Parallel Optimization of Motion Estimation for Video Coding on Cell BE Processors

This paper proposes a parallel framework to speed up video coding on the Cell Broadband Engine (Cell BE) heterogeneous multicore platform. It integrates thread-level parallelism (TLP), data level parallelism (DLP) and an innovative data prefetching scheme to exploit the implicit parallelism of the most computationally intensive motion estimation procedure in video coding. The video frame is partitioned to several slices which are processed simultaneously on multicores to exploit TLP. The sum of absolute differences (SAD) calculation in the ME process is implemented by using SIMD instructions to exploit DLP. Slices are transferred using a data prefetching scheme to hide memory access delay, which enables the process of data access and the ME process to execute concurrently. Experimental results show that the proposed parallel implementation, compared with the serial implementation of video encoding, achieves significant performance improvement.