CSS filters: fix filtered parent with animated transform child.

CSS filters: fix filtered parent with animated transform child.

When an element with a reference filter has a child element with
an animated transform, we need to invalidate the reference filter
on each animation frame. This is because the changing child position
can change the reference box used for filter region computations.

This requires disabling compositor animations for all nodes with
reference filter parents, since we need to them to run blink-side
to trigger the invalidation. It also means we must traverse up the
hierarchy to discover such filters and invalidate them.

BUG=620394
CQ_INCLUDE_TRYBOTS=master.tryserver.chromium.linux:linux_layout_tests_slimming_paint_v2

[Stephen White, 2016-08-10 20:17:47]

Description was changed from

==========
CSS filters: fix filtered parent with animated transform child.

When an element with a reference filter has a child element with
an animated transform, we need to invalidate the reference filter
on each animation frame. This is because the changing child position
can change the reference box used for filter region computations.

This requires disabling compositor animations for all nodes with
reference filter parents, since we need to them to run blink-side
to trigger the invalidation. It also means we must traverse up the
hierarchy to discover such filters and invalidate them.

BUG=620394
==========

to

==========
CSS filters: fix filtered parent with animated transform child.

When an element with a reference filter has a child element with
an animated transform, we need to invalidate the reference filter
on each animation frame. This is because the changing child position
can change the reference box used for filter region computations.

This requires disabling compositor animations for all nodes with
reference filter parents, since we need to them to run blink-side
to trigger the invalidation. It also means we must traverse up the
hierarchy to discover such filters and invalidate them.

BUG=620394
==========

[Stephen White, 2016-08-10 20:17:47]

senorblanco@chromium.org changed reviewers:
+ chrishtr@chromium.org

[Stephen White, 2016-08-10 20:19:54]

chrishtr@: PTAL. This is a followup patch to
https://codereview.chromium.org/2230593002/ which handles
the animated case. Let me know if you'd like me to merge them, or if you'd
prefer to land the other patch first.

I'm treading in areas in which I don't usually go, so feel 
free to point out if something could be done better.

[chrishtr, 2016-08-11 01:23:59]

This patch addresses what looks like two bugs:

1. can't run reference filters on the compositor thread because they may involve
computing
size of transformed children and storing them in the filter graph.

2. Need to update reference filters whenever geometry of the subtree changes.

#1 is solved I think via the CompositorAnimations change.

Marking for compositing is probably already taken care of by existing triggers.

How does your cl fix #2 for paint? In particular, how does it mark objects which
contain one that changes style as needing paint when the object's bounds change?

...

This whole thing could avoided if was not required to compute the bounds of the
reference box for filters. can Skia be taught how to do that at raster time? cc
could do something similar at draw time.

https://codereview.chromium.org/2232013002/diff/1/third_party/WebKit/Source/c...
File third_party/WebKit/Source/core/animation/CompositorAnimations.cpp (right):

https://codereview.chromium.org/2232013002/diff/1/third_party/WebKit/Source/c...
third_party/WebKit/Source/core/animation/CompositorAnimations.cpp:248: if
(layoutObject->styleRef().filter().hasReferenceFilter()) {
All reference filters need css-specified geometry?

[chrishtr, 2016-08-11 02:54:27]

On 2016/08/11 at 01:23:59, chrishtr wrote:
> This patch addresses what looks like two bugs:
> 
> 1. can't run reference filters on the compositor thread because they may
involve computing
> size of transformed children and storing them in the filter graph.
> 
> 2. Need to update reference filters whenever geometry of the subtree changes.
> 
> #1 is solved I think via the CompositorAnimations change.
> 
> Marking for compositing is probably already taken care of by existing
triggers.
> 
> How does your cl fix #2 for paint? In particular, how does it mark objects
which contain one that changes style as needing paint when the object's bounds
change?
> 
> ...
> 
> This whole thing could avoided if was not required to compute the bounds of
the reference box for filters. can Skia be taught how to do that at raster time?
cc could do something similar at draw time.
> 
>
https://codereview.chromium.org/2232013002/diff/1/third_party/WebKit/Source/c...
> File third_party/WebKit/Source/core/animation/CompositorAnimations.cpp
(right):
> 
>
https://codereview.chromium.org/2232013002/diff/1/third_party/WebKit/Source/c...
> third_party/WebKit/Source/core/animation/CompositorAnimations.cpp:248: if
(layoutObject->styleRef().filter().hasReferenceFilter()) {
> All reference filters need css-specified geometry?

