Research On Parallel Implementation Of Concordance Coefficient Based On CUDA

The study on correlation analysis was started from the foundation of statistics,and it became one of the important branches of statistics.Even now topics about correlation analysis are still the hot spots in many subjects,especially in the field of statistical signal processing.The correlation coefficient is a common tool for measuring the statistical relationship between two random variables or two signals.There are four commonly used correlation coefficients,which are Pearson's Product Moment Correlation Coefficient,Spearman's rho,Kendall's tau,and Gini correlation.However,in practical applications,we often need to measure the correlation degree between multiple channel signals.In such a multi-channel signal,it is natural to use the average of the correlation coefficients between the signals of two different channels in the total number of channels as a quantitative measure of the degree of correlation between multiple channel signals.Based on these four common correlation coefficients,relevant researchers put forward the average Pearson's Product Moment Correlation Coefficient,average Spearman's rho,average Kendall's tau and average Gini Correlation.Inspired by Kendall's concordance coefficient people will describe the index of the degree of correlation between multiple channels,which is called the concordance coefficient.With the coming of the era of big data,the scale of data to be processed by correlation analysis is increasing.At present,GPU parallel computing is widely applied in large-scale data computing.Influenced by these two aspects,this paper proposes a unified implementation framework for parallel computation of concordance coefficient.And do the following aspects of the work :Firstly,by comparing several popular parallel computing programming environments,CUDA is selected as the implementation tool for GPU parallel computing.And through programming examples of matrix multiplication,the importance of using shared memory instead of global memory access for performance optimization is illustrated.Secondly,by equivalently transforming the four kinds of correlation coefficient definition expressions into a uniform and similar structure,the operations between the numerator and the denominator are independent of each other and are similar in the process of programming.Then we propose a unified framework for the parallel computation of concordance coefficients.As long as a small number of parameters are adjusted,different concordance coefficients can be switched.Finally,this chapter passes an experiment to test the performance of the four concordance coefficients on the two CPU and CUDA platforms to test the performance advantage of the four concordance coefficients calculated under the framework of the parallel calculation of concordance proposed in this paper.Two conclusions are drawn in the experiment: First,when the number of channels is large,the CUDA-based parallel computing proposed in this paper has a clear performance advantage in calculating the concordance coefficients under the unified framework.Second,AKT and AGC is more suitable for parallel computing than APPMCC and ASR.