Prevents multiple queues in the same family from racing on frame data.
Removal of vk_family_data also fixes broken capacity calculation. Two
queue families can share the same index if one is "protected-capable"
and the other isn't.
Make fixes mostly to stabilize CTS. Tested on Windows with RTX 2080 Ti.
Use separate linked-list maps for devices and queues.
Process queues upfront in OBS_CreateDevice, and not lazily in
OBS_GetDeviceQueue.
Handle failed node allocation for OBS_CreateInstance, OBS_CreateDevice,
and OBS_CreateSwapchainKHR.
Handle VK_NULL_HANDLE for OBS_DestroySurfaceKHR and
OBS_DestroySwapchainKHR.
Add temporary code path CTS_WORKAROUND to deal with CTS sharing surfaces
between instances.
Fix return value for OBS_DestroyInstance.
Add missing _freea calls, required for _malloca.
Make sure the queue passed to OBS_QueuePresentKHR supports transfer to
safely call vkCmdCopyImage.
Update GetInstanceProcAddr/GetDeviceProcAddr to conform closer to spec.
Remove extra macros because I dislike indirection more than verbosity.
Reorganize data to avoid best practices layer warning. Sort of a false
positive in our case because we only have one buffer per pool now, but
implicit layers should be clean citizens.
Use general layout and 0 access masks for external synchronization as
specified in the spec.
Also set pipeline stages for maximum synchronization just in case
because it doesn't seem like the pipeline stages are specified.
"Vulkan-incompatible APIs will require the image to be in the GENERAL
layout whenever they are accessing them."
"Whilst it is not invalid to provide destination or source access masks
for memory barriers used for release or acquire operations,
respectively, they have no practical effect. Access after a release
operation has undefined results, and so visibility for those accesses
has no practical effect. Similarly, write access before an acquire
operation will produce undefined results for future access, so
availability of those writes has no practical use. In an earlier version
of the specification, these were required to match on both sides - but
this was subsequently relaxed. These masks should be set to 0."
Don't allow unsupported Vulkan formats to fall back to B8G8R8A8.
Probably better to fail completely than do an illegal copy.
Also remove bad conversion for VK_FORMAT_A2R10G10B10_UNORM_PACK32.
Red and blue channels were reversed, and there's no DXGI equivalent.
Addresses #2796. We can do more later if justified.
Make sure HWND tracking is cleaned up when Vulkan surfaces are
destroyed. Also use unbounded linked list to fix games that leak
surfaces on Alt+Tab like Doom.
Also replace CRITICAL_SECTION with SRWLOCK, both for claimed speed
benefit, and to remove initialization code.
Sometimes the width/height would initially be at 0x0 when the swap chain
was alt-tabbing, causing the capture to fail full-stop when trying to
reacquire.
Game capture assumes the top-level ancestor HWND, not a child HWND.
This fixes bugs such as the VLC window not capturing anymore due to the
fact that the capture HWND of VLC is a child window of the main VLC
window and not the main VLC window itself.
Checks the hook version to ensure compatibility with hook DLL. It's
unlikely it'll ever be necessary to increment the hook version, but this
is just a precautionary thing that allows a hook DLL to make sure it's
rejected by an older OBS version if needed. Again however, very
unlikely that the major version will ever be incremented.
Prevents multiple separate hook DLLs from being loaded in to a target at
once. This is done just in the off-chance that someone might add
another hook to the Vulkan layer registry.
Allows automatically outputting the function name as part of the hook
logging. This really doesn't need to be a manual process. Makes code a
bit cleaner when used.
Before this change, after a game capture source would send a signal to
init or restart a graphics hook, it would respond to any and all hook
ready signals.
With multiple game capture sources in the same scene, a source could
receive the signal intended for another source, and show the wrong
texture.
This change adds the window handle to the name for shared data with the
hook, resulting in hooks for other sources being ignored.
**Commit message modified and clarified by Jim**
When hooking a program that has both DirectX and OpenGL contexts in use,
it is possible to cause a crash on shutdown due to capture_active()
returning true when an OpenGL context is deleted. Normally, when
capturing an OpenGL program, this would not happen because the 'active'
variable would not be set due to OpenGL capture not being initialized,
but if DirectX is captured while an OpenGL context is available, and
OpenGL could not load these required functions, then GL can crash due to
trying to use unavailable functions.
This case is extremely rare and doesn't happen under normal
circumstances; only if a program is using both DirectX and OpenGL within
the same program simultaneously, and *only* if OpenGL could not load the
required functions. This likely almost never happens under normal
programs, games, and hardware. This was apparently produced by hooking
a GL Qt program that used QWebEngine, which used multiple contexts at
once.
Code submissions have continually suffered from formatting
inconsistencies that constantly have to be addressed. Using
clang-format simplifies this by making code formatting more consistent,
and allows automation of the code formatting so that maintainers can
focus more on the code itself instead of code formatting.
