Projeto de Hardware de Alto Desempenho para Ferramentas da Predição Inter Quadros do Codificador AV1

Abstract

Currently, digital videos are increasingly present all over the world, whether they are intended for leisure or professional activities. Consequently, a significant growth in the number of mobile devices capable of capturing, manipulating and transmitting these digital videos can be observed. These mobile devices, in turn, have serious restrictions in terms of energy consumption and computational power. In this context, video encoding is responsible for an important part of these restrictions. New encoders are developed with the aim of compressing these videos more and more efficiently and for ever higher resolutions, thus making the support to digital videos a very costly operation from the point of view of computational power and energy consumption. In this scenario, the AOMedia Video 1 (AV1) encoder has gained prominence. AV1 is an encoder developed by a consortium of large companies in the area, such as Google, Apple, Intel, AMD, among others, with the objective of having high compression efficiency in a royalty-free solution. To this end, AV1 adopted new coding tools with a high computational cost, which increases the challenges for mobile devices to support this encoder. Among the stages of AV1 that brought greater innovations is the inter-frame prediction. Due to these innovations, the interframe prediction is responsible for the highest computational cost in the AV1 encoder, with about 58% of its the execution time, especially when we take into account that the interframe is one step among the eight within the AV1 encoder. This work presents a set of dedicated hardware solutions for AV1 inter-frame prediction. Efficient architectures were developed for Fractional Motion Estimation, Conventional Motion Compensation and Local and Global Warped Motion Compensations. All proposed architectures sought an efficient hardware design in terms of area and power. The architectures were synthesized to ASIC using the 40 nm TSMC library and they are able to process ultra high definition videos such as UHD 4K and UHD 8K at frame rates ranging between 30 and 60 frames per second. The architectures designed in this work are the first in the literature presenting solutions for the inter-frame prediction tools of the AV1 encoder.