WebGPU on EC2 Windows - requestAdapter Failed to create WebGPU Context Provider

[wenhan chong]

Hi All,

While trying to run Chrome with WebGPU on EC2 Windows instances unattended, I noticed inconsistent behaviors for WebGPU. Each browser session is a coin flip between having and not having WebGPU. The code fails to get adapter from navigator.gpu.requestAdapter. Errors show "Failed to create WebGPU Context Provider". This only happens when I run the tests unattended. After connecting to the EC2 instances, the same set of automated tests are able to render WebGPU consistently without issues. I also noticed several error messages including GPU process crash in chrome://gpu when running unattended.

[5788:5792:0722/161415.057:ERROR:angle_platform_impl.cc(44)] : SwapChain11.cpp:636 (virtual rx::SwapChain11::reset): Could not create additional swap chains or offscreen surfaces, HRESULT: 0x887A0022
[5788:5792:0722/161415.057:ERROR:angle_platform_impl.cc(44)] : SwapChain11.cpp:643 (virtual rx::SwapChain11::reset): Device lost in SwapChain11::reset HRESULT: 0x887A0022, reason: HRESULT: 0x00000000
[5788:5792:0722/161415.057:ERROR:gl_display.cc(520)] : EGL Driver message (Critical) eglCreateWindowSurface: Context lost.
[5788:5792:0722/161415.057:ERROR:gl_surface_egl.cc(431)] : eglCreateWindowSurface failed with error EGL_CONTEXT_LOST
[5788:5792:0722/161415.058:ERROR:gl_display.cc(520)] : EGL Driver message (Critical) eglSetBlobCacheFuncsANDROID: display had a context loss
[5788:5792:0722/161415.065:ERROR:shared_context_state.cc(1200)] : SharedContextState context lost via ARB/EXT_robustness. Reset status = GL_UNKNOWN_CONTEXT_RESET_KHR
GpuProcessHost: The GPU process crashed!
[5492:6120:0722/161415.439:ERROR:shared_image_manager.cc(328)] : SharedImageManager::ProduceMemory: Trying to Produce a Memory representation from a non-existent mailbox.

Trying to understand what is causing the inconsistency and if there are additional options required to run WebGPU on unattended EC2 windows instance.

Regards,
Wen Han

[Kai Ninomiya]

This sounds like a Chromium bug to me - either we shouldn't crash, or we should blocklist this configuration. It's possible that some flags could work around the issue, but flags should never be required to make Chrome work correctly.

Could you please file a bug on https://crbug.com containing the following:

- The info you provided above

- What type of EC2 instance it is

- Full chrome://gpu report (ideally two copies - one when it works and one when it doesn't - if possible)

- Crash report IDs from chrome://crashes if there are any which might be related to this problem

And then send the bug link here once you have done that, so we can triage it.

Thank you!

-Kai (he/they)

[wenhan chong]

Bug report filed:  Inconsistent behavior for WebGPU on Amazon EC2 Windows Server 2022 when running tests unattended [355397968] - Chromium