How does this compare with cuFFT?

[cherrybajan]

Hello!

Thanks for this amazing work. I was wondering how this compares with cuFFT in terms of speed. Thanks!!

[jeromeku]

@cherrybajan

torch uses cuFFT underneath the hood. See here.

If you use nsys to profile torch.fft per the reference FFT implementation in test_flashfftconv , you'll see a kernel regular_fft_r2c which corresponds to the function linked above.