Data sharing across contexts is possible by mapping the same resource into multiple contexts. Synchronization of multiple contexts is the client’s responsibility, any support is provided for shared-memory access: are there thread-safe CAL\IL instructions ? are there atomic CAL \ IL instructions such as try&set or atomic swap?
In order to limit communication between local and remote memory and perform as much calculus as possible on the GPU, I think it should be very important to have such support.