It seems to me that "data parallel model" just means that the software programmer's work stops at decomposition. Assignment, orchestration and mapping are all handled by compiler, runtime and/or hardware.

I'd agree with the above comment. However, what happens if tasks are declared as being 'data parallel' but actually aren't? Are these systems smart enough to intervene/reject compilation? It seems relatively difficult to infer that the above operation is 'data parallel' for a compiler?