實現於RISC架構之H.264視訊編碼複雜度控制

 

隨著行動多媒體的蓬勃發展,相關研究及應用引起越來越多的關注與討論。行動裝置輕薄短小、攜帶方便,同時卻被賦予能夠支援多種多媒體功能的期望,如:影像擷取與播放、即時視訊通訊、電影與電視之播放……等等。為了使行動裝置在有限電源供應下同時能夠支援如此高運算量、高頻寬之要求,電力知曉 (Power Aware) 的設計觀念被應用在電力最佳化。

電力知曉視訊編碼系統不僅是低功耗的設計,亦能針對無線視訊編碼系統,依據電池可供電剩餘時間調整編碼之複雜度動態調整功耗,以達到延長電池可供電時間之目的。

本研究提出一參數化的視訊編碼架構。除了可以完全調整功耗,同時引入以人類主觀視覺之視訊品質評估方式,建立一個功率-位元率-失真 (Power-Rate-Distortion) 之視訊編碼系統模型。該系統可以調整其複雜度控制參數以符合行動裝置所能提供之能量,同時將視訊品質提升至最高。

本研究在Intel Xscale處理器開發平台上實際探討,並針對平台撰寫SAD(sum of absolute difference)組合語言,以減少編碼時間。

實驗結果顯示,此視訊編碼模型可減少系統高達75%之功耗。

 

 

Complexity Control for H.264 Video Coding on RISC processors

ABSTRACT

Mobile multimedia is a rising trend. Related research and application topics attract more and more attention. The expectation for a mobile device is light, thin, small, but powerful enough to support the increasing number of multimedia functions, such as photo capture and display, real-time video communications, and movies and TV display. To support these operations with high computing and high bandwidth requirements on a mobile device with limited energy, the power aware design concept is applied for achieving power optimization.

A power aware video coding system is not only a conventional low power design, but also a design that can adaptively adjust its encoding complexity based on the remaining battery supply to prolong the battery life.

We develop a parametric video encoding architecture, which is fully scalable in power consumption, and import a quality metric for subjective video quality measure. we establish a power-rate-distortion model of the video encoding system. The system is able to adjust its complexity control parameters based on the available energy supply of the mobile device while maximizing the picture quality.

A development platform based on the Intel Xscale microprocessor has been developed. SAD (sum of absolute difference) function is accelerated by assembly code to reduce encoding time. Experiments show that the system is able to save up to 75% power consumption in video coding based on this model.