All Zero Block Detection Algorithm Research And AVX Instruction Optimization For AVS Encoder

Towards the new century,video compression technology has a rapidly development in order to effectively solve the difficulty from the transmission and storage of massive video data. AVS as China’s independent design of video codec standard, its emergence break the foreign monopoly in the field of video codec. There are so many innovative technology applied in AVS that it has a good compression performance. However, the superiority of AVS encoding performance is at the cost of high computational complexity, thus it is a hot research topic currently that how to reduce the computational complexity of AVS without affecting the subjective and objective quality of videos.Based on the original open-source x AVS encoder, this thesis aims to further improve the code speed using algorithm and instruction optimization. The work we have done are listed as follows:Firstly, based on the characteristics of the transformation and quantization, also combining with the concept of all-zero block detection, a condition for detecting all-zero block on the characteristics of transformation and quantization is deduced. Meanwhile, according to the theory that the image residual coefficients obey Laplace distribution, we derive the pre-judgment conditions for all-zero block based on the statistical distribution of the Laplace. Combining the above two conditions, an hybird all-zero block detection algorithm is given. The experimental results show that the detectable rate of all-zero blocks of this algorithm is improved by 20% in average compared with that of the existing algorithms.Secondly, Based on above-mentioned all zero block detection algorithm, an inter mode selection early terminate based all zero block is given. This algorithm is applied to the mode selection of AVS interframe so as to terminate the mode selection early as well as skip the transformation and quantization, thus saving the encoding time. The experimental results show that the algorithm improves the encoding speed by an average of 27% while the PSNR decreases an average of 0. 05 d B and the transmission rate increases an average of 2.04%.Finally, the time complexity of each key modules is evaluated. According to the profiling results, the modules which are largely time-consuming and computationally intensive, such as the computing SAD, sub-pixel interpolation, intra prediction and transform quantization, are optimized respectively on the platform that supported AVX instructions set. Compared with the C code modules, the optimized modules can deal with multiple data with a single instruction simultaneously. Through the comparison of HD 1080 P video sequence tests with the original C code encoder, we have analyzed three indicators for the optimization performance: the peak signal-to-noise ratio, the code rate and the coding speed. The experimental results show that its average coding speed is increased by 3.11 times than that of the original x AVS encoder, while maintaining the peak signal-to-noise ratio and the code rate. Finally, combining the inter mode selection early termination algorithm based on all zero block with AVX instructions set, the optimized coding speed raise 3.37 times as against that of the original x AVS encoder.