clfft benchmarks

Question

clfft benchmarks

psteinb opened this issue 4 years ago · comments

Hi folks,
great to see this coming to live. Just wanted to share our benchmark for FFT libraries:
https://github.com/mpicbg-scicomp/gearshifft
and the viz of the current results are available here
https://www.kcod.de/gearshifft/
which include numbers for clfft.

In the distant future, clij2-fft and other projects might be a good subject for further study. For example, I'd be interested in the overhead that is incurred by calling clfft from java.

Robert Haase · Answer 1 · Fri May 29 2020 17:03:57 GMT+0800 (China Standard Time)

Hey Peter @psteinb ,

just a link. There was some discussion about benchmarks @bnorthan performed earlier. You may find insightful:
https://forum.image.sc/t/high-performance-fft-based-algorithms-deconvolution/31710/8

Btw. I also think that clij2-fft has huge potential. It's so good what Brian implemented!

Cheers,
Robert

Peter Steinbach · Answer 2 · Fri May 29 2020 17:20:20 GMT+0800 (China Standard Time)

:-D interesting thread. Great to hear that people are interested in FFTs. Our data doesn't support a few of the claims made, but that motivates us to continue. It's a shame that clfft doesn't see continued investment. We've kept it in the loop for backwards compatibility. With Intel going for oneAPI/Sycl/DPC++ and AMD investing in ROCM, I fear that we soon see more FFT libs coming to live. rocFFT is already a thing: https://github.com/ROCmSoftwarePlatform/rocFFT (Linux only for now AFAIK) Fortunately, @zyzzyxdonta is working on extending gearshifft (with an HPC focus though). So if this community has any wishes on what we can improve, let us know! Our motivation for gearshifft always was to spare people the time to benchmark their beloved FFT code to competitors.