James Trotter: Automated code generation and GPU acceleration for finite element methods in FEniCS
Abstract: Graphics processing units, or GPUs, offer significantly increased performance for some scientific computing workloads. But in the case of finite element simulations on unstructured meshes, the benefits of using GPUs are still the subject of an ongoing discussion for which there is no clear conclusion. We describe our work on improving the GPU acceleration of a finite element solver framework called FEniCS, where code is automatically generated for the user from a high-level description of their finite element problem. We use automated code generation to offload the assembly of linear systems to a GPU, while taking care that data transfers between CPU and GPU do not become a performance bottleneck. We provide examples to show that GPUs and automated code generation can be used to accelerate finite element solvers. Even though more work is needed to find efficient GPU-based linear solvers, our improvements to FEniCS can be used as a starting point for exploring the potential of GPU acceleration for finite element simulations.
This talk is part of the Mechanics Lunch Seminar series. That means 20min talks plus discussion in an informal setting.
Zoom: To obtain the Zoom meeting details please contact Timo Koch (timokoch at math.uio.no).