Large Sizes Segfault about rocfft HOT 3 CLOSED

Hello; thanks for the report.

I think that this may have to do with the fact that some larger problem sizes use work memory; this isn't provided in the example code. Could you try looking at the examples in

https://github.com/ROCmSoftwarePlatform/rocFFT/tree/develop/docs/samples

and see if you still have an issue?

from rocfft.

Hello; thanks for the report.

I think that this may have to do with the fact that some larger problem sizes use work memory; this isn't provided in the example code. Could you try looking at the examples in

https://github.com/ROCmSoftwarePlatform/rocFFT/tree/develop/docs/samples

and see if you still have an issue?

Hi @malcolmroberts, that worked, thanks! I switched to the first sample and I could use larger sizes. On a side note, do you expect this library to be on par with cuFFT for most sizes? I tried a very large FFT size + batch (2^18 with batch of 100), and it's about 2x slower when I account for the fact that my Radeon VII has about 75% of the cores the V100 does.

from rocfft.

Glad that worked!

We have done some work on performance analysis, and large 1D sizes is a target. We have some ideas, so stay tuned! But our goal isn't to be only just as fast as cuFFT. ;)

from rocfft.