Huthwaite, Peter; Imperial College; United Kingdom
Huthwaite, P.; Imperial College London; United Kingdom
Session: Mathematical Modelling - UT
Time: 15:40 - 16:00
The fast generation of accurate numerical wave propagation data is important for research in the fields of geophysics, medicine and non-destructive testing, allowing predictions to be made prior to lab tests, and is also vital in inversion imaging methods. The majority of numerical wave calculation methods are highly parallelisable, with the same calculation being performed many times with different input values. Graphical Processing Units (GPUs) are widely recognised to possess an architecture well suited to such calculations, by allowing many small, lightweight threads to be run in parallel.
The finite element (FE) method is a powerful, flexible method for performing such simulations, having the advantage that arbitrary geometry can be discretised using a free, unstructured mesh. While efficiently allocating a uniform structured mesh to GPU memory is relatively trivial, this key benefit of being able to use free meshes with FE presents a challenge to optimal solution on a GPU.
This presentation discusses Pogo, a finite element solver written to calculate the propagation of elastic waves on a graphics card. The underlying approach for Pogo is outlined, with a solution developed to enable arbitrary structured meshes to be allocated to GPU memory while maintaining efficient memory access. Pogo has been found to allow a speed improvement of approximately two orders of magnitude over typical finite element packages run on the CPU, as will be outlined in the talk.