Date: Mon, 10 Feb 2025 12:48:31 -0500 From: enh <enh@google.com> To: sgk@troutmask.apl.washington.edu Cc: Pedro Giffuni <pfg@freebsd.org>, "freebsd-numerics@FreeBSD.org" <freebsd-numerics@freebsd.org> Subject: Re: Time to evaluate the CORE-MATH project? Message-ID: <CAJgzZorVAqtnMvGOp5A=mUtu21ODnLFVyFPpi5poguf8x22vPA@mail.gmail.com> In-Reply-To: <Z6o2e8zQJLOdMKun@troutmask.apl.washington.edu> References: <1749328077.10705244.1739204040803.ref@mail.yahoo.com> <1749328077.10705244.1739204040803@mail.yahoo.com> <Z6o2e8zQJLOdMKun@troutmask.apl.washington.edu>
next in thread | previous in thread | raw e-mail | index | archive | help
yeah, i'm interested in switching Android's libm over to the llvm-libc routines, but not until i can do it by taking a whole family of f/d/ld variants at a time... On Mon, Feb 10, 2025 at 12:25 PM Steve Kargl <sgk@troutmask.apl.washington.edu> wrote: > > On Mon, Feb 10, 2025 at 04:14:00PM +0000, Pedro Giffuni wrote: > > Quoting a recent posting in hackers@:https://lists.freebsd.org/archives/freebsd-hackers/2025-February/004304.txt > > From: Paul Zimmermann <Paul.Zimmermann_at_inria.fr>Date: Mon, 10 Feb 2025 08:53:53 UTC Dear all, > > a new update of our comparison is available, with FreeBSD 14.2: > > https://members.loria.fr/PZimmermann/papers/accuracy.pdf > > Main changes in this update: > > * GNU libc now provides 21 correctly-rounded functions in single precision (integrating the CORE-MATH code)* LLVM-libc provides more correctly-rounded functions > > Brian GladmanVincenzo InnocenteJohn MatherPaul Zimmermann______ > > The code is under an MIT license:https://core-math.gitlabpages.inria.fr/ > > From the table, code is available for float and for double except > lgamma. There are only 6 functions implemented for LD80 long > double and no code for LD128 long double. This suggest that if > someone starts importing parts of core-math, then freebsd's libm > will be a mixture of msun and core-math, i.e, maintenance will > become more challenging. Someone has already imported a few > routines from Cephus to cover missing functionality in LD128. > I know bde would not look at cephus code, and I seldom look at > it as there is no one to review a change if I made one. > > PS: The core-math LD80 routines would need to be tested on IA32 > due to the way FreeBSD sets the FPU to 53-bit precision. Of > course, FreeBSD may drop IA32 support, so this might not be an > issue. > > -- > Steve >
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?CAJgzZorVAqtnMvGOp5A=mUtu21ODnLFVyFPpi5poguf8x22vPA>