Huawei_Technology
/
mindspore
Not watched
1
0
Fork 0
Code
Releases 13 Wiki evaluate Activity
Issues 0 Pull Requests 0 Datasets Model Cloudbrain HPC
Browse Source

!6878 [MS][LITE][CPU]optimize fp16 winograd 

Merge pull request !6878 from fuzhiye/tmp
tags/v1.1.0
mindspore-ci-bot Gitee 5 years ago
parent
338c27f9b3 f57b55014a
commit
0fdb359775
14 changed files with 2466 additions and 913 deletions
Split View
Diff Options
Show Stats Download Patch File Download Diff File
  1. +8
    -346
      mindspore/lite/nnacl/fp16/conv_fp16.c
  2. +2
    -21
      mindspore/lite/nnacl/fp16/conv_fp16.h
  3. +52
    -25
      mindspore/lite/nnacl/fp16/winograd_transform_fp16.c
  4. +3
    -3
      mindspore/lite/nnacl/fp16/winograd_transform_fp16.h
  5. +2085
    -0
      mindspore/lite/nnacl/fp16/winograd_utils_fp16.c
  6. +274
    -0
      mindspore/lite/nnacl/fp16/winograd_utils_fp16.h
  7. +6
    -7
      mindspore/lite/nnacl/fp32/conv.c
  8. +1
    -8
      mindspore/lite/src/runtime/kernel/arm/fp16/convolution_fp16.cc
  9. +0
    -236
      mindspore/lite/src/runtime/kernel/arm/fp16/convolution_sw_fp16.cc
  10. +0
    -72
      mindspore/lite/src/runtime/kernel/arm/fp16/convolution_sw_fp16.h
  11. +27
    -148
      mindspore/lite/src/runtime/kernel/arm/fp16/convolution_winograd_fp16.cc
  12. +3
    -19
      mindspore/lite/src/runtime/kernel/arm/fp16/convolution_winograd_fp16.h
  13. +4
    -22
      mindspore/lite/src/runtime/kernel/arm/fp32/convolution_winograd.cc
  14. +1
    -6
      mindspore/lite/src/runtime/kernel/arm/fp32/convolution_winograd.h

+ 8
- 346
mindspore/lite/nnacl/fp16/conv_fp16.c View File

@@ -120,208 +120,6 @@ void IndirectGemmFp16_16x8_c8(float16_t *output, float16_t *input, float16_t *we
}
#endif

void SWBorderPixel(float16_t *dst, const float16_t *src, const float16_t *weight, const float16_t *bias, int height,
int width, int in_kh_step, int in_kw_step, int kernel_h, int kernel_w, int ic, bool is_relu,
bool is_relu6) {
int ic8 = ic / C8NUM;
int ic8_res = ic8 % C8NUM;
int ic4 = ic8_res / C4NUM;
for (int c = 0; c < C4NUM; c++) {
dst[c] = 0;
}
const float16_t *weight_oc = weight;
for (int oc = 0; oc < C4NUM; ++oc) {
const float16_t *weight_kh = weight_oc;
const float16_t *src_kh = src;
for (int kh = 0; kh < height; kh++) {
const float16_t *src_kw = src_kh;
const float16_t *weight_kw = weight_kh;
for (int kw = 0; kw < width; kw++) {
const float16_t *src_ic8 = src_kw;
const float16_t *weight_ic8 = weight_kw;

for (int rc = 0; rc < ic8; ++rc) {
for (int c = 0; c < C8NUM; c++) {
dst[oc] += src_ic8[c] * weight_ic8[c];
}
src_ic8 += C8NUM;
weight_ic8 += C8NUM;
} // ic8 loop

const float16_t *src_ic4 = src_ic8;
const float16_t *weight_ic4 = weight_ic8;
for (int rc = 0; rc < ic4; ++rc) {
for (int c = 0; c < C4NUM; c++) {
dst[oc] += src_ic4[c] * weight_ic4[c];
}
src_ic4 += C4NUM;
weight_ic4 += C4NUM;
} // ic4 loop

src_kw += in_kw_step;
weight_kw += ic4 * C4NUM;
} // kernel_w loop
src_kh += in_kh_step;
weight_kh += kernel_w * ic4 * C4NUM;
} // kernel_h loop
dst[oc] += bias[oc];
dst[oc] = (is_relu) ? (MSMAX(0, dst[oc])) : (dst[oc]);
dst[oc] = (is_relu6) ? (MSMIN(6, MSMAX(0, dst[oc]))) : (dst[oc]);
weight_oc += kernel_h * kernel_w * ic4 * C4NUM;
} // oc loop
}

void SWBorderFp16(float16_t *dst, const float16_t *src, const float16_t *weight, const float16_t *bias, int top,
int bottom, int left, int right, const ConvParameter *conv_param, const SlidingWindowParam *sliding) {
bool relu = conv_param->act_type_ == ActType_Relu;
bool relu6 = conv_param->act_type_ == ActType_Relu6;
float16_t *dst_h = dst + top * sliding->out_h_step_;
for (int oh = top; oh < bottom; oh++) {
int ih = oh * conv_param->stride_h_ - conv_param->pad_u_;
int start_kh = MSMAX(0, UP_DIV(-ih, conv_param->dilation_h_));
int end_kh = MSMIN(conv_param->kernel_h_, UP_DIV(conv_param->input_h_ - ih, conv_param->dilation_h_));
const float16_t *src_h = src + ih * sliding->in_h_step_;

float16_t *dst_kernel = dst_h + left * sliding->block_channel_;
for (int ow = left; ow < right; ow++) {
int iw = ow * conv_param->stride_w_ - conv_param->pad_l_;
int start_kw = MSMAX(0, UP_DIV(-iw, conv_param->dilation_w_));
int end_kw = MSMIN(conv_param->kernel_w_, UP_DIV(conv_param->input_w_ - iw, conv_param->dilation_w_));
const float16_t *src_w = src_h + iw * sliding->ic4_channel_;

const float16_t *src_kernel = src_w + start_kh * sliding->in_kh_step_ + start_kw * sliding->in_kw_step_;
const float16_t *weight_kernel = weight + (start_kh * conv_param->kernel_w_ + start_kw) * sliding->ic4_channel_;

SWBorderPixel(dst_kernel, src_kernel, weight_kernel, bias, end_kh - start_kh, end_kw - start_kw,
sliding->in_kh_step_, sliding->in_kw_step_, conv_param->kernel_h_, conv_param->kernel_w_,
sliding->ic4_channel_, relu, relu6);

dst_kernel += sliding->block_channel_;
} // width loop
dst_h += sliding->out_h_step_;
} // height loop
}

void SWCenterFp16(float16_t *dst, const float16_t *src, const float16_t *weight, const float16_t *bias, int height,
int width, int kernel_h, int kernel_w, int out_h_step, int block_channel, int ic, int in_sh_step,
int in_sw_step, int in_kh_step, int in_kw_step, bool is_relu, bool is_relu6) {
int ic8 = ic / C8NUM;
int ic8_res = ic % C8NUM;
int ic4 = ic8_res / C4NUM;
float16_t *dst_h = dst;
const float16_t *src_h = src;
for (int oh = 0; oh < height; oh++) {
float16_t *dst_w = dst_h;
const float16_t *src_w = src_h;
for (int ow = 0; ow < width; ow++) {
const float16_t *weight_oc = weight;
for (int c = 0; c < C4NUM; c++) {
dst_w[c] = 0;
}

for (int oc = 0; oc < C4NUM; oc++) {
const float16_t *weight_kh = weight_oc;
const float16_t *src_kh = src_w;
for (int kh = 0; kh < kernel_h; kh++) {
const float16_t *src_kw = src_kh;
const float16_t *weight_kw = weight_kh;
for (int kw = 0; kw < kernel_w; kw++) {
const float16_t *src_ic8 = src_kw;
const float16_t *weight_ic8 = weight_kw;

for (int rc = 0; rc < ic8; ++rc) {
for (int c = 0; c < C8NUM; c++) {
dst_w[oc] += src_ic8[c] * weight_ic8[c];
}

src_ic8 += C8NUM;
weight_ic8 += C8NUM;
} // ic8 loop

const float16_t *src_ic4 = src_ic8;
const float16_t *weight_ic4 = weight_ic8;
for (int rc = 0; rc < ic4; ++rc) {
for (int c = 0; c < C4NUM; c++) {
dst_w[oc] += src_ic4[c] * weight_ic4[c];
}

src_ic4 += C4NUM;
weight_ic4 += C4NUM;
} // ic4 loop

src_kw += in_kw_step;
weight_kw += ic4 * C4NUM;
} // kernel_w loop
src_kh += in_kh_step;
weight_kh += kernel_w * ic4 * C4NUM;
} // kernel_h loop
// add biad relu

dst_w[oc] += bias[oc];
dst_w[oc] = (is_relu) ? (MSMAX(0, dst_w[oc])) : (dst_w[oc]);
dst_w[oc] = (is_relu6) ? (MSMIN(6, MSMAX(0, dst_w[oc]))) : (dst_w[oc]);
weight_oc += kernel_h * kernel_w * ic4 * C4NUM;
} // oc block

dst_w += block_channel;
src_w += in_sw_step;
} // dst_width loop
dst_h += out_h_step;
src_h += in_sh_step;
} // dst_height loop
}

// fp16 conv sliding window
void ConvSWFp16(const float16_t *input_data, const float16_t *packed_weight, const float16_t *bias_data,
float16_t *tmp_out_block, float16_t *output_data, int task_id, ConvParameter *conv_param,
SlidingWindowParam *slidingWindow_param) {
bool relu = conv_param->act_type_ == ActType_Relu;
bool relu6 = conv_param->act_type_ == ActType_Relu6;
int oc4_res = conv_param->output_channel_ % C4NUM;
const float16_t *src = input_data;
float16_t *dst = NULL;
if (oc4_res == 0) {
dst = output_data;
} else {
dst = tmp_out_block;
}

for (int b = 0; b < conv_param->output_batch_; b++) {
for (int oc = task_id; oc < slidingWindow_param->c_block_; oc += conv_param->thread_num_) {
const float16_t *src_data = src;
float16_t *dst_data = dst + oc * C4NUM;
const float16_t *weight = packed_weight + oc * slidingWindow_param->kernel_step_;
const float16_t *bias = bias_data + oc * C4NUM;
SWBorderFp16(dst_data, src_data, weight, bias, 0, slidingWindow_param->top_, 0, conv_param->output_w_, conv_param,
slidingWindow_param);
SWBorderFp16(dst_data, src_data, weight, bias, slidingWindow_param->bottom_, conv_param->output_h_, 0,
conv_param->output_w_, conv_param, slidingWindow_param);
SWBorderFp16(dst_data, src_data, weight, bias, slidingWindow_param->top_, slidingWindow_param->bottom_, 0,
slidingWindow_param->left_, conv_param, slidingWindow_param);
SWBorderFp16(dst_data, src_data, weight, bias, slidingWindow_param->top_, slidingWindow_param->bottom_,
slidingWindow_param->right_, conv_param->output_w_, conv_param, slidingWindow_param);

if (slidingWindow_param->right_ > slidingWindow_param->left_ &&
slidingWindow_param->bottom_ > slidingWindow_param->top_) {
int in_h_start = slidingWindow_param->top_ * conv_param->stride_h_ - conv_param->pad_u_;
int in_w_start = slidingWindow_param->left_ * conv_param->stride_w_ - conv_param->pad_l_;
const float16_t *in_t =
src_data + in_h_start * slidingWindow_param->in_h_step_ + in_w_start * slidingWindow_param->ic4_channel_;
float16_t *out_t = dst_data + slidingWindow_param->top_ * slidingWindow_param->out_h_step_ +
slidingWindow_param->left_ * slidingWindow_param->block_channel_;
SWCenterFp16(out_t, in_t, weight, bias, slidingWindow_param->bottom_ - slidingWindow_param->top_,
slidingWindow_param->right_ - slidingWindow_param->left_, conv_param->kernel_h_,
conv_param->kernel_w_, slidingWindow_param->out_h_step_, slidingWindow_param->block_channel_,
slidingWindow_param->ic4_channel_, slidingWindow_param->in_sh_step_,
slidingWindow_param->in_sw_step_, slidingWindow_param->in_kh_step_,
slidingWindow_param->in_kw_step_, relu, relu6);
}
} // output C4 loop
src += slidingWindow_param->in_step_;
dst += slidingWindow_param->out_step_;
} // batch loop
}

