Donald Bren School of Information and Computer Sciences

Parallelizing a loop for MIMD machines can be described as a process of partitioning it into a number of relatively independent subloops. Previous approaches to partitioning non-vectorizable loops were mainly based on iteration pipelining which partitioned a loop based on iteration number and exploited parallelism by overlapping the execution of iterations. However, the amount of parallelism exploited this way is limited because the parallelism inside iterations has been ignored. In this paper, we present a new loop partitioning technique which can exploit both forms of parallelism - inside and across iterations. While inspired by the VLIW approach, our method is designed for more general, asynchronous, MIMD machines. In particular, our schedule takes the cost of communication into account, and attempts to balance it with respect to parallelism. We show our method is correct, efficient, and produces better schedules than previous iteration level approaches.