Parallel Optimization For H.264 Encoding

For video coding algorithm based on macro-block partition, motion estimation occupies most of the encoding time because of its huge amount of computations. In order to speed up the computation of motion compensation, a lot of efforts have been made on two main directions:one is to develop new algorithms, and the other one is to optimize available methods. The paper deals mainly with the optimization of motion estimation in H.264.This paper focuses on the parallel of motion estimation over CPU and GPU frameworks. As for CPU based optimization, SSE instruction set supported by most Intel CPU is used to speed up the computation. With respect to GPU based optimization, this paper first provides a brief introduction to the development history and parallel structure of GPU. Then CUDA program structure and underlying principle are discussed. In order to speed up the computation of motion estimation, this paper has made a deep investigation to some common used motion estimation algorithms, and reached a general conclusion that the Full Search (FS) algorithm has the greatest potential to get performance improvements by parallel optimization. It also reveals the details about how to make full use of parallel properties of CPU and GPU respectively.Simulation on a PC equipped with an Intel E5400 CPU and NVidia GeForce GT240 GPU shows that SSE and CUDA can improve the computation of full search based motion estimation block by a ratio of 2 and 8 respectively. As for its total contribution to H.264 encoding, at least 2 times acceleration can be achieved by only optimizing the motion estimation component.