high performance computing on graphics processing units: hgpu.org

Graphic processing units (GPUs) have gained popularity in scientific computing the recent years. This is because of the massive computing power they can provide for parallel tasks, and while GPUs are powerful, it is also hard to fully utilize their power. A part of this difficulty comes from the many parameters available, and tuning of these is necessary to maximize performance. Tuning consists of finding the best possible combination of those parameters. Separate tuning is needed even for GPUs from the same vendor and same hardware generation. Manually tuning these parameters is tedious and time consuming, and we therefore explore automatic tuning of such parameters. In this thesis we explore this problem for a shallow water simulator. We have successfully applied auto-tuning, making the program do the required tuning by itself. This has yielded an increase of 10-30% in performance, over manual tuning, on different NVIDIA GPU models. In addition we have implemented other mathematical approaches for solving the equations, and shown that which approach that is the fastest is different for different GPUs. A way of automatically selecting the best approach was also implemented.