Re: open64 versus gcc

A.L. wrote:
> On 1 Dec 2006 09:49:03 -0500, "A.L." <alewando@fala2005.com> wrote:
>> [GCC is fine for systems programming. I've never done serious
>> numerical work in it, so you may well be right about that. -John]
>
> Disclaimer: My expiments with gcc and numerical computations ended
> in 2003. Maybe since this time something changedIf there is somebody
> here who is using the up-to-date version of gcc for large scale,
> intensive floating point number crunching, please share the
> experience.


Well, it's benchmarks rather than a real-world experience, but
Polyhedron's Fortran benchmarks are specifically aimed at "large-scale
intensive floating-point number crunching". On those, gfortran (the gcc
Fortran front-end) comes in, on average, about 25-30% slower than the
best compilers -- but there's a fair bit of variation; in a fair number
of cases it's about as fast as anything else.


It should also be noted that this is testing GCC's Fortran front end
(gfortran), of course, rather than the far-more-mature C or C++
compilers; that probably accounts for some of the slowness. (Compare,
for instance, the g95 compiler, which also uses the gcc backend but
doesn't have nearly as much optimization work on the frontend.) In
particular, I believe the notable slowness on the AERMOD benchmark has
been traced to one particular inner loop optimization that's missed in
the front end.


Actual results (for P4 on Linux) here, and this page links AMD results:
          http://www.polyhedron.com/pb05/linux/f90bench_p4.html


At the very least, it's certainly quite usable for serious numerical
work; the OpenFOAM CFD package (which I use) uses gcc, and seems to work
quite acceptably well, though that doesn't mean too much since I haven't
actually compared the results with other compilers.


- Brooks




--
The "bmoses-nospam" address is valid; no unmunging needed.

Return to the comp.compilers page.
Search the comp.compilers archives again.