Hide Forgot
Arpack currently builds against reference LAPACK. I believe it should be built against ATLAS instead, which should bring a notable speed benefit.
Are there inline optimizations in atlas that lapack misses which arpack makes use of? Otherwise there should not be any speedups compared to simply replacing lapack with atlas or atlas-sse2.
Among others, yes. Normally one gets around an order or more speed when using a tuned library. See for instance http://www.ats.ucla.edu/clusters/common/software/libraries/lapack_benchmark.htm A patch to follow up soon.
Created attachment 490958 [details] Patch against rawhide spec Switches to compilation against ATLAS and also fixes some bugs in the spec file.
(In reply to comment #2) > Among others, yes. Normally one gets around an order or more speed when using a > tuned library. See for instance > > http://www.ats.ucla.edu/clusters/common/software/libraries/lapack_benchmark.htm Jussi, thanks for the patch. But please consider the following: The *GESV solvers are typically never *inlined* in any binding or library BLAS implementation I have seen, they are just calls to external functions. This means that you will still get the speed benefits by simply swaping lapack with the proper atlas or any other blas library. There is no benefit to compiling against ATLAS. Situations where this is not true are inline DFT or expression template coding which is not the case here. I wouldn't mind being proven wrong, just use gcc -S against a minimal program with a single SGSEV call and compare building against LAPACK vs ATLAS headers.
Is this still an open issue or should we close it?
I haven't had time to look at this. Anyway, whenever BLAS functions are used, using an optimized library instead of the reference version results in notable speed benefits. This would just need to be benchmarked, if there is a benefit. However, I don't think one would lose anything by replacing the link against LAPACK to the version provided by ATLAS...
We don't disagree, I use ATLAS myself. But the speed benefits come from the run-time resolution of where the library functions are offered from, not against what library you built against. Anyway, I'll flag this as needmoreinfo, and when/if you want to check it, please confirm or not the above claim. :)
... but it really isn't resolved at runtime. For instance the difference in Octave linked against reference LAPACK vs that linked with ATLAS is orders of magnitude.
This message is a notice that Fedora 15 is now at end of life. Fedora has stopped maintaining and issuing updates for Fedora 15. It is Fedora's policy to close all bug reports from releases that are no longer maintained. At this time, all open bugs with a Fedora 'version' of '15' have been closed as WONTFIX. (Please note: Our normal process is to give advanced warning of this occurring, but we forgot to do that. A thousand apologies.) Package Maintainer: If you wish for this bug to remain open because you plan to fix it in a currently maintained version, feel free to reopen this bug and simply change the 'version' to a later Fedora version. Bug Reporter: Thank you for reporting this issue and we are sorry that we were unable to fix it before Fedora 15 reached end of life. If you would still like to see this bug fixed and are able to reproduce it against a later version of Fedora, you are encouraged to click on "Clone This Bug" (top right of this page) and open it against that version of Fedora. Although we aim to fix as many bugs as possible during every release's lifetime, sometimes those efforts are overtaken by events. Often a more recent Fedora release includes newer upstream software that fixes bugs or makes them obsolete. The process we are following is described here: http://fedoraproject.org/wiki/BugZappers/HouseKeeping