Home
last modified time | relevance | path

Searched refs:output_channel (Results 1 – 25 of 50) sorted by relevance

12

/third_party/mindspore/mindspore/lite/micro/coder/wrapper/int8/
Dconv1x1_init_int8_wrapper.c22 int32_t output_channel, int32_t input_zp, bool support_optimize, bool filter_peroc, in Conv1x1Init() argument
30 size_t size = UP_ROUND(input_channel, C16NUM) * UP_ROUND(output_channel, C2NUM) * sizeof(int8_t); in Conv1x1Init()
36 RowMajor2Row2x16MajorInt8(src_weight, packed_weight_, output_channel, input_channel); in Conv1x1Init()
38 size = (size_t)UP_ROUND(output_channel, C2NUM); in Conv1x1Init()
46 memcpy(bias_data_, src_bias, (size_t)output_channel * sizeof(int32_t)); in Conv1x1Init()
51 …size_t size = support_optimize ? UP_ROUND(input_channel, C4NUM) * UP_ROUND(output_channel, C16NUM)… in Conv1x1Init()
52 … : UP_ROUND(input_channel, C16NUM) * UP_ROUND(output_channel, C4NUM) * sizeof(int8_t); in Conv1x1Init()
59 RowMajor2Row4x16MajorInt8(src_weight, packed_weight_, output_channel, input_channel); in Conv1x1Init()
61 RowMajor2Row16x4MajorInt8(src_weight, packed_weight_, output_channel, input_channel); in Conv1x1Init()
64 size = support_optimize ? UP_ROUND(output_channel, C16NUM) : UP_ROUND(output_channel, C4NUM); in Conv1x1Init()
[all …]
Dconv_init_int8_wrapper.c23 … int kernel_w, int input_channel, int output_channel, int32_t input_zp, bool filter_peroc, in ConvInit() argument
31 up_round_oc = UP_ROUND(output_channel, C2NUM); in ConvInit()
35 up_round_oc = UP_ROUND(output_channel, C8NUM); in ConvInit()
38 up_round_oc = UP_ROUND(output_channel, C4NUM); in ConvInit()
52 …RowMajor2Row2x16MajorInt8(origin_weight, packed_weight_, output_channel, input_channel * kernel_pl… in ConvInit()
55 …RowMajor2Row8x4MajorInt8(origin_weight, packed_weight_, output_channel, input_channel * kernel_pla… in ConvInit()
57 …RowMajor2Row16x4MajorInt8(origin_weight, packed_weight_, output_channel, input_channel * kernel_pl… in ConvInit()
69 memcpy(bias_data_, ori_bias, (unsigned int)output_channel * sizeof(int32_t)); in ConvInit()
72 for (int oc = 0; oc < output_channel; oc++) { in ConvInit()
/third_party/mindspore/tests/st/fusion/
Dtest_conv_bn1_fusion.py26 output_channel = 512 variable
60 self.conv = nn.Conv2d(input_channel, output_channel,
62 self.conv1 = nn.Conv2d(input_channel, output_channel,
64 self.bn = nn.BatchNorm2d(output_channel, momentum=0.1, eps=0.0001)
69 self.dense = nn.Dense(output_channel, num_class)
77 output = self.reshape(output, (batch_size, output_channel))
91 self.conv = nn.Conv2d(input_channel, output_channel,
93 self.bn = nn.BatchNorm2d(output_channel, momentum=0.1, eps=0.0001)
97 self.dense = nn.Dense(output_channel, num_class)
104 output = self.reshape(output, (batch_size, output_channel))
[all …]
/third_party/mindspore/mindspore/ccsrc/backend/kernel_compiler/cpu/nnacl/fp32/
Dcommon_func_fp32.c19 …stConvFuncComm(const float *src_ptr_, float *out_ptr, const float *bias_ptr, size_t output_channel, in PostConvFuncComm() argument
24 for (size_t oc = 0; oc < output_channel; oc++) { in PostConvFuncComm()
42 …nvFuncFp32C8(const float *c8_out_ptr, float *out_ptr, const float *bias_ptr, size_t output_channel, in PostConvFuncFp32C8() argument
