| // The MIT License (MIT) |
| // |
| // Copyright (c) 2016 Northeastern University |
| // |
| // Permission is hereby granted, free of charge, to any person obtaining a copy |
| // of this software and associated documentation files (the "Software"), to deal |
| // in the Software without restriction, including without limitation the rights |
| // to use, copy, modify, merge, publish, distribute, sublicense, and/or sell |
| // copies of the Software, and to permit persons to whom the Software is |
| // furnished to do so, subject to the following conditions: |
| // |
| // The above copyright notice and this permission notice shall be included in |
| // all copies or substantial portions of the Software. |
| // |
| // THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR |
| // IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, |
| // FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE |
| // AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER |
| // LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, |
| // OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE |
| // SOFTWARE. |
| |
| #ifndef CORE_INCLUDE_LAYERS_CONV_LAYER_H_ |
| #define CORE_INCLUDE_LAYERS_CONV_LAYER_H_ |
| |
| #include "dnn_layer.h" |
| #include <iostream> |
| |
| namespace dnnmark { |
| |
| template <typename T> |
| class ConvolutionLayer : public Layer<T> { |
| // using declaration for calling member from base class |
| using Layer<T>::p_dnnmark_; |
| using Layer<T>::layer_id_; |
| using Layer<T>::previous_layer_name_; |
| using Layer<T>::input_dim_; |
| using Layer<T>::output_dim_; |
| using Layer<T>::bottom_desc_; |
| using Layer<T>::top_desc_; |
| using Layer<T>::data_manager_; |
| |
| using Layer<T>::num_bottoms_; |
| using Layer<T>::bottoms_; |
| using Layer<T>::bottom_chunk_ids_; |
| using Layer<T>::bottom_diffs_; |
| using Layer<T>::bottom_diff_chunk_ids_; |
| |
| using Layer<T>::num_tops_; |
| using Layer<T>::tops_; |
| using Layer<T>::top_chunk_ids_; |
| using Layer<T>::top_diffs_; |
| using Layer<T>::top_diff_chunk_ids_; |
| |
| private: |
| ConvolutionParam conv_param_; |
| |
| // Convolution specific descriptor |
| ConvolutionDesc<T> desc_; |
| |
| // Layer weights |
| Data<T> *weights_; |
| int weights_chunk_id_; |
| Data<T> *weights_diff_; |
| int weights_diff_chunk_id_; |
| |
| // Algorithm specific parameters |
| ConvAlgo<T> conv_algo_; |
| size_t fwd_workspace_size_; |
| size_t bwd_data_workspace_size_; |
| size_t bwd_filter_workspace_size_; |
| Data<T> *fwd_workspace_; |
| int fwd_workspace_id_; |
| Data<T> *bwd_data_workspace_; |
| int bwd_data_workspace_id_; |
| Data<T> *bwd_filter_workspace_; |
| int bwd_filter_workspace_id_; |
| bool has_fwd_workspace_; |
| bool has_bwd_data_workspace_; |
| bool has_bwd_filter_workspace_; |
| public: |
| ConvolutionLayer(DNNMark<T> *p_dnnmark) |
| : Layer<T>(p_dnnmark), |
| conv_param_(), desc_(), conv_algo_() { |
| Layer<T>::has_learnable_params_ = true; |
| fwd_workspace_size_ = 0; |
| bwd_data_workspace_size_ = 0; |
| bwd_filter_workspace_size_ = 0; |
| has_fwd_workspace_ = false; |
| has_bwd_data_workspace_ = false; |
| has_bwd_filter_workspace_ = false; |
| } |
| |
| ~ConvolutionLayer() { |
| // Free the workspace |
| if (has_fwd_workspace_) { |
| data_manager_->RemoveData(fwd_workspace_id_); |
| has_fwd_workspace_ = false; |
| } |
| if (has_bwd_data_workspace_) { |
| data_manager_->RemoveData(bwd_data_workspace_id_); |
| has_bwd_data_workspace_ = false; |
| } |
| if (has_bwd_filter_workspace_) { |
| data_manager_->RemoveData(bwd_filter_workspace_id_); |
| has_bwd_filter_workspace_ = false; |
| } |
| } |
| |
| ConvolutionParam *getConvParam() { return &conv_param_; } |
| |
| void Setup() { |
| // Set up indispensable stuff here |
| Layer<T>::Setup(); |
| |
