IN040-05
Efficient implementations for matrix-free solutions of PDEs with libCEED

Tuesday, 15 December 2020: 19:12
Virtual
Valeria Barra, University of Colorado at Boulder, Computer Science, Boulder, CO, United States; California Institute of Technology, Global Environmental Center, Pasadena, CA, United States, Jed Brown, Argonne National Laboratory, Argonne, IL, United States, Jeremy Thompson, University of Colorado at Boulder, Applied Mathematics, Boulder, United States, Leila Ghaffari, University of Colorado at Boulder, Computer Science, Boulder, United States, Yohann Dudouit, Lawrence Livermore National Laboratory, Livermore, CA, United States and Natalie Beams, University of Tennessee, Knoxville, United States
Abstract:
The recent advances in computational resources from current and next generation HPC systems have enabled new modeling strategies for physical systems that now feature finer spatial resolution, higher-order methods and additional complexity, compared to past frameworks. Often, high-performance implementations of high-order numerical methods have been application-specific and architecture fine-tuned, thus difficult to port to new applications or machines.

libCEED is a new open-source library providing efficient evaluation of operators for high-order numerical methods. It offers a purely algebraic interface that is independent of a particular application and is minimally intrusive, enabling rapid development and retro fitting for new and legacy codes. libCEED uses a matrix-free representation and enhances performance-portability by providing run-time selection of implementations tuned for a variety of computational devices including CPUs and GPUs (with both CUDA and HIP support). We introduce libCEED’s interface and demonstrate its usage and performance through examples in computational fluid dynamics and solid elasticity that are of interest to the Earth sciences community at large.

<< Previous Abstract | Next Abstract >>