From: @YeFeng_24 Reviewed-by: @hangangqiang,@zhanghaibo5 Signed-off-by: @hangangqiangtags/v1.2.0-rc1
| @@ -93,8 +93,7 @@ int Convolution1x1FP16CPUKernel::InitWeightBias() { | |||||
| MS_LOG(ERROR) << "Conv1x1 Malloc bias_ptr_ error!"; | MS_LOG(ERROR) << "Conv1x1 Malloc bias_ptr_ error!"; | ||||
| return RET_ERROR; | return RET_ERROR; | ||||
| } | } | ||||
| auto bias_tensor = in_tensors_.at(kBiasIndex); | |||||
| if (bias_tensor->data_type() == kNumberTypeFloat16) { | |||||
| if (origin_bias_data_type_ == kNumberTypeFloat16) { | |||||
| memcpy(bias_data_, origin_bias_, output_channel * sizeof(float16_t)); | memcpy(bias_data_, origin_bias_, output_channel * sizeof(float16_t)); | ||||
| } else { | } else { | ||||
| Float32ToFloat16(reinterpret_cast<float *>(origin_bias_), reinterpret_cast<float16_t *>(bias_data_), | Float32ToFloat16(reinterpret_cast<float *>(origin_bias_), reinterpret_cast<float16_t *>(bias_data_), | ||||
| @@ -112,7 +111,7 @@ int Convolution1x1FP16CPUKernel::InitWeightBias() { | |||||
| } | } | ||||
| memset(reinterpret_cast<char *>(weight_ptr_) + down_size, 0, size - down_size); | memset(reinterpret_cast<char *>(weight_ptr_) + down_size, 0, size - down_size); | ||||
| ColMajor2Row8MajorFp16(origin_weight_, weight_ptr_, input_channel, output_channel, | ColMajor2Row8MajorFp16(origin_weight_, weight_ptr_, input_channel, output_channel, | ||||
| weight_tensor->data_type() == kNumberTypeFloat16); | |||||
| origin_weight_data_type_ == kNumberTypeFloat16); | |||||
| return RET_OK; | return RET_OK; | ||||
| } | } | ||||
| @@ -30,10 +30,13 @@ class Convolution1x1FP16CPUKernel : public ConvolutionBaseFP16CPUKernel { | |||||
| public: | public: | ||||
| Convolution1x1FP16CPUKernel(OpParameter *parameter, const std::vector<lite::Tensor *> &inputs, | Convolution1x1FP16CPUKernel(OpParameter *parameter, const std::vector<lite::Tensor *> &inputs, | ||||
| const std::vector<lite::Tensor *> &outputs, const InnerContext *ctx, | const std::vector<lite::Tensor *> &outputs, const InnerContext *ctx, | ||||
| const mindspore::lite::PrimitiveC *primitive, void *origin_weight, void *origin_bias) | |||||
| const mindspore::lite::PrimitiveC *primitive, void *origin_weight, void *origin_bias, | |||||
| TypeId origin_weight_data_type, TypeId origin_bias_data_type) | |||||
| : ConvolutionBaseFP16CPUKernel(parameter, inputs, outputs, ctx, primitive), | : ConvolutionBaseFP16CPUKernel(parameter, inputs, outputs, ctx, primitive), | ||||
| origin_weight_(origin_weight), | origin_weight_(origin_weight), | ||||
| origin_bias_(origin_bias) {} | |||||
| origin_bias_(origin_bias), | |||||
| origin_weight_data_type_(origin_weight_data_type), | |||||
| origin_bias_data_type_(origin_bias_data_type) {} | |||||
| ~Convolution1x1FP16CPUKernel() override; | ~Convolution1x1FP16CPUKernel() override; | ||||
| int Init() override; | int Init() override; | ||||
| @@ -62,6 +65,8 @@ class Convolution1x1FP16CPUKernel : public ConvolutionBaseFP16CPUKernel { | |||||
| float16_t *pack_input_ = nullptr; | float16_t *pack_input_ = nullptr; | ||||
| float16_t *output_ptr_ = nullptr; | float16_t *output_ptr_ = nullptr; | ||||
