p47068915
/
mindspore
forked from mindspore-Ecosystem/mindspore
1
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

!7766 [MSLITE][Develop]fp16 activation support swish 

Merge pull request !7766 from ling/deconv
This commit is contained in:
mindspore-ci-bot 2020-11-02 14:04:55 +08:00 committed by Gitee
parent 29697187e2 cd1596b05a
commit 2124ab1a8c
11 changed files with 99 additions and 146 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

12
mindspore/lite/nnacl/fp16/activation_fp16.c
View File

@ -94,3 +94,15 @@ int HSwishFp16(const float16_t *src, float16_t *dst, int ele_num) {
}
return NNACL_OK;
}
int SwishFp16(const float16_t *src, float16_t *dst, int ele_num) {
int ret = SigmoidFp16(src, dst, ele_num);
if (ret != NNACL_OK) {
return NNACL_ERR;
}
int index = 0;
for (; index < ele_num; index++) {
dst[index] = src[index] * dst[index];
}
return NNACL_OK;
}

1
mindspore/lite/nnacl/fp16/activation_fp16.h
View File

@ -38,6 +38,7 @@ int LReluFp16(const float16_t *src, float16_t *dst, int ele_num, float16_t alpha
int SigmoidFp16(const float16_t *src, float16_t *dst, int ele_num);
int TanhFp16(const float16_t *src, float16_t *dst, int ele_num);
int HSwishFp16(const float16_t *src, float16_t *dst, int ele_num);
int SwishFp16(const float16_t *src, float16_t *dst, int ele_num);
#ifdef __cplusplus
}
#endif

15
mindspore/lite/nnacl/fp16/deconv_winograd_fp16.c
View File

@ -110,7 +110,7 @@ void DeConvWgMergeFp16(const float16_t *src, float16_t *dst, size_t src_stride,
return;
}
void _deConvWinogradFp16(float16_t *tile_in, float16_t *tile_out, float16_t *weight_buf, float16_t *tmp_buf,
void DeConvWgCalWgFp16(float16_t *tile_in, float16_t *tile_out, float16_t *weight_buf, float16_t *tmp_buf,
float16_t *at_buf, float16_t *a_mid_buf, float16_t *trans_a_buf, bool *transfered,
float16_t *bt_buf, float16_t *b_tmp_buf, int unit_size, int w_start, int h_start,
ConvParameter *conv_param, DeConvParam *deconv_param) {
@ -151,7 +151,7 @@ void _deConvWinogradFp16(float16_t *tile_in, float16_t *tile_out, float16_t *wei
return;
}
void _deConvCommonFp16(float16_t *tile_in, float16_t *tile_out, float16_t *weight, float16_t *tmp_buf, int h_start,
void DeConvWgCalCommFp16(float16_t *tile_in, float16_t *tile_out, float16_t *weight, float16_t *tmp_buf, int h_start,
int w_start, int h_size, int w_size, ConvParameter *conv_param, DeConvParam *deconv_param) {
int count = deconv_param->oc_div4_ * w_size * h_size;
int in_stride = DECONV_WINOGRAD_DEFAULT_TILE * deconv_param->ic_up4_;
@ -218,6 +218,7 @@ int PackDeConvWgDataFp16(float16_t *nhwc_weight, DeConvComputeUnit *unit, ConvPa
unit->winograd_.AT_ = malloc(unit->winograd_.i_ * unit->winograd_.o_ * sizeof(float16_t));
if (unit->winograd_.AT_ == NULL) {
free(current_unit_weight);
current_unit_weight = NULL;
return NNACL_NULL_PTR;
}
Float32ToFloat16(matrix_at, unit->winograd_.AT_, unit->winograd_.i_ * unit->winograd_.o_);
@ -227,6 +228,8 @@ int PackDeConvWgDataFp16(float16_t *nhwc_weight, DeConvComputeUnit *unit, ConvPa
if (unit->winograd_.BT_ == NULL) {
free(current_unit_weight);
free(unit->winograd_.AT_);
current_unit_weight = NULL;
unit->winograd_.AT_ = NULL;
return NNACL_NULL_PTR;
}
Float32ToFloat16(matrix_bt, unit->winograd_.BT_, unit->winograd_.o_ * unit->winograd_.o_);
@ -238,6 +241,9 @@ int PackDeConvWgDataFp16(float16_t *nhwc_weight, DeConvComputeUnit *unit, ConvPa
free(current_unit_weight);
free(unit->winograd_.AT_);
free(unit->winograd_.BT_);
current_unit_weight = NULL;
unit->winograd_.AT_ = NULL;
unit->winograd_.BT_ = NULL;
return NNACL_NULL_PTR;
}
@ -268,6 +274,7 @@ int PackDeConvWgDataFp16(float16_t *nhwc_weight, DeConvComputeUnit *unit, ConvPa
}
free(current_unit_weight);
current_unit_weight = NULL;
return NNACL_OK;
}
@ -320,13 +327,13 @@ void DeconvWgFp16(float16_t *nhwc_input_, float16_t *tile_in, float16_t *tile_ou
float16_t *tmp_b = (float16_t *)unit->winograd_.b_buffer_ + task_id * unit->winograd_.kh_ * unit->winograd_.kw_ *
DECONV_WINOGRAD_DEFAULT_TILE *
deconv_param->oc_up4_;
_deConvWinogradFp16(tile_in, tile_out, (float16_t *)unit->weight_, tmp_buf, unit->winograd_.AT_, mid_a, dst_a,
DeConvWgCalWgFp16(tile_in, tile_out, (float16_t *)unit->weight_, tmp_buf, unit->winograd_.AT_, mid_a, dst_a,
transfered, unit->winograd_.BT_, tmp_b, unit->winograd_.kh_, unit->w_start_, unit->h_start_,
conv_param, deconv_param);
} else {
float16_t *tmp_buf = (float16_t *)unit->tmp_buffer_ + task_id * deconv_param->oc_div4_ * unit->w_size_ *
unit->h_size_ * DECONV_WINOGRAD_DEFAULT_TILE * C4NUM;
_deConvCommonFp16(tile_in, tile_out, (float16_t *)unit->weight_, tmp_buf, unit->h_start_, unit->w_start_,
DeConvWgCalCommFp16(tile_in, tile_out, (float16_t *)unit->weight_, tmp_buf, unit->h_start_, unit->w_start_,
unit->h_size_, unit->w_size_, conv_param, deconv_param);
}
}

