Bug 275043 – [WebGPU] white screen on noclip.website

• Status: RESOLVED
• Resolution: FIXED
• Component: WebGPU

[Jasper St. Pierre]

You can try out a preview WebGPU version of noclip.website here:

https://noclip.website/?webgpu#mkds/cross_course

My version of Safari Technology Preview supports a large number of WebGPU samples and examples, but it seems to break on this with a white screen and no error message, while it works on Firefox and Chrome. Some relatively strange things I'm doing in noclip's renderer that you might not have experienced elsewhere:

* I use one giant dynamic uniform buffer for my entire scene.
* I render to an offscreen texture, and then copyTextureToTexture to the onscreen texture.
* I'm using naga to generate the shaders, which might lead to some strangely-written shaders.

Happy to provide advice or help in case you need help narrowing the issue down.

[Radar WebKit Bug Importer]

<rdar://problem/129149508>

[Mike Wyrzykowski]

Is this reproducing with the latest version of Safari Technology Preview 195 https://developer.apple.com/safari/resources/ ?

If it reproduces on STP 195, can you please update this issue with the Mac Model Identifier, e.g., Mac14,14 as it may be device specific? I do not reproduce on a Mac14,14

[Mike Wyrzykowski]

Also would be interesting to know if you see any errors via the shader compilation, GPUShaderModule.getCompilationInfo, or any validation errors via GPUDevice.pushErrorScope("validation") / GPUDevice.popErrorScope("validation")

[Mike Wyrzykowski]

I tried STP 195 on a Mac14,14 and I don't reproduce, but maybe I am doing something wrong.

[Jasper St. Pierre]

Oh, interesting, I made some small changes last night and didn't re-test on my MacBook. Seems like the change I made fixed it.

It seems like it's related to holes in the render pass attachments. I've put back the code [0] which has broken Safari again.

I tried to recreate it in a simpler repro and wasn't able to.

[0] https://github.com/magcius/noclip.website/commit/38952d24352ef21c232019de2524a213d3c7fd43

[Jasper St. Pierre]

For reference, my hardware is a 13-inch M1 2020, Apple M1, running Sonoma 14.4.1, Safari Technology Preview 195.

[Mike Wyrzykowski]

Oh that seems like the bug, thanks for narrowing that down, I will fix this.

I reproduce now with the change rolled-back

[Mike Wyrzykowski]

Created attachment 471577 [details]
upper right corner of the website

[Mike Wyrzykowski]

Looks better now

[Mike Wyrzykowski]

Pull request: https://github.com/WebKit/WebKit/pull/29470

[Jasper St. Pierre]

Thanks! I'm surprised this doesn't have a CTS test for it.

[Mike Wyrzykowski]

https://gpuweb.github.io/cts/standalone/?q=webgpu:api,validation,render_pass,attachment_compatibility,* attempts to test this, but I think it is incomplete as STP 195 passes all the cases locally

[EWS]

Committed 279704@main (858f0966395a): <https://commits.webkit.org/279704@main>

Reviewed commits have been landed. Closing PR #29470 and removing active labels.