Ping! Any specific insights for this benchmark?

Updated instructions:
http://trac.webkit.org/export/HEAD/trunk/PerformanceTests/Animometer/developer.html
Select 'HTML suite > CSS bouncing blend circles'
Select "maintain target framerate" at 50fps

Results on Linux 64bit, browser maximised on 1 out of 2 monitors, v51.0.2704.63 (stable)
1. score 226.16, time complexity 226.16 +/-  0.30%, FPS 42.60 +/-  2.82%
2. score 188.27, time complexity 188.27 +/-  1.45%, FPS 51.64 +/- 10.96%
3. score 138.25, time complexity 138.25 +/-  3.68%, FPS 50.17 +/-  6.77%

Ditto but v52.0.2743.19 (dev)
1. score 392.36, time complexity 392.36 +/-  7.11%, FPS 50.05 +/-  5.13%
2. score 396.02, time complexity 396.02 +/-  4.51%, FPS 49.99 +/-  5.22%
3. score 388.20, time complexity 388.20 +/-  6.98%, FPS 49.84 +/-  6.76%

Results on Chromebook Pixel 2 v50.0.2661.104
1. score  37.00, time complexity  37.00 +/-  0.00%, FPS 50.06 +/-  4.26%
2. score  41.81, time complexity  41.81 +/-  1.72%, FPS 51.10 +/-  3.48%
3. score  35.51, time complexity  35.51 +/-  8.49%, FPS 49.27 +/-  4.99%

Results on Nexus 7, android 6.0.1, v50.0.2661.89 (32 bit)
1. score   3.74, time complexity   3.74 +/- 25.84%, FPS 53.75 +/-  9.56%
2. score   3.00, time complexity   3.00 +/-  0.00%, FPS 53.28 +/-  6.99%
3. score   4.00, time complexity   4.00 +/-  0.00%, FPS 54.67 +/-  6.26%

Now I understand a little more about what's going on, the poor numbers on the Pixel 2 are not surprising (high DPI screen).

With alancutter, I've dug into traces for a few runs of this benchmark on a Lenovo X1 Carbon (also high DPI). Here are a few observations:

- Little time is being spent in main thread animations (~2ms per frame).
- Most time is being spent in rasterize paint on the tile workers (~14ms per frame) and is causing us to miss the target frame rate deadline. We guess that mix-blend-mode is not being hardware accelerated.
- Turning on Rendering > Paint Flashing in devtools shows the occasional flash of the whole area being painted, and when that's not the case individual circles are often broken up into multiple painted rectangles. But we're seeing the same thing with the CSS bouncing circles benchmark.

We think it's best for the paint team to take it from here for further investigations.

We are done with the bulk of the optimizations for this benchmark, closing the task.