[Impeller] Investigate "DlCanvas implementation wrapping Aiks canvas" 90th percentile increase

[gaaclarke]

flutter/engine#44248 seems to have improved the average rasterizer time for the backdrop filter benchmark, but potentially raised the floor on the gallery 90th percentile test (albeit the attribution is tenuous). Can we double check the implementation to make sure there isn't something we missed that might negatively effect the 90th percentile

backdrop filter average raster time benchmark drop

gallery 90th percentile increase

(notice that the floor never reaches 10ms around the time that PR landed)

cc @dnfield

[gaaclarke]

I think it's easiest if @dnfield just gives it a look to see if there is some easy tweak we could make. He won't be available for a while though.

[jonahwilliams]

Here are some CPU usage numbers from the iOS:

https://flutter-flutter-perf.skia.org/e/?formulas=filter(%22sub_result%3Daverage_cpu_usage%22)&selected=commit%3D36135%26name%3D%252Carch%253Dintel%252Cbranch%253Dmaster%252Cconfig%253Ddefault%252Cdevice_type%253DiPhone_11%252Cdevice_version%253Dnone%252Chost_type%253Dmac%252Csub_result%253Daverage_cpu_usage%252Ctest%253Ddraw_points_perf_ios__timeline_summary%252C

drawPoints dropped by 50%, a few regressed, a few improved. I think we should track down the regressions.

[jason-simmons]

Some other benchmarks regressed noticeably in that engine roll.

For example, this query shows post_backdrop_filter_perf_ios__timeline_summary/average_cpu_usage up by 25% and backdrop_filter_perf_ios__timeline_summary/average_gpu_usage up by 40%:

https://flutter-flutter-perf.skia.org/e/?begin=1688572077&end=1691427625&keys=X8a443eafed2166315999f8d7eb6fd852&requestType=0&selected=commit%3D36136%26name%3D%252Carch%253Dintel%252Cbranch%253Dmaster%252Cconfig%253Ddefault%252Cdevice_type%253DiPhone_11%252Cdevice_version%253Dnone%252Chost_type%253Dmac%252Csub_result%253Daverage_cpu_usage%252Ctest%253Dpost_backdrop_filter_perf_ios__timeline_summary%252C&xbaroffset=36136

[gaaclarke]

@zanderso @chinmaygarde what do you want to do about that PR? Looks like its benefits are complicated and I looked at the PR and it still had some unanswered comments on it. Dan isn't going to be available for a minute to investigate.

[zanderso]

Let's revert. @dnfield can re-land next week after investigating the regressions noted here.

[zanderso]

The improvement in the drawPoints benchmark is so large as to be suspicious. We should verify that it was still rendering something with the change.

[chinmaygarde]

The revert has landed and I have filed an issue tracking re-landing with references #132416. Closing this.

[github-actions]

This thread has been automatically locked since there has not been any recent activity after it was closed. If you are still experiencing a similar issue, please open a new bug, including the output of flutter doctor -v and a minimal reproduction of the issue.