andersbll / cudarray

CUDA-based NumPy
MIT License
233 stars 61 forks source link

Restrictions to CudNN and CUDA toolkit version. #81

Open Erik-Koning opened 6 years ago

Erik-Koning commented 6 years ago

I am trying to run neural_artistic_style with the Nvidia CUDA toolkit 9.1 and 7.1 cudNN. Are these newer version from nividia not supported? I have ran into many problems such as the GPU architecture not supported error in this project, and other errors in neural_artistic_style.

Erik-Koning commented 6 years ago

When making cudarray specifically i also recieve the error: src/nnet/cudnn.cpp:169:5: error: too few arguments to function ‘cudnnStatus_t cudnnSetConvolution2dDescriptor(cudnnConvolutionDescriptor_t, int, int, int, int, int, int, cudnnConvolutionMode_t, cudnnDataType_t)'));

prannayk commented 6 years ago

Yes it does seem that this library does not work with CUDA 8.0 or above

The-Oracle commented 5 years ago

For cudNN 7.x and CUDA 9.x compatibility, make the following change in cudnn.cpp, replace lines 167-169 with the following:

CUDNN_CHECK(cudnnSetConvolution2dDescriptor( conv_desc, pad_y, pad_x, stride_y, stride_x, 1, 1, CUDNN_CONVOLUTION, CUDNN_DATA_FLOAT ));

This will get you past the compilation error, I have not yet had the chance to test the library with cudNN 7 and CUDA 9.

yueanga commented 3 years ago

For cudNN 7.x and CUDA 9.x compatibility, make the following change in cudnn.cpp, replace lines 167-169 with the following:

CUDNN_CHECK(cudnnSetConvolution2dDescriptor( conv_desc, pad_y, pad_x, stride_y, stride_x, 1, 1, CUDNN_CONVOLUTION, CUDNN_DATA_FLOAT ));

This will get you past the compilation error, I have not yet had the chance to test the library with cudNN 7 and CUDA 9.

it worked with cuda10 as well thank you so much!!!!

MobtgZhang commented 3 years ago

I am trying to install cudarry,but the version of cuda-toolkit is cuda10.2 and cuDNN 8.2.1.32,I meet the same problems as you,how can I solve this problem?

src/nnet/cudnn.cpp:169:5: error: too few arguments to function ‘cudnnStatus_t cudnnSetConvolution2dDescriptor(cudnnConvolutionDescriptor_t, int, int, int, int, int, int, cudnnConvolutionMode_t, cudnnDataType_t)’ )); ^

MobtgZhang commented 3 years ago

and how about the installation of cuda10.x and cuda11.x for cudarray?

Bernadette-Mohr commented 1 year ago

and how about the installation of cuda10.x and cuda11.x for cudarray?

For me, the change in the code also worked for cuda 11.6.

SWEDEN1003 commented 1 year ago

For cudNN 7.x and CUDA 9.x compatibility, make the following change in cudnn.cpp, replace lines 167-169 with the following:

CUDNN_CHECK(cudnnSetConvolution2dDescriptor( conv_desc, pad_y, pad_x, stride_y, stride_x, 1, 1, CUDNN_CONVOLUTION, CUDNN_DATA_FLOAT ));

This will get you past the compilation error, I have not yet had the chance to test the library with cudNN 7 and CUDA 9.

For me, it can work for cuda -11.3