Multi-swarm particle swarm optimization based on CUDA for sparse reconstruction

Particle swarm optimization (PSO) has been successfully applied to the sparse reconstruction problem and achieved good results. With the dimension of the problem increases, parallelizing PSO is an effective method to reduce its running time. This paper proposes a parallel PSO framework to solve the sparse reconstruction problem based on Compute Unified Device Architecture (CUDA) platform on Graphics Processing Unit (GPU). In order to further utilize potential computing resources in the GPU and improve the performance of the algorithm, each particle is launched by CUDA threads and the swarm is divided into multiple sub-swarms in CUDA streams. A local search strategy based on gradient and a particle coding strategy is combined into PSO for the purposes of achieving better reconstruction accuracy and accelerating convergence. In addition, in order to further optimize the parallel execution process of CUDA, the reduction algorithm and dynamic parallelism are incorporated into the proposed method. In the performance experiments, the proposed algorithm achieves a maximum speedup ratio of 25 times compared to the serial version in the signal reconstruction tasks.