[WebGPU] Access violation when initializing the WebGPU context

[shiyi9801]

This issue happens in WebNN ORT WebGPU EP.

reproduce steps

1. Download the latest chrome canary
2. Put the ORT dlls built with WebGPU (onnxruntime.dll, onnxruntime_providers_shared.dll) under "C:\Program Files\ORT" (dxcompiler.dll and dxil.dll are not required because they are packed into chrome)
3. start the browser .\chrome.exe --enable-features=WebMachineLearningNeuralNetwork,WebNNOnnxRuntime --allow-third-party-modules --webnn-ort-library-path-for-testing="C:\Program Files\ORT" --webnn-ort-ep-library-path-for-testing=fakepath --webnn-ort-logging-level=VERBOSE
4. navigate to https://webmachinelearning.github.io/webnn-samples/image_classification/
5. Select WebNN (GPU) Float32 MobileNet V2, the test should run successfully
6. Refresh the page, repeat step 5, the process will crash. Navigate to chrome://gpu and scroll down to the bottom, you can see the logs GpuProcessHost: The GPU process crashed! Exit code: STATUS_ACCESS_VIOLATION.

debugging
From my debugging, it looks like the access violation is caused by accessing a null dawn::native::InstanceBase*.

You can put the pdbs along with the ORT dlls and attach to the chrome GPU process if you want to debug.

[shiyi9801]

@fs-eire @guschmue @fdwr , do you have any ideas about this issue? thanks!

/cc @huningxin @qwu16

[fs-eire]

when you create the ONNX Runtime session, do you specify any EP options (dawnProcTable, webgpuInstance, webgpuDevice)? They are called via onnxruntime API SessionOptionsAppendExecutionProvider.

[shiyi9801]

We didn't specify any WebGPU options I believe, the WebGPU EP should be defaulted.

/cc @mingmingtasd

[shiyi9801]

The code of session options: https://source.chromium.org/chromium/chromium/src/+/main:services/webnn/ort/ort_session_options.cc;l=230