diff options
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8u32_dot_16x4/generic.cpp')
-rw-r--r-- | src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8u32_dot_16x4/generic.cpp | 1852 |
1 files changed, 905 insertions, 947 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8u32_dot_16x4/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8u32_dot_16x4/generic.cpp index 6301fa5657..23d919a64c 100644 --- a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8u32_dot_16x4/generic.cpp +++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8u32_dot_16x4/generic.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2019 Arm Limited. + * Copyright (c) 2018-2019 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -25,14 +25,17 @@ #include <algorithm> +#include "arm_gemm.hpp" #include <cstdint> #include "../../asmlib.hpp" #include "../../utils.hpp" namespace arm_gemm { -void a64_hybrid_u8u32_dot_16x4(const uint8_t *A, int lda, const uint8_t *B, uint32_t *C, int ldc, uint32_t beta, int M, int N, int K) { - const long beta0 = (beta == 0u); +void a64_hybrid_u8u32_dot_16x4(const uint8_t *A, int lda, const uint8_t *B, uint32_t *C, int ldc, int M, int N, int K, const uint32_t *bias, Activation act, bool append) { + UNUSED(bias); + UNUSED(act); + const int K_stride = ((K + 3) / 4) * 4; const long loops_count = ((K + 16) / 32) - 1; K -= loops_count * 32; @@ -49,7 +52,6 @@ void a64_hybrid_u8u32_dot_16x4(const uint8_t *A, int lda, const uint8_t *B, uint for (int x0=0; x0<N; x0+=16ul) { const long width = std::min((unsigned long)N-x0, 16ul); - const uint32_t *betaptr = β long loops = loops_count; long regs = regs_count; long blocks = blocks_count; @@ -60,7 +62,7 @@ void a64_hybrid_u8u32_dot_16x4(const uint8_t *A, int lda, const uint8_t *B, uint uint32_t result_buffer[64]; const unsigned long ldcb = (use_result_buffer ? 16 : ldc) * sizeof(uint32_t); uint32_t *c_ptr_real = c_ptr0; - if (use_result_buffer && !beta0) { + if (use_result_buffer && append) { 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]; @@ -74,7 +76,7 @@ void a64_hybrid_u8u32_dot_16x4(const uint8_t *A, int lda, const uint8_t *B, uint switch(M-y) { case 1: __asm __volatile ( - "cbz %[beta0], 1f\n" + "cbnz %[append], 1f\n" "movi v16.4s, #0\n" "ldr q0, [%[a_ptr0]]\n" "movi v17.4s, #0\n" @@ -92,93 +94,88 @@ void a64_hybrid_u8u32_dot_16x4(const uint8_t *A, int lda, const uint8_t *B, uint "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" + "add %[a_ptr0], %[a_ptr0], #0x10\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 0x6f80e110 // udot v16.4s, v8.16b, v0.4b[0]\n" + ".inst 0x6f80e110 // udot v16.4s, v8.16b, v0.4b[0]\n" "ldr q15, [%[b_ptr0], #-0x10]\n" - ".word 0x6f80e131 // udot v17.4s, v9.16b, v0.4b[0]\n" + ".inst 0x6f80e131 // udot v17.4s, v9.16b, v0.4b[0]\n" "ldr q4, [%[a_ptr0]]\n" - ".word 0x6f80e152 // udot v18.4s, v10.16b, v0.4b[0]\n" + ".inst 0x6f80e152 // udot v18.4s, v10.16b, v0.4b[0]\n" "ldr q8, [%[b_ptr0]]\n" - ".word 0x6f80e173 // udot v19.4s, v11.16b, v0.4b[0]\n" + ".inst 0x6f80e173 // udot v19.4s, v11.16b, v0.4b[0]\n" "ldr q9, [%[b_ptr0], #0x10]\n" - ".word 0x6fa0e190 // udot v16.4s, v12.16b, v0.4b[1]\n" + ".inst 0x6fa0e190 // udot v16.4s, v12.16b, v0.4b[1]\n" "ldr q10, [%[b_ptr0], #0x20]\n" - ".word 0x6fa0e1b1 // udot v17.4s, v13.16b, v0.4b[1]\n" + ".inst 0x6fa0e1b1 // udot v17.4s, v13.16b, v0.4b[1]\n" "ldr q11, [%[b_ptr0], #0x30]\n" - ".word 0x6fa0e1d2 // udot v18.4s, v14.16b, v0.4b[1]\n" + ".inst 0x6fa0e1d2 // udot v18.4s, v14.16b, v0.4b[1]\n" "ldr q12, [%[b_ptr0], #0x40]\n" - ".word 0x6fa0e1f3 // udot v19.4s, v15.16b, v0.4b[1]\n" + ".inst 0x6fa0e1f3 // udot v19.4s, v15.16b, v0.4b[1]\n" "ldr q13, [%[b_ptr0], #0x50]\n" - ".word 0x6f80e910 // udot v16.4s, v8.16b, v0.4b[2]\n" + ".inst 0x6f80e910 // udot v16.4s, v8.16b, v0.4b[2]\n" "ldr q14, [%[b_ptr0], #0x60]\n" - ".word 0x6f80e931 // udot v17.4s, v9.16b, v0.4b[2]\n" + ".inst 0x6f80e931 // udot v17.4s, v9.16b, v0.4b[2]\n" "ldr q15, [%[b_ptr0], #0x70]\n" - ".word 0x6f80e952 // udot v18.4s, v10.16b, v0.4b[2]\n" + ".inst 0x6f80e952 // udot v18.4s, v10.16b, v0.4b[2]\n" "subs %[loops], %[loops], #0x1\n" - ".word 0x6f80e973 // udot v19.4s, v11.16b, v0.4b[2]\n" + ".inst 0x6f80e973 // udot v19.4s, v11.16b, v0.4b[2]\n" "prfm PLDL1KEEP, [%[a_ptr0], #0x40]\n" - ".word 0x6fa0e990 // udot v16.4s, v12.16b, v0.4b[3]\n" + ".inst 0x6fa0e990 // udot v16.4s, v12.16b, v0.4b[3]\n" "add %[b_ptr0], %[b_ptr0], #0x100\n" - ".word 0x6fa0e9b1 // udot v17.4s, v13.16b, v0.4b[3]\n" + ".inst 0x6fa0e9b1 // udot v17.4s, v13.16b, v0.4b[3]\n" "ldr q8, [%[b_ptr0], #-0x80]\n" - ".word 0x6fa0e9d2 // udot v18.4s, v14.16b, v0.4b[3]\n" + ".inst 0x6fa0e9d2 // udot v18.4s, v14.16b, v0.4b[3]\n" "ldr q9, [%[b_ptr0], #-0x70]\n" - ".word 0x6fa0e9f3 // udot v19.4s, v15.16b, v0.4b[3]\n" + ".inst 0x6fa0e9f3 // udot 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 0x6f84e110 // udot v16.4s, v8.16b, v4.4b[0]\n" + ".inst 0x6f84e110 // udot v16.4s, v8.16b, v4.4b[0]\n" "ldr q12, [%[b_ptr0], #-0x40]\n" - ".word 0x6f84e131 // udot v17.4s, v9.16b, v4.4b[0]\n" + ".inst 0x6f84e131 // udot v17.4s, v9.16b, v4.4b[0]\n" "ldr q13, [%[b_ptr0], #-0x30]\n" - ".word 0x6f84e152 // udot v18.4s, v10.16b, v4.4b[0]\n" + ".inst 0x6f84e152 // udot v18.4s, v10.16b, v4.4b[0]\n" "ldr q14, [%[b_ptr0], #-0x20]\n" - ".word 0x6f84e173 // udot v19.4s, v11.16b, v4.4b[0]\n" + ".inst 0x6f84e173 // udot v19.4s, v11.16b, v4.4b[0]\n" "ldr q15, [%[b_ptr0], #-0x10]\n" - ".word 0x6fa4e190 // udot v16.4s, v12.16b, v4.4b[1]\n" + ".inst 0x6fa4e190 // udot v16.4s, v12.16b, v4.4b[1]\n" "ldr q0, [%[a_ptr0], #-0x10]\n" - ".word 0x6fa4e1b1 // udot v17.4s, v13.16b, v4.4b[1]\n" + ".inst 0x6fa4e1b1 // udot v17.4s, v13.16b, v4.4b[1]\n" "ldr q8, [%[b_ptr0]]\n" - ".word 0x6fa4e1d2 // udot v18.4s, v14.16b, v4.4b[1]\n" + ".inst 0x6fa4e1d2 // udot v18.4s, v14.16b, v4.4b[1]\n" "ldr q9, [%[b_ptr0], #0x10]\n" - ".word 0x6fa4e1f3 // udot v19.4s, v15.16b, v4.4b[1]\n" + ".inst 0x6fa4e1f3 // udot v19.4s, v15.16b, v4.4b[1]\n" "ldr q10, [%[b_ptr0], #0x20]\n" "ldr q11, [%[b_ptr0], #0x30]\n" - ".word 0x6f84e910 // udot v16.4s, v8.16b, v4.4b[2]\n" + ".inst 0x6f84e910 // udot v16.4s, v8.16b, v4.4b[2]\n" "ldr q12, [%[b_ptr0], #0x40]\n" - ".word 0x6f84e931 // udot v17.4s, v9.16b, v4.4b[2]\n" + ".inst 0x6f84e931 // udot v17.4s, v9.16b, v4.4b[2]\n" "ldr q13, [%[b_ptr0], #0x50]\n" - ".word 0x6f84e952 // udot v18.4s, v10.16b, v4.4b[2]\n" + ".inst 0x6f84e952 // udot v18.4s, v10.16b, v4.4b[2]\n" "ldr q14, [%[b_ptr0], #0x60]\n" - ".word 0x6f84e973 // udot v19.4s, v11.16b, v4.4b[2]\n" + ".inst 0x6f84e973 // udot v19.4s, v11.16b, v4.4b[2]\n" "ldr q15, [%[b_ptr0], #0x70]\n" - ".word 0x6fa4e990 // udot v16.4s, v12.16b, v4.4b[3]\n" + ".inst 0x6fa4e990 // udot v16.4s, v12.16b, v4.4b[3]\n" "add %[b_ptr0], %[b_ptr0], #0x100\n" - ".word 0x6fa4e9b1 // udot v17.4s, v13.16b, v4.4b[3]\n" + ".inst 0x6fa4e9b1 // udot v17.4s, v13.16b, v4.4b[3]\n" "ldr q8, [%[b_ptr0], #-0x80]\n" - ".word 0x6fa4e9d2 // udot v18.4s, v14.16b, v4.4b[3]\n" + ".inst 0x6fa4e9d2 // udot v18.4s, v14.16b, v4.4b[3]\n" "ldr q9, [%[b_ptr0], #-0x70]\n" - ".word 0x6fa4e9f3 // udot v19.4s, v15.16b, v4.4b[3]\n" + ".inst 0x6fa4e9f3 // udot 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" @@ -189,93 +186,93 @@ void a64_hybrid_u8u32_dot_16x4(const uint8_t *A, int lda, const uint8_t *B, uint "ldr q15, [%[b_ptr0], #-0x10]\n" "prfm PSTL1KEEP, [%[c_ptr0]]\n" "cbz %[regs], 4f\n" - ".word 0x6f80e110 // udot v16.4s, v8.16b, v0.4b[0]\n" + ".inst 0x6f80e110 // udot v16.4s, v8.16b, v0.4b[0]\n" "ldr q4, [%[a_ptr0]]\n" - ".word 0x6f80e131 // udot v17.4s, v9.16b, v0.4b[0]\n" + ".inst 0x6f80e131 // udot v17.4s, v9.16b, v0.4b[0]\n" "ldr q8, [%[b_ptr0]]\n" - ".word 0x6f80e152 // udot v18.4s, v10.16b, v0.4b[0]\n" + ".inst 0x6f80e152 // udot v18.4s, v10.16b, v0.4b[0]\n" "ldr q9, [%[b_ptr0], #0x10]\n" - ".word 0x6f80e173 // udot v19.4s, v11.16b, v0.4b[0]\n" + ".inst 0x6f80e173 // udot v19.4s, v11.16b, v0.4b[0]\n" "ldr q10, [%[b_ptr0], #0x20]\n" - ".word 0x6fa0e190 // udot v16.4s, v12.16b, v0.4b[1]\n" + ".inst 0x6fa0e190 // udot v16.4s, v12.16b, v0.4b[1]\n" "ldr q11, [%[b_ptr0], #0x30]\n" - ".word 0x6fa0e1b1 // udot v17.4s, v13.16b, v0.4b[1]\n" + ".inst 0x6fa0e1b1 // udot v17.4s, v13.16b, v0.4b[1]\n" "ldr q12, [%[b_ptr0], #0x40]\n" - ".word 0x6fa0e1d2 // udot v18.4s, v14.16b, v0.4b[1]\n" + ".inst 0x6fa0e1d2 // udot v18.4s, v14.16b, v0.4b[1]\n" "ldr q13, [%[b_ptr0], #0x50]\n" - ".word 0x6fa0e1f3 // udot v19.4s, v15.16b, v0.4b[1]\n" + ".inst 0x6fa0e1f3 // udot v19.4s, v15.16b, v0.4b[1]\n" "ldr q14, [%[b_ptr0], #0x60]\n" - ".word 0x6f80e910 // udot v16.4s, v8.16b, v0.4b[2]\n" + ".inst 0x6f80e910 // udot v16.4s, v8.16b, v0.4b[2]\n" "ldr q15, [%[b_ptr0], #0x70]\n" - ".word 0x6f80e931 // udot v17.4s, v9.16b, v0.4b[2]\n" + ".inst 0x6f80e931 // udot v17.4s, v9.16b, v0.4b[2]\n" "add %[b_ptr0], %[b_ptr0], #0x100\n" - ".word 0x6f80e952 // udot v18.4s, v10.16b, v0.4b[2]\n" + ".inst 0x6f80e952 // udot v18.4s, v10.16b, v0.4b[2]\n" "ldr q8, [%[b_ptr0], #-0x80]\n" - ".word 0x6f80e973 // udot v19.4s, v11.16b, v0.4b[2]\n" + ".inst 0x6f80e973 // udot v19.4s, v11.16b, v0.4b[2]\n" "ldr q9, [%[b_ptr0], #-0x70]\n" - ".word 0x6fa0e990 // udot v16.4s, v12.16b, v0.4b[3]\n" + ".inst 0x6fa0e990 // udot v16.4s, v12.16b, v0.4b[3]\n" "ldr q10, [%[b_ptr0], #-0x60]\n" - ".word 0x6fa0e9b1 // udot v17.4s, v13.16b, v0.4b[3]\n" + ".inst 0x6fa0e9b1 // udot v17.4s, v13.16b, v0.4b[3]\n" "ldr q11, [%[b_ptr0], #-0x50]\n" - ".word 0x6fa0e9d2 // udot v18.4s, v14.16b, v0.4b[3]\n" + ".inst 0x6fa0e9d2 // udot v18.4s, v14.16b, v0.4b[3]\n" "ldr q12, [%[b_ptr0], #-0x40]\n" - ".word 0x6fa0e9f3 // udot v19.4s, v15.16b, v0.4b[3]\n" + ".inst 0x6fa0e9f3 // udot v19.4s, v15.16b, v0.4b[3]\n" "ldr q13, [%[b_ptr0], #-0x30]\n" - ".word 0x6f84e110 // udot v16.4s, v8.16b, v4.4b[0]\n" + ".inst 0x6f84e110 // udot v16.4s, v8.16b, v4.4b[0]\n" "ldr q14, [%[b_ptr0], #-0x20]\n" - ".word 0x6f84e131 // udot v17.4s, v9.16b, v4.4b[0]\n" + ".inst 0x6f84e131 // udot v17.4s, v9.16b, v4.4b[0]\n" "ldr q15, [%[b_ptr0], #-0x10]\n" - ".word 0x6f84e152 // udot v18.4s, v10.16b, v4.4b[0]\n" + ".inst 0x6f84e152 // udot v18.4s, v10.16b, v4.4b[0]\n" "ldr q8, [%[b_ptr0]]\n" - ".word 0x6f84e173 // udot v19.4s, v11.16b, v4.4b[0]\n" + ".inst 0x6f84e173 // udot v19.4s, v11.16b, v4.4b[0]\n" "ldr q9, [%[b_ptr0], #0x10]\n" - ".word 0x6fa4e190 // udot v16.4s, v12.16b, v4.4b[1]\n" + ".inst 0x6fa4e190 // udot v16.4s, v12.16b, v4.4b[1]\n" "ldr q10, [%[b_ptr0], #0x20]\n" - ".word 0x6fa4e1b1 // udot v17.4s, v13.16b, v4.4b[1]\n" + ".inst 0x6fa4e1b1 // udot v17.4s, v13.16b, v4.4b[1]\n" "ldr q11, [%[b_ptr0], #0x30]\n" - ".word 0x6fa4e1d2 // udot v18.4s, v14.16b, v4.4b[1]\n" + ".inst 0x6fa4e1d2 // udot v18.4s, v14.16b, v4.4b[1]\n" "ldr q12, [%[b_ptr0], #0x40]\n" - ".word 0x6fa4e1f3 // udot v19.4s, v15.16b, v4.4b[1]\n" + ".inst 0x6fa4e1f3 // udot v19.4s, v15.16b, v4.4b[1]\n" "ldr q13, [%[b_ptr0], #0x50]\n" - ".word 0x6f84e910 // udot v16.4s, v8.16b, v4.4b[2]\n" + ".inst 0x6f84e910 // udot v16.4s, v8.16b, v4.4b[2]\n" "ldr q14, [%[b_ptr0], #0x60]\n" - ".word 0x6f84e931 // udot v17.4s, v9.16b, v4.4b[2]\n" + ".inst 0x6f84e931 // udot v17.4s, v9.16b, v4.4b[2]\n" "ldr q15, [%[b_ptr0], #0x70]\n" - ".word 0x6f84e952 // udot v18.4s, v10.16b, v4.4b[2]\n" + ".inst 0x6f84e952 // udot v18.4s, v10.16b, v4.4b[2]\n" "add %[b_ptr0], %[b_ptr0], #0x80\n" - ".word 0x6f84e973 // udot v19.4s, v11.16b, v4.4b[2]\n" + ".inst 0x6f84e973 // udot v19.4s, v11.16b, v4.4b[2]\n" "add %[a_ptr0], %[a_ptr0], #0x10\n" - ".word 0x6fa4e990 // udot v16.4s, v12.16b, v4.4b[3]\n" - ".word 0x6fa4e9b1 // udot v17.4s, v13.16b, v4.4b[3]\n" - ".word 0x6fa4e9d2 // udot v18.4s, v14.16b, v4.4b[3]\n" - ".word 0x6fa4e9f3 // udot v19.4s, v15.16b, v4.4b[3]\n" + ".inst 0x6fa4e990 // udot v16.4s, v12.16b, v4.4b[3]\n" + ".inst 0x6fa4e9b1 // udot v17.4s, v13.16b, v4.4b[3]\n" + ".inst 0x6fa4e9d2 // udot v18.4s, v14.16b, v4.4b[3]\n" + ".inst 0x6fa4e9f3 // udot v19.4s, v15.16b, v4.4b[3]\n" "b 5f\n" "4:\n" - ".word 0x6f80e110 // udot v16.4s, v8.16b, v0.4b[0]\n" + ".inst 0x6f80e110 // udot v16.4s, v8.16b, v0.4b[0]\n" "ldr q8, [%[b_ptr0]]\n" - ".word 0x6f80e131 // udot v17.4s, v9.16b, v0.4b[0]\n" + ".inst 0x6f80e131 // udot v17.4s, v9.16b, v0.4b[0]\n" "ldr q9, [%[b_ptr0], #0x10]\n" - ".word 0x6f80e152 // udot v18.4s, v10.16b, v0.4b[0]\n" + ".inst 0x6f80e152 // udot v18.4s, v10.16b, v0.4b[0]\n" "ldr q10, [%[b_ptr0], #0x20]\n" - ".word 0x6f80e173 // udot v19.4s, v11.16b, v0.4b[0]\n" + ".inst 0x6f80e173 // udot v19.4s, v11.16b, v0.4b[0]\n" "ldr q11, [%[b_ptr0], #0x30]\n" - ".word 0x6fa0e190 // udot v16.4s, v12.16b, v0.4b[1]\n" + ".inst 0x6fa0e190 // udot v16.4s, v12.16b, v0.4b[1]\n" "ldr q12, [%[b_ptr0], #0x40]\n" - ".word 0x6fa0e1b1 // udot v17.4s, v13.16b, v0.4b[1]\n" + ".inst 0x6fa0e1b1 // udot v17.4s, v13.16b, v0.4b[1]\n" "ldr q13, [%[b_ptr0], #0x50]\n" - ".word 0x6fa0e1d2 // udot v18.4s, v14.16b, v0.4b[1]\n" + ".inst 0x6fa0e1d2 // udot v18.4s, v14.16b, v0.4b[1]\n" "ldr q14, [%[b_ptr0], #0x60]\n" - ".word 0x6fa0e1f3 // udot v19.4s, v15.16b, v0.4b[1]\n" + ".inst 0x6fa0e1f3 // udot v19.4s, v15.16b, v0.4b[1]\n" "ldr q15, [%[b_ptr0], #0x70]\n" - ".word 0x6f80e910 // udot v16.4s, v8.16b, v0.4b[2]\n" + ".inst 0x6f80e910 // udot v16.4s, v8.16b, v0.4b[2]\n" "add %[b_ptr0], %[b_ptr0], #0x80\n" - ".word 0x6f80e931 // udot v17.4s, v9.16b, v0.4b[2]\n" - ".word 0x6f80e952 // udot v18.4s, v10.16b, v0.4b[2]\n" - ".word 0x6f80e973 // udot v19.4s, v11.16b, v0.4b[2]\n" - ".word 0x6fa0e990 // udot v16.4s, v12.16b, v0.4b[3]\n" - ".word 0x6fa0e9b1 // udot v17.4s, v13.16b, v0.4b[3]\n" - ".word 0x6fa0e9d2 // udot v18.4s, v14.16b, v0.4b[3]\n" - ".word 0x6fa0e9f3 // udot v19.4s, v15.16b, v0.4b[3]\n" + ".inst 0x6f80e931 // udot v17.4s, v9.16b, v0.4b[2]\n" + ".inst 0x6f80e952 // udot v18.4s, v10.16b, v0.4b[2]\n" + ".inst 0x6f80e973 // udot v19.4s, v11.16b, v0.4b[2]\n" + ".inst 0x6fa0e990 // udot v16.4s, v12.16b, v0.4b[3]\n" + ".inst 0x6fa0e9b1 // udot v17.4s, v13.16b, v0.4b[3]\n" + ".inst 0x6fa0e9d2 // udot v18.4s, v14.16b, v0.4b[3]\n" + ".inst 0x6fa0e9f3 // udot v19.4s, v15.16b, v0.4b[3]\n" "5:\n" "cbz %[blocks], 6f\n" "7:\n" @@ -287,10 +284,10 @@ void a64_hybrid_u8u32_dot_16x4(const uint8_t *A, int lda, const uint8_t *B, uint "add %[a_ptr0], %[a_ptr0], #0x4\n" "ldr q11, [%[b_ptr0], #0x30]\n" "add %[b_ptr0], %[b_ptr0], #0x40\n" - ".word 0x6f80e110 // udot v16.4s, v8.16b, v0.4b[0]\n" - ".word 0x6f80e131 // udot v17.4s, v9.16b, v0.4b[0]\n" - ".word 0x6f80e152 // udot v18.4s, v10.16b, v0.4b[0]\n" - ".word 0x6f80e173 // udot v19.4s, v11.16b, v0.4b[0]\n" + ".inst 0x6f80e110 // udot v16.4s, v8.16b, v0.4b[0]\n" + ".inst 0x6f80e131 // udot v17.4s, v9.16b, v0.4b[0]\n" + ".inst 0x6f80e152 // udot v18.4s, v10.16b, v0.4b[0]\n" + ".inst 0x6f80e173 // udot v19.4s, v11.16b, v0.4b[0]\n" "b.ne 7b\n" "6:\n" "cbz %[odds], 8f\n" @@ -306,10 +303,10 @@ void a64_hybrid_u8u32_dot_16x4(const uint8_t *A, int lda, const uint8_t *B, uint "ldr q9, [%[b_ptr0], #0x10]\n" "ldr q10, [%[b_ptr0], #0x20]\n" "ldr