| 903 | } |
| 904 | |
| 905 | void tensor_conv:: |
| 906 | choose_forward_algorithm ( |
| 907 | const tensor& data, |
| 908 | const tensor& output |
| 909 | ) |
| 910 | { |
| 911 | if (forward_algo >= 0) |
| 912 | return; |
| 913 | |
| 914 | const auto find = [&](bool& benchmarked) { |
| 915 | cudnnConvolutionFwdAlgo_t forward_best_algo; |
| 916 | #if CUDNN_MAJOR >= 8 |
| 917 | int num_possible_algorithms = 0; |
| 918 | CHECK_CUDNN(cudnnGetConvolutionForwardAlgorithmMaxCount(context(), &num_possible_algorithms)); |
| 919 | std::vector<cudnnConvolutionFwdAlgoPerf_t> perf_results(num_possible_algorithms); |
| 920 | int num_algorithms = 0; |
| 921 | cudnnStatus_t status = cudnnFindConvolutionForwardAlgorithm( |
| 922 | context(), |
| 923 | descriptor(data), |
| 924 | (const cudnnFilterDescriptor_t)filter_handle, |
| 925 | (const cudnnConvolutionDescriptor_t)conv_handle, |
| 926 | descriptor(output), |
| 927 | num_possible_algorithms, |
| 928 | &num_algorithms, |
| 929 | perf_results.data()); |
| 930 | // Benchmarking needs memory that the GPU may not have to spare at the moment, so |
| 931 | // fall back to the heuristics of cuDNN then. |
| 932 | if (is_out_of_memory(status)) |
| 933 | { |
| 934 | benchmarked = false; |
| 935 | status = cudnnGetConvolutionForwardAlgorithm_v7( |
| 936 | context(), |
| 937 | descriptor(data), |
| 938 | (const cudnnFilterDescriptor_t)filter_handle, |
| 939 | (const cudnnConvolutionDescriptor_t)conv_handle, |
| 940 | descriptor(output), |
| 941 | num_possible_algorithms, |
| 942 | &num_algorithms, |
| 943 | perf_results.data()); |
| 944 | } |
| 945 | CHECK_CUDNN(status); |
| 946 | perf_results.resize(num_algorithms); |
| 947 | forward_best_algo = pick_best_algorithm(perf_results); |
| 948 | #else |
| 949 | CHECK_CUDNN(cudnnGetConvolutionForwardAlgorithm( |
| 950 | context(), |
| 951 | descriptor(data), |
| 952 | (const cudnnFilterDescriptor_t)filter_handle, |
| 953 | (const cudnnConvolutionDescriptor_t)conv_handle, |
| 954 | descriptor(output), |
| 955 | dnn_prefer_fastest_algorithms()?CUDNN_CONVOLUTION_FWD_PREFER_FASTEST:CUDNN_CONVOLUTION_FWD_NO_WORKSPACE, |
| 956 | std::numeric_limits<size_t>::max(), |
| 957 | &forward_best_algo)); |
| 958 | #endif |
| 959 | return static_cast<int>(forward_best_algo); |
| 960 | }; |
| 961 | |
| 962 | const auto get_workspace_size = [&](int algo) { |
nothing calls this directly
no test coverage detected