| /third_party/mindspore/mindspore/ccsrc/backend/kernel_compiler/cpu/nnacl/fp16/ |
| D | winograd_transform_fp16.c | 28 int input_unit = conv_param->input_unit_; in WinogradInputTransformFp16() local
44 int src_x_e = src_x_s + input_unit; in WinogradInputTransformFp16()
45 int src_y_e = src_y_s + input_unit; in WinogradInputTransformFp16()
46 int interval_x_e = src_x_e < input_w ? input_unit : (input_w - src_x_s); in WinogradInputTransformFp16()
47 int interval_y_e = src_y_e < input_h ? input_unit : (input_h - src_y_s); in WinogradInputTransformFp16()
53 memset(tmp_data, 0, input_unit * input_unit * C8NUM * sizeof(float16_t)); in WinogradInputTransformFp16()
63 int dst_y_offset = interval * input_unit * C8NUM + interval_x_s * C8NUM; in WinogradInputTransformFp16()
81 int dst_y_offset = interval * input_unit * C8NUM + interval_x_s * C8NUM; in WinogradInputTransformFp16()
105 int dst_y_offset = interval * input_unit * C8NUM + interval_x_s * C8NUM; in WinogradInputTransformFp16()
136 int input_unit = conv_param->input_unit_; in WinogradOutputNHWCTransformFp16() local
[all …]
|
| D | winograd_transform_fp16.h | 46 … const float *matrix_gt, int oc_block, int input_unit, int kernel_unit,
|
| D | winograd_utils_fp16.c | 82 InputTransFp16Func GetInputTransFp16Func(int input_unit) { return InputTransFp16FuncList[input_unit… in GetInputTransFp16Func() argument
432 OutputTransFp16Func GetOutputTransFp16Func(int input_unit, int output_unit, ActType act_type) { in GetOutputTransFp16Func() argument
433 if (input_unit == 4 && output_unit < 4) { in GetOutputTransFp16Func()
441 } else if (input_unit == 6 && output_unit < 6) { in GetOutputTransFp16Func()
449 } else if (input_unit == 8 && output_unit < 8) { in GetOutputTransFp16Func()
2906 int input_unit = i + kernel_w - 1; in SelectOutputUnitFp16() local
2907 if (!GetOutputTransFp16Func(input_unit, i, ActType_No)) { in SelectOutputUnitFp16()
2910 float penalty = ((float)input_unit * input_unit) / ((float)kernel_h * kernel_w) * 0.12f; in SelectOutputUnitFp16()
2911 …float wino_cost = ((2 + out_c) * (float)input_unit * input_unit * in_c + ((float)input_unit + i) *… in SelectOutputUnitFp16()
|
| D | winograd_utils_fp16.h | 279 InputTransFp16Func GetInputTransFp16Func(int input_unit);
287 OutputTransFp16Func GetOutputTransFp16Func(int input_unit, int output_unit, ActType act_type);
|
| /third_party/mindspore/mindspore/ccsrc/backend/kernel_compiler/cpu/nnacl/base/ |
| D | conv_common_base.c | 32 bool CheckWinogradInputOutputUnit(int input_unit, int output_unit) { in CheckWinogradInputOutputUnit() argument
33 if (input_unit != 4 && input_unit != 6 && input_unit != 8) { in CheckWinogradInputOutputUnit()
36 if ((output_unit >= input_unit) || (output_unit < 2)) { in CheckWinogradInputOutputUnit()
65 int input_unit = i + kernel_w - 1; in SelectOutputUnit() local
66 if (!CheckWinogradInputOutputUnit(input_unit, i)) { in SelectOutputUnit()
69 float penalty = ((float)input_unit * input_unit) / ((float)kernel_h * kernel_w) * 0.12f; in SelectOutputUnit()
70 …float wino_cost = ((2 + out_c) * (float)input_unit * input_unit * in_c + ((float)input_unit + i) *… in SelectOutputUnit()
|
| D | minimal_filtering_generator.c | 256 … int oc_block, int input_unit, int kernel_unit, int channel, int batch, bool pack) { in WinogradWeightTransform() argument
267 float *tmp_data = (float *)(malloc(channel * input_unit * kernel_unit * sizeof(float))); in WinogradWeightTransform()
271 float *trans_out_data = (float *)(malloc(channel * input_unit * input_unit * sizeof(float))); in WinogradWeightTransform()
278 float *tmp_data1 = (float *)(malloc(channel * input_unit * kernel_unit * sizeof(float))); in WinogradWeightTransform()
284 float *trans_out_data1 = (float *)(malloc(channel * input_unit * input_unit * sizeof(float))); in WinogradWeightTransform()
301 …ograd(weight_data + i * input_oz_offset, matrix_gt, tmp_data, kernel_unit, kernel_unit, input_unit, in WinogradWeightTransform()
304 PackHWCToWHC(tmp_data, tmp_data1, kernel_unit, input_unit, channel); in WinogradWeightTransform()
306 …rixMultiplyWinograd(tmp_data1, matrix_gt, trans_out_data1, input_unit, kernel_unit, input_unit, ch… in WinogradWeightTransform()
