high performance computing on graphics processing units: hgpu.org

In this paper we compare a Haskell system that exploits a GPU back end using Obsidian against a number of other GPU/parallel processing systems. Our examples demonstrate two major results. Firstly they show that the Haskell system allows the applications programmer to exploit GPUs in a manner that eases the development of parallel code by abstracting from the hardware. Secondly we show that the performance results from generating the GPU code from Haskell are acceptably comparable to expert hand written GPU code in most cases; and permit very significant performance benefits over single and multi-threaded implementations whilst maintaining ease of development. Where our results differ from expert hand written GPU (CUDA) code we consider the reasons for this and discuss possible developments that may mitigate these differences. We conclude with a discussion of a domain specific example that benefits directly and significantly from these results.