// fp16 convolution common (im2col+gemm)
void ConvFp16(float16_t *input_data, float16_t *packed_input, float16_t *packed_weight, float16_t *bias_data,
float16_t *tmp_out_block, float16_t *output_data, int task_id, ConvParameter *conv_param) {
@@ -537,8 +335,9 @@ void UnPack3x3Relu6OutputFp16(const float16_t *src, float16_t *dst, int batch, i

// fp16 convolution winograd
void ConvWinogardFp16(float16_t *input_data, float16_t *trans_weight, const float16_t *bias_data,
TmpBufferAddressFp16 *buffer_list, int task_id, ConvParameter *conv_param,
MatricesFp16 *matrices) {
float16_t *output_data, TmpBufferAddressFp16 *buffer_list, int task_id, ConvParameter *conv_param,
InputTransFp16Func in_func, OutputTransFp16Func out_func) {
const int tile_num = 16;
int thread_num = conv_param->thread_num_;
int input_unit = conv_param->input_unit_;
int in_batch = conv_param->input_batch_;
@@ -547,7 +346,6 @@ void ConvWinogardFp16(float16_t *input_data, float16_t *trans_weight, const floa
int out_unit = conv_param->output_unit_;
int out_w_block = UP_DIV(conv_param->output_w_, out_unit);
int out_h_block = UP_DIV(conv_param->output_h_, out_unit);
const int tile_num = 16;
int output_count = out_w_block * out_h_block;
int output_tile_count = UP_DIV(output_count, tile_num);
int out_channel = conv_param->output_channel_;
@@ -557,8 +355,7 @@ void ConvWinogardFp16(float16_t *input_data, float16_t *trans_weight, const floa

float16_t *trans_input = buffer_list[0];
float16_t *gemm_out = buffer_list[1];
float16_t *tmp_out_data = buffer_list[2];
float16_t *tmp_data = buffer_list[3];
float16_t *tmp_data = buffer_list[2];
int trans_input_offset = tile_num * input_unit_square * ic8 * C8NUM;
int gemm_out_offset = tile_num * input_unit_square * oc8 * C8NUM;
int tmp_data_offset = input_unit_square * C8NUM;
@@ -566,156 +363,21 @@ void ConvWinogardFp16(float16_t *input_data, float16_t *trans_weight, const floa
// step 2 : input transform (online)
for (int b = 0; b < in_batch; b++) {
int in_batch_offset = b * ic8 * C8NUM * conv_param->input_h_ * conv_param->input_w_;
int tmp_out_batch_offset = b * out_w_block * out_h_block * out_unit * out_unit * oc8 * C8NUM;
int out_batch_offset = b * out_channel * conv_param->output_h_ * conv_param->output_w_;
for (int thread_id = task_id; thread_id < output_tile_count; thread_id += thread_num) {
int out_tile_index = thread_id * tile_num;
int cal_num = output_count - thread_id * tile_num;
cal_num = cal_num > tile_num ? tile_num : cal_num;
WinogradInputTransformFp16(input_data + in_batch_offset, trans_input + task_id * trans_input_offset,
tmp_data + task_id * tmp_data_offset, cal_num, out_tile_index, out_w_block, conv_param,
matrices[2], matrices[3]);
in_func);
// step 3 : gemm
IndirectGemmFp16_16x8(gemm_out + task_id * gemm_out_offset, trans_input + task_id * trans_input_offset,
trans_weight, NULL, input_unit_square, ic8 * 2, oc8 * C8NUM, output_offset, 1, 1, 0, 0);

// step 4 : output transform
WinogradOutputTransformFp16(gemm_out + task_id * gemm_out_offset, tmp_out_data + tmp_out_batch_offset, bias_data,
cal_num, out_tile_index, out_w_block, conv_param, matrices[0], matrices[1]);
}
}
}

void UnPackWinogradOutputFp16(const float16_t *src, float16_t *dst, int batch, int height, int width, int channel,
int output_unit) {
int out_h_block_num = UP_DIV(height, output_unit);
int out_w_block_num = UP_DIV(width, output_unit);
int c8 = UP_DIV(channel, C8NUM);
int c8_block = C8NUM * out_h_block_num * output_unit * out_w_block_num * output_unit;
for (int b = 0; b < batch; b++) {
int src_batch_offset = b * c8 * c8_block;
int dst_batch_offset = b * height * width * channel;
for (int h = 0; h < height; h++) {
int src_h_offset = src_batch_offset + C8NUM * (h * out_w_block_num * output_unit);
const int dst_h_offset = dst_batch_offset + h * width * channel;
for (int w = 0; w < width; w++) {
int src_w_offset = src_h_offset + w * C8NUM;
int dst_w_offset = dst_h_offset + w * channel;
for (int c = 0; c < c8 - 1; c++) {
int src_c8_offset = src_w_offset + c * c8_block;
int dst_c8_offset = dst_w_offset + c * C8NUM;
#ifdef ENABLE_NEON
vst1q_f16(dst + dst_c8_offset, vld1q_f16(src + src_c8_offset));
#else
for (int i = 0; i < C8NUM; ++i) {
dst[dst_c8_offset + i] = src[src_c8_offset + i];
}
#endif
}
int c_res = channel - (c8 - 1) * C8NUM;
int src_c_res_offset = (c8 - 1) * c8_block;
int dst_c_res_offset = (c8 - 1) * C8NUM;
for (int c = 0; c < c_res; c++) {
int src_c8_res_offset = src_w_offset + src_c_res_offset + c;
int dst_c8_res_offset = dst_w_offset + dst_c_res_offset + c;
dst[dst_c8_res_offset] = src[src_c8_res_offset];
}
}
}
}
}

void UnPackWinogradReluOutputFp16(const float16_t *src, float16_t *dst, int batch, int height, int width, int channel,
int output_unit) {
int out_h_block_num = UP_DIV(height, output_unit);
int out_w_block_num = UP_DIV(width, output_unit);
int c8 = UP_DIV(channel, C8NUM);
int c8_block = C8NUM * out_h_block_num * output_unit * out_w_block_num * output_unit;
for (int b = 0; b < batch; b++) {
int src_batch_offset = b * c8 * c8_block;
int dst_batch_offset = b * height * width * channel;
for (int h = 0; h < height; h++) {
int src_h_offset = src_batch_offset + C8NUM * (h * out_w_block_num * output_unit);
const int dst_h_offset = dst_batch_offset + h * width * channel;
for (int w = 0; w < width; w++) {
int src_w_offset = src_h_offset + w * C8NUM;
int dst_w_offset = dst_h_offset + w * channel;
for (int c = 0; c < c8 - 1; c++) {
int src_c8_offset = src_w_offset + c * c8_block;
int dst_c8_offset = dst_w_offset + c * C8NUM;
#ifdef ENABLE_NEON
float16x8_t input_ptr = vld1q_f16(src + src_c8_offset);
float16x8_t zero = vdupq_n_f16(0);
input_ptr = vmaxq_f16(zero, input_ptr);
vst1q_f16(dst + dst_c8_offset, input_ptr);
#else
for (int i = 0; i < C8NUM; ++i) {
float16_t input_data = src[src_c8_offset + i];
input_data = input_data < 0 ? 0 : input_data;
dst[dst_c8_offset + i] = input_data;
}
#endif
}
int c_res = channel - (c8 - 1) * C8NUM;
int src_c_res_offset = (c8 - 1) * c8_block;
int dst_c_res_offset = (c8 - 1) * C8NUM;
for (int c = 0; c < c_res; c++) {
int src_c8_res_offset = src_w_offset + src_c_res_offset + c;
int dst_c8_res_offset = dst_w_offset + dst_c_res_offset + c;
float16_t input_data = src[src_c8_res_offset];
input_data = input_data < 0 ? 0 : input_data;
dst[dst_c8_res_offset] = input_data;
}
}
}
}
}

void UnPackWinogradRelu6OutputFp16(const float16_t *src, float16_t *dst, int batch, int height, int width, int channel,
int output_unit) {
int out_h_block_num = UP_DIV(height, output_unit);
int out_w_block_num = UP_DIV(width, output_unit);
int c8 = UP_DIV(channel, C8NUM);
int c8_block = C8NUM * out_h_block_num * output_unit * out_w_block_num * output_unit;
for (int b = 0; b < batch; b++) {
int src_batch_offset = b * c8 * c8_block;
int dst_batch_offset = b * height * width * channel;
for (int h = 0; h < height; h++) {
int src_h_offset = src_batch_offset + C8NUM * (h * out_w_block_num * output_unit);
const int dst_h_offset = dst_batch_offset + h * width * channel;
for (int w = 0; w < width; w++) {
int src_w_offset = src_h_offset + w * C8NUM;
int dst_w_offset = dst_h_offset + w * channel;
for (int c = 0; c < c8 - 1; c++) {
int src_c8_offset = src_w_offset + c * c8_block;
int dst_c8_offset = dst_w_offset + c * C8NUM;
#ifdef ENABLE_NEON
float16x8_t input_ptr = vld1q_f16(src + src_c8_offset);
float16x8_t zero = vdupq_n_f16(0);
float16x8_t six = vdupq_n_f16(6);
input_ptr = vmaxq_f16(zero, input_ptr);
input_ptr = vminq_f16(six, input_ptr);
vst1q_f16(dst + dst_c8_offset, input_ptr);
#else
for (int i = 0; i < C8NUM; ++i) {
float16_t input_data = src[src_c8_offset + i];
input_data = input_data < 0 ? 0 : input_data;
input_data = input_data > 6 ? 6 : input_data;
dst[dst_c8_offset + i] = input_data;
}
#endif
}
int c_res = channel - (c8 - 1) * C8NUM;
int src_c_res_offset = (c8 - 1) * c8_block;
int dst_c_res_offset = (c8 - 1) * C8NUM;
for (int c = 0; c < c_res; c++) {
int src_c8_res_offset = src_w_offset + src_c_res_offset + c;
int dst_c8_res_offset = dst_w_offset + dst_c_res_offset + c;
float16_t input_data = src[src_c8_res_offset];
input_data = input_data < 0 ? 0 : input_data;
input_data = input_data > 6 ? 6 : input_data;
dst[dst_c8_res_offset] = input_data;
}
}
WinogradOutputTransformFp16(gemm_out + task_id * gemm_out_offset, output_data + out_batch_offset, bias_data,
cal_num, out_tile_index, out_w_block, conv_param, out_func);
}
}
}

