Abstract: Floating-point computing programs are widely used in aerospace,artificial intelligence,national defense and military,financial settlement and other fields.The computing accuracy and performance of floating-point programs are directly related to the safety and effectiveness of related applications.The maximum floating-point error is the key indicator to measure the accuracy of floating-point computing programs,and the cumulative effect of floating-point errors will also lead to unbearable disasters,so it is necessary to develop an accurate and efficient float-point maximum error detection tool to provide support for researchers to take timely optimization and intervention measures.The proposed algorithm transforms the problem of maximum error detection into the problem of searching for the maximum value of the objective function,gives full play to the computing power advantages of the master-slave architecture two-level parallel computing mode of the domestic Sunway platform,deeply excavates the performance and accuracy potential of the particle swarm heuristic search algorithm,and optimizes the particle swarm algorithm with the idea of grid search,independent cultivation,hierarchical convergence and dynamic adaptation.According to the different stages of the search process,the relevant search parameters are set,so that the improved algorithm achieves improvement in both search accuracy and search performance.This provides a new practical tool and thinking reference for accurately detecting the maximum error of floating-point numbers,and further enriches the tool library of domestic Sunway platform.

Key words: Floating-point, Error detection, Particle swarm optimization algorithm, Parallel computing, Sunway platform