We do uploads asynchronously, and we map the software frames in
order to avoid 2-stage copying. However, whilst we added a dependency
upon the mapped buffers, we did not add the original frame backing
those buffers as a dependency.
This caused issues on RADV, particularly with RGB images.
Push descriptors are in theory slightly faster, but come with
limitations for which we have to check.
Either way, they're not difficult to implement, so even though
no one should be using peasant-tier descriptors, do it anyway.
This permits:
- The use of Vulkan filtering on many more devices
- Better debugging due to lack of descriptor buffer support in layers
Much of the changes here are due to a requirement that updates to
descriptors must happen between the command buffer being waited on,
and the pipeline not being bound.
We routinely did it the other way around, by updating only after
we bind the pipeline.
More recent kernel versions allow for users to extract a sync_file
handle from a DMA-BUF, which can then be imported into Vulkan as a
binary semaphore.
This finally allows for synchronization between Vulkan and DMA-BUF
images, such as those from screen capture software, or VAAPI,
avoiding any corruption artifacts.
This is done fully asynchronously, where we use the kernel's
given binary semaphores as a dependency to increment the image's
usual VkSemaphores we allocate. The old imported binary semaphores
are cleaned up after execution as usual.
In the future, hwcontext_drm should receive support for explicitly
synchronized images as well, which would make the synchronization
more robust and portable.
The issue is that we ask for storage images by default if
available, but because that is gated by the format supporting
storage images, and the check for the format supporting storage
images is gated by the usage, this resulted in a catch-22.
Vulkan encoding was designed in a very... consolidated way.
You had to know the exact codec and profile that the image was going to
eventually be encoded as at... image creation time. Unfortunately, as good
as our code is, glimpsing into the exact future isn't what its capable of.
video_maintenance1 removed that requirement, which only then made encoding
images practically possible.
The issue is that enabling features requires that the device
extension is supported. The extensions bitfield was set later,
so it was always 0, leading to no features being added.
The validation layer option only supported GPU-assisted validation.
This is mutually exclusive with shader debug printfs, so we need to
differentiate between the two.
This also fixes issues with user-given layers, and leaks in case of
errors.
Fixes:
vkCreateDevice(): pCreateInfo->pNext<VkPhysicalDeviceOpticalFlowFeaturesNV> includes a
pointer to a VkPhysicalDeviceOpticalFlowFeaturesNV, but when creating VkDevice, the
parent extension (VK_NV_optical_flow) was not included in ppEnabledExtensionNames.
The Vulkan spec states: Each pNext member of any structure (including this one) in
the pNext chain must be either NULL or a pointer to a valid struct for extending
VkDeviceCreateInfo.
This commit was long overdue. The old transfer dubiously tried to
merge as much code as possible, and had very little in the way
of optimizations, apart from basic host-mapping.
The new code uses buffer pools for any temporary bufflers, and
handles falling back to buffer-based uploads if host-mapping fails.
Roundtrip performance difference:
ffmpeg -init_hw_device "vulkan=vk:0,debug=0,disable_multiplane=1" -f lavfi \
-i color=red:s=3840x2160 -vf hwupload,hwdownload,format=yuv420p -f null -
7900XTX:
Before: 224fps
After: 502fps
Ada, with proprietary drivers:
Before: 29fps
After: 54fps
Alder Lake:
Before: 85fps
After: 108fps
With the host-mapping codepath disabled:
Before: 32fps
After: 51fps
