mmap()ing device memory


As a complement to the zero-copy capabilities of 2.3, is it likely that it will become possible to mmap() GPU device memory into a host-side process’s address space?

(Or perhaps it’s already possible and I’ve missed something obvious…)



Another question is wether is it possible to copy thing using cudaMemCpy() from host to gpu, or reverse WHILE a kernel is running!

Anyway all that may be emulated using Pinned Mapped Memory, but it’s interesting to have other tools :-)