ATLAS on AMD Geode LX?

[M. Edward (Ed) Borasky]

I'm getting one of the Give 1 Get 1 laptops (One Laptop Per Child),
which has an AMD Geode LX processor in it. As far as I can tell, it's an
Athlon (MMX plus 3DNow! only). Are there special assembler kernels for
the Geode, or should the Athlon ATLAS be "as good as it gets?"

-------------------------------------------------------------------------
SF.Net email is sponsored by: The Future of Linux Business White Paper
from Novell.  From the desktop to the data center, Linux is going
mainstream.  Let it simplify your IT future.
http://altfarm.mediaplex.com/ad/ck/8857-50307-18918-4
_______________________________________________
Math-atlas-devel mailing list
Math-atlas-devel@...
https://lists.sourceforge.net/lists/listinfo/math-atlas-devel

[Tim Mattox]

Depending on what you are wanting to use the BLAS for, you could enable
the use of the 3DNow! optimized assembly SGEMM.  Make sure you understand
the disclaimers associated with using the non-IEEE compliant 3DNow!
instructions.
(no denormals for one...)   You can find a good disclaimer here:
http://www.cs.utk.edu/~rwhaley/ATLAS/errata3.4.2.html#3DNowNotSSE
But if you are willing to deal with that (and single precision), then the
ATL_smm_3dnow_90.c SGEMM will likely get you the highest performance
on that machine.

However, I don't think its an Athlon core... it could be based on the
K6-2 or K6-3 series, since it appears that AMD is claiming that one of the big
enhancements to the Geode NX is that it is based on the 7th generation CPU
core of the Athlon....  I optimized the code in ATL_smm_3dnow_90.c for
the Athlon,
so I don't know if it will do all that well on a K6-2/3 pipeline.
It's still worth a try.

On Dec 1, 2007 4:48 AM, M. Edward (Ed) Borasky <znmeb@...> wrote:

> I'm getting one of the Give 1 Get 1 laptops (One Laptop Per Child),
> which has an AMD Geode LX processor in it. As far as I can tell, it's an
> Athlon (MMX plus 3DNow! only). Are there special assembler kernels for
> the Geode, or should the Athlon ATLAS be "as good as it gets?"
>
> -------------------------------------------------------------------------
> SF.Net email is sponsored by: The Future of Linux Business White Paper
> from Novell.  From the desktop to the data center, Linux is going
> mainstream.  Let it simplify your IT future.
> http://altfarm.mediaplex.com/ad/ck/8857-50307-18918-4
> _______________________________________________
> Math-atlas-devel mailing list
> Math-atlas-devel@...
> https://lists.sourceforge.net/lists/listinfo/math-atlas-devel
>

--
Tim Mattox, Ph.D. - http://homepage.mac.com/tmattox/
 tmattox@... || timattox@...
    I'm a bright... http://www.the-brights.net/

-------------------------------------------------------------------------
SF.Net email is sponsored by: The Future of Linux Business White Paper
from Novell.  From the desktop to the data center, Linux is going
mainstream.  Let it simplify your IT future.
http://altfarm.mediaplex.com/ad/ck/8857-50307-18918-4
_______________________________________________
Math-atlas-devel mailing list
Math-atlas-devel@...
https://lists.sourceforge.net/lists/listinfo/math-atlas-devel

[M. Edward (Ed) Borasky]

Tim Mattox wrote:

> Depending on what you are wanting to use the BLAS for, you could enable
> the use of the 3DNow! optimized assembly SGEMM.  Make sure you understand
> the disclaimers associated with using the non-IEEE compliant 3DNow!
> instructions.
> (no denormals for one...)   You can find a good disclaimer here:
> http://www.cs.utk.edu/~rwhaley/ATLAS/errata3.4.2.html#3DNowNotSSE
> But if you are willing to deal with that (and single precision), then the
> ATL_smm_3dnow_90.c SGEMM will likely get you the highest performance
> on that machine.
>
> However, I don't think its an Athlon core... it could be based on the
> K6-2 or K6-3 series, since it appears that AMD is claiming that one of the big
> enhancements to the Geode NX is that it is based on the 7th generation CPU
> core of the Athlon....  I optimized the code in ATL_smm_3dnow_90.c for
> the Athlon,
> so I don't know if it will do all that well on a K6-2/3 pipeline.
> It's still worth a try.

