr q10, [%[b_ptr0], #0x20]\n"
+ ".word 0x4f80e173 // sdot v19.4s, v11.16b, v0.4b[0]\n"
+ ".word 0x4f81e177 // sdot v23.4s, v11.16b, v1.4b[0]\n"
+ ".word 0x4f82e17b // sdot v27.4s, v11.16b, v2.4b[0]\n"
+ "ldr q11, [%[b_ptr0], #0x30]\n"
+ ".word 0x4fa0e190 // sdot v16.4s, v12.16b, v0.4b[1]\n"
+ ".word 0x4fa1e194 // sdot v20.4s, v12.16b, v1.4b[1]\n"
+ ".word 0x4fa2e198 // sdot v24.4s, v12.16b, v2.4b[1]\n"
+ "ldr q12, [%[b_ptr0], #0x40]\n"
+ ".word 0x4fa0e1b1 // sdot v17.4s, v13.16b, v0.4b[1]\n"
+ ".word 0x4fa1e1b5 // sdot v21.4s, v13.16b, v1.4b[1]\n"
+ ".word 0x4fa2e1b9 // sdot v25.4s, v13.16b, v2.4b[1]\n"
+ "ldr q13, [%[b_ptr0], #0x50]\n"
+ ".word 0x4fa0e1d2 // sdot v18.4s, v14.16b, v0.4b[1]\n"
+ ".word 0x4fa1e1d6 // sdot v22.4s, v14.16b, v1.4b[1]\n"
+ ".word 0x4fa2e1da // sdot v26.4s, v14.16b, v2.4b[1]\n"
+ "ldr q14, [%[b_ptr0], #0x60]\n"
+ ".word 0x4fa0e1f3 // sdot v19.4s, v15.16b, v0.4b[1]\n"
+ ".word 0x4fa1e1f7 // sdot v23.4s, v15.16b, v1.4b[1]\n"
+ ".word 0x4fa2e1fb // sdot v27.4s, v15.16b, v2.4b[1]\n"
+ "ldr q15, [%[b_ptr0], #0x70]\n"
+ ".word 0x4f80e910 // sdot v16.4s, v8.16b, v0.4b[2]\n"
+ "add %[b_ptr0], %[b_ptr0], #0x100\n"
+ ".word 0x4f81e914 // sdot v20.4s, v8.16b, v1.4b[2]\n"
+ ".word 0x4f82e918 // sdot v24.4s, v8.16b, v2.4b[2]\n"
+ "ldr q8, [%[b_ptr0], #-0x80]\n"
+ ".word 0x4f80e931 // sdot v17.4s, v9.16b, v0.4b[2]\n"
+ ".word 0x4f81e935 // sdot v21.4s, v9.16b, v1.4b[2]\n"
+ ".word 0x4f82e939 // sdot v25.4s, v9.16b, v2.4b[2]\n"
+ "ldr q9, [%[b_ptr0], #-0x70]\n"
+ ".word 0x4f80e952 // sdot v18.4s, v10.16b, v0.4b[2]\n"
+ ".word 0x4f81e956 // sdot v22.4s, v10.16b, v1.4b[2]\n"
+ ".word 0x4f82e95a // sdot v26.4s, v10.16b, v2.4b[2]\n"
+ "ldr q10, [%[b_ptr0], #-0x60]\n"
+ ".word 0x4f80e973 // sdot v19.4s, v11.16b, v0.4b[2]\n"
+ ".word 0x4f81e977 // sdot v23.4s, v11.16b, v1.4b[2]\n"
+ ".word 0x4f82e97b // sdot v27.4s, v11.16b, v2.4b[2]\n"
+ "ldr q11, [%[b_ptr0], #-0x50]\n"
+ ".word 0x4fa0e990 // sdot v16.4s, v12.16b, v0.4b[3]\n"
+ ".word 0x4fa1e994 // sdot v20.4s, v12.16b, v1.4b[3]\n"
+ ".word 0x4fa2e998 // sdot v24.4s, v12.16b, v2.4b[3]\n"
+ "ldr q12, [%[b_ptr0], #-0x40]\n"
+ ".word 0x4fa0e9b1 // sdot v17.4s, v13.16b, v0.4b[3]\n"
+ ".word 0x4fa1e9b5 // sdot v21.4s, v13.16b, v1.4b[3]\n"
+ ".word 0x4fa2e9b9 // sdot v25.4s, v13.16b, v2.4b[3]\n"
+ "ldr q13, [%[b_ptr0], #-0x30]\n"
+ ".word 0x4fa0e9d2 // sdot v18.4s, v14.16b, v0.4b[3]\n"
+ ".word 0x4fa1e9d6 // sdot v22.4s, v14.16b, v1.4b[3]\n"
+ ".word 0x4fa2e9da // sdot v26.4s, v14.16b, v2.4b[3]\n"
+ "ldr q14, [%[b_ptr0], #-0x20]\n"
+ ".word 0x4fa0e9f3 // sdot v19.4s, v15.16b, v0.4b[3]\n"
+ ".word 0x4fa1e9f7 // sdot v23.4s, v15.16b, v1.4b[3]\n"
+ ".word 0x4fa2e9fb // sdot v27.4s, v15.16b, v2.4b[3]\n"
+ "ldr q15, [%[b_ptr0], #-0x10]\n"
+ ".word 0x4f84e110 // sdot v16.4s, v8.16b, v4.4b[0]\n"
+ ".word 0x4f85e114 // sdot v20.4s, v8.16b, v5.4b[0]\n"
+ ".word 0x4f86e118 // sdot v24.4s, v8.16b, v6.4b[0]\n"
+ "ldr q8, [%[b_ptr0]]\n"
+ ".word 0x4f84e131 // sdot v17.4s, v9.16b, v4.4b[0]\n"
+ ".word 0x4f85e135 // sdot v21.4s, v9.16b, v5.4b[0]\n"
+ ".word 0x4f86e139 // sdot v25.4s, v9.16b, v6.4b[0]\n"
+ "ldr q9, [%[b_ptr0], #0x10]\n"
+ ".word 0x4f84e152 // sdot v18.4s, v10.16b, v4.4b[0]\n"
+ ".word 0x4f85e156 // sdot v22.4s, v10.16b, v5.4b[0]\n"
+ ".word 0x4f86e15a // sdot v26.4s, v10.16b, v6.4b[0]\n"
+ "ldr q10, [%[b_ptr0], #0x20]\n"
+ ".word 0x4f84e173 // sdot v19.4s, v11.16b, v4.4b[0]\n"
+ ".word 0x4f85e177 // sdot v23.4s, v11.16b, v5.4b[0]\n"
+ ".word 0x4f86e17b // sdot v27.4s, v11.16b, v6.4b[0]\n"
+ "ldr q11, [%[b_ptr0], #0x30]\n"
+ ".word 0x4fa4e190 // sdot v16.4s, v12.16b, v4.4b[1]\n"
+ ".word 0x4fa5e194 // sdot v20.4s, v12.16b, v5.4b[1]\n"
+ ".word 0x4fa6e198 // sdot v24.4s, v12.16b, v6.4b[1]\n"
+ "ldr q12, [%[b_ptr0], #0x40]\n"
+ ".word 0x4fa4e1b1 // sdot v17.4s, v13.16b, v4.4b[1]\n"
+ ".word 0x4fa5e1b5 // sdot v21.4s, v13.16b, v5.4b[1]\n"
+ ".word 0x4fa6e1b9 // sdot v25.4s, v13.16b, v6.4b[1]\n"
+ "ldr q13, [%[b_ptr0], #0x50]\n"
+ ".word 0x4fa4e1d2 // sdot v18.4s, v14.16b, v4.4b[1]\n"
+ ".word 0x4fa5e1d6 // sdot v22.4s, v14.16b, v5.4b[1]\n"
+ ".word 0x4fa6e1da // sdot v26.4s, v14.16b, v6.4b[1]\n"
+ "ldr q14, [%[b_ptr0], #0x60]\n"
+ ".word 0x4fa4e1f3 // sdot v19.4s, v15.16b, v4.4b[1]\n"
+ ".word 0x4fa5e1f7 // sdot v23.4s, v15.16b, v5.4b[1]\n"
+ ".word 0x4fa6e1fb // sdot v27.4s, v15.16b, v6.4b[1]\n"
+ "ldr q15, [%[b_ptr0], #0x70]\n"
+ ".word 0x4f84e910 // sdot v16.4s, v8.16b, v4.4b[2]\n"
+ "add %[b_ptr0], %[b_ptr0], #0x80\n"
+ ".word 0x4f85e914 // sdot v20.4s, v8.16b, v5.4b[2]\n"
+ ".word 0x4f86e918 // sdot v24.4s, v8.16b, v6.4b[2]\n"
+ ".word 0x4f84e931 // sdot v17.4s, v9.16b, v4.4b[2]\n"
+ ".word 0x4f85e935 // sdot v21.4s, v9.16b, v5.4b[2]\n"
+ ".word 0x4f86e939 // sdot v25.4s, v9.16b, v6.4b[2]\n"
+ ".word 0x4f84e952 // sdot v18.4s, v10.16b, v4.4b[2]\n"
+ ".word 0x4f85e956 // sdot v22.4s, v10.16b, v5.4b[2]\n"
+ ".word 0x4f86e95a // sdot v26.4s, v10.16b, v6.4b[2]\n"
+ ".word 0x4f84e973 // sdot v19.4s, v11.16b, v4.4b[2]\n"
+ ".word 0x4f85e977 // sdot v23.4s, v11.16b, v5.4b[2]\n"
+ ".word 0x4f86e97b // sdot v27.4s, v11.16b, v6.4b[2]\n"
+ ".word 0x4fa4e990 // sdot v16.4s, v12.16b, v4.4b[3]\n"
+ ".word 0x4fa5e994 // sdot v20.4s, v12.16b, v5.4b[3]\n"
+ ".word 0x4fa6e998 // sdot v24.4s, v12.16b, v6.4b[3]\n"
+ ".word 0x4fa4e9b1 // sdot v17.4s, v13.16b, v4.4b[3]\n"
+ ".word 0x4fa5e9b5 // sdot v21.4s, v13.16b, v5.4b[3]\n"
+ ".word 0x4fa6e9b9 // sdot v25.4s, v13.16b, v6.4b[3]\n"
+ ".word 0x4fa4e9d2 // sdot v18.4s, v14.16b, v4.4b[3]\n"
+ ".word 0x4fa5e9d6 // sdot v22.4s, v14.16b, v5.4b[3]\n"
+ ".word 0x4fa6e9da // sdot v26.4s, v14.16b, v6.4b[3]\n"
+ ".word 0x4fa4e9f3 // sdot v19.4s, v15.16b, v4.4b[3]\n"
+ ".word 0x4fa5e9f7 // sdot v23.4s, v15.16b, v5.4b[3]\n"
+ ".word 0x4fa6e9fb // sdot v27.4s, v15.16b, v6.4b[3]\n"
+ "b 5f\n"
+ "4:\n"
+ ".word 0x4f80e110 // sdot v16.4s, v8.16b, v0.4b[0]\n"
+ ".word 0x4f81e114 // sdot v20.4s, v8.16b, v1.4b[0]\n"
+ ".word 0x4f82e118 // sdot v24.4s, v8.16b, v2.4b[0]\n"
+ "ldr q8, [%[b_ptr0]]\n"
+ ".word 0x4f80e131 // sdot v17.4s, v9.16b, v0.4b[0]\n"
+ ".word 0x4f81e135 // sdot v21.4s, v9.16b, v1.4b[0]\n"
+ ".word 0x4f82e139 // sdot v25.4s, v9.16b, v2.4b[0]\n"
+ "ldr q9, [%[b_ptr0], #0x10]\n"
+ ".word 0x4f80e152 // sdot v18.4s, v10.16b, v0.4b[0]\n"
+ ".word 0x4f81e156 // sdot v22.4s, v10.16b, v1.4b[0]\n"
+ ".word 0x4f82e15a // sdot v26.4s, v10.16b, v2.4b[0]\n"
+ "ldr q10, [%[b_ptr0], #0x20]\n"
+ ".word 0x4f80e173 // sdot v19.4s, v11.16b, v0.4b[0]\n"
+ ".word 0x4f81e177 // sdot v23.4s, v11.16b, v1.4b[0]\n"
+ ".word 0x4f82e17b // sdot v27.4s, v11.16b, v2.4b[0]\n"
+ "ldr q11, [%[b_ptr0], #0x30]\n"
+ ".word 0x4fa0e190 // sdot v16.4s, v12.16b, v0.4b[1]\n"
+ ".word 0x4fa1e194 // sdot v20.4s, v12.16b, v1.4b[1]\n"
+ ".word 0x4fa2e198 // sdot v24.4s, v12.16b, v2.4b[1]\n"
+ "ldr q12, [%[b_ptr0], #0x40]\n"
+ ".word 0x4fa0e1b1 // sdot v17.4s, v13.16b, v0.4b[1]\n"
+ ".word 0x4fa1e1b5 // sdot v21.4s, v13.16b, v1.4b[1]\n"
+ ".word 0x4fa2e1b9 // sdot v25.4s, v13.16b, v2.4b[1]\n"
+ "ldr q13, [%[b_ptr0], #0x50]\n"
+ ".word 0x4fa0e1d2 // sdot v18.4s, v14.16b, v0.4b[1]\n"
+ ".word 0x4fa1e1d6 // sdot v22.4s, v14.16b, v1.4b[1]\n"
+ ".word 0x4fa2e1da // sdot v26.4s, v14.16b, v2.4b[1]\n"
+ "ldr q14, [%[b_ptr0], #0x60]\n"
+ ".word 0x4fa0e1f3 // sdot v19.4s, v15.16b, v0.4b[1]\n"
+ ".word 0x4fa1e1f7 // sdot v23.4s, v15.16b, v1.4b[1]\n"
+ ".word 0x4fa2e1fb // sdot v27.4s, v15.16b, v2.4b[1]\n"
+ "ldr q15, [%[b_ptr0], #0x70]\n"
+ ".word 0x4f80e910 // sdot v16.4s, v8.16b, v0.4b[2]\n"
+ "add %[b_ptr0], %[b_ptr0], #0x80\n"
+ ".word 0x4f81e914 // sdot v20.4s, v8.16b, v1.4b[2]\n"
+ ".word 0x4f82e918 // sdot v24.4s, v8.16b, v2.4b[2]\n"
+ ".word 0x4f80e931 // sdot v17.4s, v9.16b, v0.4b[2]\n"
+ ".word 0x4f81e935 // sdot v21.4s, v9.16b, v1.4b[2]\n"
+ ".word 0x4f82e939 // sdot v25.4s, v9.16b, v2.4b[2]\n"
+ ".word 0x4f80e952 // sdot v18.4s, v10.16b, v0.4b[2]\n"
+ ".word 0x4f81e956 // sdot v22.4s, v10.16b, v1.4b[2]\n"
+ ".word 0x4f82e95a // sdot v26.4s, v10.16b, v2.4b[2]\n"
+ ".word 0x4f80e973 // sdot v19.4s, v11.16b, v0.4b[2]\n"
+ ".word 0x4f81e977 // sdot v23.4s, v11.16b, v1.4b[2]\n"
+ ".word 0x4f82e97b // sdot v27.4s, v11.16b, v2.4b[2]\n"
+ ".word 0x4fa0e990 // sdot v16.4s, v12.16b, v0.4b[3]\n"
+ ".word 0x4fa1e994 // sdot v20.4s, v12.16b, v1.4b[3]\n"
