Reusable adapters?

[kainino0x]

In the WebGPU spec, adapters are not reusable. Creating a device consumes the adapter. This was done to make sure that people check adapters again if they need to reinitialize.

In https://dawn-review.googlesource.com/c/dawn/+/237274/comment/a9335ef5_172b283b/ it was suggested that this shouldn't be true in webgpu.h (unless the backend is Wasm), or it should be configurable (and we validate it on Wasm).

cc @Kangz

[kainino0x]

I think the same arguments that motivated making adapters single-use in JS apply equally to C. If we had some kind of PhysicalDevice concept then maybe it should be a reusable singleton (also checkable for pointer equality). But Adapter isn't really that.

I'm not sure how much existing code will break if we implement this in Dawn.

[eliemichel]

At least the default behavior should match the Web/WASM case. It is always possible to add an extension to ask not to consume the adapter (I did not know about that consumption BTW!)

[kainino0x]

May 1 meeting:

• CF: Use cases for making multiple devices from the same adapter…
• CF: Does Dawn cache the enumerated adapters? So you can call requestAdapter many times and it's fine. wgpu does not cache though. If there is caching, fine to make it single-use.
• LK: François had a prototype for the single-use. Corentin suggested maybe it should be an extension and not the default behavior - maybe a Dawn-specific issue, but tests and things depend on being able to create a single adapter and get multiple devices from it. Could be temporary Dawn thing.
• CF: Instance feature to enable adapter reuse?
• KN: Think it should be in RequestAdapterOptions. Right now that only affects which adapter, not the state of the adapter you get. But
• CF: Or could be in RequestDevice, controls whether it poisons the adapter.
• KN: I like having it match the web API. Corentin feels more strongly about it
• LK: Also don't really want to make the change. Would lean toward not doing it.
• CF: Would default to poisoning (same as web api)
• all: in favor of flag
• KN: Where to put it?
• CF: Device…
• LK: Might have to be instance-level because Wasm can't support it, have to validate it
• KN: Can validate that at RequestDevice time
• LK: Think it's weird that you can RequestDevice(please expire = false), RequestDevice(please expire = true) [ambiguous whether the second one will succeed]
• KN: Think it's OK, just manipulates the adapter.destroyed flag that we already spec (in the web spec).
• CF: Phrase it as performing a task in RequestDevice (destroy the adapter).
• KN: More fragile, if you miss it in one RequestDevice call, then your adapter gets destroyed forever.
• LK: More a property of the adapter … Think it's easier to understand that way, than as a property of the operation. And set in fewer places, instead of every time you call RequestDevice.
  • KN: Especially because it defaults to expiring.
• …
• KN: Think it will be a bool member of either RequestAdapterOptions or DeviceDescriptor.
• LK: Don't think you'll ever actually want to control it that finely. Prefer instance feature.
• CF: Instance feature does seem good. Similar to other features where Emscripten can reject it early.
• KN: When you enable this feature, all adapters on the device don't auto-expire?
  • LK/CF: Yes.
• KN: Probably fine for Dawn, but I think we share one instance for both internal stuff and web content.
  • LK: Pretty sure we have an instance for each WebGPUDecoderImpl now. [KN: confirmed this seems to be true]
  • KN: Makes sense, also because we have toggles at that level.
• Add a new WGPUInstanceFeatureName that causes adapters to be multi-use instead of single-use. Single-use by default.