基于CUDA的柔软织物仿真研究

发布时间：2019-04-02 07:41

【摘要】：柔软织物与我们的生活息息相关,因此它成为了虚拟现实技术中一类重要的仿真对象。早在19世纪80年代,研究人员便开始关注基于物理的柔软织物的仿真研究。经过30多年的发展,广泛应用于游戏娱乐、影视动画、服装设计等多个领域。尽管柔软织物仿真的相关技术已经非常丰富,在某些方面,依旧面临巨大挑战,比如如何提升模拟织物的褶皱细节时的时间性能。而为了在不损失真实度的条件下显著提高仿真的时间性能,人们开始越来越多地关注基于GPUs对布料仿真进行并行化仿真的研究。本文便旨在基于CUDA架构对柔软织物仿真的相关算法进行并行化,优化仿真系统的时间性能。论文从以下几个方面进行展开:首先,介绍了柔软织物仿真相关的软件和硬件技术以及并行化编程的基础,例如GPU并行编程的硬件架构及并行编程思想等。为后面对数值解算及碰撞检测算法的并行化提供了理论依据。然后分为几个部分介绍仿真系统的主要构成部分及其算法研究。第一步为对柔软织物物理建模。并通过对比分析目前主流的三种柔软织物建模方法,选择了物理建模技术作为本文的重点研究方法。接下来详细描述物理建模方法中的质点-弹簧系统理论,从模型的构建、系统的受力分析两个方面阐述质点-弹簧系统,并分析解决质点-弹簧系统的“超弹性”问题。然后,对比质点弹簧模型的几种动力学积分解算方法,选择了易于进行并行化设计且相对较为稳定、高效的Verlet积分法。最后对Verlet积分法的并行化做了研究,提出共享内存与线程的动态映射方法以保证了并行化计算时的负载均衡。接下来是柔软织物仿真系统中的碰撞检测部分进行并行化。首先介绍并对比几种常用的碰撞检测方法,结合柔软织物易形变的特性,最终选择了 AABB层次包围盒方法。然后针对AABB层次树,提出了一种改进的二叉基数布局方法,以便能够用完全并行的方法创建AABB包围盒的层次结构,以及AABB包围盒的计算。另外,通过改进遍历树的并行递归算法,实现了并行迭代遍历算法,极大提升了该部分的时间性能。最后,给出基于CUDA柔软织物仿真系统的实现结果。为了便于分析论文中提出的数值解算及碰撞检测算法各自对系统性能的影响,通过设计合适的仿真场景,分别针对这两个算法进行了实验,并给出了用于对比的串行实现结果。最后是整个系统的仿真实现结果。
[Abstract]:Soft fabric is closely related to our lives, so it has become an important simulation object in virtual reality technology. As early as the 1880s, researchers began to focus on physical-based simulation of soft fabrics. After more than 30 years of development, widely used in games and entertainment, film and television animation, clothing design and many other fields. Although the technology of soft fabric simulation is very rich, there are still great challenges in some aspects, such as how to improve the time performance of the wrinkle details of the simulated fabric. In order to improve the time performance of simulation without loss of authenticity, people pay more and more attention to the parallel simulation of cloth simulation based on GPUs. The aim of this paper is to parallelize the algorithms of soft fabric simulation based on CUDA architecture to optimize the time performance of the simulation system. The thesis is carried out from the following aspects: firstly, the software and hardware technologies related to soft fabric simulation and the basis of parallel programming, such as the hardware architecture of GPU parallel programming and the idea of parallel programming, are introduced. It provides a theoretical basis for later numerical solution and parallelization of collision detection algorithm. Then it is divided into several parts to introduce the main components of the simulation system and its algorithm research. The first step is the physical modeling of soft fabrics. By comparing and analyzing the three mainstream soft fabric modeling methods, the physical modeling technology is chosen as the key research method in this paper. Secondly, the theory of particle-spring system in physical modeling method is described in detail. The particle-spring system is expounded from two aspects: the construction of model and the force analysis of the system, and the problem of "hyperelasticity" of particle-spring system is analyzed and solved. Then, by comparing several dynamic integral solutions of particle spring model, a relatively stable and efficient Verlet integration method is selected, which is easy to parallel design and is relatively stable. Finally, the parallelization of Verlet integration method is studied, and a dynamic mapping method between shared memory and thread is proposed to ensure the load balancing in parallel computation. Then the collision detection part of the soft fabric simulation system is parallelized. Firstly, several commonly used collision detection methods are introduced and compared. Combined with the deformation-prone characteristics of soft fabrics, the AABB hierarchical bounding box method is finally selected. Then, an improved binary cardinality layout method for AABB hierarchical tree is proposed, so that the hierarchical structure of AABB bounding box and the calculation of AABB bounding box can be created by completely parallel method. In addition, the parallel iterative traversal algorithm is implemented by improving the parallel recursive algorithm of traversal tree, which greatly improves the time performance of this part. Finally, the results of soft fabric simulation system based on CUDA are given. In order to analyze the influence of the numerical solution and collision detection algorithm on the performance of the system, the experiments of the two algorithms are carried out by designing suitable simulation scenes, and the serial implementation results for comparison are given. Finally, the simulation results of the whole system are given.
【学位授予单位】：哈尔滨工程大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.9;TS101.9

【参考文献】