Re: open64 versus gcc

"Jonathan Thornburg -- remove -animal to reply" <jthorn@aei.mpg-zebra.de>
3 Dec 2006 21:30:19 -0500

          From comp.compilers

Related articles
[3 earlier articles]
Re: open64 versus gcc stevenb.gcc@gmail.com (Steven Bosscher) (2006-11-26)
Re: open64 versus gcc Sid-pasdespam.Touati@inria.fr (Sid Touati) (2006-11-27)
Re: open64 versus gcc drizzle76@gmail.com (dz) (2006-11-29)
Re: open64 versus gcc dnovillo@redhat.com (Diego Novillo) (2006-11-29)
Re: open64 versus gcc alewando@fala2005.com (A.L.) (2006-12-01)
Re: open64 versus gcc alewando@fala2005.com (A.L.) (2006-12-01)
Re: open64 versus gcc jthorn@aei.mpg-zebra.de (Jonathan Thornburg -- remove -animal to reply) (2006-12-03)
Re: open64 versus gcc dnovillo@redhat.com (Diego Novillo) (2006-12-03)
Re: open64 versus gcc lindahl@pbm.com (Greg Lindahl) (2006-12-03)
Re: open64 versus gcc bmoses-nospam@cits1.stanford.edu (Brooks Moses) (2006-12-03)
Re: open64 versus gcc mwso@earthlink.net (Gary Oblock) (2006-12-03)
Re: open64 versus gcc Sid-Touati@inria.fr (ST) (2006-12-06)
| List of all articles for this month |

From: "Jonathan Thornburg -- remove -animal to reply" <jthorn@aei.mpg-zebra.de>
Newsgroups: comp.compilers
Date: 3 Dec 2006 21:30:19 -0500
Organization: Compilers Central
References: 06-11-094 06-11-100 06-11-104 06-11-113 06-11-120 06-11-124 06-12-015
Keywords: arithmetic, performance, GCC

A.L. <alewando@fala2005.com> wrote:
[[about gcc]]
> One of the feature out of the list of "rich features" is that the
> results of numerical computations (such as inverting large matrix or
> solving large set of linear equations) strongly depends on activated
> options, especially optimization level.


This is just what you would expect for a computation which is
highly sensitive to small differences in roundoff errors.




> If you are a hobbyist, game programmer or GUI programmer, pretty
> likely gcc is good enough. If you do mission critical application,
> intensive number crunching or both, stay away from gcc.


I disagree. I've done heavy-duty number-crunching for a living for
over 20 years (simulating black hole collisions & similar problems).
I'm very happy with gcc, and would recommend it as an excellent compiler
for numerical work. It's not quite as good an optimizer as (say) the
Intel compiler, but it's usually within 10 or 20%. My colleagues and
I definitely find that the Intel compilers generate wrong code, or
abort with internal errors in valid code, considerably more often
than gcc. This is for both gcc 3.3* and gcc4, code usually a mix
of C, C++, Fortran 77, and Fortran 90.


ciao,


--
--
"Jonathan Thornburg -- remove -animal to reply" <jthorn@aei.mpg-zebra.de>
      Max-Planck-Institut fuer Gravitationsphysik (Albert-Einstein-Institut),
      Golm, Germany, "Old Europe" http://www.aei.mpg.de/~jthorn/home.html


Post a followup to this message

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