Hi guys, A couple of days back the Khronos group released their OpenCL 2.0 preview: http://www.khronos.org/registry/cl/
They pretty much extend their memory model, including shared virtual memory for APU-like processors sharing CPU and accelerator hardware on the same chip and also (try to) standardize an LLVM-like IR. As I was discussing with Matt that the OpenCL compilation on the CPU is completely messed up as-is, and considering that Jed played a bit with LLVM recently, it may be worth targeting such an IR for our basic kernels.
Of course this does not solve the memory problems of NUMA and dealing properly with the increasing number of memory channels...
Best regards, Karli
