Exploiting Parallelism in Physically-Based Simulations on Multi-Core Processor Architectures

Abstract

As multi-core processor systems become more and more widespread, the demand for designing efficient parallel algorithms propagates also into the field of computer graphics. This is especially true for the physically-based simulation, which is notorious for expensive numerical methods. In this paper we explore possibilities for accelerating these algorithms on modern multi-core architectures. As an application we focus on physically-based cloth simulation. In this context, two distinct problems can be identified: the physical model and the collision handling stage both bearing potential bottlenecks for the simulation. From the parallelization point of view these two components are substantially different. The physical model can be treated efficiently using static problem decomposition. The collision handling problem, however, requires a different approach, due to its dynamically changing structure. We address this problem using multi-threaded programming with fully dynamic task decomposition. Furthermore, we propose a new task splitting approach based on a robust work estimate. The associated data is derived from temporal coherence. Altogether, the combination of different parallelization techniques leads to a concise and yet versatile framework for highly efficient physical simulation.