A convolution implementation based on cuda extension for PyTorch. The source code reference to the PyTorch's inefficient implementation here. See here for the accompanying tutorial.
- Build CUDA extensions by going into the
conv_cuda/
folder and executingpython setup.py install
, - JIT-compile CUDA extensions by going into the
conv_cuda/
folder and callingpython jit.py
, which will JIT-compile the extension and load it, - Check the result of the convolution by running
python test.py