Automated Design of Parallel Programs for Heterogeneous Platforms Using Algebra-Algorithmic Tools

Abstract

The further development of software tools based on algebra-algorithmic approach and term rewriting technique in the direction of automated design of programs for heterogeneous platforms using OpenCL is proposed. The method for semi-automatic parallelization of loop operators is developed. The particular feature of the approach consists in using high-level algebra-algorithmic program specifications. The tools automate design of parallel programs starting with construction of an algorithm scheme by superposition of operations of Glushkov’s system of algorithmic algebra and then synthesize corresponding source code in a target programming language on the basis of the scheme. The parallelization technique is based on loop tiling and data serialization and uses rewriting rules to transform programs. The application of the approach is illustrated by an example of developing an OpenCL interpolation program used in numerical weather forecasting. The results of the experiment consisting in executing the generated OpenCL program on a graphics processing unit are given.