Just tested this locally, and I'm only seeing a small (10%) decrease due to GPU raster. However, either way, I don't really feel like this is a real scenario. The test in question:

- Sets up 500 divs on top of each-other, each with 0.5 opacity.
- Adds a transform-z to these so we create a CC layer for each.
- Tries to smoothly change the color of all 500 divs at the same time.

This ends up turning into rasterizing / compositing 1000 tiles a frame, which swamps the compositor is very inefficient in terms of tile rasterization (lots of skia / draw call overhead, very little batching).

As a point of comparison, if we get rid of the translate-z and let Skia handle all this overdraw internally, GPU raster runs at 60hz and is 2.5x faster than SW.

This test may be useful to catch compositor regressions, but I'm not sure it's a useful comparison point for GPU vs SW raster.

There may be ways to optimize this, but I'm not sure it's worth it?