RE: 3dnow

["M. Edward Borasky" <znmeb@aracnet.com>]

No, I haven't tried that, but I can certainly do it and compare the results.
I'll post the comparison when I do run it; it'll probably be done tomorrow.
--
M. Edward (Ed) Borasky, Chief Scientist, Borasky Research
http://www.borasky-research.net  http://www.aracnet.com/~znmeb
mailto:znmeb@borasky-research.com  mailto:znmeb@aracnet.com

Q: How do you get an elephant out of a theatre?
A: You can't. It's in their blood.

> -----Original Message-----
> From: R Clint Whaley [mailto:rwhaley@cs.utk.edu]
> Sent: Saturday, June 30, 2001 10:31 AM
> To: znmeb@aracnet.com
> Cc: atlas-comm@cs.utk.edu
> Subject: RE: 3dnow
>
>
> >2. I downloaded the Atlas 3.3.1 developer snapshot and built it using GCC
> >3.0 on the aforementioned Athlon, selecting the "3DNow2" option
> and letting
> >it grind overnight through all of its tests. The "SUMMARY.LOG" file is
> >attached; I have all the other files if anyone is interested. It
> takes about
> >eight hours at 1.333 GHz, BTW.
>
> Have you tried to say "yes" to architectural defaults?  My guess
> is that it
> will get you better performance than what you are reporting.  The search
> script needs some hand-holding on Athlons to get good results.  This is
> due to the fact that the hardware uses seperate multiply and add units,
> but due to the register pressure and renaming stuff, you nonetheless need
> to program matmul like you have a combined muladd FPU.  Since
> ATLAS detects
> the seperate multiply/add unit, the gemm search is non-optimal.
>
> Cheers,
> Clint
>
>