Diffuse Raytracing example freezes Gnome window manager on Ubuntu 23.04 LTS with NVidia driver 535.86.05

[s-macke]

Diffuse Raytracing example freezes Gnome window manager on Ubuntu 23.04 LTS with NVidia driver 535.86.05

[s-macke]

Hi dcerisano,

I have exactly the same issue with my Ubuntu and Chrome Unstable and had to restart my computer several times.

The issue was somehow introduced with
59f1670#diff-4ee3a8ff57d5db88a6253467fc7959e596761b5978fb884f1b02d0db06a5d3cb

The commit fixed the demos for Windows.

[s-macke]

Thanks to https://toji.dev/webgpu-best-practices/error-handling.html I have improved the error handling and stop immediately after the first error.

Also I have refactored my code for the diffuse raytracing.

Now, it works flawless under Linux.

[dcerisano]

Have you moved those changes to the WebGPU-Lab website yet? I pinged the WebGPU matrix group about your site - it is very cool. Possibly the changes take a while to propagate through git pages. Will try later and close this if it passes.

I use it to smoke-test the WebGPU stack for linux, which seems to be fairly stable but still being staged for release.

I notice some changes you made to the demo can affect when the next animation frame is processed. Possibly a sync issue (race) with the Vulkan swap-chain on linux, affecting both AMD and NVidia drivers somewhere in the API stack.

[dcerisano]

Tested in the wild, passing.

Should probably raise a bug against WebGPU API. It should never be able to crash a desktop. Possibly something in the API stack about Vulkan swap-chain management (maybe a race condition, but it is not intermittent).

The crash only occurs with NVidia Vulkan driver, not AMD's (just a temporary freeze). A difference there is that AMD Vulkan driver is part of the Linux kernel, but NVidia's is not (yet). Another difference is that on my AMD system Wayland is default, but on NVidia X11 is default (have not been able to get Wayland running with NVidia on 23.04).

Maybe it is a bug against kernel.org? That would be interesting. Seems to be something upstream anyway.