1803375 - Geometry generated by wpf-gpu-raster fails to rasterize as intended on macOS AMD

Status: RESOLVED FIXED

(Core :: Graphics: Canvas2D, defect, P2)

Description

[Jeff Muizelaar [:jrmuizel]]

It looks like there's a precision issue when drawing long pixel high triangles

Comment 1

[Jeff Muizelaar [:jrmuizel]]

Attachment: Recording that shows the problem

Comment 2

[Jeff Muizelaar [:jrmuizel]]

Interestingly enough, I'm not able to reproduce the problem when using the software rasterizer

Comment 3

[Jeff Muizelaar [:jrmuizel]]

The problem doesn't seem to show up on Windows either.

Comment 4

[Jeff Muizelaar [:jrmuizel]]

Attachment: Screen Shot 2022-11-30 at 10.54.41 AM.png

Comment 5

[Andrew Osmond [:aosmond] (he/him)]

It renders fine for me on Ubuntu 20.04, NVIDIA GeForce GTX 950, driver 510.85.02, GNOME + X11.

Comment 6

[Jeff Muizelaar [:jrmuizel]]

The player is here: https://htmlpreview.github.io/?https://github.com/kdashg/canvas-rr/blob/master/player.html

Comment 7

[Jeff Muizelaar [:jrmuizel]]

This seems like it might be a shader compiler bug.

Comment 8

[Jeff Muizelaar [:jrmuizel]]

It looks like I only see this when using the AMD Radeon Pro 5300M GPU

Comment 9

[Jeff Muizelaar [:jrmuizel]]

I can reproduce this with a standalone test case.

Comment 10

[Jeff Muizelaar [:jrmuizel]]

That test case is here: https://github.com/jrmuizel/amd-rasterization-bug/. The test case seems to run fine on AMD on Linux. I changed the mesa driver to set the ROUND_MODE which "Controls conversion of X,Y coordinates from IEEE to fixed-point" in PA_SU_VTX_CNTL to Truncate (OGL) instead of Round to Even (D3D). With that change, I reproduced a symptom very similar to what I see on macOS. My current theory is that macOS is using the Truncate value.

Comment 11

[Jeff Muizelaar [:jrmuizel]]

Attachment: Bug 1803375. Use different coordinates when running on AMD.

This updates the version wpf-gpu-raster which adds support for
GPUs/drivers that use truncation instead of rounding when converting
vertices to fixed point.

It also adds the GL vendor to InitContextResult so that we can detect
AMD on macOS and tell wpf-gpu-raster that truncation is going to happen.

Comment 12

[Pulsebot]

Pushed by [e-mail]:
https://hg.mozilla.org/integration/autoland/rev/bf51418892a8
Use different coordinates when running on AMD. r=lsalzman,jgilbert

Comment 13

[Narcis Beleuzu [:NarcisB]]

https://hg.mozilla.org/mozilla-central/rev/bf51418892a8

Comment 14

[Jeff Muizelaar [:jrmuizel]]

Comment on attachment 9313484 [details]
Bug 1803375. Use different coordinates when running on AMD.

Beta/Release Uplift Approval Request

• User impact if declined: Canvas will draw with artifacts on AMD gpus on macOS
• Is this code covered by automated tests?: No
• Has the fix been verified in Nightly?: Yes
• Needs manual test from QE?: No
• If yes, steps to reproduce:
• List of other uplifts needed: None
• Risk to taking this patch: Low
• Why is the change risky/not risky? (and alternatives if risky): The code is conditional to AMD gpus on macOS
• String changes made/needed:
• Is Android affected?: No

Comment 15

[Pascal Chevrel:pascalc]

Comment on attachment 9313484 [details]
Bug 1803375. Use different coordinates when running on AMD.

Approved for 110 beta 8, thanks.

Comment 16

[Pascal Chevrel:pascalc]

https://hg.mozilla.org/releases/mozilla-beta/rev/7f692b5f626c

Comment 17

[Jeff Muizelaar [:jrmuizel]]

This didn't actually get hooked up properly. I've filed bug 1813774 to do that.