Larger FFTs coming?

Are the FFT size restrictions in the current CUDA release final? I am personally hoping for much larger FFT support.

The current FFT size restrictions are not final. We plan to increase supported sizes in a future release, however we don’t have specific numbers at this time. What size limits are you interested in?

At least up to 2048*2048 (1D transform, not a 2048x2048 2D transform) would be nice since I have some existing code that has that limitation since that was the largest size texture we could reliably get on ATI cards under DX9 for some time.

I don’t know exactly how you’re coding the FFTs, but even if you only supported power of 2 FFTs for large sizes, that would still be extremely useful.

Yes, something in the 4M area would be nice. I don’t know if you have heard of GPUFFTW, but I was able to use it to get very good performance for 1D FFTs up to 8M. There appears to be a lot of overhead in using it, however, which I would hope that CUDA could avoid. At the moment, I have some hacky code that uses CUDA for FFTs up to 16K, and then GPUFFTW above that. Personally, I’m really only interested in 1D, power of 2 FFTs.

Steven, can you run GPUFFTW on 8800 correctly?

I use FC-6 (up to date), glDrawbuffers() makes some fault segmentation.

it is running ok after changed to multiple glDrawbuffer().

but the results seem not correct

1D FFTs limited only by available card memory would be ideal, especially on something like the Quadro 5600. Are you implementing something like Cooley-Tukey for decomposition of larger DFTs or is this something you’ll be expecting CUDA users to do on their own for 2^24 - 2^30 sorts of FFT lengths?

I know that there are people trying to develop CUDA implementations for the Seti@Home project and the size limit is not enough atm… FFT caculations for seti needs FFT length : 8, 16, 32, 64, 128, 256, 512, 1024, 2048, 4096, 8192, 16384, 32768, 65536, and 131072

Without having the proper lengths needed there are struggles in making optimized caculations for S@H and making it utilize the board to its maximum so in this department the community can only wait and hope.


Well i sincerly hope that someone here is a keen programmer and want to develop a Cuda version of the Seti@Home client…

The code is out there and if someone is skilled enough to convert the client from FFTW libraries to CUFFT libraries then jump aboard and talk to Simon here…

You’re help is truly appreciated…

Kind Regards Vyper