[parsec-users] performance improvement for fluidanimate
Christian A Bienia
cbienia at CS.Princeton.EDU
Mon Jun 28 11:04:40 EDT 2010
I'm very interested in all patches, including your optimization for fluidanimate. :) I rewrote the program for the next version of PARSEC, so your change might not apply anymore. Please send me a patch and I'll take a look.
----- Ursprüngliche Mail -----
Von: "Julia Fedorova" <Julia.Fedorova at intel.com>
An: "PARSEC Users" <parsec-users at lists.cs.princeton.edu>
Gesendet: Montag, 28. Juni 2010 07:14:55 GMT +01:00 Amsterdam/Berlin/Bern/Rom/Stockholm/Wien
Betreff: [parsec-users] performance improvement for fluidanimate
Hi Chris, all
I have some performance improvement for fluidanimante application.
They are some small code changes in ComputeForces (in case of serial version) and in
ComputeForcesMT & ComputeDensitiesMT functions (in pthreads version)
that lead to ~ 18% speed up as measured for serial and pthreads, native input with 500 frames on Xeon Core i7 single socket platform.
Gcc-serial (indeed 1 thread):
Initial – 358.6 sec
Optimized – 293.4 sec
gcc-pthreads, 4 threads:
Initial – 109.7 sec
Optimized – 89 sec
It is related to excessive branch misprediction cost in the code and putting additional “if” upper by the code flow gives visible performance boost.
Although above numbers are for the intel platform – I believe the changes will bring benefit for other platforms/CPU also – as they minimize wasted work that created by branch misprediction.
Is there interest to such changes?
I will send the patches then.
Regards, Julia Fedorova
Closed Joint Stock Company Intel A/O
Registered legal address: Krylatsky Hills Business Park,
17 Krylatskaya Str., Bldg 4, Moscow 121614,
This e-mail and any attachments may contain confidential material for
the sole use of the intended recipient(s). Any review or distribution
by others is strictly prohibited. If you are not the intended
recipient, please contact the sender and delete all copies.
parsec-users mailing list
parsec-users at lists.cs.princeton.edu
More information about the parsec-users