Well ... it turns out that i386 Atlas 3.6.0 is already installed on the
system, right out of Fedora Core 7. I haven't figured out what they're
doing with Atlas or the BLAS -- this is supposed to be a kids' machine.
:) I won't get an actual hardware unit until mid-January, so I have
plenty of time to do research before that.

-------------------------------------------------------------------------
SF.Net email is sponsored by: The Future of Linux Business White Paper
from Novell.  From the desktop to the data center, Linux is going
mainstream.  Let it simplify your IT future.
http://altfarm.mediaplex.com/ad/ck/8857-50307-18918-4
_______________________________________________
Math-atlas-devel mailing list
Math-atlas-devel@...
https://lists.sourceforge.net/lists/listinfo/math-atlas-devel

[Clint Whaley]

Guys,

>I'm getting one of the Give 1 Get 1 laptops (One Laptop Per Child),
>which has an AMD Geode LX processor in it. As far as I can tell, it's an
>Athlon (MMX plus 3DNow! only). Are there special assembler kernels for
>the Geode, or should the Athlon ATLAS be "as good as it gets?"

>From looking at the specs, I'm not sure what its pedigree actually is,
but ATLAS should use the best kernel automatically.  You'll have to do
the full search, since ATLAS doesn't have any arch defs for it.  In fact,
I no longer have access to any classic athlons (i.e. K7, not hammer
descendence), so ATLAS has no arch defs for non-hammer athlons.

>From the specs, it looks like Geode NX is a descendent of the Athlon XP,
but no real info is given on the LX . . .
instructions.

My guess is that the excellent athlon/x87 kernels inspired or written by
Julien Ruhe will do very well (they got something like 90% of peak on
classic athlons).

>http://www.cs.utk.edu/~rwhaley/ATLAS/errata3.4.2.html#3DNowNotSSE
>But if you are willing to deal with that (and single precision), then the
>ATL_smm_3dnow_90.c SGEMM will likely get you the highest performance
>on that machine.

Yeah, I don't recommend 3DNow due to the overflow/underflow issues.  But,
if single precision performance is you main goal, and you are sure you
don't have these accuracy problems, you'll need to tell ATLAS that you
want to use the 3DNow for flops despite the accuracy problems.
I'm not sure, but I think adding
   -D c -DATL_3DNow -D c -DATL_3DNowFLOPS
to configure should make ATLAS try the 3DNow kernels, and use them if they
are faster.

>Well ... it turns out that i386 Atlas 3.6.0 is already installed on the
>system, right out of Fedora Core 7. I haven't figured out what they're
>doing with Atlas or the BLAS -- this is supposed to be a kids' machine.
>:) I won't get an actual hardware unit until mid-January, so I have
>plenty of time to do research before that.

ATLAS provides fast linear algebra, which helps with a lot of stuff.  For
instance, I know Apple uses it in OS X to speedup their spam filtering, etc.
Maybe something like that going on here . . .

I doubt 3.8 will be faster than 3.6 on this old architecture, but you may
nonetheless get speedup, since preinstalls often use generic architecture
settings, rather than being tuned for a particular system.

Regards,
Clint

**************************************************************************
** R. Clint Whaley, PhD ** Assist Prof, UTSA ** www.cs.utsa.edu/~whaley **
**************************************************************************

-------------------------------------------------------------------------
SF.Net email is sponsored by: The Future of Linux Business White Paper
from Novell.  From the desktop to the data center, Linux is going
mainstream.  Let it simplify your IT future.
http://altfarm.mediaplex.com/ad/ck/8857-50307-18918-4
_______________________________________________
Math-atlas-devel mailing list
Math-atlas-devel@...
https://lists.sourceforge.net/lists/listinfo/math-atlas-devel