q11, [%[b_ptr0], #0x30]\n" - ".word 0x6f80e110 // udot v16.4s, v8.16b, v0.4b[0]\n" - ".word 0x6f80e131 // udot v17.4s, v9.16b, v0.4b[0]\n" - ".word 0x6f80e152 // udot v18.4s, v10.16b, v0.4b[0]\n" - ".word 0x6f80e173 // udot v19.4s, v11.16b, v0.4b[0]\n" + ".inst 0x6f80e110 // udot v16.4s, v8.16b, v0.4b[0]\n" + ".inst 0x6f80e131 // udot v17.4s, v9.16b, v0.4b[0]\n" + ".inst 0x6f80e152 // udot v18.4s, v10.16b, v0.4b[0]\n" + ".inst 0x6f80e173 // udot v19.4s, v11.16b, v0.4b[0]\n" "8:\n" "str q16, [%[c_ptr0]]\n" "str q17, [%[c_ptr0], #0x10]\n" @@ -317,7 +314,7 @@ void a64_hybrid_u8u32_dot_16x4(const uint8_t *A, int lda, const uint8_t *B, uint "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) + : [width] "r" (width), [append] "r" (static_cast<uint64_t>(append)), [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; @@ -327,7 +324,7 @@ void a64_hybrid_u8u32_dot_16x4(const uint8_t *A, int lda, const uint8_t *B, uint "c_ptr1 .req X1\n" "add a_ptr1, %[a_ptr0], %[lda]\n" "add c_ptr1, %[c_ptr0], %[ldc]\n" - "cbz %[beta0], 1f\n" + "cbnz %[append], 1f\n" "movi v16.4s, #0\n" "ldr q0, [%[a_ptr0]]\n" "movi v17.4s, #0\n" @@ -351,287 +348,278 @@ void a64_hybrid_u8u32_dot_16x4(const uint8_t *A, int lda, const uint8_t *B, uint "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" + "add %[a_ptr0], %[a_ptr0], #0x10\n" "ldr q1, [a_ptr1]\n" - "mul v22.4s, v22.4s, v15.4s\n" + "add a_ptr1, a_ptr1, #0x10\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 0x6f80e110 // udot v16.4s, v8.16b, v0.4b[0]\n" + ".inst 0x6f80e110 // udot v16.4s, v8.16b, v0.4b[0]\n" "ldr q15, [%[b_ptr0], #-0x10]\n" - ".word 0x6f81e114 // udot v20.4s, v8.16b, v1.4b[0]\n" + ".inst 0x6f81e114 // udot v20.4s, v8.16b, v1.4b[0]\n" "ldr q4, [%[a_ptr0]]\n" - ".word 0x6f80e131 // udot v17.4s, v9.16b, v0.4b[0]\n" + ".inst 0x6f80e131 // udot v17.4s, v9.16b, v0.4b[0]\n" "ldr q5, [a_ptr1]\n" - ".word 0x6f81e135 // udot v21.4s, v9.16b, v1.4b[0]\n" + ".inst 0x6f81e135 // udot v21.4s, v9.16b, v1.4b[0]\n" "ldr q8, [%[b_ptr0]]\n" - ".word 0x6f80e152 // udot v18.4s, v10.16b, v0.4b[0]\n" + ".inst 0x6f80e152 // udot v18.4s, v10.16b, v0.4b[0]\n" "ldr q9, [%[b_ptr0], #0x10]\n" - ".word 0x6f81e156 // udot v22.4s, v10.16b, v1.4b[0]\n" + ".inst 0x6f81e156 // udot v22.4s, v10.16b, v1.4b[0]\n" "ldr q10, [%[b_ptr0], #0x20]\n" - ".word 0x6f80e173 // udot v19.4s, v11.16b, v0.4b[0]\n" + ".inst 0x6f80e173 // udot v19.4s, v11.16b, v0.4b[0]\n" "subs %[loops], %[loops], #0x1\n" - ".word 0x6f81e177 // udot v23.4s, v11.16b, v1.4b[0]\n" + ".inst 0x6f81e177 // udot v23.4s, v11.16b, v1.4b[0]\n" "ldr q11, [%[b_ptr0], #0x30]\n" - ".word 0x6fa0e190 // udot v16.4s, v12.16b, v0.4b[1]\n" + ".inst 0x6fa0e190 // udot v16.4s, v12.16b, v0.4b[1]\n" "prfm PLDL1KEEP, [%[a_ptr0], #0x40]\n" - ".word 0x6fa1e194 // udot v20.4s, v12.16b, v1.4b[1]\n" + ".inst 0x6fa1e194 // udot v20.4s, v12.16b, v1.4b[1]\n" "ldr q12, [%[b_ptr0], #0x40]\n" - ".word 0x6fa0e1b1 // udot v17.4s, v13.16b, v0.4b[1]\n" + ".inst 0x6fa0e1b1 // udot v17.4s, v13.16b, v0.4b[1]\n" "add %[a_ptr0], %[a_ptr0], #0x20\n" - ".word 0x6fa1e1b5 // udot v21.4s, v13.16b, v1.4b[1]\n" + ".inst 0x6fa1e1b5 // udot v21.4s, v13.16b, v1.4b[1]\n" "ldr q13, [%[b_ptr0], #0x50]\n" - ".word 0x6fa0e1d2 // udot v18.4s, v14.16b, v0.4b[1]\n" + ".inst 0x6fa0e1d2 // udot v18.4s, v14.16b, v0.4b[1]\n" "add a_ptr1, a_ptr1, #0x20\n" - ".word 0x6fa1e1d6 // udot v22.4s, v14.16b, v1.4b[1]\n" + ".inst 0x6fa1e1d6 // udot v22.4s, v14.16b, v1.4b[1]\n" "ldr q14, [%[b_ptr0], #0x60]\n" - ".word 0x6fa0e1f3 // udot v19.4s, v15.16b, v0.4b[1]\n" + ".inst 0x6fa0e1f3 // udot v19.4s, v15.16b, v0.4b[1]\n" "prfm PLDL1KEEP, [a_ptr1, #0x40]\n" - ".word 0x6fa1e1f7 // udot v23.4s, v15.16b, v1.4b[1]\n" + ".inst 0x6fa1e1f7 // udot v23.4s, v15.16b, v1.4b[1]\n" "ldr q15, [%[b_ptr0], #0x70]\n" - ".word 0x6f80e910 // udot v16.4s, v8.16b, v0.4b[2]\n" + ".inst 0x6f80e910 // udot v16.4s, v8.16b, v0.4b[2]\n" "add %[b_ptr0], %[b_ptr0], #0x100\n" - ".word 0x6f81e914 // udot v20.4s, v8.16b, v1.4b[2]\n" + ".inst 0x6f81e914 // udot v20.4s, v8.16b, v1.4b[2]\n" "ldr q8, [%[b_ptr0], #-0x80]\n" - ".word 0x6f80e931 // udot v17.4s, v9.16b, v0.4b[2]\n" - ".word 0x6f81e935 // udot v21.4s, v9.16b, v1.4b[2]\n" + ".inst 0x6f80e931 // udot v17.4s, v9.16b, v0.4b[2]\n" + ".inst 0x6f81e935 // udot v21.4s, v9.16b, v1.4b[2]\n" "ldr q9, [%[b_ptr0], #-0x70]\n" - ".word 0x6f80e952 // udot v18.4s, v10.16b, v0.4b[2]\n" - ".word 0x6f81e956 // udot v22.4s, v10.16b, v1.4b[2]\n" + ".inst 0x6f80e952 // udot v18.4s, v10.16b, v0.4b[2]\n" + ".inst 0x6f81e956 // udot v22.4s, v10.16b, v1.4b[2]\n" "ldr q10, [%[b_ptr0], #-0x60]\n" - ".word 0x6f80e973 // udot v19.4s, v11.16b, v0.4b[2]\n" - ".word 0x6f81e977 // udot v23.4s, v11.16b, v1.4b[2]\n" + ".inst 0x6f80e973 // udot v19.4s, v11.16b, v0.4b[2]\n" + ".inst 0x6f81e977 // udot v23.4s, v11.16b, v1.4b[2]\n" "ldr q11, [%[b_ptr0], #-0x50]\n" - ".word 0x6fa0e990 // udot v16.4s, v12.16b, v0.4b[3]\n" - ".word 0x6fa1e994 // udot v20.4s, v12.16b, v1.4b[3]\n" + ".inst 0x6fa0e990 // udot v16.4s, v12.16b, v0.4b[3]\n" + ".inst 0x6fa1e994 // udot v20.4s, v12.16b, v1.4b[3]\n" "ldr q12, [%[b_ptr0], #-0x40]\n" - ".word 0x6fa0e9b1 // udot v17.4s, v13.16b, v0.4b[3]\n" - ".word 0x6fa1e9b5 // udot v21.4s, v13.16b, v1.4b[3]\n" + ".inst 0x6fa0e9b1 // udot v17.4s, v13.16b, v0.4b[3]\n" + ".inst 0x6fa1e9b5 // udot v21.4s, v13.16b, v1.4b[3]\n" "ldr q13, [%[b_ptr0], #-0x30]\n" - ".word 0x6fa0e9d2 // udot v18.4s, v14.16b, v0.4b[3]\n" - ".word 0x6fa1e9d6 // udot v22.4s, v14.16b, v1.4b[3]\n" + ".inst 0x6fa0e9d2 // udot v18.4s, v14.16b, v0.4b[3]\n" + ".inst 0x6fa1e9d6 // udot v22.4s, v14.16b, v1.4b[3]\n" "ldr q14, [%[b_ptr0], #-0x20]\n" - ".word 0x6fa0e9f3 // udot v19.4s, v15.16b, v0.4b[3]\n" + ".inst 0x6fa0e9f3 // udot v19.4s, v15.16b, v0.4b[3]\n" "ldr q0, [%[a_ptr0], #-0x10]\n" - ".word 0x6fa1e9f7 // udot v23.4s, v15.16b, v1.4b[3]\n" + ".inst 0x6fa1e9f7 // udot v23.4s, v15.16b, v1.4b[3]\n" "ldr q15, [%[b_ptr0], #-0x10]\n" - ".word 0x6f84e110 // udot v16.4s, v8.16b, v4.4b[0]\n" + ".inst 0x6f84e110 // udot v16.4s, v8.16b, v4.4b[0]\n" "ldr q1, [a_ptr1, #-0x10]\n" - ".word 0x6f85e114 // udot v20.4s, v8.16b, v5.4b[0]\n" + ".inst 0x6f85e114 // udot v20.4s, v8.16b, v5.4b[0]\n" "ldr q8, [%[b_ptr0]]\n" - ".word 0x6f84e131 // udot v17.4s, v9.16b, v4.4b[0]\n" - ".word 0x6f85e135 // udot v21.4s, v9.16b, v5.4b[0]\n" + ".inst 0x6f84e131 // udot v17.4s, v9.16b, v4.4b[0]\n" + ".inst 0x6f85e135 // udot v21.4s, v9.16b, v5.4b[0]\n" "ldr q9, [%[b_ptr0], #0x10]\n" - ".word 0x6f84e152 // udot v18.4s, v10.16b, v4.4b[0]\n" - ".word 0x6f85e156 // udot v22.4s, v10.16b, v5.4b[0]\n" + ".inst 0x6f84e152 // udot v18.4s, v10.16b, v4.4b[0]\n" + ".inst 0x6f85e156 // udot v22.4s, v10.16b, v5.4b[0]\n" "ldr q10, [%[b_ptr0], #0x20]\n" - ".word 0x6f84e173 // udot v19.4s, v11.16b, v4.4b[0]\n" - ".word 0x6f85e177 // udot v23.4s, v11.16b, v5.4b[0]\n" + ".inst 0x6f84e173 // udot v19.4s, v11.16b, v4.4b[0]\n" + ".inst 0x6f85e177 // udot v23.4s, v11.16b, v5.4b[0]\n" "ldr q11, [%[b_ptr0], #0x30]\n" - ".word 0x6fa4e190 // udot v16.4s, v12.16b, v4.4b[1]\n" - ".word 0x6fa5e194 // udot v20.4s, v12.16b, v5.4b[1]\n" + ".inst 0x6fa4e190 // udot v16.4s, v12.16b, v4.4b[1]\n" + ".inst 0x6fa5e194 // udot v20.4s, v12.16b, v5.4b[1]\n" "ldr q12, [%[b_ptr0], #0x40]\n" - ".word 0x6fa4e1b1 // udot v17.4s, v13.16b, v4.4b[1]\n" - ".word 0x6fa5e1b5 // udot v21.4s, v13.16b, v5.4b[1]\n" + ".inst 0x6fa4e1b1 // udot v17.4s, v13.16b, v4.4b[1]\n" + ".inst 0x6fa5e1b5 // udot v21.4s, v13.16b, v5.4b[1]\n" "ldr q13, [%[b_ptr0], #0x50]\n" - ".word 0x6fa4e1d2 // udot v18.4s, v14.16b, v4.4b[1]\n" - ".word 0x6fa5e1d6 // udot v22.4s, v14.16b, v5.4b[1]\n" + ".inst 0x6fa4e1d2 // udot v18.4s, v14.16b, v4.4b[1]\n" + ".inst 0x6fa5e1d6 // udot v22.4s, v14.16b, v5.4b[1]\n" "ldr q14, [%[b_ptr0], #0x60]\n" - ".word 0x6fa4e1f3 // udot v19.4s, v15.16b, v4.4b[1]\n" - ".word 0x6fa5e1f7 // udot v23.4s, v15.16b, v5.4b[1]\n" + ".inst 0x6fa4e1f3 // udot v19.4s, v15.16b, v4.4b[1]\n" + ".inst 0x6fa5e1f7 // udot v23.4s, v15.16b, v5.4b[1]\n" "ldr q15, [%[b_ptr0], #0x70]\n" - ".word 0x6f84e910 // udot v16.4s, v8.16b, v4.4b[2]\n" + ".inst 0x6f84e910 // udot v16.4s, v8.16b, v4.4b[2]\n" "add %[b_ptr0], %[b_ptr0], #0x100\n" - ".word 0x6f85e914 // udot v20.4s, v8.16b, v5.4b[2]\n" + ".inst 0x6f85e914 // udot v20.4s, v8.16b, v5.4b[2]\n" "ldr q8, [%[b_ptr0], #-0x80]\n" - ".word 0x6f84e931 // udot v17.4s, v9.16b, v4.4b[2]\n" - ".word 0x6f85e935 // udot v21.4s, v9.16b, v5.4b[2]\n" + ".inst 0x6f84e931 // udot v17.4s, v9.16b, v4.4b[2]\n" + ".inst 0x6f85e935 // udot v21.4s, v9.16b, v5.4b[2]\n" "ldr q9, [%[b_ptr0], #-0x70]\n" - ".word 0x6f84e952 // udot v18.4s, v10.16b, v4.4b[2]\n" - ".word 0x6f85e956 // udot v22.4s, v10.16b, v5.4b[2]\n" + ".inst 0x6f84e952 // udot v18.4s, v10.16b, v4.4b[2]\n" + ".inst 0x6f85e956 // udot v22.4s, v10.16b, v5.4b[2]\n" "ldr q10, [%[b_ptr0], #-0x60]\n" - ".word 0x6f84e973 // udot v19.4s, v11.16b, v4.4b[2]\n" - ".word 0x6f85e977 // udot v23.4s, v11.16b, v5.4b[2]\n" + ".inst 0x6f84e973 // udot v19.4s, v11.16b, v4.4b[2]\n" + ".inst 0x6f85e977 // udot v23.4s, v11.16b, v5.4b[2]\n" "ldr q11, [%[b_ptr0], #-0x50]\n" - ".word 0x6fa4e990 // udot v16.4s, v12.16b, v4.4b[3]\n" - ".word 0x6fa5e994 // udot v20.4s, v12.16b, v5.4b[3]\n" + ".inst 0x6fa4e990 // udot v16.4s, v12.16b, v4.4b[3]\n" + ".inst 0x6fa5e994 // udot v20.4s, v12.16b, v5.4b[3]\n" "ldr q12, [%[b_ptr0], #-0x40]\n" - ".word 0x6fa4e9b1 // udot v17.4s, v13.16b, v4.4b[3]\n" - ".word 0x6fa5e9b5 // udot v21.4s, v13.16b, v5.4b[3]\n" + ".inst 0x6fa4e9b1 // udot v17.4s, v13.16b, v4.4b[3]\n" + ".inst 0x6fa5e9b5 // udot v21.4s, v13.16b, v5.4b[3]\n" "ldr q13, [%[b_ptr0], #-0x30]\n" - ".word 0x6fa4e9d2 // udot v18.4s, v14.16b, v4.4b[3]\n" - ".word 0x6fa5e9d6 // udot v22.4s, v14.16b, v5.4b[3]\n" + ".inst 0x6fa4e9d2 // udot v18.4s, v14.16b, v4.4b[3]\n" + ".inst 0x6fa5e9d6 // udot v22.4s, v14.16b, v5.4b[3]\n" "ldr q14, [%[b_ptr0], #-0x20]\n" - ".word 0x6fa4e9f3 // udot v19.4s, v15.16b, v4.4b[3]\n" - ".word 0x6fa5e9f7 // udot v23.4s, v15.16b, v5.4b[3]\n" + ".inst 0x6fa4e9f3 // udot v19.4s, v15.16b, v4.4b[3]\n" + ".inst 0x6fa5e9f7 // udot 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 0x6f80e110 // udot v16.4s, v8.16b, v0.4b[0]\n" + ".inst 0x6f80e110 // udot v16.4s, v8.16b, v0.4b[0]\n" "ldr q4, [%[a_ptr0]]\n" - ".word 0x6f81e114 // udot v20.4s, v8.16b, v1.4b[0]\n" + ".inst 0x6f81e114 // udot v20.4s, v8.16b, v1.4b[0]\n" "ldr q5, [a_ptr1]\n" - ".word 0x6f80e131 // udot v17.4s, v9.16b, v0.4b[0]\n" + ".inst 0x6f80e131 // udot v17.4s, v9.16b, v0.4b[0]\n" "ldr q8, [%[b_ptr0]]\n" - ".word 0x6f81e135 // udot v21.4s, v9.16b, v1.4b[0]\n" + ".inst 0x6f81e135 // udot v21.4s, v9.16b, v1.4b[0]\n" "ldr q9, [%[b_ptr0], #0x10]\n" - ".word 0x6f80e152 // udot v18.4s, v10.16b, v0.4b[0]\n" + ".inst 0x6f80e152 // udot v18.4s, v10.16b, v0.4b[0]\n" "add %[a_ptr0], %[a_ptr0], #0x10\n" - ".word 0x6f81e156 // udot v22.4s, v10.16b, v1.4b[0]\n" + ".inst 0x6f81e156 // udot v22.4s, v10.16b, v1.4b[0]\n" "ldr q10, [%[b_ptr0], #0x20]\n" - ".word 0x6f80e173 // udot v19.4s, v11.16b, v0.4b[0]\n" + ".inst 0x6f80e173 // udot v19.4s, v11.16b, v0.4b[0]\n" "add a_ptr1, a_ptr1, #0x10\n" - ".word 0x6f81e177 // udot v23.4s, v11.16b, v1.4b[0]\n" + ".inst 0x6f81e177 // udot v23.4s, v11.16b, v1.4b[0]\n" "ldr q11, [%[b_ptr0], #0x30]\n" - ".word 0x6fa0e190 // udot v16.4s, v12.16b, v0.4b[1]\n" - ".word 0x6fa1e194 // udot v20.4s, v12.16b, v1.4b[1]\n" + ".inst 0x6fa0e190 // udot v16.4s, v12.16b, v0.4b[1]\n" + ".inst 0x6fa1e194 // udot v20.4s, v12.16b, v1.4b[1]\n" "ldr q12, [%[b_ptr0], #0x40]\n" - ".word 0x6fa0e1b1 // udot v17.4s, v13.16b, v0.4b[1]\n" - ".word 0x6fa1e1b5 // udot v21.4s, v13.16b, v1.4b[1]\n" + ".inst 0x6fa0e1b1 // udot v17.4s, v13.16b, v0.4b[1]\n" + ".inst 0x6fa1e1b5 // udot v21.4s, v13.16b, v1.4b[1]\n" "ldr q13, [%[b_ptr0], #0x50]\n" - ".word 0x6fa0e1d2 // udot v18.4s, v14.16b, v0.4b[1]\n" - ".word 0x6fa1e1d6 // udot v22.4s, v14.16b, v1.4b[1]\n" + ".inst 0x6fa0e1d2 // udot v18.4s, v14.16b, v0.4b[1]\n" + ".inst 0x6fa1e1d6 // udot v22.4s, v14.16b, v1.4b[1]\n" "ldr q14, [%[b_ptr0], #0x60]\n" - ".word 0x6fa0e1f3 // udot v19.4s, v15.16b, v0.4b[1]\n" - ".word 0x6fa1e1f7 // udot v23.4s, v15.16b, v1.4b[1]\n" + ".inst 0x6fa0e1f3 // udot v19.4s, v15.16b, v0.4b[1]\n" + ".inst 0x6fa1e1f7 // udot v23.4s, v15.16b, v1.4b[1]\n" "ldr q15, [%[b_ptr0], #0x70]\n" - ".word 0x6f80e910 // udot v16.4s, v8.16b, v0.4b[2]\n" + ".inst 0x6f80e910 // udot v16.4s, v8.16b, v0.4b[2]\n" "add %[b_ptr0], %[b_ptr0], #0x100\n" - ".word 0x6f81e914 // udot v20.4s, v8.16b, v1.4b[2]\n" + ".inst 0x6f81e914 // udot v20.4s, v8.16b, v1.4b[2]\n" "ldr q8, [%[b_ptr0], #-0x80]\n" - ".word 0x6f80e931 // udot v17.4s, v9.16b, v0.4b[2]\n" - ".word 0x6f81e935 // udot v21.4s, v9.16b, v1.4b[2]\n" + ".inst 0x6f80e931 // udot v17.4s, v9.16b, v0.4b[2]\n" + ".inst 0x6f81e935 // udot v21.4s, v9.16b, v1.4b[2]\n" "ldr q9, [%[b_ptr0], #-0x70]\n" - ".word 0x6f80e952 // udot v18.4s, v10.16b, v0.4b[2]\n" - ".word 0x6f81e956 // udot v22.4s, v10.16b, v1.4b[2]\n" + ".inst 0x6f80e952 // udot v18.4s, v10.16b, v0.4b[2]\n" + ".inst 0x6f81e956 // udot v22.4s, v10.16b, v1.4b[2]\n" "ldr q10, [%[b_ptr0], #-0x60]\n" - ".word 0x6f80e973 // udot v19.4s, v11.16b, v0.4b[2]\n" - ".word 0x6f81e977 // udot v23.4s, v11.16b, v1.4b[2]\n" + ".inst 0x6f80e973 // udot v19.4s, v11.16b, v0.4b[2]\n" + ".inst 0x6f81e977 // udot v23.4s, v11.16b, v1.4b[2]\n" "ldr q11, [%[b_ptr0], #-0x50]\n" - ".word 0x6fa0e990 // udot v16.4s, v12.16b, v0.4b[3]\n" - ".word 0x6fa1e994 // udot v20.4s, v12.16b, v1.4b[3]\n" + ".inst 0x6fa0e990 // udot v16.4s, v12.16b, v0.4b[3]\n" + ".inst 0x6fa1e994 // udot v20.4s, v12.16b, v1.4b[3]\n" "ldr q12, [%[b_ptr0], #-0x40]\n" - ".word 0x6fa0e9b1 // udot v17.4s, v13.16b, v0.4b[3]\n" - ".word 0x6fa1e9b5 // udot v21.4s, v13.16b, v1.4b[3]\n" + ".inst 0x6fa0e9b1 // udot v17.4s, v13.16b, v0.4b[3]\n" + ".inst 0x6fa1e9b5 // udot v21.4s, v13.16b, v1.4b[3]\n" "ldr q13, [%[b_ptr0], #-0x30]\n" - ".word 0x6fa0e9d2 // udot v18.4s, v14.16b, v0.4b[3]\n" - ".word 0x6fa1e9d6 // udot v22.4s, v14.16b, v1.4b[3]\n" + ".inst 0x6fa0e9d2 // udot v18.4s, v14.16b, v0.4b[3]\n" + ".inst 0x6fa1e9d6 // udot v22.4s, v14.16b, v1.4b[3]\n" "ldr q14, [%[b_ptr0], #-0x20]\n" - ".word 0x6fa0e9f3 // udot v19.4s, v15.16b, v0.4b[3]\n" - ".word 0x6fa1e9f7 // udot v23.4s, v15.16b, v1.4b[3]\n" + ".inst 0x6fa0e9f3 // udot v19.4s, v15.16b, v0.4b[3]\n" + ".inst 0x6fa1e9f7 // udot v23.4s, v15.16b, v1.4b[3]\n" "ldr q15, [%[b_ptr0], #-0x10]\n" - ".word 0x6f84e110 // udot v16.4s, v8.16b, v4.4b[0]\n" - ".word 0x6f85e114 // udot v20.4s, v8.16b, v5.4b[0]\n" + ".inst 0x6f84e110 // udot v16.4s, v8.16b, v4.4b[0]\n" + ".inst 0x6f85e114 // udot v20.4s, v8.16b, v5.4b[0]\n" "ldr q8, [%[b_ptr0]]\n" - ".word 0x6f84e131 // udot v17.4s, v9.16b, v4.4b[0]\n" - ".word 0x6f85e135 // udot v21.4s, v9.16b, v5.4b[0]\n" + ".inst 0x6f84e131 // udot v17.4s, v9.16b, v4.4b[0]\n" + ".inst 0x6f85e135 // udot v21.4s, v9.16b, v5.4b[0]\n" "ldr