A new method of video data optimization for motion estimation
Jiangyun Liu, Chenglin Zhao · 2011
Efficiency of motion estimation is affected by serious cache miss and cache split load. Firstly, the data to be used in the motion estimation process have poor spatial locality due to the layout format of video data. Secondly, the motion estimation process accesses memory irregularly and sometimes it accesses the data that spread across two cache lines. Irregular memory access pattern brings difficulties for a processor to pre-fetch data into the cache. The cache split load, on the other hand, forces the processor to load more data into the crowded cache. In this paper, a new layout format of video data, named overlapped tiled format, is proposed. The new format addressed the aforementioned two issues. By organizing the video data in the proposed way, the issue of cache split load is completely removed from the motion estimation process and the data cache miss is also greatly reduced. The full search motion estimation is speed up by 28% when organizing the video data in an overlapped tiled way.