RDE-AGMO: seleção automática de pontos de controle eficientes para codificadores HEVC utilizando algoritmos genéticos

• Autor(a) principal: Machado, ítalo Dombrowski
• Data de Publicação: 2020
• Tipo de documento: Dissertação
• Idioma: por
• Título da fonte: Repositório Institucional da UFPel - Guaiaca
• Texto Completo: http://guaiaca.ufpel.edu.br/handle/prefix/6087
• topic: Computação; HEVC; AGMO; Problema de otimização; Controle de energia; Optmization problem; Energy control; CNPQ::CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO

Due to the high energy consumption of the video encoding process of HEVC encoders and the fact that they must be executed on billions of mobile devices that have limited energy reserves, it is of extreme interest to reduce the energy consumption of these applications. Aiming to achieve reduction, energy controllers that manage the energy consumption of HEVC encoders have been proposed. Thereunto, these controls have a set of configurations called Control Points (PCs) that are dynamically selected to reduce or increase their consumption when necessary. This set of PCs is extremely important for the performance of the controller and is assembled through the sensitivity analysis process. Despite this being a complex task of multiobjective optimization and crucial for the efficiency of the controller, the methods applied to carry out this technical selection analysis presents aspects that can be improved. Since the current sensitivity analysis methodologies require an expert to manually create and select PCs, it is impossible to evaluate an acceptable set of configurations and for the developer to accurately choose the consumption heterogeneity of their PCs. Therefore, this dissertation presents the RDE-AGMO Rate-Distortion-Energy Multiobjective Genetic Algorithm, a multiobjective genetic algorithm (MOGA) capable of selecting efficient configurations with specific energy consumptions for video encoders. The RDEAGMO is able to carry out the sensitivity analysis process in a faster, autonomous and comprehensive way than the currently used methodologies while providing benefits to the controller’s behavior. Furthermore, due to the generality of the algorithm, it can be applied to any coder and controller variable (energy, time, arithmetic complexity) without requiring major modifications. When observing the impact of the RDE-AGMO applicated in the development of a controller, it was possible to notice that there was an improvement in the energy distribution of its PC table, with the mean square error of the configurations decreasing from 41% to 16%. This better distribution in the consumption of PCs consequently increased the possible control range from a maximum of 30% of energy reduction to 70%. Additionally, the average BD-BR caused by the energy controller decreased from 4.07% to 3.45% in exchange for an increase in the consumption error obtained in relation to the desired of only 2.45%.