Flexible Hierarchical Parallel Processing for AVS3 Video Coding

Hannong Zheng, Yuhuai Zhang, Jian Zhang, Hengyu Man, Xuan Deng, Siwei Ma

Abstract: Video coding standards adopt a series of time-consuming techniques to improve coding performance. Acceleration is vital to the application of real-time video coding. However, conventional frame-level parallelisms usually fail to accelerate effectively, since the coding order is barely adjusted. This motivates more fine-grained parallelisms adopted in industry codecs, while the coding performance loss is inevitably introduced due to the coding restrictions. To benefit parallelism in Random Access (RA) configuration, we propose a novel parallel strategy named Flexible Hierarchical Parallel Processing (FHPP). Based on level order traversal, FHPP achieves cross-GOP and cross-intra-period parallelism in a flexible and hierarchical way. Furthermore, we analyze the upper bound of the FHPP speedup ratio theoretically. Extensive experimental results on the AVS3 standard show that FHPP can achieve an average acceleration of 10 times for 4K sequences under 16 threads, and the maximum acceleration can exceed 12 times, which is significantly better than the traditional frame-level parallelisms.