| MatMulParameter *matmul_param_ = nullptr; | MatMulParameter *matmul_param_ = nullptr; | ||||
| TypeId origin_weight_data_type_; | |||||
| TypeId origin_bias_data_type_; | |||||
| }; | }; | ||||
| } // namespace mindspore::kernel | } // namespace mindspore::kernel | ||||
| @@ -67,15 +67,17 @@ int ConvolutionDelegateFP16CPUKernel::Init() { | |||||
| if (in_tensors_.size() == 3) { | if (in_tensors_.size() == 3) { | ||||
| origin_bias_ = CopyData(in_tensors_.at(kBiasIndex)); | origin_bias_ = CopyData(in_tensors_.at(kBiasIndex)); | ||||
| need_free_ = need_free_ | BIAS_NEED_FREE; | need_free_ = need_free_ | BIAS_NEED_FREE; | ||||
| origin_bias_data_type_ = in_tensors_.at(kBiasIndex)->data_type(); | |||||
| } | } | ||||
| origin_weight_data_type_ = in_tensors_[1]->data_type(); | |||||
| origin_weight_data_type_ = in_tensors_.at(kWeightIndex)->data_type(); | |||||
| return RET_OK; | return RET_OK; | ||||
| } | } | ||||
| origin_weight_ = in_tensors_.at(kWeightIndex)->data_c(); | origin_weight_ = in_tensors_.at(kWeightIndex)->data_c(); | ||||
| if (in_tensors_.size() == 3) { | if (in_tensors_.size() == 3) { | ||||
| origin_bias_ = in_tensors_.at(kBiasIndex)->data_c(); | origin_bias_ = in_tensors_.at(kBiasIndex)->data_c(); | ||||
| origin_bias_data_type_ = in_tensors_.at(kBiasIndex)->data_type(); | |||||
| } | } | ||||
| origin_weight_data_type_ = in_tensors_[1]->data_type(); | |||||
| origin_weight_data_type_ = in_tensors_.at(kWeightIndex)->data_type(); | |||||
| return ReSize(); | return ReSize(); | ||||
| } | } | ||||
| @@ -84,8 +86,9 @@ int ConvolutionDelegateFP16CPUKernel::ReSize() { | |||||
| SetInputOutputShapeInfo(reinterpret_cast<ConvParameter *>(op_parameter_), in_tensors_.front(), out_tensors_.front(), | SetInputOutputShapeInfo(reinterpret_cast<ConvParameter *>(op_parameter_), in_tensors_.front(), out_tensors_.front(), | ||||
| context_); | context_); | ||||
| if (fp16_conv_kernel_ == nullptr) { | if (fp16_conv_kernel_ == nullptr) { | ||||
| fp16_conv_kernel_ = CpuConvFp16KernelSelect(in_tensors_, out_tensors_, op_parameter_, context_, primitive_, | |||||
| origin_weight_, origin_bias_, origin_weight_data_type_); | |||||
| fp16_conv_kernel_ = | |||||
| CpuConvFp16KernelSelect(in_tensors_, out_tensors_, op_parameter_, context_, primitive_, origin_weight_, | |||||
| origin_bias_, origin_weight_data_type_, origin_bias_data_type_); | |||||
| if (fp16_conv_kernel_ == nullptr) { | if (fp16_conv_kernel_ == nullptr) { | ||||
| MS_LOG(ERROR) << "Selecting execute kernel failed for conv_kernel, got a nullptr."; | MS_LOG(ERROR) << "Selecting execute kernel failed for conv_kernel, got a nullptr."; | ||||
| return RET_ERROR; | return RET_ERROR; | ||||
| @@ -109,7 +112,8 @@ ConvParameter *CreateNewConvParameterFp16(ConvParameter *parameter) { | |||||
| kernel::LiteKernel *CpuConvFp16KernelSelect(const std::vector<lite::Tensor *> &inputs, | kernel::LiteKernel *CpuConvFp16KernelSelect(const std::vector<lite::Tensor *> &inputs, | ||||
| const std::vector<lite::Tensor *> &outputs, OpParameter *op_parameter, | const std::vector<lite::Tensor *> &outputs, OpParameter *op_parameter, | ||||
