A growing number of scientific and medical applications are now producing large-scale data, ranging from gigabytes to even terabytes, on a daily basis. To analyze and understand this enormous amount of data, scientific visualization has become an indispensable tool. However, as the size of data increases, it can easily overwhelm the underlying computer system with limited computation power, storage space and network bandwidth. The interactivity of traditional visualization approaches is severely challenged. More advanced solutions are needed. This dissertation is focused on designing efficient visibility culling schemes for scalable visualization systems running in a massively parallel environment. The key idea is to efficiently estimate visible portions of data before the parallel visualization process starts. By utilizing parallel computing power, we are able to speed up the visualization process and visualize large-scale data that cannot be easily handled by one single PC. Visibility culling techniques provide further acceleration for a visualization algorithm by reducing the amount of data sent to the visualization pipeline. Achieving effective visibility culling in a scalable parallel visualization system is the main goal of this research.; In this dissertation, we present several efficient and scalable visibility culling schemes for parallel visualization algorithms. First, we developed a data management and distribution mechanism to ensure the balanced workload with minimal run-time data communication overhead. Second, we proposed a multi-pass visibility culling scheme designed especially for parallel view-dependent isosurface extraction. To speed up the visibility estimation, we introduced a hardware accelerated solution that takes advantage of the occlusion query capability supported by up-to-date graphics hardware. Finally, to minimize the synchronization overhead in a multi-pass solution, we devised a highly scalable visibility culling framework using Plenoptic Opacity Function (POF), an effective way to encode the occluding capability of data in a spatial partition. We showed that such scheme is efficient, effective and scalable in both parallel isosurface extraction and parallel volume rendering. Using the temporal occlusion coherence, we further extended the framework to perform visibility culling in a parallel time-varying volume rendering system with high efficiency.
展开▼
