Developments of the FEniCSx libraries for GPU-accelerated computations
Please login to view abstract download link
The FEniCS libraries are widely used for computing solutions to partial differential equations using the finite element method, providing both C++ and Python interfaces. We present a range of recent developments, including high-performance mixed-topology meshes, automated creation of 'multi-physics' problems with different function spaces on parts of a domain and with couplings between domains (e.g., fluid--structure interaction), and high-performance computations on GPUs from both C++ and Python. In particular, we examine closely the performance of highly optimised kernels on latest generation GPUs, assessed against the hardware limits of the devices. We will then step back and examine how very high performing kernels can be used from high-level, programmable Python interfaces.
