The PE build uses FlsAlloc(), which for our purposes makes no difference vs TlsAlloc(), and allows the use of a destruction callback.
--
v8: vkd3d: Replace the descriptor object cache with a thread-local implementation.
https://gitlab.winehq.org/wine/vkd3d/-/merge_requests/384
An alternative approach to `EnumAdapters2` that does not directly call to GDI driver and therefore does not require GDI driver functions to be expanded. The implementation now lives in `sysparams.c` to make it closer to win32u caches it accesses. It still uses `OpenAdapterFromLuid` internally because that's where adapter handles are actually assigned and to make sure returned handles are usable with other functions that rely on GDI driver being aware of what the handle represents to be able to function correctly (like `QueryVideoMemoryInfo`).
See !4791 for some background and why this version might be preferable over the one submitted there.
--
v5: win32u: Implement NtGdiDdDDIEnumAdapters2.
win32u: Maintain a list of GPUs in cache.
gdi32,win32u: Add stubs for D3DKMTEnumAdapters2 and NtGdiDdDDIEnumAdapters2.
gdi32/tests: Add D3DKMTEnumAdapters2 tests.
https://gitlab.winehq.org/wine/wine/-/merge_requests/4857