45 …PostConvFuncComm(c8_out_ptr, out_ptr, bias_ptr, output_channel, plane_size, plane_size, stride, re… in PostConvFuncFp32C8()
47 size_t oc8mod = output_channel % C8NUM; in PostConvFuncFp32C8()
48 size_t oc8div = output_channel - oc8mod; in PostConvFuncFp32C8()
55 …nvFuncFp32C4(const float *c4_out_ptr, float *out_ptr, const float *bias_ptr, size_t output_channel, in PostConvFuncFp32C4() argument
58 size_t oc4mod = output_channel % C4NUM; in PostConvFuncFp32C4()
59 size_t oc4div = output_channel - oc4mod; in PostConvFuncFp32C4()
63 …PostConvFuncComm(c4_out_ptr, out_ptr, bias_ptr, output_channel, plane_size, plane_stride, output_c… in PostConvFuncFp32C4()
Dwinograd_transform.c104 int output_channel = conv_param->output_channel_; in WinogradOutputNHWCTransform() local
106 int oc4 = UP_DIV(output_channel, C4NUM); in WinogradOutputNHWCTransform()
108 int oc8 = UP_DIV(output_channel, C8NUM); in WinogradOutputNHWCTransform()
125 int dst_tile_offset = output_channel * (dst_x_s + dst_y_s * output_w); in WinogradOutputNHWCTransform()
131 int r_c = output_channel - j * C4NUM; in WinogradOutputNHWCTransform()
138 func(src_ptr, dst_ptr, bias_ptr, C8NUM, output_w, output_channel, r_w, r_h, r_c); in WinogradOutputNHWCTransform()
143 int r_c = output_channel - j * C8NUM; in WinogradOutputNHWCTransform()
150 func(src_ptr, dst_ptr, bias_ptr, C8NUM, output_w, output_channel, r_w, r_h, r_c); in WinogradOutputNHWCTransform()
164 int output_channel = conv_param->output_channel_; in WinogradOutputNC4HW4Transform() local
166 int oc4 = UP_DIV(output_channel, C4NUM); in WinogradOutputNC4HW4Transform()
[all …]
Ddeconv_fp32.c19 void PackDeConvWeightFp32(const float *weight, float *dst, int input_channel, int output_channel, i… in PackDeConvWeightFp32() argument
22 for (int oc = 0; oc < output_channel; oc++) { in PackDeConvWeightFp32()
27 int src_index = ic * plane * output_channel + hw * output_channel + oc; in PackDeConvWeightFp32()
36 …d DeConvPostFp32C8(const float *src, float *tmp, const float *bias, float *dst, int output_channel, in DeConvPostFp32C8() argument
43 int oc8 = UP_ROUND(output_channel, C8NUM); in DeConvPostFp32C8()
108 …PostConvFuncFp32C8(tmp, dst, bias, output_channel, output_plane, conv_param->output_channel_, conv… in DeConvPostFp32C8()
Ddeconv_fp32.h30 void PackDeConvWeightFp32(const float *weight, float *dst, int input_channel, int output_channel, i…
31 …ConvPostFp32C8(const float *src, float *tmp_out, const float *bias, float *dst, int output_channel,
/third_party/mindspore/mindspore/lite/src/runtime/kernel/arm/int8/
Dconvolution_int8.cc53 auto output_channel = filter_tensor->Batch(); in InitWeightBias() local
56 conv_param_->output_channel_ = output_channel; in InitWeightBias()
60 up_round_oc = UP_ROUND(output_channel, C2NUM); in InitWeightBias()
64 up_round_oc = UP_ROUND(output_channel, C8NUM); in InitWeightBias()
67 up_round_oc = UP_ROUND(output_channel, C4NUM); in InitWeightBias()
85 …RowMajor2Row2x16MajorInt8(origin_weight, packed_weight_, output_channel, input_channel * kernel_pl… in InitWeightBias()
