General Purpose Parallel Computation On The Modern Graphics Processor Units

With the introduction of the conception stream processor, modern graphics processor units (GPUs) have become a powerful parallel computing hardware. This thesis studies general purpose computing on GPUs in the application of CFD and real-time media processing. The main work of the thesis is as follows:1) Second order projection method in CFD was introduced into parallel computing on GPUs. We define their boundary representation, data structure and their organization. A whole computing flow was designed in the thesis. We also proposed a 3-level programming model to organize our code of general purpose computing on GPUs, which is based on OpenGL system. A set of CFD kernel functions were given.2) We introduced volume fraction multigrid method to this parallel computing model to accelerate convergence rate of the pressure Poisson equation. We improved the method to parallel forms and optimized implementation by using new features of modern graphics processor units. For example, z-cull was used to quickly ignore the computing of boundary cells, linear texture scaling was used to avoid computing texture offsets in fragment programs or vertex programs. We demonstrate the validity and efficiency of the method by numerical experiments.3) We compared the performances of the above computation on GPUs with on CPU. The results illustrated the advantage, disadvantage and features of general purpose computing on GPUs. We analysed its bottleneck and pointed out where is needed to be further improved.4) In the media computing area, we compared the performance of discrete cosine transform (DCT) and its fast algorithm on GPUs, and we get the result that traditional fast algorithms cannot be sure of fastness on GPU parallel computing.5) We proposed a fast and simplified depth of field effect method based on post-filtering method. GPUs presented high performance and more flexibility in this media stream processing.