high performance computing on graphics processing units: hgpu.org

Parallel programming is difficult. For regular computation on central processing units application programming interfaces such as OpenMP, which augment normal sequential programs with preprocessor directives to achieve parallelism, have proven to be easy for programmers and they provide good multithreaded performance. OpenACC is a fork of the OpenMP project, which aims to provide a similar interface for heterogeneous multiprocessing as an alternative for low-level programming models such as CUDA and OpenCL. In this paper the general structure of existing compilers for directive-based programming for GPU’s is presented and some used program analysis and optimization methods are explained. Just like with low-level programming models, optimizing the parallel regions for efficient memory usage is the most important task to achieve good performance.