osr: windows: Add support for accessing gpu renderer information from chromium

[csmoe]

Sharing textures cross gpu cards is not "possible", so If a PC has two more gpu cards, chromium and the client app might have different card chosen.

This kind of api should be helpful for client app to align gpu selection with chromium. cc @reitowo

[csmoe]

Someone encountered this too https://magpcss.org/ceforum/viewtopic.php?f=6&t=19456

[reitowo]

I recommend to directly guide users or write reg table to force apps to use specified gpu. (You'll need to do that anyway if two apps mismatch.)

[magreenblatt]

I agree that pre-configuration of the target GPU is a better approach here.

[csmoe]

Really thanks for your guide, but chromium has its own gpu selection algorithm based on GpuPreferences, the final result may not be the system's config at my first glance, so pre gpu targeting might not work in some cases?

[reitowo]

I think that selection prefers to use high performance GPU

if (gpu_default == gpu_high_perf) {
  // If the default GPU is already the high performance GPU, then it's better
  // for Chrome to always use this GPU.
  gl::SetGpuPreferenceEGL(gl::GpuPreference::kDefault,
                          system_device_id_high_perf);
  return;
}

I remember if you choose GPU explicitly in reg the default GPU will be the selected one.

[dsh0416]

FYI, for our integraton of using osr, we found forcing the usage of high performance GPU might not be enough. We use some hook technique to force the CEF to use the GPU from our main renderer. This could somehow fix multiple-gpu selection issues. See https://github.com/dsh0416/godot-cef/blob/59617963c2ca193d984a73ead7d951c4a08f1372/gdcef_helper/src/dxgi_hook.rs

[csmoe]

I used an internal cef fork to notify the host app about the gpuinfo chromium's gpu process using while launching finally.

[magreenblatt]

Reopening as it looks like OS capabilities may not be enough here. @csmoe can you share a PR with your changes?

[reitowo]

The SharedImage gets D3D device from here, here, here. I think you can control the selection by setting gpu selection flags at here without hacking system API. And when none of them are set, Chromium uses this to detect device's perference. There're also workaround flags --force_high_performance_gpu for force Chromium use high performance GPU at here. Also read here, here

Could you give more details about fail cases?

[dsh0416]

I think I have missed that there are actually kGpuVendorId and kGpuDeviceId switches at first, since one of our users report that even when the helper process is set to use high performance GPU from the Windows settings, it still got somehow ignored.
But I am still wondering if exposing a physical-device-accurate API (like LUID on Windows or UUID on Linux) is necessary if multiple GPUs with identical device ids are possible (for example, on a workstation).