cc @liyuqian who wrote the original version.

@knopp : do you happen to also have a sample Flutter app that exhibits the difference before and after this PR? I'm very curious about what triggered you to discover this issue, and I'd like to dive deep to see if there are similar issues in other places (e.g., in PictureLayer). I personally tried the gallery app but couldn't find anything. A sample app would also help us see whether someone could rely on the old behavior which makes this a breaking change.

@liyuqian, This is related to partial redraw, which is work in progress. The issue here is that if I compute a damage area between last frame and current and then set it as cull-rect in preroll, some layers will be culled out even though they are within damage area due to wrongly calculated screen-space coordinates (as the unit test shows).

Given that this only causes issue when doing partial repaint you can't really reproduce it with current engine code. But I also can't imagine how any app would rely on wrong screen-space layer coordinates during pre-roll.

There are few other subtle issues that manifest while doing partial repaint related to Backdrop and ImageFilterLayers bounds, but I plan to address them in separate PRs.

In both cases (opacity and picture layer), these layers use SkMatrix::postTranslate in their preroll, but leaf_nodes_canvas->translate in their Paint methods. The canvas translate method uses SkM44::preTranslate to do the operation.

(Actually, opacity uses a mixture of leaf_nodes_canvas and internal_nodes_canvas in its Paint method whereas picture uses only the leaf version. I don't know if that is another issue to consider...?)

The matrix in the Preroll method is only used for rasterizing layers AFAICT, and since we ignore the translation when we rasterize the layer, it won't affect the resulting image. During paint we then work out where to place the rasterized image and at that point the translation is correct and so we put it in the right position. So, even raster caching wouldn't detect this issue.

But, we do integer CTM adjustments on the transforms which might round in different directions if the offset was a fractional pixel, so this could lead to "off by 1" errors in the placement of cached layers?

Good point with the integer CTM. I think both ::Diff and ::Preroll needs be consistent with ::Paint for partial redraws.

The off by 1 error should be fine as we've already relaxed the constraint in

Since this translation only affects the raster cache slightly, I'm ready to approve it. Let's also correct the usage in PictureLayer to make it consistent. Then I think it's ready to merge.

Since this translation only affects the raster cache slightly, I'm ready to approve it. Let's also correct the usage in PictureLayer to make it consistent. Then I think it's ready to merge.

Not entirely sure what you mean. The patch already changes PictureLayer to preTranslate here. Is there something else you have in mind?