Research And Implementation Of Optimization And Acceleration Of Airborne Ice Sounding Radar Field Algorithm Based On Heterogeneous Platform

During 2018-2019,I was fortunate to be a member of the 35 th China Antarctic Scientific Expedition and participated in the completion of on-site data processing of the "Snow Eagle 601" airborne observation platform.During on-site work,data processing of airborne ice sounding radar is the most important and critical part of the work.However,the current airborne ice sounding radar processing method adopted in China is based on the central processing unit(CPU)framework,and the ice sounding radar data collected based on airborne observations has a wide coverage and a complicated processing process.The more time-consuming data processing will greatly limit the efficiency of on-site data analysis,which in turn will affect the development of subsequent airborne observation tasks.Through in-depth investigation,it is found that in the ice sounding radar processing algorithm,the vast majority of steps can be split into concurrent execution,so we can consider using an efficient parallel processing platform to improve processing efficiency.Of the many parallel data processing platforms,Graphics Processing Unit(GPU)was initially used in the field of images and graphics,but because of its excellent parallel computing performance,it has been widely used by researchers in the field of general computing.It is particularly worth mentioning that in the process of the processor moving from the multi-core era to the heterogeneous era,NVIDIA Corporation pioneered the CUDA(Compute Unified Device Architecture,CUDA),which enabled programmers It can make more convenient and full use of the GPU's computing performance and improve the efficiency of parallel computing.In view of the above reasons,this article will start with the analysis of the characteristics of the airborne ice sounding radar algorithm.By optimizing the algorithm structure,improving the parallelism of the algorithm,and improving the efficiency of I / O data transmission,a series of methods are designed and implemented based on the CUDA programming framework Optimize the acceleration program.The main content of this article can be summarized as the following four aspects:1.Taking the CUDA and airborne ice sounding radar algorithms as the main research lines,introduced the CUDA thread model and micro-architecture model,and conducted an in-depth study of the CUDA architecture.Subsequently,a theoretical analysis of the parallelism of the airborne ice sounding radar algorithm is carried out based on the characteristics of the CUDA architecture.Through analysis of the experimental results of the existing optimization platform,performance bottlenecks that affect the efficiency of the optimization program are found,which lays a good foundation for subsequent optimization methods.Theoretical basis.2.In order to solve the problem of insufficient concurrency of existing processing algorithms,we analyzed the processing dimensions of the airborne ice sounding radar algorithm,combined with the GPU's own architectural features,and used overlapping operations and loop expansions for the range and azimuth processing steps of the airborne ice sounding radar.Instruction-level optimization method.3.In order to solve the problem of transmission efficiency of the GPU processing process caused by the large amount of airborne ice sounding radar data,the stream is used to hide the data transmission delay between the CPU and the GPU;then by analyzing the spatial locality of the airborne ice sounding radar processing algorithm,combined with the GPU cache hit mechanism,Cyclic adjustments for the purpose of improving the GPU cache hit rate can effectively improve the spatial locality of the azimuth processing steps.4.According to the needs of on-site analysis of airborne ice sounding radar,a method for quickly extracting the interface between ice surface and bedrock in Antarctic airborne ice sounding radar data is proposed.Combining the above optimization methods,a set of implementation schemes based on optimization and analysis methods is proposed.Finally,the relevant optimization results and optimal resource allocation schemes are given,and a set of application programs serving onsite data processing and analysis is realized,which is of great significance to the development of China Antarctic aero-geophysics survey.