| // Set convolution related descriptors |
| desc_.Set(conv_param_, input_dim_.c_); |
| |
| // Set up convolution related data |
| if (input_dim_.n_ != 0 && input_dim_.c_ != 0 && |
| input_dim_.h_ != 0 && input_dim_.w_ != 0) { |
| // |
| // Standalone mode |
| // |
| |
| // Compute dimension of output data |
| ComputeOutputDim(); |
| |
| // Set top tensor |
| top_desc_.Set(output_dim_.n_, |
| output_dim_.c_, |
| output_dim_.h_, |
| output_dim_.w_); |
| |
| // Prepare top data |
| int top_size = output_dim_.n_ * |
| output_dim_.c_ * |
| output_dim_.h_ * |
| output_dim_.w_; |
| for (int i = 0; i < num_tops_; i++) { |
| top_chunk_ids_.push_back( |
| data_manager_->CreateData(top_size)); |
| tops_.push_back( |
| data_manager_->GetData(top_chunk_ids_[i])); |
| top_diff_chunk_ids_.push_back( |
| data_manager_->CreateData(top_size)); |
| top_diffs_.push_back( |
| data_manager_->GetData(top_diff_chunk_ids_[i])); |
| } |
| |
| } |
| |
| // Only one set of weights is considered |
| |
| int weights_size = conv_param_.output_num_ * |
| input_dim_.c_ * |
| conv_param_.kernel_size_h_ * |
| conv_param_.kernel_size_w_; |
| weights_chunk_id_ = data_manager_->CreateData(weights_size); |
| weights_ = data_manager_->GetData(weights_chunk_id_); |
| weights_diff_chunk_id_ = |
| data_manager_->CreateData(weights_size); |
| weights_diff_ = data_manager_->GetData(weights_diff_chunk_id_); |
| |
| // Fill the weight data |
| weights_->Filler(); |
| |
| // Set convolution forward algorithm |
| // Use default algorithm for now |
| conv_algo_.SetFwdAlgo(conv_param_.algo_); |
| |
| // Allocate workspace |
| conv_algo_.GetFwdWorkspaceSize(*(p_dnnmark_->GetHandle()), |
| p_dnnmark_->getRunMode(), layer_id_, |
| bottom_desc_, |
| top_desc_, |
| desc_, |
| &fwd_workspace_size_); |
| if (fwd_workspace_size_ > 0) { |
| fwd_workspace_id_ = data_manager_->CreateData(fwd_workspace_size_); |
| fwd_workspace_ = data_manager_->GetData(fwd_workspace_id_); |
| has_fwd_workspace_ = true; |
| } |
| |
| #ifdef NVIDIA_CUDNN |
| // Set convolution backward filter/weights algorithm |
| if (!conv_param_.algo_.compare("cudnn")) { |
| // Chainer default behaviour |
| // Use cuDNN function cudnnGetConvolutionBackwardFilterAlgorithm |
| conv_algo_.SetBwdFilterAlgo(*(p_dnnmark_->GetHandle()), |
| p_dnnmark_->getRunMode(), layer_id_, |
| bottom_desc_, |
| top_desc_, |
| desc_, |
| conv_param_.conv_bwd_filter_pref_); |
| LOG(INFO) << "Set cuDNN recommended conv. bwd filter alg. to " << conv_algo_.GetBwdFilterAlgo(); |
| std::cout << "cuDNN recommended bwd convolution filter algorithm:"<<conv_algo_.GetBwdFilterAlgo()<<"\n"; |
| } else if (conv_param_.algo_ == "auto" ) { |
| // Query cuDNN for the fastest BWD convolution filter gradient algorithm. |
| // Use cuDNN function cudnnFindConvolutionBackwardFilterAlgorithm (called inside FindBwdFilterAlgo()) |
| |
| // NOTE: The below code selects algorithms prior to run, during setup phase. |
| // FindBwdFilterAlgoEx must be called during run phase through dnn_wrapper. |
| //conv_algo_.SetBwdFilterAlgo("autoex"); |
| conv_algo_.FindBwdFilterAlgo(*(p_dnnmark_->GetHandle()), |
| p_dnnmark_->getRunMode(), layer_id_, |
| bottom_desc_, |
| desc_, |
| top_desc_); |
| LOG(INFO) << "cuDNN fastest bwd conv. filter algo.:" << conv_algo_.GetBwdFilterAlgo(); |
| std::cout << "cuDNN fastest bwd conv. filter algorithm:"<<conv_algo_.GetBwdFilterAlgo()<<"\n"; |
| } else { |
| // Use default algorithm for now |
| LOG(INFO) << "Setting Bwd Filter Algo to " << conv_param_.algo_; |
| conv_algo_.SetBwdFilterAlgo(conv_param_.algo_); |
| } |
| #endif |
| #ifdef AMD_MIOPEN |
| // Use default algorithm for now |
| LOG(INFO) << "Setting Bwd Filter Algo to " << conv_param_.algo_; |
| conv_algo_.SetBwdFilterAlgo(conv_param_.algo_); |
| #endif |
| |
| |
| // Allocate workspace |
