#ifndef CAFFE2_OPERATORS_CUDNN_OP_UTILS_H_ #define CAFFE2_OPERATORS_CUDNN_OP_UTILS_H_ #include "caffe2/core/cudnn_wrappers.h" namespace caffe2 { // Earlier in the days Caffe sets the default cudnn workspace to 8MB. We bump // it up to 64MB in Caffe2, as this enables the use of Winograd in many cases, // something very beneficial to more recent CNN models. static constexpr size_t kCONV_CUDNN_WORKSPACE_LIMIT_BYTES = 64 * 1024 * 1024; // Manually specified number of algorithms implemented in CuDNN. // This does not have any performance implications, as we will always find the // fastest algorithm; setting them to the right number of algorithms will enable // us to best report the statistics when doing an exhaustive search, though. #if CUDNN_VERSION_MIN(7, 0, 0) // Note: Double each of these due to potential // tensorcore + non-tensorcore versions // which are treated as separate returned algos static constexpr size_t kNUM_CUDNN_FWD_ALGS = 2 * CUDNN_CONVOLUTION_FWD_ALGO_COUNT; static constexpr size_t kNUM_CUDNN_BWD_FILTER_ALGS = 2 * CUDNN_CONVOLUTION_BWD_FILTER_ALGO_COUNT; static constexpr size_t kNUM_CUDNN_BWD_DATA_ALGS = 2 * CUDNN_CONVOLUTION_BWD_DATA_ALGO_COUNT; #else static constexpr size_t kNUM_CUDNN_FWD_ALGS = 7; static constexpr size_t kNUM_CUDNN_BWD_FILTER_ALGS = 4; static constexpr size_t kNUM_CUDNN_BWD_DATA_ALGS = 5; #endif namespace { template inline void LogCuDNNPerfStats( const ArrayOfcudnnConvolutionAlgoPerf_t& perf_stat, int returned_algo_count) { VLOG(1) << "Perf result: (algo: stat, time, memory)"; for (int i = 0; i < returned_algo_count; ++i) { const auto& stat = perf_stat[i]; VLOG(1) << stat.algo << ": " << stat.status << " " << stat.time << " " << stat.memory; } } } // namespace // Easier indexing into force_algo_ vector, // shared by CudnnConvTransposeOpBase and CudnnConvOpBase to force // usage of a particular algortihm instead of searching enum { ALGO_FWD = 0, ALGO_WGRAD = 1, ALGO_DGRAD = 2 }; } // namespace caffe2 #endif // CAFFE2_OPERATORS_CUDNN_OP_UTILS_H_