[d3x0r]

SwapChainD3D hardcoded DXGI_ALPHA_MODE_IGNORE even though PhysicalDeviceD3D advertises Premultiplied as a supported alpha mode, so a surface configured for per-pixel transparency was validated and then composited as opaque.

DXGI rejects a non-opaque alpha mode on CreateSwapChainForHwnd, so a premultiplied HWND surface is instead created with CreateSwapChainForComposition and bound to the window through a DirectComposition visual whose device, target and visual hang off the swapchain and are moved across when a swapchain is recycled - leaving them behind would destroy the target while Present kept succeeding, and the window would go blank. DXGI_SWAP_CHAIN_FLAG_ALLOW_MODE_SWITCH is masked out for composition swapchains, which DXGI refuses outright, and MakeWindowAssociation is skipped since a composition swapchain is not associated with a window. Recycling an existing swapchain now also requires the alpha mode to be unchanged: AlphaMode is fixed at creation, so without that check a runtime configure() with a different alpha mode silently adopted the old swapchain and dropped the request. dcomp.lib is added to both the CMake and GN builds.

For the transparency to be visible the application's window must have been created with WS_EX_NOREDIRECTIONBITMAP; otherwise the window's redirection surface sits behind the DirectComposition visual and stays opaque.

Split out of google#69 at review request so the Windows side can be looked at separately; the Vulkan half of that change is unaffected by this one.

Verified on Windows 11 + NVIDIA with a Node addon embedding Dawn, driving three.js WebGPURenderer over a WS_EX_NOREDIRECTIONBITMAP window on D3D12. This cannot be demonstrated through ManualSurfaceTest: GLFW implements GLFW_TRANSPARENT_FRAMEBUFFER with DwmEnableBlurBehindWindow, so the window keeps a redirection surface, and WS_EX_NOREDIRECTIONBITMAP has no effect when applied after CreateWindowEx.

[d3x0r]

Thanks for the earlier comments on the combined change - carrying them over as known follow-ups for this PR rather than pre-empting them: the code-history comment, break → if {} else {}, whether UsesComposition's HWND check is redundant with PhysicalDevice::GetSurfaceCapabilities (and whether the capability computation should move here so the logic is local), and caching the DComp device on DeviceD3D through per-backend virtual overrides so the right IUnknown is passed. I have not checked dcomp.lib availability for UWP yet.

Two notes on what's actually on the branch: it's based on current upstream/main (PR google#69 is still on the older base, and they touch disjoint files, so no interaction), and ManualSurfaceTest.cpp is deliberately absent — the note explaining the GLFW limitation belongs on top of the T key from google#69, so it should follow once that lands rather than duplicate approved work here.