| conv_algo_.GetBwdFilterWorkspaceSize(*(p_dnnmark_->GetHandle()), |
| p_dnnmark_->getRunMode(), layer_id_, |
| bottom_desc_, |
| top_desc_, |
| desc_, |
| &bwd_filter_workspace_size_); |
| if (bwd_filter_workspace_size_ > 0) { |
| bwd_filter_workspace_id_ = data_manager_-> |
| CreateData(bwd_filter_workspace_size_); |
| bwd_filter_workspace_ = data_manager_->GetData(bwd_filter_workspace_id_); |
| has_bwd_filter_workspace_ = true; |
| } |
| |
| // Set convolution backward data algorithm |
| // Use default algorithm for now |
| conv_algo_.SetBwdDataAlgo(conv_param_.algod_); |
| #ifdef NVIDIA_CUDNN |
| LOG(INFO) << "BWD conv. data algo set to:"<< static_cast<int>(conv_algo_.getDataAlgo()); |
| // std::cout << "cuDNN recommended BWD convolution data algorithm:"<<conv_algo_.GetBwdDataAlgo()<<"\n"; |
| #endif |
| |
| // Allocate workspace |
| conv_algo_.GetBwdDataWorkspaceSize(*(p_dnnmark_->GetHandle()), |
| p_dnnmark_->getRunMode(), layer_id_, |
| bottom_desc_, |
| top_desc_, |
| desc_, |
| &bwd_data_workspace_size_); |
| if (bwd_data_workspace_size_ > 0) { |
| bwd_data_workspace_id_ = data_manager_-> |
| CreateData(bwd_data_workspace_size_); |
| bwd_data_workspace_ = data_manager_->GetData(bwd_data_workspace_id_); |
| has_bwd_data_workspace_ = true; |
| } |
| |
| } |
| |
| void ComputeOutputDim() { |
| output_dim_.n_ = input_dim_.n_; |
| output_dim_.c_ = conv_param_.output_num_; |
| output_dim_.h_ = (input_dim_.h_ + |
| 2 * conv_param_.pad_h_ - conv_param_.kernel_size_h_) / |
| conv_param_.stride_u_ + 1; |
| output_dim_.w_ = (input_dim_.w_ + |
| 2 * conv_param_.pad_w_ - conv_param_.kernel_size_w_) / |
| conv_param_.stride_v_ + 1; |
| } |
| |
| void ForwardPropagation() { |
| // Fill the bottom data |
| if (p_dnnmark_->getRunMode() == STANDALONE || |
| !previous_layer_name_.compare("null")) { |
| for (int i = 0; i < num_bottoms_; i++) { |
| bottoms_[i]->Filler(); |
| } |
| } |
| // Convolution forward computation |
| for (int i = 0; i < num_bottoms_; i++) { |
| dnnmarkConvolutionForward( |
| *(p_dnnmark_->GetHandle()), |
| p_dnnmark_->getRunMode(), layer_id_, |
| p_dnnmark_->GetTimer(), |
| DataType<T>::one, |
| bottom_desc_, bottoms_[i]->Get(), |
| desc_, weights_->Get(), |
| &conv_algo_, |
| has_fwd_workspace_? fwd_workspace_->Get() : nullptr, |
| fwd_workspace_size_, |
| DataType<T>::zero, |
| top_desc_, tops_[i]->Get()); |
| } |
| } |
| |
| void BackwardPropagation() { |
| if (p_dnnmark_->getRunMode() == STANDALONE || |
| !previous_layer_name_.compare("null")) { |
| // Fill the top data and top diff data |
| for (int i = 0; i < num_tops_; i++) { |
| tops_[i]->Filler(); |
| top_diffs_[i]->Filler(); |
| } |
| // Fill the bottom data |
| for (int i = 0; i < num_bottoms_; i++) { |
| bottoms_[i]->Filler(); |
| } |
| } |
| |
| // Convolution forward computation |
| for (int i = 0; i < num_tops_; i++) { |
| dnnmarkConvolutionBackwardFilter( |
| *(p_dnnmark_->GetHandle()), |
| p_dnnmark_->getRunMode(), layer_id_, |
| p_dnnmark_->GetTimer(), |
| DataType<T>::one, |
| bottom_desc_, bottoms_[i]->Get(), |
| top_desc_, top_diffs_[i]->Get(), |
| desc_, |
| &conv_algo_, |
| has_bwd_filter_workspace_? bwd_filter_workspace_->Get() : nullptr, |
| bwd_filter_workspace_size_, |
| DataType<T>::zero, |
| weights_diff_->Get()); |
| if (conv_param_.propagation_) { |
| dnnmarkConvolutionBackwardData( |
| *(p_dnnmark_->GetHandle()), |
| p_dnnmark_->getRunMode(), layer_id_, |
| p_dnnmark_->GetTimer(), |
| DataType<T>::one, |
| top_desc_, top_diffs_[i]->Get(), |
| desc_, weights_->Get(), |
| &conv_algo_, |
| has_bwd_data_workspace_? bwd_data_workspace_->Get() : nullptr, |
| bwd_data_workspace_size_, |
| DataType<T>::zero, |
| bottom_desc_, bottoms_[i]->Get()); |
| } |
| } |
| } |
| |
| }; |
| |
| } // namespace dnnmark |
| |
| #endif // CORE_INCLUDE_LAYERS_CONV_LAYER_H_ |