mmr/tkDNN
Archived
1
0
Fork 0
Code Issues 46 Pull Requests 12 Releases 4 Activity

Patch for CUDNN 8.0.1

Browse Source 
Signed-off-by: tk <micaelaverucchi@gmail.com>
This commit is contained in:
tk
2020-07-16 18:16:09 +02:00
parent b2df9fc110
commit c4aad7fe95
2 changed files with 15 additions and 15 deletions
Download Patch File Download Diff File Expand all files Collapse all files
include/tkDNN/Layer.h
+2 -2
View File
@@ -273,8 +273,8 @@ public:
protected: protected:
cudnnFilterDescriptor_t filterDesc; cudnnFilterDescriptor_t filterDesc;
cudnnConvolutionDescriptor_t convDesc; cudnnConvolutionDescriptor_t convDesc;
cudnnConvolutionFwdAlgo_t algo; cudnnConvolutionFwdAlgoPerf_t algo;
cudnnConvolutionBwdDataAlgo_t bwAlgo; cudnnConvolutionBwdDataAlgoPerf_t bwAlgo;
cudnnTensorDescriptor_t biasTensorDesc; cudnnTensorDescriptor_t biasTensorDesc;
void initCUDNN(bool back = false); void initCUDNN(bool back = false);
src/Conv2d.cpp
+13 -13
View File
@@ -63,23 +63,23 @@ void Conv2d::initCUDNN(bool back) {
workSpace = NULL; workSpace = NULL;
ws_sizeInBytes = 0; ws_sizeInBytes = 0;
if(back) { if(back) {
checkCUDNN( cudnnGetConvolutionBackwardDataAlgorithm(net->cudnnHandle, checkCUDNN( cudnnGetConvolutionBackwardDataAlgorithm_v7(net->cudnnHandle,
filterDesc, dstTensor, convDesc, srcTensor, filterDesc, dstTensor, convDesc, srcTensor, 1, 0, &bwAlgo) );
CUDNN_CONVOLUTION_BWD_DATA_PREFER_FASTEST, 0, &bwAlgo) );
checkCUDNN(cudnnGetConvolutionBackwardDataWorkspaceSize(net->cudnnHandle, checkCUDNN(cudnnGetConvolutionBackwardDataWorkspaceSize(net->cudnnHandle,
filterDesc, dstTensor, convDesc, srcTensor, filterDesc, dstTensor, convDesc, srcTensor,
bwAlgo, &ws_sizeInBytes)); bwAlgo.algo, &ws_sizeInBytes));
// invert tensors // invert tensors
srcTensorDesc = dstTensor; srcTensorDesc = dstTensor;
dstTensorDesc = srcTensor; dstTensorDesc = srcTensor;
} else { } else {
checkCUDNN( cudnnGetConvolutionForwardAlgorithm(net->cudnnHandle, checkCUDNN( cudnnGetConvolutionForwardAlgorithm_v7(net->cudnnHandle,
srcTensor, filterDesc, convDesc, dstTensor, srcTensor, filterDesc, convDesc, dstTensor,
CUDNN_CONVOLUTION_FWD_PREFER_FASTEST, 0, &algo) ); 1, 0, &algo) );
checkCUDNN(cudnnGetConvolutionForwardWorkspaceSize(net->cudnnHandle, checkCUDNN(cudnnGetConvolutionForwardWorkspaceSize(net->cudnnHandle,
srcTensor, filterDesc, convDesc, dstTensor, srcTensor, filterDesc, convDesc, dstTensor,
algo, &ws_sizeInBytes)); algo.algo, &ws_sizeInBytes));
} }
} }
@@ -91,12 +91,12 @@ void Conv2d::inferCUDNN(dnnType* srcData, bool back) {
checkCUDNN(cudnnConvolutionBackwardData(net->cudnnHandle, checkCUDNN(cudnnConvolutionBackwardData(net->cudnnHandle,
&alpha, filterDesc, data_d, &alpha, filterDesc, data_d,
srcTensorDesc, srcData, srcTensorDesc, srcData,
convDesc, bwAlgo, workSpace, ws_sizeInBytes, convDesc, bwAlgo.algo, workSpace, ws_sizeInBytes,
&beta, dstTensorDesc, dstData)); &beta, dstTensorDesc, dstData));
} else { } else {
checkCUDNN(cudnnConvolutionForward(net->cudnnHandle, checkCUDNN(cudnnConvolutionForward(net->cudnnHandle,
&alpha, srcTensorDesc, srcData, filterDesc, &alpha, srcTensorDesc, srcData, filterDesc,
data_d, convDesc, algo, workSpace, ws_sizeInBytes, data_d, convDesc, algo.algo, workSpace, ws_sizeInBytes,
&beta, dstTensorDesc, dstData)); &beta, dstTensorDesc, dstData));
} }