309 PackHWCToWHC(trans_out_data1, trans_out_data, input_unit, input_unit, channel); in WinogradWeightTransform()
312 …ograd(weight_data + i * input_oz_offset, matrix_gt, tmp_data, kernel_unit, kernel_unit, input_unit, in WinogradWeightTransform()
[all …]
|
| D | conv_common_base.h | 30 bool CheckWinogradInputOutputUnit(int input_unit, int output_unit);
|
| /third_party/mindspore/mindspore/ccsrc/backend/kernel_compiler/cpu/nnacl/fp32/ |
| D | winograd_transform.c | 24 int input_unit = conv_param->input_unit_; in WinogradInputTransform() local
39 int src_x_e = src_x_s + input_unit; in WinogradInputTransform()
40 int src_y_e = src_y_s + input_unit; in WinogradInputTransform()
41 int interval_x_e = src_x_e < input_w ? input_unit : (input_w - src_x_s); in WinogradInputTransform()
42 int interval_y_e = src_y_e < input_h ? input_unit : (input_h - src_y_s); in WinogradInputTransform()
48 memset(tmp_data, 0, input_unit * input_unit * C4NUM * (int)(sizeof(float))); in WinogradInputTransform()
57 int dst_y_offset = interval * input_unit * C4NUM + interval_x_s * C4NUM; in WinogradInputTransform()
75 int dst_y_offset = interval * input_unit * C4NUM + interval_x_s * C4NUM; in WinogradInputTransform()
109 int input_unit = conv_param->input_unit_; in WinogradOutputNHWCTransform() local
124 int src_tile_offset = i * oc8 * C8NUM * input_unit * input_unit; in WinogradOutputNHWCTransform()
[all …]
|
| D | winograd_utils.h | 156 InputTransFunc GetInputTransFunc(int input_unit);
164 OutputTransFunc GetOutputTransFunc(int input_unit, int output_unit, ActType act_type);
|
| D | winograd_utils.c | 53 InputTransFunc GetInputTransFunc(int input_unit) { return InputTransFuncList[input_unit]; } in GetInputTransFunc() argument
307 OutputTransFunc GetOutputTransFunc(int input_unit, int output_unit, ActType act_type) { in GetOutputTransFunc() argument
308 if (!CheckWinogradInputOutputUnit(input_unit, output_unit)) { in GetOutputTransFunc()
311 int in_index = (input_unit - 4) / 2; in GetOutputTransFunc()
324 return OutputTransFuncList[index + (input_unit - 2) * act_index + output_unit - 2]; in GetOutputTransFunc()
|
| /third_party/mindspore/mindspore/lite/micro/coder/opcoders/nnacl/fp32/ |
| D | convolution_winograd_fp32_coder.cc | 182 std::string ConvolutionWinogradFP32Coder::GetInputTransFunc(int input_unit) { in GetInputTransFunc() argument
184 return InputTransFuncList.at(input_unit); in GetInputTransFunc()
187 std::string ConvolutionWinogradFP32Coder::GetOutputTransFunc(int input_unit, int output_unit, ActTy… in GetOutputTransFunc() argument
189 if (input_unit == DIMENSION_4D && output_unit < DIMENSION_4D) { in GetOutputTransFunc()
197 } else if (input_unit == DIMENSION_6D && output_unit < DIMENSION_6D) { in GetOutputTransFunc()
205 } else if (input_unit == DIMENSION_8D && output_unit < DIMENSION_8D) { in GetOutputTransFunc()
|
| D | convolution_winograd_fp32_coder.h | 50 std::string GetInputTransFunc(int input_unit);
52 std::string GetOutputTransFunc(int input_unit, int output_unit, ActType act_type);
|
| /third_party/mindspore/mindspore/ccsrc/backend/kernel_compiler/cpu/nnacl/int8/ |
| D | conv3x3_int8.c | 190 const int input_unit = 4; in Conv3x3Int8FilterTransform() local
196 … int dst_oc_offset = oc4_block_num * C4NUM * iC8 * C8NUM * input_unit * input_unit + oc4_block_rem; in Conv3x3Int8FilterTransform()
754 const int input_unit = 4; in Conv3x3Int8OutputTransform() local
761 int src_tile_offset = i * oc4 * C4NUM * input_unit * input_unit; in Conv3x3Int8OutputTransform()
765 int src_oc4_offset = src_tile_offset + j * input_unit * input_unit * C4NUM; in Conv3x3Int8OutputTransform()
792 const int input_unit = 4; in Conv3x3Int8InputTransform() local
801 int real_x_end = (origin_x + input_unit) < input_width ? input_unit : (input_width - origin_x); in Conv3x3Int8InputTransform()
803 … int real_y_end = (origin_y + input_unit) < input_height ? input_unit : (input_height - origin_y); in Conv3x3Int8InputTransform()
809 for (int i = 0; i < input_unit * input_unit * TILE_NUM; i++) tmp_data[i] = input_zp; in Conv3x3Int8InputTransform()
|