基于GPU的HASM动态模拟与实时渲染方法

 基于微分几何曲面论的高精度曲面模拟(high accuracy surface modeling, HASM)需要大量的复杂密集计算,在CPU上模拟极为耗时,使得在现有的硬件条件下,实时动态模拟曲面并实时可视化极具挑战性。论文提出了GPU加速的HASM方法,充分利用现代显示适配器(graphic processor unit, GPU)技术,运用GPU最新发展起来的并行计算能力,使用并行化的预处理共轭梯度方法解算曲面,完成曲面模拟,并同时利用GPU的高速缓存架构,对渲染操作进行充分优化,以实现高效实时可视化。HASM需要的有限差分离散和高速解算操作,均充分利用现代GPU架构,所具有的多处理器和众多的流处理器所产生的强大并行计算能力,可视化也用GPU高速缓存技术和三角条带方法进行充分优化。数值实验和实际项目区高程模拟实验均表明, 在GPU为NVIDIA quadro 2000和CPU为DualCore Intel Core 2 Duo E8400的硬件配置下,GPU并行化的曲面模拟方法比普通方法速度提高了约10倍,使得动态模拟与可视化算法可以达到交互式的帧速及实时可视化的要求。

A Novel Method for Dynamic Modelling and Real-time Rendering Based on GPU

High accuracy surface modeling （HASM） is a method for building surface with high accuracy in terms of sampled points, which is based on the fundamental theorem of surfaces and gives a solution to error problem in geographic information system （GIS）. Previous studies show this method can model surface with much higher accuracy than other classical methods widely used in GIS, while its speed is limited because of its huge computational cost. In order to accelerate its computational speed, a novel parallel and interactive method of HASM for real-time rendering on GPU using compute unified development architecture （CUDA） is presented, which allows for efficient and high quality visualization. The computational task of HASM is parallelized and run simultaneously on many cores of modern GPUs which have multiprocessors and many stream processors, which can improve the performance significantly. Coupled with an efficient rendering method, dynamic surface simulation and real-time rendering is done concurrently on GPU. Preconditioned conjugate gradients methods are used to solve the huge linear systems arising from HASM. Fully harnessing the processing power of modern GPUs with a highly parallel architecture and multiprocessors and many stream processors, we can simulate the surface dynamically and post it to ren- dering pipeline simultaneously. Making use of state-of-the-art GPU techniques such as vertex buffer ob- ject, texture buffer, the rendering can be carried out with a very high efficiency. A few experiments were carried out including some digital elevation model constructions and the tests results showed that our method can construct surface dynamically and visualize it at very high frame rates.