10
mindspore/lite/nnacl/fp32/deconv_winograd.c
View File

@ -340,7 +340,7 @@ void DeConvWgMerge(const float *src, float *dst, size_t src_stride, size_t dst_s
return;
}
void _deConvWinograd(const float *tile_in, float *tile_out, float *weight_buf, float *tmp_buf, const float *at_buf,
void DeConvWgCalWgFp32(const float *tile_in, float *tile_out, float *weight_buf, float *tmp_buf, float *at_buf,
float *a_mid_buf, float *trans_a_buf, bool *transfered, const float *bt_buf, float *b_tmp_buf,
int unit_size, int w_start, int h_start, ConvParameter *conv_param, DeConvParam *deconv_param) {
int winograd_plane = unit_size * unit_size;
@ -380,7 +380,7 @@ void _deConvWinograd(const float *tile_in, float *tile_out, float *weight_buf, f
return;
}
void _deConvCommon(float *tile_in, float *tile_out, const float *weight, float *tmp_buf, int h_start, int w_start,
void DeConvWgCalCommFp32(float *tile_in, float *tile_out, const float *weight, float *tmp_buf, int h_start, int w_start,
int h_size, int w_size, ConvParameter *conv_param, DeConvParam *deconv_param) {
int count = deconv_param->oc_div4_ * w_size * h_size;
int in_stride = DECONV_WINOGRAD_DEFAULT_TILE * deconv_param->ic_up4_;
@ -461,14 +461,14 @@ void DeconvWg(float *nhwc_input_, float *tile_in, float *tile_out, int start_ind
DECONV_WINOGRAD_DEFAULT_TILE * deconv_param->ic_up4_;
float *tmp_b_buf = (float *)unit->winograd_.b_buffer_ + task_id * unit->winograd_.kh_ * unit->winograd_.kw_ *
deconv_param->oc_up4_ * DECONV_WINOGRAD_DEFAULT_TILE;
_deConvWinograd(tile_in, tile_out, (float *)unit->weight_, tmp_buf, unit->winograd_.AT_, wg_mid_a_buf,
DeConvWgCalWgFp32(tile_in, tile_out, (float *)unit->weight_, tmp_buf, unit->winograd_.AT_, wg_mid_a_buf,
wg_dst_a_buf, transfered, unit->winograd_.BT_, tmp_b_buf, unit->winograd_.kh_, unit->w_start_,
unit->h_start_, conv_param, deconv_param);
} else {
float *tmp_buf = (float *)unit->tmp_buffer_ + task_id * deconv_param->oc_div4_ * unit->w_size_ * unit->h_size_ *
DECONV_WINOGRAD_DEFAULT_TILE * C4NUM;
_deConvCommon(tile_in, tile_out, (float *)unit->weight_, tmp_buf, unit->h_start_, unit->w_start_, unit->h_size_,
unit->w_size_, conv_param, deconv_param);
DeConvWgCalCommFp32(tile_in, tile_out, (float *)unit->weight_, tmp_buf, unit->h_start_, unit->w_start_,
unit->h_size_, unit->w_size_, conv_param, deconv_param);
}
}
return;

