m53297601
/
mindspore
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

add sse support for winograd transform

Browse Source
pull/9980/head
lixian 4 years ago
parent 25f3866be7
commit 560d174d30
5 changed files with 1353 additions and 1307 deletions
Show Stats Download Patch File Download Diff File

2
mindspore/lite/nnacl/assembly/avx/MatmulAvx.S
Unescape View File

@ -840,7 +840,7 @@ LoopRow:
vmovups %ymm12, (%rdx)
addq %r12, %rdx
vmovups %ymm14, (%rdx)
cmpq $-8, %rbx
cmpq $8, %rbx
je WriteEnd
movq %rax, %rdx
addq %r13, %rax

47
mindspore/lite/nnacl/fp32/pooling_fp32.c
Unescape View File

@ -17,6 +17,7 @@
#include "nnacl/fp32/pooling_fp32.h"
#include <float.h>
#include "nnacl/errorcode.h"
#include "nnacl/op_base.h"
int AvgPooling(const float *input_ptr, float *output_ptr, const PoolingParameter *pooling_param, int task_id,
float minf, float maxf) {
@ -32,9 +33,9 @@ int AvgPooling(const float *input_ptr, float *output_ptr, const PoolingParameter
int out_tile_count = UP_DIV(out_plane, TILE_NUM);
int window = win_w * win_h;
#ifdef ENABLE_NEON
float32x4_t min_value = vdupq_n_f32(minf);
float32x4_t max_value = vdupq_n_f32(maxf);
#if defined(ENABLE_NEON) || defined(ENALBE_SSE)
MS_FLOAT32X4 min_value = MS_MOVQ_F32(minf);
MS_FLOAT32X4 max_value = MS_MOVQ_F32(maxf);
#endif
for (int batch = 0; batch < pooling_param->output_batch_; batch++) {
@ -61,8 +62,8 @@ int AvgPooling(const float *input_ptr, float *output_ptr, const PoolingParameter
for (int ci = 0; ci < c4; ci++) {
const float *src_c_ptr = src_plane_ptr + ci * C4NUM;
float *dst_c_ptr = dst_plane_ptr + ci * C4NUM;
#ifdef ENABLE_NEON
float32x4_t tmp_avg = vdupq_n_f32(0);
#if defined(ENABLE_NEON) || defined(ENALBE_SSE)
MS_FLOAT32X4 tmp_avg = MS_MOVQ_F32(0);
#else
float tmp_avg1 = 0;
float tmp_avg2 = 0;
@ -73,8 +74,8 @@ int AvgPooling(const float *input_ptr, float *output_ptr, const PoolingParameter
for (int h = real_win_h_start; h < real_win_h_end; h++) {
for (int w = real_win_w_start; w < real_win_w_end; w++) {
const float *src_win_ptr = src_c_ptr + ((in_h_index + h) * in_w + in_w_index + w) * channel;
#ifdef ENABLE_NEON
tmp_avg = vaddq_f32(tmp_avg, vld1q_f32(src_win_ptr));
#if defined(ENABLE_NEON) || defined(ENALBE_SSE)
tmp_avg = MS_ADDQ_F32(tmp_avg, MS_LDQ_F32(src_win_ptr));
#else
tmp_avg1 += src_win_ptr[0];
tmp_avg2 += src_win_ptr[1];
@ -90,11 +91,11 @@ int AvgPooling(const float *input_ptr, float *output_ptr, const PoolingParameter
if (real_count == 0) {
return NNACL_ERR;
}
#ifdef ENABLE_NEON
tmp_avg = tmp_avg / vdupq_n_f32(real_count);
tmp_avg = vmaxq_f32(tmp_avg, min_value);
tmp_avg = vminq_f32(tmp_avg, max_value);
vst1q_f32(dst_c_ptr, tmp_avg);
#if defined(ENABLE_NEON) || defined(ENALBE_SSE)
tmp_avg = tmp_avg / MS_MOVQ_F32(real_count);
tmp_avg = MS_MAXQ_F32(tmp_avg, min_value);
tmp_avg = MS_MINQ_F32(tmp_avg, max_value);
MS_STQ_F32(dst_c_ptr, tmp_avg);
#else
tmp_avg1 /= (float)real_count;
tmp_avg2 /= (float)real_count;
@ -158,9 +159,9 @@ void MaxPooling(const float *input_ptr, float *output_ptr, const PoolingParamete
