Research On GPU-based Radar Clutter Simulation Method

Large amounts of clutter data are required for analysis in radar research and design.However,the continuous use of measured clutter data will consume a lot of manpower and material resources,affecting the research process of modern radar.The clutter simulation method based on statistical models can simulate clutter data in a shorter time,making clutter simulation into an important research direction.Zero Memory Non-Linerity(ZMNL)and Spherically Invariant Random Processes(SIRP)are two main methods of clutter simulation based on statistical models.They are widely used in the simulation of clutter data.With the continuous development of modern radar research,the accuracy and breadth of radar detection continue to increase,which means that the data received by the radar at one time has increased significantly.The clutter simulation method based on the traditional CPU can no longer meet the real-time requirements of clutter data of modern radar.The high-performance implementation of the clutter simulation method is an urgent problem.According to the current research and analysis of CPU-GPU heterogeneous system in high-performance computing platform,this thesis uses CPU-GPU heterogeneous platform to study the parallel strategy of ZMNL and SIRP,and analyzes the optimization effect in the experiment.The bottlenecks of ZMNL method and SIRP method in real time are analyzed in this thesis.According to the computing characteristics of the two methods,it optimizes the ZMNL method from instruction set,shared memory,convolution and so on to get ZMNL-CU method.From the aspect of integral calculation and task scheduling,the coarse-grained optimization of SIRP method is carried out,and the SIRP-CU method is proposed.Finally,the optimization results of the two methods are analyzed by reference to the actual radar parameters.Aiming at the time-consuming problem of the convolution calculation in the ZMNL method,Conv C-a matrix variable dimension convolution algorithm based on cu BLAS is proposed in this thesis,which is campared with the direct convolution method,the gemm method,and the Fast Fourier Transform(FFT)convolution method on multiple high-performance platforms.The experimental results show that the Conv C method has more significant computing performance than other methods when the amount of data is large.Aiming at the problem of unbalanced load of CPU-GPU heterogeneous platformin SIRP method,a coarse-grained task flow parallel model based on Open MP + CUDA is proposed in this thesis,and deduces the calculation efficiency evaluation function of the model.Finally,the effect of efficiency improvement is calculated through experimental analysis.Analyzing the optimization effect obtained by the two methods in the experiments,the ZMNL-CU method achieved a 41 times speedup compared to the CPU,and nearly doubled the performance compared to the traditional GPU optimization method.The SIRP-CU method achieves a speedup of 108 times compared to CPU computing,and improves the computing efficiency by nearly 61%compared to traditional GPU optimization methods.