Hello. I am have had an occasion to use FFTs extensively in my projects.
However, I used torch.backends.cudnn.enabled=True to speed up my code.
In doing so, I wonder if the Fourier Transforms are also sped up by this setting or only the convolutions.
After all, the FFTs are in the cuFFT library, not the cuDNN library.
However, I would be most grateful to know if this is so.
This is because the data being Fourier transformed is of different sizes while the image data is obtained by cropping the Fourier Transformed data to a fixed size.
Thank you in advance for anyone who can help out.
cuFFT should be called by _fft, if you are running your code on a GPU.
EDIT: Sorry, I misunderstood the question (or rather skipped the cudnn part).
The cuFFT methods should not depend on cudnn, as you already mentioned. 
Thank you! You have saved me from quite a lot of worry.