+ 2
- 21
mindspore/lite/nnacl/fp16/conv_fp16.h View File

@@ -40,17 +40,6 @@ void IndirectGemmFp16_16x8_c8(float16_t *output, float16_t *input, float16_t *we
#ifdef __cplusplus
extern "C" {
#endif
void SWBorderFp16(float16_t *dst, const float16_t *src, const float16_t *weight, const float16_t *bias, int top,
int bottom, int left, int right, const ConvParameter *conv_param, const SlidingWindowParam *sliding);

void SWCenterFp16(float16_t *dst, const float16_t *src, const float16_t *weight, const float16_t *bias, int height,
int width, int kernel_h, int kernel_w, int out_h_step, int block_channel, int ic, int in_sh_step,
int in_sw_step, int in_kh_step, int in_kw_step, bool is_relu, bool is_relu6);

// fp16 sliding window
void ConvSWFp16(const float16_t *input_data, const float16_t *packed_weight, const float16_t *bias_data,
float16_t *tmp_out_block, float16_t *output_data, int task_id, ConvParameter *conv_param,
SlidingWindowParam *slidingWindow_param);

// fp16 convolution common (im2col+gemm)
void ConvFp16(float16_t *input_data, float16_t *packed_input, float16_t *packed_weight, float16_t *bias_data,
@@ -69,17 +58,9 @@ void UnPack3x3Relu6OutputFp16(const float16_t *src, float16_t *dst, int batch, i

// fp16 convolution winograd
void ConvWinogardFp16(float16_t *input_data, float16_t *trans_weight, const float16_t *bias_data,
TmpBufferAddressFp16 *buffer_list, int task_id, ConvParameter *conv_param,
MatricesFp16 *matrices);

void UnPackWinogradOutputFp16(const float16_t *src, float16_t *dst, int batch, int height, int width, int channel,
int output_unit);

void UnPackWinogradReluOutputFp16(const float16_t *src, float16_t *dst, int batch, int height, int width, int channel,
int output_unit);
float16_t *output_data, TmpBufferAddressFp16 *buffer_list, int task_id, ConvParameter *conv_param,
InputTransFp16Func in_func, OutputTransFp16Func out_func);

void UnPackWinogradRelu6OutputFp16(const float16_t *src, float16_t *dst, int batch, int height, int width, int channel,
int output_unit);
#ifdef __cplusplus
}
#endif


+ 52
- 25
mindspore/lite/nnacl/fp16/winograd_transform_fp16.c View File

@@ -569,8 +569,8 @@ void Conv3x3Fp16OutputTransform(const float16_t *gemm_out, float16_t *out_data,

// fp16 common winograd
void WinogradInputTransformFp16(const float16_t *input_data, float16_t *trans_input, float16_t *tmp_data, int cal_num,
int out_tile_index, int out_w_block_num, ConvParameter *conv_param, float16_t *matrix_b,
float16_t *matrix_bt) {
int out_tile_index, int out_w_block_num, ConvParameter *conv_param,
InputTransFp16Func func) {
const int tile_num = 16;
int input_unit = conv_param->input_unit_;
int output_unit = conv_param->output_unit_;
@@ -593,36 +593,56 @@ void WinogradInputTransformFp16(const float16_t *input_data, float16_t *trans_in
int interval_x_e = src_x_e < input_w ? input_unit : (input_w - src_x_s);
int interval_y_e = src_y_e < input_h ? input_unit : (input_h - src_y_s);

int src_plane_offset = ic8 * C8NUM * (src_y_s * input_w + src_x_s);
int src_plane_offset = in_channel * (src_y_s * input_w + src_x_s);
int dst_plane_offset = c * C4NUM;
for (int ic = 0; ic < ic8; ic++) {
// clear tmp buffer
memset(tmp_data, 0, input_unit * input_unit * C8NUM * sizeof(float16_t));

// get real input block with padding
int real_c = in_channel - ic * C8NUM;
real_c = real_c > C8NUM ? C8NUM : real_c;
int src_ic8_offset = src_plane_offset + ic * C8NUM;
for (int interval = interval_y_s; interval < interval_y_e; interval++) {
int src_y_offset = src_ic8_offset + (interval * input_w + interval_x_s) * ic8 * C8NUM;
int dst_y_offset = interval * input_unit * C8NUM + interval_x_s * C8NUM;
for (int j = 0; j < (interval_x_e - interval_x_s); j++) {
int src_x_offset = src_y_offset + j * ic8 * C8NUM;
int dst_x_offset = dst_y_offset + j * C8NUM;
const float16_t *src_addr = input_data + src_x_offset;
float16_t *dst_addr = tmp_data + dst_x_offset;

// get real input block with padding
if (real_c == C8NUM) {
for (int interval = interval_y_s; interval < interval_y_e; interval++) {
int src_y_offset = src_ic8_offset + (interval * input_w + interval_x_s) * in_channel;
int dst_y_offset = interval * input_unit * C8NUM + interval_x_s * C8NUM;
for (int j = 0; j < (interval_x_e - interval_x_s); j++) {
int src_x_offset = src_y_offset + j * in_channel;
int dst_x_offset = dst_y_offset + j * C8NUM;
const float16_t *src_addr = input_data + src_x_offset;
float16_t *dst_addr = tmp_data + dst_x_offset;
#ifdef ENABLE_NEON
vst1q_f16(dst_addr, vld1q_f16(src_addr));
vst1q_f16(dst_addr, vld1q_f16(src_addr));
#else
for (int k = 0; k < C8NUM; k++) {
dst_addr[k] = src_addr[k];
}
for (int k = 0; k < C8NUM; k++) {
dst_addr[k] = src_addr[k];
}
#endif
}
}
} else {
for (int interval = interval_y_s; interval < interval_y_e; interval++) {
int src_y_offset = src_ic8_offset + (interval * input_w + interval_x_s) * in_channel;
int dst_y_offset = interval * input_unit * C8NUM + interval_x_s * C8NUM;
for (int j = 0; j < (interval_x_e - interval_x_s); j++) {
int src_x_offset = src_y_offset + j * in_channel;
int dst_x_offset = dst_y_offset + j * C8NUM;
const float16_t *src_addr = input_data + src_x_offset;
float16_t *dst_addr = tmp_data + dst_x_offset;
for (int k = 0; k < real_c; k++) {
dst_addr[k] = src_addr[k];
}
}
}
}

// input transform
int dst_ic8_offset = dst_plane_offset + ic * tile_num * C8NUM;
size_t dst_step = ic8 * C8NUM * tile_num;
float16_t *trans_input_ptr = trans_input + dst_ic8_offset;
GeneralInputTransformUnitFp16(tmp_data, trans_input_ptr, matrix_b, matrix_bt, C8NUM, dst_step, input_unit);
func(tmp_data, trans_input_ptr, C8NUM, dst_step);
}
out_tile_index++;
} // cal_tile_num loop
@@ -630,12 +650,10 @@ void WinogradInputTransformFp16(const float16_t *input_data, float16_t *trans_in

void WinogradOutputTransformFp16(const float16_t *gemm_out, float16_t *tmp_out_data, const float16_t *bias_data,
int cal_num, int out_tile_index, int output_unit_num, ConvParameter *conv_param,
float16_t *matrix_a, float16_t *matrix_at) {
OutputTransFp16Func func) {
int output_unit = conv_param->output_unit_;
int output_w = conv_param->output_w_;
int output_h = conv_param->output_h_;
int output_w_unit_block = UP_DIV(output_w, output_unit);
int output_h_unit_block = UP_DIV(output_h, output_unit);
int output_channel = conv_param->output_channel_;
int oc8 = UP_DIV(output_channel, C8NUM);
int input_unit = conv_param->input_unit_;
@@ -645,18 +663,27 @@ void WinogradOutputTransformFp16(const float16_t *gemm_out, float16_t *tmp_out_d
for (int i = 0; i < cal_num; i++) {
int dst_x_s = out_tile_index % output_unit_num;
int dst_y_s = out_tile_index / output_unit_num;
int r_w = output_w - dst_x_s * output_unit;
r_w = r_w > output_unit ? output_unit : r_w;
int r_h = output_h - dst_y_s * output_unit;
r_h = r_h > output_unit ? output_unit : r_h;
int tmp_ix = dst_x_s * output_unit;
dst_x_s = tmp_ix > output_w ? output_w : tmp_ix;
int tmp_iy = dst_y_s * output_unit;
dst_y_s = tmp_iy > output_h ? output_h : tmp_iy;

int src_tile_offset = i * oc8 * C8NUM * input_unit * input_unit;
int dst_tile_offset = C8NUM * output_unit * (dst_x_s + dst_y_s * output_w_unit_block * output_unit);
int dst_tile_offset = output_channel * (dst_x_s + dst_y_s * output_w);

for (int j = 0; j < oc8; j++) {
int r_c = output_channel - j * C8NUM;
r_c = r_c > C8NUM ? C8NUM : r_c;
int src_oc8_offset = src_tile_offset + j * input_unit * input_unit * C8NUM;
int dst_oc8_offset =
dst_tile_offset + j * C8NUM * output_h_unit_block * output_w_unit_block * output_unit * output_unit;
int dst_oc8_offset = dst_tile_offset + j * C8NUM;
const float16_t *src_ptr = gemm_out + src_oc8_offset;
const float16_t *bias_ptr = bias_data + j * C8NUM;
float16_t *dst_ptr = tmp_out_data + dst_oc8_offset;
GeneralOutputTransformUnitFp16(src_ptr, dst_ptr, bias_ptr, matrix_a, matrix_at, C8NUM,
output_w_unit_block * output_unit, input_unit, output_unit);
func(src_ptr, dst_ptr, bias_ptr, C8NUM, output_w, output_channel, r_w, r_h, r_c);
}
out_tile_index++;
}


+ 3
- 3
mindspore/lite/nnacl/fp16/winograd_transform_fp16.h View File

@@ -43,12 +43,12 @@ void Conv3x3Fp16OutputTransform(const float16_t *gemm_out, float16_t *out_data,

// fp16 common winograd
void WinogradInputTransformFp16(const float16_t *input_data, float16_t *trans_input, float16_t *tmp_data, int cal_num,
int out_tile_index, int out_w_block_num, ConvParameter *conv_param, float16_t *matrix_b,
float16_t *matrix_bt);
int out_tile_index, int out_w_block_num, ConvParameter *conv_param,
InputTransFp16Func func);

void WinogradOutputTransformFp16(const float16_t *gemm_out, float16_t *tmp_out_data, const float16_t *bias_data,
int cal_num, int out_tile_index, int output_unit_num, ConvParameter *conv_param,
float16_t *matrix_a, float16_t *matrix_at);
OutputTransFp16Func func);
#ifdef __cplusplus
}
#endif


+ 2085
- 0
mindspore/lite/nnacl/fp16/winograd_utils_fp16.c
File diff suppressed because it is too large
View File


