Thanks for your reply. This is definitely a bug in the AMD driver if it cannot free memory that has been marked for release and there is a pending memory request that needs satisfying, i.e. a giant memory leak. Might this be fixed or should I give up and wait for Vulkan drivers?
(As a discussion aside on C's free(): while there is no guarantee the C library will return the memory to the OS, the OS is perfectly capable of paging out that (unused) memory. Unless the address space is exhausted (extremely unlikely in 64 bit), the memory that has not been returned to the OS can sit out of the way. Now if AMD's drivers actually paged out memory from the device which would then allow further allocations to occur successfully then we'd be in a different place.....)