Given that the concurrent L1-minimization(L1-min)problem is often required in some real applications,we investigate how to solve it in parallel on GPUs in this paper.First,we propose a novel self-adaptive warp impleme...Given that the concurrent L1-minimization(L1-min)problem is often required in some real applications,we investigate how to solve it in parallel on GPUs in this paper.First,we propose a novel self-adaptive warp implementation of the matrix-vector multiplication(Ax)and a novel self-adaptive thread implementation of the matrix-vector multiplication(ATx),respectively,on the GPU.The vector-operation and inner-product decision trees are adopted to choose the optimal vector-operation and inner-product kernels for vectors of any size.Second,based on the above proposed kernels,the iterative shrinkage-thresholding algorithm is utilized to present two concurrent L1-min solvers from the perspective of the streams and the thread blocks on a GPU,and optimize their performance by using the new features of GPU such as the shuffle instruction and the read-only data cache.Finally,we design a concurrent L1-min solver on multiple GPUs.The experimental results have validated the high effectiveness and good performance of our proposed methods.展开更多
Two new versions of accelerated first-order methods for minimizing convex composite functions are proposed. In this paper, we first present an accelerated first-order method which chooses the step size 1/ Lk to be 1/ ...Two new versions of accelerated first-order methods for minimizing convex composite functions are proposed. In this paper, we first present an accelerated first-order method which chooses the step size 1/ Lk to be 1/ L0 at the beginning of each iteration and preserves the computational simplicity of the fast iterative shrinkage-thresholding algorithm. The first proposed algorithm is a non-monotone algorithm. To avoid this behavior, we present another accelerated monotone first-order method. The proposed two accelerated first-order methods are proved to have a better convergence rate for minimizing convex composite functions. Numerical results demonstrate the efficiency of the proposed two accelerated first-order methods.展开更多
The sparse recovery algorithms formulate synthetic aperture radar （SAR） imaging problem in terms of sparse representation （SR） of a small number of strong scatters＇ positions among a much large number of potentia...The sparse recovery algorithms formulate synthetic aperture radar （SAR） imaging problem in terms of sparse representation （SR） of a small number of strong scatters＇ positions among a much large number of potential scatters＇ positions, and provide an effective approach to improve the SAR image resolution. Based on the attributed scatter center model, several experiments were performed with different practical considerations to evaluate the performance of five representative SR techniques, namely, sparse Bayesian learning （SBL）, fast Bayesian matching pursuit （FBMP）, smoothed 10 norm method （SL0）, sparse reconstruction by separable approximation （SpaRSA）, fast iterative shrinkage-thresholding algorithm （FISTA）, and the parameter settings in five SR algorithms were discussed. In different situations, the performances of these algorithms were also discussed. Through the comparison of MSE and failure rate in each algorithm simulation, FBMP and SpaRSA are found suitable for dealing with problems in the SAR imaging based on attributed scattering center model. Although the SBL is time-consuming, it always get better performance when related to failure rate and high SNR.展开更多
基金The research has been supported by the Natural Science Foundation of China under great number 61872422the Natural Science Foundation of Zhejiang Province,China under great number LY19F020028.
文摘Given that the concurrent L1-minimization(L1-min)problem is often required in some real applications,we investigate how to solve it in parallel on GPUs in this paper.First,we propose a novel self-adaptive warp implementation of the matrix-vector multiplication(Ax)and a novel self-adaptive thread implementation of the matrix-vector multiplication(ATx),respectively,on the GPU.The vector-operation and inner-product decision trees are adopted to choose the optimal vector-operation and inner-product kernels for vectors of any size.Second,based on the above proposed kernels,the iterative shrinkage-thresholding algorithm is utilized to present two concurrent L1-min solvers from the perspective of the streams and the thread blocks on a GPU,and optimize their performance by using the new features of GPU such as the shuffle instruction and the read-only data cache.Finally,we design a concurrent L1-min solver on multiple GPUs.The experimental results have validated the high effectiveness and good performance of our proposed methods.
基金Sponsored by the National Natural Science Foundation of China(Grant No.11461021)the Natural Science Basic Research Plan in Shaanxi Province of China(Grant No.2017JM1014)
文摘Two new versions of accelerated first-order methods for minimizing convex composite functions are proposed. In this paper, we first present an accelerated first-order method which chooses the step size 1/ Lk to be 1/ L0 at the beginning of each iteration and preserves the computational simplicity of the fast iterative shrinkage-thresholding algorithm. The first proposed algorithm is a non-monotone algorithm. To avoid this behavior, we present another accelerated monotone first-order method. The proposed two accelerated first-order methods are proved to have a better convergence rate for minimizing convex composite functions. Numerical results demonstrate the efficiency of the proposed two accelerated first-order methods.
基金Project(61171133)supported by the National Natural Science Foundation of ChinaProject(11JJ1010)supported by the Natural Science Fund for Distinguished Young Scholars of Hunan Province,ChinaProject(61101182)supported by National Natural Science Foundation for Young Scientists of China
文摘The sparse recovery algorithms formulate synthetic aperture radar （SAR） imaging problem in terms of sparse representation （SR） of a small number of strong scatters＇ positions among a much large number of potential scatters＇ positions, and provide an effective approach to improve the SAR image resolution. Based on the attributed scatter center model, several experiments were performed with different practical considerations to evaluate the performance of five representative SR techniques, namely, sparse Bayesian learning （SBL）, fast Bayesian matching pursuit （FBMP）, smoothed 10 norm method （SL0）, sparse reconstruction by separable approximation （SpaRSA）, fast iterative shrinkage-thresholding algorithm （FISTA）, and the parameter settings in five SR algorithms were discussed. In different situations, the performances of these algorithms were also discussed. Through the comparison of MSE and failure rate in each algorithm simulation, FBMP and SpaRSA are found suitable for dealing with problems in the SAR imaging based on attributed scattering center model. Although the SBL is time-consuming, it always get better performance when related to failure rate and high SNR.