+ ".word 0x4fa2e998 // sdot v24.4s, v12.16b, v2.4b[3]\n"
+ ".word 0x4fa0e9b1 // sdot v17.4s, v13.16b, v0.4b[3]\n"
+ ".word 0x4fa1e9b5 // sdot v21.4s, v13.16b, v1.4b[3]\n"
+ ".word 0x4fa2e9b9 // sdot v25.4s, v13.16b, v2.4b[3]\n"
+ ".word 0x4fa0e9d2 // sdot v18.4s, v14.16b, v0.4b[3]\n"
+ ".word 0x4fa1e9d6 // sdot v22.4s, v14.16b, v1.4b[3]\n"
+ ".word 0x4fa2e9da // sdot v26.4s, v14.16b, v2.4b[3]\n"
+ ".word 0x4fa0e9f3 // sdot v19.4s, v15.16b, v0.4b[3]\n"
+ ".word 0x4fa1e9f7 // sdot v23.4s, v15.16b, v1.4b[3]\n"
+ ".word 0x4fa2e9fb // sdot v27.4s, v15.16b, v2.4b[3]\n"
+ "5:\n"
+ "cbz %[blocks], 6f\n"
+ "7:\n"
+ "ldr q8, [%[b_ptr0]]\n"
+ "subs %[blocks], %[blocks], #0x1\n"
+ "ldr q9, [%[b_ptr0], #0x10]\n"
+ "ldr s0, [%[a_ptr0]]\n"
+ "ldr q10, [%[b_ptr0], #0x20]\n"
+ "add %[a_ptr0], %[a_ptr0], #0x4\n"
+ "ldr q11, [%[b_ptr0], #0x30]\n"
+ "add %[b_ptr0], %[b_ptr0], #0x40\n"
+ ".word 0x4f80e110 // sdot v16.4s, v8.16b, v0.4b[0]\n"
+ "ldr s1, [a_ptr1]\n"
+ ".word 0x4f80e131 // sdot v17.4s, v9.16b, v0.4b[0]\n"
+ "add a_ptr1, a_ptr1, #0x4\n"
+ ".word 0x4f80e152 // sdot v18.4s, v10.16b, v0.4b[0]\n"
+ "ldr s2, [a_ptr2]\n"
+ ".word 0x4f81e114 // sdot v20.4s, v8.16b, v1.4b[0]\n"
+ "add a_ptr2, a_ptr2, #0x4\n"
+ ".word 0x4f81e135 // sdot v21.4s, v9.16b, v1.4b[0]\n"
+ ".word 0x4f82e118 // sdot v24.4s, v8.16b, v2.4b[0]\n"
+ ".word 0x4f82e139 // sdot v25.4s, v9.16b, v2.4b[0]\n"
+ ".word 0x4f81e156 // sdot v22.4s, v10.16b, v1.4b[0]\n"
+ ".word 0x4f82e15a // sdot v26.4s, v10.16b, v2.4b[0]\n"
+ ".word 0x4f80e173 // sdot v19.4s, v11.16b, v0.4b[0]\n"
+ ".word 0x4f81e177 // sdot v23.4s, v11.16b, v1.4b[0]\n"
+ ".word 0x4f82e17b // sdot v27.4s, v11.16b, v2.4b[0]\n"
+ "b.ne 7b\n"
+ "6:\n"
+ "cbz %[odds], 8f\n"
+ "ld1 {v0.b}[0], [%[a_ptr0]], #1\n"
+ "ld1 {v1.b}[0], [a_ptr1], #1\n"
+ "ld1 {v2.b}[0], [a_ptr2], #1\n"
+ "subs %[odds], %[odds], #0x1\n"
+ "b.eq 9f\n"
+ "ld1 {v0.b}[1], [%[a_ptr0]], #1\n"
+ "ld1 {v1.b}[1], [a_ptr1], #1\n"
+ "ld1 {v2.b}[1], [a_ptr2], #1\n"
+ "subs %[odds], %[odds], #0x1\n"
+ "b.eq 9f\n"
+ "ld1 {v0.b}[2], [%[a_ptr0]]\n"
+ "ld1 {v1.b}[2], [a_ptr1]\n"
+ "ld1 {v2.b}[2], [a_ptr2]\n"
+ "9:\n"
+ "ldr q8, [%[b_ptr0]]\n"
+ "ldr q9, [%[b_ptr0], #0x10]\n"
+ "ldr q10, [%[b_ptr0], #0x20]\n"
+ "ldr q11, [%[b_ptr0], #0x30]\n"
+ ".word 0x4f80e110 // sdot v16.4s, v8.16b, v0.4b[0]\n"
+ ".word 0x4f81e114 // sdot v20.4s, v8.16b, v1.4b[0]\n"
+ ".word 0x4f82e118 // sdot v24.4s, v8.16b, v2.4b[0]\n"
+ ".word 0x4f80e131 // sdot v17.4s, v9.16b, v0.4b[0]\n"
+ ".word 0x4f81e135 // sdot v21.4s, v9.16b, v1.4b[0]\n"
+ ".word 0x4f82e139 // sdot v25.4s, v9.16b, v2.4b[0]\n"
+ ".word 0x4f80e152 // sdot v18.4s, v10.16b, v0.4b[0]\n"
+ ".word 0x4f81e156 // sdot v22.4s, v10.16b, v1.4b[0]\n"
+ ".word 0x4f82e15a // sdot v26.4s, v10.16b, v2.4b[0]\n"
+ ".word 0x4f80e173 // sdot v19.4s, v11.16b, v0.4b[0]\n"
+ ".word 0x4f81e177 // sdot v23.4s, v11.16b, v1.4b[0]\n"
+ ".word 0x4f82e17b // sdot v27.4s, v11.16b, v2.4b[0]\n"
+ "8:\n"
+ "str q16, [%[c_ptr0]]\n"
+ "str q17, [%[c_ptr0], #0x10]\n"
+ "str q18, [%[c_ptr0], #0x20]\n"
+ "str q19, [%[c_ptr0], #0x30]\n"
+ "add %[c_ptr0], %[c_ptr0], #0x40\n"
+ "str q20, [c_ptr1]\n"
+ "str q21, [c_ptr1, #0x10]\n"
+ "str q22, [c_ptr1, #0x20]\n"
+ "str q23, [c_ptr1, #0x30]\n"
+ "str q24, [c_ptr2]\n"
+ "str q25, [c_ptr2, #0x10]\n"
+ "str q26, [c_ptr2, #0x20]\n"
+ "str q27, [c_ptr2, #0x30]\n"
+ ".unreq a_ptr1\n"
+ ".unreq a_ptr2\n"
+ ".unreq c_ptr1\n"
+ ".unreq c_ptr2\n"
+ : [a_ptr0] "+r" (a_ptr0), [b_ptr0] "+r" (b_ptr0), [c_ptr0] "+r" (c_ptr0), [loops] "+r" (loops), [regs] "+r" (regs), [blocks] "+r" (blocks), [odds] "+r" (odds)
+ : [betaptr] "r" (betaptr), [width] "r" (width), [beta0] "r" (beta0), [lda] "r" (ldab), [ldc] "r" (ldcb)
+ : "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x0", "x1", "x2", "x3", "cc", "memory"
);
break;
default:
case 4:
__asm __volatile (
- "a_ptr1 .req X0\n"
- "a_ptr2 .req X1\n"
- "a_ptr3 .req X2\n"
- "c_ptr1 .req X3\n"
- "c_ptr2 .req X4\n"
- "c_ptr3 .req X5\n"
- "add a_ptr1, %[a_ptr0], %[lda]\n"
- "add c_ptr1, %[c_ptr0], %[ldc]\n"
- "add a_ptr2, a_ptr1, %[lda]\n"
- "add c_ptr2, c_ptr1, %[ldc]\n"
- "add a_ptr3, a_ptr2, %[lda]\n"
- "add c_ptr3, c_ptr2, %[ldc]\n"
- "cbz %[beta0], 1f\n"
- "movi v16.4s, #0\n"
- "ldr q0, [%[a_ptr0]]\n"
- "movi v17.4s, #0\n"
- "ldr q1, [a_ptr1]\n"
- "movi v18.4s, #0\n"
- "ldr q2, [a_ptr2]\n"
- "movi v19.4s, #0\n"
- "ldr q3, [a_ptr3]\n"
- "movi v20.4s, #0\n"
- "ldr q8, [%[b_ptr0]]\n"
- "movi v21.4s, #0\n"
- "ldr q9, [%[b_ptr0], #0x10]\n"
- "movi v22.4s, #0\n"
- "ldr q10, [%[b_ptr0], #0x20]\n"
- "movi v23.4s, #0\n"
- "ldr q11, [%[b_ptr0], #0x30]\n"
- "movi v24.4s, #0\n"
- "ldr q12, [%[b_ptr0], #0x40]\n"
- "movi v25.4s, #0\n"
- "ldr q13, [%[b_ptr0], #0x50]\n"
- "movi v26.4s, #0\n"
- "ldr q14, [%[b_ptr0], #0x60]\n"
- "movi v27.4s, #0\n"
- "add %[a_ptr0], %[a_ptr0], #0x10\n"
- "movi v28.4s, #0\n"
- "add a_ptr1, a_ptr1, #0x10\n"
- "movi v29.4s, #0\n"
- "add a_ptr2, a_ptr2, #0x10\n"
- "movi v30.4s, #0\n"
- "add a_ptr3, a_ptr3, #0x10\n"
- "movi v31.4s, #0\n"
- "add %[b_ptr0], %[b_ptr0], #0x80\n"
- "cbz %[loops], 2f\n"
- "b 3f\n"
- "1:\n"
- "ld1r {v15.4s}, [%[betaptr]]\n"
- "ldr q16, [%[c_ptr0]]\n"
- "ldr q17, [%[c_ptr0], #0x10]\n"
- "ldr q18, [%[c_ptr0], #0x20]\n"
- "ldr q19, [%[c_ptr0], #0x30]\n"
- "mul v16.4s, v16.4s, v15.4s\n"
- "ldr q20, [c_ptr1]\n"
- "mul v17.4s, v17.4s, v15.4s\n"
- "ldr q21, [c_ptr1, #0x10]\n"
- "mul v18.4s, v18.4s, v15.4s\n"
- "ldr q22, [c_ptr1, #0x20]\n"
- "mul v19.4s, v19.4s, v15.4s\n"
- "ldr q23, [c_ptr1, #0x30]\n"
- "mul v20.4s, v20.4s, v15.4s\n"
- "ldr q24, [c_ptr2]\n"
- "mul v21.4s, v21.4s, v15.4s\n"
- "ldr q25, [c_ptr2, #0x10]\n"
- "mul v22.4s, v22.4s, v15.4s\n"
- "ldr q26, [c_ptr2, #0x20]\n"
- "mul v23.4s, v23.4s, v15.4s\n"
- "ldr q27, [c_ptr2, #0x30]\n"
- "mul v24.4s, v24.4s, v15.4s\n"
- "ldr q28, [c_ptr3]\n"
- "mul v25.4s, v25.4s, v15.4s\n"
- "ldr q29, [c_ptr3, #0x10]\n"
- "mul v26.4s, v26.4s, v15.4s\n"
- "ldr q30, [c_ptr3, #0x20]\n"
- "mul v27.4s, v27.4s, v15.4s\n"
- "ldr q31, [c_ptr3, #0x30]\n"
- "mul v28.4s, v28.4s, v15.4s\n"
- "ldr q0, [%[a_ptr0]]\n"
- "mul v29.4s, v29.4s, v15.4s\n"
- "ldr q1, [a_ptr1]\n"
- "mul v30.4s, v30.4s, v15.4s\n"
- "ldr q2, [a_ptr2]\n"
- "mul v31.4s, v31.4s, v15.4s\n"
- "ldr q3, [a_ptr3]\n"
- "ldr q8, [%[b_ptr0]]\n"
- "add %[a_ptr0], %[a_ptr0], #0x10\n"
- "ldr q9, [%[b_ptr0], #0x10]\n"
- "add a_ptr1, a_ptr1, #0x10\n"
- "ldr q10, [%[b_ptr0], #0x20]\n"
- "add a_ptr2, a_ptr2, #0x10\n"
- "ldr q11, [%[b_ptr0], #0x30]\n"
- "add a_ptr3, a_ptr3, #0x10\n"
- "ldr q12, [%[b_ptr0], #0x40]\n"
- "ldr q13, [%[b_ptr0], #0x50]\n"
- "ldr q14, [%[b_ptr0], #0x60]\n"
- "add %[b_ptr0], %[b_ptr0], #0x80\n"
- "cbz %[loops], 2f\n"
- "3:\n"
- ".word 0x4f80e110 // sdot v16.4s, v8.16b, v0.4b[0]\n"
- "ldr q15, [%[b_ptr0], #-0x10]\n"
- ".word 0x4f81e114 // sdot v20.4s, v8.16b, v1.4b[0]\n"
- "ldr q4, [%[a_ptr0]]\n"
- ".word 0x4f82e118 // sdot v24.4s, v8.16b, v2.4b[0]\n"
- "ldr q5, [a_ptr1]\n"
- ".word 0x4f83e11c // sdot v28.4s, v8.16b, v3.4b[0]\n"
- "ldr q6, [a_ptr2]\n"
- ".word 0x4f80e131 // sdot v17.4s, v9.16b, v0.4b[0]\n"
- "ldr q7, [a_ptr3]\n"
- ".word 0x4f81e135 // sdot v21.4s, v9.16b, v1.4b[0]\n"
- "ldr q8, [%[b_ptr0]]\n"
- ".word 0x4f82e139 // sdot v25.4s, v9.16b, v2.4b[0]\n"
- "subs %[loops], %[loops], #0x1\n"
- ".word 0x4f83e13d // sdot v29.4s, v9.16b, v3.4b[0]\n"
- "ldr q9, [%[b_ptr0], #0x10]\n"
- ".word 0x4f80e152 // sdot v18.4s, v10.16b, v0.4b[0]\n"
- "prfm PLDL1KEEP, [%[a_ptr0], #0x40]\n"
- ".word 0x4f81e156 // sdot v22.4s, v10.16b, v1.4b[0]\n"
- "add %[a_ptr0], %[a_ptr0], #0x20\n"
- ".word 0x4f82e15a // sdot v26.4s, v10.16b, v2.4b[0]\n"
- "add a_ptr1, a_ptr1, #0x20\n"
- ".word 0x4f83e15e // sdot v30.4s, v10.16b, v3.4b[0]\n"
- "ldr q10, [%[b_ptr0], #0x20]\n"
- ".word 0x4f80e173 // sdot v19.4s, v11.16b, v0.4b[0]\n"
- "add a_ptr2, a_ptr2, #0x20\n"
- ".word 0x4f81e177 // sdot v23.4s, v11.16b, v1.4b[0]\n"
- "add a_ptr3, a_ptr3, #0x20\n"
- ".word 0x4f82e17b // sdot v27.4s, v11.16b, v2.4b[0]\n"
- "prfm PLDL1KEEP, [a_ptr1, #0x40]\n"
- ".word 0x4f83e17f // sdot v31.4s, v11.16b, v3.4b[0]\n"
- "ldr q11, [%[b_ptr0], #0x30]\n"
- ".word 0x4fa0e190 // sdot v16.4s, v12.16b, v0.4b[1]\n"
- "prfm PLDL1KEEP, [a_ptr2, #0x40]\n"
- ".word 0x4fa1e194 // sdot v20.4s, v12.16b, v1.4b[1]\n"
- "prfm PLDL1KEEP, [a_ptr3, #0x40]\n"
