enable tensor cores for fp16 convolutions

This commit is contained in:
YashasSamaga 2019-12-16 15:38:12 +05:30
parent c2b6c67431
commit cf93df41fc

View File

@ -224,6 +224,8 @@ namespace cv { namespace dnn { namespace cuda4dnn { namespace csl { namespace cu
);
}
CUDA4DNN_CHECK_CUDNN(cudnnSetConvolutionGroupCount(descriptor, group_count));
if (std::is_same<T, half>::value)
CUDA4DNN_CHECK_CUDNN(cudnnSetConvolutionMathType(descriptor, CUDNN_TENSOR_OP_MATH));
} catch (...) {
/* cudnnDestroyConvolutionDescriptor will not fail for a valid desriptor object */
CUDA4DNN_CHECK_CUDNN(cudnnDestroyConvolutionDescriptor(descriptor));