Run-time characterization of irregular accesses applied to parallelization of irregular reductions

Abstract

Irregular reduction operations are the core of many large scientific and engineering applications. There are, in the literature, different methods to solve these operations in parallel. In this paper we discuss a new technique which improves performance significantly, both in terms of execution time and memory overhead. These improvements are achieved in the preprocessing as well as in the resulting parallel code. Our proposal is based on the use of the Irregular Access Region Descriptor (IARD). This data structure is a compact characterization of indirectly accessed arrays that can be used for the efficient parallelization of a wide spectrum of irregular codes. In this paper we present its application to parallelize irregular reduction operations.