| const lite::InnerContext *ctx, const mindspore::lite::PrimitiveC *primitive, | const lite::InnerContext *ctx, const mindspore::lite::PrimitiveC *primitive, | ||||
| void *origin_weight, void *origin_bias, TypeId origin_weight_data_type) { | |||||
| void *origin_weight, void *origin_bias, TypeId origin_weight_data_type, | |||||
| TypeId origin_bias_data_type) { | |||||
| auto conv_param = reinterpret_cast<ConvParameter *>(op_parameter); | auto conv_param = reinterpret_cast<ConvParameter *>(op_parameter); | ||||
| bool use_winograd = false; | bool use_winograd = false; | ||||
| int out_unit; | int out_unit; | ||||
| @@ -117,13 +121,15 @@ kernel::LiteKernel *CpuConvFp16KernelSelect(const std::vector<lite::Tensor *> &i | |||||
| kernel::LiteKernel *kernel = nullptr; | kernel::LiteKernel *kernel = nullptr; | ||||
| if (conv_param->kernel_h_ == 1 && conv_param->kernel_w_ == 1) { | if (conv_param->kernel_h_ == 1 && conv_param->kernel_w_ == 1) { | ||||
| kernel = new (std::nothrow) | kernel = new (std::nothrow) | ||||
| kernel::Convolution1x1FP16CPUKernel(op_parameter, inputs, outputs, ctx, primitive, origin_weight, origin_bias); | |||||
| kernel::Convolution1x1FP16CPUKernel(op_parameter, inputs, outputs, ctx, primitive, origin_weight, origin_bias, | |||||
| origin_weight_data_type, origin_bias_data_type); | |||||
| } else if (use_winograd) { | } else if (use_winograd) { | ||||
| kernel = new (std::nothrow) kernel::ConvolutionWinogradFP16CPUKernel(op_parameter, inputs, outputs, ctx, primitive, | |||||
| out_unit, origin_weight, origin_bias); | |||||
| kernel = new (std::nothrow) kernel::ConvolutionWinogradFP16CPUKernel( | |||||
| op_parameter, inputs, outputs, ctx, primitive, out_unit, origin_weight, origin_bias, origin_bias_data_type); | |||||
| } else { | } else { | ||||
| kernel = new (std::nothrow) kernel::ConvolutionFP16CPUKernel(op_parameter, inputs, outputs, ctx, primitive, | |||||
| origin_weight, origin_bias, origin_weight_data_type); | |||||
| kernel = | |||||
| new (std::nothrow) kernel::ConvolutionFP16CPUKernel(op_parameter, inputs, outputs, ctx, primitive, origin_weight, | |||||
| origin_bias, origin_weight_data_type, origin_bias_data_type); | |||||
| } | } | ||||
| // Once kernel is selected, init func will invoke InitWeightAndBias | // Once kernel is selected, init func will invoke InitWeightAndBias | ||||
| auto ret = kernel->Init(); | auto ret = kernel->Init(); | ||||
| @@ -55,12 +55,14 @@ class ConvolutionDelegateFP16CPUKernel : public LiteKernel { | |||||
| void *origin_bias_ = nullptr; | void *origin_bias_ = nullptr; | ||||
| kernel::LiteKernel *fp16_conv_kernel_ = nullptr; | kernel::LiteKernel *fp16_conv_kernel_ = nullptr; | ||||
| TypeId origin_weight_data_type_; | TypeId origin_weight_data_type_; | ||||
| TypeId origin_bias_data_type_; | |||||
| }; | }; | ||||
| kernel::LiteKernel *CpuConvFp16KernelSelect(const std::vector<lite::Tensor *> &inputs, | kernel::LiteKernel *CpuConvFp16KernelSelect(const std::vector<lite::Tensor *> &inputs, | ||||
| const std::vector<lite::Tensor *> &outputs, OpParameter *op_parameter, | const std::vector<lite::Tensor *> &outputs, OpParameter *op_parameter, | ||||
