DOACROSS: BEYOND VECTORIZATION FOR MULTIPROCESSORS.

Abstract

Although vector constructs could be executed with speedup on a multiprocessor, the resources of a multiprocessor are wasted unless there are constructs for which it outperforms a vector processor. The author describes such a method for executing iterative loops on a multiprocessor. Loops are classified by their parallelism and a class of loops is exposed that must execute sequentially on vector machines yet can execute with speedup on a multiprocessor.