+ 274
- 0
mindspore/lite/nnacl/fp16/winograd_utils_fp16.h View File

@@ -26,12 +26,286 @@
#ifdef __cplusplus
extern "C" {
#endif
typedef void (*InputTransFp16Func)(const float16_t *src_data, float16_t *dst_data, int src_step, int dst_step);

typedef void (*OutputTransFp16Func)(const float16_t *src_data, float16_t *dst_data, const float16_t *bias_data,
int src_step, int dst_step, int out_c, int r_w, int r_h, int r_c);

void GeneralInputTransformUnitFp16(const float16_t *src_data, float16_t *dst_data, float16_t *matrix_b,
float16_t *matrix_bt, int src_step, int dst_step, int in_unit);

void GeneralOutputTransformUnitFp16(const float16_t *src_data, float16_t *dst_data, const float16_t *bias_data,
float16_t *matrix_a, float16_t *matrix_at, int src_step, int dst_step, int in_unit,
int out_unit);

#define Load16DataFp16 \
src[0] = vld1q_f16(src_data + 0 * src_step); \
src[1] = vld1q_f16(src_data + 1 * src_step); \
src[2] = vld1q_f16(src_data + 2 * src_step); \
src[3] = vld1q_f16(src_data + 3 * src_step); \
src[4] = vld1q_f16(src_data + 4 * src_step); \
src[5] = vld1q_f16(src_data + 5 * src_step); \
src[6] = vld1q_f16(src_data + 6 * src_step); \
src[7] = vld1q_f16(src_data + 7 * src_step); \
src[8] = vld1q_f16(src_data + 8 * src_step); \
src[9] = vld1q_f16(src_data + 9 * src_step); \
src[10] = vld1q_f16(src_data + 10 * src_step); \
src[11] = vld1q_f16(src_data + 11 * src_step); \
src[12] = vld1q_f16(src_data + 12 * src_step); \
src[13] = vld1q_f16(src_data + 13 * src_step); \
src[14] = vld1q_f16(src_data + 14 * src_step); \
src[15] = vld1q_f16(src_data + 15 * src_step);

#define Load36DataFp16 \
src[0] = vld1q_f16(src_data + 0 * src_step); \
src[1] = vld1q_f16(src_data + 1 * src_step); \
src[2] = vld1q_f16(src_data + 2 * src_step); \
src[3] = vld1q_f16(src_data + 3 * src_step); \
src[4] = vld1q_f16(src_data + 4 * src_step); \
src[5] = vld1q_f16(src_data + 5 * src_step); \
src[6] = vld1q_f16(src_data + 6 * src_step); \
src[7] = vld1q_f16(src_data + 7 * src_step); \
src[8] = vld1q_f16(src_data + 8 * src_step); \
src[9] = vld1q_f16(src_data + 9 * src_step); \
src[10] = vld1q_f16(src_data + 10 * src_step); \
src[11] = vld1q_f16(src_data + 11 * src_step); \
src[12] = vld1q_f16(src_data + 12 * src_step); \
src[13] = vld1q_f16(src_data + 13 * src_step); \
src[14] = vld1q_f16(src_data + 14 * src_step); \
src[15] = vld1q_f16(src_data + 15 * src_step); \
src[16] = vld1q_f16(src_data + 16 * src_step); \
src[17] = vld1q_f16(src_data + 17 * src_step); \
src[18] = vld1q_f16(src_data + 18 * src_step); \
src[19] = vld1q_f16(src_data + 19 * src_step); \
src[20] = vld1q_f16(src_data + 20 * src_step); \
src[21] = vld1q_f16(src_data + 21 * src_step); \
src[22] = vld1q_f16(src_data + 22 * src_step); \
src[23] = vld1q_f16(src_data + 23 * src_step); \
src[24] = vld1q_f16(src_data + 24 * src_step); \
src[25] = vld1q_f16(src_data + 25 * src_step); \
src[26] = vld1q_f16(src_data + 26 * src_step); \
src[27] = vld1q_f16(src_data + 27 * src_step); \
src[28] = vld1q_f16(src_data + 28 * src_step); \
src[29] = vld1q_f16(src_data + 29 * src_step); \
src[30] = vld1q_f16(src_data + 30 * src_step); \
src[31] = vld1q_f16(src_data + 31 * src_step); \
src[32] = vld1q_f16(src_data + 32 * src_step); \
src[33] = vld1q_f16(src_data + 33 * src_step); \
src[34] = vld1q_f16(src_data + 34 * src_step); \
src[35] = vld1q_f16(src_data + 35 * src_step);

#define Load64DataFp16 \
src[0] = vld1q_f16(src_data + 0 * src_step); \
src[1] = vld1q_f16(src_data + 1 * src_step); \
src[2] = vld1q_f16(src_data + 2 * src_step); \
src[3] = vld1q_f16(src_data + 3 * src_step); \
src[4] = vld1q_f16(src_data + 4 * src_step); \
src[5] = vld1q_f16(src_data + 5 * src_step); \
src[6] = vld1q_f16(src_data + 6 * src_step); \
src[7] = vld1q_f16(src_data + 7 * src_step); \
src[8] = vld1q_f16(src_data + 8 * src_step); \
src[9] = vld1q_f16(src_data + 9 * src_step); \
src[10] = vld1q_f16(src_data + 10 * src_step); \
src[11] = vld1q_f16(src_data + 11 * src_step); \
src[12] = vld1q_f16(src_data + 12 * src_step); \
src[13] = vld1q_f16(src_data + 13 * src_step); \
src[14] = vld1q_f16(src_data + 14 * src_step); \
src[15] = vld1q_f16(src_data + 15 * src_step); \
src[16] = vld1q_f16(src_data + 16 * src_step); \
src[17] = vld1q_f16(src_data + 17 * src_step); \
src[18] = vld1q_f16(src_data + 18 * src_step); \
src[19] = vld1q_f16(src_data + 19 * src_step); \
src[20] = vld1q_f16(src_data + 20 * src_step); \
src[21] = vld1q_f16(src_data + 21 * src_step); \
src[22] = vld1q_f16(src_data + 22 * src_step); \
src[23] = vld1q_f16(src_data + 23 * src_step); \
src[24] = vld1q_f16(src_data + 24 * src_step); \
src[25] = vld1q_f16(src_data + 25 * src_step); \
src[26] = vld1q_f16(src_data + 26 * src_step); \
src[27] = vld1q_f16(src_data + 27 * src_step); \
src[28] = vld1q_f16(src_data + 28 * src_step); \
src[29] = vld1q_f16(src_data + 29 * src_step); \
src[30] = vld1q_f16(src_data + 30 * src_step); \
src[31] = vld1q_f16(src_data + 31 * src_step); \
src[32] = vld1q_f16(src_data + 32 * src_step); \
src[33] = vld1q_f16(src_data + 33 * src_step); \
src[34] = vld1q_f16(src_data + 34 * src_step); \
src[35] = vld1q_f16(src_data + 35 * src_step); \
src[36] = vld1q_f16(src_data + 36 * src_step); \
src[37] = vld1q_f16(src_data + 37 * src_step); \
src[38] = vld1q_f16(src_data + 38 * src_step); \
src[39] = vld1q_f16(src_data + 39 * src_step); \
src[40] = vld1q_f16(src_data + 40 * src_step); \
src[41] = vld1q_f16(src_data + 41 * src_step); \
src[42] = vld1q_f16(src_data + 42 * src_step); \
src[43] = vld1q_f16(src_data + 43 * src_step); \
src[44] = vld1q_f16(src_data + 44 * src_step); \
src[45] = vld1q_f16(src_data + 45 * src_step); \
src[46] = vld1q_f16(src_data + 46 * src_step); \
src[47] = vld1q_f16(src_data + 47 * src_step); \
src[48] = vld1q_f16(src_data + 48 * src_step); \
src[49] = vld1q_f16(src_data + 49 * src_step); \
src[50] = vld1q_f16(src_data + 50 * src_step); \
src[51] = vld1q_f16(src_data + 51 * src_step); \
src[52] = vld1q_f16(src_data + 52 * src_step); \
src[53] = vld1q_f16(src_data + 53 * src_step); \
src[54] = vld1q_f16(src_data + 54 * src_step); \
src[55] = vld1q_f16(src_data + 55 * src_step); \
src[56] = vld1q_f16(src_data + 56 * src_step); \
src[57] = vld1q_f16(src_data + 57 * src_step); \
src[58] = vld1q_f16(src_data + 58 * src_step); \
src[59] = vld1q_f16(src_data + 59 * src_step); \
src[60] = vld1q_f16(src_data + 60 * src_step); \
src[61] = vld1q_f16(src_data + 61 * src_step); \
src[62] = vld1q_f16(src_data + 62 * src_step); \
src[63] = vld1q_f16(src_data + 63 * src_step);

InputTransFp16Func GetInputTransFp16Func(int input_unit);

void InputTransform4x4UnitFp16(const float16_t *src_data, float16_t *dst_data, int src_step, int dst_step);

void InputTransform6x6UnitFp16(const float16_t *src_data, float16_t *dst_data, int src_step, int dst_step);

void InputTransform8x8UnitFp16(const float16_t *src_data, float16_t *dst_data, int src_step, int dst_step);

OutputTransFp16Func GetOutputTransFp16Func(int input_unit, int output_unit, ActType act_type);

#define Store4DataFp16 \
vst1q_f16(dst_data, m[0]); \
vst1q_f16(dst_data + out_c, m[1]); \
vst1q_f16(dst_data + dst_step * out_c, m[2]); \
vst1q_f16(dst_data + dst_step * out_c + out_c, m[3]);

#define Store9DataFp16 \
vst1q_f16(dst_data, m[0]); \
vst1q_f16(dst_data + out_c, m[1]); \
vst1q_f16(dst_data + 2 * out_c, m[2]); \
vst1q_f16(dst_data + dst_step * out_c, m[3]); \
vst1q_f16(dst_data + dst_step * out_c + out_c, m[4]); \
vst1q_f16(dst_data + dst_step * out_c + 2 * out_c, m[5]); \
vst1q_f16(dst_data + 2 * dst_step * out_c, m[6]); \
vst1q_f16(dst_data + 2 * dst_step * out_c + out_c, m[7]); \
vst1q_f16(dst_data + 2 * dst_step * out_c + 2 * out_c, m[8]);

#define Store16DataFp16 \
vst1q_f16(dst_data, m[0]); \
vst1q_f16(dst_data + out_c, m[1]); \
vst1q_f16(dst_data + 2 * out_c, m[2]); \
vst1q_f16(dst_data + 3 * out_c, m[3]); \
vst1q_f16(dst_data + dst_step * out_c, m[4]); \
vst1q_f16(dst_data + dst_step * out_c + out_c, m[5]); \
vst1q_f16(dst_data + dst_step * out_c + 2 * out_c, m[6]); \
vst1q_f16(dst_data + dst_step * out_c + 3 * out_c, m[7]); \
vst1q_f16(dst_data + 2 * dst_step * out_c, m[8]); \
vst1q_f16(dst_data + 2 * dst_step * out_c + out_c, m[9]); \
vst1q_f16(dst_data + 2 * dst_step * out_c + 2 * out_c, m[10]); \
vst1q_f16(dst_data + 2 * dst_step * out_c + 3 * out_c, m[11]); \
vst1q_f16(dst_data + 3 * dst_step * out_c, m[12]); \
vst1q_f16(dst_data + 3 * dst_step * out_c + out_c, m[13]); \
vst1q_f16(dst_data + 3 * dst_step * out_c + 2 * out_c, m[14]); \
vst1q_f16(dst_data + 3 * dst_step * out_c + 3 * out_c, m[15]);

