Data Layout Optimization For Motion Compensation Of Embedded Video Decoders

Improving the performance and reducing the energy consumption in embedded systems has become a hot spot issue in most domains. This is particularly important for multimedia applications since they are data-intensive. Therefore, the special Digital Signal Processor (DSP) for multimedia has been used in embedded systems more and more. In order to diminish the gap between computation speed and data access speed of embedded systems, ScratchPad Memory (SPM) has been proposed. Because the SPM is located on chip, it is called memory on-chip as well. The optimization of data layout for SPM will obtain the goal of enhancing the performance and decreasing the energy consumption of embedded systems.In the premise of guaranteeing the video quality and timeliness, with the thought of stream computing, this paper proposed a scheme of data layout optimization for motion compensation. A memory structure of ping-pang buffer is set in SPM, then the data for motion compensation is allocated in this buffer through pre-fetching. Using the thought of stream computing that processing data and accessing data simultaneously, the data in the buffer useless for successive motion compensation is replaced by the data useful for successive motion compensation at the same time of executing motion compensation for macroblocks. This can improve the reusability of data on-chip and reduce the number of main memory access. In addition, data index algorithm is used to access the data for motion compensation.Based on TMS320DM642 of TI Company, with the MPEG-4 video decoder as research object, various standard sequences of CIF format are tested after the optimization of data layout for motion compensation. The experiment result shows that the performance improvement of MPEG-4 video decoder is 6.7% on average and the energy consumption reduction of DM642 level 2 memory on-chip is 18.5% on average after optimization. Moreover, the miss rate of level 2 cache decreases 1.94% in fact after optimization, and 31.8% with respect to non-optimization. It can safely conclude that the optimization of data layout for motion compensation could enhance the performance and reduce the energy consumption.This thesis is supported by the project of National Natural Science Foundation of China (No.60873029) and the project of Innovation Research Foundation of Huazhong University of Science and Technology (No.2010MS014).