q9, [%[b_ptr0], #0x10]\n" - ".word 0x6f84e152 // udot v18.4s, v10.16b, v4.4b[0]\n" - ".word 0x6f85e156 // udot v22.4s, v10.16b, v5.4b[0]\n" + ".inst 0x6f84e152 // udot v18.4s, v10.16b, v4.4b[0]\n" + ".inst 0x6f85e156 // udot v22.4s, v10.16b, v5.4b[0]\n" "ldr q10, [%[b_ptr0], #0x20]\n" - ".word 0x6f84e173 // udot v19.4s, v11.16b, v4.4b[0]\n" - ".word 0x6f85e177 // udot v23.4s, v11.16b, v5.4b[0]\n" + ".inst 0x6f84e173 // udot v19.4s, v11.16b, v4.4b[0]\n" + ".inst 0x6f85e177 // udot v23.4s, v11.16b, v5.4b[0]\n" "ldr q11, [%[b_ptr0], #0x30]\n" - ".word 0x6fa4e190 // udot v16.4s, v12.16b, v4.4b[1]\n" - ".word 0x6fa5e194 // udot v20.4s, v12.16b, v5.4b[1]\n" + ".inst 0x6fa4e190 // udot v16.4s, v12.16b, v4.4b[1]\n" + ".inst 0x6fa5e194 // udot v20.4s, v12.16b, v5.4b[1]\n" "ldr q12, [%[b_ptr0], #0x40]\n" - ".word 0x6fa4e1b1 // udot v17.4s, v13.16b, v4.4b[1]\n" - ".word 0x6fa5e1b5 // udot v21.4s, v13.16b, v5.4b[1]\n" + ".inst 0x6fa4e1b1 // udot v17.4s, v13.16b, v4.4b[1]\n" + ".inst 0x6fa5e1b5 // udot v21.4s, v13.16b, v5.4b[1]\n" "ldr q13, [%[b_ptr0], #0x50]\n" - ".word 0x6fa4e1d2 // udot v18.4s, v14.16b, v4.4b[1]\n" - ".word 0x6fa5e1d6 // udot v22.4s, v14.16b, v5.4b[1]\n" + ".inst 0x6fa4e1d2 // udot v18.4s, v14.16b, v4.4b[1]\n" + ".inst 0x6fa5e1d6 // udot v22.4s, v14.16b, v5.4b[1]\n" "ldr q14, [%[b_ptr0], #0x60]\n" - ".word 0x6fa4e1f3 // udot v19.4s, v15.16b, v4.4b[1]\n" - ".word 0x6fa5e1f7 // udot v23.4s, v15.16b, v5.4b[1]\n" + ".inst 0x6fa4e1f3 // udot v19.4s, v15.16b, v4.4b[1]\n" + ".inst 0x6fa5e1f7 // udot v23.4s, v15.16b, v5.4b[1]\n" "ldr q15, [%[b_ptr0], #0x70]\n" - ".word 0x6f84e910 // udot v16.4s, v8.16b, v4.4b[2]\n" + ".inst 0x6f84e910 // udot v16.4s, v8.16b, v4.4b[2]\n" "add %[b_ptr0], %[b_ptr0], #0x80\n" - ".word 0x6f85e914 // udot v20.4s, v8.16b, v5.4b[2]\n" - ".word 0x6f84e931 // udot v17.4s, v9.16b, v4.4b[2]\n" - ".word 0x6f85e935 // udot v21.4s, v9.16b, v5.4b[2]\n" - ".word 0x6f84e952 // udot v18.4s, v10.16b, v4.4b[2]\n" - ".word 0x6f85e956 // udot v22.4s, v10.16b, v5.4b[2]\n" - ".word 0x6f84e973 // udot v19.4s, v11.16b, v4.4b[2]\n" - ".word 0x6f85e977 // udot v23.4s, v11.16b, v5.4b[2]\n" - ".word 0x6fa4e990 // udot v16.4s, v12.16b, v4.4b[3]\n" - ".word 0x6fa5e994 // udot v20.4s, v12.16b, v5.4b[3]\n" - ".word 0x6fa4e9b1 // udot v17.4s, v13.16b, v4.4b[3]\n" - ".word 0x6fa5e9b5 // udot v21.4s, v13.16b, v5.4b[3]\n" - ".word 0x6fa4e9d2 // udot v18.4s, v14.16b, v4.4b[3]\n" - ".word 0x6fa5e9d6 // udot v22.4s, v14.16b, v5.4b[3]\n" - ".word 0x6fa4e9f3 // udot v19.4s, v15.16b, v4.4b[3]\n" - ".word 0x6fa5e9f7 // udot v23.4s, v15.16b, v5.4b[3]\n" + ".inst 0x6f85e914 // udot v20.4s, v8.16b, v5.4b[2]\n" + ".inst 0x6f84e931 // udot v17.4s, v9.16b, v4.4b[2]\n" + ".inst 0x6f85e935 // udot v21.4s, v9.16b, v5.4b[2]\n" + ".inst 0x6f84e952 // udot v18.4s, v10.16b, v4.4b[2]\n" + ".inst 0x6f85e956 // udot v22.4s, v10.16b, v5.4b[2]\n" + ".inst 0x6f84e973 // udot v19.4s, v11.16b, v4.4b[2]\n" + ".inst 0x6f85e977 // udot v23.4s, v11.16b, v5.4b[2]\n" + ".inst 0x6fa4e990 // udot v16.4s, v12.16b, v4.4b[3]\n" + ".inst 0x6fa5e994 // udot v20.4s, v12.16b, v5.4b[3]\n" + ".inst 0x6fa4e9b1 // udot v17.4s, v13.16b, v4.4b[3]\n" + ".inst 0x6fa5e9b5 // udot v21.4s, v13.16b, v5.4b[3]\n" + ".inst 0x6fa4e9d2 // udot v18.4s, v14.16b, v4.4b[3]\n" + ".inst 0x6fa5e9d6 // udot v22.4s, v14.16b, v5.4b[3]\n" + ".inst 0x6fa4e9f3 // udot v19.4s, v15.16b, v4.4b[3]\n" + ".inst 0x6fa5e9f7 // udot v23.4s, v15.16b, v5.4b[3]\n" "b 5f\n" "4:\n" - ".word 0x6f80e110 // udot v16.4s, v8.16b, v0.4b[0]\n" - ".word 0x6f81e114 // udot v20.4s, v8.16b, v1.4b[0]\n" + ".inst 0x6f80e110 // udot v16.4s, v8.16b, v0.4b[0]\n" + ".inst 0x6f81e114 // udot v20.4s, v8.16b, v1.4b[0]\n" "ldr q8, [%[b_ptr0]]\n" - ".word 0x6f80e131 // udot v17.4s, v9.16b, v0.4b[0]\n" - ".word 0x6f81e135 // udot v21.4s, v9.16b, v1.4b[0]\n" + ".inst 0x6f80e131 // udot v17.4s, v9.16b, v0.4b[0]\n" + ".inst 0x6f81e135 // udot v21.4s, v9.16b, v1.4b[0]\n" "ldr q9, [%[b_ptr0], #0x10]\n" - ".word 0x6f80e152 // udot v18.4s, v10.16b, v0.4b[0]\n" - ".word 0x6f81e156 // udot v22.4s, v10.16b, v1.4b[0]\n" + ".inst 0x6f80e152 // udot v18.4s, v10.16b, v0.4b[0]\n" + ".inst 0x6f81e156 // udot v22.4s, v10.16b, v1.4b[0]\n" "ldr q10, [%[b_ptr0], #0x20]\n" - ".word 0x6f80e173 // udot v19.4s, v11.16b, v0.4b[0]\n" - ".word 0x6f81e177 // udot v23.4s, v11.16b, v1.4b[0]\n" + ".inst 0x6f80e173 // udot v19.4s, v11.16b, v0.4b[0]\n" + ".inst 0x6f81e177 // udot v23.4s, v11.16b, v1.4b[0]\n" "ldr q11, [%[b_ptr0], #0x30]\n" - ".word 0x6fa0e190 // udot v16.4s, v12.16b, v0.4b[1]\n" - ".word 0x6fa1e194 // udot v20.4s, v12.16b, v1.4b[1]\n" + ".inst 0x6fa0e190 // udot v16.4s, v12.16b, v0.4b[1]\n" + ".inst 0x6fa1e194 // udot v20.4s, v12.16b, v1.4b[1]\n" "ldr q12, [%[b_ptr0], #0x40]\n" - ".word 0x6fa0e1b1 // udot v17.4s, v13.16b, v0.4b[1]\n" - ".word 0x6fa1e1b5 // udot v21.4s, v13.16b, v1.4b[1]\n" + ".inst 0x6fa0e1b1 // udot v17.4s, v13.16b, v0.4b[1]\n" + ".inst 0x6fa1e1b5 // udot v21.4s, v13.16b, v1.4b[1]\n" "ldr q13, [%[b_ptr0], #0x50]\n" - ".word 0x6fa0e1d2 // udot v18.4s, v14.16b, v0.4b[1]\n" - ".word 0x6fa1e1d6 // udot v22.4s, v14.16b, v1.4b[1]\n" + ".inst 0x6fa0e1d2 // udot v18.4s, v14.16b, v0.4b[1]\n" + ".inst 0x6fa1e1d6 // udot v22.4s, v14.16b, v1.4b[1]\n" "ldr q14, [%[b_ptr0], #0x60]\n" - ".word 0x6fa0e1f3 // udot v19.4s, v15.16b, v0.4b[1]\n" - ".word 0x6fa1e1f7 // udot v23.4s, v15.16b, v1.4b[1]\n" + ".inst 0x6fa0e1f3 // udot v19.4s, v15.16b, v0.4b[1]\n" + ".inst 0x6fa1e1f7 // udot v23.4s, v15.16b, v1.4b[1]\n" "ldr q15, [%[b_ptr0], #0x70]\n" - ".word 0x6f80e910 // udot v16.4s, v8.16b, v0.4b[2]\n" + ".inst 0x6f80e910 // udot v16.4s, v8.16b, v0.4b[2]\n" "add %[b_ptr0], %[b_ptr0], #0x80\n" - ".word 0x6f81e914 // udot v20.4s, v8.16b, v1.4b[2]\n" - ".word 0x6f80e931 // udot v17.4s, v9.16b, v0.4b[2]\n" - ".word 0x6f81e935 // udot v21.4s, v9.16b, v1.4b[2]\n" - ".word 0x6f80e952 // udot v18.4s, v10.16b, v0.4b[2]\n" - ".word 0x6f81e956 // udot v22.4s, v10.16b, v1.4b[2]\n" - ".word 0x6f80e973 // udot v19.4s, v11.16b, v0.4b[2]\n" - ".word 0x6f81e977 // udot v23.4s, v11.16b, v1.4b[2]\n" - ".word 0x6fa0e990 // udot v16.4s, v12.16b, v0.4b[3]\n" - ".word 0x6fa1e994 // udot v20.4s, v12.16b, v1.4b[3]\n" - ".word 0x6fa0e9b1 // udot v17.4s, v13.16b, v0.4b[3]\n" - ".word 0x6fa1e9b5 // udot v21.4s, v13.16b, v1.4b[3]\n" - ".word 0x6fa0e9d2 // udot v18.4s, v14.16b, v0.4b[3]\n" - ".word 0x6fa1e9d6 // udot v22.4s, v14.16b, v1.4b[3]\n" - ".word 0x6fa0e9f3 // udot v19.4s, v15.16b, v0.4b[3]\n" - ".word 0x6fa1e9f7 // udot v23.4s, v15.16b, v1.4b[3]\n" + ".inst 0x6f81e914 // udot v20.4s, v8.16b, v1.4b[2]\n" + ".inst 0x6f80e931 // udot v17.4s, v9.16b, v0.4b[2]\n" + ".inst 0x6f81e935 // udot v21.4s, v9.16b, v1.4b[2]\n" + ".inst 0x6f80e952 // udot v18.4s, v10.16b, v0.4b[2]\n" + ".inst 0x6f81e956 // udot v22.4s, v10.16b, v1.4b[2]\n" + ".inst 0x6f80e973 // udot v19.4s, v11.16b, v0.4b[2]\n" + ".inst 0x6f81e977 // udot v23.4s, v11.16b, v1.4b[2]\n" + ".inst 0x6fa0e990 // udot v16.4s, v12.16b, v0.4b[3]\n" + ".inst 0x6fa1e994 // udot v20.4s, v12.16b, v1.4b[3]\n" + ".inst 0x6fa0e9b1 // udot v17.4s, v13.16b, v0.4b[3]\n" + ".inst 0x6fa1e9b5 // udot v21.4s, v13.16b, v1.4b[3]\n" + ".inst 0x6fa0e9d2 // udot v18.4s, v14.16b, v0.4b[3]\n" + ".inst 0x6fa1e9d6 // udot v22.4s, v14.16b, v1.4b[3]\n" + ".inst 0x6fa0e9f3 // udot v19.4s, v15.16b, v0.4b[3]\n" + ".inst 0x6fa1e9f7 // udot v23.4s, v15.16b, v1.4b[3]\n" "5:\n" "cbz %[blocks], 6f\n" "7:\n" @@ -643,16 +631,16 @@ void a64_hybrid_u8u32_dot_16x4(const uint8_t *A, int lda, const uint8_t *B, uint "add %[a_ptr0], %[a_ptr0], #0x4\n" "ldr q11, [%[b_ptr0], #0x30]\n" "add %[b_ptr0], %[b_ptr0], #0x40\n" - ".word 0x6f80e110 // udot v16.4s, v8.16b, v0.4b[0]\n" + ".inst 0x6f80e110 // udot v16.4s, v8.16b, v0.4b[0]\n" "ldr s1, [a_ptr1]\n" - ".word 0x6f80e131 // udot v17.4s, v9.16b, v0.4b[0]\n" + ".inst 0x6f80e131 // udot v17.4s, v9.16b, v0.4b[0]\n" "add a_ptr1, a_ptr1, #0x4\n" - ".word 0x6f80e152 // udot v18.4s, v10.16b, v0.4b[0]\n" - ".word 0x6f81e114 // udot v20.4s, v8.16b, v1.4b[0]\n" - ".word 0x6f81e135 // udot v21.4s, v9.16b, v1.4b[0]\n" - ".word 0x6f81e156 // udot v22.4s, v10.16b, v1.4b[0]\n" - ".word 0x6f80e173 // udot v19.4s, v11.16b, v0.4b[0]\n" - ".word 0x6f81e177 // udot v23.4s, v11.16b, v1.4b[0]\n" + ".inst 0x6f80e152 // udot v18.4s, v10.16b, v0.4b[0]\n" + ".inst 0x6f81e114 // udot v20.4s, v8.16b, v1.4b[0]\n" + ".inst 0x6f81e135 // udot v21.4s, v9.16b, v1.4b[0]\n" + ".inst 0x6f81e156 // udot v22.4s, v10.16b, v1.4b[0]\n" + ".inst 0x6f80e173 // udot v19.4s, v11.16b, v0.4b[0]\n" + ".inst 0x6f81e177 // udot v23.4s, v11.16b, v1.4b[0]\n" "b.ne 7b\n" "6:\n" "cbz %[odds], 8f\n" @@ -671,14 +659,14 @@ void a64_hybrid_u8u32_dot_16x4(const uint8_t *A, int lda, const uint8_t *B, uint "ldr q9, [%[b_ptr0], #0x10]\n" "ldr q10, [%[b_ptr0], #0x20]\n" "ldr q11, [%[b_ptr0], #0x30]\n" - ".word 0x6f80e110 // udot v16.4s, v8.16b, v0.4b[0]\n" - ".word 0x6f81e114 // udot v20.4s, v8.16b, v1.4b[0]\n" - ".word 0x6f80e131 // udot v17.4s, v9.16b, v0.4b[0]\n" - ".word 0x6f81e135 // udot v21.4s, v9.16b, v1.4b[0]\n" - ".word 0x6f80e152 // udot v18.4s, v10.16b, v0.4b[0]\n" - ".word 0x6f81e156 // udot v22.4s, v10.16b, v1.4b[0]\n" - ".word 0x6f80e173 // udot v19.4s, v11.16b, v0.4b[0]\n" - ".word 0x6f81e177 // udot v23.4s, v11.16b, v1.4b[0]\n" + ".inst 0x6f80e110 // udot v16.4s, v8.16b, v0.4b[0]\n" + ".inst 0x6f81e114 // udot v20.4s, v8.16b, v1.4b[0]\n" + ".inst 0x6f80e131 // udot v17.4s, v9.16b, v0.4b[0]\n" + ".inst 0x6f81e135 // udot v21.4s, v9.16b, v1.4b[0]\n" + ".inst 0x6f80e152 // udot v18.4s, v10.16b, v0.4b[0]\n" + ".inst 0x6f81e156 // udot v22.4s, v10.16b, v1.4b[0]\n" + ".inst 0x6f80e173 // udot v19.4s, v11.16b, v0.4b[0]\n" + ".inst 0x6f81e177 // udot v23.4s, v11.16b, v1.4b[0]\n" "8:\n" "str q16, [%[c_ptr0]]\n" "str q17, [%[c_ptr0], #0x10]\n" @@ -692,7 +680,7 @@ void a64_hybrid_u8u32_dot_16x4(const uint8_t *A, int lda, const uint8_t *B, uint ".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) + : [width] "r" (width), [append] "r" (static_cast<uint64_t>(append)), [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; @@ -706,7 +694,7 @@ void a64_hybrid_u8u32_dot_16x4(const uint8_t *A, int lda, const uint8_t *B, uint "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" + "cbnz %[append], 1f\n" "movi v16.4s, #0\n" "ldr q0, [%[a_ptr0]]\n" "movi v17.4s, #0\n" @@ -736,190 +724,177 @@ void a64_hybrid_u8u32_dot_16x4(const uint8_t *A, int lda, const uint8_t *B, uint "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" + "add %[a_ptr0], %[a_ptr0], #0x10\n" "ldr q1, [a_ptr1]\n" - "mul v26.4s, v26.4s, v15.4s\n" + "add a_ptr1, a_ptr1, #0x10\n" "ldr q2, [a_ptr2]\n" - "mul v27.4s, v27.4s, v15.4s\n" + "add a_ptr2, a_ptr2, #0x10\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 0x6f80e110 // udot v16.4s, v8.16b, v0.4b[0]\n" + ".inst 0x6f80e110 // udot v16.4s, v8.16b, v0.4b[0]\n" "ldr q15, [%[b_ptr0], #-0x10]\n" - ".word 0x6f81e114 // udot v20.4s, v8.16b, v1.4b[0]\n" + ".inst 0x6f81e114 // udot v20.4s, v8.16b, v1.4b[0]\n" "ldr q4, [%[a_ptr0]]\n" - ".word 0x6f82e118 // udot v24.4s, v8.16b, v2.4b[0]\n" + ".inst 0x6f82e118 // udot v24.4s, v8.16b, v2.4b[0]\n" "ldr q5, [a_ptr1]\n" - ".word 0x6f80e131 // udot v17.4s, v9.16b, v0.4b[0]\n" + ".inst 0x6f80e131 // udot v17.4s, v9.16b, v0.4b[0]\n" "ldr q6, [a_ptr2]\n" - ".word 0x6f81e135 // udot v21.4s, v9.16b, v1.4b[0]\n" + ".inst 0x6f81e135 // udot v21.4s, v9.16b, v1.4b[0]\n" "ldr q8, [%[b_ptr0]]\n" - ".word 0x6f82e139 // udot v25.4s, v9.16b, v2.4b[0]\n" + ".inst 0x6f82e139 // udot v25.4s, v9.16b, v2.4b[0]\n" "ldr q9, [%[b_ptr0], #0x10]\n" - ".word 0x6f80e152 // udot v18.4s, v10.16b, v0.4b[0]\n" + ".inst 0x6f80e152 // udot v18.4s, v10.16b, v0.4b[0]\n" "subs %[loops], %[loops], #0x1\n" - ".word 0x6f81e156 // udot v22.4s, v10.16b, v1.4b[0]\n" + ".inst 0x6f81e156 // udot v22.4s, v10.16b, v1.4b[0]\n" "prfm PLDL1KEEP, [%[a_ptr0], #0x40]\n" - ".word 0x6f82e15a // udot v26.4s, v10.16b, v2.4b[0]\n" + ".inst 0x6f82e15a // udot v26.4s, v10.16b, v2.4b[0]\n" "ldr q10, [%[b_ptr0], #0x20]\n" - ".word 0x6f80e173 // udot v19.4s, v11.16b, v0.4b[0]\n" + ".inst 0x6f80e173 // udot v19.4s, v11.16b, v0.4b[0]\n" "add %[a_ptr0], %[a_ptr0], #0x20\n" - ".word 0x6f81e177 // udot v23.4s, v11.16b, v1.4b[0]\n" + ".inst 0x6f81e177 // udot v23.4s, v11.16b, v1.4b[0]\n" "add a_ptr1, a_ptr1, #0x20\n" - ".word 0x6f82e17b // udot v27.4s, v11.16b, v2.4b[0]\n" + ".inst 0x6f82e17b // udot v27.4s, v11.16b, v2.4b[0]\n" "ldr q11, [%[b_ptr0], #0x30]\n" - ".word 0x6fa0e190 // udot v16.4s, v12.16b, v0.4b[1]\n" + ".inst 0x6fa0e190 // udot v16.4s, v12.16b, v0.4b[1]\n" "add a_ptr2, a_ptr2, #0x20\n" - ".word 0x6fa1e194 // udot v20.4s, v12.16b, v1.4b[1]\n" + ".inst 0x6fa1e194 // udot v20.4s, v12.16b, v1.4b[1]\n" "prfm PLDL1KEEP, [a_ptr1, #0x40]\n" - ".word 0x6fa2e198 // udot v24.4s, v12.16b, v2.4b[1]\n" + ".inst 0x6fa2e198 // udot v24.4s, v12.16b, v2.4b[1]\n" "ldr q12, [%[b_ptr0], #0x40]\n" - ".word 0x6fa0e1b1 // udot v17.4s, v13.16b, v0.4b[1]\n" + ".inst 0x6fa0e1b1 // udot v17.4s, v13.16b, v0.4b[1]\n" "prfm PLDL1KEEP, [a_ptr2, #0x40]\n" - ".word 0x6fa1e1b5 // udot v21.4s, v13.16b, v1.4b[1]\n" - ".word 0x6fa2e1b9 // udot v25.4s, v13.16b, v2.4b[1]\n" + ".inst 0x6fa1e1b5 // udot v21.4s, v13.16b, v1.4b[1]\n" + ".inst 0x6fa2e1b9 // udot v25.4s, v13.16b, v2.4b[1]\n" "ldr q13, [%[b_ptr0], #0x50]\n" - ".word 0x6fa0e1d2 // udot v18.4s, v14.16b, v0.4b[1]\n" - ".word 0x6fa1e1d6 // udot v22.4s, v14.16b, v1.4b[1]\n" - ".word 0x6fa2e1da // udot v26.4s, v14.16b, v2.4b[1]\n" + ".inst 0x6fa0e1d2 // udot v18.4s, v14.16b, v0.4b[1]\n" + ".inst 0x6fa1e1d6 // udot v22.4s, v14.16b, v1.4b[1]\n" + ".inst 0x6fa2e1da // udot v26.4s, v14.16b, v2.4b[1]\n" "ldr q14, [%[b_ptr0], #0x60]\n" - ".word 0x6fa0e1f3 // udot v19.4s, v15.16b, v0.4b[1]\n" - ".word 0x6fa1e1f7 // udot v23.4s, v15.16b, v1.4b[1]\n" - ".word 0x6fa2e1fb // udot v27.4s, v15.16b, v2.4b[1]\n" + ".inst 0x6fa0e1f3 // udot v19.4s, v15.16b, v0.4b[1]\n" + ".inst 0x6fa1e1f7 // udot v23.4s, v15.16b, v1.4b[1]\n" + ".inst 0x6fa2e1fb // udot v27.4s, v15.16b, v2.4b[1]\n" "ldr q15, [%[b_ptr0], #0x70]\n" - ".word 0x6f80e910 // udot v16.4s, v8.16b, v0.4b[2]\n" + ".inst 0x6f80e910 // udot v16.4s, v8.16b, v0.4b[2]\n" "add %[b_ptr0], %[b_ptr0], #0x100\n" - ".word 0x6f81e914 // udot v20.4s, v8.16b, v1.4b[2]\n" - ".word 0x6f82e918 // udot v24.4s, v8.16b, v2.4b[2]\n" + ".inst 0x6f81e914 // udot v20.4s, v8.16b, v1.4b[2]\n" + ".inst 0x6f82e918 // udot v24.4s, v8.16b, v2.4b[2]\n" "ldr q8, [%[b_ptr0], #-0x80]\n" - ".word 0x6f80e931 // udot v17.4s, v9.16b, v0.4b[2]\n" - ".word 0x6f81e935 // udot v21.4s, v9.16b, v1.4b[2]\n" - ".word 0x6f82e939 // udot v25.4s, v9.16b, v2.4b[2]\n" + ".inst 0x6f80e931 // udot v17.4s, v9.16b, v0.4b[2]\n" + ".inst 0x6f81e935 // udot v21.4s, v9.16b, v1.4b[2]\n" + ".inst 0x6f82e939 // udot v25.4s, v9.16b, v2.4b[2]\n" "ldr q9, [%[b_ptr0], #-0x70]\n" - ".word 0x6f80e952 // udot v18.4s, v10.16b, v0.4b[2]\n" - ".word 0x6f81e956 // udot v22.4s, v10.16b, v1.4b[2]\n" - ".word 