Software Pipelines Design for Variable Block-Size Motion Estimation with Large Search Range

This paper presents some techniques for efficient motion estimation (ME) implementation on fixed-point digital signal processor (DSP) for high resolution video coding. First, challenges in large search-range ME are discussed. Then, based on the statistics analysis of the best matched block, correlations in variable block-size ME are removed on algorithm level while keeping almost the same performance as the original fast search strategy. Based on this modification and data reuse technique, this paper proposes an "index search" method for DSP implementation, which is a good solution to balance both high coding efficiency and high coding speed on the DSP platform. In "index search", ME is divided into three steps and highly parallel pipelines are designed for each step. The fully use of pipelines brings about 85% time reduction during ME. "Index search" can also be easily extended to achieve different search range and search patterns.