[prev in list] [next in list] [prev in thread] [next in thread] 

List:       sbcl-devel
Subject:    Re: [Sbcl-devel] sb-simd is (almost) ready
From:       Stas Boukarev <stassats () gmail ! com>
Date:       2022-06-21 13:12:28
Message-ID: CAF63=13yWbcyuzbBdBo_tbGkh69efxAVRE59SsHK+fFC99jOjg () mail ! gmail ! com
[Download RAW message or body]

It doesn't seem like there's a test for avx1 when building, so it
would fail on both kinds of machines?

On Tue, Jun 21, 2022 at 4:11 PM Stas Boukarev <stassats@gmail.com> wrote:
> 
> Just don't build when (zerop (sb-alien:extern-alien "avx2_supported" int))?
> 
> On Tue, Jun 21, 2022 at 3:59 PM Heisig, Marco <marco.heisig@fau.de> wrote:
> > 
> > I just pushed a fix for the sb-unicode issue, and I'm now looking into that build \
> > failure on mac. 
> > Judging from the logfile, the problem is that one of the functions in \
> > define-fake-vops.lisp contains a 256-bit SIMD constant that goes through \
> > load-int-avx2-immediate, and that function uses an AVX2 instruction, triggering a \
> > SIGILL.  So although sb-simd is causing the problem, the actual culprit is SBCL \
> > in assuming there are no systems that have AVX but no AVX2, when clearly there is \
> > a CI system with such a CPU.  I will try to find a sensible solution for this \
> > problem.  Is there any way I can find out more information about the hardware of \
> > the mac runner, especially about its CPU? 
> > > Is there an assumption in sb-simd that it will be *built* on a host that
> > > supports all the simd instruction sets?
> > 
> > There is no such assumption in sb-simd.  In theory, sb-simd should even load on \
> > platforms with no SIMD support at all, in which case it would just be a fancy \
> > instruction database and where each SIMD function signals an error once it is \
> > invoked.  In practice, there are still some portability issues, which is why \
> > sb-simd is only enabled on x86-64 for now.  But an update with ARM and maybe even \
> > PPC support is definitely planned. 
> > Best regards,
> > Marco


_______________________________________________
Sbcl-devel mailing list
Sbcl-devel@lists.sourceforge.net
https://lists.sourceforge.net/lists/listinfo/sbcl-devel


[prev in list] [next in list] [prev in thread] [next in thread]