0x6f82e95a // udot v26.4s, v10.16b, v2.4b[2]\n" + ".inst 0x6f80e952 // udot v18.4s, v10.16b, v0.4b[2]\n" + ".inst 0x6f81e956 // udot v22.4s, v10.16b, v1.4b[2]\n" + ".inst 0x6f82e95a // udot v26.4s, v10.16b, v2.4b[2]\n" "ldr q10, [%[b_ptr0], #-0x60]\n" - ".word 0x6f80e973 // udot v19.4s, v11.16b, v0.4b[2]\n" - ".word 0x6f81e977 // udot v23.4s, v11.16b, v1.4b[2]\n" - ".word 0x6f82e97b // udot v27.4s, v11.16b, v2.4b[2]\n" + ".inst 0x6f80e973 // udot v19.4s, v11.16b, v0.4b[2]\n" + ".inst 0x6f81e977 // udot v23.4s, v11.16b, v1.4b[2]\n" + ".inst 0x6f82e97b // udot v27.4s, v11.16b, v2.4b[2]\n" "ldr q11, [%[b_ptr0], #-0x50]\n" - ".word 0x6fa0e990 // udot v16.4s, v12.16b, v0.4b[3]\n" - ".word 0x6fa1e994 // udot v20.4s, v12.16b, v1.4b[3]\n" - ".word 0x6fa2e998 // udot v24.4s, v12.16b, v2.4b[3]\n" + ".inst 0x6fa0e990 // udot v16.4s, v12.16b, v0.4b[3]\n" + ".inst 0x6fa1e994 // udot v20.4s, v12.16b, v1.4b[3]\n" + ".inst 0x6fa2e998 // udot v24.4s, v12.16b, v2.4b[3]\n" "ldr q12, [%[b_ptr0], #-0x40]\n" - ".word 0x6fa0e9b1 // udot v17.4s, v13.16b, v0.4b[3]\n" - ".word 0x6fa1e9b5 // udot v21.4s, v13.16b, v1.4b[3]\n" - ".word 0x6fa2e9b9 // udot v25.4s, v13.16b, v2.4b[3]\n" + ".inst 0x6fa0e9b1 // udot v17.4s, v13.16b, v0.4b[3]\n" + ".inst 0x6fa1e9b5 // udot v21.4s, v13.16b, v1.4b[3]\n" + ".inst 0x6fa2e9b9 // udot v25.4s, v13.16b, v2.4b[3]\n" "ldr q13, [%[b_ptr0], #-0x30]\n" - ".word 0x6fa0e9d2 // udot v18.4s, v14.16b, v0.4b[3]\n" - ".word 0x6fa1e9d6 // udot v22.4s, v14.16b, v1.4b[3]\n" - ".word 0x6fa2e9da // udot v26.4s, v14.16b, v2.4b[3]\n" + ".inst 0x6fa0e9d2 // udot v18.4s, v14.16b, v0.4b[3]\n" + ".inst 0x6fa1e9d6 // udot v22.4s, v14.16b, v1.4b[3]\n" + ".inst 0x6fa2e9da // udot v26.4s, v14.16b, v2.4b[3]\n" "ldr q14, [%[b_ptr0], #-0x20]\n" - ".word 0x6fa0e9f3 // udot v19.4s, v15.16b, v0.4b[3]\n" + ".inst 0x6fa0e9f3 // udot v19.4s, v15.16b, v0.4b[3]\n" "ldr q0, [%[a_ptr0], #-0x10]\n" - ".word 0x6fa1e9f7 // udot v23.4s, v15.16b, v1.4b[3]\n" + ".inst 0x6fa1e9f7 // udot v23.4s, v15.16b, v1.4b[3]\n" "ldr q1, [a_ptr1, #-0x10]\n" - ".word 0x6fa2e9fb // udot v27.4s, v15.16b, v2.4b[3]\n" + ".inst 0x6fa2e9fb // udot v27.4s, v15.16b, v2.4b[3]\n" "ldr q15, [%[b_ptr0], #-0x10]\n" - ".word 0x6f84e110 // udot v16.4s, v8.16b, v4.4b[0]\n" + ".inst 0x6f84e110 // udot v16.4s, v8.16b, v4.4b[0]\n" "ldr q2, [a_ptr2, #-0x10]\n" - ".word 0x6f85e114 // udot v20.4s, v8.16b, v5.4b[0]\n" - ".word 0x6f86e118 // udot v24.4s, v8.16b, v6.4b[0]\n" + ".inst 0x6f85e114 // udot v20.4s, v8.16b, v5.4b[0]\n" + ".inst 0x6f86e118 // udot v24.4s, v8.16b, v6.4b[0]\n" "ldr q8, [%[b_ptr0]]\n" - ".word 0x6f84e131 // udot v17.4s, v9.16b, v4.4b[0]\n" - ".word 0x6f85e135 // udot v21.4s, v9.16b, v5.4b[0]\n" - ".word 0x6f86e139 // udot v25.4s, v9.16b, v6.4b[0]\n" + ".inst 0x6f84e131 // udot v17.4s, v9.16b, v4.4b[0]\n" + ".inst 0x6f85e135 // udot v21.4s, v9.16b, v5.4b[0]\n" + ".inst 0x6f86e139 // udot v25.4s, v9.16b, v6.4b[0]\n" "ldr q9, [%[b_ptr0], #0x10]\n" - ".word 0x6f84e152 // udot v18.4s, v10.16b, v4.4b[0]\n" - ".word 0x6f85e156 // udot v22.4s, v10.16b, v5.4b[0]\n" - ".word 0x6f86e15a // udot v26.4s, v10.16b, v6.4b[0]\n" + ".inst 0x6f84e152 // udot v18.4s, v10.16b, v4.4b[0]\n" + ".inst 0x6f85e156 // udot v22.4s, v10.16b, v5.4b[0]\n" + ".inst 0x6f86e15a // udot v26.4s, v10.16b, v6.4b[0]\n" "ldr q10, [%[b_ptr0], #0x20]\n" - ".word 0x6f84e173 // udot v19.4s, v11.16b, v4.4b[0]\n" - ".word 0x6f85e177 // udot v23.4s, v11.16b, v5.4b[0]\n" - ".word 0x6f86e17b // udot v27.4s, v11.16b, v6.4b[0]\n" + ".inst 0x6f84e173 // udot v19.4s, v11.16b, v4.4b[0]\n" + ".inst 0x6f85e177 // udot v23.4s, v11.16b, v5.4b[0]\n" + ".inst 0x6f86e17b // udot v27.4s, v11.16b, v6.4b[0]\n" "ldr q11, [%[b_ptr0], #0x30]\n" - ".word 0x6fa4e190 // udot v16.4s, v12.16b, v4.4b[1]\n" - ".word 0x6fa5e194 // udot v20.4s, v12.16b, v5.4b[1]\n" - ".word 0x6fa6e198 // udot v24.4s, v12.16b, v6.4b[1]\n" + ".inst 0x6fa4e190 // udot v16.4s, v12.16b, v4.4b[1]\n" + ".inst 0x6fa5e194 // udot v20.4s, v12.16b, v5.4b[1]\n" + ".inst 0x6fa6e198 // udot v24.4s, v12.16b, v6.4b[1]\n" "ldr q12, [%[b_ptr0], #0x40]\n" - ".word 0x6fa4e1b1 // udot v17.4s, v13.16b, v4.4b[1]\n" - ".word 0x6fa5e1b5 // udot v21.4s, v13.16b, v5.4b[1]\n" - ".word 0x6fa6e1b9 // udot v25.4s, v13.16b, v6.4b[1]\n" + ".inst 0x6fa4e1b1 // udot v17.4s, v13.16b, v4.4b[1]\n" + ".inst 0x6fa5e1b5 // udot v21.4s, v13.16b, v5.4b[1]\n" + ".inst 0x6fa6e1b9 // udot v25.4s, v13.16b, v6.4b[1]\n" "ldr q13, [%[b_ptr0], #0x50]\n" - ".word 0x6fa4e1d2 // udot v18.4s, v14.16b, v4.4b[1]\n" - ".word 0x6fa5e1d6 // udot v22.4s, v14.16b, v5.4b[1]\n" - ".word 0x6fa6e1da // udot v26.4s, v14.16b, v6.4b[1]\n" + ".inst 0x6fa4e1d2 // udot v18.4s, v14.16b, v4.4b[1]\n" + ".inst 0x6fa5e1d6 // udot v22.4s, v14.16b, v5.4b[1]\n" + ".inst 0x6fa6e1da // udot v26.4s, v14.16b, v6.4b[1]\n" "ldr q14, [%[b_ptr0], #0x60]\n" - ".word 0x6fa4e1f3 // udot v19.4s, v15.16b, v4.4b[1]\n" - ".word 0x6fa5e1f7 // udot v23.4s, v15.16b, v5.4b[1]\n" - ".word 0x6fa6e1fb // udot v27.4s, v15.16b, v6.4b[1]\n" + ".inst 0x6fa4e1f3 // udot v19.4s, v15.16b, v4.4b[1]\n" + ".inst 0x6fa5e1f7 // udot v23.4s, v15.16b, v5.4b[1]\n" + ".inst 0x6fa6e1fb // udot v27.4s, v15.16b, v6.4b[1]\n" "ldr q15, [%[b_ptr0], #0x70]\n" - ".word 0x6f84e910 // udot v16.4s, v8.16b, v4.4b[2]\n" + ".inst 0x6f84e910 // udot v16.4s, v8.16b, v4.4b[2]\n" "add %[b_ptr0], %[b_ptr0], #0x100\n" - ".word 0x6f85e914 // udot v20.4s, v8.16b, v5.4b[2]\n" - ".word 0x6f86e918 // udot v24.4s, v8.16b, v6.4b[2]\n" + ".inst 0x6f85e914 // udot v20.4s, v8.16b, v5.4b[2]\n" + ".inst 0x6f86e918 // udot v24.4s, v8.16b, v6.4b[2]\n" "ldr q8, [%[b_ptr0], #-0x80]\n" - ".word 0x6f84e931 // udot v17.4s, v9.16b, v4.4b[2]\n" - ".word 0x6f85e935 // udot v21.4s, v9.16b, v5.4b[2]\n" - ".word 0x6f86e939 // udot v25.4s, v9.16b, v6.4b[2]\n" + ".inst 0x6f84e931 // udot v17.4s, v9.16b, v4.4b[2]\n" + ".inst 0x6f85e935 // udot v21.4s, v9.16b, v5.4b[2]\n" + ".inst 0x6f86e939 // udot v25.4s, v9.16b, v6.4b[2]\n" "ldr q9, [%[b_ptr0], #-0x70]\n" - ".word 0x6f84e952 // udot v18.4s, v10.16b, v4.4b[2]\n" - ".word 0x6f85e956 // udot v22.4s, v10.16b, v5.4b[2]\n" - ".word 0x6f86e95a // udot v26.4s, v10.16b, v6.4b[2]\n" + ".inst 0x6f84e952 // udot v18.4s, v10.16b, v4.4b[2]\n" + ".inst 0x6f85e956 // udot v22.4s, v10.16b, v5.4b[2]\n" + ".inst 0x6f86e95a // udot v26.4s, v10.16b, v6.4b[2]\n" "ldr q10, [%[b_ptr0], #-0x60]\n" - ".word 0x6f84e973 // udot v19.4s, v11.16b, v4.4b[2]\n" - ".word 0x6f85e977 // udot v23.4s, v11.16b, v5.4b[2]\n" - ".word 0x6f86e97b // udot v27.4s, v11.16b, v6.4b[2]\n" + ".inst 0x6f84e973 // udot v19.4s, v11.16b, v4.4b[2]\n" + ".inst 0x6f85e977 // udot v23.4s, v11.16b, v5.4b[2]\n" + ".inst 0x6f86e97b // udot v27.4s, v11.16b, v6.4b[2]\n" "ldr q11, [%[b_ptr0], #-0x50]\n" - ".word 0x6fa4e990 // udot v16.4s, v12.16b, v4.4b[3]\n" - ".word 0x6fa5e994 // udot v20.4s, v12.16b, v5.4b[3]\n" - ".word 0x6fa6e998 // udot v24.4s, v12.16b, v6.4b[3]\n" + ".inst 0x6fa4e990 // udot v16.4s, v12.16b, v4.4b[3]\n" + ".inst 0x6fa5e994 // udot v20.4s, v12.16b, v5.4b[3]\n" + ".inst 0x6fa6e998 // udot v24.4s, v12.16b, v6.4b[3]\n" "ldr q12, [%[b_ptr0], #-0x40]\n" - ".word 0x6fa4e9b1 // udot v17.4s, v13.16b, v4.4b[3]\n" - ".word 0x6fa5e9b5 // udot v21.4s, v13.16b, v5.4b[3]\n" - ".word 0x6fa6e9b9 // udot v25.4s, v13.16b, v6.4b[3]\n" + ".inst 0x6fa4e9b1 // udot v17.4s, v13.16b, v4.4b[3]\n" + ".inst 0x6fa5e9b5 // udot v21.4s, v13.16b, v5.4b[3]\n" + ".inst 0x6fa6e9b9 // udot v25.4s, v13.16b, v6.4b[3]\n" "ldr q13, [%[b_ptr0], #-0x30]\n" - ".word 0x6fa4e9d2 // udot v18.4s, v14.16b, v4.4b[3]\n" - ".word 0x6fa5e9d6 // udot v22.4s, v14.16b, v5.4b[3]\n" - ".word 0x6fa6e9da // udot v26.4s, v14.16b, v6.4b[3]\n" + ".inst 0x6fa4e9d2 // udot v18.4s, v14.16b, v4.4b[3]\n" + ".inst 0x6fa5e9d6 // udot v22.4s, v14.16b, v5.4b[3]\n" + ".inst 0x6fa6e9da // udot v26.4s, v14.16b, v6.4b[3]\n" "ldr q14, [%[b_ptr0], #-0x20]\n" - ".word 0x6fa4e9f3 // udot v19.4s, v15.16b, v4.4b[3]\n" - ".word 0x6fa5e9f7 // udot v23.4s, v15.16b, v5.4b[3]\n" - ".word 0x6fa6e9fb // udot v27.4s, v15.16b, v6.4b[3]\n" + ".inst 0x6fa4e9f3 // udot v19.4s, v15.16b, v4.4b[3]\n" + ".inst 0x6fa5e9f7 // udot v23.4s, v15.16b, v5.4b[3]\n" + ".inst 0x6fa6e9fb // udot v27.4s, v15.16b, v6.4b[3]\n" "b.ne 3b\n" "2:\n" "ldr q15, [%[b_ptr0], #-0x10]\n" @@ -927,193 +902,193 @@ void a64_hybrid_u8u32_dot_16x4(const uint8_t *A, int lda, const uint8_t *B, uint "prfm PSTL1KEEP, [c_ptr1]\n" "prfm PSTL1KEEP, [c_ptr2]\n" "cbz %[regs], 4f\n" - ".word 0x6f80e110 // udot v16.4s, v8.16b, v0.4b[0]\n" + ".inst 0x6f80e110 // udot v16.4s, v8.16b, v0.4b[0]\n" "ldr q4, [%[a_ptr0]]\n" - ".word 0x6f81e114 // udot v20.4s, v8.16b, v1.4b[0]\n" + ".inst 0x6f81e114 // udot v20.4s, v8.16b, v1.4b[0]\n" "ldr q5, [a_ptr1]\n" - ".word 0x6f82e118 // udot v24.4s, v8.16b, v2.4b[0]\n" + ".inst 0x6f82e118 // udot v24.4s, v8.16b, v2.4b[0]\n" "ldr q6, [a_ptr2]\n" - ".word 0x6f80e131 // udot v17.4s, v9.16b, v0.4b[0]\n" + ".inst 0x6f80e131 // udot v17.4s, v9.16b, v0.4b[0]\n" "ldr q8, [%[b_ptr0]]\n" - ".word 0x6f81e135 // udot v21.4s, v9.16b, v1.4b[0]\n" + ".inst 0x6f81e135 // udot v21.4s, v9.16b, v1.4b[0]\n" "add %[a_ptr0], %[a_ptr0], #0x10\n" - ".word 0x6f82e139 // udot v25.4s, v9.16b, v2.4b[0]\n" + ".inst 0x6f82e139 // udot v25.4s, v9.16b, v2.4b[0]\n" "ldr q9, [%[b_ptr0], #0x10]\n" - ".word 0x6f80e152 // udot v18.4s, v10.16b, v0.4b[0]\n" + ".inst 0x6f80e152 // udot v18.4s, v10.16b, v0.4b[0]\n" "add a_ptr1, a_ptr1, #0x10\n" - ".word 0x6f81e156 // udot v22.4s, v10.16b, v1.4b[0]\n" + ".inst 0x6f81e156 // udot v22.4s, v10.16b, v1.4b[0]\n" "add a_ptr2, a_ptr2, #0x10\n" - ".word 0x6f82e15a // udot v26.4s, v10.16b, v2.4b[0]\n" + ".inst 0x6f82e15a // udot v26.4s, v10.16b, v2.4b[0]\n" "ldr q10, [%[b_ptr0], #0x20]\n" - ".word 0x6f80e173 // udot v19.4s, v11.16b, v0.4b[0]\n" - ".word 0x6f81e177 // udot v23.4s, v11.16b, v1.4b[0]\n" - ".word 0x6f82e17b // udot v27.4s, v11.16b, v2.4b[0]\n" + ".inst 0x6f80e173 // udot v19.4s, v11.16b, v0.4b[0]\n" + ".inst 0x6f81e177 // udot v23.4s, v11.16b, v1.4b[0]\n" + ".inst 0x6f82e17b // udot v27.4s, v11.16b, v2.4b[0]\n" "ldr q11, [%[b_ptr0], #0x30]\n" - ".word 0x6fa0e190 // udot v16.4s, v12.16b, v0.4b[1]\n" - ".word 0x6fa1e194 // udot v20.4s, v12.16b, v1.4b[1]\n" - ".word 0x6fa2e198 // udot v24.4s, v12.16b, v2.4b[1]\n" + ".inst 0x6fa0e190 // udot v16.4s, v12.16b, v0.4b[1]\n" + ".inst 0x6fa1e194 // udot v20.4s, v12.16b, v1.4b[1]\n" + ".inst 0x6fa2e198 // udot v24.4s, v12.16b, v2.4b[1]\n" "ldr q12, [%[b_ptr0], #0x40]\n" - ".word 0x6fa0e1b1 // udot v17.4s, v13.16b, v0.4b[1]\n" - ".word 0x6fa1e1b5 // udot v21.4s, v13.16b, v1.4b[1]\n" - ".word 0x6fa2e1b9 // udot v25.4s, v13.16b, v2.4b[1]\n" + ".inst 0x6fa0e1b1 // udot v17.4s, v13.16b, v0.4b[1]\n" + ".inst 0x6fa1e1b5 // udot v21.4s, v13.16b, v1.4b[1]\n" + ".inst 0x6fa2e1b9 // udot v25.4s, v13.16b, v2.4b[1]\n" "ldr q13, [%[b_ptr0], #0x50]\n" - ".word 0x6fa0e1d2 // udot v18.4s, v14.16b, v0.4b[1]\n" - ".word 0x6fa1e1d6 // udot v22.4s, v14.16b, v1.4b[1]\n" - ".word 0x6fa2e1da // udot v26.4s, v14.16b, v2.4b[1]\n" + ".inst 0x6fa0e1d2 // udot v18.4s, v14.16b, v0.4b[1]\n" + ".inst 0x6fa1e1d6 // udot v22.4s, v14.16b, v1.4b[1]\n" + ".inst 0x6fa2e1da // udot v26.4s, v14.16b, v2.4b[1]\n" "ldr q14, [%[b_ptr0], #0x60]\n" - ".word 0x6fa0e1f3 // udot v19.4s, v15.16b, v0.4b[1]\n" - ".word 0x6fa1e1f7 // udot v23.4s, v15.16b, v1.4b[1]\n" - ".word 0x6fa2e1fb // udot v27.4s, v15.16b, v2.4b[1]\n" + ".inst 0x6fa0e1f3 // udot v19.4s, v15.16b, v0.4b[1]\n" + ".inst 0x6fa1e1f7 // udot v23.4s, v15.16b, v1.4b[1]\n" + ".inst 0x6fa2e1fb // udot v27.4s, v15.16b, v2.4b[1]\n" "ldr q15, [%[b_ptr0], #0x70]\n" - ".word 0x6f80e910 // udot v16.4s, v8.16b, v0.4b[2]\n" + ".inst 0x6f80e910 // udot v16.4s, v8.16b, v0.4b[2]\n" "add %[b_ptr0], %[b_ptr0], #0x100\n" - ".word 0x6f81e914 // udot v20.4s, v8.16b, v1.4b[2]\n" - ".word 0x6f82e918 // udot v24.4s, v8.16b, v2.4b[2]\n" + ".inst 0x6f81e914 // udot v20.4s, v8.16b, v1.4b[2]\n" + ".inst 0x6f82e918 // udot v24.4s, v8.16b, v2.4b[2]\n" "ldr q8, [%[b_ptr0], #-0x80]\n" - ".word 0x6f80e931 // udot v17.4s, v9.16b, v0.4b[2]\n" - ".word 0x6f81e935 // udot v21.4s, v9.16b, v1.4b[2]\n" - ".word 0x6f82e939 // udot v25.4s, v9.16b, v2.4b[2]\n" + ".inst 0x6f80e931 // udot v17.4s, v9.16b, v0.4b[2]\n" + ".inst 0x6f81e935 // udot v21.4s, v9.16b, v1.4b[2]\n" + ".inst 0x6f82e939 // udot v25.4s, v9.16b, v2.4b[2]\n" "ldr q9, [%[b_ptr0], #-0x70]\n" - ".word 0x6f80e952 // udot v18.4s, v10.16b, v0.4b[2]\n" - ".word 0x6f81e956 // udot v22.4s, v10.16b, v1.4b[2]\n" - ".word 0x6f82e95a // udot v26.4s, v10.16b, v2.4b[2]\n" + ".inst 0x6f80e952 // udot v18.4s, v10.16b, v0.4b[2]\n" + ".inst 0x6f81e956 // udot v22.4s, v10.16b, v1.4b[2]\n" + ".inst 0x6f82e95a // udot v26.4s, v10.16b, v2.4b[2]\n" "ldr q10, [%[b_ptr0], #-0x60]\n" - ".word 0x6f80e973 // udot v19.4s, v11.16b, v0.4b[2]\n" - ".word 0x6f81e977 // udot v23.4s, v11.16b, v1.4b[2]\n" - ".word 0x6f82e97b // udot v27.4s, v11.16b, v2.4b[2]\n" + ".inst 0x6f80e973 // udot v19.4s, v11.16b, v0.4b[2]\n" + ".inst 0x6f81e977 // udot v23.4s, v11.16b, v1.4b[2]\n" + ".inst 0x6f82e97b // udot v27.4s, v11.16b, v2.4b[2]\n" "ldr q11, [%[b_ptr0], #-0x50]\n" - ".word 0x6fa0e990 // udot v16.4s, v12.16b, v0.4b[3]\n" - ".word 0x6fa1e994 // udot v20.4s, v12.16b, v1.4b[3]\n" - ".word 0x6fa2e998 // udot v24.4s, v12.16b, v2.4b[3]\n" + ".inst 0x6fa0e990 // udot v16.4s, v12.16b, v0.4b[3]\n" + ".inst 0x6fa1e994 // udot v20.4s, v12.16b, v1.4b[3]\n" + ".inst 0x6fa2e998 // udot v24.4s, v12.16b, v2.4b[3]\n" "ldr q12, [%[b_ptr0], #-0x40]\n" - ".word 0x6fa0e9b1 // udot v17.4s, v13.16b, v0.4b[3]\n" - ".word 0x6fa1e9b5 // udot v21.4s, v13.16b, v1.4b[3]\n" - ".word 0x6fa2e9b9 // udot v25.4s, v13.16b, v2.4b[3]\n" + ".inst 0x6fa0e9b1 // udot v17.4s, v13.16b, v0.4b[3]\n" + ".inst 0x6fa1e9b5 // udot v21.4s, v13.16b, v1.4b[3]\n" + ".inst 0x6fa2e9b9 // udot v25.4s, v13.16b, v2.4b[3]\n" "ldr q13, [%[b_ptr0], #-0x30]\n" - ".word 0x6fa0e9d2 // udot v18.4s, v14.16b, v0.4b[3]\n" - ".word 0x6fa1e9d6 // udot v22.4s, v14.16b, v1.4b[3]\n" - ".word 0x6fa2e9da // udot v26.4s, v14.16b, v2.4b[3]\n" + ".inst 0x6fa0e9d2 // udot v18.4s, v14.16b, v0.4b[3]\n" + ".inst 0x6fa1e9d6 // udot v22.4s, v14.16b, v1.4b[3]\n" + ".inst 0x6fa2e9da // udot v26.4s, v14.16b, v2.4b[3]\n" "ldr q14, [%[b_ptr0], #-0x20]\n" - ".word 0x6fa0e9f3 // udot v19.4s, v15.16b, v0.4b[3]\n" - ".word 0x6fa1e9f7 // udot v23.4s, v15.16b, v1.4b[3]\n" - ".word 0x6fa2e9fb // udot v27.4s, v15.16b, v2.4b[3]\n" + ".inst 0x6fa0e9f3 // udot v19.4s, v15.16b, v0.4b[3]\n" + ".inst 0x6fa1e9f7 // udot v23.4s, v15.16b, v1.4b[3]\n" + ".inst 0x6fa2e9fb // udot v27.4s, v15.16b, v2.4b[3]\n" "ldr q15, [%[b_ptr0], #-0x10]\n" - ".word 0x6f84e110 // udot v16.4s, v8.16b, v4.4b[0]\n" - ".word 0x6f85e114 // udot v20.4s, v8.16b, v5.4b[0]\n" - ".word 0x6f86e118 // udot v24.4s, v8.16b, v6.4b[0]\n" + ".inst 0x6f84e110 // udot v16.4s, v8.16b, v4.4b[0]\n" + ".inst 0x6f85e114 // udot v20.4s, v8.16b, v5.4b[0]\n" + ".inst 0x6f86e118 // udot v24.4s, v8.16b, v6.4b[0]\n" "ldr