| const lite::InnerContext *ctx, const mindspore::lite::PrimitiveC *primitive, | const lite::InnerContext *ctx, const mindspore::lite::PrimitiveC *primitive, | ||||
| void *origin_weight, void *origin_bias, TypeId origin_weight_data_type); | |||||
| void *origin_weight, void *origin_bias, TypeId origin_weight_data_type, | |||||
| TypeId origin_bias_data_type); | |||||
| } // namespace mindspore::kernel | } // namespace mindspore::kernel | ||||
| #endif // MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_FP16_CONVOLUTION_DELEGATE_FP16_H_ | #endif // MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_FP16_CONVOLUTION_DELEGATE_FP16_H_ | ||||
| @@ -62,8 +62,7 @@ int ConvolutionFP16CPUKernel::InitWeightBias() { | |||||
| } | } | ||||
| memset(bias_data_, 0, oc8 * sizeof(float16_t)); | memset(bias_data_, 0, oc8 * sizeof(float16_t)); | ||||
| if (in_tensors_.size() == kInputSize2) { | if (in_tensors_.size() == kInputSize2) { | ||||
| auto bias_tensor = in_tensors_.at(kBiasIndex); | |||||
| if (bias_tensor->data_type() == kNumberTypeFloat16) { | |||||
| if (origin_bias_data_type_ == kNumberTypeFloat16) { | |||||
| memcpy(bias_data_, origin_bias_, out_channel * sizeof(float16_t)); | memcpy(bias_data_, origin_bias_, out_channel * sizeof(float16_t)); | ||||
| } else { | } else { | ||||
| Float32ToFloat16(reinterpret_cast<float *>(origin_bias_), reinterpret_cast<float16_t *>(bias_data_), out_channel); | Float32ToFloat16(reinterpret_cast<float *>(origin_bias_), reinterpret_cast<float16_t *>(bias_data_), out_channel); | ||||
| @@ -28,11 +28,12 @@ class ConvolutionFP16CPUKernel : public ConvolutionBaseFP16CPUKernel { | |||||
| ConvolutionFP16CPUKernel(OpParameter *parameter, const std::vector<lite::Tensor *> &inputs, | ConvolutionFP16CPUKernel(OpParameter *parameter, const std::vector<lite::Tensor *> &inputs, | ||||
| const std::vector<lite::Tensor *> &outputs, const InnerContext *ctx, | const std::vector<lite::Tensor *> &outputs, const InnerContext *ctx, | ||||
| const mindspore::lite::PrimitiveC *primitive, void *origin_weight, void *origin_bias, | const mindspore::lite::PrimitiveC *primitive, void *origin_weight, void *origin_bias, | ||||
| TypeId origin_weight_data_type) | |||||
| TypeId origin_weight_data_type, TypeId origin_bias_data_type) | |||||
| : ConvolutionBaseFP16CPUKernel(parameter, inputs, outputs, ctx, primitive), | : ConvolutionBaseFP16CPUKernel(parameter, inputs, outputs, ctx, primitive), | ||||
| origin_weight_(origin_weight), | origin_weight_(origin_weight), | ||||
| origin_bias_(origin_bias), | origin_bias_(origin_bias), | ||||
| origin_weight_data_type_(origin_weight_data_type) {} | |||||
| origin_weight_data_type_(origin_weight_data_type), | |||||
| origin_bias_data_type_(origin_bias_data_type) {} | |||||
| ~ConvolutionFP16CPUKernel() override { | ~ConvolutionFP16CPUKernel() override { | ||||
| if (packed_weight_ != nullptr) { | if (packed_weight_ != nullptr) { | ||||
| free(packed_weight_); | free(packed_weight_); | ||||
| @@ -65,6 +66,7 @@ class ConvolutionFP16CPUKernel : public ConvolutionBaseFP16CPUKernel { | |||||
| float16_t *packed_weight_ = nullptr; | float16_t *packed_weight_ = nullptr; | ||||
