Multithreaded Parallel Programming Of H.264 On TILE64

In order to compress video information more effectively, MPEG and VCEG developed H.264 standard. In order to get better coding efficiency, H.264 integrates many high complexity encoding tools. It's hard for single processor to meet real-time needs. With the expanding of digital signal processor applications, the research of the video processing based on multi-core processors has become a hot spot.This paper studies the impact of video encoding with functional and data parallelism, and we use the superior TILE64 processor as the platform to research how to implement H.264 coding in parallel.Firstly, motion estimation is one of the most time-consuming modules in H.264, and if it is able to be separated out of traditional standard, then bound it to another cores, it is sure that the encoding speed will increase. During the research on the functional parallel processing, this paper analyzes the working principle of H.264 encoder and the motion estimation in-deep. In this paper, we present a way which can parallel process the motion estimation module and the other modules. The experimental results show that when compared with single processor, the frame rate after parallel processing is improved by 30% and the subjective performance of video decreases slightly. However, the approach of using the original frame instead of reconstruction frame may cause a decline in image quality, which is need to strengthen in our future research.Secondly, in the course of the studying of data parallelization, the paper analyzes the segmentation method of images with different size, and compared their advantages and finally we decided to use slice-based segmentation which can get the better speedup. However, the main reason for low speed in the traditional multithreading model is to create and destruct the threads frequently. This paper proposes the thread pool model based on slice level division to avoid the additional cost. Experiment shows that the speedup of image processing is enhanced by adopting slice level division and thread pool model and the performance on multi-core is also liner trend.Finally, the paper takes further analysis on the way of slice-based segmentation, point out its shortcoming such as the macroblock division, and proposed the idea by the dynamic macroblock segmentation for the following study.