Key Techniques Research Of Video Compression Coding Based On Stream Processing

In the information time, digital video is an important way of acquiring information. It is widly applied in video conference, distance medical treatment, distance education, high definition television broadcast, digital cenima and so on. To ensure its application, video compression encoding is a key technology of digital video processing. It's a typical intensive computing application with large amount of complicated computation. Since the digital video goes to high definition, the performance demands of video compression would achieve tera operations per second. However, current video encoder on general purpose processors could not satisfy the requirements of high performance video encoding, while special purpose ASIC video encoder is also less than satisfactory with the inflexibility, long development time and high cost. Therefore, video encoder with parallel computing model performed on the programmable processors is becoming a new trend.As an emerging parallel processing model, stream processing model made outstanding achievements in application acceleration area of media/signal processing and scientific computing including video encoding. Stream processing model improves the performance of applications by exploiting the computing intensity, abundant parallelism, and multi-level data reference locality.However, the inherent characteristics of the video encoding and the increasing requirements of video resolution, real-time performance, compression rate cause serial problems to video encoding based on stream processing. The problems include the scalarbility of video encoding, the restrictions caused by denpencies during the encoding procedure, and the low performance of high definition video encoding. To solve the problems, this paper tries to present a solution of high performance parallel video coding based on stream processing. The researches are mainly about stream processing framework of video encoding, high performance streaming video encoding algorithm, eliminating restrictions of parallelism, and efficient execution method of encoder. The main contributions are summarized as follows:1. The paper proposes Video coding Scalable Stream Framework (VSSF), a scalable stream framework for video coding, aiming for solving the problems in the aspects of parallel data granularity scalability, parallel degree, picture resolution scalability, and coding modules scalability. VSSF, derived from the stream processing, decouples computation and data access in video coding. Computation is encapsulated in the kernel engine, while data access is performed by reorganizing data into data strip. VSSF supports the scalability of parallel data granularity by storing data into off-chip memory, supports the scalability of parallel degree by changing organization of data strip, supports the scalability of picture resolution by changing number of data strips which are processed by kernel engine, supports the scalability of coding module by adding new module into the VSSF.2. This paper develops a parallel streaming intra prediction algorithm, Multiple Resolution Multiple Window Inter Prediction (MRMW). In order to reduce the costs of irregular computing and irregular data accessing, MRMW divides Inter Prediction into three individual phases of Reduced Resolution Full Search, Multi-Windows Refine Search, and motion compensation. Also, MRMW reduces the amount of computation by searching motion vectors in multiple levels of picture resolutions. The refine search in multiple windows could find the accurate motion vectors. In addition, a double dependencies MV selection method is proposed for the windows selection in multi-windows refine search.3. This paper puts forward solutions to eliminate the restrictions of parallelism caused by data dependencies, contrl dependencies, and constrains in bitstream store. First, a parallel streaming CAVLC method on basis of DLP is proposed to deal with constrains in bitstream store. The method dissevers serial execution path caused by tightly coupled computation and data access. Second, a multi-stage SIMD parallel method deals with the dependencies between neighbour blocks. It increases parallel degree and throughput. Third, a multi-group parallel method is developed to solve the problem due to multi-layer double direction data dependencies.4. This paper proposes a hardware/software method to accelerate the video coding execution by improving the cooperation speed. Aiming at solving the problem of low performance caused by software method, the proposed cooperative method is designed that software modules work with hardware logic. The hardware logic is responsible for performance sensitive tasks, while the software modules deal with complicated and flexiable tasks. And this paper implements it on the MASA-I stream processing SOPC system. The results show that hardware/software method can achieve promotion of cooperation performance by two orders of magnitude and the hardware costs only increase 2%.The dissertation implements H.264 VSSF encoder on four different platforms including stream processor, GPU, multicore CPU and DSP. HD-VideoBench is used to evaluate the encoder. The results show that H.264 VSSF encoder achieves outsanding performance promotion on the four platforms. And the 1920x1080 encoding performance of the VSSF encoder on stream processor meets the realtime encoding requirements.