8
mindspore/lite/nnacl/int8/common_func_int8.c
View File

@ -41,14 +41,6 @@ void PostConvFuncCommInt8(const int32_t *in, int8_t *out, const int32_t *bias, s
return;
}
void PostFuncInt8C8(const int32_t *in, const int32_t *bias, int8_t *out, size_t oc, size_t plane, int32_t multiplier,
int32_t left_shift, int32_t right_shift, int32_t zp, int32_t mini, int32_t maxi) {
/* ((int32_t)row8x8-major + bias) * multiplier + output_zp => (int8)relu => (int8_t)row-major */
PostConvFuncCommInt8(in, out, bias, oc, plane, oc, UP_ROUND(plane, C8NUM) * C8NUM, multiplier, mini, maxi, left_shift,
right_shift, zp, C8NUM);
return;
}
void PostFuncInt8C4(const int32_t *in, const int32_t *bias, int8_t *out, size_t oc, size_t plane, size_t stride,
int32_t multiplier, int32_t left_shift, int32_t right_shift, int32_t zp, int32_t mini,
int32_t maxi) {

58
mindspore/lite/nnacl/int8/deconv_int8.c
View File

@ -17,52 +17,6 @@
#include "nnacl/int8/deconv_int8.h"
#include "nnacl/int8/matmul_int8.h"
#include "nnacl/int8/common_func_int8.h"
int DeConvPostInt8C8(const int32_t *src, const int32_t *bias, int32_t *tmp, int8_t *out, int output_channel,
ConvParameter *conv_param) {
/* row8x8-major(ih*iw x oc*kh*kw) -> row8-major(oh*ow x oc) */
size_t input_plane = conv_param->input_w_ * conv_param->input_h_;
size_t kernel_plane = conv_param->kernel_w_ * conv_param->kernel_h_;
size_t output_plane = conv_param->output_w_ * conv_param->output_h_;
int oc8 = UP_DIV(output_channel, C8NUM);
int in_plane8 = UP_ROUND(input_plane, 8);
for (int c = 0; c < oc8; c++) {
int32_t *dst_ptr = tmp + c * output_plane * C8NUM;
const int32_t *src_ptr = src + c * in_plane8 * kernel_plane * C8NUM;
memset(dst_ptr, 0, output_plane * C8NUM * sizeof(int32_t));
for (int ih = 0; ih < conv_param->input_h_; ih++) {
for (int iw = 0; iw < conv_param->input_w_; iw++) {
int oh = ih * conv_param->stride_h_ - conv_param->pad_u_;
int ow = iw * conv_param->stride_w_ - conv_param->pad_l_;
int kh_start = MSMAX(0, UP_DIV(-oh, conv_param->dilation_h_));
int kh_end = MSMIN(conv_param->kernel_h_, UP_DIV(conv_param->output_h_ - oh, conv_param->dilation_h_));
int kw_start = MSMAX(0, UP_DIV(-ow, conv_param->dilation_w_));
int kw_end = MSMIN(conv_param->kernel_w_, UP_DIV(conv_param->output_w_ - ow, conv_param->dilation_w_));
for (int kh = kh_start; kh < kh_end; kh++) {
for (int kw = kw_start; kw < kw_end; kw++) {
int src_index = ih * conv_param->input_w_ * C8NUM + iw * C8NUM +
