high performance computing on graphics processing units: hgpu.org

Computing systems have become increasingly complex with the emergence of heterogeneous hardware combining multicore CPUs and GPUs. These parallel systems exhibit tremendous computational power at the cost of increased programming effort. This results in a tension between achieving performance and code portability. Code is either tuned using device-specific optimizations to achieve maximum performance or is written in a high-level language to achieve portability at the expense of performance. We propose a novel approach that offers high-level programming, code portability and high-performance. It is based on algorithmic pattern composition coupled with a powerful, yet simple, set of rewrite rules. This enables systematic transformation and optimization of a high-level program into a low-level hardware specific representation which leads to high performance code. We test our design in practice by describing a subset of the OpenCL programming model with low-level patterns and by implementing a compiler which generates high performance OpenCL code. Our experiments show that we can systematically derive high-performance device-specific implementations from simple high-level algorithmic expressions. The performance of the generated OpenCL code is on par with highly tuned implementations for multicore CPUs and GPUs written by experts.