diff options
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_bf16fp32_mmla_4VLx4/generic.cpp')
-rw-r--r-- | src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_bf16fp32_mmla_4VLx4/generic.cpp | 17 |
1 files changed, 14 insertions, 3 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_bf16fp32_mmla_4VLx4/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_bf16fp32_mmla_4VLx4/generic.cpp index e3debe508d..4b67d747e2 100644 --- a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_bf16fp32_mmla_4VLx4/generic.cpp +++ b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_bf16fp32_mmla_4VLx4/generic.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018-2019 Arm Limited. + * Copyright (c) 2018-2020 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -62,12 +62,23 @@ void sve_hybrid_bf16fp32_mmla_4VLx4(const bfloat16 *A, int lda, const bfloat16 * break; } - for (int y=0; y<M; y+=8) { + int rows_to_compute; + + for (int y=0; y<M; y+=rows_to_compute) { const bfloat16 * const a_ptr0_base = A + (y * lda); const unsigned long ldab = lda * sizeof(bfloat16); float *c_ptr0 = C + (y * ldc); + rows_to_compute = M-y; + if (rows_to_compute > 8) { + if (rows_to_compute % 8) { + rows_to_compute = 8 - 1; + } else { + rows_to_compute = 8; + } + } + for (int x0=0; x0<N; x0+=(2 * get_vector_length<float>())) { const long width = std::min((unsigned long)N-x0, (2 * get_vector_length<float>())); long loops = loops_count; @@ -79,7 +90,7 @@ void sve_hybrid_bf16fp32_mmla_4VLx4(const bfloat16 *A, int lda, const bfloat16 * const unsigned long ldcb = ldc * sizeof(float); const float *biasptr = bias ? bias+x0 : nullbias; - switch(M-y) { + switch(rows_to_compute) { case 1: __asm __volatile ( "whilelt p6.h, %[temp], %[leftovers]\n" |