q8, [%[b_ptr0]]\n" - ".word 0x6f84e131 // udot v17.4s, v9.16b, v4.4b[0]\n" - ".word 0x6f85e135 // udot v21.4s, v9.16b, v5.4b[0]\n" - ".word 0x6f86e139 // udot v25.4s, v9.16b, v6.4b[0]\n" + ".inst 0x6f84e131 // udot v17.4s, v9.16b, v4.4b[0]\n" + ".inst 0x6f85e135 // udot v21.4s, v9.16b, v5.4b[0]\n" + ".inst 0x6f86e139 // udot v25.4s, v9.16b, v6.4b[0]\n" "ldr q9, [%[b_ptr0], #0x10]\n" - ".word 0x6f84e152 // udot v18.4s, v10.16b, v4.4b[0]\n" - ".word 0x6f85e156 // udot v22.4s, v10.16b, v5.4b[0]\n" - ".word 0x6f86e15a // udot v26.4s, v10.16b, v6.4b[0]\n" + ".inst 0x6f84e152 // udot v18.4s, v10.16b, v4.4b[0]\n" + ".inst 0x6f85e156 // udot v22.4s, v10.16b, v5.4b[0]\n" + ".inst 0x6f86e15a // udot v26.4s, v10.16b, v6.4b[0]\n" "ldr q10, [%[b_ptr0], #0x20]\n" - ".word 0x6f84e173 // udot v19.4s, v11.16b, v4.4b[0]\n" - ".word 0x6f85e177 // udot v23.4s, v11.16b, v5.4b[0]\n" - ".word 0x6f86e17b // udot v27.4s, v11.16b, v6.4b[0]\n" + ".inst 0x6f84e173 // udot v19.4s, v11.16b, v4.4b[0]\n" + ".inst 0x6f85e177 // udot v23.4s, v11.16b, v5.4b[0]\n" + ".inst 0x6f86e17b // udot v27.4s, v11.16b, v6.4b[0]\n" "ldr q11, [%[b_ptr0], #0x30]\n" - ".word 0x6fa4e190 // udot v16.4s, v12.16b, v4.4b[1]\n" - ".word 0x6fa5e194 // udot v20.4s, v12.16b, v5.4b[1]\n" - ".word 0x6fa6e198 // udot v24.4s, v12.16b, v6.4b[1]\n" + ".inst 0x6fa4e190 // udot v16.4s, v12.16b, v4.4b[1]\n" + ".inst 0x6fa5e194 // udot v20.4s, v12.16b, v5.4b[1]\n" + ".inst 0x6fa6e198 // udot v24.4s, v12.16b, v6.4b[1]\n" "ldr q12, [%[b_ptr0], #0x40]\n" - ".word 0x6fa4e1b1 // udot v17.4s, v13.16b, v4.4b[1]\n" - ".word 0x6fa5e1b5 // udot v21.4s, v13.16b, v5.4b[1]\n" - ".word 0x6fa6e1b9 // udot v25.4s, v13.16b, v6.4b[1]\n" + ".inst 0x6fa4e1b1 // udot v17.4s, v13.16b, v4.4b[1]\n" + ".inst 0x6fa5e1b5 // udot v21.4s, v13.16b, v5.4b[1]\n" + ".inst 0x6fa6e1b9 // udot v25.4s, v13.16b, v6.4b[1]\n" "ldr q13, [%[b_ptr0], #0x50]\n" - ".word 0x6fa4e1d2 // udot v18.4s, v14.16b, v4.4b[1]\n" - ".word 0x6fa5e1d6 // udot v22.4s, v14.16b, v5.4b[1]\n" - ".word 0x6fa6e1da // udot v26.4s, v14.16b, v6.4b[1]\n" + ".inst 0x6fa4e1d2 // udot v18.4s, v14.16b, v4.4b[1]\n" + ".inst 0x6fa5e1d6 // udot v22.4s, v14.16b, v5.4b[1]\n" + ".inst 0x6fa6e1da // udot v26.4s, v14.16b, v6.4b[1]\n" "ldr q14, [%[b_ptr0], #0x60]\n" - ".word 0x6fa4e1f3 // udot v19.4s, v15.16b, v4.4b[1]\n" - ".word 0x6fa5e1f7 // udot v23.4s, v15.16b, v5.4b[1]\n" - ".word 0x6fa6e1fb // udot v27.4s, v15.16b, v6.4b[1]\n" + ".inst 0x6fa4e1f3 // udot v19.4s, v15.16b, v4.4b[1]\n" + ".inst 0x6fa5e1f7 // udot v23.4s, v15.16b, v5.4b[1]\n" + ".inst 0x6fa6e1fb // udot v27.4s, v15.16b, v6.4b[1]\n" "ldr q15, [%[b_ptr0], #0x70]\n" - ".word 0x6f84e910 // udot v16.4s, v8.16b, v4.4b[2]\n" + ".inst 0x6f84e910 // udot v16.4s, v8.16b, v4.4b[2]\n" "add %[b_ptr0], %[b_ptr0], #0x80\n" - ".word 0x6f85e914 // udot v20.4s, v8.16b, v5.4b[2]\n" - ".word 0x6f86e918 // udot v24.4s, v8.16b, v6.4b[2]\n" - ".word 0x6f84e931 // udot v17.4s, v9.16b, v4.4b[2]\n" - ".word 0x6f85e935 // udot v21.4s, v9.16b, v5.4b[2]\n" - ".word 0x6f86e939 // udot v25.4s, v9.16b, v6.4b[2]\n" - ".word 0x6f84e952 // udot v18.4s, v10.16b, v4.4b[2]\n" - ".word 0x6f85e956 // udot v22.4s, v10.16b, v5.4b[2]\n" - ".word 0x6f86e95a // udot v26.4s, v10.16b, v6.4b[2]\n" - ".word 0x6f84e973 // udot v19.4s, v11.16b, v4.4b[2]\n" - ".word 0x6f85e977 // udot v23.4s, v11.16b, v5.4b[2]\n" - ".word 0x6f86e97b // udot v27.4s, v11.16b, v6.4b[2]\n" - ".word 0x6fa4e990 // udot v16.4s, v12.16b, v4.4b[3]\n" - ".word 0x6fa5e994 // udot v20.4s, v12.16b, v5.4b[3]\n" - ".word 0x6fa6e998 // udot v24.4s, v12.16b, v6.4b[3]\n" - ".word 0x6fa4e9b1 // udot v17.4s, v13.16b, v4.4b[3]\n" - ".word 0x6fa5e9b5 // udot v21.4s, v13.16b, v5.4b[3]\n" - ".word 0x6fa6e9b9 // udot v25.4s, v13.16b, v6.4b[3]\n" - ".word 0x6fa4e9d2 // udot v18.4s, v14.16b, v4.4b[3]\n" - ".word 0x6fa5e9d6 // udot v22.4s, v14.16b, v5.4b[3]\n" - ".word 0x6fa6e9da // udot v26.4s, v14.16b, v6.4b[3]\n" - ".word 0x6fa4e9f3 // udot v19.4s, v15.16b, v4.4b[3]\n" - ".word 0x6fa5e9f7 // udot v23.4s, v15.16b, v5.4b[3]\n" - ".word 0x6fa6e9fb // udot v27.4s, v15.16b, v6.4b[3]\n" + ".inst 0x6f85e914 // udot v20.4s, v8.16b, v5.4b[2]\n" + ".inst 0x6f86e918 // udot v24.4s, v8.16b, v6.4b[2]\n" + ".inst 0x6f84e931 // udot v17.4s, v9.16b, v4.4b[2]\n" + ".inst 0x6f85e935 // udot v21.4s, v9.16b, v5.4b[2]\n" + ".inst 0x6f86e939 // udot v25.4s, v9.16b, v6.4b[2]\n" + ".inst 0x6f84e952 // udot v18.4s, v10.16b, v4.4b[2]\n" + ".inst 0x6f85e956 // udot v22.4s, v10.16b, v5.4b[2]\n" + ".inst 0x6f86e95a // udot v26.4s, v10.16b, v6.4b[2]\n" + ".inst 0x6f84e973 // udot v19.4s, v11.16b, v4.4b[2]\n" + ".inst 0x6f85e977 // udot v23.4s, v11.16b, v5.4b[2]\n" + ".inst 0x6f86e97b // udot v27.4s, v11.16b, v6.4b[2]\n" + ".inst 0x6fa4e990 // udot v16.4s, v12.16b, v4.4b[3]\n" + ".inst 0x6fa5e994 // udot v20.4s, v12.16b, v5.4b[3]\n" + ".inst 0x6fa6e998 // udot v24.4s, v12.16b, v6.4b[3]\n" + ".inst 0x6fa4e9b1 // udot v17.4s, v13.16b, v4.4b[3]\n" + ".inst 0x6fa5e9b5 // udot v21.4s, v13.16b, v5.4b[3]\n" + ".inst 0x6fa6e9b9 // udot v25.4s, v13.16b, v6.4b[3]\n" + ".inst 0x6fa4e9d2 // udot v18.4s, v14.16b, v4.4b[3]\n" + ".inst 0x6fa5e9d6 // udot v22.4s, v14.16b, v5.4b[3]\n" + ".inst 0x6fa6e9da // udot v26.4s, v14.16b, v6.4b[3]\n" + ".inst 0x6fa4e9f3 // udot v19.4s, v15.16b, v4.4b[3]\n" + ".inst 0x6fa5e9f7 // udot v23.4s, v15.16b, v5.4b[3]\n" + ".inst 0x6fa6e9fb // udot v27.4s, v15.16b, v6.4b[3]\n" "b 5f\n" "4:\n" - ".word 0x6f80e110 // udot v16.4s, v8.16b, v0.4b[0]\n" - ".word 0x6f81e114 // udot v20.4s, v8.16b, v1.4b[0]\n" - ".word 0x6f82e118 // udot v24.4s, v8.16b, v2.4b[0]\n" + ".inst 0x6f80e110 // udot v16.4s, v8.16b, v0.4b[0]\n" + ".inst 0x6f81e114 // udot v20.4s, v8.16b, v1.4b[0]\n" + ".inst 0x6f82e118 // udot v24.4s, v8.16b, v2.4b[0]\n" "ldr q8, [%[b_ptr0]]\n" - ".word 0x6f80e131 // udot v17.4s, v9.16b, v0.4b[0]\n" - ".word 0x6f81e135 // udot v21.4s, v9.16b, v1.4b[0]\n" - ".word 0x6f82e139 // udot v25.4s, v9.16b, v2.4b[0]\n" + ".inst 0x6f80e131 // udot v17.4s, v9.16b, v0.4b[0]\n" + ".inst 0x6f81e135 // udot v21.4s, v9.16b, v1.4b[0]\n" + ".inst 0x6f82e139 // udot v25.4s, v9.16b, v2.4b[0]\n" "ldr q9, [%[b_ptr0], #0x10]\n" - ".word 0x6f80e152 // udot v18.4s, v10.16b, v0.4b[0]\n" - ".word 0x6f81e156 // udot v22.4s, v10.16b, v1.4b[0]\n" - ".word 0x6f82e15a // udot v26.4s, v10.16b, v2.4b[0]\n" + ".inst 0x6f80e152 // udot v18.4s, v10.16b, v0.4b[0]\n" + ".inst 0x6f81e156 // udot v22.4s, v10.16b, v1.4b[0]\n" + ".inst 0x6f82e15a // udot v26.4s, v10.16b, v2.4b[0]\n" "ldr q10, [%[b_ptr0], #0x20]\n" - ".word 0x6f80e173 // udot v19.4s, v11.16b, v0.4b[0]\n" - ".word 0x6f81e177 // udot v23.4s, v11.16b, v1.4b[0]\n" - ".word 0x6f82e17b // udot v27.4s, v11.16b, v2.4b[0]\n" + ".inst 0x6f80e173 // udot v19.4s, v11.16b, v0.4b[0]\n" + ".inst 0x6f81e177 // udot v23.4s, v11.16b, v1.4b[0]\n" + ".inst 0x6f82e17b // udot v27.4s, v11.16b, v2.4b[0]\n" "ldr q11, [%[b_ptr0], #0x30]\n" - ".word 0x6fa0e190 // udot v16.4s, v12.16b, v0.4b[1]\n" - ".word 0x6fa1e194 // udot v20.4s, v12.16b, v1.4b[1]\n" - ".word 0x6fa2e198 // udot v24.4s, v12.16b, v2.4b[1]\n" + ".inst 0x6fa0e190 // udot v16.4s, v12.16b, v0.4b[1]\n" + ".inst 0x6fa1e194 // udot v20.4s, v12.16b, v1.4b[1]\n" + ".inst 0x6fa2e198 // udot v24.4s, v12.16b, v2.4b[1]\n" "ldr q12, [%[b_ptr0], #0x40]\n" - ".word 0x6fa0e1b1 // udot v17.4s, v13.16b, v0.4b[1]\n" - ".word 0x6fa1e1b5 // udot v21.4s, v13.16b, v1.4b[1]\n" - ".word 0x6fa2e1b9 // udot v25.4s, v13.16b, v2.4b[1]\n" + ".inst 0x6fa0e1b1 // udot v17.4s, v13.16b, v0.4b[1]\n" + ".inst 0x6fa1e1b5 // udot v21.4s, v13.16b, v1.4b[1]\n" + ".inst 0x6fa2e1b9 // udot v25.4s, v13.16b, v2.4b[1]\n" "ldr q13, [%[b_ptr0], #0x50]\n" - ".word 0x6fa0e1d2 // udot v18.4s, v14.16b, v0.4b[1]\n" - ".word 0x6fa1e1d6 // udot v22.4s, v14.16b, v1.4b[1]\n" - ".word 0x6fa2e1da // udot v26.4s, v14.16b, v2.4b[1]\n" + ".inst 0x6fa0e1d2 // udot v18.4s, v14.16b, v0.4b[1]\n" + ".inst 0x6fa1e1d6 // udot v22.4s, v14.16b, v1.4b[1]\n" + ".inst 0x6fa2e1da // udot v26.4s, v14.16b, v2.4b[1]\n" "ldr q14, [%[b_ptr0], #0x60]\n" - ".word 0x6fa0e1f3 // udot v19.4s, v15.16b, v0.4b[1]\n" - ".word 0x6fa1e1f7 // udot v23.4s, v15.16b, v1.4b[1]\n" - ".word 0x6fa2e1fb // udot v27.4s, v15.16b, v2.4b[1]\n" + ".inst 0x6fa0e1f3 // udot v19.4s, v15.16b, v0.4b[1]\n" + ".inst 0x6fa1e1f7 // udot v23.4s, v15.16b, v1.4b[1]\n" + ".inst 0x6fa2e1fb // udot v27.4s, v15.16b, v2.4b[1]\n" "ldr q15, [%[b_ptr0], #0x70]\n" - ".word 0x6f80e910 // udot v16.4s, v8.16b, v0.4b[2]\n" + ".inst 0x6f80e910 // udot v16.4s, v8.16b, v0.4b[2]\n" "add %[b_ptr0], %[b_ptr0], #0x80\n" - ".word 0x6f81e914 // udot v20.4s, v8.16b, v1.4b[2]\n" - ".word 0x6f82e918 // udot v24.4s, v8.16b, v2.4b[2]\n" - ".word 0x6f80e931 // udot v17.4s, v9.16b, v0.4b[2]\n" - ".word 0x6f81e935 // udot v21.4s, v9.16b, v1.4b[2]\n" - ".word 0x6f82e939 // udot v25.4s, v9.16b, v2.4b[2]\n" - ".word 0x6f80e952 // udot v18.4s, v10.16b, v0.4b[2]\n" - ".word 0x6f81e956 // udot v22.4s, v10.16b, v1.4b[2]\n" - ".word 0x6f82e95a // udot v26.4s, v10.16b, v2.4b[2]\n" - ".word 0x6f80e973 // udot v19.4s, v11.16b, v0.4b[2]\n" - ".word 0x6f81e977 // udot v23.4s, v11.16b, v1.4b[2]\n" - ".word 0x6f82e97b // udot v27.4s, v11.16b, v2.4b[2]\n" - ".word 0x6fa0e990 // udot v16.4s, v12.16b, v0.4b[3]\n" - ".word 0x6fa1e994 // udot v20.4s, v12.16b, v1.4b[3]\n" - ".word 0x6fa2e998 // udot v24.4s, v12.16b, v2.4b[3]\n" - ".word 0x6fa0e9b1 // udot v17.4s, v13.16b, v0.4b[3]\n" - ".word 0x6fa1e9b5 // udot v21.4s, v13.16b, v1.4b[3]\n" - ".word 0x6fa2e9b9 // udot v25.4s, v13.16b, v2.4b[3]\n" - ".word 0x6fa0e9d2 // udot v18.4s, v14.16b, v0.4b[3]\n" - ".word 0x6fa1e9d6 // udot v22.4s, v14.16b, v1.4b[3]\n" - ".word 0x6fa2e9da // udot v26.4s, v14.16b, v2.4b[3]\n" - ".word 0x6fa0e9f3 // udot v19.4s, v15.16b, v0.4b[3]\n" - ".word 0x6fa1e9f7 // udot v23.4s, v15.16b, v1.4b[3]\n" - ".word 0x6fa2e9fb // udot v27.4s, v15.16b, v2.4b[3]\n" + ".inst 0x6f81e914 // udot v20.4s, v8.16b, v1.4b[2]\n" + ".inst 0x6f82e918 // udot v24.4s, v8.16b, v2.4b[2]\n" + ".inst 0x6f80e931 // udot v17.4s, v9.16b, v0.4b[2]\n" + ".inst 0x6f81e935 // udot v21.4s, v9.16b, v1.4b[2]\n" + ".inst 0x6f82e939 // udot v25.4s, v9.16b, v2.4b[2]\n" + ".inst 0x6f80e952 // udot v18.4s, v10.16b, v0.4b[2]\n" + ".inst 0x6f81e956 // udot v22.4s, v10.16b, v1.4b[2]\n" + ".inst 0x6f82e95a // udot v26.4s, v10.16b, v2.4b[2]\n" + ".inst 0x6f80e973 // udot v19.4s, v11.16b, v0.4b[2]\n" + ".inst 0x6f81e977 // udot v23.4s, v11.16b, v1.4b[2]\n" + ".inst 0x6f82e97b // udot v27.4s, v11.16b, v2.4b[2]\n" + ".inst 0x6fa0e990 // udot v16.4s, v12.16b, v0.4b[3]\n" + ".inst 0x6fa1e994 // udot v20.4s, v12.16b, v1.4b[3]\n" + ".inst 0x6fa2e998 // udot v24.4s, v12.16b, v2.4b[3]\n" + ".inst 0x6fa0e9b1 // udot v17.4s, v13.16b, v0.4b[3]\n" + ".inst 0x6fa1e9b5 // udot v21.4s, v13.16b, v1.4b[3]\n" + ".inst 0x6fa2e9b9 // udot v25.4s, v13.16b, v2.4b[3]\n" + ".inst 0x6fa0e9d2 // udot v18.4s, v14.16b, v0.4b[3]\n" + ".inst 0x6fa1e9d6 // udot v22.4s, v14.16b, v1.4b[3]\n" + ".inst 0x6fa2e9da // udot v26.4s, v14.16b, v2.4b[3]\n" + ".inst 0x6fa0e9f3 // udot v19.4s, v15.16b, v0.4b[3]\n" + ".inst 0x6fa1e9f7 // udot v23.4s, v15.16b, v1.4b[3]\n" + ".inst 0x6fa2e9fb // udot v27.4s, v15.16b, v2.4b[3]\n" "5:\n" "cbz %[blocks], 6f\n" "7:\n" @@ -1125,22 +1100,22 @@ void a64_hybrid_u8u32_dot_16x4(const uint8_t *A, int lda, const uint8_t *B, uint "add %[a_ptr0], %[a_ptr0], #0x4\n" "ldr q11, [%[b_ptr0], #0x30]\n" "add %[b_ptr0], %[b_ptr0], #0x40\n" - ".word 0x6f80e110 // udot v16.4s, v8.16b, v0.4b[0]\n" + ".inst 0x6f80e110 // udot v16.4s, v8.16b, v0.4b[0]\n" "ldr s1, [a_ptr1]\n" - ".word 0x6f80e131 // udot v17.4s, v9.16b, v0.4b[0]\n" + ".inst 0x6f80e131 // udot v17.4s, v9.16b, v0.4b[0]\n" "add a_ptr1, a_ptr1, #0x4\n" - ".word 0x6f80e152 // udot v18.4s, v10.16b, v0.4b[0]\n" + ".inst 0x6f80e152 // udot v18.4s, v10.16b, v0.4b[0]\n" "ldr s2, [a_ptr2]\n" - ".word 0x6f81e114 // udot v20.4s, v8.16b, v1.4b[0]\n" + ".inst 0x6f81e114 // udot v20.4s, v8.16b, v1.4b[0]\n" "add a_ptr2, a_ptr2, #0x4\n" - ".word 0x6f81e135 // udot v21.4s, v9.16b, v1.4b[0]\n" - ".word 0x6f82e118 // udot v24.4s, v8.16b, v2.4b[0]\n" - ".word 0x6f82e139 // udot v25.4s, v9.16b, v2.4b[0]\n" - ".word 0x6f81e156 // udot v22.4s, v10.16b, v1.4b[0]\n" - ".word 0x6f82e15a // udot v26.4s, v10.16b, v2.4b[0]\n" - ".word 0x6f80e173 // udot v19.4s, v11.16b, v0.4b[0]\n" - ".word 0x6f81e177 // udot v23.4s, v11.16b, v1.4b[0]\n" - ".word 0x6f82e17b // udot v27.4s, v11.16b, v2.4b[0]\n" + ".inst 0x6f81e135 // udot v21.4s, v9.16b, v1.4b[0]\n" + ".inst 0x6f82e118 // udot v24.4s, v8.16b, v2.4b[0]\n" + ".inst 0x6f82e139 // udot v25.4s, v9.16b, v2.4b[0]\n" + ".inst 0x6f81e156 // udot v22.4s, v10.16b, v1.4b[0]\n" + ".inst 0x6f82e15a // udot v26.4s, v10.16b, v2.4b[0]\n" + ".inst 0x6f80e173 // udot v19.4s, v11.16b, v0.4b[0]\n" + ".inst 0x6f81e177 // udot v23.4s, v11.16b, v1.4b[0]\n" + ".inst 0x6f82e17b // udot v27.4s, v11.16b, v2.4b[0]\n" "b.ne 7b\n" "6:\n" "cbz %[odds], 8f\n" @@ -1162,18 +1137,18 @@ void a64_hybrid_u8u32_dot_16x4(const uint8_t *A, int lda, const uint8_t *B, uint "ldr q9, [%[b_ptr0], #0x10]\n" "ldr q10, [%[b_ptr0], #0x20]\n" "ldr q11, [%[b_ptr0], #0x30]\n" - ".word 0x6f80e110 // udot v16.4s, v8.16b, v0.4b[0]\n" - ".word 0x6f81e114 // udot v20.4s, v8.16b, v1.4b[0]\n" - ".word 0x6f82e118 // udot v24.4s, v8.16b, v2.4b[0]\n" - ".word 0x6f80e131 // udot v17.4s, v9.16b, v0.4b[0]\n" - ".word 0x6f81e135 // udot v21.4s, v9.16b, v1.4b[0]\n" - ".word 0x6f82e139 // udot v25.4s, v9.16b, v2.4b[0]\n" - ".word 0x6f80e152 // udot v18.4s, v10.16b, v0.4b[0]\n" - ".word 0x6f81e156 // udot v22.4s, v10.16b, v1.4b[0]\n" - ".word 0x6f82e15a // udot v26.4s, v10.16b, v2.4b[0]\n" - ".word 0x6f80e173 // udot v19.4s, v11.16b, v0.4b[0]\n" - ".word 0x6f81e177 // udot v23.4s, v11.16b, v1.4b[0]\n" - ".word 0x6f82e17b // udot v27.4s, v11.16b, v2.4b[0]\n" + ".inst 0x6f80e110 // udot v16.4s, v8.16b, v0.4b[0]\n" + ".inst 0x6f81e114 // udot v20.4s, v8.16b, v1.4b[0]\n" + ".inst 0x6f82e118 // udot v24.4s, v8.16b, v2.4b[0]\n" + ".inst 0x6f80e131 // udot v17.4s, v9.16b, v0.4b[0]\n" + ".inst 0x6f81e135 // udot v21.4s, v9.16b, v1.4b[0]\n" + ".inst 0x6f82e139 // udot v25.4s, v9.16b, v2.4b[0]\n" + ".inst 0x6f80e152 // udot v18.4s, v10.16b, v0.4b[0]\n" + ".inst 0x6f81e156 // udot v22.4s, v10.16b, v1.4b[0]\n" + ".inst 0x6f82e15a // udot v26.4s, v10.16b, v2.4b[0]\n" + ".inst 0x6f80e173 // udot v19.4s, v11.16b, v0.4b[0]\n" + ".inst 0x6f81e177 // udot v23.4s, v11.16b, v1.4b[0]\n" + ".inst 0x6f82e17b // udot v27.4s, v11.16b, v2.4b[0]\n" "8:\n" "str q16, [%[c_ptr0]]\n" "str q17, [%[c_ptr0], #0x10]\n" @@ -1193,7 +1168,7 @@ void a64_hybrid_u8u32_dot_16x4(const uint8_t *A, int lda, const uint8_t *B, uint ".