Studies On Decoding Algorithms Of LDPC Codes Based On GPU

Nowadays,with the rapid development of the computer technology and digital communication technology,how to ensure the reliability of communication technology becomes more and more important.In response to this problem,error correcting code technology arises at the historic moment.In 1948,Shannon proposed a mathematical theory of communication,which founded the field of error correcting code.In this field,the LDPC codes was proposed by Gallager in 1962,which has played an important role in digital communication system because of its excellent performance.Therefore,how to improve the throughput of LDPC decoders has become the focus of the study.The nodes of LDPC codes have the characteristics of non-interfering with the same nodes in the sum-product algorithm.Based on this characteristic,we can design the LDPC decoders with FPGA and GPU.Compared with FPGA,the GPU decoders provide more flexibility and expansibility.Especially at the appearance of Compute Unified Device Architecture--CUDA,GPU programming has become more simply and conveniently.In this paper,parallel decoding schemes based on GPU are designed according to the decoding algorithm of LDPC codes,and parallel decoders based on GPU are designed for multi codes.The main work includes the following aspects:(a)For the decoding algorithm of LDPC codes,this paper presents an edges parallel scheme to accelerate decoding speed of LDPC codes on GPU.In this scheme,each thread processes the information to be transmitted in each side of the Tanner graph,and the decoding time is compared with the decoding time of nodes parallel scheme.(b)To optimize the scheme proposed in(a),several schemes have been put forward: The update process of check nodes and variable nodes are merged into one kernel function through the on-chip cache;The decoding processes are implemented in two steams;Interactive algorithm is used to reduce the computational complexity of codewords with large row or column weight.(c)Parallel speedup of individual codeword is weak for QC-LDPC codes in the TDMP decoding algorithm,in this paper,we use multi codewords parallel decoding scheme,andoptimize the external information storage mode to achieve efficient use of global memory space and high-speed access to global memory information;(d)In order to realize the parallel decoding of multi codewords of pseudo-random LDPC code,a multi-code and multi-thread decoding scheme is presented for pseudo random LDPC codes.In this scheme,each thread handles a codeword and information storage is optimized.The information for adjacent threads can be continuously stored and accessed.Through realizing this program,GPU can process thousands or even tens of thousands of LDPC codes.The simulation experiments for all optimization schemes are carried out in this paper: With using edges parallel decoding scheme,the speed of LDPC decoders based on the GPU is 73.40~130.47 times faster than CPU;through merging kernel functions scheme,GPU obtains a speedup of 106.91~209.69 times;With using two streams asynchronous parallel,GPU ultimately achieves a speedup of 135.42~257.42 times;As for the column weight codeword by interactive program,the decoding speed of codewords relativing to nonoptimized GPU program enhances 371 ms and 1119 ms,compared to CPU,the decoding speed accelerates 101.87~154.57 times;The designs for QC-LDPC decoding and pseudorandom LDPC code decoding scheme make the decoding speed of parallel decoder based on multi code significantly improved,compared with the CPU decoding time,the accelerations are 48.93 times and 134.58 times respectively.