- ".word 0x4fa2e198 // sdot v24.4s, v12.16b, v2.4b[1]\n"
- ".word 0x4fa3e19c // sdot v28.4s, v12.16b, v3.4b[1]\n"
- "ldr q12, [%[b_ptr0], #0x40]\n"
- ".word 0x4fa0e1b1 // sdot v17.4s, v13.16b, v0.4b[1]\n"
- ".word 0x4fa1e1b5 // sdot v21.4s, v13.16b, v1.4b[1]\n"
- ".word 0x4fa2e1b9 // sdot v25.4s, v13.16b, v2.4b[1]\n"
- ".word 0x4fa3e1bd // sdot v29.4s, v13.16b, v3.4b[1]\n"
- "ldr q13, [%[b_ptr0], #0x50]\n"
- ".word 0x4fa0e1d2 // sdot v18.4s, v14.16b, v0.4b[1]\n"
- ".word 0x4fa1e1d6 // sdot v22.4s, v14.16b, v1.4b[1]\n"
- ".word 0x4fa2e1da // sdot v26.4s, v14.16b, v2.4b[1]\n"
- ".word 0x4fa3e1de // sdot v30.4s, v14.16b, v3.4b[1]\n"
- "ldr q14, [%[b_ptr0], #0x60]\n"
- ".word 0x4fa0e1f3 // sdot v19.4s, v15.16b, v0.4b[1]\n"
- ".word 0x4fa1e1f7 // sdot v23.4s, v15.16b, v1.4b[1]\n"
- ".word 0x4fa2e1fb // sdot v27.4s, v15.16b, v2.4b[1]\n"
- ".word 0x4fa3e1ff // sdot v31.4s, v15.16b, v3.4b[1]\n"
- "ldr q15, [%[b_ptr0], #0x70]\n"
- ".word 0x4f80e910 // sdot v16.4s, v8.16b, v0.4b[2]\n"
- "add %[b_ptr0], %[b_ptr0], #0x100\n"
- ".word 0x4f81e914 // sdot v20.4s, v8.16b, v1.4b[2]\n"
- ".word 0x4f82e918 // sdot v24.4s, v8.16b, v2.4b[2]\n"
- ".word 0x4f83e91c // sdot v28.4s, v8.16b, v3.4b[2]\n"
- "ldr q8, [%[b_ptr0], #-0x80]\n"
- ".word 0x4f80e931 // sdot v17.4s, v9.16b, v0.4b[2]\n"
- ".word 0x4f81e935 // sdot v21.4s, v9.16b, v1.4b[2]\n"
- ".word 0x4f82e939 // sdot v25.4s, v9.16b, v2.4b[2]\n"
- ".word 0x4f83e93d // sdot v29.4s, v9.16b, v3.4b[2]\n"
- "ldr q9, [%[b_ptr0], #-0x70]\n"
- ".word 0x4f80e952 // sdot v18.4s, v10.16b, v0.4b[2]\n"
- ".word 0x4f81e956 // sdot v22.4s, v10.16b, v1.4b[2]\n"
- ".word 0x4f82e95a // sdot v26.4s, v10.16b, v2.4b[2]\n"
- ".word 0x4f83e95e // sdot v30.4s, v10.16b, v3.4b[2]\n"
- "ldr q10, [%[b_ptr0], #-0x60]\n"
- ".word 0x4f80e973 // sdot v19.4s, v11.16b, v0.4b[2]\n"
- ".word 0x4f81e977 // sdot v23.4s, v11.16b, v1.4b[2]\n"
- ".word 0x4f82e97b // sdot v27.4s, v11.16b, v2.4b[2]\n"
- ".word 0x4f83e97f // sdot v31.4s, v11.16b, v3.4b[2]\n"
- "ldr q11, [%[b_ptr0], #-0x50]\n"
- ".word 0x4fa0e990 // sdot v16.4s, v12.16b, v0.4b[3]\n"
- ".word 0x4fa1e994 // sdot v20.4s, v12.16b, v1.4b[3]\n"
- ".word 0x4fa2e998 // sdot v24.4s, v12.16b, v2.4b[3]\n"
- ".word 0x4fa3e99c // sdot v28.4s, v12.16b, v3.4b[3]\n"
- "ldr q12, [%[b_ptr0], #-0x40]\n"
- ".word 0x4fa0e9b1 // sdot v17.4s, v13.16b, v0.4b[3]\n"
- ".word 0x4fa1e9b5 // sdot v21.4s, v13.16b, v1.4b[3]\n"
- ".word 0x4fa2e9b9 // sdot v25.4s, v13.16b, v2.4b[3]\n"
- ".word 0x4fa3e9bd // sdot v29.4s, v13.16b, v3.4b[3]\n"
- "ldr q13, [%[b_ptr0], #-0x30]\n"
- ".word 0x4fa0e9d2 // sdot v18.4s, v14.16b, v0.4b[3]\n"
- ".word 0x4fa1e9d6 // sdot v22.4s, v14.16b, v1.4b[3]\n"
- ".word 0x4fa2e9da // sdot v26.4s, v14.16b, v2.4b[3]\n"
- ".word 0x4fa3e9de // sdot v30.4s, v14.16b, v3.4b[3]\n"
- "ldr q14, [%[b_ptr0], #-0x20]\n"
- ".word 0x4fa0e9f3 // sdot v19.4s, v15.16b, v0.4b[3]\n"
- "ldr q0, [%[a_ptr0], #-0x10]\n"
- ".word 0x4fa1e9f7 // sdot v23.4s, v15.16b, v1.4b[3]\n"
- "ldr q1, [a_ptr1, #-0x10]\n"
- ".word 0x4fa2e9fb // sdot v27.4s, v15.16b, v2.4b[3]\n"
- "ldr q2, [a_ptr2, #-0x10]\n"
- ".word 0x4fa3e9ff // sdot v31.4s, v15.16b, v3.4b[3]\n"
- "ldr q15, [%[b_ptr0], #-0x10]\n"
- ".word 0x4f84e110 // sdot v16.4s, v8.16b, v4.4b[0]\n"
- "ldr q3, [a_ptr3, #-0x10]\n"
- ".word 0x4f85e114 // sdot v20.4s, v8.16b, v5.4b[0]\n"
- ".word 0x4f86e118 // sdot v24.4s, v8.16b, v6.4b[0]\n"
- ".word 0x4f87e11c // sdot v28.4s, v8.16b, v7.4b[0]\n"
- "ldr q8, [%[b_ptr0]]\n"
- ".word 0x4f84e131 // sdot v17.4s, v9.16b, v4.4b[0]\n"
- ".word 0x4f85e135 // sdot v21.4s, v9.16b, v5.4b[0]\n"
- ".word 0x4f86e139 // sdot v25.4s, v9.16b, v6.4b[0]\n"
- ".word 0x4f87e13d // sdot v29.4s, v9.16b, v7.4b[0]\n"
- "ldr q9, [%[b_ptr0], #0x10]\n"
- ".word 0x4f84e152 // sdot v18.4s, v10.16b, v4.4b[0]\n"
- ".word 0x4f85e156 // sdot v22.4s, v10.16b, v5.4b[0]\n"
- ".word 0x4f86e15a // sdot v26.4s, v10.16b, v6.4b[0]\n"
- ".word 0x4f87e15e // sdot v30.4s, v10.16b, v7.4b[0]\n"
- "ldr q10, [%[b_ptr0], #0x20]\n"
- ".word 0x4f84e173 // sdot v19.4s, v11.16b, v4.4b[0]\n"
- ".word 0x4f85e177 // sdot v23.4s, v11.16b, v5.4b[0]\n"
- ".word 0x4f86e17b // sdot v27.4s, v11.16b, v6.4b[0]\n"
- ".word 0x4f87e17f // sdot v31.4s, v11.16b, v7.4b[0]\n"
- "ldr q11, [%[b_ptr0], #0x30]\n"
- ".word 0x4fa4e190 // sdot v16.4s, v12.16b, v4.4b[1]\n"
- ".word 0x4fa5e194 // sdot v20.4s, v12.16b, v5.4b[1]\n"
- ".word 0x4fa6e198 // sdot v24.4s, v12.16b, v6.4b[1]\n"
- ".word 0x4fa7e19c // sdot v28.4s, v12.16b, v7.4b[1]\n"
- "ldr q12, [%[b_ptr0], #0x40]\n"
- ".word 0x4fa4e1b1 // sdot v17.4s, v13.16b, v4.4b[1]\n"
- ".word 0x4fa5e1b5 // sdot v21.4s, v13.16b, v5.4b[1]\n"
- ".word 0x4fa6e1b9 // sdot v25.4s, v13.16b, v6.4b[1]\n"
- ".word 0x4fa7e1bd // sdot v29.4s, v13.16b, v7.4b[1]\n"
- "ldr q13, [%[b_ptr0], #0x50]\n"
- ".word 0x4fa4e1d2 // sdot v18.4s, v14.16b, v4.4b[1]\n"
- ".word 0x4fa5e1d6 // sdot v22.4s, v14.16b, v5.4b[1]\n"
- ".word 0x4fa6e1da // sdot v26.4s, v14.16b, v6.4b[1]\n"
- ".word 0x4fa7e1de // sdot v30.4s, v14.16b, v7.4b[1]\n"
- "ldr q14, [%[b_ptr0], #0x60]\n"
- ".word 0x4fa4e1f3 // sdot v19.4s, v15.16b, v4.4b[1]\n"
- ".word 0x4fa5e1f7 // sdot v23.4s, v15.16b, v5.4b[1]\n"
- ".word 0x4fa6e1fb // sdot v27.4s, v15.16b, v6.4b[1]\n"
- ".word 0x4fa7e1ff // sdot v31.4s, v15.16b, v7.4b[1]\n"
- "ldr q15, [%[b_ptr0], #0x70]\n"
- ".word 0x4f84e910 // sdot v16.4s, v8.16b, v4.4b[2]\n"
- "add %[b_ptr0], %[b_ptr0], #0x100\n"
- ".word 0x4f85e914 // sdot v20.4s, v8.16b, v5.4b[2]\n"
- ".word 0x4f86e918 // sdot v24.4s, v8.16b, v6.4b[2]\n"
- ".word 0x4f87e91c // sdot v28.4s, v8.16b, v7.4b[2]\n"
- "ldr q8, [%[b_ptr0], #-0x80]\n"
- ".word 0x4f84e931 // sdot v17.4s, v9.16b, v4.4b[2]\n"
- ".word 0x4f85e935 // sdot v21.4s, v9.16b, v5.4b[2]\n"
- ".word 0x4f86e939 // sdot v25.4s, v9.16b, v6.4b[2]\n"
- ".word 0x4f87e93d // sdot v29.4s, v9.16b, v7.4b[2]\n"
- "ldr q9, [%[b_ptr0], #-0x70]\n"
- ".word 0x4f84e952 // sdot v18.4s, v10.16b, v4.4b[2]\n"
- ".word 0x4f85e956 // sdot v22.4s, v10.16b, v5.4b[2]\n"
- ".word 0x4f86e95a // sdot v26.4s, v10.16b, v6.4b[2]\n"
- ".word 0x4f87e95e // sdot v30.4s, v10.16b, v7.4b[2]\n"
- "ldr q10, [%[b_ptr0], #-0x60]\n"
- ".word 0x4f84e973 // sdot v19.4s, v11.16b, v4.4b[2]\n"
- ".word 0x4f85e977 // sdot v23.4s, v11.16b, v5.4b[2]\n"
- ".word 0x4f86e97b // sdot v27.4s, v11.16b, v6.4b[2]\n"
- ".word 0x4f87e97f // sdot v31.4s, v11.16b, v7.4b[2]\n"
- "ldr q11, [%[b_ptr0], #-0x50]\n"
- ".word 0x4fa4e990 // sdot v16.4s, v12.16b, v4.4b[3]\n"
- ".word 0x4fa5e994 // sdot v20.4s, v12.16b, v5.4b[3]\n"
- ".word 0x4fa6e998 // sdot v24.4s, v12.16b, v6.4b[3]\n"
- ".word 0x4fa7e99c // sdot v28.4s, v12.16b, v7.4b[3]\n"
- "ldr q12, [%[b_ptr0], #-0x40]\n"
- ".word 0x4fa4e9b1 // sdot v17.4s, v13.16b, v4.4b[3]\n"
- ".word 0x4fa5e9b5 // sdot v21.4s, v13.16b, v5.4b[3]\n"
- ".word 0x4fa6e9b9 // sdot v25.4s, v13.16b, v6.4b[3]\n"
- ".word 0x4fa7e9bd // sdot v29.4s, v13.16b, v7.4b[3]\n"
- "ldr q13, [%[b_ptr0], #-0x30]\n"
- ".word 0x4fa4e9d2 // sdot v18.4s, v14.16b, v4.4b[3]\n"
- ".word 0x4fa5e9d6 // sdot v22.4s, v14.16b, v5.4b[3]\n"
- ".word 0x4fa6e9da // sdot v26.4s, v14.16b, v6.4b[3]\n"
- ".word 0x4fa7e9de // sdot v30.4s, v14.16b, v7.4b[3]\n"
- "ldr q14, [%[b_ptr0], #-0x20]\n"
- ".word 0x4fa4e9f3 // sdot v19.4s, v15.16b, v4.4b[3]\n"
- ".word 0x4fa5e9f7 // sdot v23.4s, v15.16b, v5.4b[3]\n"
- ".word 0x4fa6e9fb // sdot v27.4s, v15.16b, v6.4b[3]\n"
- ".word 0x4fa7e9ff // sdot v31.4s, v15.16b, v7.4b[3]\n"
- "b.ne 3b\n"
- "2:\n"
- "ldr q15, [%[b_ptr0], #-0x10]\n"
- "prfm PSTL1KEEP, [%[c_ptr0]]\n"
- "prfm PSTL1KEEP, [c_ptr1]\n"
- "prfm PSTL1KEEP, [c_ptr2]\n"
- "prfm PSTL1KEEP, [c_ptr3]\n"
- "cbz %[regs], 4f\n"
- ".word 0x4f80e110 // sdot v16.4s, v8.16b, v0.4b[0]\n"
- "ldr q4, [%[a_ptr0]]\n"
- ".word 0x4f81e114 // sdot v20.4s, v8.16b, v1.4b[0]\n"
- "ldr q5, [a_ptr1]\n"
- ".word 0x4f82e118 // sdot v24.4s, v8.16b, v2.4b[0]\n"
- "ldr q6, [a_ptr2]\n"
- ".word 0x4f83e11c // sdot v28.4s, v8.16b, v3.4b[0]\n"
- "ldr q7, [a_ptr3]\n"
- ".word 0x4f80e131 // sdot v17.4s, v9.16b, v0.4b[0]\n"
- "ldr q8, [%[b_ptr0]]\n"
- ".word 0x4f81e135 // sdot v21.4s, v9.16b, v1.4b[0]\n"
- ".word 0x4f82e139 // sdot v25.4s, v9.16b, v2.4b[0]\n"
- ".word 0x4f83e13d // sdot v29.4s, v9.16b, v3.4b[0]\n"
- "ldr q9, [%[b_ptr0], #0x10]\n"
- ".word 0x4f80e152 // sdot v18.4s, v10.16b, v0.4b[0]\n"
- ".word 0x4f81e156 // sdot v22.4s, v10.16b, v1.4b[0]\n"
