Research Of Parallel Algorithm For Kalman Filter On CPU-GPU Cooperative-heterogeneous Platforms

With the extensive use of state estimation theory in scientific computing and engineer-ing applications,many industries and fields have strong demands for the application and development of state estimation software.Kalman filter algorithm is an optimal estima-tion algorithm for system states,which is a typical application of state estimation theory.Kalman Filter algorithm is an effective data processing algorithm,which has been widely used in space monitoring,wireless communications,tracking systems,financial industries,and other fields.At present,the dimension of system state attribution in Kalman filter algo-rithm is still simple one-dimensional,two-dimensional,but with the complexity of system states and the dimension of state estimation attribution increasing,the scale of data increas-es exponentially,so that the traditional Kalman filter algorithm is difficult to meet needs of the application.A CPU-GPU heterogeneous computing system has become an ideal platform for large-scale parallel computing because of its economy and high efficiency.Based on the CPU-GPU heterogeneous computing platform,the parallel optimization of the computational process is studied for the large-scale and multi-dimensional Kalman filter algorithm,and the computational performance of the Kalman filter algorithm is improved.In this paper,first of all,we study several CPU-GPU collaborative heterogeneous par-allel programming models,such as GPGPU architecture,OpenMP,CUDA,CPU-GPU syn-ergetic model.Next,the Kalman filter parallel algorithm is designed and implemented on these three platforms.The main contributions of this paper include:(1)Using the OpenMP parallel programming model on the multi-core CPU platform,a parallel Kalman filter algorithm is implemented for partitioning of matrices and vectors.Compared with the serial algorithm,the experimental results show that the performance of the parallel algorithm is improved significantly.(2)On the CPU-GPU heterogeneous platform,an algorithm of CPU-GPU task parti-tioning is proposed.We utilize shared memory to improve the efficiency of data access,and use CUDA stream to overlap calculation and transmission.Not only has it taken advan-tage of the GPU powerful parallel computing power,but it has also used CPU computing resources effectively.(3)On the Sunway TaihuLight platform,according to new architecture of the sw pro-cessor(260 cores)and new type of master and slave nuclear isomerism coding mode,we present an improved kalman filter parallel algorithm.And we provide some strategies for vector optimization and de-correlation optimization for seven-level pipelines,as well as the double buffering mechanisms,which improve the parallel efficiency of Kalman filter algo-rithm.Finally,compared with the CPU serial algorithm on five estimation experiments with different data scale,the results show that the parallel Kalman filter has a more progressive performance improvement on all the CPU multi-core platform,the CPU-GPU heteroge-neous platform,and the Sunway TaihuLight platform.And with the size of the data in-creasing,the speedup also increases,which reflects that the parallel algorithm is scalable well.