#define Store25DataFp16 \
vst1q_f16(dst_data, m[0]); \
vst1q_f16(dst_data + out_c, m[1]); \
vst1q_f16(dst_data + 2 * out_c, m[2]); \
vst1q_f16(dst_data + 3 * out_c, m[3]); \
vst1q_f16(dst_data + 4 * out_c, m[4]); \
vst1q_f16(dst_data + dst_step * out_c, m[5]); \
vst1q_f16(dst_data + dst_step * out_c + out_c, m[6]); \
vst1q_f16(dst_data + dst_step * out_c + 2 * out_c, m[7]); \
vst1q_f16(dst_data + dst_step * out_c + 3 * out_c, m[8]); \
vst1q_f16(dst_data + dst_step * out_c + 4 * out_c, m[9]); \
vst1q_f16(dst_data + 2 * dst_step * out_c, m[10]); \
vst1q_f16(dst_data + 2 * dst_step * out_c + out_c, m[11]); \
vst1q_f16(dst_data + 2 * dst_step * out_c + 2 * out_c, m[12]); \
vst1q_f16(dst_data + 2 * dst_step * out_c + 3 * out_c, m[13]); \
vst1q_f16(dst_data + 2 * dst_step * out_c + 4 * out_c, m[14]); \
vst1q_f16(dst_data + 3 * dst_step * out_c, m[15]); \
vst1q_f16(dst_data + 3 * dst_step * out_c + out_c, m[16]); \
vst1q_f16(dst_data + 3 * dst_step * out_c + 2 * out_c, m[17]); \
vst1q_f16(dst_data + 3 * dst_step * out_c + 3 * out_c, m[18]); \
vst1q_f16(dst_data + 3 * dst_step * out_c + 4 * out_c, m[19]); \
vst1q_f16(dst_data + 4 * dst_step * out_c, m[20]); \
vst1q_f16(dst_data + 4 * dst_step * out_c + out_c, m[21]); \
vst1q_f16(dst_data + 4 * dst_step * out_c + 2 * out_c, m[22]); \
vst1q_f16(dst_data + 4 * dst_step * out_c + 3 * out_c, m[23]); \
vst1q_f16(dst_data + 4 * dst_step * out_c + 4 * out_c, m[24]);

void OutputTransform4x2UnitFp16(const float16_t *src_data, float16_t *dst_data, const float16_t *bias_data,
int src_step, int dst_step, int out_c, int r_w, int r_h, int r_c);
void OutputTransform4x2ReluUnitFp16(const float16_t *src_data, float16_t *dst_data, const float16_t *bias_data,
int src_step, int dst_step, int out_c, int r_w, int r_h, int r_c);
void OutputTransform4x2Relu6UnitFp16(const float16_t *src_data, float16_t *dst_data, const float16_t *bias_data,
int src_step, int dst_step, int out_c, int r_w, int r_h, int r_c);
void OutputTransform4x3UnitFp16(const float16_t *src_data, float16_t *dst_data, const float16_t *bias_data,
int src_step, int dst_step, int out_c, int r_w, int r_h, int r_c);
void OutputTransform4x3ReluUnitFp16(const float16_t *src_data, float16_t *dst_data, const float16_t *bias_data,
int src_step, int dst_step, int out_c, int r_w, int r_h, int r_c);
void OutputTransform4x3Relu6UnitFp16(const float16_t *src_data, float16_t *dst_data, const float16_t *bias_data,
int src_step, int dst_step, int out_c, int r_w, int r_h, int r_c);

void OutputTransform6x2UnitFp16(const float16_t *src_data, float16_t *dst_data, const float16_t *bias_data,
int src_step, int dst_step, int out_c, int r_w, int r_h, int r_c);
void OutputTransform6x2ReluUnitFp16(const float16_t *src_data, float16_t *dst_data, const float16_t *bias_data,
int src_step, int dst_step, int out_c, int r_w, int r_h, int r_c);
void OutputTransform6x2Relu6UnitFp16(const float16_t *src_data, float16_t *dst_data, const float16_t *bias_data,
int src_step, int dst_step, int out_c, int r_w, int r_h, int r_c);
void OutputTransform6x3UnitFp16(const float16_t *src_data, float16_t *dst_data, const float16_t *bias_data,
int src_step, int dst_step, int out_c, int r_w, int r_h, int r_c);
void OutputTransform6x3ReluUnitFp16(const float16_t *src_data, float16_t *dst_data, const float16_t *bias_data,
int src_step, int dst_step, int out_c, int r_w, int r_h, int r_c);
void OutputTransform6x3Relu6UnitFp16(const float16_t *src_data, float16_t *dst_data, const float16_t *bias_data,
int src_step, int dst_step, int out_c, int r_w, int r_h, int r_c);
void OutputTransform6x4UnitFp16(const float16_t *src_data, float16_t *dst_data, const float16_t *bias_data,
int src_step, int dst_step, int out_c, int r_w, int r_h, int r_c);
void OutputTransform6x4ReluUnitFp16(const float16_t *src_data, float16_t *dst_data, const float16_t *bias_data,
int src_step, int dst_step, int out_c, int r_w, int r_h, int r_c);
void OutputTransform6x4Relu6UnitFp16(const float16_t *src_data, float16_t *dst_data, const float16_t *bias_data,
int src_step, int dst_step, int out_c, int r_w, int r_h, int r_c);
void OutputTransform6x5UnitFp16(const float16_t *src_data, float16_t *dst_data, const float16_t *bias_data,
int src_step, int dst_step, int out_c, int r_w, int r_h, int r_c);
void OutputTransform6x5ReluUnitFp16(const float16_t *src_data, float16_t *dst_data, const float16_t *bias_data,
int src_step, int dst_step, int out_c, int r_w, int r_h, int r_c);
void OutputTransform6x5Relu6UnitFp16(const float16_t *src_data, float16_t *dst_data, const float16_t *bias_data,
int src_step, int dst_step, int out_c, int r_w, int r_h, int r_c);

void OutputTransform8x2UnitFp16(const float16_t *src_data, float16_t *dst_data, const float16_t *bias_data,
int src_step, int dst_step, int out_c, int r_w, int r_h, int r_c);
void OutputTransform8x2ReluUnitFp16(const float16_t *src_data, float16_t *dst_data, const float16_t *bias_data,
int src_step, int dst_step, int out_c, int r_w, int r_h, int r_c);
void OutputTransform8x2Relu6UnitFp16(const float16_t *src_data, float16_t *dst_data, const float16_t *bias_data,
int src_step, int dst_step, int out_c, int r_w, int r_h, int r_c);
void OutputTransform8x3UnitFp16(const float16_t *src_data, float16_t *dst_data, const float16_t *bias_data,
int src_step, int dst_step, int out_c, int r_w, int r_h, int r_c);
void OutputTransform8x3ReluUnitFp16(const float16_t *src_data, float16_t *dst_data, const float16_t *bias_data,
int src_step, int dst_step, int out_c, int r_w, int r_h, int r_c);
void OutputTransform8x3Relu6UnitFp16(const float16_t *src_data, float16_t *dst_data, const float16_t *bias_data,
int src_step, int dst_step, int out_c, int r_w, int r_h, int r_c);
void OutputTransform8x4UnitFp16(const float16_t *src_data, float16_t *dst_data, const float16_t *bias_data,
int src_step, int dst_step, int out_c, int r_w, int r_h, int r_c);
void OutputTransform8x4ReluUnitFp16(const float16_t *src_data, float16_t *dst_data, const float16_t *bias_data,
int src_step, int dst_step, int out_c, int r_w, int r_h, int r_c);
void OutputTransform8x4Relu6UnitFp16(const float16_t *src_data, float16_t *dst_data, const float16_t *bias_data,
int src_step, int dst_step, int out_c, int r_w, int r_h, int r_c);
void OutputTransform8x5UnitFp16(const float16_t *src_data, float16_t *dst_data, const float16_t *bias_data,
int src_step, int dst_step, int out_c, int r_w, int r_h, int r_c);
void OutputTransform8x5ReluUnitFp16(const float16_t *src_data, float16_t *dst_data, const float16_t *bias_data,
int src_step, int dst_step, int out_c, int r_w, int r_h, int r_c);
void OutputTransform8x5Relu6UnitFp16(const float16_t *src_data, float16_t *dst_data, const float16_t *bias_data,
int src_step, int dst_step, int out_c, int r_w, int r_h, int r_c);
void OutputTransform8x6UnitFp16(const float16_t *src_data, float16_t *dst_data, const float16_t *bias_data,
int src_step, int dst_step, int out_c, int r_w, int r_h, int r_c);
void OutputTransform8x6ReluUnitFp16(const float16_t *src_data, float16_t *dst_data, const float16_t *bias_data,
int src_step, int dst_step, int out_c, int r_w, int r_h, int r_c);
void OutputTransform8x6Relu6UnitFp16(const float16_t *src_data, float16_t *dst_data, const float16_t *bias_data,
int src_step, int dst_step, int out_c, int r_w, int r_h, int r_c);
void OutputTransform8x7UnitFp16(const float16_t *src_data, float16_t *dst_data, const float16_t *bias_data,
int src_step, int dst_step, int out_c, int r_w, int r_h, int r_c);
void OutputTransform8x7ReluUnitFp16(const float16_t *src_data, float16_t *dst_data, const float16_t *bias_data,
int src_step, int dst_step, int out_c, int r_w, int r_h, int r_c);
void OutputTransform8x7Relu6UnitFp16(const float16_t *src_data, float16_t *dst_data, const float16_t *bias_data,
int src_step, int dst_step, int out_c, int r_w, int r_h, int r_c);
#ifdef __cplusplus
}
#endif


+ 6
- 7
mindspore/lite/nnacl/fp32/conv.c View File

@@ -41,8 +41,7 @@ void ConvFp32(float *input_data, float *packed_input, float *packed_weight, cons
#endif
int output_tile_count = UP_DIV(output_count, cal_num);
int kernel_plane = kernel_h * kernel_w;
int unit_size = kernel_plane * in_channel;
int deep = in_channel * kernel_plane;
int deep = kernel_plane * in_channel;

for (int b = 0; b < in_batch; b++) {
int in_batch_offset = b * in_channel * in_h * in_w;
@@ -50,9 +49,9 @@ void ConvFp32(float *input_data, float *packed_input, float *packed_weight, cons
for (int thread_id = task_id; thread_id < output_tile_count; thread_id += thread_count) {
int start_index = thread_id * cal_num;
int real_cal_num = (output_count - start_index) < cal_num ? (output_count - start_index) : cal_num;
float *gemm_input = packed_input + task_id * unit_size * cal_num;
float *col_major_gemm_input = col_major_input + task_id * unit_size * cal_num;
size_t packed_input_size = unit_size * cal_num * sizeof(float);
float *gemm_input = packed_input + task_id * deep * cal_num;
float *col_major_gemm_input = col_major_input + task_id * deep * cal_num;
size_t packed_input_size = deep * cal_num * sizeof(float);
memset(gemm_input, 0, packed_input_size);
memset(col_major_gemm_input, 0, packed_input_size);
Im2ColPackUnitFp32(input_data + in_batch_offset, conv_param, gemm_input, real_cal_num, start_index);
@@ -95,8 +94,8 @@ void ConvWinogardFp32(float *input_data, float *trans_weight, const float *bias_

float *trans_input = buffer_list[0];
float *gemm_out = buffer_list[1];
float *tmp_data = buffer_list[3];
float *col_buffer = buffer_list[4];
float *tmp_data = buffer_list[2];
float *col_buffer = buffer_list[3];
int trans_input_offset = tile_num * input_unit_square * ic4 * C4NUM;
int gemm_out_offset = tile_num * input_unit_square * oc8 * C8NUM;
int tmp_data_offset = input_unit_square * C4NUM;


