high performance computing on graphics processing units: hgpu.org

Standard PC hardware rapidly increases in parallel computing power in form of multicore CPUs and general purpose GPUs. To take advantage of this situation it is necessary to create specialized code. This is a very time consuming and therefore an expensive task. One approach on solving this problem is the OpenCL (Open Computing Language) standard. It offers the possibility to run the same code on different hardware platforms. OpenCL provides code portability but not performance portability. This paper introduces the concept of a new developed self-optimizing parallel programming framework that addresses the issue of performance portability. This framework provides a set of algorithm building blocks. With the help of these building blocks a wide range of algorithms can be described, that work on one- or two-dimensional objects like images, vectors and matrices. The achieved performance is demonstrated with different algorithms on standard computer hardware platforms.