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) + : [width] "r" (width), [append] "r" (static_cast<uint64_t>(append)), [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; @@ -1212,7 +1187,7 @@ void a64_hybrid_u8u32_dot_16x4(const uint8_t *A, int lda, const uint8_t *B, uint "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" + "cbnz %[append], 1f\n" "movi v16.4s, #0\n" "ldr q0, [%[a_ptr0]]\n" "movi v17.4s, #0\n" @@ -1248,236 +1223,219 @@ void a64_hybrid_u8u32_dot_16x4(const uint8_t *A, int lda, const uint8_t *B, uint "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" + "add %[a_ptr0], %[a_ptr0], #0x10\n" "ldr q1, [a_ptr1]\n" - "mul v30.4s, v30.4s, v15.4s\n" + "add a_ptr1, a_ptr1, #0x10\n" "ldr q2, [a_ptr2]\n" - "mul v31.4s, v31.4s, v15.4s\n" + "add a_ptr2, a_ptr2, #0x10\n" "ldr q3, [a_ptr3]\n" + "add a_ptr3, a_ptr3, #0x10\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 0x6f80e110 // udot v16.4s, v8.16b, v0.4b[0]\n" + ".inst 0x6f80e110 // udot v16.4s, v8.16b, v0.4b[0]\n" "ldr q15, [%[b_ptr0], #-0x10]\n" - ".word 0x6f81e114 // udot v20.4s, v8.16b, v1.4b[0]\n" + ".inst 0x6f81e114 // udot v20.4s, v8.16b, v1.4b[0]\n" "ldr q4, [%[a_ptr0]]\n" - ".word 0x6f82e118 // udot v24.4s, v8.16b, v2.4b[0]\n" + ".inst 0x6f82e118 // udot v24.4s, v8.16b, v2.4b[0]\n" "ldr q5, [a_ptr1]\n" - ".word 0x6f83e11c // udot v28.4s, v8.16b, v3.4b[0]\n" + ".inst 0x6f83e11c // udot v28.4s, v8.16b, v3.4b[0]\n" "ldr q6, [a_ptr2]\n" - ".word 0x6f80e131 // udot v17.4s, v9.16b, v0.4b[0]\n" + ".inst 0x6f80e131 // udot v17.4s, v9.16b, v0.4b[0]\n" "ldr q7, [a_ptr3]\n" - ".word 0x6f81e135 // udot v21.4s, v9.16b, v1.4b[0]\n" + ".inst 0x6f81e135 // udot v21.4s, v9.16b, v1.4b[0]\n" "ldr q8, [%[b_ptr0]]\n" - ".word 0x6f82e139 // udot v25.4s, v9.16b, v2.4b[0]\n" + ".inst 0x6f82e139 // udot v25.4s, v9.16b, v2.4b[0]\n" "subs %[loops], %[loops], #0x1\n" - ".word 0x6f83e13d // udot v29.4s, v9.16b, v3.4b[0]\n" + ".inst 0x6f83e13d // udot v29.4s, v9.16b, v3.4b[0]\n" "ldr q9, [%[b_ptr0], #0x10]\n" - ".word 0x6f80e152 // udot v18.4s, v10.16b, v0.4b[0]\n" + ".inst 0x6f80e152 // udot v18.4s, v10.16b, v0.4b[0]\n" "prfm PLDL1KEEP, [%[a_ptr0], #0x40]\n" - ".word 0x6f81e156 // udot v22.4s, v10.16b, v1.4b[0]\n" + ".inst 0x6f81e156 // udot v22.4s, v10.16b, v1.4b[0]\n" "add %[a_ptr0], %[a_ptr0], #0x20\n" - ".word 0x6f82e15a // udot v26.4s, v10.16b, v2.4b[0]\n" + ".inst 0x6f82e15a // udot v26.4s, v10.16b, v2.4b[0]\n" "add a_ptr1, a_ptr1, #0x20\n" - ".word 0x6f83e15e // udot v30.4s, v10.16b, v3.4b[0]\n" + ".inst 0x6f83e15e // udot v30.4s, v10.16b, v3.4b[0]\n" "ldr q10, [%[b_ptr0], #0x20]\n" - ".word 0x6f80e173 // udot v19.4s, v11.16b, v0.4b[0]\n" + ".inst 0x6f80e173 // udot v19.4s, v11.16b, v0.4b[0]\n" "add a_ptr2, a_ptr2, #0x20\n" - ".word 0x6f81e177 // udot v23.4s, v11.16b, v1.4b[0]\n" + ".inst 0x6f81e177 // udot v23.4s, v11.16b, v1.4b[0]\n" "add a_ptr3, a_ptr3, #0x20\n" - ".word 0x6f82e17b // udot v27.4s, v11.16b, v2.4b[0]\n" + ".inst 0x6f82e17b // udot v27.4s, v11.16b, v2.4b[0]\n" "prfm PLDL1KEEP, [a_ptr1, #0x40]\n" - ".word 0x6f83e17f // udot v31.4s, v11.16b, v3.4b[0]\n" + ".inst 0x6f83e17f // udot v31.4s, v11.16b, v3.4b[0]\n" "ldr q11, [%[b_ptr0], #0x30]\n" - ".word 0x6fa0e190 // udot v16.4s, v12.16b, v0.4b[1]\n" + ".inst 0x6fa0e190 // udot v16.4s, v12.16b, v0.4b[1]\n" "prfm PLDL1KEEP, [a_ptr2, #0x40]\n" - ".word 0x6fa1e194 // udot v20.4s, v12.16b, v1.4b[1]\n" + ".inst 0x6fa1e194 // udot v20.4s, v12.16b, v1.4b[1]\n" "prfm PLDL1KEEP, [a_ptr3, #0x40]\n" - ".word 0x6fa2e198 // udot v24.4s, v12.16b, v2.4b[1]\n" - ".word 0x6fa3e19c // udot v28.4s, v12.16b, v3.4b[1]\n" + ".inst 0x6fa2e198 // udot v24.4s, v12.16b, v2.4b[1]\n" + ".inst 0x6fa3e19c // udot v28.4s, v12.16b, v3.4b[1]\n" "ldr q12, [%[b_ptr0], #0x40]\n" - ".word 0x6fa0e1b1 // udot v17.4s, v13.16b, v0.4b[1]\n" - ".word 0x6fa1e1b5 // udot v21.4s, v13.16b, v1.4b[1]\n" - ".word 0x6fa2e1b9 // udot v25.4s, v13.16b, v2.4b[1]\n" - ".word 0x6fa3e1bd // udot v29.4s, v13.16b, v3.4b[1]\n" + ".inst 0x6fa0e1b1 // udot v17.4s, v13.16b, v0.4b[1]\n" + ".inst 0x6fa1e1b5 // udot v21.4s, v13.16b, v1.4b[1]\n" + ".inst 0x6fa2e1b9 // udot v25.4s, v13.16b, v2.4b[1]\n" + ".inst 0x6fa3e1bd // udot v29.4s, v13.16b, v3.4b[1]\n" "ldr q13, [%[b_ptr0], #0x50]\n" - ".word 0x6fa0e1d2 // udot v18.4s, v14.16b, v0.4b[1]\n" - ".word 0x6fa1e1d6 // udot v22.4s, v14.16b, v1.4b[1]\n" - ".word 0x6fa2e1da // udot v26.4s, v14.16b, v2.4b[1]\n" - ".word 0x6fa3e1de // udot v30.4s, v14.16b, v3.4b[1]\n" + ".inst 0x6fa0e1d2 // udot v18.4s, v14.16b, v0.4b[1]\n" + ".inst 0x6fa1e1d6 // udot v22.4s, v14.16b, v1.4b[1]\n" + ".inst 0x6fa2e1da // udot v26.4s, v14.16b, v2.4b[1]\n" + ".inst 0x6fa3e1de // udot v30.4s, v14.16b, v3.4b[1]\n" "ldr q14, [%[b_ptr0], #0x60]\n" - ".word 0x6fa0e1f3 // udot v19.4s, v15.16b, v0.4b[1]\n" - ".word 0x6fa1e1f7 // udot v23.4s, v15.16b, v1.4b[1]\n" - ".word 0x6fa2e1fb // udot v27.4s, v15.16b, v2.4b[1]\n" - ".word 0x6fa3e1ff // udot v31.4s, v15.16b, v3.4b[1]\n" + ".inst 0x6fa0e1f3 // udot v19.4s, v15.16b, v0.4b[1]\n" + ".inst 0x6fa1e1f7 // udot v23.4s, v15.16b, v1.4b[1]\n" + ".inst 0x6fa2e1fb // udot v27.4s, v15.16b, v2.4b[1]\n" + ".inst 0x6fa3e1ff // udot v31.4s, v15.16b, v3.4b[1]\n" "ldr q15, [%[b_ptr0], #0x70]\n" - ".word 0x6f80e910 // udot v16.4s, v8.16b, v0.4b[2]\n" + ".inst 0x6f80e910 // udot v16.4s, v8.16b, v0.4b[2]\n" "add %[b_ptr0], %[b_ptr0], #0x100\n" - ".word 0x6f81e914 // udot v20.4s, v8.16b, v1.4b[2]\n" - ".word 0x6f82e918 // udot v24.4s, v8.16b, v2.4b[2]\n" - ".word 0x6f83e91c // udot v28.4s, v8.16b, v3.4b[2]\n" + ".inst 0x6f81e914 // udot v20.4s, v8.16b, v1.4b[2]\n" + ".inst 0x6f82e918 // udot v24.4s, v8.16b, v2.4b[2]\n" + ".inst 0x6f83e91c // udot v28.4s, v8.16b, v3.4b[2]\n" "ldr q8, [%[b_ptr0], #-0x80]\n" - ".word 0x6f80e931 // udot v17.4s, v9.16b, v0.4b[2]\n" - ".word 0x6f81e935 // udot v21.4s, v9.16b, v1.4b[2]\n" - ".word 0x6f82e939 // udot v25.4s, v9.16b, v2.4b[2]\n" - ".word 0x6f83e93d // udot v29.4s, v9.16b, v3.4b[2]\n" + ".inst 0x6f80e931 // udot v17.4s, v9.16b, v0.4b[2]\n" + ".inst 0x6f81e935 // udot v21.4s, v9.16b, v1.4b[2]\n" + ".inst 0x6f82e939 // udot v25.4s, v9.16b, v2.4b[2]\n" + ".inst 0x6f83e93d // udot v29.4s, v9.16b, v3.4b[2]\n" "ldr q9, [%[b_ptr0], #-0x70]\n" - ".word 0x6f80e952 // udot v18.4s, v10.16b, v0.4b[2]\n" - ".word 0x6f81e956 // udot v22.4s, v10.16b, v1.4b[2]\n" - ".word 0x6f82e95a // udot v26.4s, v10.16b, v2.4b[2]\n" - ".word 0x6f83e95e // udot v30.4s, v10.16b, v3.4b[2]\n" + ".inst 0x6f80e952 // udot v18.4s, v10.16b, v0.4b[2]\n" + ".inst 0x6f81e956 // udot v22.4s, v10.16b, v1.4b[2]\n" + ".inst 0x6f82e95a // udot v26.4s, v10.16b, v2.4b[2]\n" + ".inst 0x6f83e95e // udot v30.4s, v10.16b, v3.4b[2]\n" "ldr q10, [%[b_ptr0], #-0x60]\n" - ".word 0x6f80e973 // udot v19.4s, v11.16b, v0.4b[2]\n" - ".word 0x6f81e977 // udot v23.4s, v11.16b, v1.4b[2]\n" - ".word 0x6f82e97b // udot v27.4s, v11.16b, v2.4b[2]\n" - ".word 0x6f83e97f // udot v31.4s, v11.16b, v3.4b[2]\n" + ".inst 0x6f80e973 // udot v19.4s, v11.16b, v0.4b[2]\n" + ".inst 0x6f81e977 // udot v23.4s, v11.16b, v1.4b[2]\n" + ".inst 0x6f82e97b // udot v27.4s, v11.16b, v2.4b[2]\n" + ".inst 0x6f83e97f // udot v31.4s, v11.16b, v3.4b[2]\n" "ldr q11, [%[b_ptr0], #-0x50]\n" - ".word 0x6fa0e990 // udot v16.4s, v12.16b, v0.4b[3]\n" - ".word 0x6fa1e994 // udot v20.4s, v12.16b, v1.4b[3]\n" - ".word 0x6fa2e998 // udot v24.4s, v12.16b, v2.4b[3]\n" - ".word 0x6fa3e99c // udot v28.4s, v12.16b, v3.4b[3]\n" + ".inst 0x6fa0e990 // udot v16.4s, v12.16b, v0.4b[3]\n" + ".inst 0x6fa1e994 // udot v20.4s, v12.16b, v1.4b[3]\n" + ".inst 0x6fa2e998 // udot v24.4s, v12.16b, v2.4b[3]\n" + ".inst 0x6fa3e99c // udot v28.4s, v12.16b, v3.4b[3]\n" "ldr q12, [%[b_ptr0], #-0x40]\n" - ".word 0x6fa0e9b1 // udot v17.4s, v13.16b, v0.4b[3]\n" - ".word 0x6fa1e9b5 // udot v21.4s, v13.16b, v1.4b[3]\n" - ".word 0x6fa2e9b9 // udot v25.4s, v13.16b, v2.4b[3]\n" - ".word 0x6fa3e9bd // udot v29.4s, v13.16b, v3.4b[3]\n" + ".inst 0x6fa0e9b1 // udot v17.4s, v13.16b, v0.4b[3]\n" + ".inst 0x6fa1e9b5 // udot v21.4s, v13.16b, v1.4b[3]\n" + ".inst 0x6fa2e9b9 // udot v25.4s, v13.16b, v2.4b[3]\n" + ".inst 0x6fa3e9bd // udot v29.4s, v13.16b, v3.4b[3]\n" "ldr q13, [%[b_ptr0], #-0x30]\n" - ".word 0x6fa0e9d2 // udot v18.4s, v14.16b, v0.4b[3]\n" - ".word 0x6fa1e9d6 // udot v22.4s, v14.16b, v1.4b[3]\n" - ".word 0x6fa2e9da // udot v26.4s, v14.16b, v2.4b[3]\n" - ".word 0x6fa3e9de // udot v30.4s, v14.16b, v3.4b[3]\n" + ".inst 0x6fa0e9d2 // udot v18.4s, v14.16b, v0.4b[3]\n" + ".inst 0x6fa1e9d6 // udot v22.4s, v14.16b, v1.4b[3]\n" + ".inst 0x6fa2e9da // udot v26.4s, v14.16b, v2.4b[3]\n" + ".inst 0x6fa3e9de // udot v30.4s, v14.16b, v3.4b[3]\n" "ldr q14, [%[b_ptr0], #-0x20]\n" - ".word 0x6fa0e9f3 // udot v19.4s, v15.16b, v0.4b[3]\n" + ".inst 0x6fa0e9f3 // udot v19.4s, v15.16b, v0.4b[3]\n" "ldr q0, [%[a_ptr0], #-0x10]\n" - ".word 0x6fa1e9f7 // udot v23.4s, v15.16b, v1.4b[3]\n" + ".inst 0x6fa1e9f7 // udot v23.4s, v15.16b, v1.4b[3]\n" "ldr q1, [a_ptr1, #-0x10]\n" - ".word 0x6fa2e9fb // udot v27.4s, v15.16b, v2.4b[3]\n" + ".inst 0x6fa2e9fb // udot v27.4s, v15.16b, v2.4b[3]\n" "ldr q2, [a_ptr2, #-0x10]\n" - ".word 0x6fa3e9ff // udot v31.4s, v15.16b, v3.4b[3]\n" + ".inst 0x6fa3e9ff // udot v31.4s, v15.16b, v3.4b[3]\n" "ldr q15, [%[b_ptr0], #-0x10]\n" - ".word 0x6f84e110 // udot v16.4s, v8.16b, v4.4b[0]\n" + ".inst 0x6f84e110 // udot v16.4s, v8.16b, v4.4b[0]\n" "ldr q3, [a_ptr3, #-0x10]\n" - ".word 0x6f85e114 // udot v20.4s, v8.16b, v5.4b[0]\n" - ".word 0x6f86e118 // udot v24.4s, v8.16b, v6.4b[0]\n" - ".word 0x6f87e11c // udot v28.4s, v8.16b, v7.4b[0]\n" + ".inst 0x6f85e114 // udot v20.4s, v8.16b, v5.4b[0]\n" + ".inst 0x6f86e118 // udot v24.4s, v8.16b, v6.4b[0]\n" + ".inst 0x6f87e11c // udot v28.4s, v8.16b, v7.4b[0]\n" "ldr q8, [%[b_ptr0]]\n" - ".word 0x6f84e131 // udot v17.4s, v9.16b, v4.4b[0]\n" - ".word 0x6f85e135 // udot v21.4s, v9.16b, v5.4b[0]\n" - ".word 0x6f86e139 // udot v25.4s, v9.16b, v6.4b[0]\n" - ".word 0x6f87e13d // udot v29.4s, v9.16b, v7.4b[0]\n" + ".inst 0x6f84e131 // udot v17.4s, v9.16b, v4.4b[0]\n" + ".inst 0x6f85e135 // udot v21.4s, v9.16b, v5.4b[0]\n" + ".inst 0x6f86e139 // udot v25.4s, v9.16b, v6.4b[0]\n" + ".inst 0x6f87e13d // udot v29.4s, v9.16b, v7.4b[0]\n" "ldr q9, [%[b_ptr0], #0x10]\n" - ".word 0x6f84e152 // udot v18.4s, v10.16b, v4.4b[0]\n" - ".word 0x6f85e156 // udot v22.4s, v10.16b, v5.4b[0]\n" - ".word 0x6f86e15a // udot v26.4s, v10.16b, v6.4b[0]\n" - ".word 0x6f87e15e // udot v30.4s, v10.16b, v7.4b[0]\n" + ".inst 0x6f84e152 // udot v18.4s, v10.16b, v4.4b[0]\n" + ".inst 0x6f85e156 // udot v22.4s, v10.16b, v5.4b[0]\n" + ".inst 0x6f86e15a // udot v26.4s, v10.16b, v6.4b[0]\n" + ".inst 0x6f87e15e // udot v30.4s, v10.16b, v7.4b[0]\n" "ldr q10, [%[b_ptr0], #0x20]\n" - ".word 0x6f84e173 // udot v19.4s, v11.16b, v4.4b[0]\n" - ".word 0x6f85e177 // udot v23.4s, v11.16b, v5.4b[0]\n" - ".word 0x6f86e17b // udot v27.4s, v11.16b, v6.4b[0]\n" - ".word 0x6f87e17f // udot v31.4s, v11.16b, v7.4b[0]\n" + ".inst 0x6f84e173 // udot v19.4s, v11.16b, v4.4b[0]\n" + ".inst 0x6f85e177 // udot v23.4s, v11.16b, v5.4b[0]\n" + ".inst 0x6f86e17b // udot v27.4s, v11.16b, v6.4b[0]\n" + ".inst 0x6f87e17f // udot v31.4s, v11.16b, v7.4b[0]\n" "ldr q11, [%[b_ptr0], #0x30]\n" - ".word 0x6fa4e190 // udot v16.4s, v12.16b, v4.4b[1]\n" - ".word 0x6fa5e194 // udot v20.4s, v12.16b, v5.4b[1]\n" - ".word 0x6fa6e198 // udot v24.4s, v12.16b, v6.4b[1]\n" - ".word 0x6fa7e19c // udot v28.4s, v12.16b, v7.4b[1]\n" + ".inst 0x6fa4e190 // udot v16.4s, v12.16b, v4.4b[1]\n" + ".inst 0x6fa5e194 // udot v20.4s, v12.16b, v5.4b[1]\n" + ".inst 0x6fa6e198 // udot v24.4s, v12.16b, v6.4b[1]\n" + ".inst 0x6fa7e19c // udot v28.4s, v12.16b, v7.4b[1]\n" "ldr q12, [%[b_ptr0], #0x40]\n" - ".word 0x6fa4e1b1 // udot v17.4s, v13.16b, v4.4b[1]\n" - ".word 0x6fa5e1b5 // udot v21.4s, v13.16b, v5.4b[1]\n" - ".word 0x6fa6e1b9 // udot v25.4s, v13.16b, v6.4b[1]\n" - ".word 0x6fa7e1bd // udot v29.4s, v13.16b, v7.4b[1]\n" + ".inst 0x6fa4e1b1 // udot v17.4s, v13.16b, v4.4b[1]\n" + ".inst 0x6fa5e1b5 // udot v21.4s, v13.16b, v5.4b[1]\n" + ".inst 0x6fa6e1b9 // udot v25.4s, v13.16b, v6.4b[1]\n" + ".inst 0x6fa7e1bd // udot v29.4s, v13.16b, v7.4b[1]\n" "ldr q13, [%[b_ptr0], #0x50]\n" - ".word 0x6fa4e1d2 // udot v18.4s, v14.16b, v4.4b[1]\n" - ".word 0x6fa5e1d6 // udot v22.4s, v14.16b, v5.4b[1]\n" - ".word 0x6fa6e1da // udot v26.4s, v14.16b, v6.4b[1]\n" - ".word 0x6fa7e1de // udot v30.4s, v14.16b, v7.4b[1]\n" + ".inst 0x6fa4e1d2 // udot v18.4s, v14.16b, v4.4b[1]\n" + ".inst 0x6fa5e1d6 // udot v22.4s, v14.16b, v5.4b[1]\n" + ".inst 0x6fa6e1da // udot v26.4s, v14.16b, v6.4b[1]\n" + ".inst 0x6fa7e1de // udot v30.4s, v14.16b, v7.4b[1]\n" "ldr q14, [%[b_ptr0], #0x60]\n" - ".word 0x6fa4e1f3 // udot v19.4s, v15.16b, v4.4b[1]\n" - ".word 0x6fa5e1f7 // udot v23.4s, v15.16b, v5.4b[1]\n" - ".word 0x6fa6e1fb // udot v27.4s, v15.16b, v6.4b[1]\n" - ".word 0x6fa7e1ff // udot v31.4s, v15.16b, v7.4b[1]\n" + ".inst 0x6fa4e1f3 // udot v19.4s, v15.16b, v4.4b[1]\n" + ".inst 0x6fa5e1f7 // udot v23.4s, v15.16b, v5.4b[1]\n" + ".inst 0x6fa6e1fb // udot v27.4s, v15.16b, v6.4b[1]\n" + ".inst 0x6fa7e1ff // udot v31.4s, v15.16b, v7.4b[1]\n" "ldr q15, [%[b_ptr0], #0x70]\n" - ".word 0x6f84e910 // udot v16.4s, v8.16b, v4.4b[2]\n" + ".inst 0x6f84e910 // udot v16.4s, v8.16b, v4.4b[2]\n" "add %[b_ptr0], %[b_ptr0], #0x100\n" - ".word 0x6f85e914 // udot v20.4s, v8.16b, v5.4b[2]\n" - ".word 0x6f86e918 // udot v24.4s, v8.16b, v6.4b[2]\n" - ".word 0x6f87e91c // udot v28.4s, v8.16b, v7.4b[2]\n" + ".inst 0x6f85e914 // udot v20.4s, v8.16b, v5.4b[2]\n" + ".inst 0x6f86e918 // udot v24.4s, v8.16b, v6.4b[2]\n" + ".inst 0x6f87e91c // udot v28.4s, v8.16b, v7.4b[2]\n" "ldr q8, [%[b_ptr0], #-0x80]\n" - ".word 0x6f84e931 // udot v17.4s, v9.16b, v4.4b[2]\n" - ".word 0x6f85e935 // udot v21.4s, v9.16b, v5.4b[2]\n" - ".word 0x6f86e939 // udot v25.4s, v9.16b, v6.4b[2]\n" - ".word 0x6f87e93d // udot v29.4s, v9.16b, v7.4b[2]\n" + ".inst 0x6f84e931 // udot v17.4s, v9.16b, v4.4b[2]\n" + ".inst 0x6f85e935 // udot v21.4s, v9.16b, v5.4b[2]\n" + ".inst 0x6f86e939 // udot v25.4s, v9.16b, v6.4b[2]\n" + ".inst 0x6f87e93d // udot v29.4s, v9.16b, v7.4b[2]\n" "ldr q9, [%[b_ptr0], #-0x70]\n" - ".word 0x6f84e952 // udot v18.4s, v10.16b, v4.4b[2]\n" - ".word 0x6f85e956 // udot v22.4s, v10.16b, v5.4b[2]\n" - ".word 0x6f86e95a // udot v26.4s, v10.16b, v6.4b[2]\n" - ".word 0x6f87e95e // udot v30.4s, v10.16b, v7.4b[2]\n" + ".inst 0x6f84e952 // udot v18.4s, v10.16b, v4.4b[2]\n" + ".inst 0x6f85e956 // udot v22.4s, v10.16b, v5.4b[2]\n" + ".inst 0x6f86e95a // udot v26.4s, v10.16b, v6.4b[2]\n" + ".inst 0x6f87e95e // udot v30.4s, v10.16b, v7.4b[2]\n" "ldr q10, [%[b_ptr0], #-0x60]\n" - ".word 0x6f84e973 // udot v19.4s, v11.16b, v4.4b[2]\n" - ".word 0x6f85e977 // udot v23.4s, v11.16b, v5.4b[2]\n" - ".word 0x6f86e97b // udot v27.4s, v11.16b, v6.4b[2]\n" - ".word 0x6f87e97f // udot v31.4s, v11.16b, v7.4b[2]\n" + ".inst 0x6f84e973 // udot v19.4s, v11.16b, v4.4b[2]\n" + ".inst 0x6f85e977 // udot v23.4s, v11.16b, v5.4b[2]\n" + ".inst 0x6f86e97b // udot v27.4s, v11.16b, v6.4b[2]\n" + ".inst 0x6f87e97f // udot v31.4s, v11.16b, v7.4b[2]\n" "ldr q11, [%[b_ptr0], #-0x50]\n" - ".word 0x6fa4e990 // udot v16.4s, v12.16b, v4.4b[3]\n" - ".word 0x6fa5e994 // udot