Some devices like Android smartphones have displays which are
rotated, meaning the compositor needs to rotate its output.
Add support for this to the compute compositor by rotating the
contents of the textures it uses for distortion lookups. This
requires postponing the calculation of that texture and adding
code to recreate it if the rotation changes (which is allowed,
but unlikely to happen in practice.)
The vk bundle currently creates a single queue, calls vkGetDeviceQueue,
and stores it in the bundle. There's only ever 1 queue from 1 queue
family right now, even when using the compute compositor.
It doesn't really make sense for the renderer to re-request this queue
instead of using the one in the bundle. If we wanted to store a local
version to it, it should probably be assigned directly (i.e. r->queue =
vk->queue) to ensure that it's the same. If we decide we want multiple
queues later, that would be a good time to add this sort of thing back.
Currently, there is a single command pool in the vk bundle, shared by
everyone. Since command pools (and command buffers allocated from those
pools) can only be used on one thread at a time, this requires locking.
However, the main point of having these annoying command pool things in
the first place is that you can use one for each thread/lifetime/area in
the app and avoid the overhead of the locks (both computational and
cognitive).
In this change I have given the rendering bits of the compositor its own
command pool. Instead of allocating and freeing a command buffer every
frame, a single command buffer is allocated from the pool during
initialization, and the pool is reset at the beginning of each frame.
Normally, multiple pools would need to be used, but this is not
necessary in monado because frames are serialized. The `TRANSIENT` and
`ONE_TIME_SUBMIT` flags have been added, which can allow for some driver
optimizations. The render code no longer takes out the command pool
mutex. The shared command pool is still there for a few remaining
places where vulkan work needs to be done outside the compositor.
I used the command buffer vulkan helpers when possible, but I would
maybe propose the idea of removing them, since they aren't really
wrapping much at this point. The `C` macro helps a lot and it's a bit
easier to see the Vulkan details in front of you instead of needing to
switch back and forth between the helper.
Later, I think it would be cool to apply and document some constraints
like "the queue is only accessed in functions XYZ, the render_resources
command pool must only be accessed in layer_commit from 1 thread" etc.
This was only there to satisfy a validation layer bug, which appears
to have since been fixed. It doesn't make sense to transition an XR
swapchain image to PRESENT_SRC, and the client transitions from
UNDEFINED anyway when acquiring the image.
This also fixes a case where monado violates the OpenXR specification:
> The list of OpenXR functions where the OpenXR runtime may access the
> VkQueue are:
> - xrBeginFrame
> - xrEndFrame
> - xrAcquireSwapchainImage
> - xrReleaseSwapchainImage
> The runtime must not access the VkQueue in any OpenXR function that is
> not listed above or in an extension definition.
`queue_upload_for_first_level_and_layer` uploads pixels to an image.
It performs a layout transition, does the copy, and does another layout
transition. There is an execution dependency between the the copy and
the second layout transition, but the memory dependency was missing.
Caught/tested via enabling synchronization validation.
