A DSP-based solution to increase the energy efficiency of real-time video encoders

摘要

Implementation of energy-efficient real-time video coders in hardware becomes a new research frontier as the demand for High-Definition(HD) video encoders growing to be a part of popular applications in hand-held smartphones and tablets. In this paper, we tackle an important problem encountered in hardware video encoder design, which wastes hardware cycles (hence loss of battery power) and introduces delays by reducing data throughput. Video conferencing applications in smartphones such as FaceTime™ and Skype™ necessitate energy-efficient real-time video encoders. In these applications, the video encoder should produce independently decodable data units (e.g., H.264/AVC NAL-units, slices) with size smaller than the maximum transmission unit (MTU) size of the network, to prevent fragmentation of packets. In this paper, we propose a method that significantly reduces the requirement to flush hardware pipeline when maximum NAL-unit size is reached by accurate estimation of the bit-rates of macroblocks at very early stages in the encoding pipeline. With this new method, we ensure that our video encoders in communication terminals to produce H.241 compliant maximum NAL-units with minimum waste of hardware cycles. Experiments show that without our method encoder hardware needs more than 16,000 pipeline flushes for one minute of video, which can be reduced by order of magnitudes with the proposed method to hundred or even less flushes.