int out_tile_count = UP_DIV(out_plane, TILE_NUM);
int c4 = channel / C4NUM; /* oc && ic */
#ifdef ENABLE_NEON
float32x4_t min_value = vdupq_n_f32(minf);
float32x4_t max_value = vdupq_n_f32(maxf);
#if defined(ENABLE_NEON) || defined(ENALBE_SSE)
MS_FLOAT32X4 min_value = MS_MOVQ_F32(minf);
MS_FLOAT32X4 max_value = MS_MOVQ_F32(maxf);
#endif
for (int batch = 0; batch < output_batch; batch++) {
@ -187,8 +188,8 @@ void MaxPooling(const float *input_ptr, float *output_ptr, const PoolingParamete
for (int ci = 0; ci < c4; ci++) {
const float *src_c_ptr = src_plane_ptr + ci * C4NUM;
float *dst_c_ptr = dst_plane_ptr + ci * C4NUM;
#ifdef ENABLE_NEON
float32x4_t tmp_max = vdupq_n_f32(-FLT_MAX);
#if defined(ENABLE_NEON) || defined(ENALBE_SSE)
MS_FLOAT32X4 tmp_max = MS_MOVQ_F32(-FLT_MAX);
#else
float tmp_max1 = -FLT_MAX;
float tmp_max2 = -FLT_MAX;
@ -199,8 +200,8 @@ void MaxPooling(const float *input_ptr, float *output_ptr, const PoolingParamete
for (int kh = real_win_h_start; kh < real_win_h_end; kh++) {
for (int kw = real_win_w_start; kw < real_win_w_end; kw++) {
const float *src_win_ptr = src_c_ptr + ((in_h_index + kh) * in_w + in_w_index + kw) * channel;
#ifdef ENABLE_NEON
tmp_max = vmaxq_f32(tmp_max, vld1q_f32(src_win_ptr));
#if defined(ENABLE_NEON) || defined(ENALBE_SSE)
tmp_max = MS_MAXQ_F32(tmp_max, MS_LDQ_F32(src_win_ptr));
#else
tmp_max1 = fmax(tmp_max1, src_win_ptr[0]);
tmp_max2 = fmax(tmp_max2, src_win_ptr[1]);
@ -209,10 +210,10 @@ void MaxPooling(const float *input_ptr, float *output_ptr, const PoolingParamete
#endif
} // win_w loop
} // win_h loop
#ifdef ENABLE_NEON
tmp_max = vmaxq_f32(tmp_max, min_value);
tmp_max = vminq_f32(tmp_max, max_value);
vst1q_f32(dst_c_ptr, tmp_max);
#if defined(ENABLE_NEON) || defined(ENALBE_SSE)
tmp_max = MS_MAXQ_F32(tmp_max, min_value);
tmp_max = MS_MINQ_F32(tmp_max, max_value);
MS_STQ_F32(dst_c_ptr, tmp_max);
#else
tmp_max1 = fmax(tmp_max1, minf);
tmp_max2 = fmax(tmp_max2, minf);

5
mindspore/lite/nnacl/winograd_transform.c
Unescape View File

@ -15,6 +15,7 @@
*/
#include "nnacl/winograd_transform.h"
#include "nnacl/op_base.h"
// fp32 conv winograd
void WinogradInputTransform(const float *input_data, float *trans_input, float *tmp_data, int cal_num,
@ -61,8 +62,8 @@ void WinogradInputTransform(const float *input_data, float *trans_input, float *
int dst_x_offset = dst_y_offset + j * C4NUM;
float *src_addr = (float *)(input_data) + src_x_offset;
float *dst_addr = tmp_data + dst_x_offset;
#ifdef ENABLE_NEON
vst1q_f32(dst_addr, vld1q_f32(src_addr));
#if defined(ENABLE_ARM) || defined(ENABLE_SSE)
MS_STQ_F32(dst_addr, MS_LDQ_F32(src_addr));
#else
for (int k = 0; k < C4NUM; k++) {
dst_addr[k] = src_addr[k];

2252
mindspore/lite/nnacl/winograd_utils.c
View File

File diff suppressed because it is too large Load Diff

354
mindspore/lite/nnacl/winograd_utils.h
View File

File diff suppressed because it is too large Load Diff
Write Preview
Loading…
Cancel
Save