Re: NEC-LIST: LAPACK LU (was Re: NEC-LIST: Xeon vs Pentium II - 450)

I have put up a new version of the note at:

http://www-users.york.ac.uk/~idf1/nec2blas

in PDF, postscript and html.

This fixes the URL for the Intel Performance Math Library and adds
some notes on using the ASCI Red BLAS with egcs. Basically don't use
the -malign-double option with egcs when compiling NEC2 against the
ASCI Red Libraries. The speed up from using these BLAS libraries with
egcs is then comparable to that using the PGI compiler, i.e. a factor
> 5 for a 2000 segment model.

Since this really does look like an alignment problem with egcs if is
difficult to determine the best compilation options. It is possible
for the performance to vary dramatically from run to run with the same
executable, though the few tests I've done with egcs not using the
-malign-double flag were reasonable consistent. Posts on the egcs
mailing list suggest that the alignment issue is currently being
looked at.

Ian

-- 
Dr Ian David Flintoft           Email:      idf1_at_ohm.york.ac.uk
Applied Electromagnetic Group   Tel:        +44 1904 432391    
Department of Electronics       Fax:        +44 1904 433224    
University of York  
Heslington          
YORK, UK               <  EMC Aspects of Radio-based Mobile  >
YO10 5DD               <     Telecommunication Systems.      >