Re: [Fink-users] IBM's compiler

[Charles Williams]

I tried out IBM's f77 compiler a while back on my 12" PowerBook G4, and  
I was pretty impressed for the simple test problem I tried.  The same  
problem using g77 took about 2.6 times as long to run.  Compilation is  
much slower, because it's probably doing more optimization.  I haven't  
yet tried to recompile atlas, and I have a suspicion that it won't  
produce a great benefit, since atlas does so much optimization on its  
own.  The problem I tried was a simple linear algebra problem (Cholesky  
decomposition of a positive definite matrix to create an inverse,  
multiplication of the inverse times the original matrix, and then  
computing the trace and the sum of the off-diagonal magnitudes of the  
result).  For reference (when I get the time), I should do the test  
problem on my new G5 using both compilers, and then compare the results  
with those obtained using BLAS calls with both the fink-installed atlas  
and Apple's vecLib (I haven't really tested out vecLib yet, so I don't  
know how it performs).  For true completeness, I should also try using  
vac and xlf to build a different version of atlas and test that as  
well.  If I manage to get all of this done sometime soon (in my  
non-existent spare time), I'll post the results for both the G4  
PowerBook and the dual G5.

Charles

On Wednesday, October 8, 2003, at 02:46  PM,  
[EMAIL PROTECTED] wrote:

Message: 3
Date: Wed, 8 Oct 2003 17:33:11 +0200
Subject: Re: [Fink-users] IBM's compiler
Cc: Timothy Reaves <[EMAIL PROTECTED]>,
        [EMAIL PROTECTED]
To: chuckr <[EMAIL PROTECTED]>
From: Martin Costabel <[EMAIL PROTECTED]>
On mercredi, oct 8, 2003, at 15:06 Europe/Paris, chuckr wrote:
[]
I'd be very interested in what config changes are needed.

Thanks.

Could you supply a link to those compilers?  I'd take a look ...
Long URL:

http://www14.software.ibm.com/webapp/download/
search.jsp?q=compiler&k=all&cat=&sb=p&go=y&sr=1&rs=&S_TACT=&S_CMP=&s=&x 
=
10&y=13&pf=Mac+OS+X&dt=BETA

On a G4, I was somewhat disappointed and gave up after having spent
quite some time to change the atlas and scilab packages so they would
use IBM's Fortran compiler. The compilation was much slower than with
gcc, and the test results showed basically no difference. Probably, in
order to make the compilations go through, I had to do things that
prevented any substantial optimization. I don't remember what it was I
had to do, but there were even weird things like some source files
(from scilab) with DOS line breaks that gave rise to strange error
messages. Also, for atlas, the optimization done by the package seems
to be so efficient that the compiler is not of much importance. On a
G5, the story is probably quite different.
--
Martin


--__--__--

Charles A. Williams
Dept. of Earth & Environmental Science
Science Center 2C01
Rensselaer Polytechnic Institute
Troy, NY  12180
Phone:  (518) 276-8463
FAX:            (518) 276-2012
e-mail: [EMAIL PROTECTED]


-------------------------------------------------------
This SF.net email is sponsored by: SF.net Giveback Program.
SourceForge.net hosts over 70,000 Open Source Projects.
See the people who have HELPED US provide better services:
Click here: http://sourceforge.net/supporters.php
_______________________________________________
Fink-users mailing list
[EMAIL PROTECTED]
https://lists.sourceforge.net/lists/listinfo/fink-users