Dubach, Christophe

Steuwer, Michel

Fumero Alfonso, Juan José

2018-02-08T16:40:43Z

2018-02-08T16:40:43Z

2017-11-30

Nowadays, most computer systems are equipped with powerful parallel devices such as Graphics Processing Units (GPUs). They are present in almost every computer system including mobile devices, tablets, desktop computers and servers. These parallel systems have unlocked the possibility for many scientists and companies to process significant amounts of data in shorter time. But the usage of these parallel systems is very challenging due to their programming complexity. The most common programming languages for GPUs, such as OpenCL and CUDA, are created for expert programmers, where developers are required to know hardware details to use GPUs. However, many users of heterogeneous and parallel hardware, such as economists, biologists, physicists or psychologists, are not necessarily expert GPU programmers. They have the need to speed up their applications, which are often written in high-level and dynamic programming languages, such as Java, R or Python. Little work has been done to generate GPU code automatically from these high-level interpreted and dynamic programming languages. This thesis presents a combination of a programming interface and a set of compiler techniques which enable an automatic translation of a subset of Java and R programs into OpenCL to execute on a GPU. The goal is to reduce the programmability and usability gaps between interpreted programming languages and GPUs. The first contribution is an Application Programming Interface (API) for programming heterogeneous and multi-core systems. This API combines ideas from functional programming and algorithmic skeletons to compose and reuse parallel operations. The second contribution is a new OpenCL Just-In-Time (JIT) compiler that automatically translates a subset of the Java bytecode to GPU code. This is combined with a new runtime system that optimises the data management and avoids data transformations between Java and OpenCL. This OpenCL framework and the runtime system achieve speedups of up to 645x compared to Java within 23% slowdown compared to the handwritten native OpenCL code. The third contribution is a new OpenCL JIT compiler for dynamic and interpreted programming languages. While the R language is used in this thesis, the developed techniques are generic for dynamic languages. This JIT compiler uniquely combines a set of existing compiler techniques, such as specialisation and partial evaluation, for OpenCL compilation together with an optimising runtime that compile and execute R code on GPUs. This JIT compiler for the R language achieves speedups of up to 1300x compared to GNU-R and 1.8x slowdown compared to native OpenCL.

http://hdl.handle.net/1842/28718

en

The University of Edinburgh

Juan J. Fumero, Michel Steuwer, Christophe Dubach. A Composable Array Function Interface for Heterogeneous Computing in Java. In Proceedings of ACM SIGPLAN International Workshop on Libraries, Languages, and Compilers for Array Programming, June 2014.

Juan J. Fumero, Toomas Remmelg, Michel Steuwer, Christophe Dubach. Runtime Code Generation and Data Management for Heterogeneous Computing in Java. In Proceedings of ACM SIGPLAN International Conference on Principles and Practices of Programming on the Java Platform, September 2015.

Juan Fumero, Michel Steuwer, Lukas Stadler, Christophe Dubach. OpenCL JIT Compilation for Dynamic Programming Languages. MoreVMs Workshop on Modern Language Runtimes, Ecosystems, and VMs, April 2017.

Juan Fumero, Michel Steuwer, Lukas Stadler, Christophe Dubach. Just-In-Time GPU Compilation for Interpreted Languages with Partial Evaluation. In Proceedings of ACM SIGPLAN/SIGOPS International Conference on Virtual Execution Environments (VEE), April 2017.

GPU Computing

interpreted programming languages

JIT compilation

runtime optiimisations

Java

R (Computer program language)

R

Accelerating interpreted programming languages on GPUs with just-in-time compilation and runtime optimisations

Thesis or Dissertation

Doctoral

PhD Doctor of Philosophy