The Speed Optimization Technology Of HEVC Based On Parallel Computing

With the rapid development of 4G and the internet technology,the requirement of people for video becomes higher,HD(High Definition)and UHD(Ultra High Definition)format videos pose great challenge on the video storage and transmission,state-of-the-art video coding standard H.264/AVC can no longer provide satisfactory compression performance.Against this background,the HEVC was proved to be the new coding standard by JCT-VT in January 2013.With the introduction of a series of new technology,HEVC achieves 50% bitrate reduction compared with H.264 under the same visual quality,but the computational complexity increases greatly,which brought difficulty to its application,so it becomes an urgent problem for the parallel acceleration of HEVC.GPU has strong operation ability with a large number of parallel processing unit,what's more,NVIDIA's CUDA programming model makes it easier for the programming of GPU.In this paper,we use CUDA parallel programing to accelerate HEVC,with the in-depth excavation of HEVC module's parallel speedup potential.In this paper,we use CUDA to accelerate motion estimation and interpolation process,which are the most time-consuming parts of HEVC.In our CUDA parallel algorithm of motion estimation,we use the MV information of the previous CTU in the time domain to predict the search start point and search range of the current CTU's PU,which reduces the computational complexity with the guarantee of coding efficiency at the same time.In our CUDA implementation of the motion estimation and interpolation,we make full use of CUDA optimization techniques,such as vector operation,avoiding branch division and advanced determination of thread blocks' size,which increase the program performance greatly.To adapt to Wavefront Parallel Processing of x265,we design a new CPU+GPU parallel computing framework with dynamic task scheduling.The MV information of the previous CTU in the time domain is used to predict the complexity of the coding region,which further schedules the computational tasks and realize the load balance better.Experiments show that the speedup brought by the framework is obvious.