Quoting Chris Wilson (2018-06-04 13:59:12) > Quoting Tvrtko Ursulin (2018-06-04 13:52:39) > > From: Tvrtko Ursulin <tvrtko.ursulin@xxxxxxxxx> > > > > As Chris has discovered on his Ivybridge, and later automated test runs > > have confirmed, on most of our platforms hrtimer faced with heavy GPU load > > ca occasionally become sufficiently imprecise to affect PMU sampling > > s/ca/can/ > > > calculations. > > > > This means we cannot assume sampling frequency is what we asked for, but > > we need to measure the interval ourselves. > > > > This patch is similar to Chris' original proposal for per-engine counters, > > but instead of introducing a new set to work around the problem with > > frequency sampling, it swaps around the way internal frequency accounting > > is done. Instead of accumulating current frequency and dividing by > > sampling frequency on readout, it accumulates frequency scaled by each > > period. > > My ulterior motive failed to win favour ;) > > > Testcase: igt/perf_pmu/*busy* # snb, ivb, hsw > > Signed-off-by: Tvrtko Ursulin <tvrtko.ursulin@xxxxxxxxx> > > Suggested-by: Chris Wilson <chris@xxxxxxxxxxxxxxxxxx> > > Cc: Chris Wilson <chris@xxxxxxxxxxxxxxxxxx> > Reviewed-by: Chris Wilson <chris@xxxxxxxxxxxxxxxxxx> I should point out that even with this fix (or rather my patch), we can still see errors of 25-30us, enough to fail the test. However, without the fix our errors can be orders of magnitude worse (e.g. reporting 80us busy instead of 500us). -Chris _______________________________________________ Intel-gfx mailing list Intel-gfx@xxxxxxxxxxxxxxxxxxxxx https://lists.freedesktop.org/mailman/listinfo/intel-gfx
