| Git Clone URL: | https://aur.archlinux.org/fftw-amd.git (read-only, click to copy) |
|---|---|
| Package Base: | fftw-amd |
| Description: | A library for computing the discrete Fourier transform (DFT) - AMD Ryzen/EPYC Optimized Version |
| Upstream URL: | https://github.com/amd/amd-fftw/ |
| Keywords: | AMD DSP epyc FFT Ryzen Zen |
| Licenses: | GPL2 |
| Conflicts: | fftw |
| Provides: | fftw, libfftw3.so, libfftw3f.so, libfftw3f_mpi.so, libfftw3f_omp.so, libfftw3f_threads.so, libfftw3l.so, libfftw3l_mpi.so, libfftw3l_omp.so, libfftw3l_threads.so, libfftw3q.so, libfftw3q_omp.so, libfftw3q_threads.so, libfftw3_mpi.so, libfftw3_omp.so, libfftw3_threads.so |
| Submitter: | tuxzz |
| Maintainer: | MeissnerEffect |
| Last Packager: | MeissnerEffect |
| Votes: | 5 |
| Popularity: | 0.014244 |
| First Submitted: | 2020-05-17 18:43 (UTC) |
| Last Updated: | 2024-04-26 02:36 (UTC) |
I've made an updated PKGBUILD for the 5.0 release, which should be easier to update in the future (since the version numbers are now variables) and no longer overwrites everyone's CFLAGS, just supplements them as needed.
There was a compile bug, currently suppressed by adding -Wno-incompatible-pointer-types
You can get/merge my updated one from here: https://gist.github.com/jclds139/5fe8db4110151a718d7fe794fdbac47e
Hi again, I cannot seem to figure out how to get the 4.x updates to work based on this PKGBUILD template. There seems to be no simple workaround at the moment.
The current version is 4.1… additionally, the “provides” seems to use old methodology for PKGBUILD stuff.
Oh, nevermind... I just realized that this is a completely separate project :-/
I will ping the upstream about upstreaming their fixes.
@tuxzz: Do the AMD optimizations have a diametral effect on using Intel CPUs?
I am wondering whether to include the flag in the official package or not.
If you are using AMD Ryzen series or EPYC series CPU, this package can boost the speed of fftw up around 30% compared to Archlinux official repository (Test on Ryzen 3700X).
Intel MKL perform very poor on non-Intel CPUs, because it detect your CPU and switch to a slow path.
I know there are some environment variables such as MKL_DEBUG_CPU_TYPE or MKL_NUM_THREADS. But Intel removed these trick since MKL 2020 Update 1.
By default, this build enables -ffast-math compiler flag, which provides better performance and works well on most applications. If you require some extra numeric percision, remove it from PKGBUILD by yourself.
If you have problem on this package, report issue to AUR comment, AMD YES ^w^.
Pinned Comments
tuxzz commented on 2020-05-31 17:38 (UTC) (edited on 2020-05-31 17:39 (UTC) by tuxzz)
Why not use Archlinux offical repository
If you are using AMD Ryzen series or EPYC series CPU, this package can boost the speed of fftw up around 30% compared to Archlinux official repository (Test on Ryzen 3700X).
Why not use Intel MKL
Intel MKL perform very poor on non-Intel CPUs, because it detect your CPU and switch to a slow path. I know there are some environment variables such as
MKL_DEBUG_CPU_TYPEorMKL_NUM_THREADS. But Intel removed these trick since MKL 2020 Update 1.WARNING
By default, this build enables
-ffast-mathcompiler flag, which provides better performance and works well on most applications. If you require some extra numeric percision, remove it fromPKGBUILDby yourself.If you have problem on this package, report issue to AUR comment, AMD YES ^w^.