kh * input_plane * conv_param->kernel_w_ * C8NUM + kw * input_plane * C8NUM;
int dst_index = oh * conv_param->output_w_ * C8NUM + ow * C8NUM +
kh * conv_param->dilation_h_ * conv_param->output_w_ * C8NUM +
kw * conv_param->dilation_w_ * C8NUM;
for (int i = 0; i < C8NUM; i++) {
dst_ptr[dst_index + i] += src_ptr[src_index + i];
}
} /*kw*/
} /*kh*/
} /*iw*/
} /*ih*/
} /*oc8*/
PostFuncInt8C8(tmp, bias, out, output_channel, output_plane, conv_param->conv_quant_arg_.quant_multiplier_[0],
conv_param->conv_quant_arg_.left_shift_[0], conv_param->conv_quant_arg_.right_shift_[0],
conv_param->conv_quant_arg_.output_quant_args_[0].zp_, conv_param->conv_quant_arg_.out_act_min_[0],
conv_param->conv_quant_arg_.out_act_max_[0]);
return NNACL_OK;
}
int DeConvPostInt8C4(const int32_t *src, const int32_t *bias, int32_t *tmp, int8_t *out, int output_channel,
ConvParameter *conv_param) {
/* row4x4-major(ih*iw x oc*kh*kw) -> row4-major(oh*ow x oc) */
@ -74,8 +28,8 @@ int DeConvPostInt8C4(const int32_t *src, const int32_t *bias, int32_t *tmp, int8
int src_iw_stride = C4NUM;
int src_ih_stride = conv_param->input_w_ * C4NUM;
int src_kw_stride = input_plane * C4NUM;
int src_kh_stride = input_plane * conv_param->kernel_w_ * C4NUM;
int src_kw_stride = in_plane4 * C4NUM;
int src_kh_stride = in_plane4 * conv_param->kernel_w_ * C4NUM;
int dst_oh_stride = conv_param->output_w_ * C4NUM;
int dst_ow_stride = C4NUM;
int dst_kh_stride = conv_param->dilation_h_ * conv_param->output_w_ * C4NUM;
@ -153,18 +107,18 @@ void DeConvWeightTransInt8(int8_t *src, int8_t *dst, int input_channel, int outp
return;
}
void DeConvPackWeightSum(int8_t *weight, int32_t *weight_sum, int32_t input_zp, int32_t filter_zp, int deep16, int col4,
void DeConvPackWeightSum(int8_t *weight, int32_t *weight_sum, int32_t input_zp, int32_t filter_zp, int deep, int col4,
bool suppport_opt) {
/* optimize normal -> same layout */
int deep16 = UP_ROUND(deep, C16NUM);
for (int c = 0; c < col4; c++) {
int c4div = c / C4NUM, c4mod = c % C4NUM;
int32_t value = 0;
for (int r = 0; r < deep16; r++) {
for (int r = 0; r < deep; r++) {
int r16div = r / C16NUM, r16mod = r % C16NUM;
int src_index = c4div * deep16 * C4NUM + r16div * C4NUM * C16NUM + c4mod * C16NUM + r16mod;
value += weight[src_index];
}
weight_sum[c] = filter_zp * input_zp * deep16 - value * input_zp;
weight_sum[c] = filter_zp * input_zp * deep - value * input_zp;
}
return;
}

