Optimize mix-blend-mode to avoid readback in most cases.

[gw3583]

(brain dump only - needs more detail added)

At the moment we have various hacks in the WR code and shaders to handle the fact that mix-blend-mode must be isolated from other stacking contexts. I think we can modify how we implement mix-blend-mode in a way that removes these hacks and also is an optimization - removing the need for framebuffer readback in most cases.

When we encounter a mix-blend-mode during flattening, we should move the parent stacking context to be a child of the mix-blend-mode picture node. Then, no readback is needed - the parent stacking context will be drawn to an off-screen surface in the correct pass.

A couple of caveats:

• We still need to do the readback if the parent stacking context is the root (framebuffer). This is rare, and in future we can also track if the root only has a simple background color and avoid the readback in this case too.
• To get correct blending, we need to ensure that the render task created for the mix-blend-mode is the maximum of the size of the backdrop and child stacking context.

This should allow for a much faster implementation of mix-blend-mode in the majority of cases, by skipping framebuffer readback. It also makes it easier to optimize some of the common cases, where the backdrop is a solid color (either by detecting this, or caching the solid color picture in the texture cache).

[kvark]

Sounds reasonable. For the first caveat, can't we introduce an "artificial" full-screen picture just so the readback path is not needed (at the cost of an extra blit/copy)?

[gw3583]

Yup, that should work.

[Gankra]

vague reminder to not introduce a side-channel attack with content-based optimizations on mix-blend-modes/filters.