TRACO parallelizing compiler

Abstract

This paper presents a source-to-source compiler, TRACO, for automatic extraction of both coarse- and fine-grained parallelism available in C/C++ loops. Parallelization techniques, implemented in TRACO, are based on the transitive closure of a relation describing all the dependences in a loop. Coarse- and fine-grained parallelism is represented with synchronization-free slices (space partitions) and a legal loop statement instance schedule (time partitions), respectively. On its output, TRACO produces compilable parallel OpenMP C/C++ and/or OpenACC C/C++ code. The effectiveness of TRACO and efficiency of parallel code produced by TRACO are evaluated by means of the NAS Parallel Benchmark and Polyhedral Benchmark suites.