Gpu Based Acceleration And Implementation Of Compressive Sensing Reconstruction

The compressive sensing(CS)theory breaks the limitation of the sampling rate based on the traditional Nyquist sampling theorem,and it can realize the compressive sampling of the signal with the low sampling rate under the condition that the signal is sparse.After it was proposed,CS had gained wide attention on many application fields such as wireless communication,array signal processing,pattern recognition and biological sensing,etc.Compressive sensing reconstruction method usually has high computational complexity.Especially when the signal scale is large,the time of compressive sensing reconstruction will increase dramatically,which is difficult to meet the need of the real-time requirement of large scale signal compressive sensing reconstruction.The accelerated implementation of compressive sensing reconstruction algorithm has become an important issue which this field focuses on.With the implementation of graphics processing unit(GPU)based on CUDA architecture,the GPU initially used for image acceleration,graphics conversion,rendering and other work is now widely used in high performance computing.Considering the high efficient parallel execution ability of the GPU and the good acceleration potential on the large scale parallel data processing,in this paper,the acceleration method and implementation technology of the compressive sensing reconstruction algorithm based on GPU are discussed.On the basis of the analysis and comparison of three kinds of compressive sensing reconstruction algorithms(i.e.,OMP algorithm,BP algorithm and A*OMP algorithm),and the feasibility analysis of the parallelism of the algorithm module,the design and implementation of parallel acceleration for two kinds of reconstruction algorithms(i.e.,OMP algorithm and A*OMP algorithm)on GPU platform are presented.The main works and innovation of this paper are as follows.(1)The acceleration method and implementation technique of OMP algorithm based on GPU was discussed.In order to reduce the transmission latency between the central processing unit and GPU,parallel processing method for transferring the iterative process of OMP to the GPU was proposed.In GPU processing module,the CUDA program was improved based on the read-write character of global memory,so that the read-write of memory meets the condition of parallel access,and thus the access delay is reduced.By using the rich resource of the Streaming Multiprocessor(SM)and assigning its shared memory,the memory access peed is increased significantly.(2)The design of parallel A*OMP algorithm and the acceleration method based on GPU was exploited.By changing the time-consuming matrix inverse operation parallel matrix / vector operations and considering the inner correlation of algorithm,an iterative method is proposed to reduce computation complexity.By mapping all parallel matrix / vector computation into the GPU parallel execution,and making use of the parallel programming of Jacket,the speed-up performance of system is promoted considerably.