Techniques for Autotuning Algorithms on Heterogenous Platforms

Proceedings of the First PhD Symposium on Sustainable Ultrascale Computing Systems (NESUS PhD 2016) Timisoara, Romania. February 8-11, 2016.Current GPUs (Graphic Processing Units) can obtain high computational performance in scientific applications. Nevertheless, programmers have to use suitable parallel algorithms for these architectures and have to consider optimization techniques in the implementation in order to achieve that performance. This thesis is focused on designing and implementing parallel prefix algorithms into GPU architectures with little effort. For that, we have developed a very optimized library called BPLG (Tuning Butterfly Processing Library for GPUs) and based on a set of building blocks that enable to easily design well-known algorithms such as FFT, tridiagonal systems solvers, scan operator, sorting or signal processing. This library is designed under a tuning methodology based on two-stages indentified as GPU resource analysis and operator string manipulation. Specifically, this strategy is focused on a set of parallel prefix algorithms that can be represented according to a set of common permutations of the digits of each of its element indices [4], denoted as Index-Digit (ID) algorithms. So far, the proposed methodology has obtained very good results with respect to state-of-art libraries, as CUFFT, CUSPARSE, CUDPP or ModernGPU.European Cooperation in Science and Technology. COS