| Recently,video applications have flourished and become an indispensable part of people’s lives.With the rise of short video applications(“Tik Tok”,“Kuaishou” and others),video is applied to record dribs and drabs of life.However,video has the defects of large storage capacity and slow transmission speed,which is especially obvious for high-definition video.Thus video compression is a key technology to reduce the storage capacity and the requirements of transmission speed,which is employed to remove redundant information of the data,retain the key information of the image,and reduce the amount of video data while ensuring visual effects.In this key technology,motion estimation is the core and the most time-consuming algorithm in video compression.Therefore,optimizing the performance of the motion estimation algorithm can effectively improve the execution efficiency of video compression,which is of great significance for the improvement of the performance of related video applications.Plenty of research has been conducted on the performance optimization of motion estimation by researchers at home and abroad and some achievements have been achieved based on this.But the following problems still need to be further solved and improved:(1)lacking of data reuse research on quick search algorithms.In the optimization of the motion estimation algorithm,strategy research of data reuse in the full research is mentioned,while the reusability of the search data in the operation of motion estimation algorithm of quick search has less consideration.Although with the faster search speed,the fast search algorithm increases the number of memory visit due to the irregularity of its memory access than the full search algorithm,thus taking more time.(2)lacking of research on the scanning order of the fast search motion estimation algorithm.In the optimization of fast search motion estimation algorithm,the search steps of the algorithm are generally optimized,while the scan order of the algorithm is not fully studied.Different scanning order of the algorithm will affect the effect of data reuse to a certain extent,which has an impact on the performance of the algorithm.Considering the above problems,the in-depth research is conducted to optimize the performance of motion estimation algorithm.And the main research contents and innovations are as follows:(1)Targeting at the insufficient research on the data reuse method of fast search algorithm,the data reuse method among search areas of fast search motion estimation is put forward by combining the parallelization characteristics of GPU platform.In the algorithm search process of motion estimation,there are overlapping data in the search area of two adjacent blocks.Then,according to the data reuse range of the adjacent block search area of the motion estimation algorithm,the on-chip memory is used to store the reusable data,which can reduce the visit numbers of the off-chip memory and the running time.As a result,in order to improve the utilization of overlapping data,a data reuse method for fast search motion estimation is proposed in this paper.The method reuses overlapping data between search regions of two adjacent blocks and further divides the overlapping data into two parts: a definite data reuse area and a possible data reuse area.The experimental results show that the proposeddata reuse method can effectively reduce the running time of the fast search motion estimation algorithm and the power dissipation of the on-chip memory.(2)Aiming at the insufficient study of the scanning order of the fast search algorithm,a performance optimization method based on scan order for fast search motion estimation is put forward.In accordance with the three different scanning orders of the full search motion estimation algorithm,this paper proposes three scanning sequences for the fast search motion estimation,the circular scanning order,the arc scanning order and the Z-shaped scanning order are proposed.Then,taking three-step search and diamond search as examples,the implementation methods of different scanning sequences and their effects on data reuse efficiency are studied.Moreover,experiments on the diamond search show that the arc scanning order is effective o reducing the running time and has low rate of change in the peak signal-to-noise ratio in comparison with the Z-shaped and circular scanning order. |
