Complexity Reduction Of H.264 Using Parallel Programming

Abstract

The H.264 encoder provides for adaptive directional intra-prediction, motion-compensated inter-prediction followed by transform, quantization, deblocking filtering and either variable length encoding or arithmetic coding. All these blocks of the H.264 encoder make it highly complex as compared to previous video coding standards. There is a need for measures to reduce the complexity. This thesis aims at reducing the complexity of the .H.264 encoder by reducing the encoding time of the H.264 standard while not sacrificing the video quality, compression efficiency and bitrate by parallel programming. There are several parallel programming models that can be used. The massively parallel Graphics Processing Units (GPUs) provided by NVIDIA Corp. are used in this thesis for parallel processing. The main focus of the thesis is to reduce the time it takes for the motion estimation during the inter prediction. Motion estimation is the most compute-intensive process of H.264 and involves basic mathematical operations like subtraction and addition between the pixels of the reference frame and the frame under prediction. The subtraction between the pixels, also known as the sum of absolute differences (SAD) is done in parallel. The frame is partitioned into smaller 8 x 8 blocks and for these blocks threads are invoked on the GPU and all the calculations are done in parallel. The largest macroblock has a size 16 x 16, if this is divided into 4 equal parts each of size 8 x 8, a set of threads called blocks are invoked that carry out the operation for each particular 8 x 8 block. Up to 50% reduction in total time is observed for various input sequences of different characteristics. For implementation of the encoder, JM 16.0 reference software is used in this thesis. The manual for this reference is available and provides reference of different encoder parameters, syntax, and additional information regarding the best practices and configurations of the software.