Jacek A. Jankowski. Bringing the principal 2Dxy scheme to a GPU, 8th UnTRIM Workshop, Trento, 9-11 May 2011.
The first part of the presentation concerns the process of finding an optimal GPU porting method for the principal two-dimensional, vertically integrated semi-implicit scheme for general free surface flows. Based on the overview of the general purpose GPU programming methodology, the reasons for eliminating the application of compiler directives to an existing code (OpenMP / OpenACC) as well as CUDA Fortran and venturing a completely anew CUDA C/C++ porting applying Thrust and CUSP template libraries are explained and the advantages of their application highlighted.
The second part of the presentation delivers some implementation details characteristic for the general purpose GPU programming and presents simple verification examples for the computational core of the scheme. The reached speedups on a GPU compared to a single-core of a CPU are up to 20-30. An outlook for the future developments on the background of the present hardware evolvement directions is provided.
Transparencies are available.