Just wondering what is the need to express data parallelism in this (formal) way? So that other ISPC-like implementations could be built using this?