[Llvm-bgq-discuss] New bgclang nighty builds (and other updates)

Sun May 10 09:25:34 CDT 2015

Hi,

Super-brief summary:

The bug occurs at high optimization (-O3) with the vec_ld() or vec_st() functions when calling the (overloaded) version with single-precision arguments (i.e., float pointers).

This is reproducible even with a trivial piece of code that tries to load from single precision memory and then write out the contents of the (double precision) vector4doubles by storing to a double variable.

- Everything works fine with the double-precision functions, regardless of optimization
- It works fine with clang in single precision without optimization 
- It works fine with xlc.

Cheers,

Erik

From: Mark Abraham <mark.j.abraham at gmail.com>
Reply: Mark Abraham <mark.j.abraham at gmail.com>>
Date: 10 May 2015 at 16:04:51
To: Hal Finkel <hfinkel at anl.gov>>
Cc: llvm-bgq-discuss at lists.alcf.anl.gov <llvm-bgq-discuss at lists.alcf.anl.gov>>, Erik Lindahl <erik.lindahl at gmail.com>>
Subject:  Re: [Llvm-bgq-discuss] New bgclang nighty builds (and other updates)  

Hi,

Bump. Erik Lindahl has also observed this on bgclang on JUQUEEN. Is it helpful if I produce a small piece of code that reproduces some of the issues on bgclang 3.6?

Mark

On Wed, Mar 25, 2015 at 4:16 PM Mark Abraham <mark.j.abraham at gmail.com> wrote:
On Tue, Mar 24, 2015 at 7:29 PM, Hal Finkel <hfinkel at anl.gov> wrote:
----- Original Message -----
> From: "Mark Abraham" <mark.j.abraham at gmail.com>
> To: "Hal Finkel" <hfinkel at anl.gov>
> Cc: llvm-bgq-discuss at lists.alcf.anl.gov
> Sent: Tuesday, March 24, 2015 12:35:20 PM
> Subject: Re: [Llvm-bgq-discuss] New bgclang nighty builds (and other updates)
>
>
> Hi Hal,
>
>
> Thanks very much for the update & effort.
>

You're very welcome.

>
> I tried out the default bgclang 3.6.0 on vesta, but found a bunch of
> the GROMACS SIMD-layer unit tests failing. These need correct QPX
> vector intrinsics available. From memory, things worked fine with
> bgclang in ~August last year, but I no longer have those results. Is
> there a simple way I can compile on vesta with older bgclang to see
> where a problem might lie? Otherwise / depending what I learn, I'll
> break out a debugger.

The old builds are still all installed. Just use the MPI wrappers from:

/home/projects/llvm/<whatever>/mpi/bgclang/bin -- I don't know exactly which build you were using in August of last year, r209570-20140527 maybe?

Yes, that was it, looking at some cruft in my former build script. In any case, the latest GROMACS SIMD single-precision unit tests pass on that old compiler version (3.5.0), and many of them fail in a release build on the default 3.6.0 bgclang on vesta. For some of the cases, it looks like some junk memory gets loaded, somehow. A few tests pass, but no theme for passing or failing tests leaps out at me. Double precision unit tests are OK, though.

Naturally, in debug mode most of the tests pass. However, the two I looked at in ddt were fine up until they called vec_extract(x, 0) e.g. on lines https://github.com/gromacs/gromacs/blob/master/src/gromacs/simd/impl_ibm_qpx/impl_ibm_qpx.h#L315 (where  a SIMD vector of 1 2 3 1 was summed to 3, rather than 7) and https://github.com/gromacs/gromacs/blob/master/src/gromacs/simd/impl_ibm_qpx/impl_ibm_qpx.h#L462 (where a SIMD vector dot product of the first 3 lanes returns a garbage answer). So maybe that is a productive lead?

I couldn't inspect the disassembly in ddt, so I'm not sure where we can take this from here, Hal. Tarball of test results attached.

Thanks,

Mark

 -Hal

>
>
> Thanks,
>
>
> Mark
>
>
>
>
> On Fri, Mar 20, 2015 at 12:15 AM, Hal Finkel < hfinkel at anl.gov >
> wrote:
>
>
> Hello everyone,
>
> First, let me apologize to everyone, this is a few months late...
> but, hopefully, this will never be a problem again...
>
> I now have a system setup which automatically pulls in upstream
> changes and tries to merge those with the bgclang-specific patches,
> and then builds the resulting suite of bgclang RPMs. When this
> succeeds, the RPMs should be posted automatically to:
>
> http://www.mcs.anl.gov/~hfinkel/bgclang/
> (note that installing a build from here now also requires both the
> 'stage1' and 'stage2' RPMs as well)
>
> The first such nightly build, r232720-20150319, has been posted to
> that page.
>
> And, for the curious, the local repositories used for version control
> are now mirrored to github:
>
> https://github.com/hfinkel/clang-bgq
> https://github.com/hfinkel/llvm-bgq
> https://github.com/hfinkel/bgclang-aux
> https://github.com/hfinkel/compiler-rt-bgq
> https://github.com/hfinkel/libcxx-bgq
> https://github.com/hfinkel/openmp-bgq
> https://github.com/hfinkel/sleef-bgq
>
> Compared to the latest "released" version (r220548-20141024), the
> most-recent nightly build does show some performance regressions,
> and there are a few things I've not even tested yet (LTO, ASan,
> etc.), but it also contains a number of bug fixes and improvements,
> so feel free to test on your applications.
>
> One particular noteworthy improvement is that our OpenMP runtime
> library now has affinity support enabled. This means that all of the
> OpenMP 4 affinity features should work, and also that the default
> thread<->core bindings are now sensible.
>
> The bgclang wrapper script no longer disables 'fast-isel' instruction
> selection at -O0, so your debug builds should now be faster too.
> Also, the automated vectorization of math functions using our SLEEF
> library adaptation is controlled using the new -fveclib flag (so the
> wrapper script contains -fveclib=SLEEF, and you can add
> -fveclib=none to turn it off if desired for whatever reason).
>
> Also, the core QPX support has been contributed upstream (although
> not yet the Clang-level intrinsics support); so if you're using LLVM
> as a library, and want to just build from upstream sources instead
> of depending on the bgclang builds, that is now possible.
>
> Thanks again everyone, and please let me know if you experience any
> difficulties,
> Hal
>
> --
> Hal Finkel
> Assistant Computational Scientist
> Leadership Computing Facility
> Argonne National Laboratory
> _______________________________________________
> llvm-bgq-discuss mailing list
> llvm-bgq-discuss at lists.alcf.anl.gov
> https://lists.alcf.anl.gov/mailman/listinfo/llvm-bgq-discuss
>
>

--
Hal Finkel
Assistant Computational Scientist
Leadership Computing Facility
Argonne National Laboratory
-- 
Erik Lindahl <erik.lindahl at gmail.com> 
Professor of Biophysics, Dept. Biochemistry & Biophysics, Stockholm University 
Professor of Theoretical biophysics, Dept. Theoretical Physics, Royal Inst. Technology 
Science for Life Laboratory, Box 1031, 17121 Solna, Sweden
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.alcf.anl.gov/pipermail/llvm-bgq-discuss/attachments/20150510/9644918e/attachment-0001.html>