Bug 1114964
|
Description
Dave Love
2014-07-01 10:56:44 UTC
If you have the time and resources to run benchmark both and compare, please do. (I started running benchfft on my personal machine, but it takes a long time.) (In reply to Conrad Meyer from comment #1) > If you have the time and resources to run benchmark both and compare, please > do. It might help to know under what circumstances the AVX is supposed to be slow. Anyhow, I'll attach results with and without AVX (selected with an environment variable, using a modified version of the library) on a sandybridge node. I'm a bit confused by them, and the bench program in the source tests directory gives better results for AVX in the cases I tried, but I don't have time to investigate now. Created attachment 914522 [details]
sandybridge benchfft real, double, avx
Created attachment 914523 [details]
sandybridge benchfft real, double, no avx
Created attachment 914524 [details]
sandybridge benchfft complex, double, avx
Created attachment 914525 [details]
sandybridge benchfft complex, double, no avx
Created attachment 914527 [details] Graph of AVX vs no-AVX (real and complex), powers of 2 > It might help to know under what circumstances the AVX is supposed to be slow. Honestly, I have no idea. Created attachment 914528 [details]
Graph of AVX vs no-AVX (real and complex), non-powers of 2
Oh, hrm, plots may be bogus... Created attachment 914531 [details]
Graph of AVX vs no-AVX (real and complex), powers of 2
Fixed the plot.
Created attachment 914532 [details]
Graph of AVX vs no-AVX (real and complex), non-powers of 2
Plot fixed. Doesn't seem to make much difference on the real mode (r2r), but for complex number FFT, AVX is the clear winner.
Thanks for collecting the data!
The original rationale was that GROMACS claims 20% worse performance with avx enabled fftw. But I guess it should be enabled in the default fftw library. It sure would be useful if one could trigger the use of avx instructions in userland. Rawhide build w/ AVX enabled: http://koji.fedoraproject.org/koji/taskinfo?taskID=7104966 (tested local build first and it finished fine). > GROMACS claims 20% worse performance with avx enabled fftw
They claim *up to* 20% worse, because they only use short transform lengths. I agree that it's not a good reason to disable AVX for the system library, especially when we can't reproduce those claims. (The attached graphs don't show anything close to 20% worse, and on larger transform sizes we see 40-60% improvement with AVX enabled.)
fftw-3.3.4-3.fc20 has been submitted as an update for Fedora 20. https://admin.fedoraproject.org/updates/fftw-3.3.4-3.fc20 fftw-3.3.4-3.fc20 has been pushed to the Fedora 20 stable repository. If problems still persist, please make note of it in this bug report. This bug appears to have been reported against 'rawhide' during the Fedora 22 development cycle. Changing version to '22'. More information and reason for this action is here: https://fedoraproject.org/wiki/Fedora_Program_Management/HouseKeeping/Fedora22 For information of anyone interested (not suggesting it needs action): There's a recent change available from Gromacs-land to address what I assume is the reported problem with small transforms: <https://github.com/FFTW/fftw3/commit/b606e3191e5b65e2e13f67ef7dad5b1e7c40206c>. Fedora 22 changed to end-of-life (EOL) status on 2016-07-19. Fedora 22 is no longer maintained, which means that it will not receive any further security or bug fix updates. As a result we are closing this bug. If you can reproduce this bug against a currently maintained version of Fedora please feel free to reopen this bug against that version. If you are unable to reopen this bug, please file a new report against the current release. If you experience problems, please add a comment to this bug. Thank you for reporting this bug and we are sorry it could not be fixed. |