wgpu: Cannot access dedicated graphics with wgpu from a docker container.

Looks like this was resolved so I don’t think there’s anything left to do here from the wgpu side. As mentioned, Docker needs to use a compatible GPU adapter or CPU adapter as a fallback. Closing this for now

I think you’re right, just realized it’s actually not using the dedicated GPU right now.

When requesting an adapter like so:

let adapter = instance
    .request_adapter(&wgpu::RequestAdapterOptions {
        power_preference: wgpu::PowerPreference::HighPerformance,
        force_fallback_adapter: false,
        compatible_surface: surface.as_ref(),
    })
...

We are getting the CPU adapter:

AdapterInfo {
    name: "llvmpipe (LLVM 12.0.0, 256 bits)",
    vendor: 65541,
    device: 0,
    device_type: Cpu,
    backend: Vulkan,
}

So I think this is probably not a wgpu issue, and just an issue with docker accessing the gpu.

@kvark I had not realised how old our version was. We are using 0.6.0 where 0.11.1 is available. I’m just porting our code and will report back here.

Yeah, this looks like a broken Vulkan setup. It was able to create an instance but not do anything useful with it.

Great that you figured it out! Seems to be a separate issue from the subject though? Anyway, we are detecting and returning a proper error when the feature isn’t enabled, and feature requirements are documented right on unclipped_depth.

You are absolutely right @kvark , I wasn’t comparing the same branch, but I get the same result on the v0.12 branch.

So maybe you are comparing hello-compute from wgpu master versus your program on wgpu-0.12? Could you try running hello-compute from https://github.com/gfx-rs/wgpu/tree/v0.12 instead?

I don’t think wgpu requires the X server at any level. Obviously, if you are creating wgpu::Surface, it needs to be presented somewhere. Try running the hello-compute example without the X server, it should be unaffected.

However we’ve realized that the issue is tied to docker Xserver access, since running xhost +local:docker on the host machine solves our problems and vulkaninfo, vkcube and wgpu-info work correctly in that case.