Research Of Rendering Acceleration Techniques For Parallel Environments

With over thirty year's development, computer graphics has made great progress both in fundamental theory research and practical application. Its application has been expanded to many areas, including medical and information visualization, computer-aided design and manufacture, city planning and culture heritage conservation, visual analysis and simulation, digital media and entertainment, education and training. Recently, many graphics applications need to deal with massive models, to render models with complex material and motion properties, to present photorealistic lighting effects, and to output to high resolution tiled display walls. All these new features, along with the requirement of real-time performance, greatly challenge current graphics rendering platforms. Three approaches are mainly employed to increase the graphics rendering ability, including improving GPU performance, assembly parallel rendering environments, and combining software rendering acceleration techniques.But most software rendering acceleration techniques are designed for stand-alone PC, instead of parallel rendering environments. Their scalability and performance are unknown for a parallel rendering environment. Combining with the characteristics of parallel rendering environments, this paper proposes novel algorithms and techniques in occlusion culling, mesh simplification, multiresolution modeling, out-of-core data management, mesh layout optimization, and integration with parallel rendering systems and distributed simulation systems. The finale integrated system supports tiled display wall, walkthrough and distributed simulation applications, and real-time rendering of massive meshes and large scale virtual scenes. The main contributions of this paper are followings.An efficient occlusion culling algorithm is proposed which utilizes hardware occlusion query and exploits both spatial and temporal coherence of visibility. A visibility predictor is designed for each geometry node to reduce the count of occlusion queries and the stalls of graphics pipeline. The algorithm can work in a conservative way for high image quality or in an approximate way for time critical rendering. Furthermore, different strategies of parallelizing this occlusion culling algorithm on a GPUs cluster are proposed. Data parallelism strategies decompose the data sets for occlusion query into disjoint parts and map these queries on different cluster nodes for parallel execution. While functionality parallelism strategies assemble an occlusion culling pipeline with multiple cluster nodes which outputs image stream steadily. A number of solutions to some special issues on parallelizing this occlusion culling algorithm are also presented, such as the transferring of data dependency and the load-balancing of occlusion culling pipeline.Simplification and construction the multiresolution representation of massive meshes often cost much time, especially for those algorithms using iterative edge collapse. This paper proposes parallel simplification and parallel multiresolution representation construction techniques which utilize a PCs cluster to speedup the pre-processing stages. For the parallel simplification, mesh cutting and mesh stream schemes are introduced to efficiently generate simplification sub-tasks. For the parallel construction of multiresolution representations, the sub-tasks are produced by dividing the sub-trees of the multiresolution representations. During the course of the parallel simplification and the parallel multiresolution representation construction, two novel techniques are designed to achieve the load balancing for the PCs cluster, i.e. benchmark based resource management and dynamic sub-tasks management. Experimental results show that the proposed parallel simplification techniques have good performance and scalability.The efficiency of storage system access has much influence on the performance of a parallel rendering environment. This paper proposes out-of-core and cache efficient techniques for the parallel rendering system and various rendering acceleration modules to efficiently access scene data among disk, main memory and video memory. Parallel rendering systems for large scale virtual scenes usually adopt retained mode. With this observation, a priority based out-of-core data management framework is presented. The framework associates each data scheduling request with a priority which is computed from the specific data scheduling strategy of a module, and execute the request in a uniform manner. This method greatly facilitates different rendering modules to manage out-of-core data.A scene graph system is designed to efficiently integrate the proposed software rendering acceleration techniques, a PCs cluster based parallel rendering system and a HLA based distributed simulation system. The scene graph system provides multiple views of the scene data with which a module can only deal with the related data structures, supports various object hierarchies which enables modules to choose the most suitable hierarchy, adopts abstract data structures and unified interface which provide good scalability, separates data structures from the visiting procedures which enhances the reusability of the core data structures. The scene graph system also supports rendering task decomposition and load balancing for the parallel rendering system and efficient data exchange between the parallel rendering system and the distributed simulation system.