Issue metadata
Sign in to add a comment
|
35-70% regression in draw properties computation time |
||||||||||||||||||||||||
Issue descriptionDraw property computation time as measured by the Compositing.Renderer.LayerTreeImpl.CalculateDrawPropertiesUs metric regressed by ~35% at the 95th percentile on Android and Windows Canary, in the range https://chromium.googlesource.com/chromium/src/+log/60.0.3077.0..60.0.3081.0?pretty=fuller&n=10000 Bisecting locally using cc_perftest CalcDrawPropsTest.TenTen, there's a ~70% regression from https://codereview.chromium.org/2840433002 ("Move LayerImpl's bounds_delta to property trees"). The regression is probably from LayerImpl::bounds() becoming more expensive because it now calls ViewportBoundsDelta(), which in turn makes calls that involve LayerById std::unordered_map lookups.
,
May 3 2017
The 35% regression on UMA is from looking at the Compositing.Renderer.LayerTreeImpl.CalculateDrawPropertiesUs timeline (see http://go/windows-cdp-uma). The 70% regression on the cc_perftests came from the result of running: cc_perftests --gtest_filter=CalcDrawPropsTest.TenTen --gtest_repeat=10 and taking the average of those results.
,
May 3 2017
ajuma@ was running cc_perftests --gtest_filter=CalcDrawPropsTest.TenTen locally, with some added code to do repeat I think. (also whenever I see QQ I think of Tencent)
,
May 3 2017
Thanks Ali and Wei! QQ also kinda looks like two crying eyes which I don't like. 😠|
|||||||||||||||||||||||||
►
Sign in to add a comment |
|||||||||||||||||||||||||
Comment 1 by pdr@chromium.org
, May 3 2017Status: Duplicate (was: Assigned)