The Optimization Of HEVC Motion Estimation Parallel Algorithm Based On CUDA

In recent years,with the improvement of the quality of life,people's demand for high-definition video is also getting higher and higher,the existing video coding standard H.264/AVC(Advanced Video Coding)has been unable to adapt to the development of the network and people's needs.Therefore,VCGE(Video Coding Experts Group)and MPEG(Moving Picture Experts Group)set up JCT-VC(Joint Collaborative Team on Video Coding),and developed a new generation of video coding standard H.265/HEVC.Its goal is to improve the efficiency of video encoding,and its compression ratio double improves than the video coding standard H.264/AVC High Profile.However,the increase of the compression rate also brings a significant improvement in the coding complexity,which makes the real-time performance of the system greatly restricted.Therefore,it is necessary to study the fast coding algorithm of HEVC.In this thesis,the HEVC(High Efficiency Video Coding)algorithm is optimized by CUDA(Compute Unified Device Architecture)to reduce the coding time.Aiming at the shortcomings of the motion estimation,such as large amount of computation and long time consuming,this thesis proposes a parallel algorithm based on CUDA,which utilizes the advantages of GPU(Graphic Processing Unit)parallel computing to realize the fast coding of the motion estimation.This thesis has mainly optimized the algorithm from the following aspects:Firstly,the TZSearch algorithm,which is kind of the integral-pixel motion estimation algorithm,is computationally large,and the data between adjacent prediction units is independent.But this algorithm is not suit for large-scale parallel computing as above features.So in this thesis,a CUDA algorithm based on the prediction unit and the search position is proposed.The algorithm not only improve the coding efficiency,but also reduces the influence of the CUDA algorithm on the image quality and bit rate by analyzing the motion vector of the adjacent coded frames to calculating the search starting point of each LCU(Largest Coding Unit)and dynamically selects the search template.Secondly,the fractional-pixel interpolation algorithm and the search algorithm require huge computation and there are a lot of repetitive calculations.In this thesis,a fractional-pixel motion estimation parallel algorithm is proposed to reduce repetitive calculations and data copy by adjusting the coding structure and managing the reference frame.It can effectively reduce the encoding time.Finally,aiming at the low CPU utilization in CPU and GPU co-operation,this thesis proposes a CPU and GPU asynchronous computing model based on integer pixel motion estimation.Through the CPU multi-thread design for the integral-pixel motion search and asynchronous computing model for the other coding modules,this thesis achieves a parallel computing with CPU and GPU,which effectively improves the CPU utilization and reduces the coding time.The experimental results show that the optimization of the HEVC parallel algorithm based on CUDA has achieved remarkable results.The integral-pixel motion estimation module achieves a speedup of about 12 times in the case of low influence on image quality and bit rate and The fractional-pixel motion estimation achieved a speedup of about 31 times by CUDA parallel optimization.The integral-pixel motion estimation module based on asynchronous mode achieves an average acceleration of about 18 times,which is obviously improved compared with the synchronous mode.