Nigel
/
mindspore
forked from mindspore-Ecosystem/mindspore
1
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

optimization for fp16 matmul kernel

This commit is contained in:
lixian 2020-09-29 10:37:51 +08:00
parent 56c3d2219e
commit 869bffe976
3 changed files with 1319 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

1311
mindspore/lite/nnacl/assembly/fp16/MatmulFp16Opt.S Normal file
View File

File diff suppressed because it is too large Load Diff

6
mindspore/lite/nnacl/fp16/matmul_fp16.c
View File

@ -87,7 +87,11 @@ void MatMul16x8(const float16_t *a, const float16_t *b, float16_t *dst, const fl
void MatMulFp16(const float16_t *a, const float16_t *b, float16_t *c, const float16_t *bias, ActType act_type,
int depth, int row, int col, int stride, bool write_nhwc) {
MatmulFp16Neon64(a, b, c, bias, (int)act_type, depth, row, col, stride, write_nhwc);
if (!write_nhwc) {
MatmulFp16Neon64(a, b, c, bias, (int)act_type, depth, row, col, stride, write_nhwc);
} else {
MatmulFp16Neon64Opt(a, b, c, bias, (int)act_type, depth, row, col, stride, 1);
}
return;
}

3
mindspore/lite/nnacl/fp16/matmul_fp16.h
View File

@ -39,6 +39,9 @@ void RowMajor2Col16MajorFp16Opt(float16_t *src_ptr, float16_t *dst_ptr, size_t r
void MatmulFp16Neon64(const float16_t *a, const float16_t *b, float16_t *c, const float16_t *bias, int act_type,
size_t depth, size_t row, size_t col, size_t stride, bool write_nhwc);
void MatmulFp16Neon64Opt(const float16_t *a, const float16_t *b, float16_t *c, const float16_t *bias, int act_type,
size_t depth, size_t row, size_t col, size_t stride, int write_nhwc);
void RowMajor2Col16MajorFp16(void *src, float16_t *dst, int row, int col, bool is_fp32_src);
void RowMajor2Row16MajorFp16(void *src, float16_t *dst, int row, int col, bool is_fp32_src);