I'm a NN developer who want to transport a pytorch Module (https://github.com/jonas-koehler/s2cnn) to chainer.
The original Module uses a homemade kernel, wrapped by pynvrtc and cupy.cuda.function.Module.get_function().
In searching information for the transportation, I found a PR which refer not to use pynvrtc in cupy.
#33 (comment)
My question is which is the better to use pynvrtc or cupy.cuda.compiler to create an original kernel used for chainer.
I'm a NN developer who want to transport a pytorch Module (https://github.com/jonas-koehler/s2cnn) to chainer.
The original Module uses a homemade kernel, wrapped by
pynvrtcandcupy.cuda.function.Module.get_function().In searching information for the transportation, I found a PR which refer not to use pynvrtc in cupy.
#33 (comment)
My question is which is the better to use
pynvrtcorcupy.cuda.compilerto create an original kernel used for chainer.