- ".word 0x4f82e15a // sdot v26.4s, v10.16b, v2.4b[0]\n"
- ".word 0x4f83e15e // sdot v30.4s, v10.16b, v3.4b[0]\n"
- "ldr q10, [%[b_ptr0], #0x20]\n"
- ".word 0x4f80e173 // sdot v19.4s, v11.16b, v0.4b[0]\n"
- ".word 0x4f81e177 // sdot v23.4s, v11.16b, v1.4b[0]\n"
- ".word 0x4f82e17b // sdot v27.4s, v11.16b, v2.4b[0]\n"
- ".word 0x4f83e17f // sdot v31.4s, v11.16b, v3.4b[0]\n"
- "ldr q11, [%[b_ptr0], #0x30]\n"
- ".word 0x4fa0e190 // sdot v16.4s, v12.16b, v0.4b[1]\n"
- ".word 0x4fa1e194 // sdot v20.4s, v12.16b, v1.4b[1]\n"
- ".word 0x4fa2e198 // sdot v24.4s, v12.16b, v2.4b[1]\n"
- ".word 0x4fa3e19c // sdot v28.4s, v12.16b, v3.4b[1]\n"
- "ldr q12, [%[b_ptr0], #0x40]\n"
- ".word 0x4fa0e1b1 // sdot v17.4s, v13.16b, v0.4b[1]\n"
- ".word 0x4fa1e1b5 // sdot v21.4s, v13.16b, v1.4b[1]\n"
- ".word 0x4fa2e1b9 // sdot v25.4s, v13.16b, v2.4b[1]\n"
- ".word 0x4fa3e1bd // sdot v29.4s, v13.16b, v3.4b[1]\n"
- "ldr q13, [%[b_ptr0], #0x50]\n"
- ".word 0x4fa0e1d2 // sdot v18.4s, v14.16b, v0.4b[1]\n"
- ".word 0x4fa1e1d6 // sdot v22.4s, v14.16b, v1.4b[1]\n"
- ".word 0x4fa2e1da // sdot v26.4s, v14.16b, v2.4b[1]\n"
- ".word 0x4fa3e1de // sdot v30.4s, v14.16b, v3.4b[1]\n"
- "ldr q14, [%[b_ptr0], #0x60]\n"
- ".word 0x4fa0e1f3 // sdot v19.4s, v15.16b, v0.4b[1]\n"
- ".word 0x4fa1e1f7 // sdot v23.4s, v15.16b, v1.4b[1]\n"
- ".word 0x4fa2e1fb // sdot v27.4s, v15.16b, v2.4b[1]\n"
- ".word 0x4fa3e1ff // sdot v31.4s, v15.16b, v3.4b[1]\n"
- "ldr q15, [%[b_ptr0], #0x70]\n"
- ".word 0x4f80e910 // sdot v16.4s, v8.16b, v0.4b[2]\n"
- "add %[b_ptr0], %[b_ptr0], #0x100\n"
- ".word 0x4f81e914 // sdot v20.4s, v8.16b, v1.4b[2]\n"
- ".word 0x4f82e918 // sdot v24.4s, v8.16b, v2.4b[2]\n"
- ".word 0x4f83e91c // sdot v28.4s, v8.16b, v3.4b[2]\n"
- "ldr q8, [%[b_ptr0], #-0x80]\n"
- ".word 0x4f80e931 // sdot v17.4s, v9.16b, v0.4b[2]\n"
- ".word 0x4f81e935 // sdot v21.4s, v9.16b, v1.4b[2]\n"
- ".word 0x4f82e939 // sdot v25.4s, v9.16b, v2.4b[2]\n"
- ".word 0x4f83e93d // sdot v29.4s, v9.16b, v3.4b[2]\n"
- "ldr q9, [%[b_ptr0], #-0x70]\n"
- ".word 0x4f80e952 // sdot v18.4s, v10.16b, v0.4b[2]\n"
- ".word 0x4f81e956 // sdot v22.4s, v10.16b, v1.4b[2]\n"
- ".word 0x4f82e95a // sdot v26.4s, v10.16b, v2.4b[2]\n"
- ".word 0x4f83e95e // sdot v30.4s, v10.16b, v3.4b[2]\n"
- "ldr q10, [%[b_ptr0], #-0x60]\n"
- ".word 0x4f80e973 // sdot v19.4s, v11.16b, v0.4b[2]\n"
- ".word 0x4f81e977 // sdot v23.4s, v11.16b, v1.4b[2]\n"
- ".word 0x4f82e97b // sdot v27.4s, v11.16b, v2.4b[2]\n"
- ".word 0x4f83e97f // sdot v31.4s, v11.16b, v3.4b[2]\n"
- "ldr q11, [%[b_ptr0], #-0x50]\n"
- ".word 0x4fa0e990 // sdot v16.4s, v12.16b, v0.4b[3]\n"
- ".word 0x4fa1e994 // sdot v20.4s, v12.16b, v1.4b[3]\n"
- ".word 0x4fa2e998 // sdot v24.4s, v12.16b, v2.4b[3]\n"
- ".word 0x4fa3e99c // sdot v28.4s, v12.16b, v3.4b[3]\n"
- "ldr q12, [%[b_ptr0], #-0x40]\n"
- ".word 0x4fa0e9b1 // sdot v17.4s, v13.16b, v0.4b[3]\n"
- ".word 0x4fa1e9b5 // sdot v21.4s, v13.16b, v1.4b[3]\n"
- ".word 0x4fa2e9b9 // sdot v25.4s, v13.16b, v2.4b[3]\n"
- ".word 0x4fa3e9bd // sdot v29.4s, v13.16b, v3.4b[3]\n"
- "ldr q13, [%[b_ptr0], #-0x30]\n"
- ".word 0x4fa0e9d2 // sdot v18.4s, v14.16b, v0.4b[3]\n"
- ".word 0x4fa1e9d6 // sdot v22.4s, v14.16b, v1.4b[3]\n"
- ".word 0x4fa2e9da // sdot v26.4s, v14.16b, v2.4b[3]\n"
- ".word 0x4fa3e9de // sdot v30.4s, v14.16b, v3.4b[3]\n"
- "ldr q14, [%[b_ptr0], #-0x20]\n"
- ".word 0x4fa0e9f3 // sdot v19.4s, v15.16b, v0.4b[3]\n"
- "ldr q0, [%[a_ptr0], #0x10]\n"
- ".word 0x4fa1e9f7 // sdot v23.4s, v15.16b, v1.4b[3]\n"
- "ldr q1, [a_ptr1, #0x10]\n"
- ".word 0x4fa2e9fb // sdot v27.4s, v15.16b, v2.4b[3]\n"
- "ldr q2, [a_ptr2, #0x10]\n"
- ".word 0x4fa3e9ff // sdot v31.4s, v15.16b, v3.4b[3]\n"
- "ldr q15, [%[b_ptr0], #-0x10]\n"
- ".word 0x4f84e110 // sdot v16.4s, v8.16b, v4.4b[0]\n"
- "ldr q3, [a_ptr3, #0x10]\n"
- ".word 0x4f85e114 // sdot v20.4s, v8.16b, v5.4b[0]\n"
- ".word 0x4f86e118 // sdot v24.4s, v8.16b, v6.4b[0]\n"
- ".word 0x4f87e11c // sdot v28.4s, v8.16b, v7.4b[0]\n"
- "ldr q8, [%[b_ptr0]]\n"
- ".word 0x4f84e131 // sdot v17.4s, v9.16b, v4.4b[0]\n"
- ".word 0x4f85e135 // sdot v21.4s, v9.16b, v5.4b[0]\n"
- ".word 0x4f86e139 // sdot v25.4s, v9.16b, v6.4b[0]\n"
- ".word 0x4f87e13d // sdot v29.4s, v9.16b, v7.4b[0]\n"
- "ldr q9, [%[b_ptr0], #0x10]\n"
- ".word 0x4f84e152 // sdot v18.4s, v10.16b, v4.4b[0]\n"
- ".word 0x4f85e156 // sdot v22.4s, v10.16b, v5.4b[0]\n"
- ".word 0x4f86e15a // sdot v26.4s, v10.16b, v6.4b[0]\n"
- ".word 0x4f87e15e // sdot v30.4s, v10.16b, v7.4b[0]\n"
- "ldr q10, [%[b_ptr0], #0x20]\n"
- ".word 0x4f84e173 // sdot v19.4s, v11.16b, v4.4b[0]\n"
- ".word 0x4f85e177 // sdot v23.4s, v11.16b, v5.4b[0]\n"
- ".word 0x4f86e17b // sdot v27.4s, v11.16b, v6.4b[0]\n"
- ".word 0x4f87e17f // sdot v31.4s, v11.16b, v7.4b[0]\n"
- "ldr q11, [%[b_ptr0], #0x30]\n"
- ".word 0x4fa4e190 // sdot v16.4s, v12.16b, v4.4b[1]\n"
- ".word 0x4fa5e194 // sdot v20.4s, v12.16b, v5.4b[1]\n"
- ".word 0x4fa6e198 // sdot v24.4s, v12.16b, v6.4b[1]\n"
- ".word 0x4fa7e19c // sdot v28.4s, v12.16b, v7.4b[1]\n"
- "ldr q12, [%[b_ptr0], #0x40]\n"
- ".word 0x4fa4e1b1 // sdot v17.4s, v13.16b, v4.4b[1]\n"
- ".word 0x4fa5e1b5 // sdot v21.4s, v13.16b, v5.4b[1]\n"
- ".word 0x4fa6e1b9 // sdot v25.4s, v13.16b, v6.4b[1]\n"
- ".word 0x4fa7e1bd // sdot v29.4s, v13.16b, v7.4b[1]\n"
- "ldr q13, [%[b_ptr0], #0x50]\n"
- ".word 0x4fa4e1d2 // sdot v18.4s, v14.16b, v4.4b[1]\n"
- ".word 0x4fa5e1d6 // sdot v22.4s, v14.16b, v5.4b[1]\n"
- ".word 0x4fa6e1da // sdot v26.4s, v14.16b, v6.4b[1]\n"
- ".word 0x4fa7e1de // sdot v30.4s, v14.16b, v7.4b[1]\n"
- "ldr q14, [%[b_ptr0], #0x60]\n"
- ".word 0x4fa4e1f3 // sdot v19.4s, v15.16b, v4.4b[1]\n"
- ".word 0x4fa5e1f7 // sdot v23.4s, v15.16b, v5.4b[1]\n"
- ".word 0x4fa6e1fb // sdot v27.4s, v15.16b, v6.4b[1]\n"
- ".word 0x4fa7e1ff // sdot v31.4s, v15.16b, v7.4b[1]\n"
- "ldr q15, [%[b_ptr0], #0x70]\n"
- ".word 0x4f84e910 // sdot v16.4s, v8.16b, v4.4b[2]\n"
- ".word 0x4f85e914 // sdot v20.4s, v8.16b, v5.4b[2]\n"
- ".word 0x4f86e918 // sdot v24.4s, v8.16b, v6.4b[2]\n"
- ".word 0x4f87e91c // sdot v28.4s, v8.16b, v7.4b[2]\n"
- ".word 0x4f84e931 // sdot v17.4s, v9.16b, v4.4b[2]\n"
- ".word 0x4f85e935 // sdot v21.4s, v9.16b, v5.4b[2]\n"
- ".word 0x4f86e939 // sdot v25.4s, v9.16b, v6.4b[2]\n"
- ".word 0x4f87e93d // sdot v29.4s, v9.16b, v7.4b[2]\n"
- ".word 0x4f84e952 // sdot v18.4s, v10.16b, v4.4b[2]\n"
- ".word 0x4f85e956 // sdot v22.4s, v10.16b, v5.4b[2]\n"
- ".word 0x4f86e95a // sdot v26.4s, v10.16b, v6.4b[2]\n"
- ".word 0x4f87e95e // sdot v30.4s, v10.16b, v7.4b[2]\n"
- ".word 0x4f84e973 // sdot v19.4s, v11.16b, v4.4b[2]\n"
- ".word 0x4f85e977 // sdot v23.4s, v11.16b, v5.4b[2]\n"
- ".word 0x4f86e97b // sdot v27.4s, v11.16b, v6.4b[2]\n"
- ".word 0x4f87e97f // sdot v31.4s, v11.16b, v7.4b[2]\n"
- ".word 0x4fa4e990 // sdot v16.4s, v12.16b, v4.4b[3]\n"
- ".word 0x4fa5e994 // sdot v20.4s, v12.16b, v5.4b[3]\n"
- ".word 0x4fa6e998 // sdot v24.4s, v12.16b, v6.4b[3]\n"
- ".word 0x4fa7e99c // sdot v28.4s, v12.16b, v7.4b[3]\n"
- ".word 0x4fa4e9b1 // sdot v17.4s, v13.16b, v4.4b[3]\n"
- ".word 0x4fa5e9b5 // sdot v21.4s, v13.16b, v5.4b[3]\n"
- ".word 0x4fa6e9b9 // sdot v25.4s, v13.16b, v6.4b[3]\n"
- ".word 0x4fa7e9bd // sdot v29.4s, v13.16b, v7.4b[3]\n"
- ".word 0x4fa4e9d2 // sdot v18.4s, v14.16b, v4.4b[3]\n"
- ".word 0x4fa5e9d6 // sdot v22.4s, v14.16b, v5.4b[3]\n"
- ".word 0x4fa6e9da // sdot v26.4s, v14.16b, v6.4b[3]\n"
- ".word 0x4fa7e9de // sdot v30.4s, v14.16b, v7.4b[3]\n"
- ".word 0x4fa4e9f3 // sdot v19.4s, v15.16b, v4.4b[3]\n"
- ".word 0x4fa5e9f7 // sdot v23.4s, v15.16b, v5.4b[3]\n"
- ".word 0x4fa6e9fb // sdot v27.4s, v15.16b, v6.4b[3]\n"
- ".word 0x4fa7e9ff // sdot v31.4s, v15.16b, v7.4b[3]\n"
- "b 5f\n"
- "4:\n"
- ".word 0x4f80e110 // sdot v16.4s, v8.16b, v0.4b[0]\n"
- "ldr q4, [%[a_ptr0]]\n"
- ".word 0x4f81e114 // sdot v20.4s, v8.16b, v1.4b[0]\n"
- "ldr q5, [a_ptr1]\n"
- ".word 0x4f82e118 // sdot v24.4s, v8.16b, v2.4b[0]\n"
- "ldr q6, [a_ptr2]\n"
- ".word 0x4f83e11c // sdot v28.4s, v8.16b, v3.4b[0]\n"
- "ldr q7, [a_ptr3]\n"
- ".word 0x4f80e131 // sdot v17.4s, v9.16b, v0.4b[0]\n"
- "ldr q8, [%[b_ptr0]]\n"
- ".word 0x4f81e135 // sdot v21.4s, v9.16b, v1.4b[0]\n"
- ".word 0x4f82e139 // sdot v25.4s, v9.16b, v2.4b[0]\n"
- ".word 0x4f83e13d // sdot v29.4s, v9.16b, v3.4b[0]\n"
- "ldr q9, [%[b_ptr0], #0x10]\n"
