Otimizações algorítmicas e desenvolvimento arquitetural para as DCTs do HEVC

Abstract

Video coding is an area in constant evolution, especially in function of its relevance for consumer electronics, since devices which are able to process digital videos are currently more and more common. The video coder efficiency improvement is function of the relation between bit-rate and video quality and, to improve the encoder efficiency, the video coding systems are using more and more complex coding tools. This causes a big impact in the coding process complexity. This work is focused in the coding tools defined in the new emerging standard called HEVC – High Eficiency Video Coding. More specifically, this work focuses in the variable size discrete cosine transforms defined in the HEVC. There are four 2-D DCT transforms sizes defined in the HEVC standard: 32x32, 16x16, 8x8 and 4x4. This is a novelty of this standard and its use increases the coder efficiency and complexity. This work present two main contributions: (a) the development of an algorithm to optimize the 2-D DCT algorithms (with the transformation of multiplications in shift-adds andthrough the common sub-expressions sharing) and the software implementation of this algorithm to generate the simplified 2-D DCT algorithms and (b) the hardware design of the optimized algorithms, focusing in low cost and high processing rates. These optimizations are especially necessary when battery powered devices are considered. The optimizations allowed, in the best case (32x32 DCT), an expressive reduction in the number of operations: from 22 thousand of multiplications and 25 thousand of additions or subtractions to 29 thousand of additions or subtractions. If the number of one-bit additions or subtractions is considered, the developed optimizations are able to generate gains higher than 59%. The architectures were designed using the simplified algorithms, focusing in Altera FPGAs. Five architectures were designed, one for each 2-D DCT size and one which is able to process all four sizes of the DCTs. In all cases, the transforms were designed to use the separabilitty property, then two instances of the 1-D DCT transforms were joined by a transposition matrix to form the 2-D DCT transforms. The 1-D transforms were designed in a fully combinational way. The final synthesis results showed that all designed 2- D DCT architectures are able to reach very high processing rates, which allow the processing of high definition videos in real time.