A Framework for Running the ADCIRC Discontinuous Galerkin Storm Surge Model on a GPU

Hybrid architectures utilizing GPUs provide a unique opportunity in a high performance computing environment. However, there are many legacy codes, particularly written in Fortran, that can not take immediate advantage of GPUs. Furthermore, many of these codes are under active development and so completely rewriting the code may not be an option. The advanced circulation and storm surge finite element model (ADCIRC) is one such code base. In this paper we present our semi-automatic methodology for porting portions of ADCIRC to run on the GPU and some preliminary scaling results of these subroutines. We have implemented a C++ array class and pre-processor macros to create a type of application framework to simplify the conversion and maintenance tasks. This allows the C++ syntax to be similar to Fortran, to provide for a more straight forward syntactical conversion from the original Fortran to C++ and simplified calling conventions between the two. After the necessary subroutines are converted to the C++ framework, the CUDA library can be easily used and also we are able to provide a simplified abstraction layer for accessing basic GPU functionality. For example, the problem of transferring the correct data on/off the GPU is addressed by our framework by a one time code change and a script to resolve data dependencies. Although it is currently specific to ADCIRC, our framework provides a starting point for utilizing GPUs with legacy Fortran codes, from which more specific GPU optimizations can be implemented.