| float16_t *col_major_input_ = nullptr; | float16_t *col_major_input_ = nullptr; | ||||
| TypeId origin_weight_data_type_; | TypeId origin_weight_data_type_; | ||||
| TypeId origin_bias_data_type_; | |||||
| }; | }; | ||||
| } // namespace mindspore::kernel | } // namespace mindspore::kernel | ||||
| @@ -93,8 +93,7 @@ int ConvolutionWinogradFP16CPUKernel::InitWeightBias() { | |||||
| memset(bias_data_, 0, oc_block_num * oc_block * sizeof(float16_t)); | memset(bias_data_, 0, oc_block_num * oc_block * sizeof(float16_t)); | ||||
| if (in_tensors_.size() == kInputSize2) { | if (in_tensors_.size() == kInputSize2) { | ||||
| auto bias_tensor = in_tensors_.at(kBiasIndex); | |||||
| if (bias_tensor->data_type() == kNumberTypeFloat16) { | |||||
| if (origin_bias_data_type_ == kNumberTypeFloat16) { | |||||
| memcpy(bias_data_, origin_bias_, out_channel * sizeof(float16_t)); | memcpy(bias_data_, origin_bias_, out_channel * sizeof(float16_t)); | ||||
| } else { | } else { | ||||
| Float32ToFloat16(reinterpret_cast<float *>(origin_bias_), reinterpret_cast<float16_t *>(bias_data_), out_channel); | Float32ToFloat16(reinterpret_cast<float *>(origin_bias_), reinterpret_cast<float16_t *>(bias_data_), out_channel); | ||||
| @@ -32,11 +32,12 @@ class ConvolutionWinogradFP16CPUKernel : public ConvolutionBaseFP16CPUKernel { | |||||
| ConvolutionWinogradFP16CPUKernel(OpParameter *parameter, const std::vector<lite::Tensor *> &inputs, | ConvolutionWinogradFP16CPUKernel(OpParameter *parameter, const std::vector<lite::Tensor *> &inputs, | ||||
| const std::vector<lite::Tensor *> &outputs, const InnerContext *ctx, | const std::vector<lite::Tensor *> &outputs, const InnerContext *ctx, | ||||
| const mindspore::lite::PrimitiveC *primitive, int out_unit, void *origin_weight, | const mindspore::lite::PrimitiveC *primitive, int out_unit, void *origin_weight, | ||||
| void *origin_bias) | |||||
| void *origin_bias, TypeId origin_bias_data_type) | |||||
| : ConvolutionBaseFP16CPUKernel(parameter, inputs, outputs, ctx, primitive), | : ConvolutionBaseFP16CPUKernel(parameter, inputs, outputs, ctx, primitive), | ||||
| output_unit_(out_unit), | output_unit_(out_unit), | ||||
| origin_weight_(origin_weight), | origin_weight_(origin_weight), | ||||
| origin_bias_(origin_bias) {} | |||||
| origin_bias_(origin_bias), | |||||
| origin_bias_data_type_(origin_bias_data_type) {} | |||||
| ~ConvolutionWinogradFP16CPUKernel() override { | ~ConvolutionWinogradFP16CPUKernel() override { | ||||
| if (trans_weight_ != nullptr) { | if (trans_weight_ != nullptr) { | ||||
| free(trans_weight_); | free(trans_weight_); | ||||
| @@ -86,6 +87,7 @@ class ConvolutionWinogradFP16CPUKernel : public ConvolutionBaseFP16CPUKernel { | |||||
| TmpBufferAddressFp16 tmp_buffer_address_list_[4]; | TmpBufferAddressFp16 tmp_buffer_address_list_[4]; | ||||
| InputTransFp16Func in_func_; | InputTransFp16Func in_func_; | ||||
| OutputTransFp16Func out_func_; | OutputTransFp16Func out_func_; | ||||
| TypeId origin_bias_data_type_; | |||||
| }; | }; | ||||
| } // namespace mindspore::kernel | } // namespace mindspore::kernel | ||||