high performance computing on graphics processing units: hgpu.org

The need to speed-up computing has introduced the interest to explore parallelism in algorithms and parallel programming. Technology is evolving fast but computing power in sequential execution is not increasing as much as earlier but CPUs contain more and more parallel computing resources. However, parallel algorithms may not be able to exploit all the parallelism in computers. The key issue is that algorithms need to be divided in independent parts to be executed at the same time. By using suitable parallel processors, such a GPU, we can address this problem and explore possibilities for higher speed-ups in computation. High performance calls for efficient parallel architecture but also the tools used to convert high level program description to parallel machine instructions, like languages, compilers. are equally important. This thesis remarks on the importance of using suitable languages to describes the algorithm to exploit the parallelism. This thesis discusses the following parallel programming languages: CUDA, OpenCL, OpenACC and Halide. The programming model and how they run on parallel processors. Each language has its different properties and we discuss portability, scalability, architectures and programming models. In the thesis these languages are compared and advantages and drawbacks are considered.