Hi Michael, Michael R. Crusoe, on 2020-11-09 17:29:11 +0100: > On Mon, 9 Nov 2020 at 17:21, Étienne Mollier <etienne.mollier@mailoo.org> > wrote: > > I'm filling a wishlist item in the bug tracker, so that the > > discussion does not disappear inside mail archives. I gave a > > try to shapeit4 autopkgtest suite with and without FMA & AVX2 > > support, but it had a run time of 1m25s in both cases on my > > machine (Ryzen 5 3600 w/ 6 cores). It is quite possible I > > neglected some other bottlenecks though, but the assembler did > > embed AVX2 instructions when I checked the build result. Out of > > curiosity, has someone figures on the performance gain for that > > software when extensions are available? > > > > Michael R. Crusoe, on 2020-11-05 21:26:30 +0100: > > > As documented at > > > https://wiki.debian.org/SIMDEverywhere > > > > shapeit4 provides a dedicated code path for "-mfma -mavx2" build > > options, and another one for generic builds. Is it still worth > > using SIMDe in this particular situation? The "use case" > > paragraph of the wiki page seems to suggest it is not strictly > > needed here. > > > > Given the fallback route that doesn't use SIMD, then implementing our own > is not necessary, however compiling the FMA+AVX2 path using SIMDe on > non-x86 archs may result in a speedup for them. Okay, thanks for your thoughts; IIUC, given enough instructions, the compiler may see fit some code paths, otherwise not possible with a strict baseline. > Would be best to get a bigger training dataset to confirm the benefit, or > at least the lack of regression :-) Agreed, besides, such a change may actually put quite some entropy in the package. I'm holding my horses until I have an idea of the order of magnitude of the gain. > If a performance benefit is observed, it might be interesting to see if the > AVX-only and "lower" SIMD levels on x86 also experience a speed up. Yep, if we can have a typical enough workload to quantify the gain, that would be of great help to see which optimisations are worth it. Kind Regards, -- Étienne Mollier <etienne.mollier@mailoo.org> Fingerprint: 8f91 b227 c7d6 f2b1 948c 8236 793c f67e 8f0d 11da Sent from /dev/pts/3, please excuse my verbosity.
Attachment:
signature.asc
Description: PGP signature