88 …RowMajor2Row8x4MajorInt8(origin_weight, packed_weight_, output_channel, input_channel * kernel_pla… in InitWeightBias()
90 …RowMajor2Row16x4MajorInt8(origin_weight, packed_weight_, output_channel, input_channel * kernel_pl… in InitWeightBias()
104 memcpy(bias_data_, ori_bias, static_cast<size_t>(output_channel) * sizeof(int32_t)); in InitWeightBias()
111 filter_zp_ptr_ = reinterpret_cast<int32_t *>(malloc(output_channel * sizeof(int32_t))); in InitWeightBias()
[all …]
Dconvolution_1x1_int8.cc154 …lution1x1Int8CPUKernel::InitBiasByzp(const void *src_weight, int input_channel, int output_channel, in InitBiasByzp() argument
162 for (int oc = 0; oc < output_channel; oc++) { in InitBiasByzp()
178 for (int fi = 0; fi < output_channel; fi++) { in InitBiasByzp()
188 memcpy(left_shift_, conv_param_->conv_quant_arg_.left_shift_, output_channel * sizeof(int32_t)); in InitBiasByzp()
196 … memcpy(right_shift_, conv_param_->conv_quant_arg_.right_shift_, output_channel * sizeof(int32_t)); in InitBiasByzp()
204 …memcpy(multiplier_, conv_param_->conv_quant_arg_.quant_multiplier_, output_channel * sizeof(int32_… in InitBiasByzp()
221 auto output_channel = filter_tensor->Batch(); in InitWeightBias() local
222 if (output_channel < 0) { in InitWeightBias()
227 …size_t size = support_optimize_ ? UP_ROUND(input_channel, C4NUM) * UP_ROUND(output_channel, C16NUM… in InitWeightBias()
228 … : UP_ROUND(input_channel, C16NUM) * UP_ROUND(output_channel, C4NUM) * sizeof(int8_t); in InitWeightBias()
[all …]
Dconvolution_3x3_int8.cc32 auto output_channel = conv_param->output_channel_; in ProcessFilterUint8() local
36 size_t tmp_size = output_channel * iC8 * C8NUM * kernel_plane * sizeof(int16_t); in ProcessFilterUint8()
43 Conv3x3Int8FilterTransform(tmp_addr, dst_weight, iC8, output_channel, kernel_plane); in ProcessFilterUint8()
88 auto output_channel = filter_tensor->Batch(); in InitWeightBias() local
89 if (output_channel < 0) { in InitWeightBias()
94 conv_param_->output_channel_ = output_channel; in InitWeightBias()
96 int oC4 = UP_DIV(output_channel, C4NUM); in InitWeightBias()
125 memcpy(bias_data_, ori_bias_addr, static_cast<size_t>(output_channel) * sizeof(int32_t)); in InitWeightBias()
/third_party/mindspore/mindspore/ccsrc/backend/kernel_compiler/cpu/nnacl/int8/
Ddeconv_int8.c20 …vPostInt8C4(const int32_t *src, const int32_t *bias, int32_t *tmp, int8_t *out, int output_channel, in DeConvPostInt8C4() argument
26 int oc4 = UP_DIV(output_channel, C4NUM); in DeConvPostInt8C4()
84 PostFuncInt8C4(tmp, bias, out, output_channel, (size_t)output_plane, conv_param->output_channel_, in DeConvPostInt8C4()
91 void DeConvWeightTransInt8(const int8_t *src, int8_t *dst, int input_channel, int output_channel, i… in DeConvWeightTransInt8() argument
95 int oc4 = UP_ROUND(output_channel, C4NUM); in DeConvWeightTransInt8()
98 for (int oc = 0; oc < output_channel; oc++) { in DeConvWeightTransInt8()
101 int src_index = ic * output_channel * plane + hw * output_channel + oc; in DeConvWeightTransInt8()
145 …onvPostInt8(const int32_t *src, const int32_t *bias, int32_t *tmp, int8_t *out, int output_channel, in DeConvPostInt8() argument
148 int error_code = DeConvPostInt8C4(src, bias, tmp, out, output_channel, conv_param); in DeConvPostInt8()
