This is to eventually allow d3d12_swapchain_queue_present() to
operate on a certain VkImage even if current_buffer_index has
already been modified since the corresponding Present() call.
This is an oversight in eec9c3a2f6:
vk_format specifies the format used for the frontend resources. The backend
images could use a different format, depending on what the swapchain
supports.
Also, introduce the terms "frontend" and "backend" for the two sides
of the swapchain, which are probably easier to understand.
Instead of using wined3d_texture_update_desc(). This is safe, because:
* ddraw never exposes wined3d textures directly, and always retrieves them
directly from wined3d when rendering.
* d3d8 and d3d9 (non-extended) will only resize buffers during a reset, and
resetting is forbidden if the application holds any references to the
backbuffers. RTVs are also replaced during a reset, so there is no concern
about retrieving the old RTVs from the device state.
* d3d9ex allows resetting while holding references to the backbuffers, but tests
(fixed by this patch) show that the backbuffers should in fact be recreated.
* dxgi forbids holding references to back buffers during ResizeBuffers(),
including indirect references via command lists or device contexts.
This swapchain is never actually exposed, and effectively only exists due to
implementation constraints. When it was introduced, it was necessary or simplest
to create a dxgi swapchain object, but currently that's no longer necessary, and
avoiding that allows some code to be simplified.
Instead of having d3d11 create the wined3d texture, create it in dxgi instead,
and let d3d11 create a d3d_texture2d object to wrap it.
This is a step towards getting rid of the create_swapchain_texture() callback,
which is mildly difficult to work with and conceptually complex.
From the tests it behaves more like a semaphore than like an event.
I'm not testing explicitly because I'm not sure it's relevant, but
another indication is that calling ReleaseSemaphore() returns
ERROR_ACCESS_DENIED, while calling SetEvent() returns
ERROR_INVALID_HANDLE.
I am still deliberately not testing after which message the swapchain
state changes from the old state to the new until we find a game that
depends on this.
ddraw, d3d8 and d3d9 do not catch recursive SetDisplayMode or Reset()
calls. They will recurse for about 16 times and then return library
specific weird return values like 0xd0000510 (NTSTATUS, but
nonexistent facility) or 0x88760868 (non-existent HRESULT facility). I
suspect it aborts when the kernel-side message processing stack is
exhausted. As such, an application depending on nested calls in those
libraries would have to deliberately do it, but limit the amount of
nesting to a non-zero amount.
Garou: Mark of the Wolves calls IDirect3D9::GetAdapterModeCount() on every
frame. This results in calling EnumDisplaySettingsExW() once per available mode,
which is a very slow operation, both on Windows and Wine.
Manual testing shows that Windows caches the mode list (as well as the adapter
list, which is already cached in Wine) in Direct3D 9 and lower. Calls to
GetAdapterModeCount() and EnumAdapterDisplayModes() are fast, and they also do
not change if monitors are added or removed.
DXGI behaves differently, however. The list of outputs attached to an adapter is
cached—that is, calls to IDXGIAdapter::EnumOutputs() are fast, and return stale
data. However, at least some other calls are slow and do not seem to be cached,
including IDXGIOutput::GetDisplayModeList() and IDXGIOutput::GetDesc().
ddraw is also slow and uncached. Since all testing was done on Windows 10 (for
lack of available older hardware to test with) it is not unlikely that ddraw was
reimplemented over dxgi on newer Windows, and that older Windows versions would
be fast and cached, but this is speculation. In any case I have not included
patches to cache ddraw modes.
Tests were done on Windows 10 21H2, both on real hardware with NVidia drivers
and on software drivers via qemu/KVM. In the latter case only speed could be
tested, but this was consistent with the results from the NVidia machine.
