Add pci bus and device uuid to AdapterInfo

[tychedelia]

Setting up some test infrastructure, and it's a bit frustrating not to be able to select a given adapter by a stable identifier when multiple GPUs with the same vendor id exist.

Adds the PCI bus id and device uuid to AdapterInfo on Vulkan.

[cwfitzgerald]

If you wouldn't mind, it shouldn't be hard to wire up these in d3d via https://learn.microsoft.com/en-us/windows/win32/api/dxgi/ns-dxgi-dxgi_adapter_desc which has all this info.

If you're feeling extra crazy, you could also wire this up on GL via https://registry.khronos.org/OpenGL/extensions/EXT/EXT_external_objects.txt, but not touching that one is also understandable lol.

[jimblandy]

@tychedelia Can you clear up the CI failures? We generally don't start reviewing PRs until CI likes them.

[tychedelia]

@tychedelia Can you clear up the CI failures? We generally don't start reviewing PRs until CI likes them.

Yes, apologies was just (slowly) doing @cwfitzgerald's suggestion.

I've removed Uuid as there doesn't appear a way to do this in the Windows API. But the bus ID is sufficient for a stable identifier and probably more useful anyway.

[cwfitzgerald]

Is the Luid in the adapter info not the same? I thought that was how you matched devices when doing cross api interop.

[tychedelia]

Is the Luid in the adapter info not the same? I thought that was how you matched devices when doing cross api interop.

I am not super familiar with these APIs but the docs here suggest it's not valid across restarts, which limits its usefullness relative to the Uuid provided by Vulkan.

https://learn.microsoft.com/en-us/windows/win32/api/d3d12/nf-d3d12-id3d12device-getadapterluid

[cwfitzgerald]

Code's fine and we can land. So the purpose of this is to differentiate between multiple cards on the same machine, across backends?

[cwfitzgerald]

How the hell did you figure out how to do this lmao

[cwfitzgerald]

@kpreid is this a problem for feature sets? I'm never quite sure how implicit features work.

[tychedelia]

oh hmm, i may have resolved a merge conflict incorrectly

[kpreid]

Looks like it — this line should be dropped. It would not hurt but it would not have any effect.

[kpreid]

I'm never quite sure how implicit features work.

The rule is that if

• a dependency is optional, and
• not mentioned using dep: syntax anywhere in [features],

then it gets an implicit feature. (When RFC 3491 is implemented in a future edition, implicit features will no longer be created and this will be an error.)

[tychedelia]

So the purpose of this is to differentiate between multiple cards on the same machine, across backends?

Yeah so the immediate thing is that I'm trying to set up benchmarking for bevy on my build server, but I need a way to have a stable reference to a particular card when I'm spawning docker containers. Iteration order alone probably would work for this but it's also important longer term for some media server/installation things that use multiple cards that you for sure know you are getting the physical outputs for cards plugged into a certain slot.