high performance computing on graphics processing units: hgpu.org

In this paper we present a new technique to dynamically adapt the first step (broad phase) of the collision detection process on hardware architecture during simulation. Our approach enables to face the unpredictable evolution of the simulation scenario (this includes addition of complex objects, deletion, split into several objects, …). Our technique of dynamic adaptation is performed on sequential CPU, multi-core, single GPU and multi-GPU architectures. We propose to use off-line simulations to determine fields of optimal performance for broad phase algorithms and use them during in-line simulation. This is achieved by a features analysis of algorithmic performances on different architectures. In this way we ensure the real time adaptation of the broad-phase algorithm during the simulation, switching it to a more appropriate candidate. We also present a study on how graphics hardware parameters (number of cores, bandwidth, …) can influence algorithmic performance. The goal of this analysis is to know if it is possible to find a link between variations of algorithms performances and hardware parameters. We test and compare our model on 1, 2, 4 and 8 cores architectures and also on 1 Quadro FX 3600M, 2 Quadro FX 4600 and 4 Quadro FX 5800. Our results show that using this technique during the collision detection process provides better performance throughout the simulation and enables to face unpredictable scenarios evolution in large-scale virtual environments.