Simple and efficient GPU parallelization of existing H-Matrix accelerated BEM code

TL;DR

This paper presents a simple GPU acceleration method for Hierarchical Matrix-based Boundary Element Method computations, achieving up to 5.5 times faster performance with minimal code modifications.

Contribution

It introduces a GPU offloading strategy for ACA-based BEM routines that can be integrated into existing codebases with minimal changes.

Findings

Speed-up of up to 5.5 times in BEM simulations.

Efficient GPU utilization with minimal code modifications.

Successful application to high-frequency sound scattering problem.

Abstract

In this paper, we demonstrate how GPU-accelerated BEM routines can be used in a simple black-box fashion to accelerate fast boundary element formulations based on Hierarchical Matrices (H-Matrices) with ACA (Adaptive Cross Approximation). In particular, we focus on the expensive evaluation of the discrete weak form of boundary operators associated with the Laplace and the Helmholtz equation in three space dimensions. The method is based on offloading the CPU assembly of elements during the ACA assembly onto a GPU device and to use threading strategies across ACA blocks to create sufficient workload for the GPU. The proposed GPU strategy is designed such that it can be implemented in existing code with minimal changes to the surrounding application structure. This is in particular interesting for existing legacy code that is not from the ground-up designed with GPU computing in mind. Our benchmark study gives realistic impressions of the benefits of GPU-accelerated BEM simulations by using state-of-the-art multi-threaded computations on modern high-performance CPUs as a reference, rather than drawing synthetic comparisons with single-threaded codes. Speed-up plots illustrate that performance gains up to a factor of 5.5 could be realized with GPU computing under these conditions. This refers to a boundary element model with about 4 million unknowns, whose H-Matrix weak form associated with a real-valued (Laplace) boundary operator is set up in only 100 minutes harnessing the two GPUs instead of 9 hours when using the 20 CPU cores at disposal only. The benchmark study is followed by a particularly demanding real-life application, where we compute the scattered high-frequency sound field of a submarine to demonstrate the increase in overall application performance from moving to a GPU-based ACA assembly.