基于IBM CELL多核处理器的快速运动估计算法

Fast ME Algorithm Based on IBM's CELL Nine Cores Processor

为了充分利用CELLBE处理器SIMD技术的数据并行处理能力,本文介绍一种运动估计算法,和已有的钻石搜索算法(DS)相比该算法能够实现更精确的运动估计和更低的运动复杂度。大多数传统的快速运动估计算法只是通过减少运动估计时的搜索点的个数来减少运算复杂度。然而对于CELLBE来说,搜索点的减少并不等于运算复杂度的减少。本文提出的算法在第一步使用一个包含35点的矩形搜索框,和DS算法相比能够提高了4.32%的压缩效率。相比DS算法,我们能够减少29.5%的运算量。然后优化了inter和intra模式选择算法,使intra的运算量从11%降到了3%。

In order to explore the data level parallelism in CELL BE＇s SIMD technique, this paper introduces a new fast motion estimation algorithm, which can achieve much better prediction accuracy and lower computational complexity compared with Diamond Search （DS） algorithm. Most of the conventional fast motion estimation algorithms just focus on reducing the number of search points. The introduced method uses a large rectangular search pattern at the first search step. Through increasing search points, it gains 4.32% coding efficiency compared with DS. It needs 29.5% less computation than DS method on CELL BE. We also optimize the mode judgment of inter and intra, and the intra part can be reduced from 11% to 3%.