« Return to Thread: results from tuning LAPACK
Re: results from tuning LAPACK
by M. Edward (Ed) Borasky
::
Rate this Message:
On Sun, 2008-08-10 at 07:50 -0500, Clint Whaley wrote:
> Guys,
>
> My paper on the benefits of empirically tuning LAPACK's ILAENV-controlled NB
> has been accepted for CANA'08. You can see a draft of it at:
> http://www.cs.utsa.edu/~whaley/papers/lanb.pdf
>
> It shows the advantage you get from tuning NB for several different BLAS
> an a couple of platforms. Depending on problem size and architecture,
> you can be leaving as much as 30% (or 75%, if on the Itanium :) on the table
> if you use stock lapack's ILAENV. It also highlights two new tools available
> in ATLAS (lapack timer/tuner and a tool that takes ATLAS timer output,
> does some simple statistics, and prints it out in a spreadsheet friendly
> format).
>
> We can apply this tuning to pretty much any blocked LAPACK routine. ATLAS
> presently handles tuning the factorizations. My guess is that the next
> most important routines are the eigencodes. Does anyone want to nominate
> particular routines that they use a lot?
Real symmetric eigensolvers are a big workhorse in lots of areas ...
even undergraduate chemistry. :) Singular Value Decomposition is another
workhorse -- gotta have it.
Question -- is anybody putting this much effort into tuning *sparse*
linear algebra? Something approaching ATLAS performance on huge sparse
matrix-vector multiplies would give significant improvements in many
areas, including my personal favorite, Markov chains.
>
> Cheers,
> Clint
>
> **************************************************************************
> ** R. Clint Whaley, PhD ** Assist Prof, UTSA ** www.cs.utsa.edu/~whaley **
> **************************************************************************
>
> -------------------------------------------------------------------------
> This SF.Net email is sponsored by the Moblin Your Move Developer's challenge
> Build the coolest Linux based applications with Moblin SDK & win great prizes
> Grand prize is a trip for two to an Open Source event anywhere in the world
> http://moblin-contest.org/redirect.php?banner_id=100&url=/
> _______________________________________________
> Math-atlas-devel mailing list
> Math-atlas-devel@...
> https://lists.sourceforge.net/lists/listinfo/math-atlas-devel--
M. Edward (Ed) Borasky
ruby-perspectives.blogspot.com
"A mathematician is a machine for turning coffee into theorems." --
Alfréd Rényi via Paul Erdős
-------------------------------------------------------------------------
This SF.Net email is sponsored by the Moblin Your Move Developer's challenge
Build the coolest Linux based applications with Moblin SDK & win great prizes
Grand prize is a trip for two to an Open Source event anywhere in the world
http://moblin-contest.org/redirect.php?banner_id=100&url=/
_______________________________________________
Math-atlas-devel mailing list
Math-atlas-devel@...
https://lists.sourceforge.net/lists/listinfo/math-atlas-devel
> Guys,
>
> My paper on the benefits of empirically tuning LAPACK's ILAENV-controlled NB
> has been accepted for CANA'08. You can see a draft of it at:
> http://www.cs.utsa.edu/~whaley/papers/lanb.pdf
>
> It shows the advantage you get from tuning NB for several different BLAS
> an a couple of platforms. Depending on problem size and architecture,
> you can be leaving as much as 30% (or 75%, if on the Itanium :) on the table
> if you use stock lapack's ILAENV. It also highlights two new tools available
> in ATLAS (lapack timer/tuner and a tool that takes ATLAS timer output,
> does some simple statistics, and prints it out in a spreadsheet friendly
> format).
>
> We can apply this tuning to pretty much any blocked LAPACK routine. ATLAS
> presently handles tuning the factorizations. My guess is that the next
> most important routines are the eigencodes. Does anyone want to nominate
> particular routines that they use a lot?
Real symmetric eigensolvers are a big workhorse in lots of areas ...
even undergraduate chemistry. :) Singular Value Decomposition is another
workhorse -- gotta have it.
Question -- is anybody putting this much effort into tuning *sparse*
linear algebra? Something approaching ATLAS performance on huge sparse
matrix-vector multiplies would give significant improvements in many
areas, including my personal favorite, Markov chains.
>
> Cheers,
> Clint
>
> **************************************************************************
> ** R. Clint Whaley, PhD ** Assist Prof, UTSA ** www.cs.utsa.edu/~whaley **
> **************************************************************************
>
> -------------------------------------------------------------------------
> This SF.Net email is sponsored by the Moblin Your Move Developer's challenge
> Build the coolest Linux based applications with Moblin SDK & win great prizes
> Grand prize is a trip for two to an Open Source event anywhere in the world
> http://moblin-contest.org/redirect.php?banner_id=100&url=/
> _______________________________________________
> Math-atlas-devel mailing list
> Math-atlas-devel@...
> https://lists.sourceforge.net/lists/listinfo/math-atlas-devel
M. Edward (Ed) Borasky
ruby-perspectives.blogspot.com
"A mathematician is a machine for turning coffee into theorems." --
Alfréd Rényi via Paul Erdős
-------------------------------------------------------------------------
This SF.Net email is sponsored by the Moblin Your Move Developer's challenge
Build the coolest Linux based applications with Moblin SDK & win great prizes
Grand prize is a trip for two to an Open Source event anywhere in the world
http://moblin-contest.org/redirect.php?banner_id=100&url=/
_______________________________________________
Math-atlas-devel mailing list
Math-atlas-devel@...
https://lists.sourceforge.net/lists/listinfo/math-atlas-devel
« Return to Thread: results from tuning LAPACK
| Free embeddable forum powered by Nabble | Forum Help |