+ 1
- 8
mindspore/lite/src/runtime/kernel/arm/fp16/convolution_fp16.cc View File

@@ -16,7 +16,6 @@

#include "src/runtime/kernel/arm/fp16/convolution_fp16.h"
#include <vector>
#include "src/runtime/kernel/arm/fp16/convolution_sw_fp16.h"
#include "src/runtime/kernel/arm/fp16/convolution_winograd_fp16.h"
#include "src/runtime/kernel/arm/fp16/convolution_3x3_fp16.h"
#include "src/runtime/kernel/arm/fp16/convolution_1x1_fp16.h"
@@ -203,19 +202,13 @@ kernel::LiteKernel *CpuConvFp16KernelCreator(const std::vector<lite::Tensor *> &
auto conv_param = reinterpret_cast<ConvParameter *>(opParameter);
int kernel_h = conv_param->kernel_h_;
int kernel_w = conv_param->kernel_w_;
int stride_h = conv_param->stride_h_;
int stride_w = conv_param->stride_w_;
int dilation_h = conv_param->dilation_h_;
int dilation_w = conv_param->dilation_w_;
conv_param->input_h_ = inputs.front()->Height();
conv_param->input_w_ = inputs.front()->Width();
conv_param->output_h_ = outputs.front()->Height();
conv_param->output_w_ = outputs.front()->Width();

kernel::LiteKernel *kernel = nullptr;
if (kernel_h == 3 && kernel_w == 3 && stride_h == 1 && stride_w == 1 && dilation_h == 1 && dilation_w == 1) {
kernel = new (std::nothrow) kernel::ConvolutionFP16CPUKernel(opParameter, inputs, outputs, ctx, primitive);
} else if (kernel_h == 1 && kernel_w == 1) {
if (kernel_h == 1 && kernel_w == 1) {
kernel = new (std::nothrow) kernel::Convolution1x1FP16CPUKernel(opParameter, inputs, outputs, ctx, primitive);
} else {
bool use_winograd = false;


+ 0
- 236
mindspore/lite/src/runtime/kernel/arm/fp16/convolution_sw_fp16.cc View File

@@ -1,236 +0,0 @@
/**
* Copyright 2020 Huawei Technologies Co., Ltd
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*/
#include "src/runtime/kernel/arm/fp16/convolution_sw_fp16.h"
#include <vector>
#include "nnacl/fp16/conv_fp16.h"
#include "nnacl/fp16/cast_fp16.h"
#include "nnacl/fp16/pack_fp16.h"
#include "nnacl/fp32/conv_depthwise.h"
#include "src/runtime/kernel/arm/fp16/layout_transform_fp16.h"
#include "schema/model_generated.h"
#include "src/kernel_registry.h"
#include "include/errorcode.h"
#include "src/runtime/runtime_api.h"

using mindspore::kernel::KERNEL_ARCH::kCPU;
using mindspore::lite::KernelRegistrar;
using mindspore::lite::RET_ERROR;
using mindspore::lite::RET_OK;
using mindspore::schema::PrimitiveType_Conv2D;

namespace mindspore::kernel {
int ConvolutionSWFP16CPUKernel::ProcessFilter() {
int kernel_h = conv_param_->kernel_h_;
int kernel_w = conv_param_->kernel_w_;
int in_channel = conv_param_->input_channel_;
int out_channel = conv_param_->output_channel_;
int ic4 = UP_DIV(in_channel, C4NUM);

auto ret = ConvolutionBaseFP16CPUKernel::GetExecuteFilter();
if (ret != RET_OK) {
MS_LOG(ERROR) << "Get Execute filter failed.";
return ret;
}

for (int oc = 0; oc < out_channel; ++oc) {
int src_oc_offset = oc * kernel_h * kernel_w * in_channel;
int dst_oc_offset = oc * kernel_h * kernel_w * ic4 * C4NUM;
for (int i = 0; i < kernel_h * kernel_w; ++i) {
const float16_t *src = execute_weight_ + src_oc_offset + i * in_channel;
float16_t *dst = packed_weight_ + dst_oc_offset + i * ic4 * C4NUM;
memcpy(dst, src, in_channel * sizeof(float16_t));
}
}

return RET_OK;
}

int ConvolutionSWFP16CPUKernel::InitWeightBias() {
auto filter_tensor = in_tensors_.at(kWeightIndex);
int kernel_h = filter_tensor->Height();
int kernel_w = filter_tensor->Width();
int in_channel = filter_tensor->Channel();
int out_channel = filter_tensor->Batch();
conv_param_->input_channel_ = in_channel;
conv_param_->output_channel_ = out_channel;
int oc4 = UP_DIV(out_channel, C4NUM);
int ic4 = UP_DIV(in_channel, C4NUM);
int kernel_plane = kernel_h * kernel_w;
int pack_weight_size = oc4 * ic4 * C4NUM * C4NUM * kernel_plane;

packed_weight_ = reinterpret_cast<float16_t *>(malloc(pack_weight_size * sizeof(float16_t)));
if (packed_weight_ == nullptr) {
MS_LOG(ERROR) << "malloc packed_weight_ failed.";
return RET_ERROR;
}
memset(packed_weight_, 0, pack_weight_size * sizeof(float16_t));
auto ret = ProcessFilter();
if (ret != RET_OK) {
MS_LOG(ERROR) << "Process filter failed.";
return ret;
}

bias_data_ = malloc(oc4 * C4NUM * sizeof(float16_t));
if (bias_data_ == nullptr) {
MS_LOG(ERROR) << "malloc bias_data_ failed.";
return RET_ERROR;
}
memset(bias_data_, 0, oc4 * C4NUM * sizeof(float16_t));
auto fp16_bias_data = reinterpret_cast<float16_t *>(bias_data_);
if (in_tensors_.size() == kInputSize2) {
auto ori_bias = reinterpret_cast<float *>(in_tensors_.at(kBiasIndex)->MutableData());
for (int i = 0; i < out_channel; ++i) {
fp16_bias_data[i] = (float16_t)ori_bias[i];
}
} else {
MS_ASSERT(in_tensor_.size() == kInputSize1);
}
return RET_OK;
}

int ConvolutionSWFP16CPUKernel::InitTmpBuffer() {
int out_channel = conv_param_->output_channel_;
int oc4 = UP_DIV(out_channel, C4NUM);

int ic4 = UP_DIV(conv_param_->input_channel_, C4NUM);
size_t nhwc4_input_size =
ic4 * C4NUM * conv_param_->input_batch_ * conv_param_->input_h_ * conv_param_->input_w_ * sizeof(float16_t);
nhwc4_input_ = ctx_->allocator->Malloc(nhwc4_input_size);
if (nhwc4_input_ == nullptr) {
MS_LOG(ERROR) << "malloc nhwc4_input_ failed.";
return RET_ERROR;
}

tmp_output_block_ = reinterpret_cast<float16_t *>(ctx_->allocator->Malloc(
conv_param_->output_batch_ * conv_param_->output_h_ * conv_param_->output_w_ * oc4 * C4NUM * sizeof(float16_t)));
if (tmp_output_block_ == nullptr) {
MS_LOG(ERROR) << "malloc tmp_output_block_ failed.";
return RET_ERROR;
}
return RET_OK;
}

void ConvolutionSWFP16CPUKernel::ConfigInputOutput() {
auto input_tensor = in_tensors_.at(kInputIndex);
auto input_format = input_tensor->GetFormat();
schema::Format execute_format = schema::Format::Format_NHWC4;
convert_func_ = LayoutTransformFp16(input_format, execute_format);
if (convert_func_ == nullptr) {
MS_LOG(ERROR) << "layout convert func is nullptr.";
return;
}
}

int ConvolutionSWFP16CPUKernel::Init() {
auto ret = InitWeightBias();
if (ret != RET_OK) {
MS_LOG(ERROR) << "Init weight bias failed.";
return RET_ERROR;
}
if (!InferShapeDone()) {
return RET_OK;
}
ConfigInputOutput();
return ReSize();
}

int ConvolutionSWFP16CPUKernel::ReSize() {
auto ret = ConvolutionBaseCPUKernel::CheckResizeValid();
if (ret != RET_OK) {
MS_LOG(ERROR) << "Resize is invalid.";
return ret;
}

if (slidingWindow_param_ != nullptr) {
delete slidingWindow_param_;
slidingWindow_param_ = nullptr;
}

ret = ConvolutionBaseCPUKernel::Init();
if (ret != RET_OK) {
MS_LOG(ERROR) << "ConvolutionBase init fail!ret: " << ret;
return ret;
}

// init sliding window param
slidingWindow_param_ = new (std::nothrow) SlidingWindowParam;
if (slidingWindow_param_ == nullptr) {
MS_LOG(ERROR) << "new SlidingWindowParam fail!";
return RET_ERROR;
}
InitSlidingParamConv(slidingWindow_param_, conv_param_, C4NUM);
return RET_OK;
}

int ConvolutionSWFP16CPUKernel::RunImpl(int task_id) {
ConvSWFp16(reinterpret_cast<float16_t *>(nhwc4_input_), packed_weight_, reinterpret_cast<float16_t *>(bias_data_),
tmp_output_block_, execute_output_, task_id, conv_param_, slidingWindow_param_);
return RET_OK;
}

static int ConvolutionSWFp16Impl(void *cdata, int task_id) {
auto conv = reinterpret_cast<ConvolutionSWFP16CPUKernel *>(cdata);
auto error_code = conv->RunImpl(task_id);
if (error_code != RET_OK) {
MS_LOG(ERROR) << "ConvolutionFp16 Run error task_id[" << task_id << "] error_code[" << error_code << "]";
return RET_ERROR;
}
return RET_OK;
}

int ConvolutionSWFP16CPUKernel::Run() {
auto ret = Prepare();
if (ret != RET_OK) {
MS_LOG(ERROR) << "Prepare failed.";
return RET_ERROR;
}
ret = ConvolutionBaseFP16CPUKernel::GetExecuteTensor();
if (ret != RET_OK) {
MS_LOG(ERROR) << "Get Execute tensor failed.";
return ret;
}
ret = InitTmpBuffer();
if (ret != RET_OK) {
MS_LOG(ERROR) << "Init tmp buffer failed.";
return RET_ERROR;
}

int in_batch = conv_param_->input_batch_;
int in_h = conv_param_->input_h_;
int in_w = conv_param_->input_w_;
int in_channel = conv_param_->input_channel_;
convert_func_(reinterpret_cast<void *>(execute_input_), nhwc4_input_, in_batch, in_h * in_w, in_channel);

int error_code = ParallelLaunch(this->context_->thread_pool_, ConvolutionSWFp16Impl, this, thread_count_);
if (error_code != RET_OK) {
MS_LOG(ERROR) << "conv fp16 error error_code[" << error_code << "]";
FreeTmpBuffer();
return RET_ERROR;
}

