Actually, disregard the memory migration API request. After looking at Intel's OpenCL optimization guide, I am now looking at avoiding copies altogether by allocating properly aligned memory :)
To explain why I initially requested memory migration APIs: Consider a GPU device. Without memory migration APIs, some drivers don't provide hard guarantees about where exactly a memory object "lives", and only that it will be made available when you launch a kernel. The driver is otherwise free to swap stuff in/out of GPU memory. This causes my application to do somewhat unpredicatable memory copies between CPU and GPU on other platforms. However, on Ivy Bridge, it looks like I can avoid memory transfers altogether if the alignment conditions are met.
