high performance computing on graphics processing units: hgpu.org

Heterogeneity is omnipresent in today’s commodity computational systems, which comprise at least one multi-core Central Processing Unit (CPU) and one Graphics Processing Unit (GPU). Nonetheless, all this computing power is not being exploited in mainstream computing, as the programming of these systems entails many details of the underlying architecture and of its distinct execution models. Current research on parallel programming is addressing these issues but, still, the systems’ heterogeneity is exposed at language level. This paper proposes a uniform framework, grounded on the Single Operation Multiple Data model, for the programming of such heterogeneous systems. The model is declarative, empowering the compiler to generate code for multiple architectures from the same source. To this extent, we designed a simple extension of the Java programming language that embodies the model, and developed a compiler that generates code for both multi-core CPUs and GPUs. A performance evaluation attests the validity of the approach that, despite being based on a simple programming model, is able to deliver performance gains on par with hand-tuned data parallel multi-threaded Java applications.