Implementação em hardware de módulo convolucional com aritmética de ponto-fixo.

Abstract

Convolution is a technique applied on mathematical signals in order to abstract new levels of semantic comprehension about their characteristics. Such operation finds use in artificial intelligence, the biomedical industry, electronic circuits design, and media processing. In an increasingly automated global context, with a growing need for faster, more precise and more efficient computational systems, the convenience of hardware-dedicated algorithm implementations, as well as of different numerical representations in order to achieve set requirements is not to be ignored. During the development of this project, a unidimensional convolutional module was implemented in hardware using fixed-point arithmetic, aiming to provide a platform that would be able to accelerate such computations in algorithms that make its recurrent use, such as digital signal processing schemes and artificial neural network structures. Implemented models were simulated and synthesized, after which they were analysed under both quantitative and qualitative lenses.