高吞吐率的高效视频编码熵编码并行硬件架构设计

A High-Throughput Parallel Very-Large-Scale Integration Architecture of Entropy Coding in High Efficiency Video Coding

针对超高清视频编码的实时性需求以及高效视频编码熵编码系统中存在的数据吞吐率瓶颈,提出了一种基于波前并行处理(WPP)技术的硬件架构,在提高编码并行度的同时保留了行与行的依赖性,提升了编码系统的数据吞吐率与压缩效率。该架构可以进行三线程编码树单元同步编码,提高了近3倍的编码并行性;采用同步更新单元保证线程之间概率模型的同步更新;单线程内的概率模型支持多个二元的符号的并行更新;初始化单元支持对多种模式下编码单元的概率初始化;支持在线模式和离线模式,可以更好地平衡熵编码与其他模块的吞吐差异。本设计实现的熵编码器单周期处理的平均数据量为5b,在SMIC 55nm工艺下,其综合频率达到300MHz,吞吐率为1 512Mb/s,与参考设计相比,数据吞吐率提升了89%。本设计提出的熵编码器性能满足超高清视频的实时编码需求。

Due to the real-time requirements of ultra-high-definition video encoding and the data throughput bottleneck in the entropy encoding system,a hardware architecture based on wavefront parallel processing(WPP)technology is proposed,which improves the parallelism of coding while retaining row-to-row dependency,and provides higher throughput and compression efficiency.This architecture allows three-threaded coding tree units to be encoded simultaneously,thus improves the parallelism by nearly three times.The synchronous update unit ensures the probability models between threads can be updated simultaneously.The probability models in a single thread support parallel updating of multiple binary symbols.Initialization unit can initialize the coding units in multiple modes.Online mode and offline mode are both supported,which can balance the throughput between entropy encoder and other modules.This entropy encoder delivers an average of 5 b per cycle.The design is synthesized in SMIC 55 nm and it can work at a speed of 300 MHz.And the throughput is 1512 Mb/s,which is 89%higher than that of the reference design.The result shows that the performance of the entropy encoder proposed in this design meets the real-time encoding requirements of ultra-high-definition video.