Accelerating and simulating detected physical interations
Chitalu, Floyd Mulenga
The aim of this doctoral thesis is to present a body of work aimed at improving performance and developing new methods for animating physical interactions using simulation in virtual environments. To this end we develop a number of novel parallel collision detection and fracture simulation algorithms. Methods for traversing and constructing bounding volume hierarchies (BVH) on graphics processing units (GPU) have had a wide success. In particular, they have been adopted widely in simulators, libraries and benchmarks as they allow applications to reach new heights in terms of performance. Even with such a development however, a thorough adoption of techniques has not occurred in commercial and practical applications. Due to this, parallel collision detection on GPUs remains a relatively niche problem and a wide number of applications could benefit from a significant boost in proclaimed performance gains. In fracture simulations, explicit surface tracking methods have a good track record of success. In particular they have been adopted thoroughly in 3D modelling and animation software like Houdini  as they allow accurate simulation of intricate fracture patterns with complex interactions, which are generated using physical laws. Even so, existing methods can pose restrictions on the geometries of simulated objects. Further, they often have tight dependencies on implicit surfaces (e.g. level sets) for representing cracks and performing cutting to produce rigid-body fragments. Due to these restrictions, catering to various geometries can be a challenge and the memory cost of using implicit surfaces can be detrimental and without guarantee on the preservation of sharp features. We present our work in four main chapters. We first tackle the problem in the accelerating collision detection on the GPU via BVH traversal - one of the most demanding components during collision detection. Secondly, we show the construction of a new representation of the BVH called the ostensibly implicit tree - a layout of nodes in memory which is encoded using the bitwise representation of the number of enclosed objects in the tree (e.g. polygons). Thirdly, we shift paradigm to the task of simulating breaking objects after collision: we show how traditional finite elements can be extended as a way to prevent frequent re-meshing during fracture evolution problems. Finally, we show how the fracture surface–represented as an explicit (e.g. triangulated) surface mesh–is used to generate rigid body fragments using a novel approach to mesh cutting.