Instead of using ERR. The intention here is to notify the user that the
Vulkan implementation isn't quite capable of doing what the application
requested, but when libvkd3d is compiled with -DVKD3D_ABORT_ON_ERR, ERR
will abort the process.
This allows running the tests on older NVIDIA GPUs with
-DVKD3D_ABORT_ON_ERR.
VK_DYNAMIC_STATE_VIEWPORT and VK_DYNAMIC_STATE_SCISSOR specify that
viewports and scissor rectangles are dynamic state, but not their
counts.
It took a while to notice this issue because the existing code seemed to
largely work as intended on hardware implementations, but tests using
the additional viewports would fail on llvmpipe.
Vulkan does support rendering without a fragment shader, and this seems
to work fine on current drivers. This commit gets rid of the last
embedded shader binary in libvkd3d, and may have a slight performance
advantage on hardware/drivers able to take advantage of the absence of a
fragment shader.
When a heap was released at the same time as the last resource on that
heap, it was possible for both to see each other's
refcount/resource_count as 0 and both would try to destroy the heap.
Avoid that by converting "resource_count" to an internal refcount, which
holds an extra +1 if the main refcount is nonzero. The final release
will then be synchronized between the two since both will operate on
"internal_refcount".
This effectively moves "null_event_cond" from struct d3d12_fence and
"latch" from struct vkd3d_waiting_event together into a separate
structure, as well as storing the signalling function in struct
vkd3d_waiting_event instead of getting it from struct d3d12_device. I
think that largely makes sense on its own, but storing the signalling
function in struct vkd3d_waiting_event also allows us to more easily
implement d3d12_device_SetEventOnMultipleFenceCompletion() in a
subsequent commit.