Ddeconv_int8.h34 void DeConvWeightTransInt8(const int8_t *src, int8_t *dst, int input_channel, int output_channel, i…
40 …onvPostInt8(const int32_t *src, const int32_t *bias, int32_t *tmp, int8_t *out, int output_channel,
/third_party/mindspore/mindspore/lite/micro/coder/opcoders/nnacl/int8/
Dconv2d_1x1_int8_coder.cc153 int32_t output_channel = filter_tensor_->Batch(); in InitWeightBias() local
156 MS_CHECK_TRUE(output_channel > 0, "output_channel should be positive"); in InitWeightBias()
177 … output_channel, input_zp, "GetSupportOptFlag()", filter_peroc_, packed_weight_str, in InitWeightBias()
181 … output_channel, input_zp, support_optimize_, filter_peroc_, packed_weight_str, in InitWeightBias()
190 int32_t output_channel = filter_tensor_->Batch(); in InitFilterPeroc() local
193 round_oc = UP_ROUND(output_channel, C2NUM); in InitFilterPeroc()
195 round_oc = MSMAX(UP_ROUND(output_channel, C16NUM), UP_ROUND(output_channel, C4NUM)); in InitFilterPeroc()
198 MS_CHECK_TRUE(conv_quant_arg_->filter_arg_num_ == static_cast<size_t>(output_channel), in InitFilterPeroc()
200 size_t output_size = output_channel * sizeof(int32_t); in InitFilterPeroc()
207 for (int fi = 0; fi < output_channel; fi++) { in InitFilterPeroc()
Dconv2d_3x3_int8_coder.cc29 int output_channel = conv_param->output_channel_; in ProcessFilterUint8() local
33 size_t tmp_size = output_channel * iC8 * C8NUM * kernel_plane * sizeof(int16_t); in ProcessFilterUint8()
43 Conv3x3Int8FilterTransform(tmp_addr, dst_weight, iC8, output_channel, kernel_plane); in ProcessFilterUint8()
49 int output_channel = conv_param_->output_channel_; in InitWeightBias() local
51 MS_CHECK_TRUE(output_channel > 0, "invalid output_channel"); in InitWeightBias()
53 int oC4 = UP_DIV(output_channel, C4NUM); in InitWeightBias()
71 …MS_CHECK_RET_CODE(memcpy_s(new_bias_addr_, new_bias_size, ori_bias_addr, output_channel * sizeof(i… in InitWeightBias()
Dconv2d_int8_coder.cc88 int32_t output_channel = filter_tensor_->Batch(); in InitWeightBias() local
92 conv_param_->output_channel_ = output_channel; in InitWeightBias()
93 auto output_channel_size = static_cast<size_t>(output_channel); in InitWeightBias()
105 for (int oc = 0; oc < output_channel; oc++) { in InitWeightBias()
113 up_round_oc = UP_ROUND(output_channel, C2NUM); in InitWeightBias()
116 up_round_oc = MSMAX(UP_ROUND(output_channel, C8NUM), UP_ROUND(output_channel, C4NUM)); in InitWeightBias()
119 up_round_oc = UP_ROUND(output_channel, C8NUM); in InitWeightBias()
154 … input_channel, output_channel, input_zp, filter_peroc_, "GetSupportOptFlag()", in InitWeightBias()
158 … input_channel, output_channel, input_zp, filter_peroc_, support_optimize_, in InitWeightBias()
/third_party/mindspore/mindspore/lite/src/runtime/kernel/arm/fp32/
Dconvolution_slidewindow_fp32.cc43 auto output_channel = filter_tensor->Batch(); in Init() local
47 int oc_block_num = UP_DIV(output_channel, oc_tile_); in Init()
191 auto output_channel = filter_tensor->Batch(); in PackWeight() local
194 int oc_block_num = UP_DIV(output_channel, oc_tile_); in PackWeight()
