2028326 - Hero Element on https://huggingface.co/spaces/CohereLabs/cohere-transcribe-03-2026 draws incorrectly

Status: VERIFIED FIXED

(Core :: Graphics: WebRender, defect, P1)

Description

[Jeff Muizelaar [:jrmuizel]]

Regression window: https://hg.mozilla.org/integration/autoland/pushloghtml?fromchange=91ac834f411d3bea953c9ee02e6c2ff5561b2f82&tochange=420c5f80a6ca3d16482b7be85a2457608ef4ebf3

Comment 1

[Jeff Muizelaar [:jrmuizel]]

Attachment: Screenshot 2026-03-31 at 9.43.36 PM.png

Comment 2

[BugBot [:suhaib / :marco]]

:nical, since you are the author of the regressor, bug 2020923, could you take a look? Also, could you set the severity field?

For more information, please visit BugBot documentation.

Comment 3

[Nicolas Silva [:nical]]

The regressing revision is https://phabricator.services.mozilla.com/D286035

Comment 4

[Nicolas Silva [:nical]]

The problem comes from the interaction of the tiled rendering strategy and repeated primitives. We set the parameter of the tiled rendering strategy based on the entire primitive and then apply it to each repetition separately. I suspect that computing the rendering strategy per repetition would fix the issue.

Comment 5

[Nicolas Silva [:nical]]

Another direction is to work towards addressing the tiling interacting poorly with repetitions. I noticed that we do a bunch of computation that affect (among other things) tile classification using the projected clip coverage rect for the entire primitive. Restricting the coverage rect to a conservative part that is closer to what the tile touches fixes the glitch.

Comment 6

[Nicolas Silva [:nical]]

Attachment: Bug 2028326 - Better handle tiles with repeated primitives. r=gw

This patch does two things to improve the interaction between tiles and repeated primitives:

• it tightens the clipped_surface_rect (which the tile classifier goes over) by interecting it against the projected local rects of the primitive. This avoids producing tiles outside of the primitive and gives the classifier more accurate information about what it has to deal with.
• it moves the computation of the number of tiles to a later stage where we have this more accurate coverage rect in device space.

This means that sometimes we'll end up picking the tiled strategy based on the coverage of the entire primitive even though each individual repetition will only use a single tile. It's a bit silly but the better fix is to reorganize the code so that we have better information about the device space coverage earlier when we chose the rendering strategy. Since this will need to get uplifted I want to land a less invasive change first.

Comment 7

[Pulsebot]

Pushed by [e-mail]:
https://github.com/mozilla-firefox/firefox/commit/5c928f1f37b1
https://hg.mozilla.org/integration/autoland/rev/2a3ce635d2be
Better handle tiles with repeated primitives. r=gw

Comment 8

[Pulsebot]

Pushed by [e-mail]:
https://github.com/mozilla-firefox/firefox/commit/e6e8ddc779a6
https://hg.mozilla.org/integration/autoland/rev/b2b9f822fc64
Revert "Bug 2028326 - Better handle tiles with repeated primitives. r=gw" for causing wrench failures

Comment 9

[Silaghi Andreea]

Backed out for causing wrench failures
Backout Link
Push with failures
Failure Log
Failure line subprocess.CalledProcessError: Command '['../target/debug/wrench', '--no-scissor', '--headless', 'reftest']' died with <Signals.SIGABRT: 6>.

Comment 10

[Pulsebot]

Pushed by [e-mail]:
https://github.com/mozilla-firefox/firefox/commit/bf66cf5c3dd9
https://hg.mozilla.org/integration/autoland/rev/4d6d95c9b9fd
Better handle tiles with repeated primitives. r=gw

Comment 11

[Atila Butkovits]

https://hg.mozilla.org/mozilla-central/rev/4d6d95c9b9fd

Comment 12

[Nicolas Silva [:nical]]

Attachment: Bug 2028326 - Better handle tiles with repeated primitives (Beta version). r=gw

This patch does two things to improve the interaction between tiles and repeated primitives:

• it tightens the clipped_surface_rect (which the tile classifier goes over) by interecting it against the projected local rects of the primitive. This avoids producing tiles outside of the primitive and gives the classifier more accurate information about what it has to deal with.
• it moves the computation of the number of tiles to a later stage where we have this more accurate coverage rect in device space.

This means that sometimes we'll end up picking the tiled strategy based on the coverage of the entire primitive even though each individual repetition will only use a single tile. It's a bit silly but the better fix is to reorganize the code so that we have better information about the device space coverage earlier when we chose the rendering strategy. Since this will need to get uplifted I want to land a less invasive change first.

Comment 13

[Phabricator Automation]

firefox-beta Uplift Approval Request

• User impact if declined/Reason for urgency: Visual glitches
• Code covered by automated testing?: yes
• Fix verified in Nightly?: yes
• Needs manual QE testing?: no
• Steps to reproduce for manual QE testing:
• Risk associated with taking this patch: low
• Explanation of risk level: Each patch fix a regression and the patches are fairly simple. These are 3 patches that already went through the uplift requests but did not apply cleanly on beta so I had to rebase them manually.
• String changes made/needed?: None.
• Is Android affected?: yes

Comment 14

[Nicolas Silva [:nical]]

Attachment: Bug 2028326 - Better handle tiles with repeated primitives (Beta version).

This patch does two things to improve the interaction between tiles and repeated primitives:

• it tightens the clipped_surface_rect (which the tile classifier goes over) by interecting it against the projected local rects of the primitive. This avoids producing tiles outside of the primitive and gives the classifier more accurate information about what it has to deal with.
• it moves the computation of the number of tiles to a later stage where we have this more accurate coverage rect in device space.

This means that sometimes we'll end up picking the tiled strategy based on the coverage of the entire primitive even though each individual repetition will only use a single tile. It's a bit silly but the better fix is to reorganize the code so that we have better information about the device space coverage earlier when we chose the rendering strategy. Since this will need to get uplifted I want to land a less invasive change first.

Comment 15

[Dianna Smith [:diannaS]]

Comment on attachment 9570928 [details]
Bug 2028326 - Better handle tiles with repeated primitives (Beta version).

Moving flag to release for consideration in the 150 dot release.

Comment 16

[Pulsebot]

https://github.com/mozilla-firefox/firefox/commit/b5af7f5a2f08
https://hg.mozilla.org/releases/mozilla-release/rev/e5093c0f0edb

Comment 17

[Brindusa Tot, DTE]

I was able to reproduce the issue on macOS 12 using Firefox Release 150.
Verified as fixed using Firefox 150.0.1 RC and latest beta 151.0b3 across Win 10, macOS 12 and Ubuntu 24.