14
mindspore/lite/src/runtime/kernel/arm/fp16/activation_fp16.cc
View File

@ -30,10 +30,20 @@ using mindspore::schema::ActivationType_HSWISH;
using mindspore::schema::ActivationType_LEAKY_RELU;
using mindspore::schema::ActivationType_RELU;
using mindspore::schema::ActivationType_RELU6;
using mindspore::schema::ActivationType_SWISH;
using mindspore::schema::PrimitiveType_Activation;
namespace mindspore::kernel {
int ActivationFp16CPUKernel::Init() { return RET_OK; }
int ActivationFp16CPUKernel::Init() {
if (type_ != schema::ActivationType_RELU && type_ != schema::ActivationType_RELU6 &&
type_ != schema::ActivationType_LEAKY_RELU && type_ != schema::ActivationType_SIGMOID &&
type_ != schema::ActivationType_TANH && type_ != schema::ActivationType_HSWISH &&
type_ != schema::ActivationType_SWISH) {
MS_LOG(ERROR) << "Activation fp16 not support type: " << type_;
return RET_ERROR;
}
return RET_OK;
}
int ActivationFp16CPUKernel::ReSize() { return RET_OK; }
@ -85,6 +95,8 @@ int ActivationFp16CPUKernel::DoActivation(int task_id) {
error_code = TanhFp16(fp16_input_ + stride * task_id, fp16_output_ + stride * task_id, count);
} else if (type_ == schema::ActivationType_HSWISH) {
error_code = HSwishFp16(fp16_input_ + stride * task_id, fp16_output_ + stride * task_id, count);
} else if (type_ == schema::ActivationType_SWISH) {
error_code = SwishFp16(fp16_input_ + stride * task_id, fp16_output_ + stride * task_id, count);
} else {
MS_LOG(ERROR) << "Activation fp16 not support type: " << type_;
return RET_ERROR;

2
mindspore/lite/src/runtime/kernel/arm/fp16/deconvolution_winograd_fp16.cc
View File

@ -79,6 +79,7 @@ void DeConvWinogradFp16CPUKernel::FreeResizeBuf() {
}
void DeConvWinogradFp16CPUKernel::FreeDeconvParam() {
if (deconv_param_ != nullptr) {
for (int i = 0; i < deconv_param_->compute_size_; i++) {
DeConvComputeUnit &unit = deconv_param_->compute_units_[i];
@ -104,7 +105,6 @@ void DeConvWinogradFp16CPUKernel::FreeDeconvParam() {
deconv_param_->compute_units_ = nullptr;
}
if (deconv_param_ != nullptr) {
delete (deconv_param_);
deconv_param_ = nullptr;
}

12
mindspore/lite/src/runtime/kernel/arm/fp32/activation.cc
View File

@ -33,7 +33,17 @@ using mindspore::schema::ActivationType_SWISH;
using mindspore::schema::PrimitiveType_Activation;
namespace mindspore::kernel {
int ActivationCPUKernel::Init() { return RET_OK; }
int ActivationCPUKernel::Init() {
if (type_ != schema::ActivationType_RELU && type_ != schema::ActivationType_RELU6 &&
type_ != schema::ActivationType_LEAKY_RELU && type_ != schema::ActivationType_SIGMOID &&
type_ != schema::ActivationType_TANH && type_ != schema::ActivationType_HSWISH &&
type_ != schema::ActivationType_SWISH && type_ != schema::ActivationType_HSIGMOID &&
type_ != schema::ActivationType_HARD_TANH) {
MS_LOG(ERROR) << "Activation fp32 not support type: " << type_;
return RET_ERROR;
}
return RET_OK;
}
int ActivationCPUKernel::ReSize() { return RET_OK; }

6
mindspore/lite/src/runtime/kernel/arm/int8/deconvolution_int8.cc
View File

@ -150,7 +150,7 @@ int DeConvInt8CPUKernel::InitBiasWeight() {
MS_LOG(ERROR) << "deconv int8 malloc weight_ptr_ error!";
return RET_ERROR;
}
memset(weight_ptr_, static_cast<int8_t>(conv_param_->conv_quant_arg_.filter_quant_args_[0].zp_), size);
memset(weight_ptr_, 0, size);
DeConvWeightTransInt8(reinterpret_cast<int8_t *>(in_tensors_[1]->MutableData()), weight_ptr_,
conv_param_->input_channel_, conv_param_->output_channel_,
conv_param_->kernel_h_ * conv_param_->kernel_w_, support_optimize_);
@ -163,8 +163,8 @@ int DeConvInt8CPUKernel::InitBiasWeight() {
}
memset(weight_sum_, 0, size * sizeof(int32_t));
DeConvPackWeightSum(weight_ptr_, weight_sum_, conv_param_->conv_quant_arg_.input_quant_args_[0].zp_,
conv_param_->conv_quant_arg_.filter_quant_args_[0].zp_, UP_ROUND(matmul_param_->deep_, C16NUM),
size, support_optimize_);
conv_param_->conv_quant_arg_.filter_quant_args_[0].zp_, matmul_param_->deep_, size,
support_optimize_);
return RET_OK;
}

