cameraCapture test doesn't reliably get a real frame

[kainino0x]

The cameraCapture test from #3874 doesn't reliably get a real video frame. In Chrome on Mac, I'm seeing that the first time I run the test, it passes, but then if I run it again, it fails:
https://gpuweb.github.io/cts/standalone/?q=webgpu:web_platform,external_texture,video:importExternalTexture,cameraCapture:dstColorSpace=%22srgb%22

If I display the image with:

    document.body.append(canvas);

it is black on the first run, and has a real image on the second run. When it has a real image, it fails, usually with a margin of around 0.05. I don't know why.

I found this because if I use --use-fake-device-for-media-stream --auto-accept-camera-and-microphone-capture in Chrome to test with a fake camera, then it gets a real frame on the first try, and that causes it to fail.

cc @shaoboyan

[kainino0x]

Another comment about the code that gets video frames:

There are three codepaths. Chrome always takes the first path, the other paths don't seem to work, even in Chrome (the ImageCapture path throws an exception, the video element path. We should not have fallbacks, we should pick the path deterministically (use one path for the main thread, which should never skip, and another path for worker threads, which may skip).

Safari: EXCEPTION: TypeError: imageCapture.grabFrame is not a function. (In 'imageCapture.grabFrame()', 'imageCapture.grabFrame' is undefined)
Firefox: SKIP: HTMLVideoElement.captureStream is not supported

[shaoboyan091]

The cameraCapture test from #3874 doesn't reliably get a real video frame. In Chrome on Mac, I'm seeing that the first time I run the test, it passes, but then if I run it again, it fails: https://gpuweb.github.io/cts/standalone/?q=webgpu:web_platform,external_texture,video:importExternalTexture,cameraCapture:dstColorSpace=%22srgb%22

If I display the image with:

    document.body.append(canvas);

it is black on the first run, and has a real image on the second run. When it has a real image, it fails, usually with a margin of around 0.05. I don't know why.

I found this because if I use --use-fake-device-for-media-stream --auto-accept-camera-and-microphone-capture in Chrome to test with a fake camera, then it gets a real frame on the first try, and that causes it to fail.

cc @shaoboyan

Thanks for ccing. FYI, my previous github account @shaoboyan has lost due to I linked all my recover methods to intel company email address and forget to change that... Now I used @shaoboyan091 instead. I see your fix PR, is there anything need to be further investigated? (e.g. #4364)?

[kainino0x]

FYI, my previous github account @shaoboyan has lost due to I linked all my recover methods to intel company email address and forget to change that... Now I used @shaoboyan091 instead.

Oops, I forgot. I've removed the old account from the gpuweb org, that should make it easier to remember.

I see your fix PR, is there anything need to be further investigated? (e.g. #4364)?

Eventually yes, we should add some more coverage here, but let's get a fix landed first.