Consider adding webgl content to our PGO "run" training set (motionmark regression when PGO is on)
Categories
(Firefox Build System :: General, defect, P4)
Tracking
(Not tracked)
People
(Reporter: Gijs, Unassigned)
References
(Blocks 1 open bug)
Details
I did a short overview of our PGO setup for some of the performance folks today, and as part of that I did a trypush to compare PGO with no-PGO (we no longer run most/all(?) of our perf tests on non-PGO builds, so I just artificially "broke" the PGO instrumentation and profile-use steps to get data).
As expected, it showed PGO gets us very significant wins on a number of talos tests.
However, surprisingly, I ran some webgl benchmarks, and the PGO builds actually do worse than no-PGO builds on MotionMark.
I'm filing this bug to radar this and figure out if we can/should add some webgl work into the training/run set to fix this, or if there's something else we could/should do about this.
Note that this was quite a limited experiment to just get some data for my presentation, so I guess it's also possible something is wrong with my try run and this is a red herring - though it might still be worth evaluating getting webgl integrated into the pgo training set anyway.
|
||
Updated•4 years ago
|
Description
•