Design And Implementation Of Parallel Algorithms For Some H.264Encoding Modules Based On GPU

Video provides us with rich amount of information, at the same time video compression technology is becoming more and more popular. H.264is an excellent video compression standard with high performance, but its high compression rate is at the cost of high encoding complexity, its computational amount is significantly increased compared to the previous video standards.With the development of science and technology, GPU’s powerful multi-thread parallel general purpose computing (GPGPU) ability attracts many researchers. We design key parallel algorithms in H.264and implement them on GPU in this paper.We make CPU and GPU work together, taking advantages of their own processor. CPU is responsible for handling file access and the tasks which require more logical units, and we use GPU’s multi-thread parallel mechanism to deal with modules which contain a large number of parallel tasks in H.264, such as intra prediction, inter prediction, transform and quantization, so we can take full advantage of CPU and GPU resources.We propose several parallel algorithms and implementation methods on CUDA with GPU for inter prediction and transform modules in H.264. Inter prediction module contains pixel interpolation, motion compensation and the hexagon motion search. According to each module’s data correlation, we design different parallel algorithms, and also do some optimization to make sure that our parallel algorithm can fit the characteristics of the GPU hardware better. In transform module, we design a paralleled algorithm on the base of the traditional butterfly algorithm, then analysis the experimental data and the parallel algorithm’s reasonableness.We make a comparison between our parallel algorithms on GPU with the traditional CPU serial algorithms through experiment. In pixel interpolation and transform module, the algorithm’s speed-up rate is more than20, and for hexagon parallel algorithm, the speedup rate reaches to more than100. Then we make a probing analysis of the experimental data.Finally, we talk about that our parallel algorithm can be further improved and that there is a bright future in this area.