Toward Performance-Portable Finite Element Methods on High-Performance Systems

摘要

We aim at performance portability of finite-element methods for solving PDE (Partial Differential Equations) on high-performance systems, i.e., ensuring that the same application code achieves good performance across a broad class of parallel architectures, including multi-core CPUs and many-core GPUs from different vendors. We achieve this goal by extending the popular PDE solving framework DUNE and integrating it with our parallel programming framework PACXX. We describe the implementation of our approach and we show how it allows the application programmers to use single C++ source code on a variety of target architectures by compiling it with different back-ends of the PACXX framework, achieving competitive performance against manually optimized original DUNE kernels.