A Multi-Pass, Multi-Stage, Multi-GPU Collision Detection Algorithm

We present the work in progress of a collision detection (CD) algorithm on a multi-threaded environment and based on two graphics processor units (GPUs); the first GPU is used as a normal graphics processor, and the second GPU is doubled as a ‘collision detection’ coprocessor. Our approach makes use of the stream parallel engine provided by modern GPUs together with their native architecture for handling vector operations; thus we aim to perform fast and reliable collision detection. Our initial work involves multi-threading and incorporates a cheap pre-processing stage for storing essential scene data in a format suitable for GPUs (i.e. as textures). It also involves multi-pass rendering in multiple stages. The first stage of collision detection is based on a simple boundary volume collision query. If the result is positive, a second stage is then executed, involving objects pair-wise computations on a per-vertex basis. The approach is being implemented on a single processor NVIDIA Single-Link Interface (SLI) [1] System with two NVIDIA GeForce 6600GT graphics cards.