Relatedly, it'd be cool if one of the AMD people involved in the OpenCL work (which was committed at the same time) could do a writeup of those optimizations.
