A-buffer causes timeout on AMD integrated graphics

[arswaw]

Hi,

On the a-buffer page, there is a driver timeout that occurs with AMD Ryzen 5 5500U on my HP Envy x360 laptop.
15m-eu0033dx

Is this a programming error or is this model of GPU not expected to be able to run the demo?

[kainino0x]

This is surprising, I wonder if there is some undefined behavior...
cc @rconde01

[rconde01]

I am pursing a similar issue with intel: https://github.com/rconde01/intel_driver_indirect_call_hang

Maybe the example is doing something wrong - I will review, but I don't think so? Or maybe the driver gets cranky in conjunction with other things that chrome is doing behind the scenes? @arswaw if you have the time it would be interesting if you could disable various pieces of the renderer and see what pushes it over the edge?

More broadly, I think perhaps chrome needs to treat these as DOS issues because if it happens too many times Windows will block the GPU from the chrome instance.

Did-you/could-you report to AMD? It may truly be a driver issue.

[arswaw]

@rconde01 I apologize that I hadn't addressed your comment sooner. I just reported the driver timeout to AMD.

I have time to look at this. Can you give me suggestions as to what to disable once I download the project locally?

[rconde01]

Hey @arswaw. I'm at conference this week but i will try to get back to you next week with some details.

[arswaw]

@rconde01 do you have any details?

[rconde01]

Sorry for the late reply. I have a few ideas on tests that would be interesting:

1. There are 3 passes. Opaque, Translucent, and Composite. You could comment out the composite pass and then the translucent pass to bound where the issue is.
2. In composite.wgsl you could try reducing maxLayers to see if that has an effect
3. When you find which pass is the culprit, you could try backing out code from the end to the start of the shader to see what triggers the issue. However you would need to be careful that you change it in such a way that the compiler doesn't just optimize everything out (i.e. the remaining code needs to have an effect on the output).

As I said previously, I've had similar issues in another context which are almost certainly driver related. I suspect it's the same here but happy to be wrong and learn something. Bisecting the issue as I've suggested would be helpful in that regard.

[rconde01]

@arswaw It occurred to me I'm not sure of your background. If trying those edits is not practical for you, let me know and i can probably put together something you can directly test.