High-Throughput Multifilter VLSI Design for the AV1 Fractional Motion Estimation

A main challenge for emerging video encoders is the high complexity introduced by their new encoding tools. In the royalty-free AV1 (AOMedia Video 1) codec, a large part of this complexity is focused on the inter prediction stage. This is particularly given to fractional motion estimation (FME), where a large number of FIR (Finite Impulse Response) type filters is used in the process of interpolation that generates fractional position samples given the integer position samples as input. Therefore, strategies to mitigate this complexity, such as designing hardware accelerators, are needed. Another recurring concern is the power dissipated as many users consume video media using battery-constrained devices. Based on that, this work introduces a dedicated multifilter hardware architecture for the AV1 codec interpolation filters with a focus on the motion estimation stage. The proposal implements the Regular, Sharp and Smooth filter families, using the operand isolation technique to avoid unnecessary power consumption. The designed architecture is capable of achieving a processing throughput of 3187.5 Msamples/sec for ME (Motion Estimation) operation, and can interpolate 8k videos resolution at 60 frames per second considering the MC (Motion Compensation) scenario.