First raster_and_record_micro_ct result:
https://ct.skia.org/results/cluster-telemetry/tasks/chromium_perf_runs/wangxianzhu-20180119164747/html/index.html
Fieldname Total Total with Patch Percentage Change Percentage Difference
record_time (ms) 964.125 1194.296 23.874% 21.328%
record_time_partial_invalidation_ms (ms) 844.429 1025.539 21.448% 19.370%
record_time_subsequence_caching_disabled (ms) 4377.054 5236.195 19.628% 17.874%
rasterize_time (ms) 38800.797 44622.719 15.005% 13.958%
record_time_painting_disabled (ms) 7361.178 8437.856 14.626% 13.630%
record_time_caching_disabled (ms) 21539.245 22768.032 5.705% 5.547%
viewport_picture_size (B) 1315134176 1363290976 3.662% 3.596%
record_time_construction_disabled (ms) 5520.815 5708.179 3.394% 3.337%
pixels_rasterized 15762321408 15782277120 0.127% 0.127%
pixels_recorded 55980770230 56009577915 0.051% 0.051%
pixels_rasterized_non_solid 13032996864 13038985216 0.046% 0.046%
Regressions of record times are expected because we move raster invalidation from PrePaint to Paint, but we should lower the regression as much as possible.
viewport_picture_size regression is not expected. This is perhaps due to bug 795017 (empty or no-op save/(transform|clip|effect)/restore pairs).
Ran the #c18 CL on ct: go/spv175perf (sheet 3).
Average time:
- Raster invalidation: about 30ms
- PaintChunksToCcLayer: about 90ms
(For comparison:
AppendToWebDisplayItemList: about 70ms)
So the regression is mainly from PaintChunksToCcLayer which should be our top priority.
Moving this to M-68 (meaning that we won't merge performance fixes to M-67) because the remaining regressions are all about only one test case in a test suite. Other tests in the test suite either don't have regressions or have progressions.
Comment 1 by wangxianzhu@chromium.org
, Jan 19 2018