Performance considerations: startup

[Vipitis]

This might be an issue on my specific old system, but perhaps it's worth a little discussion.

Running any small example (or larger project) has a couple seconds startup time. I want to run some more profiling but it looks like creating the instance and requesting the adapter are two of the main slowdowns.

One easy speedup I identified is to only request one backend using instance extras, this already saves 1-2 seconds.
Will need to look more into this, as this will likely depend on platform a lot. Async might also help, but I feel most of the startup tasks depend on others to finish first.

(this isn't a perfect comparison, but matches the subjective experience too)

[Korijn]

Maybe the automatic backend selection result can be cached and tried first upon subsequent starts?

[Vipitis]

I mean that sorta happens due to the instance singleton. I have to do a few tests to see if subsequent calls to request adapter are any faster. But I doubt that's like a common use case. My concern is mostly with cold start.

Since the webgpu.h spec allows for "undefined" which essentially means all again. So this would still require all the backends to be requested for the instance... I don't see a way to short circuit this - maybe if you do it manually (first vulkan, then primary, and then all/GL) and exit early? Would be another case to test - but I will also look this up upstream.

But it will likely make sense to use any user select backend (via the env var) for instance creation and enumerate adapters which currently defaults to all: https://github.com/gfx-rs/wgpu-native/blob/ba4deb5d935652f40c7e051b15cbb5d097219941/src/lib.rs#L2799 and we don't implement the backend option for enumerate adapters yet.

[almarklein]

This'd make it tempting to check sys.platform and map to Vulkan/Metal from there. Then fallback to auto-select when that fails.

[Vipitis]

I tried to do a benchmark comparison, but there is all kinds of caching going on. so to really compare a coldstart I need another evening to do like multiple scripts and do it end to end (request everything, compile a model, render a frame) to get a real world first frame latency.