| 976 | } |
| 977 | |
| 978 | void tensor_conv:: |
| 979 | choose_backward_data_algorithm ( |
| 980 | const tensor& gradient_input, |
| 981 | const tensor& data_gradient |
| 982 | ) |
| 983 | { |
| 984 | if (backward_data_algo >= 0) |
| 985 | return; |
| 986 | |
| 987 | const auto find = [&](bool& benchmarked) { |
| 988 | cudnnConvolutionBwdDataAlgo_t backward_data_best_algo; |
| 989 | #if CUDNN_MAJOR >= 8 |
| 990 | int num_possible_algorithms = 0; |
| 991 | CHECK_CUDNN(cudnnGetConvolutionBackwardFilterAlgorithmMaxCount(context(), &num_possible_algorithms)); |
| 992 | std::vector<cudnnConvolutionBwdDataAlgoPerf_t> perf_results(num_possible_algorithms); |
| 993 | int num_algorithms = 0; |
| 994 | cudnnStatus_t status = cudnnFindConvolutionBackwardDataAlgorithm( |
| 995 | context(), |
| 996 | (const cudnnFilterDescriptor_t)filter_handle, |
| 997 | descriptor(gradient_input), |
| 998 | (const cudnnConvolutionDescriptor_t)conv_handle, |
| 999 | descriptor(data_gradient), |
| 1000 | num_possible_algorithms, |
| 1001 | &num_algorithms, |
| 1002 | perf_results.data()); |
| 1003 | // Benchmarking needs memory that the GPU may not have to spare at the moment, so |
| 1004 | // fall back to the heuristics of cuDNN then. |
| 1005 | if (is_out_of_memory(status)) |
| 1006 | { |
| 1007 | benchmarked = false; |
| 1008 | status = cudnnGetConvolutionBackwardDataAlgorithm_v7( |
| 1009 | context(), |
| 1010 | (const cudnnFilterDescriptor_t)filter_handle, |
| 1011 | descriptor(gradient_input), |
| 1012 | (const cudnnConvolutionDescriptor_t)conv_handle, |
| 1013 | descriptor(data_gradient), |
| 1014 | num_possible_algorithms, |
| 1015 | &num_algorithms, |
| 1016 | perf_results.data()); |
| 1017 | } |
| 1018 | CHECK_CUDNN(status); |
| 1019 | perf_results.resize(num_algorithms); |
| 1020 | backward_data_best_algo = pick_best_algorithm(perf_results); |
| 1021 | #else |
| 1022 | CHECK_CUDNN(cudnnGetConvolutionBackwardDataAlgorithm( |
| 1023 | context(), |
| 1024 | (const cudnnFilterDescriptor_t)filter_handle, |
| 1025 | descriptor(gradient_input), |
| 1026 | (const cudnnConvolutionDescriptor_t)conv_handle, |
| 1027 | descriptor(data_gradient), |
| 1028 | dnn_prefer_fastest_algorithms()?CUDNN_CONVOLUTION_BWD_DATA_PREFER_FASTEST:CUDNN_CONVOLUTION_BWD_DATA_NO_WORKSPACE, |
| 1029 | std::numeric_limits<size_t>::max(), |
| 1030 | &backward_data_best_algo)); |
| 1031 | #endif |
| 1032 | return static_cast<int>(backward_data_best_algo); |
| 1033 | }; |
| 1034 | |
| 1035 | const auto get_workspace_size = [&](int algo) { |
nothing calls this directly
no test coverage detected