Date: Wed, 13 May 2026 21:10:45 +0300 From: Max Brazhnikov <makc@freebsd.org> To: Vladimir Druzenko <vvd@freebsd.org>, Charlie Li <vishwin@freebsd.org> Cc: ports-committers@freebsd.org, dev-commits-ports-all@freebsd.org, dev-commits-ports-main@freebsd.org Subject: Re: git: b0b4d9e1eae8 - main - */*: Switch to NumPy 2.x Message-ID: <nYGoYZ2GQVSQo7_w7bMvjg@freebsd.org> In-Reply-To: <90e36e0f-0d12-4074-b19e-dd7998788ab4@freebsd.org> References: <69fa4601.3fb5c.76ab51f0@gitrepo.freebsd.org> <0a2389d3-cce3-4a17-a928-d17b095dcff5@freebsd.org> <90e36e0f-0d12-4074-b19e-dd7998788ab4@freebsd.org>
index | next in thread | previous in thread | raw e-mail
On Mon, 11 May 2026 17:40:13 -0400 Charlie Li wrote: > Vladimir Druzenko wrote: > > 11.05.2026 22:29, Charlie Li пишет: > >> Max Brazhnikov wrote: > >>> -Dcpu-baseline="none" is safe to commit and should be committed if we > >>> claim support for older CPU. > >>> > >> This is not as clear-cut as it seems. CPU dispatching on top of -Dcpu- > >> baseline=none does not cover all optimised code paths that newer/ > >> feature-rich CPUs can use. > > Official documentation says otherwise: https://numpy.org/devdocs/ > > reference/simd/build-options.html > I quoted what I said from this, please read again. "Dispatched kernels > *don't* cover all code paths." (emphasis mine) This means that certain > optimisations are never available, even on CPUs that support them, when > baseline is none, which is not acceptable. In particular, C++ routines > are not able to be dispatch-optimised yet.>> Given that numpy is a very > performance-sensitive software and > >> considering its userbase, this kind of performance hit can be a > >> showstopper. > >> > >> The base system can claim support for older CPU however it wants, but > >> it does not mean individual third-party software has to do the same. > >> Of course we in ports should make our best efforts in this regard but > >> the expected majority use-cases and configurations need to take > >> precedence. > >> > >> I'm not opposed to an option that makes clear that -Dcpu-baseline=none > >> is both not the default and strongly discouraged outside of CPUs older > >> than or lack features of -Dcpu-baseline=min. > > > > How many years to wait for the "right" decision? > > > Whoever wants to implement an option following this description before > me or another python@ member gets to it is welcome to submit a diff for > consideration. We either should officially drop support for CPU without SSE for packages or add -Dcpu-baseline=none to numpy port by default. We may also add OPTIMIZED_CFLAGS or whatever option to build numpy with -Dcpu-baseline="native" and -Dcpu-dispatch="none". Cheers, Maxhome | help
Want to link to this message? Use this
URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?nYGoYZ2GQVSQo7_w7bMvjg>