Speed up GenEvent::attribute_names

Profiling indicated GenEvent::attribute_names was taking ~85% of runtime when reading a HepMC2 ASCII file. Replacing a linear-time search with a log-time search speeds up the function 150x in my tests.

added 1 commit

• 19a6bbbd - Linear -> log search time for GenEvent::attribute_names

Compare with previous version

added 1 commit

• 98664fb5 - Tab formatting cleanup

Compare with previous version

mentioned in commit 970c64f6

merged

Hi @ponyisi ,

first of all, thanks a lot for the contribution. Second, that would be interesting to see what kind of tests you are running. Maybe we could make some further improvements.

Best regards, Andrii

Hi Andrii,

For my test I read a fairly generic W -> mu nu HepMC2 file (just calling read_event). I profiled it with VTune (using the standard CERN CVMFS installation) and just looked at what the leading functions for time use were. After this patch, it seems the largest single CPU user for ReaderAsciiHepMC2 is all the atof calls, which I guess are unavoidable.

Best, Peter

Hi Peter,

Ok, I see. Thanks a lot for the information. We are using a combination of atof and strchr, so one could replace that with strtof or so. I've looked there some time ago, but the improvements were minuscule as atof was not the leading function. Maybe after your fix this will change.

Best regards, Andrii