High-Throughput Sharp Interpolation Filter Hardware Architecture for the AV1 Video Codec

Abstract

Motion Estimation (ME) is one of the most important steps of modern video encoders, due to its task of reducing temporal redundancies, but it is also highly computing intensive. The fractional part of ME is particularly complex, since it involves interpolating fractional pixels before the search. The fractional ME of AV1 encoders is even more challenging, since it supports up to 90 different interpolation filters grouped in 4 families. In this work, a dedicated interpolation hardware is proposed to mitigate this issue. The designed architecture interpolates the sharp interpolation filter family of AV1. A complexity analysis evaluates the throughput required in this process, showing that the designed architecture can process a 3840×2160 video sequences in real-time considering the motion compensation step, requiring 63.14 mW of power to operate.