The Animometer Leaves benchmark can reach ~500 Leaves on my Linux desktop with Software Raster.

chromium_leaves_reference.png shows the trace of typical frames on this test.  The test is Main Thread bound, but also does a large amount of raster thread work and GPU thread work copying tiles in Software Raster.  Main-Thread-Before-Activation scheduling (MFBA) enables the Main Thread to pipeline with the raster work.

jellyfish_leaves_2.png shows Jellyfish performance as of 10/24; over 50x worse than mainline.  This is mostly because of a few bugs in image decoding.

jellyfish_leaves_3.png is after fixes to the image decoding issues, and enabling 4 raster threads for the display compositor.

Deleted: chromium_leaves_reference.png 259 KB

	chromium_leaves_reference.png 259 KB View Download

Deleted: jellyfish_leaves_2.png 300 KB

	jellyfish_leaves_2.png 300 KB View Download

Deleted: jellyfish_leaves_3.png 321 KB

	jellyfish_leaves_3.png 321 KB View Download

In #3 jellyfish_leaves_3.png the following look like the main remaining issues:

- We're not getting the equivalent of MFBA scheduling, so about half the performance is lost due to no pipelining.

- Solid Color Analysis is slow (~10x slower than mainline CC).  This seems likely to be a culling issue with our display lists after de-serialization.

- Serialization ~1.3ms, Deserialization ~4.7ms, on critical paths.

- The raster time seems to be about ~20% higher; possibly also related to a culling issue.

The Solid Color Analysis slowdown cause is that we disabled DisplayItemListSettings::use_cached_picture mode on Jellyfish.

A fix is adding an early out in DisplayItemList::Raster().  This is already implemented in newer DisplayItemList upstream.

https://github.com/danakj/chromium/commit/eea8f1ed9d6896a66bf0d981221ab6736e986bf0