Research On Decoding Method Based On Alternating Direction Method Of Multipliers For LDPC Codes

Low-Density Parity-Check (LDPC) codes with approaching Shannon’s capacity performance are one of the important research directions in channel coding field, and have been widely applied in wireless communication and digital storage fields. With the constant development of information technology, communication system requires more quick and reliable information transmission and storage system requires higher storage density, therefor, it is very important to design encoding and decoding algorithms with excellent performance and low decoding complexity. The Linear Programming (LP) decoding of LDPC codes based on Alternating Direction Method of Multipliers (ADMM), not only has the Maximum Likelihood (ML) certification and easy analysis features of traditional LP decoding, but also greatly reduces the complexity of decoding algorithm based on optimization theory.It takes a long time to decode when LP decoding based on ADMM is applied to large-scale LDPC codes, so it is bad for the simulation analysis and actual engineering realization. Graphic Processing Unit (GPU) in terms of floating point arithmetic and memory bandwidth by leaps and bounds development has greatly promoted the research of parallel computing. With good parallel structure, LP decoding method based on ADMM can be well implemented on GPU. In order to substantially improve the decoding speed, this paper studies the LP decoding method based on ADMM with the GPU. The main works of this paper are summarized as follows:1. The theoretical basis of LDPC codes and its Tanner graphs model are introduced, the Belief Propagation (BP) and ML algorithms are outlined, the Fermi structure of GPU and programming technology of Compute Unified Device Architecture (CUDA) are presented, and the access features of texture memory and global memory of GPU are analyzed.2. The model of LP decoding based on ADMM are established, the Euclidean projection algorithm in decoding methods which has a large quantity of calculation are emphatically analyzed, and simulation and implementation based on ADMM decoding algorithm under the Additive White Gaussian Noise (AWGN) channel are given. Simulation result shows that the decoding algorithm based on ADMM performs well and its decoding speed can reach twice as BP algorithm.3. According to iterative message passing idea of ADMM decoding algorithm, a parallel decoding method based on GPU is designed, and the parallel implemental method of node message updates is analyzed in detail. Two optimization methods are designed in terms of accessing features of pinned memory and texture memory, and the implementation of parallel decoding method based on ADMM are given. Simulation result shows that the ADMM decoding method with the help of GPU implementation can gain 19 times speedup ratio compared with the method with pure CPU realization in the adjacent decoding performance condition.