TRACO: An automatic loop nest parallelizer for numerical applications

Abstract

We present the source-to-source TRACO compiler allowing for increasing program locality and parallelizing arbitrarily nested loop sequences in numerical applications. Algorithms for generation of tiled code and extracting synchronization-free slices composed of tiles are presented. Parallelism of arbitrary nested loops is obtained by creating a kernel of computations represented in the OpenMP standard to be executed independently on many CPUs. We consider benchmarks, typical from compute-intensive sequences of algebra operations or numerical computation from industry and engineering. The speed-up of programs generated by TRACO are discussed. Related compilers and techniques are considered. Future work is outlined.