high performance computing on graphics processing units: hgpu.org

Parallel architectures with node-level accelerators promise significant performance improvements over conventional homogeneous systems. To cope with the increased complexity of programming such systems various pattern-based programming libraries have become available. In this paper we present our work on providing autotuning capabilities for two runtime libraries that provide parallel programming patterns on state-of-the-art heterogeneous hardware. We present a brief overview of these runtime libraries, outline possible integration with existing tuning frameworks and present initial experimental results.