35
mindspore/lite/test/ut/src/runtime/kernel/arm/int8/deconv_int8_tests.cc
View File

@ -271,41 +271,6 @@ TEST_F(TestDeconvInt8, MatMulOptTest1) {
CompareOutputData(tmp_output, correct_tmp_output, 12 * 3 * 8, 0);
}
TEST_F(TestDeconvInt8, PostAddTest1) {
int32_t in[] = {
-4956, -3923, 868, -8880, -4089, -5179, -4526, -4527, -10464, 99, -5826, -2995, -4519, -4519, -10509, -2505,
-11272, 434, -4522, -4523, -5287, -8936, -878, 373, -4528, -4529, -1960, -6589, 1688, 2287, -8059, 926,
-2506, -6972, -2834, -8281, -8118, -3110, -4526, -4527, -4528, -4529, -4519, -4519, -4519, -4519, -4519, -4519,
-4520, -4521, -4522, -4523, -4524, -4525, -4526, -4527, -4528, -4529, -4519, -4519, -4519, -4519, -4519, -4519,
1578, 2231, -4522, -4523, -4524, -4525, -4526, -4527, -8449, -990, -4519, -4519, -4519, -4519, -4519, -4519,
-4303, -10293, -4522, -4523, -4524, -4525, -4526, -4527, -4528, -4529, -4519, -4519, -4519, -4519, -4519, -4519,
-7025, 924, -4522, -4523, -4524, -4525, -4526, -4527, -4528, -4529, -4519, -4519, -4519, -4519, -4519, -4519,
-4520, -4521, -4522, -4523, -4524, -4525, -4526, -4527, -4528, -4529, -4519, -4519, -4519, -4519, -4519, -4519};
int8_t co[] = {-8, 11, 99, -80, 8, -12, 0, 0, 112, 124, -109, 85, -24, 28, 0, 0, -110,
37, -72, 65, -124, 91, 0, 0, -14, -81, 67, 90, 4, -106, 0, 0, 47, -38,
114, 125, -65, 100, 0, 0, 37, -45, 31, -69, -66, 26, 0, 0, -46, 100};
int32_t bias[] = {1, 2, 3, 4, 5, 6, 7, 8, 9, 10};
int8_t out[50] = {0};
double multiplier = 0.0183649725490196;
int32_t quant_multiplier;
int32_t left_shift;
int32_t right_shift;
QuantizeRoundParameter(multiplier, &quant_multiplier, &left_shift, &right_shift);
int32_t zp = 83;
PostFuncInt8C8(in, bias, out, 10, 5, quant_multiplier, left_shift, right_shift, zp, -128, 127);
CompareOutputData(out, co, 50, 1);
int8_t co_relu[] = {0, 11, 99, 0, 8, 0, 0, 0, 112, 124, 0, 85, 0, 28, 0, 0, 0, 37, 0, 65, 0, 91, 0, 0, 0,
0, 67, 90, 4, 0, 0, 0, 47, 0, 114, 125, 0, 100, 0, 0, 37, 0, 31, 0, 0, 26, 0, 0, 0, 100};
PostFuncInt8C8(in, bias, out, 10, 5, quant_multiplier, left_shift, right_shift, zp, 0, 127);
CompareOutputData(out, co_relu, 50, 1);
int8_t co_relu6[] = {0, 6, 6, 0, 6, 0, 0, 0, 6, 6, 0, 6, 0, 6, 0, 0, 0, 6, 0, 6, 0, 6, 0, 0, 0,
0, 6, 6, 4, 0, 0, 0, 6, 0, 6, 6, 0, 6, 0, 0, 6, 0, 6, 0, 0, 6, 0, 0, 0, 6};
PostFuncInt8C8(in, bias, out, 10, 5, quant_multiplier, left_shift, right_shift, zp, 0, 6);
CompareOutputData(out, co_relu6, 50, 1);
}
int DeConvInt8TestInit1(std::vector<lite::Tensor *> *inputs_, std::vector<lite::Tensor *> *outputs_,
ConvParameter *conv_param, int8_t **correct) {
/* float data from deconv fp32 testcase : DeConvTestInit2 */