> If performance of NEC is what you want, replace piece of Fortran code
> which does the matrix factorization with calls to an optimized math-
> library (e.g. Lapack). With the (free download) Intel version you get
> a performance boost of about 3.
>
> The difference between Linux and MsW is much smaller, although like
> Dave, I would opt for Linux (and then use Greg Henry's math library.)
>
> Actually, maybe someone already has rewritten the (small) piece of
> factorization code for the speed improvement. Please let us know!
jep. Basically you need to replace one call of FACTRS with ZGETRF, and
one call of SOLVE with ZGETRS. Should be pretty forward. I observed
(cf. my paper at ACES: for 1944 unknowns 378s original, 72.76 Lapack's
LU) a speedup of more than 5 on AIX/IBM SP with the ESSL library. On a
DEC Alpha with optimized ZGEMM speed up was at least 3. I don't use
INTELs so I can't tell what happens there. I would be very interested
though to hear about it.
Interestingly, ZGETRS is able to solve for a TRANSPOSED matrix, so you
don't even have to transpose the matrix before factorizing it. And as
the factorizing is the same regardless of Z^T or Z, you save the time
for the transpose. For more info about ZGETRF look at the Users' guide
to LAPACK and the various Lapack Working Notices available at
www.netlib.org
cheers
juergen
Received on Thu Mar 25 1999 - 16:13:14 EST
This archive was generated by hypermail 2.2.0 : Sat Oct 02 2010 - 00:10:39 EDT