Fall back to the system D3D12 runtime when the Agility SDK device factory cannot create a device

[Gianluca Mazza (gianlucamazza)]

Problem

CreateDmlObjects (src/dml/dml_helpers.cpp) tries the Agility SDK device factory first and throws on any CreateDevice failure instead of reaching the existing system-runtime fallback branch:

if (SUCCEEDED(D3D12GetInterface(...)) && SUCCEEDED(sdk_config->CreateDeviceFactory(614, module_path, ...))) {
  THROW_IF_FAILED(d3d12_factory->CreateDevice(...));   // throws — fallback below is unreachable
} else {
  ...
  THROW_IF_FAILED(D3D12CreateDevice(...));
}

A known failure is DXGI_ERROR_ALREADY_EXISTS (0x887A0036): the process already holds a D3D12 device created with the system runtime — e.g. the XAML/WinUI compositor in a packaged app creates one at Window.Activate() — and devices from different D3D12 runtimes cannot coexist in one process. The plain D3D12CreateDevice branch handles exactly this situation, but is unreachable.

Whether this bites is OS-dependent: if the in-box D3D12 is older than the requested SDK version (614), CreateDeviceFactory itself fails and the fallback runs — an OS update silently flips the branch and breaks OgaCreateModel in XAML hosts.

Fix

Treat a factory CreateDevice failure like a factory creation failure: log a warning with the HRESULT and fall back to D3D12CreateDevice.

No regression for #612 (which introduced the Agility path): the new fallback only runs when the in-box runtime is >= the requested SDK version — i.e. the system runtime already contains the fixes the Agility dependency was added for. Related: #1054 (the pre-existing fallback branch for factory-creation failure).

Validation (real hardware)

Xbox Series S, Dev Mode UWP (OS 26100), ORT GenAI 0.13.2 DirectML + ORT 1.24.4, SmolLM2-360M INT4 (DML model built with the model builder):

Scenario	vanilla DLL	patched DLL
XAML app + DML EP	887A0036 at OgaCreateModel (reproduced 3x)	loads in 886 ms, weights resident on GPU (315 MB), full decode (739 tokens)
XAML app + CPU EP	works	works (67.2 tok/s, unchanged)
Headless (no compositor) + DML EP	works via Agility path	works via Agility path

[Copilot]

Pull request overview

This PR improves DirectML initialization robustness on Windows by allowing CreateDmlObjects to fall back to the in-box (system) D3D12 runtime when the D3D12 Agility SDK device factory is present but fails to create a device (e.g., DXGI_ERROR_ALREADY_EXISTS in XAML/WinUI hosts that already created a system-runtime D3D12 device).

Changes:

• Treat Agility ID3D12DeviceFactory::CreateDevice failures as non-fatal and fall back to D3D12CreateDevice.
• Add diagnostics that include the failing HRESULT when the Agility factory device creation fails.

[Gianluca Mazza (gianlucamazza)]

@microsoft-github-policy-service agree