v20.4s, v12.16b, v5.4b[3]\n" - ".word 0x6fa6e998 // udot v24.4s, v12.16b, v6.4b[3]\n" - ".word 0x6fa7e99c // udot v28.4s, v12.16b, v7.4b[3]\n" + ".inst 0x6fa4e990 // udot v16.4s, v12.16b, v4.4b[3]\n" + ".inst 0x6fa5e994 // udot v20.4s, v12.16b, v5.4b[3]\n" + ".inst 0x6fa6e998 // udot v24.4s, v12.16b, v6.4b[3]\n" + ".inst 0x6fa7e99c // udot v28.4s, v12.16b, v7.4b[3]\n" "ldr q12, [%[b_ptr0], #-0x40]\n" - ".word 0x6fa4e9b1 // udot v17.4s, v13.16b, v4.4b[3]\n" - ".word 0x6fa5e9b5 // udot v21.4s, v13.16b, v5.4b[3]\n" - ".word 0x6fa6e9b9 // udot v25.4s, v13.16b, v6.4b[3]\n" - ".word 0x6fa7e9bd // udot v29.4s, v13.16b, v7.4b[3]\n" + ".inst 0x6fa4e9b1 // udot v17.4s, v13.16b, v4.4b[3]\n" + ".inst 0x6fa5e9b5 // udot v21.4s, v13.16b, v5.4b[3]\n" + ".inst 0x6fa6e9b9 // udot v25.4s, v13.16b, v6.4b[3]\n" + ".inst 0x6fa7e9bd // udot v29.4s, v13.16b, v7.4b[3]\n" "ldr q13, [%[b_ptr0], #-0x30]\n" - ".word 0x6fa4e9d2 // udot v18.4s, v14.16b, v4.4b[3]\n" - ".word 0x6fa5e9d6 // udot v22.4s, v14.16b, v5.4b[3]\n" - ".word 0x6fa6e9da // udot v26.4s, v14.16b, v6.4b[3]\n" - ".word 0x6fa7e9de // udot v30.4s, v14.16b, v7.4b[3]\n" + ".inst 0x6fa4e9d2 // udot v18.4s, v14.16b, v4.4b[3]\n" + ".inst 0x6fa5e9d6 // udot v22.4s, v14.16b, v5.4b[3]\n" + ".inst 0x6fa6e9da // udot v26.4s, v14.16b, v6.4b[3]\n" + ".inst 0x6fa7e9de // udot v30.4s, v14.16b, v7.4b[3]\n" "ldr q14, [%[b_ptr0], #-0x20]\n" - ".word 0x6fa4e9f3 // udot v19.4s, v15.16b, v4.4b[3]\n" - ".word 0x6fa5e9f7 // udot v23.4s, v15.16b, v5.4b[3]\n" - ".word 0x6fa6e9fb // udot v27.4s, v15.16b, v6.4b[3]\n" - ".word 0x6fa7e9ff // udot v31.4s, v15.16b, v7.4b[3]\n" + ".inst 0x6fa4e9f3 // udot v19.4s, v15.16b, v4.4b[3]\n" + ".inst 0x6fa5e9f7 // udot v23.4s, v15.16b, v5.4b[3]\n" + ".inst 0x6fa6e9fb // udot v27.4s, v15.16b, v6.4b[3]\n" + ".inst 0x6fa7e9ff // udot v31.4s, v15.16b, v7.4b[3]\n" "b.ne 3b\n" "2:\n" "ldr q15, [%[b_ptr0], #-0x10]\n" @@ -1486,243 +1444,243 @@ void a64_hybrid_u8u32_dot_16x4(const uint8_t *A, int lda, const uint8_t *B, uint "prfm PSTL1KEEP, [c_ptr2]\n" "prfm PSTL1KEEP, [c_ptr3]\n" "cbz %[regs], 4f\n" - ".word 0x6f80e110 // udot v16.4s, v8.16b, v0.4b[0]\n" + ".inst 0x6f80e110 // udot v16.4s, v8.16b, v0.4b[0]\n" "ldr q4, [%[a_ptr0]]\n" - ".word 0x6f81e114 // udot v20.4s, v8.16b, v1.4b[0]\n" + ".inst 0x6f81e114 // udot v20.4s, v8.16b, v1.4b[0]\n" "ldr q5, [a_ptr1]\n" - ".word 0x6f82e118 // udot v24.4s, v8.16b, v2.4b[0]\n" + ".inst 0x6f82e118 // udot v24.4s, v8.16b, v2.4b[0]\n" "ldr q6, [a_ptr2]\n" - ".word 0x6f83e11c // udot v28.4s, v8.16b, v3.4b[0]\n" + ".inst 0x6f83e11c // udot v28.4s, v8.16b, v3.4b[0]\n" "ldr q7, [a_ptr3]\n" - ".word 0x6f80e131 // udot v17.4s, v9.16b, v0.4b[0]\n" + ".inst 0x6f80e131 // udot v17.4s, v9.16b, v0.4b[0]\n" "ldr q8, [%[b_ptr0]]\n" - ".word 0x6f81e135 // udot v21.4s, v9.16b, v1.4b[0]\n" + ".inst 0x6f81e135 // udot v21.4s, v9.16b, v1.4b[0]\n" "add %[a_ptr0], %[a_ptr0], #0x10\n" - ".word 0x6f82e139 // udot v25.4s, v9.16b, v2.4b[0]\n" + ".inst 0x6f82e139 // udot v25.4s, v9.16b, v2.4b[0]\n" "add a_ptr1, a_ptr1, #0x10\n" - ".word 0x6f83e13d // udot v29.4s, v9.16b, v3.4b[0]\n" + ".inst 0x6f83e13d // udot v29.4s, v9.16b, v3.4b[0]\n" "ldr q9, [%[b_ptr0], #0x10]\n" - ".word 0x6f80e152 // udot v18.4s, v10.16b, v0.4b[0]\n" + ".inst 0x6f80e152 // udot v18.4s, v10.16b, v0.4b[0]\n" "add a_ptr2, a_ptr2, #0x10\n" - ".word 0x6f81e156 // udot v22.4s, v10.16b, v1.4b[0]\n" + ".inst 0x6f81e156 // udot v22.4s, v10.16b, v1.4b[0]\n" "add a_ptr3, a_ptr3, #0x10\n" - ".word 0x6f82e15a // udot v26.4s, v10.16b, v2.4b[0]\n" - ".word 0x6f83e15e // udot v30.4s, v10.16b, v3.4b[0]\n" + ".inst 0x6f82e15a // udot v26.4s, v10.16b, v2.4b[0]\n" + ".inst 0x6f83e15e // udot v30.4s, v10.16b, v3.4b[0]\n" "ldr q10, [%[b_ptr0], #0x20]\n" - ".word 0x6f80e173 // udot v19.4s, v11.16b, v0.4b[0]\n" - ".word 0x6f81e177 // udot v23.4s, v11.16b, v1.4b[0]\n" - ".word 0x6f82e17b // udot v27.4s, v11.16b, v2.4b[0]\n" - ".word 0x6f83e17f // udot v31.4s, v11.16b, v3.4b[0]\n" + ".inst 0x6f80e173 // udot v19.4s, v11.16b, v0.4b[0]\n" + ".inst 0x6f81e177 // udot v23.4s, v11.16b, v1.4b[0]\n" + ".inst 0x6f82e17b // udot v27.4s, v11.16b, v2.4b[0]\n" + ".inst 0x6f83e17f // udot v31.4s, v11.16b, v3.4b[0]\n" "ldr q11, [%[b_ptr0], #0x30]\n" - ".word 0x6fa0e190 // udot v16.4s, v12.16b, v0.4b[1]\n" - ".word 0x6fa1e194 // udot v20.4s, v12.16b, v1.4b[1]\n" - ".word 0x6fa2e198 // udot v24.4s, v12.16b, v2.4b[1]\n" - ".word 0x6fa3e19c // udot v28.4s, v12.16b, v3.4b[1]\n" + ".inst 0x6fa0e190 // udot v16.4s, v12.16b, v0.4b[1]\n" + ".inst 0x6fa1e194 // udot v20.4s, v12.16b, v1.4b[1]\n" + ".inst 0x6fa2e198 // udot v24.4s, v12.16b, v2.4b[1]\n" + ".inst 0x6fa3e19c // udot v28.4s, v12.16b, v3.4b[1]\n" "ldr q12, [%[b_ptr0], #0x40]\n" - ".word 0x6fa0e1b1 // udot v17.4s, v13.16b, v0.4b[1]\n" - ".word 0x6fa1e1b5 // udot v21.4s, v13.16b, v1.4b[1]\n" - ".word 0x6fa2e1b9 // udot v25.4s, v13.16b, v2.4b[1]\n" - ".word 0x6fa3e1bd // udot v29.4s, v13.16b, v3.4b[1]\n" + ".inst 0x6fa0e1b1 // udot v17.4s, v13.16b, v0.4b[1]\n" + ".inst 0x6fa1e1b5 // udot v21.4s, v13.16b, v1.4b[1]\n" + ".inst 0x6fa2e1b9 // udot v25.4s, v13.16b, v2.4b[1]\n" + ".inst 0x6fa3e1bd // udot v29.4s, v13.16b, v3.4b[1]\n" "ldr q13, [%[b_ptr0], #0x50]\n" - ".word 0x6fa0e1d2 // udot v18.4s, v14.16b, v0.4b[1]\n" - ".word 0x6fa1e1d6 // udot v22.4s, v14.16b, v1.4b[1]\n" - ".word 0x6fa2e1da // udot v26.4s, v14.16b, v2.4b[1]\n" - ".word 0x6fa3e1de // udot v30.4s, v14.16b, v3.4b[1]\n" + ".inst 0x6fa0e1d2 // udot v18.4s, v14.16b, v0.4b[1]\n" + ".inst 0x6fa1e1d6 // udot v22.4s, v14.16b, v1.4b[1]\n" + ".inst 0x6fa2e1da // udot v26.4s, v14.16b, v2.4b[1]\n" + ".inst 0x6fa3e1de // udot v30.4s, v14.16b, v3.4b[1]\n" "ldr q14, [%[b_ptr0], #0x60]\n" - ".word 0x6fa0e1f3 // udot v19.4s, v15.16b, v0.4b[1]\n" - ".word 0x6fa1e1f7 // udot v23.4s, v15.16b, v1.4b[1]\n" - ".word 0x6fa2e1fb // udot v27.4s, v15.16b, v2.4b[1]\n" - ".word 0x6fa3e1ff // udot v31.4s, v15.16b, v3.4b[1]\n" + ".inst 0x6fa0e1f3 // udot v19.4s, v15.16b, v0.4b[1]\n" + ".inst 0x6fa1e1f7 // udot v23.4s, v15.16b, v1.4b[1]\n" + ".inst 0x6fa2e1fb // udot v27.4s, v15.16b, v2.4b[1]\n" + ".inst 0x6fa3e1ff // udot v31.4s, v15.16b, v3.4b[1]\n" "ldr q15, [%[b_ptr0], #0x70]\n" - ".word 0x6f80e910 // udot v16.4s, v8.16b, v0.4b[2]\n" + ".inst 0x6f80e910 // udot v16.4s, v8.16b, v0.4b[2]\n" "add %[b_ptr0], %[b_ptr0], #0x100\n" - ".word 0x6f81e914 // udot v20.4s, v8.16b, v1.4b[2]\n" - ".word 0x6f82e918 // udot v24.4s, v8.16b, v2.4b[2]\n" - ".word 0x6f83e91c // udot v28.4s, v8.16b, v3.4b[2]\n" + ".inst 0x6f81e914 // udot v20.4s, v8.16b, v1.4b[2]\n" + ".inst 0x6f82e918 // udot v24.4s, v8.16b, v2.4b[2]\n" + ".inst 0x6f83e91c // udot v28.4s, v8.16b, v3.4b[2]\n" "ldr q8, [%[b_ptr0], #-0x80]\n" - ".word 0x6f80e931 // udot v17.4s, v9.16b, v0.4b[2]\n" - ".word 0x6f81e935 // udot v21.4s, v9.16b, v1.4b[2]\n" - ".word 0x6f82e939 // udot v25.4s, v9.16b, v2.4b[2]\n" - ".word 0x6f83e93d // udot v29.4s, v9.16b, v3.4b[2]\n" + ".inst 0x6f80e931 // udot v17.4s, v9.16b, v0.4b[2]\n" + ".inst 0x6f81e935 // udot v21.4s, v9.16b, v1.4b[2]\n" + ".inst 0x6f82e939 // udot v25.4s, v9.16b, v2.4b[2]\n" + ".inst 0x6f83e93d // udot v29.4s, v9.16b, v3.4b[2]\n" "ldr q9, [%[b_ptr0], #-0x70]\n" - ".word 0x6f80e952 // udot v18.4s, v10.16b, v0.4b[2]\n" - ".word 0x6f81e956 // udot v22.4s, v10.16b, v1.4b[2]\n" - ".word 0x6f82e95a // udot v26.4s, v10.16b, v2.4b[2]\n" - ".word 0x6f83e95e // udot v30.4s, v10.16b, v3.4b[2]\n" + ".inst 0x6f80e952 // udot v18.4s, v10.16b, v0.4b[2]\n" + ".inst 0x6f81e956 // udot v22.4s, v10.16b, v1.4b[2]\n" + ".inst 0x6f82e95a // udot v26.4s, v10.16b, v2.4b[2]\n" + ".inst 0x6f83e95e // udot v30.4s, v10.16b, v3.4b[2]\n" "ldr q10, [%[b_ptr0], #-0x60]\n" - ".word 0x6f80e973 // udot v19.4s, v11.16b, v0.4b[2]\n" - ".word 0x6f81e977 // udot v23.4s, v11.16b, v1.4b[2]\n" - ".word 0x6f82e97b // udot v27.4s, v11.16b, v2.4b[2]\n" - ".word 0x6f83e97f // udot v31.4s, v11.16b, v3.4b[2]\n" + ".inst 0x6f80e973 // udot v19.4s, v11.16b, v0.4b[2]\n" + ".inst 0x6f81e977 // udot v23.4s, v11.16b, v1.4b[2]\n" + ".inst 0x6f82e97b // udot v27.4s, v11.16b, v2.4b[2]\n" + ".inst 0x6f83e97f // udot v31.4s, v11.16b, v3.4b[2]\n" "ldr q11, [%[b_ptr0], #-0x50]\n" - ".word 0x6fa0e990 // udot v16.4s, v12.16b, v0.4b[3]\n" - ".word 0x6fa1e994 // udot v20.4s, v12.16b, v1.4b[3]\n" - ".word 0x6fa2e998 // udot v24.4s, v12.16b, v2.4b[3]\n" - ".word 0x6fa3e99c // udot v28.4s, v12.16b, v3.4b[3]\n" + ".inst 0x6fa0e990 // udot v16.4s, v12.16b, v0.4b[3]\n" + ".inst 0x6fa1e994 // udot v20.4s, v12.16b, v1.4b[3]\n" + ".inst 0x6fa2e998 // udot v24.4s, v12.16b, v2.4b[3]\n" + ".inst 0x6fa3e99c // udot v28.4s, v12.16b, v3.4b[3]\n" "ldr q12, [%[b_ptr0], #-0x40]\n" - ".word 0x6fa0e9b1 // udot v17.4s, v13.16b, v0.4b[3]\n" - ".word 0x6fa1e9b5 // udot v21.4s, v13.16b, v1.4b[3]\n" - ".word 0x6fa2e9b9 // udot v25.4s, v13.16b, v2.4b[3]\n" - ".word 0x6fa3e9bd // udot v29.4s, v13.16b, v3.4b[3]\n" + ".inst 0x6fa0e9b1 // udot v17.4s, v13.16b, v0.4b[3]\n" + ".inst 0x6fa1e9b5 // udot v21.4s, v13.16b, v1.4b[3]\n" + ".inst 0x6fa2e9b9 // udot v25.4s, v13.16b, v2.4b[3]\n" + ".inst 0x6fa3e9bd // udot v29.4s, v13.16b, v3.4b[3]\n" "ldr q13, [%[b_ptr0], #-0x30]\n" - ".word 0x6fa0e9d2 // udot v18.4s, v14.16b, v0.4b[3]\n" - ".word 0x6fa1e9d6 // udot v22.4s, v14.16b, v1.4b[3]\n" - ".word 0x6fa2e9da // udot v26.4s, v14.16b, v2.4b[3]\n" - ".word 0x6fa3e9de // udot v30.4s, v14.16b, v3.4b[3]\n" + ".inst 0x6fa0e9d2 // udot v18.4s, v14.16b, v0.4b[3]\n" + ".inst 0x6fa1e9d6 // udot v22.4s, v14.16b, v1.4b[3]\n" + ".inst 0x6fa2e9da // udot v26.4s, v14.16b, v2.4b[3]\n" + ".inst 0x6fa3e9de // udot v30.4s, v14.16b, v3.4b[3]\n" "ldr q14, [%[b_ptr0], #-0x20]\n" - ".word 0x6fa0e9f3 // udot v19.4s, v15.16b, v0.4b[3]\n" - ".word 0x6fa1e9f7 // udot v23.4s, v15.16b, v1.4b[3]\n" - ".word 0x6fa2e9fb // udot v27.4s, v15.16b, v2.4b[3]\n" - ".word 0x6fa3e9ff // udot v31.4s, v15.16b, v3.4b[3]\n" + ".inst 0x6fa0e9f3 // udot v19.4s, v15.16b, v0.4b[3]\n" + ".inst 0x6fa1e9f7 // udot v23.4s, v15.16b, v1.4b[3]\n" + ".inst 0x6fa2e9fb // udot v27.4s, v15.16b, v2.4b[3]\n" + ".inst 0x6fa3e9ff // udot v31.4s, v15.16b, v3.4b[3]\n" "ldr q15, [%[b_ptr0], #-0x10]\n" - ".word 0x6f84e110 // udot v16.4s, v8.16b, v4.4b[0]\n" - ".word 0x6f85e114 // udot v20.4s, v8.16b, v5.4b[0]\n" - ".word 0x6f86e118 // udot v24.4s, v8.16b, v6.4b[0]\n" - ".word 0x6f87e11c // udot v28.4s, v8.16b, v7.4b[0]\n" + ".inst 0x6f84e110 // udot v16.4s, v8.16b, v4.4b[0]\n" + ".inst 0x6f85e114 // udot v20.4s, v8.16b, v5.4b[0]\n" + ".inst 0x6f86e118 // udot v24.4s, v8.16b, v6.4b[0]\n" + ".inst 0x6f87e11c // udot v28.4s, v8.16b, v7.4b[0]\n" "ldr q8, [%[b_ptr0]]\n" - ".word 0x6f84e131 // udot v17.4s, v9.16b, v4.4b[0]\n" - ".word 0x6f85e135 // udot v21.4s, v9.16b, v5.4b[0]\n" - ".word 0x6f86e139 // udot v25.4s, v9.16b, v6.4b[0]\n" - ".word 0x6f87e13d // udot v29.4s, v9.16b, v7.4b[0]\n" + ".inst 0x6f84e131 // udot v17.4s, v9.16b, v4.4b[0]\n" + ".inst 0x6f85e135 // udot v21.4s, v9.16b, v5.4b[0]\n" + ".inst 0x6f86e139 // udot v25.4s, v9.16b, v6.4b[0]\n" + ".inst 0x6f87e13d // udot v29.4s, v9.16b, v7.4b[0]\n" "ldr q9, [%[b_ptr0], #0x10]\n" - ".word 0x6f84e152 // udot v18.4s, v10.16b, v4.4b[0]\n" - ".word 0x6f85e156 // udot v22.4s, v10.16b, v5.4b[0]\n" - ".word 0x6f86e15a // udot v26.4s, v10.16b, v6.4b[0]\n" - ".word 0x6f87e15e // udot v30.4s, v10.16b, v7.4b[0]\n" + ".inst 0x6f84e152 // udot v18.4s, v10.16b, v4.4b[0]\n" + ".inst 0x6f85e156 // udot v22.4s, v10.16b, v5.4b[0]\n" + ".inst 0x6f86e15a // udot v26.4s, v10.16b, v6.4b[0]\n" + ".inst 0x6f87e15e // udot v30.4s, v10.16b, v7.4b[0]\n" "ldr q10, [%[b_ptr0], #0x20]\n" - ".word 0x6f84e173 // udot v19.4s, v11.16b, v4.4b[0]\n" - ".word 0x6f85e177 // udot v23.4s, v11.16b, v5.4b[0]\n" - ".word 0x6f86e17b // udot v27.4s, v11.16b, v6.4b[0]\n" - ".word 0x6f87e17f // udot v31.4s, v11.16b, v7.4b[0]\n" + ".inst 0x6f84e173 // udot v19.4s, v11.16b, v4.4b[0]\n" + ".inst 0x6f85e177 // udot v23.4s, v11.16b, v5.4b[0]\n" + ".inst 0x6f86e17b // udot v27.4s, v11.16b, v6.4b[0]\n" + ".inst 0x6f87e17f // udot v31.4s, v11.16b, v7.4b[0]\n" "ldr q11, [%[b_ptr0], #0x30]\n" - ".word 0x6fa4e190 // udot v16.4s, v12.16b, v4.4b[1]\n" - ".word 0x6fa5e194 // udot v20.4s, v12.16b, v5.4b[1]\n" - ".word 0x6fa6e198 // udot v24.4s, v12.16b, v6.4b[1]\n" - ".word 0x6fa7e19c // udot v28.4s, v12.16b, v7.4b[1]\n" + ".inst 0x6fa4e190 // udot v16.4s, v12.16b, v4.4b[1]\n" + ".inst 0x6fa5e194 // udot v20.4s, v12.16b, v5.4b[1]\n" + ".inst 0x6fa6e198 // udot v24.4s, v12.16b, v6.4b[1]\n" + ".inst 0x6fa7e19c // udot v28.4s, v12.16b, v7.4b[1]\n" "ldr q12, [%[b_ptr0], #0x40]\n" - ".word 0x6fa4e1b1 // udot v17.4s, v13.16b, v4.4b[1]\n" - ".word 0x6fa5e1b5 // udot v21.4s, v13.16b, v5.4b[1]\n" - ".word 0x6fa6e1b9 // udot v25.4s, v13.16b, v6.4b[1]\n" - ".word 0x6fa7e1bd // udot v29.4s, v13.16b, v7.4b[1]\n" + ".inst 0x6fa4e1b1 // udot v17.4s, v13.16b, v4.4b[1]\n" + ".inst 0x6fa5e1b5 // udot v21.4s, v13.16b, v5.4b[1]\n" + ".inst 0x6fa6e1b9 // udot v25.4s, v13.16b, v6.4b[1]\n" + ".inst 0x6fa7e1bd // udot v29.4s, v13.16b, v7.4b[1]\n" "ldr q13, [%[b_ptr0], #0x50]\n" - ".word 0x6fa4e1d2 // udot v18.4s, v14.16b, v4.4b[1]\n" - ".word 0x6fa5e1d6 // udot v22.4s, v14.16b, v5.4b[1]\n" - ".word 0x6fa6e1da // udot v26.4s, v14.16b, v6.4b[1]\n" - ".word 0x6fa7e1de // udot v30.4s, v14.16b, v7.4b[1]\n" + ".inst 0x6fa4e1d2 // udot v18.4s, v14.16b, v4.4b[1]\n" + ".inst 0x6fa5e1d6 // udot v22.4s, v14.16b, v5.4b[1]\n" + ".inst 0x6fa6e1da // udot v26.4s, v14.16b, v6.4b[1]\n" + ".inst 0x6fa7e1de // udot v30.4s, v14.16b, v7.4b[1]\n" "ldr q14, [%[b_ptr0], #0x60]\n" - ".word 0x6fa4e1f3 // udot v19.4s, v15.16b, v4.4b[1]\n" - ".word 0x6fa5e1f7 // udot v23.4s, v15.16b, v5.4b[1]\n" - ".word 0x6fa6e1fb // udot v27.4s, v15.16b, v6.4b[1]\n" - ".word 0x6fa7e1ff // udot v31.4s, v15.16b, v7.4b[1]\n" + ".inst 0x6fa4e1f3 // udot v19.4s, v15.16b, v4.4b[1]\n" + ".inst 0x6fa5e1f7 // udot v23.4s, v15.16b, v5.4b[1]\n" + ".inst 0x6fa6e1fb // udot v27.4s, v15.16b, v6.4b[1]\n" + ".inst 0x6fa7e1ff // udot v31.4s, v15.16b, v7.4b[1]\n" "ldr q15, [%[b_ptr0], #0x70]\n" - ".word 0x6f84e910 // udot v16.4s, v8.16b, v4.4b[2]\n" + ".inst 0x6f84e910 // udot v16.4s, v8.16b, v4.4b[2]\n" "add %[b_ptr0], %[b_ptr0], #0x80\n" - ".word 0x6f85e914 // udot v20.4s, v8.16b, v5.4b[2]\n" - ".word 0x6f86e918 // udot v24.4s, v8.16b, v6.4b[2]\n" - ".word 0x6f87e91c // udot v28.4s, v8.16b, v7.4b[2]\n" - ".word 0x6f84e931 // udot v17.4s, v9.16b, v4.4b[2]\n" - ".word 0x6f85e935 // udot v21.4s, v9.16b, v5.4b[2]\n" - ".word 0x6f86e939 // udot v25.4s, v9.16b, v6.4b[2]\n" - ".word 0x6f87e93d // udot v29.4s, v9.16b, v7.4b[2]\n" - ".word 0x6f84e952 // udot v18.4s, v10.16b, v4.4b[2]\n" - ".word 0x6f85e956 // udot v22.4s, v10.16b, v5.4b[2]\n" - ".word 0x6f86e95a // udot v26.4s, v10.16b, v6.4b[2]\n" - ".word 0x6f87e95e // udot v30.4s, v10.16b, v7.4b[2]\n" - ".word 0x6f84e973 // udot v19.4s, v11.16b, v4.4b[2]\n" - ".word 0x6f85e977 // udot v23.4s, v11.16b, v5.4b[2]\n" - ".word 0x6f86e97b // udot