Re: getCurrentTexture::createView() fails on dedicated GPU

Hey Markus,

WebGPU support on Linux is not exactly ready yet. In particular,
interactions between WebGPU and the rest of the page (with the canvas,
copyExternalImageToTexture, etc) are known to be somewhat broken. We want
to ship WebGPU on Linux without flags (and these bugs fixed) eventually but
it is not the top priority at the moment. See
https://issues.chromium.org/u/1/issues/40218893 for the tracking issue for
WebGPU on Linux. So for now I suggest sticking to the Intel GPU if you need
to show something on the swapchain.

Cheers,

Corentin

On Tue, Aug 13, 2024 at 5:15 PM Markus Gnauck <gnauckm@gmail.com> wrote:

> Hi everyone,
>
> Not sure if the right place to ask for dev support. I did not find
> anything to help me when searching for the problem below, so I thought to
> ask here. My apologies if this is the wrong place!
>
> I'm on Ubuntu 22.04 with the latest Chrome (stable and unstable). My
> notebook has a dedicated GPU which is a GeForce RTX 3050 6GB (Laptop) and
> an integrated Intel Iris Xe. I tried different driver versions etc. in
> troubleshooting the problem, but am currently on nvidia-driver-550.
>
> My WebGPU program runs fine on the integrated GPU (power-preference:
> "low-power"), but it does not on the dedicated RTX ("high-performance").
> Already a minimal WebGPU init will trigger the problem. It happens when the
> context is asked to provide a view of the current texture
> (context.getCurrentTexture().createView()). The error printed in the
> console of Chrome is the following:
>
> -----------
>
> Requested allocation size (1228800) is smaller than the image requires
> (1310720).
>     at ImportMemory
> (../../third_party/dawn/src/dawn/native/vulkan/external_memory/MemoryServiceImplementationOpaqueFD.cpp:131)
>
> localhost/:1 [Invalid Texture] is invalid.
>  - While calling [Invalid Texture].CreateView([TextureViewDescriptor]).
>
> -----------
>
> Canvas resolution is (in this example) 640x480. Preferred canvas format is
> "bgra8unorm". The same is true for the GPUTexture that comes back from
> context..getCurrentTexture(). I fail to understand where the extra size
> comes from that is required for the image. What am I not providing? Or is
> it something with my drivers?
>
> Interestingly, if I remove the presentation logic (i.e. no blit pass) from
> my program, the nvidia GPU will happily execute the compute shaders etc.
>
> I'm somewhat at loss. And since the program is running fine on the
> integrated GPU, I feel something this basic should not go wrong. But since
> I've been using WebGPU only for about a year so far, my experience is
> pretty limited. So anything that points me in the right direction is very
> much appreciated!
>
> Thank you!
>
> Regards,
> Markus
>
>
> Here comes a minimal code example that already triggers above error for me:
>
>   1 async function main()
>   2 {
>   3   if(!navigator.gpu)
>   4     throw new Error("No WebGPU");
>   5
>   6   // "low-power" does work
>   7   let adapter = await navigator.gpu..requestAdapter({
>   8     powerPreference: "high-performance" } );
>   9
>  10   if(!adapter)
>  11     throw new Error("Failed to request adapter");
>  12
>  13   let device = await adapter.requestDevice();
>  14   if(!device)
>  15     throw new Error("Failed to request device");
>  16
>  17   let canvas = document.querySelector("canvas");
>  18   canvas.width = 640;
>  19   canvas.height = 480;
>  20
>  21   context = canvas.getContext("webgpu");
>  22   context.configure({ device, format:
> navigator.gpu.getPreferredCanvasFormat(),
>  23     alphaMode: "opaque" });
>  24
>  25   let renderPassDescriptor = {
>  26     colorAttachments: [
>  27       { view: context.getCurrentTexture().createView(), // Fails with
> error
>  28         clearValue: { r: 1, g: 0, b: 0, a: 1.0 },
>  29         loadOp: "clear",
>  30         storeOp: "store" } ]
>  31   };
>  32 }
>  33
>  34 main();
>