Fast Algorithm and VLSI Architecture Design of Rough Mode Decision for AVS3

AVS3 video coding standard is the third generation standard developed by China AVS working group. Its first phase profile achieves approximately a 30% bit-rate savings than AVS2 for 4K ultra-high-resolution videos, but also brings a lot of encoding time complexity. In this paper, a sub-block splicing method for intra prediction is proposed. We simplify the calculation process of Rate Distortion Optimization (RDO) cost for different Coding Units (CUs) to increase the parallelism of the Rough Mode Decision (RMD) process and reduce the encoding time. Together a hardware architecture corresponding to the algorithm is proposed. Experiments show that the proposed algorithm can save 20% encoding time with an average of 0.63% BD-rate increase. And the proposed hardware architecture can support the real-time encoding of 4K videos at 30fps with an area around 0.54mm².

I. Introduction

With the development of the video industry, video data and high definition video contents have increased dramatically. To further improving coding efficiency, the third generation of Video coding Standard (AVS3) is developed [1]. The standard adopts a series of coding tools showing large coding gain with relatively limited complexity increment. Specifically, in the process of intra prediction, there are three splitting methods of coding unit (CU) based on quad-tree (QT), binary-tree (BT), and extended quad-tree (EQT). The sizes of CUs range from to . For each partition block, the number of intra prediction modes is set as 35. In reference software, the optimal partition method and the optimal prediction mode of each partition block are determined during rate distortion process (RDO) in an iteration way, which brings high computational complexity to intra prediction.