cufft: use CUFFT_COMPATIBILITY_FFTW_PADDING instead of CUFFT_COMPATIBILITY_NATIVE

[godsic]

CUFFT_COMPATIBILITY_NATIVE mode is broken since cuda7.0+. Since CUFFT_COMPATIBILITY_NATIVE has been marked as DEPRECATED since cuda6.0, Nvidia developers are not willing to fix it. Therefore mumax3 should switch to the default CUFFT_COMPATIBILITY_FFTW_PADDING mode. This is rather big change.

Since Nvidia now suggests to use out-of-place r2c and c2r transforms, i.e. see page 20 of the "CUFFT LIBRARY USER'S GUIDE DU-06707-001_v7.5", then I would prefer to take this route. This will marginally increase mumax3 memory consumption, but is way less harmful, then say touching mighty demag kernel code.

@barnex any thoughts?

[barnex]

Yes, I agree. I don't want to touch the kernel layout at any cost.
I'll give it a shot this weekend.

[godsic]

If you are busy, I am happy to work on this next week.

On 7 November 2015 at 21:09, Arne Vansteenkiste [email protected]
wrote:

Yes, I agree. I don't want to touch the kernel layout at any cost.
I'll give it a shot this weekend.

—
Reply to this email directly or view it on GitHub
#52 (comment).

Mykola

[barnex]

Sounds good. Let me know if you need any help.
e0c3a28 re-introduces the convolution self-test (against brute-force , for a sparse random magnetization). It is enabled by the flag -paranoid, and is on in all tests. This should help in testing the fix.

[godsic]

Thanks, this should help indeed.

[godsic]

de985bf
7313693

[barnex]

Thanks @godsic! All tests pass on my 680.