v27.4s, v11.16b, v6.4b[2]\n" - ".word 0x6f87e97f // udot v31.4s, v11.16b, v7.4b[2]\n" - ".word 0x6fa4e990 // udot v16.4s, v12.16b, v4.4b[3]\n" - ".word 0x6fa5e994 // udot v20.4s, v12.16b, v5.4b[3]\n" - ".word 0x6fa6e998 // udot v24.4s, v12.16b, v6.4b[3]\n" - ".word 0x6fa7e99c // udot v28.4s, v12.16b, v7.4b[3]\n" - ".word 0x6fa4e9b1 // udot v17.4s, v13.16b, v4.4b[3]\n" - ".word 0x6fa5e9b5 // udot v21.4s, v13.16b, v5.4b[3]\n" - ".word 0x6fa6e9b9 // udot v25.4s, v13.16b, v6.4b[3]\n" - ".word 0x6fa7e9bd // udot v29.4s, v13.16b, v7.4b[3]\n" - ".word 0x6fa4e9d2 // udot v18.4s, v14.16b, v4.4b[3]\n" - ".word 0x6fa5e9d6 // udot v22.4s, v14.16b, v5.4b[3]\n" - ".word 0x6fa6e9da // udot v26.4s, v14.16b, v6.4b[3]\n" - ".word 0x6fa7e9de // udot v30.4s, v14.16b, v7.4b[3]\n" - ".word 0x6fa4e9f3 // udot v19.4s, v15.16b, v4.4b[3]\n" - ".word 0x6fa5e9f7 // udot v23.4s, v15.16b, v5.4b[3]\n" - ".word 0x6fa6e9fb // udot v27.4s, v15.16b, v6.4b[3]\n" - ".word 0x6fa7e9ff // udot v31.4s, v15.16b, v7.4b[3]\n" + ".inst 0x6f85e914 // udot v20.4s, v8.16b, v5.4b[2]\n" + ".inst 0x6f86e918 // udot v24.4s, v8.16b, v6.4b[2]\n" + ".inst 0x6f87e91c // udot v28.4s, v8.16b, v7.4b[2]\n" + ".inst 0x6f84e931 // udot v17.4s, v9.16b, v4.4b[2]\n" + ".inst 0x6f85e935 // udot v21.4s, v9.16b, v5.4b[2]\n" + ".inst 0x6f86e939 // udot v25.4s, v9.16b, v6.4b[2]\n" + ".inst 0x6f87e93d // udot v29.4s, v9.16b, v7.4b[2]\n" + ".inst 0x6f84e952 // udot v18.4s, v10.16b, v4.4b[2]\n" + ".inst 0x6f85e956 // udot v22.4s, v10.16b, v5.4b[2]\n" + ".inst 0x6f86e95a // udot v26.4s, v10.16b, v6.4b[2]\n" + ".inst 0x6f87e95e // udot v30.4s, v10.16b, v7.4b[2]\n" + ".inst 0x6f84e973 // udot v19.4s, v11.16b, v4.4b[2]\n" + ".inst 0x6f85e977 // udot v23.4s, v11.16b, v5.4b[2]\n" + ".inst 0x6f86e97b // udot v27.4s, v11.16b, v6.4b[2]\n" + ".inst 0x6f87e97f // udot v31.4s, v11.16b, v7.4b[2]\n" + ".inst 0x6fa4e990 // udot v16.4s, v12.16b, v4.4b[3]\n" + ".inst 0x6fa5e994 // udot v20.4s, v12.16b, v5.4b[3]\n" + ".inst 0x6fa6e998 // udot v24.4s, v12.16b, v6.4b[3]\n" + ".inst 0x6fa7e99c // udot v28.4s, v12.16b, v7.4b[3]\n" + ".inst 0x6fa4e9b1 // udot v17.4s, v13.16b, v4.4b[3]\n" + ".inst 0x6fa5e9b5 // udot v21.4s, v13.16b, v5.4b[3]\n" + ".inst 0x6fa6e9b9 // udot v25.4s, v13.16b, v6.4b[3]\n" + ".inst 0x6fa7e9bd // udot v29.4s, v13.16b, v7.4b[3]\n" + ".inst 0x6fa4e9d2 // udot v18.4s, v14.16b, v4.4b[3]\n" + ".inst 0x6fa5e9d6 // udot v22.4s, v14.16b, v5.4b[3]\n" + ".inst 0x6fa6e9da // udot v26.4s, v14.16b, v6.4b[3]\n" + ".inst 0x6fa7e9de // udot v30.4s, v14.16b, v7.4b[3]\n" + ".inst 0x6fa4e9f3 // udot v19.4s, v15.16b, v4.4b[3]\n" + ".inst 0x6fa5e9f7 // udot v23.4s, v15.16b, v5.4b[3]\n" + ".inst 0x6fa6e9fb // udot v27.4s, v15.16b, v6.4b[3]\n" + ".inst 0x6fa7e9ff // udot v31.4s, v15.16b, v7.4b[3]\n" "b 5f\n" "4:\n" - ".word 0x6f80e110 // udot v16.4s, v8.16b, v0.4b[0]\n" - ".word 0x6f81e114 // udot v20.4s, v8.16b, v1.4b[0]\n" - ".word 0x6f82e118 // udot v24.4s, v8.16b, v2.4b[0]\n" - ".word 0x6f83e11c // udot v28.4s, v8.16b, v3.4b[0]\n" + ".inst 0x6f80e110 // udot v16.4s, v8.16b, v0.4b[0]\n" + ".inst 0x6f81e114 // udot v20.4s, v8.16b, v1.4b[0]\n" + ".inst 0x6f82e118 // udot v24.4s, v8.16b, v2.4b[0]\n" + ".inst 0x6f83e11c // udot v28.4s, v8.16b, v3.4b[0]\n" "ldr q8, [%[b_ptr0]]\n" - ".word 0x6f80e131 // udot v17.4s, v9.16b, v0.4b[0]\n" - ".word 0x6f81e135 // udot v21.4s, v9.16b, v1.4b[0]\n" - ".word 0x6f82e139 // udot v25.4s, v9.16b, v2.4b[0]\n" - ".word 0x6f83e13d // udot v29.4s, v9.16b, v3.4b[0]\n" + ".inst 0x6f80e131 // udot v17.4s, v9.16b, v0.4b[0]\n" + ".inst 0x6f81e135 // udot v21.4s, v9.16b, v1.4b[0]\n" + ".inst 0x6f82e139 // udot v25.4s, v9.16b, v2.4b[0]\n" + ".inst 0x6f83e13d // udot v29.4s, v9.16b, v3.4b[0]\n" "ldr q9, [%[b_ptr0], #0x10]\n" - ".word 0x6f80e152 // udot v18.4s, v10.16b, v0.4b[0]\n" - ".word 0x6f81e156 // udot v22.4s, v10.16b, v1.4b[0]\n" - ".word 0x6f82e15a // udot v26.4s, v10.16b, v2.4b[0]\n" - ".word 0x6f83e15e // udot v30.4s, v10.16b, v3.4b[0]\n" + ".inst 0x6f80e152 // udot v18.4s, v10.16b, v0.4b[0]\n" + ".inst 0x6f81e156 // udot v22.4s, v10.16b, v1.4b[0]\n" + ".inst 0x6f82e15a // udot v26.4s, v10.16b, v2.4b[0]\n" + ".inst 0x6f83e15e // udot v30.4s, v10.16b, v3.4b[0]\n" "ldr q10, [%[b_ptr0], #0x20]\n" - ".word 0x6f80e173 // udot v19.4s, v11.16b, v0.4b[0]\n" - ".word 0x6f81e177 // udot v23.4s, v11.16b, v1.4b[0]\n" - ".word 0x6f82e17b // udot v27.4s, v11.16b, v2.4b[0]\n" - ".word 0x6f83e17f // udot v31.4s, v11.16b, v3.4b[0]\n" + ".inst 0x6f80e173 // udot v19.4s, v11.16b, v0.4b[0]\n" + ".inst 0x6f81e177 // udot v23.4s, v11.16b, v1.4b[0]\n" + ".inst 0x6f82e17b // udot v27.4s, v11.16b, v2.4b[0]\n" + ".inst 0x6f83e17f // udot v31.4s, v11.16b, v3.4b[0]\n" "ldr q11, [%[b_ptr0], #0x30]\n" - ".word 0x6fa0e190 // udot v16.4s, v12.16b, v0.4b[1]\n" - ".word 0x6fa1e194 // udot v20.4s, v12.16b, v1.4b[1]\n" - ".word 0x6fa2e198 // udot v24.4s, v12.16b, v2.4b[1]\n" - ".word 0x6fa3e19c // udot v28.4s, v12.16b, v3.4b[1]\n" + ".inst 0x6fa0e190 // udot v16.4s, v12.16b, v0.4b[1]\n" + ".inst 0x6fa1e194 // udot v20.4s, v12.16b, v1.4b[1]\n" + ".inst 0x6fa2e198 // udot v24.4s, v12.16b, v2.4b[1]\n" + ".inst 0x6fa3e19c // udot v28.4s, v12.16b, v3.4b[1]\n" "ldr q12, [%[b_ptr0], #0x40]\n" - ".word 0x6fa0e1b1 // udot v17.4s, v13.16b, v0.4b[1]\n" - ".word 0x6fa1e1b5 // udot v21.4s, v13.16b, v1.4b[1]\n" - ".word 0x6fa2e1b9 // udot v25.4s, v13.16b, v2.4b[1]\n" - ".word 0x6fa3e1bd // udot v29.4s, v13.16b, v3.4b[1]\n" + ".inst 0x6fa0e1b1 // udot v17.4s, v13.16b, v0.4b[1]\n" + ".inst 0x6fa1e1b5 // udot v21.4s, v13.16b, v1.4b[1]\n" + ".inst 0x6fa2e1b9 // udot v25.4s, v13.16b, v2.4b[1]\n" + ".inst 0x6fa3e1bd // udot v29.4s, v13.16b, v3.4b[1]\n" "ldr q13, [%[b_ptr0], #0x50]\n" - ".word 0x6fa0e1d2 // udot v18.4s, v14.16b, v0.4b[1]\n" - ".word 0x6fa1e1d6 // udot v22.4s, v14.16b, v1.4b[1]\n" - ".word 0x6fa2e1da // udot v26.4s, v14.16b, v2.4b[1]\n" - ".word 0x6fa3e1de // udot v30.4s, v14.16b, v3.4b[1]\n" + ".inst 0x6fa0e1d2 // udot v18.4s, v14.16b, v0.4b[1]\n" + ".inst 0x6fa1e1d6 // udot v22.4s, v14.16b, v1.4b[1]\n" + ".inst 0x6fa2e1da // udot v26.4s, v14.16b, v2.4b[1]\n" + ".inst 0x6fa3e1de // udot v30.4s, v14.16b, v3.4b[1]\n" "ldr q14, [%[b_ptr0], #0x60]\n" - ".word 0x6fa0e1f3 // udot v19.4s, v15.16b, v0.4b[1]\n" - ".word 0x6fa1e1f7 // udot v23.4s, v15.16b, v1.4b[1]\n" - ".word 0x6fa2e1fb // udot v27.4s, v15.16b, v2.4b[1]\n" - ".word 0x6fa3e1ff // udot v31.4s, v15.16b, v3.4b[1]\n" + ".inst 0x6fa0e1f3 // udot v19.4s, v15.16b, v0.4b[1]\n" + ".inst 0x6fa1e1f7 // udot v23.4s, v15.16b, v1.4b[1]\n" + ".inst 0x6fa2e1fb // udot v27.4s, v15.16b, v2.4b[1]\n" + ".inst 0x6fa3e1ff // udot v31.4s, v15.16b, v3.4b[1]\n" "ldr q15, [%[b_ptr0], #0x70]\n" - ".word 0x6f80e910 // udot v16.4s, v8.16b, v0.4b[2]\n" + ".inst 0x6f80e910 // udot v16.4s, v8.16b, v0.4b[2]\n" "add %[b_ptr0], %[b_ptr0], #0x80\n" - ".word 0x6f81e914 // udot v20.4s, v8.16b, v1.4b[2]\n" - ".word 0x6f82e918 // udot v24.4s, v8.16b, v2.4b[2]\n" - ".word 0x6f83e91c // udot v28.4s, v8.16b, v3.4b[2]\n" - ".word 0x6f80e931 // udot v17.4s, v9.16b, v0.4b[2]\n" - ".word 0x6f81e935 // udot v21.4s, v9.16b, v1.4b[2]\n" - ".word 0x6f82e939 // udot v25.4s, v9.16b, v2.4b[2]\n" - ".word 0x6f83e93d // udot v29.4s, v9.16b, v3.4b[2]\n" - ".word 0x6f80e952 // udot v18.4s, v10.16b, v0.4b[2]\n" - ".word 0x6f81e956 // udot v22.4s, v10.16b, v1.4b[2]\n" - ".word 0x6f82e95a // udot v26.4s, v10.16b, v2.4b[2]\n" - ".word 0x6f83e95e // udot v30.4s, v10.16b, v3.4b[2]\n" - ".word 0x6f80e973 // udot v19.4s, v11.16b, v0.4b[2]\n" - ".word 0x6f81e977 // udot v23.4s, v11.16b, v1.4b[2]\n" - ".word 0x6f82e97b // udot v27.4s, v11.16b, v2.4b[2]\n" - ".word 0x6f83e97f // udot v31.4s, v11.16b, v3.4b[2]\n" - ".word 0x6fa0e990 // udot v16.4s, v12.16b, v0.4b[3]\n" - ".word 0x6fa1e994 // udot v20.4s, v12.16b, v1.4b[3]\n" - ".word 0x6fa2e998 // udot v24.4s, v12.16b, v2.4b[3]\n" - ".word 0x6fa3e99c // udot v28.4s, v12.16b, v3.4b[3]\n" - ".word 0x6fa0e9b1 // udot v17.4s, v13.16b, v0.4b[3]\n" - ".word 0x6fa1e9b5 // udot v21.4s, v13.16b, v1.4b[3]\n" - ".word 0x6fa2e9b9 // udot v25.4s, v13.16b, v2.4b[3]\n" - ".word 0x6fa3e9bd // udot v29.4s, v13.16b, v3.4b[3]\n" - ".word 0x6fa0e9d2 // udot v18.4s, v14.16b, v0.4b[3]\n" - ".word 0x6fa1e9d6 // udot v22.4s, v14.16b, v1.4b[3]\n" - ".word 0x6fa2e9da // udot v26.4s, v14.16b, v2.4b[3]\n" - ".word 0x6fa3e9de // udot v30.4s, v14.16b, v3.4b[3]\n" - ".word 0x6fa0e9f3 // udot v19.4s, v15.16b, v0.4b[3]\n" - ".word 0x6fa1e9f7 // udot v23.4s, v15.16b, v1.4b[3]\n" - ".word 0x6fa2e9fb // udot v27.4s, v15.16b, v2.4b[3]\n" - ".word 0x6fa3e9ff // udot v31.4s, v15.16b, v3.4b[3]\n" + ".inst 0x6f81e914 // udot v20.4s, v8.16b, v1.4b[2]\n" + ".inst 0x6f82e918 // udot v24.4s, v8.16b, v2.4b[2]\n" + ".inst 0x6f83e91c // udot v28.4s, v8.16b, v3.4b[2]\n" + ".inst 0x6f80e931 // udot v17.4s, v9.16b, v0.4b[2]\n" + ".inst 0x6f81e935 // udot v21.4s, v9.16b, v1.4b[2]\n" + ".inst 0x6f82e939 // udot v25.4s, v9.16b, v2.4b[2]\n" + ".inst 0x6f83e93d // udot v29.4s, v9.16b, v3.4b[2]\n" + ".inst 0x6f80e952 // udot v18.4s, v10.16b, v0.4b[2]\n" + ".inst 0x6f81e956 // udot v22.4s, v10.16b, v1.4b[2]\n" + ".inst 0x6f82e95a // udot v26.4s, v10.16b, v2.4b[2]\n" + ".inst 0x6f83e95e // udot v30.4s, v10.16b, v3.4b[2]\n" + ".inst 0x6f80e973 // udot v19.4s, v11.16b, v0.4b[2]\n" + ".inst 0x6f81e977 // udot v23.4s, v11.16b, v1.4b[2]\n" + ".inst 0x6f82e97b // udot v27.4s, v11.16b, v2.4b[2]\n" + ".inst 0x6f83e97f // udot v31.4s, v11.16b, v3.4b[2]\n" + ".inst 0x6fa0e990 // udot v16.4s, v12.16b, v0.4b[3]\n" + ".inst 0x6fa1e994 // udot v20.4s, v12.16b, v1.4b[3]\n" + ".inst 0x6fa2e998 // udot v24.4s, v12.16b, v2.4b[3]\n" + ".inst 0x6fa3e99c // udot v28.4s, v12.16b, v3.4b[3]\n" + ".inst 0x6fa0e9b1 // udot v17.4s, v13.16b, v0.4b[3]\n" + ".inst 0x6fa1e9b5 // udot v21.4s, v13.16b, v1.4b[3]\n" + ".inst 0x6fa2e9b9 // udot v25.4s, v13.16b, v2.4b[3]\n" + ".inst 0x6fa3e9bd // udot v29.4s, v13.16b, v3.4b[3]\n" + ".inst 0x6fa0e9d2 // udot v18.4s, v14.16b, v0.4b[3]\n" + ".inst 0x6fa1e9d6 // udot v22.4s, v14.16b, v1.4b[3]\n" + ".inst 0x6fa2e9da // udot v26.4s, v14.16b, v2.4b[3]\n" + ".inst 0x6fa3e9de // udot v30.4s, v14.16b, v3.4b[3]\n" + ".inst 0x6fa0e9f3 // udot v19.4s, v15.16b, v0.4b[3]\n" + ".inst 0x6fa1e9f7 // udot v23.4s, v15.16b, v1.4b[3]\n" + ".inst 0x6fa2e9fb // udot v27.4s, v15.16b, v2.4b[3]\n" + ".inst 0x6fa3e9ff // udot v31.4s, v15.16b, v3.4b[3]\n" "5:\n" "cbz %[blocks], 6f\n" "7:\n" @@ -1734,28 +1692,28 @@ void a64_hybrid_u8u32_dot_16x4(const uint8_t *A, int lda, const uint8_t *B, uint "add %[a_ptr0], %[a_ptr0], #0x4\n" "ldr q11, [%[b_ptr0], #0x30]\n" "add %[b_ptr0], %[b_ptr0], #0x40\n" - ".word 0x6f80e110 // udot v16.4s, v8.16b, v0.4b[0]\n" + ".inst 0x6f80e110 // udot v16.4s, v8.16b, v0.4b[0]\n" "ldr s1, [a_ptr1]\n" - ".word 0x6f80e131 // udot v17.4s, v9.16b, v0.4b[0]\n" + ".inst 0x6f80e131 // udot v17.4s, v9.16b, v0.4b[0]\n" "add a_ptr1, a_ptr1, #0x4\n" - ".word 0x6f80e152 // udot v18.4s, v10.16b, v0.4b[0]\n" + ".inst 0x6f80e152 // udot v18.4s, v10.16b, v0.4b[0]\n" "ldr s2, [a_ptr2]\n" - ".word 0x6f81e114 // udot v20.4s, v8.16b, v1.4b[0]\n" + ".inst 0x6f81e114 // udot v20.4s, v8.16b, v1.4b[0]\n" "add a_ptr2, a_ptr2, #0x4\n" - ".word 0x6f81e135 // udot v21.4s, v9.16b, v1.4b[0]\n" + ".inst 0x6f81e135 // udot v21.4s, v9.16b, v1.4b[0]\n" "ldr s3, [a_ptr3]\n" - ".word 0x6f82e118 // udot v24.4s, v8.16b, v2.4b[0]\n" + ".inst 0x6f82e118 // udot v24.4s, v8.16b, v2.4b[0]\n" "add a_ptr3, a_ptr3, #0x4\n" - ".word 0x6f82e139 // udot v25.4s, v9.16b, v2.4b[0]\n" - ".word 0x6f83e11c // udot v28.4s, v8.16b, v3.4b[0]\n" - ".word 0x6f83e13d // udot v29.4s, v9.16b, v3.4b[0]\n" - ".word 0x6f81e156 // udot v22.4s, v10.16b, v1.4b[0]\n" - ".word 0x6f82e15a // udot v26.4s, v10.16b, v2.4b[0]\n" - ".word 0x6f83e15e // udot v30.4s, v10.16b, v3.4b[0]\n" - ".word 0x6f80e173 // udot v19.4s, v11.16b, v0.4b[0]\n" - ".word 0x6f81e177 // udot v23.4s, v11.16b, v1.4b[0]\n" - ".word 0x6f82e17b // udot v27.4s, v11.16b, v2.4b[0]\n" - ".word 0x6f83e17f // udot v31.4s, v11.16b, v3.4b[0]\n" + ".inst 0x6f82e139 // udot v25.4s, v9.16b, v2.4b[0]\n" + ".inst 0x6f83e11c // udot v28.4s, v8.16b, v3.4b[0]\n" + ".inst 0x6f83e13d // udot v29.4s, v9.16b, v3.4b[0]\n" + ".inst 0x6f81e156 // udot v22.4s, v10.16b, v1.4b[0]\n" + ".inst 0x6f82e15a // udot v26.4s, v10.16b, v2.4b[0]\n" + ".inst 0x6f83e15e // udot v30.4s, v10.16b, v3.4b[0]\n" + ".inst 0x6f80e173 // udot v19.4s, v11.16b, v0.4b[0]\n" + ".inst 0x6f81e177 // udot v23.4s, v11.16b, v1.4b[0]\n" + ".inst 0x6f82e17b // udot v27.4s, v11.16b, v2.4b[0]\n" + ".inst 0x6f83e17f // udot v31.4s, v11.16b, v3.4b[0]\n" "b.ne 7b\n" "6:\n" "cbz %[odds], 8f\n" @@ -1780,22 +1738,22 @@ void a64_hybrid_u8u32_dot_16x4(const uint8_t *A, int lda, const uint8_t *B, uint "ldr q9, [%[b_ptr0], #0x10]\n" "ldr q10, [%[b_ptr0], #0x20]\n" "ldr q11, [%[b_ptr0], #0x30]\n" - ".word 0x6f80e110 // udot v16.4s, v8.16b, v0.4b[0]\n" - ".word 0x6f81e114 // udot v20.4s, v8.16b, v1.4b[0]\n" - ".word 0x6f82e118 // udot v24.4s, v8.16b, v2.4b[0]\n" - ".word 0x6f83e11c // udot v28.4s, v8.16b, v3.4b[0]\n" - ".word 0x6f80e131 // udot v17.4s, v9.16b, v0.4b[0]\n" - ".word 0x6f81e135 // udot v21.4s, v9.16b, v1.4b[0]\n" - ".word 0x6f82e139 // udot v25.4s, v9.16b, v2.4b[0]\n" - ".word 0x6f83e13d // udot v29.4s, v9.16b, v3.4b[0]\n" - ".word 0x6f80e152 // udot v18.4s, v10.16b, v0.4b[0]\n" - ".word 0x6f81e156 // udot v22.4s, v10.16b, v1.4b[0]\n" - ".word 0x6f82e15a // udot v26.4s, v10.16b, v2.4b[0]\n" - ".word 0x6f83e15e // udot v30.4s, v10.16b, v3.4b[0]\n" - ".word 0x6f80e173 // udot v19.4s, v11.16b, v0.4b[0]\n" - ".word 0x6f81e177 // udot v23.4s, v11.16b, v1.4b[0]\n" - ".word 0x6f82e17b // udot v27.4s, v11.16b, v2.4b[0]\n" - ".word 0x6f83e17f // udot v31.4s, v11.16b, v3.4b[0]\n" + ".inst 0x6f80e110 // udot v16.4s, v8.16b, v0.4b[0]\n" + ".inst 0x6f81e114 // udot v20.4s, v8.16b, v1.4b[0]\n" + ".inst 0x6f82e118 // udot v24.4s, v8.16b, v2.4b[0]\n" + ".inst 0x6f83e11c // udot v28.4s, v8.16b, v3.4b[0]\n" + ".inst 0x6f80e131 // udot v17.4s, v9.16b, v0.4b[0]\n" + ".inst 0x6f81e135 // udot v21.4s, v9.16b, v1.4b[0]\n" + ".inst 0x6f82e139 // udot v25.4s, v9.16b, v2.4b[0]\n" + ".inst 0x6f83e13d // udot v29.4s, v9.16b, v3.4b[0]\n" + ".inst 0x6f80e152 // udot v18.4s, v10.16b, v0.4b[0]\n" + ".inst 0x6f81e156 // udot v22.4s, v10.16b, v1.4b[0]\n" + ".inst 0x6f82e15a // udot v26.4s, v10.16b, v2.4b[0]\n" + ".inst 0x6f83e15e // udot v30.4s, v10.16b, v3.4b[0]\n" + ".inst 0x6f80e173 // udot v19.4s, v11.16b, v0.4b[0]\n" + ".inst 0x6f81e177 // udot v23.4s, v11.16b, v1.4b[0]\n" + ".inst 0x6f82e17b // udot v27.4s, v11.16b, v2.4b[0]\n" + ".inst 0x6f83e17f // udot v31.4s, v11.16b, v3.4b[0]\n" "8:\n" "str q16, [%[c_ptr0]]\n" "str q17, [%[c_ptr0], #0x10]\n" @@ -1821,7 +1779,7 @@ void a64_hybrid_u8u32_dot_16x4(const uint8_t *A, int lda, const uint8_t *B, uint ".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) + : [width] "r" (width), [append] "r" (static_cast<uint64_t>(append)), [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; |