Automated design and parallelization of programs for heterogeneous platforms using algebra-algorithmic tools

• Datum: 2020
• Hauptverfasser: Doroshenko, А.Yu., Beketov, O.G., Bondarenko, M.M., Yatsenko, О.А.
• Format: Artikel
• Sprache: Ukrainisch
• Veröffentlicht: PROBLEMS IN PROGRAMMING 2020
• Schlagworte: algorithm algebra; automated software design; graphics processing unit; heterogeneous systems; image convolution; OpenCL; parallel computation; software synthesis; UDC 004.4’24
• Online Zugang: https://pp.isofts.kiev.ua/index.php/ojs1/article/view/402
• Назва журналу: Problems in programming

Methods and software tools for automated design and generation of OpenCL programs based on the algebra of algorithms are proposed. OpenCL is a framework for developing parallel software that executes across heterogeneous platforms consisting of general-purpose processors and/or hardware accelerators. The proposed approach consists in using high-level algebra-algorithmic specifications of programs represented in natural linguistic form and rewriting rules. The developed software tools provide the automated design of algorithm schemes based on a superposition of Glushkov algebra constructs that are considered as reusable components. The tools automatically generate code in a target programming language on the basis of the specifications. In most computing problems, a large part of hardware resources is utilized by computations inside loops, therefore the use of automatic parallelization of cyclic operators is most efficient for them. However, the existing automatic code parallelizing tools, such as Par4All, don’t account the limited amount of accelerator’s onboard memory space while real-life problems demand huge amounts of data to be processed. Thus, there is a need for the development of a parallelization technique embracing the cases of massive computational tasks involving big data. In the paper, a method and a software tool for semi-automatic parallelization of cyclic operators based on loop tiling and data serialization are developed. The parallelization technique uses rewriting rules system to transform programs. The framework for parallelization of loops for optimization of computations using graphics processing units allows semi-automatic parallelization of sequential programs. The approach is illustrated on an example of developing a parallel OpenCL image convolution program.Problems in programming 2020; 2-3: 103-114