[Llvm-bgq-discuss] QPX clang BGQ

[Hal Finkel]

Rich, et al.,

I've found the fast-math bug (which actually was not fast-math specific*, but was almost always triggered by the Newton-iteration-based QPX division code that is now generated in fast-math mode). I've fixed this issue in my local repository, and also rebased to pick up some upstream bug fixes that could affect us. vesta has been updated with the new version. Unfortunately, my standard testing has revealed a new upstream bug. I've filed a bug report (PR15952), and as soon as that is fixed, I'll rebase again and make a new release.

(*) The problem was specifically that there was an incorrect optimization pattern for qvfnmsub[s].

Thanks again,
Hal

----- Original Message -----
> Rich,
> 
> Thanks for sending this! I can confirm that there appears to be a
> fast-math bug (enabled by passing -ffast-math or -Ofast), and I've
> found some local codes that also display this anomalous behavior.
> Tracking this down is next on my TODO list.
> 
>  -Hal
> 
> ----- Original Message -----
> > Folks,
> > 
> > I have recently taken Hal's latest compiler updates (version
> > r176829-20130309) out for a spin on a suite of loop kernels I have
> > developed for checking compiler optimizations.  What I found is
> > that
> > when
> > using the -ffast-math option, my result checksums are very
> > different
> > from
> > when I do not use -ffast-math (for all 30 loops in the suite except
> > one).
> > An earlier version (r176751-20130307) does not show these
> > differences.
> > Also, I did not see any significant performance differences across
> > the
> > loop suite between these compiler versions.  For completeness, I
> > also
> > compared the checksums generated when the code was compiled with
> > the
> > xlc
> > 12.1 compiler.  The checksums for -O0 vs. -O3 -qnostrict vs. -O3
> > -qstrict
> > are all essentially the same.  From this I concluded that something
> > is
> > broken in the latest version of the optimizer.
> > 
> > So, if you're going to use the latest compiler version, make sure
> > you
> > check your answers.
> > 
> > --Rich
> > 
> > 
> > 
> > 
> > 
> > On 5/8/13 9:42 AM, "Jeff Hammond" <jhammond at alcf.anl.gov> wrote:
> > 
> > >(copying LLVM discuss - you should sign up for this list)
> > >
> > >Hi Ivo,
> > >
> > >I don't know the answer, but if you're on BGQ, you should use
> > >Hal's
> > >version of LLVM anyways because I think it has other important
> > >patches, not just QPX ones.
> > >
> > >Best,
> > >
> > >Jeff
> > >
> > >2013/5/8 Ivo Kabadshow <i.kabadshow at fz-juelich.de>:
> > >> Jeff,
> > >>
> > >>  does clang work with QPX on BGQ or do I need Hals clang
> > >>  version?
> > >>
> > >> Best
> > >>  Ivo
> > >>
> > >>
> > >> 
> > >>-------------------------------------------------------------------------
> > >>-----------------------
> > >> 
> > >>-------------------------------------------------------------------------
> > >>-----------------------
> > >> Forschungszentrum Juelich GmbH
> > >> 52425 Juelich
> > >> Sitz der Gesellschaft: Juelich
> > >> Eingetragen im Handelsregister des Amtsgerichts Dueren Nr. HR B
> > >> 3498
> > >> Vorsitzender des Aufsichtsrats: MinDir Dr. Karl Eugen Huthmacher
> > >> Geschaeftsfuehrung: Prof. Dr. Achim Bachem (Vorsitzender),
> > >> Karsten Beneke (stellv. Vorsitzender), Prof. Dr.-Ing. Harald
> > >> Bolt,
> > >> Prof. Dr. Sebastian M. Schmidt
> > >> 
> > >>-------------------------------------------------------------------------
> > >>-----------------------
> > >> 
> > >>-------------------------------------------------------------------------
> > >>-----------------------
> > >
> > >
> > >
> > >--
> > >Jeff Hammond
> > >Argonne Leadership Computing Facility
> > >University of Chicago Computation Institute
> > >jhammond at alcf.anl.gov / (630) 252-5381
> > >http://www.linkedin.com/in/jeffhammond
> > >https://wiki.alcf.anl.gov/parts/index.php/User:Jhammond
> > >ALCF docs: http://www.alcf.anl.gov/user-guides
> > >_______________________________________________
> > >llvm-bgq-discuss mailing list
> > >llvm-bgq-discuss at lists.alcf.anl.gov
> > >https://lists.alcf.anl.gov/mailman/listinfo/llvm-bgq-discuss
> > 
> > _______________________________________________
> > llvm-bgq-discuss mailing list
> > llvm-bgq-discuss at lists.alcf.anl.gov
> > https://lists.alcf.anl.gov/mailman/listinfo/llvm-bgq-discuss
> > 
> _______________________________________________
> llvm-bgq-discuss mailing list
> llvm-bgq-discuss at lists.alcf.anl.gov
> https://lists.alcf.anl.gov/mailman/listinfo/llvm-bgq-discuss
>