From octave-maintainers-request at bevo dot che dot wisc dot edu Wed Mar 31 11:55:00 2004
Subject: Re: memory alignment and blas
From: Paul Thomas <paulthomas2 at wanadoo dot fr>
To: Paul Kienzle <pkienzle at users dot sourceforge dot net>
Cc: octave-maintainers-list list <octave-maintainers at bevo dot che dot wisc dot edu>
Date: Wed, 31 Mar 2004 19:54:49 +0200

On an Athlon 1700 running RH9, I see at most a 10% worsening with 
DOFFSET = 1 for example.  Replacing octave_numeric_buffer with double, 
effects a 20% improvement in time.  With DOFFSET=0 and DALIGN=16, the 
execution time is almost exactly 50ms.

PaulT

Paul Kienzle wrote:

> I was playing a bit with memory alignment to see what
> effect it would have on the performance of DGEMM.
>
> For pentium II, I didn't see any performance effect.
> This is with Debian and atlas3-base installed.
>
> For PPC G4, unaligned memory shows 7.5x worse
> performance than aligned memory.  4 byte aligned
> memory shows < 10% performance drop.  8 byte
> alignment is indistinguishable from 128 byte alignment.
>
> As far as I can tell, the OS X allocator returns 8 byte
> aligned memory, so no special effort is required
> to support it.
>
> My octave version is too out of date on MIPS to give
> any performance numbers.
>
> I'm attaching the code I used to test this.
>
> Paul Kienzle
> pkienzle at users dot sf dot net
>