- ".word 0x4f80e152 // sdot v18.4s, v10.16b, v0.4b[0]\n"
- ".word 0x4f81e156 // sdot v22.4s, v10.16b, v1.4b[0]\n"
- ".word 0x4f82e15a // sdot v26.4s, v10.16b, v2.4b[0]\n"
- ".word 0x4f83e15e // sdot v30.4s, v10.16b, v3.4b[0]\n"
- "ldr q10, [%[b_ptr0], #0x20]\n"
- ".word 0x4f80e173 // sdot v19.4s, v11.16b, v0.4b[0]\n"
- ".word 0x4f81e177 // sdot v23.4s, v11.16b, v1.4b[0]\n"
- ".word 0x4f82e17b // sdot v27.4s, v11.16b, v2.4b[0]\n"
- ".word 0x4f83e17f // sdot v31.4s, v11.16b, v3.4b[0]\n"
- "ldr q11, [%[b_ptr0], #0x30]\n"
- ".word 0x4fa0e190 // sdot v16.4s, v12.16b, v0.4b[1]\n"
- ".word 0x4fa1e194 // sdot v20.4s, v12.16b, v1.4b[1]\n"
- ".word 0x4fa2e198 // sdot v24.4s, v12.16b, v2.4b[1]\n"
- ".word 0x4fa3e19c // sdot v28.4s, v12.16b, v3.4b[1]\n"
- "ldr q12, [%[b_ptr0], #0x40]\n"
- ".word 0x4fa0e1b1 // sdot v17.4s, v13.16b, v0.4b[1]\n"
- ".word 0x4fa1e1b5 // sdot v21.4s, v13.16b, v1.4b[1]\n"
- ".word 0x4fa2e1b9 // sdot v25.4s, v13.16b, v2.4b[1]\n"
- ".word 0x4fa3e1bd // sdot v29.4s, v13.16b, v3.4b[1]\n"
- "ldr q13, [%[b_ptr0], #0x50]\n"
- ".word 0x4fa0e1d2 // sdot v18.4s, v14.16b, v0.4b[1]\n"
- ".word 0x4fa1e1d6 // sdot v22.4s, v14.16b, v1.4b[1]\n"
- ".word 0x4fa2e1da // sdot v26.4s, v14.16b, v2.4b[1]\n"
- ".word 0x4fa3e1de // sdot v30.4s, v14.16b, v3.4b[1]\n"
- "ldr q14, [%[b_ptr0], #0x60]\n"
- ".word 0x4fa0e1f3 // sdot v19.4s, v15.16b, v0.4b[1]\n"
- ".word 0x4fa1e1f7 // sdot v23.4s, v15.16b, v1.4b[1]\n"
- ".word 0x4fa2e1fb // sdot v27.4s, v15.16b, v2.4b[1]\n"
- ".word 0x4fa3e1ff // sdot v31.4s, v15.16b, v3.4b[1]\n"
- "ldr q15, [%[b_ptr0], #0x70]\n"
- ".word 0x4f80e910 // sdot v16.4s, v8.16b, v0.4b[2]\n"
- ".word 0x4f81e914 // sdot v20.4s, v8.16b, v1.4b[2]\n"
- ".word 0x4f82e918 // sdot v24.4s, v8.16b, v2.4b[2]\n"
- ".word 0x4f83e91c // sdot v28.4s, v8.16b, v3.4b[2]\n"
- ".word 0x4f80e931 // sdot v17.4s, v9.16b, v0.4b[2]\n"
- ".word 0x4f81e935 // sdot v21.4s, v9.16b, v1.4b[2]\n"
- ".word 0x4f82e939 // sdot v25.4s, v9.16b, v2.4b[2]\n"
- ".word 0x4f83e93d // sdot v29.4s, v9.16b, v3.4b[2]\n"
- ".word 0x4f80e952 // sdot v18.4s, v10.16b, v0.4b[2]\n"
- ".word 0x4f81e956 // sdot v22.4s, v10.16b, v1.4b[2]\n"
- ".word 0x4f82e95a // sdot v26.4s, v10.16b, v2.4b[2]\n"
- ".word 0x4f83e95e // sdot v30.4s, v10.16b, v3.4b[2]\n"
- ".word 0x4f80e973 // sdot v19.4s, v11.16b, v0.4b[2]\n"
- ".word 0x4f81e977 // sdot v23.4s, v11.16b, v1.4b[2]\n"
- ".word 0x4f82e97b // sdot v27.4s, v11.16b, v2.4b[2]\n"
- ".word 0x4f83e97f // sdot v31.4s, v11.16b, v3.4b[2]\n"
- ".word 0x4fa0e990 // sdot v16.4s, v12.16b, v0.4b[3]\n"
- ".word 0x4fa1e994 // sdot v20.4s, v12.16b, v1.4b[3]\n"
- ".word 0x4fa2e998 // sdot v24.4s, v12.16b, v2.4b[3]\n"
- ".word 0x4fa3e99c // sdot v28.4s, v12.16b, v3.4b[3]\n"
- ".word 0x4fa0e9b1 // sdot v17.4s, v13.16b, v0.4b[3]\n"
- ".word 0x4fa1e9b5 // sdot v21.4s, v13.16b, v1.4b[3]\n"
- ".word 0x4fa2e9b9 // sdot v25.4s, v13.16b, v2.4b[3]\n"
- ".word 0x4fa3e9bd // sdot v29.4s, v13.16b, v3.4b[3]\n"
- ".word 0x4fa0e9d2 // sdot v18.4s, v14.16b, v0.4b[3]\n"
- ".word 0x4fa1e9d6 // sdot v22.4s, v14.16b, v1.4b[3]\n"
- ".word 0x4fa2e9da // sdot v26.4s, v14.16b, v2.4b[3]\n"
- ".word 0x4fa3e9de // sdot v30.4s, v14.16b, v3.4b[3]\n"
- ".word 0x4fa0e9f3 // sdot v19.4s, v15.16b, v0.4b[3]\n"
- ".word 0x4fa1e9f7 // sdot v23.4s, v15.16b, v1.4b[3]\n"
- ".word 0x4fa2e9fb // sdot v27.4s, v15.16b, v2.4b[3]\n"
- ".word 0x4fa3e9ff // sdot v31.4s, v15.16b, v3.4b[3]\n"
- "5:\n"
- "str q16, [%[c_ptr0]]\n"
- "str q17, [%[c_ptr0], #0x10]\n"
- "str q18, [%[c_ptr0], #0x20]\n"
- "str q19, [%[c_ptr0], #0x30]\n"
- "add %[c_ptr0], %[c_ptr0], #0x40\n"
- "str q20, [c_ptr1]\n"
- "str q21, [c_ptr1, #0x10]\n"
- "str q22, [c_ptr1, #0x20]\n"
- "str q23, [c_ptr1, #0x30]\n"
- "str q24, [c_ptr2]\n"
- "str q25, [c_ptr2, #0x10]\n"
- "str q26, [c_ptr2, #0x20]\n"
- "str q27, [c_ptr2, #0x30]\n"
- "str q28, [c_ptr3]\n"
- "str q29, [c_ptr3, #0x10]\n"
- "str q30, [c_ptr3, #0x20]\n"
- "str q31, [c_ptr3, #0x30]\n"
- ".unreq a_ptr1\n"
- ".unreq a_ptr2\n"
- ".unreq a_ptr3\n"
- ".unreq c_ptr1\n"
- ".unreq c_ptr2\n"
- ".unreq c_ptr3\n"
- : [a_ptr0] "+r" (a_ptr0), [b_ptr0] "+r" (b_ptr0), [c_ptr0] "+r" (c_ptr0), [loops] "+r" (loops), [regs] "+r" (regs)
- : [betaptr] "r" (betaptr), [width] "r" (width), [beta0] "r" (beta0), [lda] "r" (ldab), [ldc] "r" (ldcb)
- : "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x0", "x1", "x2", "x3", "x4", "x5", "cc", "memory"
+ "a_ptr1 .req X0\n"
+ "a_ptr2 .req X1\n"
+ "a_ptr3 .req X2\n"
+ "c_ptr1 .req X3\n"
+ "c_ptr2 .req X4\n"
+ "c_ptr3 .req X5\n"
+ "add a_ptr1, %[a_ptr0], %[lda]\n"
+ "add c_ptr1, %[c_ptr0], %[ldc]\n"
+ "add a_ptr2, a_ptr1, %[lda]\n"
+ "add c_ptr2, c_ptr1, %[ldc]\n"
+ "add a_ptr3, a_ptr2, %[lda]\n"
+ "add c_ptr3, c_ptr2, %[ldc]\n"
+ "cbz %[beta0], 1f\n"
+ "movi v16.4s, #0\n"
+ "ldr q0, [%[a_ptr0]]\n"
+ "movi v17.4s, #0\n"
+ "ldr q1, [a_ptr1]\n"
+ "movi v18.4s, #0\n"
+ "ldr q2, [a_ptr2]\n"
+ "movi v19.4s, #0\n"
+ "ldr q3, [a_ptr3]\n"
+ "movi v20.4s, #0\n"
+ "ldr q8, [%[b_ptr0]]\n"
+ "movi v21.4s, #0\n"
+ "ldr q9, [%[b_ptr0], #0x10]\n"
+ "movi v22.4s, #0\n"
+ "ldr q10, [%[b_ptr0], #0x20]\n"
+ "movi v23.4s, #0\n"
+ "ldr q11, [%[b_ptr0], #0x30]\n"
+ "movi v24.4s, #0\n"
+ "ldr q12, [%[b_ptr0], #0x40]\n"
+ "movi v25.4s, #0\n"
+ "ldr q13, [%[b_ptr0], #0x50]\n"
+ "movi v26.4s, #0\n"
+ "ldr q14, [%[b_ptr0], #0x60]\n"
+ "movi v27.4s, #0\n"
+ "add %[a_ptr0], %[a_ptr0], #0x10\n"
+ "movi v28.4s, #0\n"
+ "add a_ptr1, a_ptr1, #0x10\n"
+ "movi v29.4s, #0\n"
+ "add a_ptr2, a_ptr2, #0x10\n"
+ "movi v30.4s, #0\n"
+ "add a_ptr3, a_ptr3, #0x10\n"
+ "movi v31.4s, #0\n"
+ "add %[b_ptr0], %[b_ptr0], #0x80\n"
+ "cbz %[loops], 2f\n"
+ "b 3f\n"
+ "1:\n"
+ "ld1r {v15.4s}, [%[betaptr]]\n"
+ "ldr q16, [%[c_ptr0]]\n"
+ "ldr q17, [%[c_ptr0], #0x10]\n"
+ "ldr q18, [%[c_ptr0], #0x20]\n"
+ "ldr q19, [%[c_ptr0], #0x30]\n"
+ "mul v16.4s, v16.4s, v15.4s\n"
+ "ldr q20, [c_ptr1]\n"
+ "mul v17.4s, v17.4s, v15.4s\n"
+ "ldr q21, [c_ptr1, #0x10]\n"
+ "mul v18.4s, v18.4s, v15.4s\n"
+ "ldr q22, [c_ptr1, #0x20]\n"
+ "mul v19.4s, v19.4s, v15.4s\n"
+ "ldr q23, [c_ptr1, #0x30]\n"
+ "mul v20.4s, v20.4s, v15.4s\n"
+ "ldr q24, [c_ptr2]\n"
+ "mul v21.4s, v21.4s, v15.4s\n"
+ "ldr q25, [c_ptr2, #0x10]\n"
+ "mul v22.4s, v22.4s, v15.4s\n"
+ "ldr q26, [c_ptr2, #0x20]\n"
+ "mul v23.4s, v23.4s, v15.4s\n"
+ "ldr q27, [c_ptr2, #0x30]\n"
+ "mul v24.4s, v24.4s, v15.4s\n"
+ "ldr q28, [c_ptr3]\n"
+ "mul v25.4s, v25.4s, v15.4s\n"
+ "ldr q29, [c_ptr3, #0x10]\n"
+ "mul v26.4s, v26.4s, v15.4s\n"
+ "ldr q30, [c_ptr3, #0x20]\n"
+ "mul v27.4s, v27.4s, v15.4s\n"
+ "ldr q31, [c_ptr3, #0x30]\n"
+ "mul v28.4s, v28.4s, v15.4s\n"
+ "ldr q0, [%[a_ptr0]]\n"
+ "mul v29.4s, v29.4s, v15.4s\n"
+ "ldr q1, [a_ptr1]\n"
+ "mul v30.4s, v30.4s, v15.4s\n"
+ "ldr q2, [a_ptr2]\n"
+ "mul v31.4s, v31.4s, v15.4s\n"
+ "ldr q3, [a_ptr3]\n"
+ "ldr q8, [%[b_ptr0]]\n"
+ "add %[a_ptr0], %[a_ptr0], #0x10\n"
+ "ldr q9, [%[b_ptr0], #0x10]\n"
+ "add a_ptr1, a_ptr1, #0x10\n"
+ "ldr q10, [%[b_ptr0], #0x20]\n"
+ "add a_ptr2, a_ptr2, #0x10\n"
+ "ldr q11, [%[b_ptr0], #0x30]\n"
+ "add a_ptr3, a_ptr3, #0x10\n"
+ "ldr q12, [%[b_ptr0], #0x40]\n"
+ "ldr q13, [%[b_ptr0], #0x50]\n"
+ "ldr q14, [%[b_ptr0], #0x60]\n"
+ "add %[b_ptr0], %[b_ptr0], #0x80\n"
+ "cbz %[loops], 2f\n"
+ "3:\n"
+ ".word 0x4f80e110 // sdot v16.4s, v8.16b, v0.4b[0]\n"
+ "ldr q15, [%[b_ptr0], #-0x10]\n"
+ ".word 0x4f81e114 // sdot v20.4s, v8.16b, v1.4b[0]\n"
+ "ldr q4, [%[a_ptr0]]\n"
+ ".word 0x4f82e118 // sdot v24.4s, v8.16b, v2.4b[0]\n"
+ "ldr q5, [a_ptr1]\n"
+ ".word 0x4f83e11c // sdot v28.4s, v8.16b, v3.4b[0]\n"
+ "ldr q6, [a_ptr2]\n"
+ ".word 0x4f80e131 // sdot v17.4s, v9.16b, v0.4b[0]\n"
+ "ldr q7, [a_ptr3]\n"
+ ".word 0x4f81e135 // sdot v21.4s, v9.16b, v1.4b[0]\n"
+ "ldr q8, [%[b_ptr0]]\n"
+ ".word 0x4f82e139 // sdot v25.4s, v9.16b, v2.4b[0]\n"
+ "subs %[loops], %[loops], #0x1\n"
+ ".word 0x4f83e13d // sdot v29.4s, v9.16b, v3.4b[0]\n"
+ "ldr q9, [%[b_ptr0], #0x10]\n"
+ ".word 0x4f80e152 // sdot v18.4s, v10.16b, v0.4b[0]\n"
+ "prfm PLDL1KEEP, [%[a_ptr0], #0x40]\n"
+ ".word 0x4f81e156 // sdot v22.4s, v10.16b, v1.4b[0]\n"
+ "add %[a_ptr0], %[a_ptr0], #0x20\n"
+ ".word 0x4f82e15a // sdot v26.4s, v10.16b, v2.4b[0]\n"
+ "add a_ptr1, a_ptr1, #0x20\n"
+ ".word 0x4f83e15e // sdot v30.4s, v10.16b, v3.4b[0]\n"
+ "ldr q10, [%[b_ptr0], #0x20]\n"
