GPU : `host_supports_target_device` ends up creating a device

[slomp]

When host_supports_target_device gets called on a GPU target, it ends up creating a device and running a "check" with some data upload and copy.

Would it be possible to do query a device runtime in a different way that would not entail creating a device?

The reason I am asking is because the d3d12 debug layer is kind of a total wreck when it comes to creating devices:

• D3D12 devices (D3D12Device) are singletons per adapter (IDXGIAdapter)
  https://docs.microsoft.com/en-us/windows/win32/api/d3d12/nf-d3d12-d3d12createdevice

  "If a Direct3D 12 device already exists in the current process for a given adapter, then a subsequent call to D3D12CreateDevice returns the existing device."

• ID3D12Debug::EnableDebugLayer() must be called before any device gets created:
  https://learn.microsoft.com/en-us/windows/win32/api/d3d12sdklayers/nf-d3d12sdklayers-id3d12debug-enabledebuglayer

  "Calling this API (ID3D12Debug::EnableDebugLayer) after creating the D3D12 device will cause the D3D12 runtime to remove the device."

This puts us on a precarious situation: if the non-debug d3d12 runtime module is used first, followed by a use of the -debug module, D3D12CreateDevice in the -debug module will fail with DXGI_ERROR_DEVICE_RESET because a device for the same adapter is being created again, but this time it is after EnableDebugLayer() has been called.

Tutorial 12, for example, calls host_supports_target_device() to check that the host supports the d3d12 target, and then may go ahead and create a -debug runtime later to actually run the tutorial.

AFAIK, there's no way around it.

[slomp]

Ok, so it looks like I'd need to release the device with halide_device_release(nullptr, i) before returning from host_supports_target_device().

It doesn't look like halide_device_release() is an exported symbol, though (getting linker error, despite it being declared in HalideRuntime.h)
Should i->device_release(nullptr, i) be used instead?
We might need some device usage reference counting to implement device_release() safely.

[slomp]

I was just brought to the attention of this:
https://microsoft.github.io/DirectX-Specs/d3d/IndependentDevices.html

In a nutshell, it would eliminate the "singleton" nature of D3D12 devices, and let us have non-debug and debug devices from the same adapter coexist together.

The catch: only currently available on Win11 insider builds, with possible release on retail Win10 and 11.