Runtime Translation of the Java Bytecode to OpenCL and GPU Execution of the Resulted Code

Abstract

Modern GPUs provide considerable computation power, often in the range of teraflops. By using open standards such as OpenCL, which provide an abstraction layer over the GPUs physical characteristics, these can be employed to solve general computation tasks. Massively parallel algorithms used in domains such as simulation, graphics, artificial intelligence can greatly expand their application range. It is of great importance for an application to run parts of itself on GPUs and in this respect a number of issues such as OpenCL code generation, data serialization and synchronization between application and GPU must be observed. At the same time, the GPU limitations impose some limits on their applicability to general computation tasks, so an application must carefully assess what parts are suitable for this kind of execution. The computing resources must be abstracted and when possible these should be interchangeable without modifying the source code. We present a new algorithm and library which dynamically generates OpenCL code at runtime for parts of its application in order to run these parts on GPU. Our library automatically handles tasks such as data serialization and synchronization. The practical results are significant and we succeeded in obtaining important speedups using only a straightforward Java implementation of the test algorithm, without any platform specific constructs.