+ ".word 0x4f80e173 // sdot v19.4s, v11.16b, v0.4b[0]\n"
+ "add a_ptr2, a_ptr2, #0x20\n"
+ ".word 0x4f81e177 // sdot v23.4s, v11.16b, v1.4b[0]\n"
+ "add a_ptr3, a_ptr3, #0x20\n"
+ ".word 0x4f82e17b // sdot v27.4s, v11.16b, v2.4b[0]\n"
+ "prfm PLDL1KEEP, [a_ptr1, #0x40]\n"
+ ".word 0x4f83e17f // sdot v31.4s, v11.16b, v3.4b[0]\n"
+ "ldr q11, [%[b_ptr0], #0x30]\n"
+ ".word 0x4fa0e190 // sdot v16.4s, v12.16b, v0.4b[1]\n"
+ "prfm PLDL1KEEP, [a_ptr2, #0x40]\n"
+ ".word 0x4fa1e194 // sdot v20.4s, v12.16b, v1.4b[1]\n"
+ "prfm PLDL1KEEP, [a_ptr3, #0x40]\n"
+ ".word 0x4fa2e198 // sdot v24.4s, v12.16b, v2.4b[1]\n"
+ ".word 0x4fa3e19c // sdot v28.4s, v12.16b, v3.4b[1]\n"
+ "ldr q12, [%[b_ptr0], #0x40]\n"
+ ".word 0x4fa0e1b1 // sdot v17.4s, v13.16b, v0.4b[1]\n"
+ ".word 0x4fa1e1b5 // sdot v21.4s, v13.16b, v1.4b[1]\n"
+ ".word 0x4fa2e1b9 // sdot v25.4s, v13.16b, v2.4b[1]\n"
+ ".word 0x4fa3e1bd // sdot v29.4s, v13.16b, v3.4b[1]\n"
+ "ldr q13, [%[b_ptr0], #0x50]\n"
+ ".word 0x4fa0e1d2 // sdot v18.4s, v14.16b, v0.4b[1]\n"
+ ".word 0x4fa1e1d6 // sdot v22.4s, v14.16b, v1.4b[1]\n"
+ ".word 0x4fa2e1da // sdot v26.4s, v14.16b, v2.4b[1]\n"
+ ".word 0x4fa3e1de // sdot v30.4s, v14.16b, v3.4b[1]\n"
+ "ldr q14, [%[b_ptr0], #0x60]\n"
+ ".word 0x4fa0e1f3 // sdot v19.4s, v15.16b, v0.4b[1]\n"
+ ".word 0x4fa1e1f7 // sdot v23.4s, v15.16b, v1.4b[1]\n"
+ ".word 0x4fa2e1fb // sdot v27.4s, v15.16b, v2.4b[1]\n"
+ ".word 0x4fa3e1ff // sdot v31.4s, v15.16b, v3.4b[1]\n"
+ "ldr q15, [%[b_ptr0], #0x70]\n"
+ ".word 0x4f80e910 // sdot v16.4s, v8.16b, v0.4b[2]\n"
+ "add %[b_ptr0], %[b_ptr0], #0x100\n"
+ ".word 0x4f81e914 // sdot v20.4s, v8.16b, v1.4b[2]\n"
+ ".word 0x4f82e918 // sdot v24.4s, v8.16b, v2.4b[2]\n"
+ ".word 0x4f83e91c // sdot v28.4s, v8.16b, v3.4b[2]\n"
+ "ldr q8, [%[b_ptr0], #-0x80]\n"
+ ".word 0x4f80e931 // sdot v17.4s, v9.16b, v0.4b[2]\n"
+ ".word 0x4f81e935 // sdot v21.4s, v9.16b, v1.4b[2]\n"
+ ".word 0x4f82e939 // sdot v25.4s, v9.16b, v2.4b[2]\n"
+ ".word 0x4f83e93d // sdot v29.4s, v9.16b, v3.4b[2]\n"
+ "ldr q9, [%[b_ptr0], #-0x70]\n"
+ ".word 0x4f80e952 // sdot v18.4s, v10.16b, v0.4b[2]\n"
+ ".word 0x4f81e956 // sdot v22.4s, v10.16b, v1.4b[2]\n"
+ ".word 0x4f82e95a // sdot v26.4s, v10.16b, v2.4b[2]\n"
+ ".word 0x4f83e95e // sdot v30.4s, v10.16b, v3.4b[2]\n"
+ "ldr q10, [%[b_ptr0], #-0x60]\n"
+ ".word 0x4f80e973 // sdot v19.4s, v11.16b, v0.4b[2]\n"
+ ".word 0x4f81e977 // sdot v23.4s, v11.16b, v1.4b[2]\n"
+ ".word 0x4f82e97b // sdot v27.4s, v11.16b, v2.4b[2]\n"
+ ".word 0x4f83e97f // sdot v31.4s, v11.16b, v3.4b[2]\n"
+ "ldr q11, [%[b_ptr0], #-0x50]\n"
+ ".word 0x4fa0e990 // sdot v16.4s, v12.16b, v0.4b[3]\n"
+ ".word 0x4fa1e994 // sdot v20.4s, v12.16b, v1.4b[3]\n"
+ ".word 0x4fa2e998 // sdot v24.4s, v12.16b, v2.4b[3]\n"
+ ".word 0x4fa3e99c // sdot v28.4s, v12.16b, v3.4b[3]\n"
+ "ldr q12, [%[b_ptr0], #-0x40]\n"
+ ".word 0x4fa0e9b1 // sdot v17.4s, v13.16b, v0.4b[3]\n"
+ ".word 0x4fa1e9b5 // sdot v21.4s, v13.16b, v1.4b[3]\n"
+ ".word 0x4fa2e9b9 // sdot v25.4s, v13.16b, v2.4b[3]\n"
+ ".word 0x4fa3e9bd // sdot v29.4s, v13.16b, v3.4b[3]\n"
+ "ldr q13, [%[b_ptr0], #-0x30]\n"
+ ".word 0x4fa0e9d2 // sdot v18.4s, v14.16b, v0.4b[3]\n"
+ ".word 0x4fa1e9d6 // sdot v22.4s, v14.16b, v1.4b[3]\n"
+ ".word 0x4fa2e9da // sdot v26.4s, v14.16b, v2.4b[3]\n"
+ ".word 0x4fa3e9de // sdot v30.4s, v14.16b, v3.4b[3]\n"
+ "ldr q14, [%[b_ptr0], #-0x20]\n"
+ ".word 0x4fa0e9f3 // sdot v19.4s, v15.16b, v0.4b[3]\n"
+ "ldr q0, [%[a_ptr0], #-0x10]\n"
+ ".word 0x4fa1e9f7 // sdot v23.4s, v15.16b, v1.4b[3]\n"
+ "ldr q1, [a_ptr1, #-0x10]\n"
+ ".word 0x4fa2e9fb // sdot v27.4s, v15.16b, v2.4b[3]\n"
+ "ldr q2, [a_ptr2, #-0x10]\n"
+ ".word 0x4fa3e9ff // sdot v31.4s, v15.16b, v3.4b[3]\n"
+ "ldr q15, [%[b_ptr0], #-0x10]\n"
+ ".word 0x4f84e110 // sdot v16.4s, v8.16b, v4.4b[0]\n"
+ "ldr q3, [a_ptr3, #-0x10]\n"
+ ".word 0x4f85e114 // sdot v20.4s, v8.16b, v5.4b[0]\n"
+ ".word 0x4f86e118 // sdot v24.4s, v8.16b, v6.4b[0]\n"
+ ".word 0x4f87e11c // sdot v28.4s, v8.16b, v7.4b[0]\n"
+ "ldr q8, [%[b_ptr0]]\n"
+ ".word 0x4f84e131 // sdot v17.4s, v9.16b, v4.4b[0]\n"
+ ".word 0x4f85e135 // sdot v21.4s, v9.16b, v5.4b[0]\n"
+ ".word 0x4f86e139 // sdot v25.4s, v9.16b, v6.4b[0]\n"
+ ".word 0x4f87e13d // sdot v29.4s, v9.16b, v7.4b[0]\n"
+ "ldr q9, [%[b_ptr0], #0x10]\n"
+ ".word 0x4f84e152 // sdot v18.4s, v10.16b, v4.4b[0]\n"
+ ".word 0x4f85e156 // sdot v22.4s, v10.16b, v5.4b[0]\n"
+ ".word 0x4f86e15a // sdot v26.4s, v10.16b, v6.4b[0]\n"
+ ".word 0x4f87e15e // sdot v30.4s, v10.16b, v7.4b[0]\n"
+ "ldr q10, [%[b_ptr0], #0x20]\n"
+ ".word 0x4f84e173 // sdot v19.4s, v11.16b, v4.4b[0]\n"
+ ".word 0x4f85e177 // sdot v23.4s, v11.16b, v5.4b[0]\n"
+ ".word 0x4f86e17b // sdot v27.4s, v11.16b, v6.4b[0]\n"
+ ".word 0x4f87e17f // sdot v31.4s, v11.16b, v7.4b[0]\n"
+ "ldr q11, [%[b_ptr0], #0x30]\n"
+ ".word 0x4fa4e190 // sdot v16.4s, v12.16b, v4.4b[1]\n"
+ ".word 0x4fa5e194 // sdot v20.4s, v12.16b, v5.4b[1]\n"
+ ".word 0x4fa6e198 // sdot v24.4s, v12.16b, v6.4b[1]\n"
+ ".word 0x4fa7e19c // sdot v28.4s, v12.16b, v7.4b[1]\n"
+ "ldr q12, [%[b_ptr0], #0x40]\n"
+ ".word 0x4fa4e1b1 // sdot v17.4s, v13.16b, v4.4b[1]\n"
+ ".word 0x4fa5e1b5 // sdot v21.4s, v13.16b, v5.4b[1]\n"
+ ".word 0x4fa6e1b9 // sdot v25.4s, v13.16b, v6.4b[1]\n"
+ ".word 0x4fa7e1bd // sdot v29.4s, v13.16b, v7.4b[1]\n"
+ "ldr q13, [%[b_ptr0], #0x50]\n"
+ ".word 0x4fa4e1d2 // sdot v18.4s, v14.16b, v4.4b[1]\n"
+ ".word 0x4fa5e1d6 // sdot v22.4s, v14.16b, v5.4b[1]\n"
+ ".word 0x4fa6e1da // sdot v26.4s, v14.16b, v6.4b[1]\n"
+ ".word 0x4fa7e1de // sdot v30.4s, v14.16b, v7.4b[1]\n"
+ "ldr q14, [%[b_ptr0], #0x60]\n"
+ ".word 0x4fa4e1f3 // sdot v19.4s, v15.16b, v4.4b[1]\n"
+ ".word 0x4fa5e1f7 // sdot v23.4s, v15.16b, v5.4b[1]\n"
+ ".word 0x4fa6e1fb // sdot v27.4s, v15.16b, v6.4b[1]\n"
+ ".word 0x4fa7e1ff // sdot v31.4s, v15.16b, v7.4b[1]\n"
+ "ldr q15, [%[b_ptr0], #0x70]\n"
+ ".word 0x4f84e910 // sdot v16.4s, v8.16b, v4.4b[2]\n"
+ "add %[b_ptr0], %[b_ptr0], #0x100\n"
+ ".word 0x4f85e914 // sdot v20.4s, v8.16b, v5.4b[2]\n"
+ ".word 0x4f86e918 // sdot v24.4s, v8.16b, v6.4b[2]\n"
+ ".word 0x4f87e91c // sdot v28.4s, v8.16b, v7.4b[2]\n"
+ "ldr q8, [%[b_ptr0], #-0x80]\n"
+ ".word 0x4f84e931 // sdot v17.4s, v9.16b, v4.4b[2]\n"
+ ".word 0x4f85e935 // sdot v21.4s, v9.16b, v5.4b[2]\n"
+ ".word 0x4f86e939 // sdot v25.4s, v9.16b, v6.4b[2]\n"
+ ".word 0x4f87e93d // sdot v29.4s, v9.16b, v7.4b[2]\n"
+ "ldr q9, [%[b_ptr0], #-0x70]\n"
+ ".word 0x4f84e952 // sdot v18.4s, v10.16b, v4.4b[2]\n"
+ ".word 0x4f85e956 // sdot v22.4s, v10.16b, v5.4b[2]\n"
+ ".word 0x4f86e95a // sdot v26.4s, v10.16b, v6.4b[2]\n"
+ ".word 0x4f87e95e // sdot v30.4s, v10.16b, v7.4b[2]\n"
+ "ldr q10, [%[b_ptr0], #-0x60]\n"
+ ".word 0x4f84e973 // sdot v19.4s, v11.16b, v4.4b[2]\n"
+ ".word 0x4f85e977 // sdot v23.4s, v11.16b, v5.4b[2]\n"
+ ".word 0x4f86e97b // sdot v27.4s, v11.16b, v6.4b[2]\n"
+ ".word 0x4f87e97f // sdot v31.4s, v11.16b, v7.4b[2]\n"
+ "ldr q11, [%[b_ptr0], #-0x50]\n"
+ ".word 0x4fa4e990 // sdot v16.4s, v12.16b, v4.4b[3]\n"
+ ".word 0x4fa5e994 // sdot v20.4s, v12.16b, v5.4b[3]\n"
+ ".word 0x4fa6e998 // sdot v24.4s, v12.16b, v6.4b[3]\n"
+ ".word 0x4fa7e99c // sdot v28.4s, v12.16b, v7.4b[3]\n"
+ "ldr q12, [%[b_ptr0], #-0x40]\n"
+ ".word 0x4fa4e9b1 // sdot v17.4s, v13.16b, v4.4b[3]\n"
+ ".word 0x4fa5e9b5 // sdot v21.4s, v13.16b, v5.4b[3]\n"
+ ".word 0x4fa6e9b9 // sdot v25.4s, v13.16b, v6.4b[3]\n"
+ ".word 0x4fa7e9bd // sdot v29.4s, v13.16b, v7.4b[3]\n"
+ "ldr q13, [%[b_ptr0], #-0x30]\n"
+ ".word 0x4fa4e9d2 // sdot v18.4s, v14.16b, v4.4b[3]\n"
+ ".word 0x4fa5e9d6 // sdot v22.4s, v14.16b, v5.4b[3]\n"
+ ".word 0x4fa6e9da // sdot v26.4s, v14.16b, v6.4b[3]\n"
+ ".word 0x4fa7e9de // sdot v30.4s, v14.16b, v7.4b[3]\n"
+ "ldr q14, [%[b_ptr0], #-0x20]\n"
+ ".word 0x4fa4e9f3 // sdot v19.4s, v15.16b, v4.4b[3]\n"
+ ".word 0x4fa5e9f7 // sdot v23.4s, v15.16b, v5.4b[3]\n"
+ ".word 0x4fa6e9fb // sdot v27.4s, v15.16b, v6.4b[3]\n"
+ ".word 0x4fa7e9ff // sdot v31.4s, v15.16b, v7.4b[3]\n"
+ "b.ne 3b\n"
+ "2:\n"
+ "ldr q15, [%[b_ptr0], #-0x10]\n"
+ "prfm PSTL1KEEP, [%[c_ptr0]]\n"
+ "prfm PSTL1KEEP, [c_ptr1]\n"
+ "prfm PSTL1KEEP, [c_ptr2]\n"
+ "prfm PSTL1KEEP, [c_ptr3]\n"
+ "cbz %[regs], 4f\n"
+ ".word 0x4f80e110 // sdot v16.4s, v8.16b, v0.4b[0]\n"
+ "ldr q4, [%[a_ptr0]]\n"
+ ".word 0x4f81e114 // sdot v20.4s, v8.16b, v1.4b[0]\n"
