__global memory on-chip

Hi folks,

Is it possible to have in some way a __global memory element or buffer implemented in on-chip registers or memory?

It would be very useful for performing atomic operations from many work-items. Otherwise, the atomic functions on data in off-chip RAM is way too slow. It would also be great for transferring some data from kernel to kernel without having to go to off-chip RAM.

It could be doable in OpenCL 2.0 with the program scope variables in global address space. In fact, Xilinx has it...