|
|
Blitz Devel : |
From: Steve Karmesin (karmesin_at_[hidden])
Date: 1998-08-14 11:17:18
At 11:58 AM 8/14/98 -0400, Todd Veldhuizen wrote:
>> Performance results for the above expressions, using N=100000:
>>
>> Unfused Fused Chunky
>> T3E/KCC 27.4 43.2 34.0 (mcurie.nersc.gov)
>> alpha/DECcxx 8.7 12.7 11.9 (hgar1.cwru.edu)
>> i686/egcs 14.4 10.4 19.0 (n2001.lbl.gov)
>> O2000/KCC 14.4 22.3 18.4 (bluemountain.acl.lanl.gov)
>>
>
>Err, these numbers are Mflops/s. Sorry about that omission.
>
>Cheers,
>Todd
>
This is very interesting, because this is precisely the optimization we are
chasing with the threaded infrastructure of POOMA II. With that, we plan
to break up input expressions into chunks and schedule the resulting iterates.
There are still some important uses for fuse though. The main one in my
mind comes from the fact that phrasing things in whole-array notation can
lead to promoting things from scalars to arrays.
It could be that the best way to handle that is with stencil objects rather
than with fuse though.
-srk
--------------------- blitz-dev list --------------------------------
* To subscribe/unsubscribe: mail to majordomo_at_[hidden], with
"subscribe blitz-dev" or "unsubscribe blitz-dev" in the body of the message
* Blitz++ web page: http://oonumerics.org/blitz/