+ "ldr q5, [a_ptr1]\n"
+ ".word 0x4f82e118 // sdot v24.4s, v8.16b, v2.4b[0]\n"
+ "ldr q6, [a_ptr2]\n"
+ ".word 0x4f83e11c // sdot v28.4s, v8.16b, v3.4b[0]\n"
+ "ldr q7, [a_ptr3]\n"
+ ".word 0x4f80e131 // sdot v17.4s, v9.16b, v0.4b[0]\n"
+ "ldr q8, [%[b_ptr0]]\n"
+ ".word 0x4f81e135 // sdot v21.4s, v9.16b, v1.4b[0]\n"
+ "add %[a_ptr0], %[a_ptr0], #0x10\n"
+ ".word 0x4f82e139 // sdot v25.4s, v9.16b, v2.4b[0]\n"
+ "add a_ptr1, a_ptr1, #0x10\n"
+ ".word 0x4f83e13d // sdot v29.4s, v9.16b, v3.4b[0]\n"
+ "ldr q9, [%[b_ptr0], #0x10]\n"
+ ".word 0x4f80e152 // sdot v18.4s, v10.16b, v0.4b[0]\n"
+ "add a_ptr2, a_ptr2, #0x10\n"
+ ".word 0x4f81e156 // sdot v22.4s, v10.16b, v1.4b[0]\n"
+ "add a_ptr3, a_ptr3, #0x10\n"
+ ".word 0x4f82e15a // sdot v26.4s, v10.16b, v2.4b[0]\n"
+ ".word 0x4f83e15e // sdot v30.4s, v10.16b, v3.4b[0]\n"
+ "ldr q10, [%[b_ptr0], #0x20]\n"
+ ".word 0x4f80e173 // sdot v19.4s, v11.16b, v0.4b[0]\n"
+ ".word 0x4f81e177 // sdot v23.4s, v11.16b, v1.4b[0]\n"
+ ".word 0x4f82e17b // sdot v27.4s, v11.16b, v2.4b[0]\n"
+ ".word 0x4f83e17f // sdot v31.4s, v11.16b, v3.4b[0]\n"
+ "ldr q11, [%[b_ptr0], #0x30]\n"
+ ".word 0x4fa0e190 // sdot v16.4s, v12.16b, v0.4b[1]\n"
+ ".word 0x4fa1e194 // sdot v20.4s, v12.16b, v1.4b[1]\n"
+ ".word 0x4fa2e198 // sdot v24.4s, v12.16b, v2.4b[1]\n"
+ ".word 0x4fa3e19c // sdot v28.4s, v12.16b, v3.4b[1]\n"
+ "ldr q12, [%[b_ptr0], #0x40]\n"
+ ".word 0x4fa0e1b1 // sdot v17.4s, v13.16b, v0.4b[1]\n"
+ ".word 0x4fa1e1b5 // sdot v21.4s, v13.16b, v1.4b[1]\n"
+ ".word 0x4fa2e1b9 // sdot v25.4s, v13.16b, v2.4b[1]\n"
+ ".word 0x4fa3e1bd // sdot v29.4s, v13.16b, v3.4b[1]\n"
+ "ldr q13, [%[b_ptr0], #0x50]\n"
+ ".word 0x4fa0e1d2 // sdot v18.4s, v14.16b, v0.4b[1]\n"
+ ".word 0x4fa1e1d6 // sdot v22.4s, v14.16b, v1.4b[1]\n"
+ ".word 0x4fa2e1da // sdot v26.4s, v14.16b, v2.4b[1]\n"
+ ".word 0x4fa3e1de // sdot v30.4s, v14.16b, v3.4b[1]\n"
+ "ldr q14, [%[b_ptr0], #0x60]\n"
+ ".word 0x4fa0e1f3 // sdot v19.4s, v15.16b, v0.4b[1]\n"
+ ".word 0x4fa1e1f7 // sdot v23.4s, v15.16b, v1.4b[1]\n"
+ ".word 0x4fa2e1fb // sdot v27.4s, v15.16b, v2.4b[1]\n"
+ ".word 0x4fa3e1ff // sdot v31.4s, v15.16b, v3.4b[1]\n"
+ "ldr q15, [%[b_ptr0], #0x70]\n"
+ ".word 0x4f80e910 // sdot v16.4s, v8.16b, v0.4b[2]\n"
+ "add %[b_ptr0], %[b_ptr0], #0x100\n"
+ ".word 0x4f81e914 // sdot v20.4s, v8.16b, v1.4b[2]\n"
+ ".word 0x4f82e918 // sdot v24.4s, v8.16b, v2.4b[2]\n"
+ ".word 0x4f83e91c // sdot v28.4s, v8.16b, v3.4b[2]\n"
+ "ldr q8, [%[b_ptr0], #-0x80]\n"
+ ".word 0x4f80e931 // sdot v17.4s, v9.16b, v0.4b[2]\n"
+ ".word 0x4f81e935 // sdot v21.4s, v9.16b, v1.4b[2]\n"
+ ".word 0x4f82e939 // sdot v25.4s, v9.16b, v2.4b[2]\n"
+ ".word 0x4f83e93d // sdot v29.4s, v9.16b, v3.4b[2]\n"
+ "ldr q9, [%[b_ptr0], #-0x70]\n"
+ ".word 0x4f80e952 // sdot v18.4s, v10.16b, v0.4b[2]\n"
+ ".word 0x4f81e956 // sdot v22.4s, v10.16b, v1.4b[2]\n"
+ ".word 0x4f82e95a // sdot v26.4s, v10.16b, v2.4b[2]\n"
+ ".word 0x4f83e95e // sdot v30.4s, v10.16b, v3.4b[2]\n"
+ "ldr q10, [%[b_ptr0], #-0x60]\n"
+ ".word 0x4f80e973 // sdot v19.4s, v11.16b, v0.4b[2]\n"
+ ".word 0x4f81e977 // sdot v23.4s, v11.16b, v1.4b[2]\n"
+ ".word 0x4f82e97b // sdot v27.4s, v11.16b, v2.4b[2]\n"
+ ".word 0x4f83e97f // sdot v31.4s, v11.16b, v3.4b[2]\n"
+ "ldr q11, [%[b_ptr0], #-0x50]\n"
+ ".word 0x4fa0e990 // sdot v16.4s, v12.16b, v0.4b[3]\n"
+ ".word 0x4fa1e994 // sdot v20.4s, v12.16b, v1.4b[3]\n"
+ ".word 0x4fa2e998 // sdot v24.4s, v12.16b, v2.4b[3]\n"
+ ".word 0x4fa3e99c // sdot v28.4s, v12.16b, v3.4b[3]\n"
+ "ldr q12, [%[b_ptr0], #-0x40]\n"
+ ".word 0x4fa0e9b1 // sdot v17.4s, v13.16b, v0.4b[3]\n"
+ ".word 0x4fa1e9b5 // sdot v21.4s, v13.16b, v1.4b[3]\n"
+ ".word 0x4fa2e9b9 // sdot v25.4s, v13.16b, v2.4b[3]\n"
+ ".word 0x4fa3e9bd // sdot v29.4s, v13.16b, v3.4b[3]\n"
+ "ldr q13, [%[b_ptr0], #-0x30]\n"
+ ".word 0x4fa0e9d2 // sdot v18.4s, v14.16b, v0.4b[3]\n"
+ ".word 0x4fa1e9d6 // sdot v22.4s, v14.16b, v1.4b[3]\n"
+ ".word 0x4fa2e9da // sdot v26.4s, v14.16b, v2.4b[3]\n"
+ ".word 0x4fa3e9de // sdot v30.4s, v14.16b, v3.4b[3]\n"
+ "ldr q14, [%[b_ptr0], #-0x20]\n"
+ ".word 0x4fa0e9f3 // sdot v19.4s, v15.16b, v0.4b[3]\n"
+ ".word 0x4fa1e9f7 // sdot v23.4s, v15.16b, v1.4b[3]\n"
+ ".word 0x4fa2e9fb // sdot v27.4s, v15.16b, v2.4b[3]\n"
+ ".word 0x4fa3e9ff // sdot v31.4s, v15.16b, v3.4b[3]\n"
+ "ldr q15, [%[b_ptr0], #-0x10]\n"
+ ".word 0x4f84e110 // sdot v16.4s, v8.16b, v4.4b[0]\n"
+ ".word 0x4f85e114 // sdot v20.4s, v8.16b, v5.4b[0]\n"
+ ".word 0x4f86e118 // sdot v24.4s, v8.16b, v6.4b[0]\n"
+ ".word 0x4f87e11c // sdot v28.4s, v8.16b, v7.4b[0]\n"
+ "ldr q8, [%[b_ptr0]]\n"
+ ".word 0x4f84e131 // sdot v17.4s, v9.16b, v4.4b[0]\n"
+ ".word 0x4f85e135 // sdot v21.4s, v9.16b, v5.4b[0]\n"
+ ".word 0x4f86e139 // sdot v25.4s, v9.16b, v6.4b[0]\n"
+ ".word 0x4f87e13d // sdot v29.4s, v9.16b, v7.4b[0]\n"
+ "ldr q9, [%[b_ptr0], #0x10]\n"
+ ".word 0x4f84e152 // sdot v18.4s, v10.16b, v4.4b[0]\n"
+ ".word 0x4f85e156 // sdot v22.4s, v10.16b, v5.4b[0]\n"
+ ".word 0x4f86e15a // sdot v26.4s, v10.16b, v6.4b[0]\n"
+ ".word 0x4f87e15e // sdot v30.4s, v10.16b, v7.4b[0]\n"
+ "ldr q10, [%[b_ptr0], #0x20]\n"
+ ".word 0x4f84e173 // sdot v19.4s, v11.16b, v4.4b[0]\n"
+ ".word 0x4f85e177 // sdot v23.4s, v11.16b, v5.4b[0]\n"
+ ".word 0x4f86e17b // sdot v27.4s, v11.16b, v6.4b[0]\n"
+ ".word 0x4f87e17f // sdot v31.4s, v11.16b, v7.4b[0]\n"
+ "ldr q11, [%[b_ptr0], #0x30]\n"
+ ".word 0x4fa4e190 // sdot v16.4s, v12.16b, v4.4b[1]\n"
+ ".word 0x4fa5e194 // sdot v20.4s, v12.16b, v5.4b[1]\n"
+ ".word 0x4fa6e198 // sdot v24.4s, v12.16b, v6.4b[1]\n"
+ ".word 0x4fa7e19c // sdot v28.4s, v12.16b, v7.4b[1]\n"
+ "ldr q12, [%[b_ptr0], #0x40]\n"
+ ".word 0x4fa4e1b1 // sdot v17.4s, v13.16b, v4.4b[1]\n"
+ ".word 0x4fa5e1b5 // sdot v21.4s, v13.16b, v5.4b[1]\n"
+ ".word 0x4fa6e1b9 // sdot v25.4s, v13.16b, v6.4b[1]\n"
+ ".word 0x4fa7e1bd // sdot v29.4s, v13.16b, v7.4b[1]\n"
+ "ldr q13, [%[b_ptr0], #0x50]\n"
+ ".word 0x4fa4e1d2 // sdot v18.4s, v14.16b, v4.4b[1]\n"
+ ".word 0x4fa5e1d6 // sdot v22.4s, v14.16b, v5.4b[1]\n"
+ ".word 0x4fa6e1da // sdot v26.4s, v14.16b, v6.4b[1]\n"
+ ".word 0x4fa7e1de // sdot v30.4s, v14.16b, v7.4b[1]\n"
+ "ldr q14, [%[b_ptr0], #0x60]\n"
+ ".word 0x4fa4e1f3 // sdot v19.4s, v15.16b, v4.4b[1]\n"
+ ".word 0x4fa5e1f7 // sdot v23.4s, v15.16b, v5.4b[1]\n"
+ ".word 0x4fa6e1fb // sdot v27.4s, v15.16b, v6.4b[1]\n"
+ ".word 0x4fa7e1ff // sdot v31.4s, v15.16b, v7.4b[1]\n"
+ "ldr q15, [%[b_ptr0], #0x70]\n"
+ ".word 0x4f84e910 // sdot v16.4s, v8.16b, v4.4b[2]\n"
+ "add %[b_ptr0], %[b_ptr0], #0x80\n"
+ ".word 0x4f85e914 // sdot v20.4s, v8.16b, v5.4b[2]\n"
+ ".word 0x4f86e918 // sdot v24.4s, v8.16b, v6.4b[2]\n"
+ ".word 0x4f87e91c // sdot v28.4s, v8.16b, v7.4b[2]\n"
+ ".word 0x4f84e931 // sdot v17.4s, v9.16b, v4.4b[2]\n"
+ ".word 0x4f85e935 // sdot v21.4s, v9.16b, v5.4b[2]\n"
+ ".word 0x4f86e939 // sdot v25.4s, v9.16b, v6.4b[2]\n"
+ ".word 0x4f87e93d // sdot v29.4s, v9.16b, v7.4b[2]\n"
+ ".word 0x4f84e952 // sdot v18.4s, v10.16b, v4.4b[2]\n"
+ ".word 0x4f85e956 // sdot v22.4s, v10.16b, v5.4b[2]\n"
+ ".word 0x4f86e95a // sdot v26.4s, v10.16b, v6.4b[2]\n"
+ ".word 0x4f87e95e // sdot v30.4s, v10.16b, v7.4b[2]\n"
+ ".word 0x4f84e973 // sdot v19.4s, v11.16b, v4.4b[2]\n"
+ ".word 0x4f85e977 // sdot v23.4s, v11.16b, v5.4b[2]\n"
+ ".word 0x4f86e97b // sdot v27.4s, v11.16b, v6.4b[2]\n"
+ ".word 0x4f87e97f // sdot v31.4s, v11.16b, v7.4b[2]\n"
+ ".word 0x4fa4e990 // sdot v16.4s, v12.16b, v4.4b[3]\n"
+ ".word 0x4fa5e994 // sdot v20.4s, v12.16b, v5.4b[3]\n"
+ ".word 0x4fa6e998 // sdot v24.4s, v12.16b, v6.4b[3]\n"
+ ".word 0x4fa7e99c // sdot v28.4s, v12.16b, v7.4b[3]\n"
+ ".word 0x4fa4e9b1 // sdot v17.4s, v13.16b, v4.4b[3]\n"
+ ".word 0x4fa5e9b5 // sdot v21.4s, v13.16b, v5.4b[3]\n"
+ ".word 0x4fa6e9b9 // sdot v25.4s, v13.16b, v6.4b[3]\n"
+ ".word 0x4fa7e9bd // sdot v29.4s, v13.16b, v7.4b[3]\n"
+ ".word 0x4fa4e9d2 // sdot v18.4s, v14.16b, v4.4b[3]\n"
+ ".word 0x4fa5e9d6 // sdot v22.4s, v14.16b, v5.4b[3]\n"
+ ".word 0x4fa6e9da // sdot v26.4s, v14.16b, v6.4b[3]\n"
+ ".word 0x4fa7e9de // sdot v30.4s, v14.16b, v7.4b[3]\n"
+ ".word 0x4fa4e9f3 // sdot v19.4s, v15.16b, v4.4b[3]\n"
+ ".word 0x4fa5e9f7 // sdot v23.4s, v15.16b, v5.4b[3]\n"
+ ".word 0x4fa6e9fb // sdot v27.4s, v15.16b, v6.4b[3]\n"
+ ".word 0x4fa7e9ff // sdot v31.4s, v15.16b, v7.4b[3]\n"