// output nhwc4
int oc4_res = conv_param_->output_channel_ % C4NUM;
if (oc4_res != 0) {
PackNHWC4ToNHWCFp16(reinterpret_cast<const void *>(tmp_output_block_), reinterpret_cast<void *>(execute_output_),
conv_param_->output_batch_, conv_param_->output_h_ * conv_param_->output_w_,
conv_param_->output_channel_);
}
ConvolutionBaseFP16CPUKernel::IfCastOutput();
ConvolutionBaseFP16CPUKernel::FreeTmpBuffer();
FreeTmpBuffer();
return RET_OK;
}
} // namespace mindspore::kernel

+ 0
- 72
mindspore/lite/src/runtime/kernel/arm/fp16/convolution_sw_fp16.h View File

@@ -1,72 +0,0 @@
/**
* Copyright 2020 Huawei Technologies Co., Ltd
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*/
#ifndef MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_FP16_CONVOLUTION_SW_FP16_H_
#define MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_FP16_CONVOLUTION_SW_FP16_H_

#include <arm_neon.h>
#include <vector>
#include "src/lite_kernel.h"
#include "src/runtime/kernel/arm/fp16/convolution_base_fp16.h"

namespace mindspore::kernel {
class ConvolutionSWFP16CPUKernel : public ConvolutionBaseFP16CPUKernel {
public:
ConvolutionSWFP16CPUKernel(OpParameter *parameter, const std::vector<lite::Tensor *> &inputs,
const std::vector<lite::Tensor *> &outputs, const InnerContext *ctx,
const mindspore::lite::PrimitiveC *primitive)
: ConvolutionBaseFP16CPUKernel(parameter, inputs, outputs, ctx, primitive) {}
~ConvolutionSWFP16CPUKernel() override {
if (fp16_weight_ != nullptr) {
free(fp16_weight_);
fp16_weight_ = nullptr;
}
if (packed_weight_ != nullptr) {
free(packed_weight_);
packed_weight_ = nullptr;
}
if (slidingWindow_param_ != nullptr) {
delete slidingWindow_param_;
slidingWindow_param_ = nullptr;
}
}

int Init() override;
int ReSize() override;
int Run() override;
int RunImpl(int task_id);
int InitWeightBias();
int InitTmpBuffer();
void ConfigInputOutput();
int ProcessFilter();

private:
void FreeTmpBuffer() {
if (nhwc4_input_ != nullptr) {
ctx_->allocator->Free(nhwc4_input_);
nhwc4_input_ = nullptr;
}
if (tmp_output_block_ != nullptr) {
ctx_->allocator->Free(tmp_output_block_);
tmp_output_block_ = nullptr;
}
}
float16_t *packed_weight_ = nullptr;
float16_t *tmp_output_block_ = nullptr;
SlidingWindowParam *slidingWindow_param_ = nullptr;
};
} // namespace mindspore::kernel

#endif // MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_FP16_CONVOLUTION_SW_FP16_H_

+ 27
- 148
mindspore/lite/src/runtime/kernel/arm/fp16/convolution_winograd_fp16.cc View File

@@ -138,50 +138,6 @@ int ConvolutionWinogradFP16CPUKernel::WinogradFilterTransformFp16(const float16_
return RET_OK;
}

int ConvolutionWinogradFP16CPUKernel::MallocTransformMatrices() {
matrix_a_ = reinterpret_cast<float16_t *>(malloc(input_unit_ * output_unit_ * sizeof(float16_t)));
if (matrix_a_ == nullptr) {
MS_LOG(ERROR) << "malloc matrix_a_ failed.";
return RET_ERROR;
}
matrix_at_ = reinterpret_cast<float16_t *>(malloc(input_unit_ * output_unit_ * sizeof(float16_t)));
if (matrix_at_ == nullptr) {
MS_LOG(ERROR) << "malloc matrix_at_ failed.";
return RET_ERROR;
}
matrix_b_ = reinterpret_cast<float16_t *>(malloc(input_unit_ * input_unit_ * sizeof(float16_t)));
if (matrix_b_ == nullptr) {
MS_LOG(ERROR) << "malloc matrix_b_ failed.";
return RET_ERROR;
}
matrix_bt_ = reinterpret_cast<float16_t *>(malloc(input_unit_ * input_unit_ * sizeof(float16_t)));
if (matrix_bt_ == nullptr) {
MS_LOG(ERROR) << "malloc matrix_bt_ failed.";
return RET_ERROR;
}
return RET_OK;
}

void ConvolutionWinogradFP16CPUKernel::FreeTransformMatrices() {
if (matrix_a_ != nullptr) {
free(matrix_a_);
matrix_a_ = nullptr;
}
if (matrix_at_ != nullptr) {
free(matrix_at_);
matrix_at_ = nullptr;
}
if (matrix_b_ != nullptr) {
free(matrix_b_);
matrix_b_ = nullptr;
}
if (matrix_bt_ != nullptr) {
free(matrix_bt_);
matrix_bt_ = nullptr;
}
return;
}

int ConvolutionWinogradFP16CPUKernel::InitWeightBias() {
auto filter_tensor = in_tensors_.at(kWeightIndex);
int in_channel = filter_tensor->Channel();
@@ -190,9 +146,8 @@ int ConvolutionWinogradFP16CPUKernel::InitWeightBias() {
conv_param_->input_channel_ = in_channel;
conv_param_->output_channel_ = out_channel;

int oc_block, oc_block_num;
oc_block = C8NUM;
oc_block_num = UP_DIV(out_channel, C8NUM);
const int oc_block = C8NUM;
int oc_block_num = UP_DIV(out_channel, C8NUM);

// init weight
auto ret = ConvolutionBaseFP16CPUKernel::GetExecuteFilter();
@@ -209,49 +164,24 @@ int ConvolutionWinogradFP16CPUKernel::InitWeightBias() {
return RET_ERROR;
}
memset(trans_weight_, 0, trans_matrix_data_size);
auto *matrix_g = reinterpret_cast<float *>(malloc(input_unit_ * kernel_unit_ * sizeof(float)));
if (matrix_g == nullptr) {
MS_LOG(ERROR) << "malloc matrix_g failed.";
return RET_ERROR;
}
auto matrix_gt = reinterpret_cast<float *>(malloc(input_unit_ * kernel_unit_ * sizeof(float)));
if (matrix_gt == nullptr) {
free(matrix_g);
MS_LOG(ERROR) << "malloc matrix_gt failed.";
return RET_ERROR;
}
ret = MallocTransformMatrices();
if (ret != RET_OK) {
free(matrix_g);
free(matrix_gt);
MS_LOG(ERROR) << "Malloc transform matrices failed.";
return ret;
}

float matrix_a[MAX_LEN];
float matrix_at[MAX_LEN];
float matrix_b[MAX_LEN];
float matrix_bt[MAX_LEN];
ret = CookToomFilter(matrix_a, matrix_at, matrix_b, matrix_bt, matrix_g, matrix_gt, 0.5f, output_unit_, kernel_unit_);
float matrix_g[64];
float matrix_gt[64];
float matrix_a[64];
float matrix_at[64];
float matrix_b[64];
float matrix_bt[64];
float coef = 1.0f;
if (input_unit_ == 8) {
coef = 0.5f;
}
ret = CookToomFilter(matrix_a, matrix_at, matrix_b, matrix_bt, matrix_g, matrix_gt, coef, output_unit_, kernel_unit_);
if (ret != RET_OK) {
free(matrix_g);
free(matrix_gt);
MS_LOG(ERROR) << "get matrix g from CookToomFilter failed.";
return ret;
}
Float32ToFloat16(matrix_a, matrix_a_, input_unit_ * output_unit_);
Float32ToFloat16(matrix_at, matrix_at_, input_unit_ * output_unit_);
Float32ToFloat16(matrix_b, matrix_b_, input_unit_ * input_unit_);
Float32ToFloat16(matrix_bt, matrix_bt_, input_unit_ * input_unit_);
matrices_[0] = matrix_a_;
matrices_[1] = matrix_at_;
matrices_[2] = matrix_b_;
matrices_[3] = matrix_bt_;

ret = WinogradFilterTransformFp16(execute_weight_, matrix_g, matrix_gt, oc_block);
if (ret != RET_OK) {
free(matrix_g);
free(matrix_gt);
MS_LOG(ERROR) << "winograd filter transfrom failed.";
return ret;
}
@@ -259,8 +189,6 @@ int ConvolutionWinogradFP16CPUKernel::InitWeightBias() {
// init bias
bias_data_ = malloc(oc_block_num * oc_block * sizeof(float16_t));
if (bias_data_ == nullptr) {
free(matrix_g);
free(matrix_gt);
MS_LOG(ERROR) << "malloc bias_data_ failed.";
return RET_ERROR;
}
@@ -274,27 +202,15 @@ int ConvolutionWinogradFP16CPUKernel::InitWeightBias() {
} else {
MS_ASSERT(inputs_.size() == kInputSize1);
}
free(matrix_g);
free(matrix_gt);
return RET_OK;
}

int ConvolutionWinogradFP16CPUKernel::InitTmpBuffer() {
int cal_num = 16;
const int cal_num = 16;
int channel_out = conv_param_->output_channel_;
int output_h = conv_param_->output_h_;
int output_w = conv_param_->output_w_;
int oc8 = UP_DIV(channel_out, C8NUM);
int ic8 = UP_DIV(conv_param_->input_channel_, C8NUM);

size_t nhwc8_input_size =
ic8 * C8NUM * conv_param_->input_batch_ * conv_param_->input_h_ * conv_param_->input_w_ * sizeof(float16_t);
nhwc4_input_ = ctx_->allocator->Malloc(nhwc8_input_size);
if (nhwc4_input_ == nullptr) {
MS_LOG(ERROR) << "malloc nhwc4_input_ failed.";
return RET_ERROR;
}

size_t tile_buffer_size = thread_count_ * cal_num * input_unit_ * input_unit_ * ic8 * C8NUM * sizeof(float16_t);
trans_input_ = reinterpret_cast<float16_t *>(ctx_->allocator->Malloc(tile_buffer_size));
if (trans_input_ == nullptr) {
@@ -309,16 +225,6 @@ int ConvolutionWinogradFP16CPUKernel::InitTmpBuffer() {
return RET_ERROR;
}

int out_w_block = UP_DIV(output_w, output_unit_);
int out_h_block = UP_DIV(output_h, output_unit_);
tmp_out_data_ = reinterpret_cast<float16_t *>(
ctx_->allocator->Malloc(conv_param_->output_batch_ * out_w_block * out_h_block * output_unit_ * output_unit_ * oc8 *
C8NUM * sizeof(float16_t)));
if (tmp_out_data_ == nullptr) {
MS_LOG(ERROR) << "malloc tmp_out_data_ failed.";
return RET_ERROR;
}

tmp_data_ = reinterpret_cast<float16_t *>(
ctx_->allocator->Malloc(thread_count_ * C8NUM * input_unit_ * input_unit_ * sizeof(float16_t)));
if (tmp_data_ == nullptr) {
@@ -328,14 +234,21 @@ int ConvolutionWinogradFP16CPUKernel::InitTmpBuffer() {

tmp_buffer_address_list_[0] = trans_input_;
tmp_buffer_address_list_[1] = gemm_out_;
tmp_buffer_address_list_[2] = tmp_out_data_;
tmp_buffer_address_list_[3] = tmp_data_;
tmp_buffer_address_list_[2] = tmp_data_;
return RET_OK;
}

int ConvolutionWinogradFP16CPUKernel::ConfigInputOutput() {
auto output_tensor = out_tensors_.at(kOutputIndex);
output_tensor->SetFormat(schema::Format_NHWC);
in_func_ = GetInputTransFp16Func(input_unit_);
if (in_func_ == nullptr) {
MS_LOG(ERROR) << "in_func_ is null.";
return RET_ERROR;
}
out_func_ = GetOutputTransFp16Func(input_unit_, output_unit_, conv_param_->act_type_);
if (out_func_ == nullptr) {
MS_LOG(ERROR) << "out_func_ is null.";
return RET_ERROR;
}
return RET_OK;
}

