Re-evaluating MPEG motion compensation search criteria

Abstract

As general-purpose processors evolve from single-issue machines to more advanced multiple-issue designs using dynamic instruction scheduling, speed-critical algorithms should be re-examined to ensure that the optimal strategy is being used. One important example of this class of algorithms is MPEG encoding. In MPEG encoding, the mean square error search criterion achieves better compression quality than the mean absolute distance criterion. However, the mean absolute distance approach is recognized in the literature as more efficient in terms of speed. Our research indicates that the mean square error approach becomes much more attractive as instruction-level parallelism (ILP) increases. In a general-purpose processor that can exploit a high degree of instruction-level parallelism, the mean square error approach yields both better performance and better compression quality. Additional algorithms should likewise be re-evaluated to ensure efficient execution on high-ILP machines.