* In heavy D3D11 workloads the refcounting overhead especially during fast
binding changes was significant. Refactoring the refcounting to work on a
different model and deferring destruction of objects removes most of the
overhead.
* The spec says we only return a function pointer for device or device-child
functions. In practice the loader wraps instances and physical devices so when
calling a direct GDPA returned function the loader won't unwrap it so we won't
get our proper wrapped objects and will crash.
* It's possible that a frame capture could copy from a query and then reset it,
without then recording something valid into the query. Then the next replay
the copy would be unavailable and if WAIT_BIT is set that would lead to a
device lost.
* We already fill out queries with dummy valid data on create time, so now we
record any queries that are reset and re-fill them with valid data again.
* This helps catches cases where a discarded image is accidentally used and in
many cases may still have valid data. Particularly on Vulkan this is relevant
for DONT_CARE renderpass load and store ops.
- Separate the pixel history copy pixel shader into two separate
shaders, one for colour copy and one for depth
- Allocate and update descriptor sets on demand
- Add another compute shader for pixel history depth copy
* If we don't do this, the application could create a buffer/memory and use it
then destroy it and create another one in the same capture. The driver could
assign the same opaque capture address to both buffers even though they
"overlap" in the capture because they don't overlap in actual execution.
* Slightly artificially extending the life of the resource to the end of the
capture ensures the driver gives them non-overlapping device addresses.
* We preserve each API's interpretation of bit order for packed formats like
RGBA4 or R5G6B5 when displaying the raw data in the UI, but when we need to
proxy it or save to disk, we always transform to D3D's order as standard.
* This allows us to proxy them reliably because we always have a standard bit
order and APIs that need a different order transform when fetching data to the
standard format, or setting proxy data from the standard format.
- For counting the number of fragments, also need to disable the depth
bounds test, since we are not initializing the depth value currently
- Reset depth to 0.0f for shader out, and set it to always pass. So that
we can get depth values from just one fragment.
- Initialize premod value for individual fragment events. This is not
surfaced in the UI, but available from the python API and checked in
tests.
* This also fixes a case where pipelines with dynamic stencil masks wouldn't
have the masks properly set for stencil counting and we wouldn't get shader
output properly.
baldurk