cuFFT doubt.

Hi. I’m new with this cuda stuff. I appreciate any advice you could give me.

I need to calculate 2^k sized FFT from a big input (several times (20-100k times) 2^k). What’s the way to go with cuFFT? Batches? I’ve tried with fftw over an i7 but it’s taking too much time. Also I have a tesla avaiable so I thought this is a good opportunity to make it work. Is the performance boost noticeable?

As you can see, I’m really new here so I’ll appreciate help of any kind. In the meanwhile I’ll be reading the cuFFT documentation.

Thx for your time!

batching is a good idea. For larger FFTs (k is larger) it probably won’t matter much whether you batch or not. I don’t know what to tell regarding perf compared to FFTW, that would definitely depend on a lot of factors you haven’t specified, like k, and which i7 exactly, and which GPU exactly. If you want best performance, use newest CUDA version (6.5 or 7.0 RC) and it’s good to use power-of-2 sized FFTs as you have indicated (2^k)