197 PackNHWCToNXHWCXFp32(kernel_h, kernel_w, output_channel, oc_block_num, input_channel, in PackWeight()
204 auto output_channel = filter_tensor->Batch(); in MallocWeightBiasData() local
208 conv_param_->output_channel_ = output_channel; in MallocWeightBiasData()
210 int oc_block_num = UP_DIV(output_channel, oc_tile_); in MallocWeightBiasData()
Dconvolution_1x1_fp32.cc123 auto output_channel = filter_tensor->Batch(); in Init() local
124 int size = input_channel * UP_ROUND(output_channel, col_tile_) * sizeof(float); in Init()
278 auto output_channel = filter_tensor->Batch(); in PackWeight() local
279 if (output_channel < 0) { in PackWeight()
288 output_channel, input_channel); in PackWeight()
291 output_channel, input_channel); in PackWeight()
294 output_channel, input_channel); in PackWeight()
301 auto output_channel = filter_tensor->Batch(); in MallocWeightBiasData() local
302 int size = input_channel * UP_ROUND(output_channel, col_tile_) * sizeof(float); in MallocWeightBiasData()
312 size = UP_ROUND(output_channel, col_tile_) * sizeof(float); in MallocWeightBiasData()
Ddeconvolution_fp32.cc59 auto output_channel = weight_tensor->Channel(); in MallocWeightBiasData() local
62 int output_aligned_size = UP_ROUND(output_channel, C8NUM); in MallocWeightBiasData()
84 auto output_channel = weight_tensor->Channel(); in PackWeight() local
91 input_channel, kernel_w * kernel_h, output_channel); in PackWeight()
94 input_channel, kernel_w * kernel_h, output_channel); in PackWeight()
179 auto output_channel = weight_tensor->Channel(); in Init() local
182 int output_aligned_size = UP_ROUND(output_channel, C8NUM); in Init()
/third_party/mindspore/mindspore/ccsrc/backend/kernel_compiler/cpu/nnacl/intrinsics/sse/
DConvDwFp32Row_sse.c22 size_t output_channel, size_t input_step) { in ConvDwFp32Row() argument
23 size_t out_c16 = DOWN_DIV(output_channel, C16NUM) * C16NUM; in ConvDwFp32Row()
24 size_t out_c8 = DOWN_DIV(output_channel, C8NUM) * C8NUM; in ConvDwFp32Row()
25 size_t out_c4 = DOWN_DIV(output_channel, C4NUM) * C4NUM; in ConvDwFp32Row()
80 for (; out_c < output_channel; out_c++) { in ConvDwFp32Row()
/third_party/mindspore/mindspore/lite/src/runtime/kernel/arm/fp16/
Dconvolution_1x1_fp16.cc86 auto output_channel = weight_tensor->Batch(); in MallocWeightBiasData() local
88 size_t size = input_channel * UP_ROUND(output_channel, col_tile_) * sizeof(float16_t); in MallocWeightBiasData()
101 size = UP_ROUND(output_channel, col_tile_) * sizeof(float16_t); in MallocWeightBiasData()
117 auto output_channel = weight_tensor->Batch(); in PackWeight() local
122 …orFp16(weight_origin, reinterpret_cast<float16_t *>(packed_weight_), input_channel, output_channel, in PackWeight()
126 … reinterpret_cast<float16_t *>(packed_weight_), output_channel, input_channel); in PackWeight()
129 …orFp16(weight_origin, reinterpret_cast<float16_t *>(packed_weight_), input_channel, output_channel, in PackWeight()
154 auto output_channel = weight_tensor->Batch(); in Init() local
155 size_t size = input_channel * UP_ROUND(output_channel, col_tile_) * sizeof(float16_t); in Init()
Ddeconvolution_fp16.cc59 auto output_channel = weight_tensor->Channel(); in PackWeight() local
