* On get we return valid but basically empty caches.
* We prevent any caches making it to the driver.
* Also our UUID is a bit more sane, it's "rdocTIMESTAMP" instead of
just random numbers. Still not going to clash and still unique every
run to prevent valid apps from trying to reuse a cache.
* This is better as VK_INCOMPLETE is actually the wrong error code to be
returning.
* I don't like the fact that it doesn't "just work" but this is mostly
limited by design decisions on the side of the vulkan loader.
* There is no good way with the loader to say 'please also include this
layer in the enumerated lists'. There is an all-or-nothing override of
layer searching, but that would break any layers the application might
actually rely on.
* On balance I've decided to go with this method, as it's a one-off
interruption for the user (unless someone is constantly switching
between installs).
* Rather than explicitly having vkQueueSubmit as a node with command
buffer children, and those with the contents, we now inline everything
and just add labels at beginning and end.
* Also tweaked slightly the fake pass algorithm to handle labels being
present and merging a bit more aggressively to merge adjacent command
buffers that are doing the same pass.
* The linux json file has a different name but this isn't important.
Currently I'm not sure how the json/lib will be placed (will it be
installed somewhere and the json point at /opt/renderdoc/ or what?)
* The reason being that the actual recorded events are interleaved with
others (either on the queue or the cmd buffer), so we have to skip
over those. The events are duplicated where they are inserted by EID
and refer to the original offsets.
* Any uses of this for replaying subsections of the frame are to be
implemented under the replay driver level, as it's too onerous to
require that the replay driver be able to replay arbitrary subsections
or to communicate granularity up to the above layer.
baldurk