All graphs for this bug:
  https://chromeperf.appspot.com/group_report?bug_id=910352

(For debugging:) Original alerts at time of bug-filing:
  https://chromeperf.appspot.com/group_report?sid=340182ec589e8100bf4e78a2c7119d39119ece7949c0ff4c7f02b4d0e96dce53


Bot(s) for this bug's original alert(s):

mac-10_12_laptop_low_end-perf
mac-10_13_laptop_high_end-perf

memory.desktop - Benchmark documentation link:
  None

system_health.memory_desktop - Benchmark documentation link:
  https://bit.ly/system-health-benchmarks

📍 Pinpoint job started.
https://pinpoint-dot-chromeperf.appspot.com/job/1239c352140000

📍 Found a significant difference after 1 commit.
https://pinpoint-dot-chromeperf.appspot.com/job/1239c352140000

gpu: Disable CCPR for OOP raster. by khushalsagar@chromium.org
https://chromium.googlesource.com/chromium/src/+/193eda157bdb1c66d9afdc9c54a36e0ccc249293
memory:chrome:all_processes:reported_by_chrome:skia:effective_size: 2.327e+06 → 5.973e+06 (+3.646e+06)

Understanding performance regressions:
  http://g.co/ChromePerformanceRegressions

Benchmark documentation link:
  https://bit.ly/system-health-benchmarks

All that patch did was disable CCPR using GrContextOptions. We had an issue with losing some caching with CCPR, which probably lowered skia's memory usage, and since the alternate path renderer is getting that caching we are seeing a higher memory usage.

Assigning to Chris for confirmation but I think this is a WontFix if its the case above.

I agree with your assessment. Also note that if animating simple paths, GrSoftwarePathRenderer renderer will *not* clean up old path masks. They will just pile up until the cache feels memory pressure.

How determined are we to make Skia cache individual path masks? Tiles are already a cache in and of themselves. Maybe a second level cache isn't worth the memory?

There is a context option to disable caching of path masks altogether if we want to investigate.

The caching strategy of evicting entries once you hit the cache limit, even if the entries might never be reused, is fine. Its the strategy we adopt for all other caching too. We also evict all of skia's caching on idle cleanup, so this memory wouldn't sit around forever unless there is continuous raster.

 Issue 911483  has been merged into this issue.