Parallel Decoding Of Ldpc Codes Based On The Cuda Platform

Unavoidable interference and unexpected error exists when signal is transmitted through noisy channel,so it is significant for digital communication system to transmit signal with high speed and reliability.Channel encoding for error control is crucial to achieve the target.Low Density Parity Check(LDPC) codes are a kind of linear block code which is defined in terms of sparse parity check matrices and whose most important innovation was that an iterative decoding algorithm suitable for parallel implementation is introduced.It is possible especially for long codes to gain better decoding performance.It has been shown that LDPC codes can achieve remarkable performance that is very close to the Shannon limit and thus become strong competitors to many other well-known codes including Turbo codes.There are two types of LDPC codes,irregular codes and regular codes.In general decoding performance of irregular LDPC codes are better than regular ones,but lower algorithm complexity the latter ones have.When designing the encoding structure for LDPC codes, eliminating short loop is helpful for improving their decoding performance.As a kind of soft-decision decoding algorithm,Message Passing Algorithm can gain better implementation performance than hard-decision decoding algorithm,but have higher complexity.In a word,LDPC codes have good performance and their decoding algorithm can be implemented in parallel,so have shown large potential in many application fields such as personal digital devices,mobile wireless communication and so on.It is probable to be one of main channel encoding schemes in 4G system.On the other hand,with the fast development of semiconductor industry,processing units within computer system become more and more powerful.in comparison with multicore CPU, GPUs have much more cores and computing horse.Since GPU is designed that more transistors are devoted to data processing rather than flow control,the number of Floating-Point operations per second for GPU is much larger than for CPU.Thus,people proposed the conception of GPGPU(General Purpose GPU) to release GPUs' remarkable computing performance in lots of non-graphics fields.Modern GPU can be viewed as powerful parallel device based on PC platform,which enables programmers to develop parallel GPU code for general tasks to finish analyzing and processing massive data.Because hundreds of on-chip processor cores can simultaneously work and cooperate,GPU is especially suitable to solve problems which can be describe as parallel computing.However,early GPU programming was not very user-friendly which was with graphics shader languages based on Direct3D or OpenGL APIs and assumed some basic knowledge on graphics hardware architecture and computer graphics.NVIDIA's CUDA(Compute Unified Device Architecture) is a new hardware and software architecture for general-purpose computation on GPU.On CUDA programmers can write C-like codes to run multiple GPU threads simultaneously.Since decoding algorithm of LDPC codes can be implemented in parallel,in this study we based on CUDA investigate decoding implementation of LDPC codes.The performance between on traditional CPU and on CUDA is compared according to simulation results.It is shown that CUDA can bring drastic acceleration.