Low bandwidth fractional motion estimation in H.264 design for mobile devices

The conventional fractional motion estimation(FME) approach needs a two-step interpolation for half-pel and quarter-pel refinements. Though two-sequential-step brings high encoding performance, it introduces a huge computational load and memory bandwidth requirement for searching seventeen fractional points for each of the forty-one motion vectors. These obstacles become more difficult when running the applications on mobile device with the limitation of power and hardware resource. In this paper, a low bandwidth FME design is proposed with two techniques. Based on high correlation between motion vector of a block and its up-layer as well as relationship of integer candidates, one-step FME algorithm is proposed to reduce computation complexity and memory bandwidth requirement. In addition, a memory saving scheme is proposed while carefully considering the trade-off between hardware resource and memory saving. Experimental results show that the proposed design just needs 66% of gate counts and saves 88% of memory bandwidth when compared with previous design.