+ "b 5f\n"
+ "4:\n"
+ ".word 0x4f80e110 // sdot v16.4s, v8.16b, v0.4b[0]\n"
+ ".word 0x4f81e114 // sdot v20.4s, v8.16b, v1.4b[0]\n"
+ ".word 0x4f82e118 // sdot v24.4s, v8.16b, v2.4b[0]\n"
+ ".word 0x4f83e11c // sdot v28.4s, v8.16b, v3.4b[0]\n"
+ "ldr q8, [%[b_ptr0]]\n"
+ ".word 0x4f80e131 // sdot v17.4s, v9.16b, v0.4b[0]\n"
+ ".word 0x4f81e135 // sdot v21.4s, v9.16b, v1.4b[0]\n"
+ ".word 0x4f82e139 // sdot v25.4s, v9.16b, v2.4b[0]\n"
+ ".word 0x4f83e13d // sdot v29.4s, v9.16b, v3.4b[0]\n"
+ "ldr q9, [%[b_ptr0], #0x10]\n"
+ ".word 0x4f80e152 // sdot v18.4s, v10.16b, v0.4b[0]\n"
+ ".word 0x4f81e156 // sdot v22.4s, v10.16b, v1.4b[0]\n"
+ ".word 0x4f82e15a // sdot v26.4s, v10.16b, v2.4b[0]\n"
+ ".word 0x4f83e15e // sdot v30.4s, v10.16b, v3.4b[0]\n"
+ "ldr q10, [%[b_ptr0], #0x20]\n"
+ ".word 0x4f80e173 // sdot v19.4s, v11.16b, v0.4b[0]\n"
+ ".word 0x4f81e177 // sdot v23.4s, v11.16b, v1.4b[0]\n"
+ ".word 0x4f82e17b // sdot v27.4s, v11.16b, v2.4b[0]\n"
+ ".word 0x4f83e17f // sdot v31.4s, v11.16b, v3.4b[0]\n"
+ "ldr q11, [%[b_ptr0], #0x30]\n"
+ ".word 0x4fa0e190 // sdot v16.4s, v12.16b, v0.4b[1]\n"
+ ".word 0x4fa1e194 // sdot v20.4s, v12.16b, v1.4b[1]\n"
+ ".word 0x4fa2e198 // sdot v24.4s, v12.16b, v2.4b[1]\n"
+ ".word 0x4fa3e19c // sdot v28.4s, v12.16b, v3.4b[1]\n"
+ "ldr q12, [%[b_ptr0], #0x40]\n"
+ ".word 0x4fa0e1b1 // sdot v17.4s, v13.16b, v0.4b[1]\n"
+ ".word 0x4fa1e1b5 // sdot v21.4s, v13.16b, v1.4b[1]\n"
+ ".word 0x4fa2e1b9 // sdot v25.4s, v13.16b, v2.4b[1]\n"
+ ".word 0x4fa3e1bd // sdot v29.4s, v13.16b, v3.4b[1]\n"
+ "ldr q13, [%[b_ptr0], #0x50]\n"
+ ".word 0x4fa0e1d2 // sdot v18.4s, v14.16b, v0.4b[1]\n"
+ ".word 0x4fa1e1d6 // sdot v22.4s, v14.16b, v1.4b[1]\n"
+ ".word 0x4fa2e1da // sdot v26.4s, v14.16b, v2.4b[1]\n"
+ ".word 0x4fa3e1de // sdot v30.4s, v14.16b, v3.4b[1]\n"
+ "ldr q14, [%[b_ptr0], #0x60]\n"
+ ".word 0x4fa0e1f3 // sdot v19.4s, v15.16b, v0.4b[1]\n"
+ ".word 0x4fa1e1f7 // sdot v23.4s, v15.16b, v1.4b[1]\n"
+ ".word 0x4fa2e1fb // sdot v27.4s, v15.16b, v2.4b[1]\n"
+ ".word 0x4fa3e1ff // sdot v31.4s, v15.16b, v3.4b[1]\n"
+ "ldr q15, [%[b_ptr0], #0x70]\n"
+ ".word 0x4f80e910 // sdot v16.4s, v8.16b, v0.4b[2]\n"
+ "add %[b_ptr0], %[b_ptr0], #0x80\n"
+ ".word 0x4f81e914 // sdot v20.4s, v8.16b, v1.4b[2]\n"
+ ".word 0x4f82e918 // sdot v24.4s, v8.16b, v2.4b[2]\n"
+ ".word 0x4f83e91c // sdot v28.4s, v8.16b, v3.4b[2]\n"
+ ".word 0x4f80e931 // sdot v17.4s, v9.16b, v0.4b[2]\n"
+ ".word 0x4f81e935 // sdot v21.4s, v9.16b, v1.4b[2]\n"
+ ".word 0x4f82e939 // sdot v25.4s, v9.16b, v2.4b[2]\n"
+ ".word 0x4f83e93d // sdot v29.4s, v9.16b, v3.4b[2]\n"
+ ".word 0x4f80e952 // sdot v18.4s, v10.16b, v0.4b[2]\n"
+ ".word 0x4f81e956 // sdot v22.4s, v10.16b, v1.4b[2]\n"
+ ".word 0x4f82e95a // sdot v26.4s, v10.16b, v2.4b[2]\n"
+ ".word 0x4f83e95e // sdot v30.4s, v10.16b, v3.4b[2]\n"
+ ".word 0x4f80e973 // sdot v19.4s, v11.16b, v0.4b[2]\n"
+ ".word 0x4f81e977 // sdot v23.4s, v11.16b, v1.4b[2]\n"
+ ".word 0x4f82e97b // sdot v27.4s, v11.16b, v2.4b[2]\n"
+ ".word 0x4f83e97f // sdot v31.4s, v11.16b, v3.4b[2]\n"
+ ".word 0x4fa0e990 // sdot v16.4s, v12.16b, v0.4b[3]\n"
+ ".word 0x4fa1e994 // sdot v20.4s, v12.16b, v1.4b[3]\n"
+ ".word 0x4fa2e998 // sdot v24.4s, v12.16b, v2.4b[3]\n"
+ ".word 0x4fa3e99c // sdot v28.4s, v12.16b, v3.4b[3]\n"
+ ".word 0x4fa0e9b1 // sdot v17.4s, v13.16b, v0.4b[3]\n"
+ ".word 0x4fa1e9b5 // sdot v21.4s, v13.16b, v1.4b[3]\n"
+ ".word 0x4fa2e9b9 // sdot v25.4s, v13.16b, v2.4b[3]\n"
+ ".word 0x4fa3e9bd // sdot v29.4s, v13.16b, v3.4b[3]\n"
+ ".word 0x4fa0e9d2 // sdot v18.4s, v14.16b, v0.4b[3]\n"
+ ".word 0x4fa1e9d6 // sdot v22.4s, v14.16b, v1.4b[3]\n"
+ ".word 0x4fa2e9da // sdot v26.4s, v14.16b, v2.4b[3]\n"
+ ".word 0x4fa3e9de // sdot v30.4s, v14.16b, v3.4b[3]\n"
+ ".word 0x4fa0e9f3 // sdot v19.4s, v15.16b, v0.4b[3]\n"
+ ".word 0x4fa1e9f7 // sdot v23.4s, v15.16b, v1.4b[3]\n"
+ ".word 0x4fa2e9fb // sdot v27.4s, v15.16b, v2.4b[3]\n"
+ ".word 0x4fa3e9ff // sdot v31.4s, v15.16b, v3.4b[3]\n"
+ "5:\n"
+ "cbz %[blocks], 6f\n"
+ "7:\n"
+ "ldr q8, [%[b_ptr0]]\n"
+ "subs %[blocks], %[blocks], #0x1\n"
+ "ldr q9, [%[b_ptr0], #0x10]\n"
+ "ldr s0, [%[a_ptr0]]\n"
+ "ldr q10, [%[b_ptr0], #0x20]\n"
+ "add %[a_ptr0], %[a_ptr0], #0x4\n"
+ "ldr q11, [%[b_ptr0], #0x30]\n"
+ "add %[b_ptr0], %[b_ptr0], #0x40\n"
+ ".word 0x4f80e110 // sdot v16.4s, v8.16b, v0.4b[0]\n"
+ "ldr s1, [a_ptr1]\n"
+ ".word 0x4f80e131 // sdot v17.4s, v9.16b, v0.4b[0]\n"
+ "add a_ptr1, a_ptr1, #0x4\n"
+ ".word 0x4f80e152 // sdot v18.4s, v10.16b, v0.4b[0]\n"
+ "ldr s2, [a_ptr2]\n"
+ ".word 0x4f81e114 // sdot v20.4s, v8.16b, v1.4b[0]\n"
+ "add a_ptr2, a_ptr2, #0x4\n"
+ ".word 0x4f81e135 // sdot v21.4s, v9.16b, v1.4b[0]\n"
+ "ldr s3, [a_ptr3]\n"
+ ".word 0x4f82e118 // sdot v24.4s, v8.16b, v2.4b[0]\n"
+ "add a_ptr3, a_ptr3, #0x4\n"
+ ".word 0x4f82e139 // sdot v25.4s, v9.16b, v2.4b[0]\n"
+ ".word 0x4f83e11c // sdot v28.4s, v8.16b, v3.4b[0]\n"
+ ".word 0x4f83e13d // sdot v29.4s, v9.16b, v3.4b[0]\n"
+ ".word 0x4f81e156 // sdot v22.4s, v10.16b, v1.4b[0]\n"
+ ".word 0x4f82e15a // sdot v26.4s, v10.16b, v2.4b[0]\n"
+ ".word 0x4f83e15e // sdot v30.4s, v10.16b, v3.4b[0]\n"
+ ".word 0x4f80e173 // sdot v19.4s, v11.16b, v0.4b[0]\n"
+ ".word 0x4f81e177 // sdot v23.4s, v11.16b, v1.4b[0]\n"
+ ".word 0x4f82e17b // sdot v27.4s, v11.16b, v2.4b[0]\n"
+ ".word 0x4f83e17f // sdot v31.4s, v11.16b, v3.4b[0]\n"
+ "b.ne 7b\n"
+ "6:\n"
+ "cbz %[odds], 8f\n"
+ "ld1 {v0.b}[0], [%[a_ptr0]], #1\n"
+ "ld1 {v1.b}[0], [a_ptr1], #1\n"
+ "ld1 {v2.b}[0], [a_ptr2], #1\n"
+ "ld1 {v3.b}[0], [a_ptr3], #1\n"
+ "subs %[odds], %[odds], #0x1\n"
+ "b.eq 9f\n"
+ "ld1 {v0.b}[1], [%[a_ptr0]], #1\n"
+ "ld1 {v1.b}[1], [a_ptr1], #1\n"
+ "ld1 {v2.b}[1], [a_ptr2], #1\n"
+ "ld1 {v3.b}[1], [a_ptr3], #1\n"
+ "subs %[odds], %[odds], #0x1\n"
+ "b.eq 9f\n"
+ "ld1 {v0.b}[2], [%[a_ptr0]]\n"
+ "ld1 {v1.b}[2], [a_ptr1]\n"
+ "ld1 {v2.b}[2], [a_ptr2]\n"
+ "ld1 {v3.b}[2], [a_ptr3]\n"
+ "9:\n"
+ "ldr q8, [%[b_ptr0]]\n"
+ "ldr q9, [%[b_ptr0], #0x10]\n"
+ "ldr q10, [%[b_ptr0], #0x20]\n"
+ "ldr q11, [%[b_ptr0], #0x30]\n"
+ ".word 0x4f80e110 // sdot v16.4s, v8.16b, v0.4b[0]\n"
+ ".word 0x4f81e114 // sdot v20.4s, v8.16b, v1.4b[0]\n"
+ ".word 0x4f82e118 // sdot v24.4s, v8.16b, v2.4b[0]\n"
+ ".word 0x4f83e11c // sdot v28.4s, v8.16b, v3.4b[0]\n"
+ ".word 0x4f80e131 // sdot v17.4s, v9.16b, v0.4b[0]\n"
+ ".word 0x4f81e135 // sdot v21.4s, v9.16b, v1.4b[0]\n"
+ ".word 0x4f82e139 // sdot v25.4s, v9.16b, v2.4b[0]\n"
+ ".word 0x4f83e13d // sdot v29.4s, v9.16b, v3.4b[0]\n"
+ ".word 0x4f80e152 // sdot v18.4s, v10.16b, v0.4b[0]\n"
+ ".word 0x4f81e156 // sdot v22.4s, v10.16b, v1.4b[0]\n"
+ ".word 0x4f82e15a // sdot v26.4s, v10.16b, v2.4b[0]\n"
+ ".word 0x4f83e15e // sdot v30.4s, v10.16b, v3.4b[0]\n"
+ ".word 0x4f80e173 // sdot v19.4s, v11.16b, v0.4b[0]\n"
+ ".word 0x4f81e177 // sdot v23.4s, v11.16b, v1.4b[0]\n"
+ ".word 0x4f82e17b // sdot v27.4s, v11.16b, v2.4b[0]\n"
+ ".word 0x4f83e17f // sdot v31.4s, v11.16b, v3.4b[0]\n"
+ "8:\n"
+ "str q16, [%[c_ptr0]]\n"
+ "str q17, [%[c_ptr0], #0x10]\n"
+ "str q18, [%[c_ptr0], #0x20]\n"
+ "str q19, [%[c_ptr0], #0x30]\n"
+ "add %[c_ptr0], %[c_ptr0], #0x40\n"
+ "str q20, [c_ptr1]\n"
+ "str q21, [c_ptr1, #0x10]\n"
+ "str q22, [c_ptr1, #0x20]\n"
+ "str q23, [c_ptr1, #0x30]\n"
+ "str q24, [c_ptr2]\n"
+ "str q25, [c_ptr2, #0x10]\n"
+ "str q26, [c_ptr2, #0x20]\n"
+ "str q27, [c_ptr2, #0x30]\n"
+ "str q28, [c_ptr3]\n"
+ "str q29, [c_ptr3, #0x10]\n"
+ "str q30, [c_ptr3, #0x20]\n"
+ "str q31, [c_ptr3, #0x30]\n"
+ ".unreq a_ptr1\n"
+ ".unreq a_ptr2\n"
+ ".unreq a_ptr3\n"
+ ".unreq c_ptr1\n"
+ ".unreq c_ptr2\n"
+ ".unreq c_ptr3\n"
+ : [a_ptr0] "+r" (a_ptr0), [b_ptr0] "+r" (b_ptr0), [c_ptr0] "+r" (c_ptr0), [loops] "+r" (loops), [regs] "+r" (regs), [blocks] "+r" (blocks), [odds] "+r" (odds)
+ : [betaptr] "r" (betaptr), [width] "r" (width), [beta0] "r" (beta0), [lda] "r" (ldab), [ldc] "r" (ldcb)
+ : "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x0", "x1", "x2", "x3", "x4", "x5", "cc", "memory"
);
break;
}
+ if (use_result_buffer) {
+ for(int cy=0; cy<std::min(M-y, 4); cy++) {
+ for(unsigned int cx=0; cx<width; cx++) {
+ c_ptr_real[cy * ldc + cx] = result_buffer[cy * 16 + cx];
+ }
+ }
+ }
}
}
}
} // namespace arm_gemm
-#endif // __aarch64__
+#endif // __aarch64__ \ No newline at end of file