high performance computing on graphics processing units: hgpu.org

Work involving the use of application acceleration devices is showing great promise, however, there are still major obstacles preventing their widespread adoption. Currently the process of porting applications to an accelerator requires expertise in both the computer science and application domains, due to the lack of abstraction available. We present our work associated with the development of a novel solution to this abstraction problem; an intelligent semi-automatic application porting system, that will allow a higher level of abstraction, to be presented to the end user, while maintaining reasonable performance levels. A prototype system has been constructed that can successfully port applications to Graphics Processing Units (GPUs) and shows promising results in terms of performance comparisons between CPU and GPU. We are presently extending our prototype to other application acceleration devices and to allow the automatic selection of the most appropriate device for an application using Machine Learning techniques. We expect our work and results will be of widespread interest to the increasing community involved in porting code to application accelerators.