A Systematical Partitioning Mechanism for Nested Loops with Non-uniform Dependences

Abstract

This paper proposes a generalized and systematical mechanism to exploit parallelism from nested loops with non-uniform dependences. This mechanism could partition all kinds of non-uniform dependence loops effectively. Our approach, based on the dependence convex theory and the optimized integer programming technique, divides loops into parallel partitions with variable size. All non-uniform dependence loops could be classified into four types. Adaptive mechanisms are proposed for each type to minimize the number of parallel regions in the iteration space. Consequently, our mechanism outperforms previous ones not only in the number of parallel partitions but also in the performance in the real machine