Precision of FFT with non-power-of-2 sizes starts to deteriorate around size 65536

fjarri / reikna

Pure Python GPGPU library

http://reikna.publicfields.net/

MIT License

164 stars 16 forks source link

Precision of FFT with non-power-of-2 sizes starts to deteriorate around size 65536 #51

Open fjarri opened 5 years ago

fjarri commented 5 years ago

There seems to be some accuracy problem with Bluestein's transform. The threshold for the size is around 65536, after which the error grows quickly. The effect appears both with single and double precision, and with fast_math set to either True or False. There is no error with OpenCL and Intel\ UHD Graphics 630.

fjarri commented 5 years ago

Sorry for taking so long to reply, I was pretty busy the last few days. I will try to look into that during the weekend, but I'm not sure if it's the same problem - the array sizes are pretty small. Would it be possible for you to post the reproducing code?

jyhmiinlin commented 5 years ago

I checked: it is not related to reikna.fft. I am deleting my post for now.