65 input_channel, kernel_w * kernel_h, output_channel); in PackWeight()
71 auto output_channel = weight_tensor->Channel(); in MallocWeightBiasData() local
74 …size_t weight_pack_size = input_channel * kernel_w * kernel_h * UP_ROUND(output_channel, C8NUM) * … in MallocWeightBiasData()
83 auto bias_size = UP_ROUND(output_channel, C8NUM) * sizeof(float16_t); in MallocWeightBiasData()
89 memset(bias_data_, 0, UP_ROUND(output_channel, C8NUM) * sizeof(float16_t)); in MallocWeightBiasData()
193 auto output_channel = weight_tensor->Channel(); in Init() local
196 …size_t weight_pack_size = input_channel * kernel_w * kernel_h * UP_ROUND(output_channel, C8NUM) * … in Init()
/third_party/mindspore/mindspore/ccsrc/backend/kernel_compiler/cpu/nnacl/fp16/
Dwinograd_transform_fp16.c134 int output_channel = conv_param->output_channel_; in WinogradOutputNHWCTransformFp16() local
135 int oc8 = UP_DIV(output_channel, C8NUM); in WinogradOutputNHWCTransformFp16()
151 int dst_tile_offset = output_channel * (dst_x_s + dst_y_s * output_w); in WinogradOutputNHWCTransformFp16()
154 int r_c = output_channel - j * C8NUM; in WinogradOutputNHWCTransformFp16()
161 func(src_ptr, dst_ptr, bias_ptr, C8NUM, output_w, output_channel, r_w, r_h, r_c); in WinogradOutputNHWCTransformFp16()
174 int output_channel = conv_param->output_channel_; in WinogradOutputNC8HW8TransformFp16() local
175 int oc8 = UP_DIV(output_channel, C8NUM); in WinogradOutputNC8HW8TransformFp16()
194 int r_c = output_channel - j * C8NUM; in WinogradOutputNC8HW8TransformFp16()
Ddeconv_winograd_fp16.c245 int output_channel = conv_param->output_channel_; in PackDeConvWgDataFp16() local
246 int size = conv_param->input_channel_ * output_channel * tmp_kernel_plane; in PackDeConvWgDataFp16()
252 const float16_t *src_ic = nhwc_weight + deconv_param->kernel_plane_ * output_channel * ic; in PackDeConvWgDataFp16()
253 float16_t *dst_ic = current_unit_weight + tmp_kernel_plane * output_channel * ic; in PackDeConvWgDataFp16()
258 …loat16_t *src_hw = src_ic + (src_h_offset * conv_param->kernel_w_ + src_w_offset) * output_channel; in PackDeConvWgDataFp16()
259 float16_t *dst_hw = dst_ic + (uhi * unit->w_size_ + uwi) * output_channel; in PackDeConvWgDataFp16()
260 memcpy(dst_hw, src_hw, output_channel * sizeof(float16_t)); in PackDeConvWgDataFp16()
301 size = conv_param->input_channel_ * output_channel * unit->winograd_.kh_ * unit->winograd_.kw_; in PackDeConvWgDataFp16()
314 … unit->winograd_.kh_, unit->h_size_, output_channel, conv_param->input_channel_, false); in PackDeConvWgDataFp16()
328 for (int oc = 0; oc < output_channel; oc++) { in PackDeConvWgDataFp16()
[all …]
Ddeconv_fp16.c19 …16(const float16_t *src, float16_t *tmp, const float16_t *bias, float16_t *dst, int output_channel, in DeConvPostFp16() argument
25 int oc8 = UP_ROUND(output_channel, C8NUM); in DeConvPostFp16()
82 …PostConvFuncFp16C8(tmp, dst, bias, output_channel, output_plane, conv_param->output_channel_, conv… in DeConvPostFp16()
Dcommon_func_fp16.h27 size_t output_channel, size_t plane_size, size_t stride, ActType act_type);
32 …cFp16C4(const float16_t *c4_out, float16_t *nhwc_out, const float16_t *bias, size_t output_channel,

12