Re: Polyhedron benchmark results on AMD64

[Tobias Burnus <burnus at net-b dot de>]

Hi François-Xavier,

François-Xavier Coudert a ecrit:
> I've been doing (for professional use) a run of the Polyhedron
> benchmark of the Fortran 95 compilers I have at hand on one of my
> x86_64-linux machines (this is: GNU, Intel, Portland, Sun and g95; I
> would really like to have Pathscale as it seems so fast, but I don't).
> I think people might find it interesting so here's the link:
> http://www.eleves.ens.fr/home/coudert/benchmark.pdf
Do you also know of my daily Polyhedron runs?
http://www.physik.fu-berlin.de/~tburnus/gcc-trunk/benchmark/#rt

There you can find also the result for Pathscale 2.9.99 besides those of
ifort 9.1, ifort 10beta, sunf95 (free also for commercial use!), NAG f95
and g95.

Pathscale is ~20% faster than gfortran. (The ifort10 regression in
channel only occurs on amd64 and not on intel64 and will probably be
fixed soon.)

Contrary to you I used -ftree-vectorize which seems to make things
faster, but breaks two g95 tests.

[sunf95]
> I especially wonder how they do get fatigue so fast...
This is indeed a good question, especially since they don't use sincos
(which ifort and newer gfortran 4.3 use). Pathscale is even faster!

gfortran can improve by ~25% using one of Richard's patches, I think it
is the following patch:
http://gcc.gnu.org/ml/gcc-patches/2007-01/msg02335.html

In addition, I think we can gain a bit more by vectorizing sincos (it is
not yet vectorized, if I understood Richard correctly).

Actually, one of the problems of gfortran is that we create much too
many temporary variables in the middle end as Richard mentioned
recently. --- Has anyone actually looked at his patch at
http://gcc.gnu.org/ml/fortran/2007-02/msg00132.html

> Feel free to comment or suggest other ideas. I'd also be interested in
> OpenMP benchmarks, although I've not yet found if there's any such
> benchmark suite available freely.
Well, there is a OpenMP validation suite available, but this is not
really a benchmark:
http://www.hlrs.de/organization/amt/projects/openmpvalidation/
(I think the version on the webpage is slightly outdated, however.
gfortran passes all tests.)

Tobias