Runtime dependency issue on CUDA 10.0

NVIDIA / FastPhotoStyle

Style transfer, deep learning, feature transform

Other

11.15k stars 1.2k forks source link

Runtime dependency issue on CUDA 10.0 #79

Open sleebapaul opened 5 years ago

sleebapaul commented 5 years ago

Did anyone try to reproduce the code in Azure ML Pipeline? It has a default Conda environment with CUDA 10.0 version.

On research, I found that,

pynvrtc library doesn't even have an update on CUDA latest version. So mostly, I can't migrate the code to the latest version of CUDA even if I modify the header files in smooth_filter.py.

And why there is a tight coupling between CUDA 9.1 and the code? Any help is greatly appreciated.

Adlinga commented 5 years ago

To solve this try 0) check your CUDA paths in the system enviroment 1) In smoothfilter.py: Change path in line 6 to your CUDA directory. For me it was #include "C:/NVIDIA/CUDA/10/include/mathfunctions.h" 2) In smooth_filter.py: Change line 336 to ptx = program.compile() 3) In your Anaconda or python path find ...\Lib\site-packages\pynvrtc and there interface.py. In interface.py: Change line 96 to _def_lib_name = 'nvrtc64_1000.dll' 4) Enjoy!

jverygood commented 5 years ago

If you are using CUDA10.0 under Ubuntu, you should also modify the export LD_LIBRARY_PATH=${ANACONDA}/lib:${CUDA_PATH}/bin64:$LD_LIBRARY_PATH to export LD_LIBRARY_PATH=${ANACONDA}/lib:${CUDA_PATH}/lib64 :$LD_LIBRARY_PATH