Failure to move invariant WL out of FOR loop kills ipbb performance
| Bugzilla Link | 524 |
| Created on | Jul 14, 2009 04:10 |
| Version | 1.00beta |
| OS | Linux |
| Architecture | PC |
| Attachments | ipbb5.slow.sac, bug.funcond.sac, bug.compress.breaks.sac, iotaonly7.sac, crud.slow, crud.fast |
Extended Description
Created an attachment (id=551) Shorter source code to reproduce fault I've been chasing a problem with the performance of a Boolean-Boolean inner product code, apex/ipbb/ipbb.sac. Although I think there are some problems with my extrema code in this area, due to differences in performance in the current system, the attached shorter example has the interesting property that the compiler does not appear to be moving a loop-invariant Wl out of a FOR-loop. I say this because, if I move the loop out by hand, everything runs bags faster in both the extrema and non-extrema world. I am off to bed, so have not looked into this yet... This all with Build #16193. With xrow inside FOR-loop: sac2c -O3: ipbb5.slow.sac.exe.O3.papiex.rattler.28893:PAPI_TOT_INS: 99065594 sac2c -O3 -extrema -nowlf -doswlf: ipbb5.slow.sac.exe.swlf.papiex.rattler.28928:PAPI_TOT_INS: 1003065073 With xow outside FOR-loop: sac2c -O3: ipbb5.slow.sac.exe.O3.papiex.rattler.28998:PAPI_TOT_INS: 765065354 sac2c -O3 -extrema -nowlf -doswlf: ipbb5.slow.sac.exe.swlf.papiex.rattler.28970:PAPI_TOT_INS: 765065175
To upload designs, you'll need to enable LFS and have an admin enable hashed storage. More information