This removes hacking the canvas CTM and wrapping the paint's image
filter in a special MatrixTransform that computed a post-transform
instead of its documented pre-transform effect. Performance-wise, the
computed layer sizes should be about the same, but we avoid one less
render target switch because we apply the transformation while drawing
to the dst device, vs. transforming into another temporary layer and
then drawing that to the dst device.
Several important changes in behavior here:
1. The DeviceCM record no longer has a stashed matrix to restore and
holds its restoration paint directly.
2. Devices for image filter inputs can now have device-to-global
transforms that are not integer translates.
3. The MatrixTransform hack punted when there was perspective because it
could produce excessively large temporary images, but the new version
appears to work around that. We now impose a maximum layer size to
protect against that and automatically scale the layer to prevent it.
Perspective image filters otherwise now draw correctly.
6. Updated layer sizing code to use the new image filter APIs
7. Updated backdrop filter and restore filters to go through the same
code paths, although restore filters skip the intermediate image
transform.
- layer bounds and transforms now go through the updated skif API
and is hopefully more straight forward to understand.
8. Now we can optimize root color filter nodes of a filter DAG, even if
the entire DAG can't be represented as a color filter. The last node
is pulled off and composed with the restoration paint instead.
Bug: skia:9074,skia:9283
Change-Id: I1fa1d50135b9d6d453b02f89aa3cc3b54deab678
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/328376
Commit-Queue: Michael Ludwig <michaelludwig@google.com>
Reviewed-by: Brian Salomon <bsalomon@google.com>
f8f23b2030