GitLab

The issue with fences was identified a long time ago, but it did not
concern Linux before. As long as we use surfaceless drawing and
insert fences between each buffer swap, we noticed some perf
regression.

On a powerful laptop (dell 9560) with Intel 630 graphics, gpu::AsyncSwap
took ~8ms with fences and ~1ms without fences.

On Raspberry Pi3 with Broadcom GPU, swap took ~105ms and
after the applied workaround it took 77ms. The FPS rate
became 16-17FPS instead of 11-12FPS in aquarium webgl demo.

https://docs.google.com/spreadsheets/d/1dKLIQ6dkql-X0ppdc5s5y-ws2hPM-RLgqgGzkFpgUe0/edit?usp=sharing

Bug: 938286
Change-Id: I9fc616976a6ee3f9b4544ad7d8effc41e3ad4a91
Reviewed-on: https://chromium-review.googlesource.com/c/chromium/src/+/1503576
Commit-Queue: Maksim Sisov <msisov@igalia.com>
Reviewed-by: Robert Kroeger <rjkroege@chromium.org>
Reviewed-by: Antoine Labour <piman@chromium.org>
Cr-Commit-Position: refs/heads/master@{#638037}