@@ -381,9 +294,8 @@ int ConvolutionWinogradFP16CPUKernel::ReSize() {
}

int ConvolutionWinogradFP16CPUKernel::RunImpl(int task_id) {
ConvWinogardFp16(reinterpret_cast<float16_t *>(nhwc4_input_), trans_weight_,
reinterpret_cast<const float16_t *>(bias_data_), tmp_buffer_address_list_, task_id, conv_param_,
matrices_);
ConvWinogardFp16(execute_input_, trans_weight_, reinterpret_cast<const float16_t *>(bias_data_), execute_output_,
tmp_buffer_address_list_, task_id, conv_param_, in_func_, out_func_);
return RET_OK;
}

@@ -397,28 +309,6 @@ static int ConvolutionWinogradFp16Impl(void *cdata, int task_id) {
return RET_OK;
}

int ConvolutionWinogradFP16CPUKernel::PostProcess() {
auto act_type = conv_param_->act_type_;
switch (act_type) {
case ActType_No:
UnPackWinogradOutputFp16(tmp_out_data_, execute_output_, conv_param_->output_batch_, conv_param_->output_h_,
conv_param_->output_w_, conv_param_->output_channel_, output_unit_);
break;
case ActType_Relu:
UnPackWinogradReluOutputFp16(tmp_out_data_, execute_output_, conv_param_->output_batch_, conv_param_->output_h_,
conv_param_->output_w_, conv_param_->output_channel_, output_unit_);
break;
case ActType_Relu6:
UnPackWinogradRelu6OutputFp16(tmp_out_data_, execute_output_, conv_param_->output_batch_, conv_param_->output_h_,
conv_param_->output_w_, conv_param_->output_channel_, output_unit_);
break;
default:
MS_LOG(ERROR) << "Unsupport activation type.";
return RET_ERROR;
}
return RET_OK;
}

int ConvolutionWinogradFP16CPUKernel::Run() {
auto prepare_ret = Prepare();
if (prepare_ret != RET_OK) {
@@ -438,12 +328,6 @@ int ConvolutionWinogradFP16CPUKernel::Run() {
return RET_ERROR;
}

int in_batch = conv_param_->input_batch_;
int in_h = conv_param_->input_h_;
int in_w = conv_param_->input_w_;
int in_channel = conv_param_->input_channel_;
PackNHWCToNHWC8Fp16(execute_input_, nhwc4_input_, in_batch, in_h * in_w, in_channel);

int error_code = ParallelLaunch(this->context_->thread_pool_, ConvolutionWinogradFp16Impl, this, thread_count_);
if (error_code != RET_OK) {
MS_LOG(ERROR) << "conv winograd error error_code[" << error_code << "]";
@@ -451,11 +335,6 @@ int ConvolutionWinogradFP16CPUKernel::Run() {
return RET_ERROR;
}

ret = PostProcess();
if (ret != RET_OK) {
MS_LOG(ERROR) << "Post process failed.";
return ret;
}
ConvolutionBaseFP16CPUKernel::IfCastOutput();
ConvolutionBaseFP16CPUKernel::FreeTmpBuffer();
FreeTmpBuffer();


+ 3
- 19
mindspore/lite/src/runtime/kernel/arm/fp16/convolution_winograd_fp16.h View File

@@ -42,7 +42,6 @@ class ConvolutionWinogradFP16CPUKernel : public ConvolutionBaseFP16CPUKernel {
free(trans_weight_);
trans_weight_ = nullptr;
}
FreeTransformMatrices();
}

int Init() override;
@@ -50,19 +49,12 @@ class ConvolutionWinogradFP16CPUKernel : public ConvolutionBaseFP16CPUKernel {
int Run() override;
int RunImpl(int task_id);
int InitWeightBias();
int MallocTransformMatrices();
void FreeTransformMatrices();
int InitTmpBuffer();
int ConfigInputOutput();
int PostProcess();
int WinogradFilterTransformFp16(const float16_t *weight_data, float *matrix_g, float *matrix_gt, int oc_block);

private:
void FreeTmpBuffer() {
if (nhwc4_input_ != nullptr) {
ctx_->allocator->Free(nhwc4_input_);
nhwc4_input_ = nullptr;
}
if (trans_input_ != nullptr) {
ctx_->allocator->Free(trans_input_);
trans_input_ = nullptr;
@@ -75,10 +67,6 @@ class ConvolutionWinogradFP16CPUKernel : public ConvolutionBaseFP16CPUKernel {
ctx_->allocator->Free(gemm_out_);
gemm_out_ = nullptr;
}
if (tmp_out_data_ != nullptr) {
ctx_->allocator->Free(tmp_out_data_);
tmp_out_data_ = nullptr;
}
}
int kernel_unit_;
int input_unit_;
@@ -86,14 +74,10 @@ class ConvolutionWinogradFP16CPUKernel : public ConvolutionBaseFP16CPUKernel {
float16_t *tmp_data_ = nullptr;
float16_t *trans_input_ = nullptr;
float16_t *gemm_out_ = nullptr;
float16_t *tmp_out_data_ = nullptr;
float16_t *matrix_a_ = nullptr;
float16_t *matrix_at_ = nullptr;
float16_t *matrix_b_ = nullptr;
float16_t *matrix_bt_ = nullptr;
float16_t *trans_weight_ = nullptr;
TmpBufferAddressFp16 tmp_buffer_address_list_[4];
MatricesFp16 matrices_[4];
TmpBufferAddressFp16 tmp_buffer_address_list_[3];
InputTransFp16Func in_func_;
OutputTransFp16Func out_func_;
};
} // namespace mindspore::kernel



+ 4
- 22
mindspore/lite/src/runtime/kernel/arm/fp32/convolution_winograd.cc View File

@@ -117,9 +117,8 @@ int ConvolutionWinogradCPUKernel::InitWeightBias() {
conv_param_->output_channel_ = out_channel;

int oc4 = UP_DIV(out_channel, C4NUM);
int oc_block, oc_block_num;
oc_block = C8NUM;
oc_block_num = UP_DIV(out_channel, C8NUM);
const int oc_block = C8NUM;
int oc_block_num = UP_DIV(out_channel, C8NUM);

// set data
auto trans_matrix_data_size = input_unit_ * input_unit_ * ic4 * C4NUM * oc_block_num * oc_block * sizeof(float);
@@ -172,9 +171,6 @@ int ConvolutionWinogradCPUKernel::InitWeightBias() {

int ConvolutionWinogradCPUKernel::InitTmpBuffer() {
int channel_out = conv_param_->output_channel_;
int output_h = conv_param_->output_h_;
int output_w = conv_param_->output_w_;
int oc4 = UP_DIV(channel_out, C4NUM);
int oc8 = UP_DIV(channel_out, C8NUM);
int ic4 = UP_DIV(conv_param_->input_channel_, C4NUM);
#ifdef ENABLE_ARM32
@@ -198,16 +194,6 @@ int ConvolutionWinogradCPUKernel::InitTmpBuffer() {
return RET_ERROR;
}

int out_w_block = UP_DIV(output_w, output_unit_);
int out_h_block = UP_DIV(output_h, output_unit_);
tmp_out_data_ =
reinterpret_cast<float *>(ctx_->allocator->Malloc(conv_param_->output_batch_ * out_w_block * out_h_block *
output_unit_ * output_unit_ * oc4 * C4NUM * sizeof(float)));
if (tmp_out_data_ == nullptr) {
MS_LOG(ERROR) << "malloc tmp_out_data_ failed.";
return RET_MEMORY_FAILED;
}

tmp_data_ = reinterpret_cast<float *>(
ctx_->allocator->Malloc(thread_count_ * C4NUM * input_unit_ * input_unit_ * sizeof(float)));
if (tmp_data_ == nullptr) {
@@ -224,16 +210,12 @@ int ConvolutionWinogradCPUKernel::InitTmpBuffer() {

tmp_buffer_address_list_[0] = trans_input_;
tmp_buffer_address_list_[1] = gemm_out_;
tmp_buffer_address_list_[2] = tmp_out_data_;
tmp_buffer_address_list_[3] = tmp_data_;
tmp_buffer_address_list_[4] = col_buffer_;
tmp_buffer_address_list_[2] = tmp_data_;
tmp_buffer_address_list_[3] = col_buffer_;
return RET_OK;
}

int ConvolutionWinogradCPUKernel::ConfigInputOutput() {
auto output_tensor = out_tensors_.at(kOutputIndex);
output_tensor->SetFormat(schema::Format::Format_NHWC);

in_func_ = GetInputTransFunc(input_unit_);
if (in_func_ == nullptr) {
MS_LOG(ERROR) << "in_func_ is null.";


+ 1
- 6
mindspore/lite/src/runtime/kernel/arm/fp32/convolution_winograd.h View File

@@ -61,10 +61,6 @@ class ConvolutionWinogradCPUKernel : public ConvolutionBaseCPUKernel {
ctx_->allocator->Free(gemm_out_);
gemm_out_ = nullptr;
}
if (tmp_out_data_ != nullptr) {
ctx_->allocator->Free(tmp_out_data_);
tmp_out_data_ = nullptr;
}
if (col_buffer_ != nullptr) {
ctx_->allocator->Free(col_buffer_);
col_buffer_ = nullptr;
@@ -76,10 +72,9 @@ class ConvolutionWinogradCPUKernel : public ConvolutionBaseCPUKernel {
float *tmp_data_ = nullptr;
float *trans_input_ = nullptr;
float *gemm_out_ = nullptr;
float *tmp_out_data_ = nullptr;
float *col_buffer_ = nullptr;
float *trans_weight_ = nullptr;
TmpBufferAddress tmp_buffer_address_list_[5];
TmpBufferAddress tmp_buffer_address_list_[4];
InputTransFunc in_func_;
OutputTransFunc out_func_;
};


Write Preview
Loading…
Cancel
Save