And to clarify, I don't mean teaching Skia/cc about SVG layout and CSS, I mean
teaching it how to use the bounds of the render surface the effect
applies to as the reference box for the filter graph. Or in the case of the
non-composited path, the equivalent of a render surface.

[Stephen White, 2016-08-11 15:17:12]

https://codereview.chromium.org/2232013002/diff/1/third_party/WebKit/Source/c...
File third_party/WebKit/Source/core/animation/CompositorAnimations.cpp (right):

https://codereview.chromium.org/2232013002/diff/1/third_party/WebKit/Source/c...
third_party/WebKit/Source/core/animation/CompositorAnimations.cpp:248: if
(layoutObject->styleRef().filter().hasReferenceFilter()) {
On 2016/08/11 01:23:59, chrishtr wrote:
> All reference filters need css-specified geometry?

Yes. All reference filters include the "filter region" (the x/y/width/height
attributes on the <filter> node, see
https://drafts.fxtf.org/filters/#filter-region), which is expressed in terms of
the "fractions or percentages of the bounding box on the referencing element".
For hysterical reasons, it defaults to 10% larger than the bounding box (x=-10%,
y=-10%, width=120%, height=120%). So
even the default requires that we know the size.

[Stephen White, 2016-08-11 15:20:54]

On 2016/08/11 01:23:59, chrishtr wrote:
> This patch addresses what looks like two bugs:
> 
> 1. can't run reference filters on the compositor thread because they may
involve
> computing
> size of transformed children and storing them in the filter graph.
> 
> 2. Need to update reference filters whenever geometry of the subtree changes.
> 
> #1 is solved I think via the CompositorAnimations change.

Correct.

> Marking for compositing is probably already taken care of by existing
triggers.
> 
> How does your cl fix #2 for paint? In particular, how does it mark objects
which
> contain one that changes style as needing paint when the object's bounds
change?

That's the CompositedLayerMapping change: it rebuilds the filter graphs of all
parent elements when a child transform changes.

> 
> ...
> 
> This whole thing could avoided if was not required to compute the bounds of
the
> reference box for filters. can Skia be taught how to do that at raster time?
cc
> could do something similar at draw time.

Skia currently expresses its crop rect (the equivalent of filter region) in
terms 
of (essentially) CSS pixels.

As noted above, the filter region (on the <filter> node, and also used to
compute
the filter primitive subregion for each filter primitive) is expressed as a
fraction 
of the bounding box of the filtered element.

The only ways I can see to make this work are:

1) Convert those fractions (and other SVG units) to CSS pixels at paint time
relative to the computed total bounds of the filtered element, which is what I'm

doing here.

2) Teach Skia how to apply a fractional crop rect, and then apply those
fractions to the layer size received in cc. This assumes that the layer size
by the time we get to cc is exactly the correct bounding box.

[Stephen White, 2016-08-11 22:03:17]

Description was changed from

==========
CSS filters: fix filtered parent with animated transform child.

When an element with a reference filter has a child element with
an animated transform, we need to invalidate the reference filter
on each animation frame. This is because the changing child position
can change the reference box used for filter region computations.

This requires disabling compositor animations for all nodes with
reference filter parents, since we need to them to run blink-side
to trigger the invalidation. It also means we must traverse up the
hierarchy to discover such filters and invalidate them.

BUG=620394
==========

to

==========
CSS filters: fix filtered parent with animated transform child.

When an element with a reference filter has a child element with
an animated transform, we need to invalidate the reference filter
on each animation frame. This is because the changing child position
can change the reference box used for filter region computations.

This requires disabling compositor animations for all nodes with
reference filter parents, since we need to them to run blink-side
to trigger the invalidation. It also means we must traverse up the
hierarchy to discover such filters and invalidate them.

BUG=620394
CQ_INCLUDE_TRYBOTS=master.tryserver.chromium.linux:linux_layout_tests_slimming_paint_v2
==========