Device lost handling

[valaphee]

Bevy version and features

baddb7e

[Optional] Relevant system information

SystemInfo { os: "Windows 11 Pro", kernel: "26100", cpu: "Intel(R) Core(TM) i9-14900HX", core_count: "24", memory: "127.6 GiB" }
AdapterInfo { name: "NVIDIA GeForce RTX 4090 Laptop GPU", vendor: 4318, device: 10007, device_type: DiscreteGpu, driver: "NVIDIA", driver_info: "576.52", backend: Vulkan }

What you did

cargo run --example 3d_scene

causing a device lost by disabling the dedicated gpu in device manager

What went wrong

There is no handling for render device migration, e.g. when switching render devices, Bevy stack overflows or panics.

Additional information

.log

https://learn.microsoft.com/en-us/windows/uwp/gaming/handling-device-lost-scenarios

[alice-i-cecile]

I would be very happy to see improvements here, but I don't have concrete suggestions on where to start beyond "make it crash and then handle the error gracefully".

[alice-i-cecile]

#21335 may require this, but @tychedelia will know more.

[tychedelia]

@atlv24 has tried this before and run into issues. It will be necessary for #21335. Basically, I think the solve here is to move device creation into the ECS such that we could recreate things on-demand, but @atlv24 wasn't sure if this was generally possible. More testing needed here, agree the crash is non-ideal.

[atlv24]

I have a reaaaalllyyyyy old branch on my local that had some refactors for this, i landed most of the important ones but still have not ever been able to create a new device once the current one was lost. I fear there may be some application-level device poisoning. I can't recall if i was recreating the adapter too though - it may just be a question of needing to go a step higher

[atlv24]

I will say though, the easiest way to lose a device is to run a shader with while(true) {} or some equivalent in it