May in some cases improve performance when under memory pressure:
If a dedicated allocation is preferred, but the alloaction fails,
try to allocate memory from an already allocated chunk instead
of falling back to system memory right away.
The default behaviour is now to allow overallocation by default,
which makes sense especially when VK_EXT_memory_priority is used.
Does not seem to affect Nvidia drivers.
The priority defines the importance of any given allocation being
resident in device-local memory rather than system memory, with
values between 0 (low) and 1 (high).
Suballocations will only work on chunks with the requested priority.
This ensures that high-priority allocations are reserved solely for
high-priority resources, and that high-priority resources will be
allocated from a high-priority memory chunk.
This reverts commit 83ae39f727d7aeb43cef277082bcb42093530779.
Does not show any considerable advantage over the 16 MiB chunk size
and reduces the effectiveness of the host-visible device-local memory
type on AMD cards.
Command submission now does not synchronize with the device every single
time. Instead, the command list and the fence that was created for it are
added to a queue. A separate thread will then wait for the execution to
complete and return the command list to the device.
