Uma ferramenta para integração de aplicações Java à computação paralela de propósito geral em processadores gráficos

Abstract

The purpose of this work is to develop a software development tool for generalpurpose computing on graphics processing units (GPGPU). This kind of computation allows the execution of parallelizable tasks with superior performance relative to current CPUs. In these architectures, the programming tools most widely used are the proprietary ones: CUDA and Stream SDK, which are offered by the same companies that make the hardware (GPUs): NVIDIA and AMD/ATI, respectively. It is estimated that those are the programming solutions with the best performance on it’s respective hardware. However, the generated programs, by being tied to the manufacturer’s hardware, lose portability. In this context, the tool proposed in this work consists of a package with Java classes that use the proprietary SDK’s code as part of the solution, to make available a flexible, productivity increasing solution that doesn’t compromise performance. TAJMAHAL, the developed tool, contains two examples of parallel programs: bitonic sort and matrix multiplication. Each of these examples has their own TAJMAHAL project containing a CUDA implementation and a Stream SDK implementation. To test the efficiency of this tool, quantitative and qualitative analyses were made.