1352654 - Enable building gradient display items for nsDisplayBackgroundImage by default

Status: RESOLVED DUPLICATE of bug 1359242

(Core :: Graphics: WebRender, enhancement)

Description

[Ryan Hunt [:rhunt] (on leave until early May)]

This is just a tracking bug for the work to enable gradient display item building from nsDisplayBackgroundImage by default. This is just about gradients, other work will be done for images (bug 1351242).

Comment 1

[Ryan Hunt [:rhunt] (on leave until early May)]

Here is a try run with the pref and some more recent patches affecting gradients,
https://treeherder.mozilla.org/#/jobs?repo=try&revision=f33df7c2d57d521475649fd77c3c9faed1e97a94

Another try run based off the previous,
https://treeherder.mozilla.org/#/jobs?repo=try&revision=f33df7c2d57d521475649fd77c3c9faed1e97a94

There are about 40-50 reftest failures to look into. They mostly seem to be gradient tests so that is nice.

Comment 2

[Ryan Hunt [:rhunt] (on leave until early May)]

I should also note, with the pref enabled I am noticing some rendering artifacts on the chrome of the browser that I am looking into.

Comment 3

[Ryan Hunt [:rhunt] (on leave until early May)]

A lot of the failures were due to dithering one gradient and not the other. I've updated the try run with dithering disabled.

https://treeherder.mozilla.org/#/jobs?repo=try&revision=a9fb9840cc6c468d3275da06c420df68cc898bd5&selectedJob=88389564

It seems like there are new failures that sneak in.

Comment 4

[Ryan Hunt [:rhunt] (on leave until early May)]

https://treeherder.mozilla.org/#/jobs?repo=try&revision=8babcac9491f25da2653cdb5ba3a8c4966dfcf10&selectedJob=93588720

The remaining failures:
1. border gradient tests - differences between WR path and non WR border path (~33 tests)
2. wrong positioning of bounds in nsCSSRenderingGradients (~2 tests) - have fix for this
3. background-blend-mode tests (~2 tests) - triaging this one still
4. linear gradient tests (~5 tests) - differences between WR path and canvas path
5. radial shape tests (~6 tests) - problems with -moz-transforms in reference html
6. zero length radial gradient tests (~10 tests) - have a fix for this
7. two stop tests (~2 tests) - have a fix for this

When border gradients are using WebRender, 1. should be fixed, until then it might need fuzzing. 4 is more difficult because of the canvas path. For both of them, I think the differences between the gradients should be investigated. We might be able to improve the accuracy and not need fuzzing or as much.

Comment 5

[Ryan Hunt [:rhunt] (on leave until early May)]

The latest try run with some PR's cherry picked from upstream webrender. I only see the rotated test and some problems with display item transforms as remaining. The rest are fuzziness issues.

https://treeherder.mozilla.org/#/jobs?repo=try&revision=81bba142e762c16ee02379f5e0e93e47d5cae092

I also kicked off a bunch of try runs with different attempts to increase accuracy through two different ways,

1. Increase the GradientData table size - doesn't seem to have a strong effect, but noticeable when the table size reaches 1024, one test resolves itself. I don't think that increasing the table size to 1024 is worth it for just that

2. Change the lerp used in filling the GradientData table size - doesn't seem to have a noticeable effect


https://treeherder.mozilla.org/#/jobs?repo=try&revision=9ddebf7e876518513b0fbfb1df136a2b74c6892d
https://treeherder.mozilla.org/#/jobs?repo=try&revision=45b1dd1bed79b0ee1dc6da75fe6e4cc8931b468c
https://treeherder.mozilla.org/#/jobs?repo=try&revision=d64ad2c2e13548ba2a3e95669357ec5ba1bc6188
https://treeherder.mozilla.org/#/jobs?repo=try&revision=ae6cc8a364a0fe8f3e781e1145cba48fc3cacbb7

Comment 6

[Ryan Hunt [:rhunt] (on leave until early May)]

I'm not actively working on this.

Comment 7

[Kevin Chen[:kechen] (UTC + 8)]

Hello Ryan, do you have any WIP for this bug?

Comment 8

[Ryan Hunt [:rhunt] (on leave until early May)]

Hi Kevin, no any code I had for webrender gradients is checked in to central. I also don't know much about what needs to be done still because a lot has changed since I worked on this.

Comment 9

[Kevin Chen[:kechen] (UTC + 8)]

I think we've enable nsDisplayBackgroundImage in Bug 1359242.
Some fuzzy conditions for gradient reftest are added in that bug, but the errors are minor and I think we do need to fix it right now; therefore, I would close this bug.