Can the FFT’s in CUDA via the CUFFT library be depended on to be deterministic? That is if the same fft input data is
run multiple times do you get the same results exactly?
FFTW can pick different plan algorithms at different times for the same input data. CUFFT seems to be somewhat modelled or
based on FFTW. The way around this is to run a sample FFT of each size, create wisdom data and save that off. The later FFTs of a particular size will use the
plan suggested by the wisdom information. I’ve used this to make FFTW and Jfftw3 FFTs deterministic but the CUFFT library doesn’t expose any wisdom interface. I’m wondering if this would be a problem.