Make VKFFT the default FFT implementation
About
This MR is a first step towards getting rid of scikit-cuda which brings more and more trouble.
VKFFT has several advantages:
- Faster and much less memory-hungry than scikit-cuda/CUFFT
- Maintained and still under active development
- OpenCL backend comes for free
The only drawback is that it requires nvrtc for the cuda backend.
To do
-
Make vkfftthe default FFT implementation everywhere -
Remove obsolete code -
Ensure unit tests are passing, even when scikit-cuda or vkfft is absent -
Update changelog/documentation -
End-to-end reconstruction test
Edited by Pierre Paleo