Parallel computing method for directly solving structured triangular sparse linear equation set

Abstract

The invention provides a parallel computing method for directly solving a structured triangular sparse linear equation set, and belongs to the field of heterogeneous multi-core platform communication optimization and high-performance numerical computing. The method comprises the steps of inputting a solving problem scale and a grid calculation template of a structured linear equation set, receiving a solving matrix and a right-end vector, adaptively selecting a solving mapping scheme, and starting multi-core parallel processing; dividing the solution matrix and the right-end vector into a plurality of batches according to the selected solution mapping scheme, and mapping the solution matrix and the right-end vector to a slave core array; based on the computing task of each slave core, opening up a space in a local storage space of the slave core to store dependency data; and based on the calculation in batches, mapping the solution matrix and the right-end vector to the slave core array in batches for calculation, and obtaining dependency data by the slave core array in each batch according to a pipeline mode to complete calculation and communication until the structured linear equation set problem is completely and correctly solved. By adopting the method, the structured linear equation set can be efficiently solved.

Description

technical field [0001] The invention relates to the fields of heterogeneous multi-core platform communication optimization and high-performance numerical calculation, in particular to a parallel calculation method for directly solving structured triangular sparse linear equations. Background technique [0002] In recent years, structured grid problems have been widely used in many scientific fields, including global forecasting systems, ocean simulation, atmospheric simulation, radiation propagation, electromagnetics, and the important High Performance Geometric Multigrid (HPGMG) benchmark. Such problems are usually formulated as the discretization of a system of partial differential equations on a structured grid representing a surface or space. The most central step in solving these practically applied systems of equations involves solving the resulting sparse matrix linear equations. [0003] There are mainly direct method and iterative method to solve sparse matrix line...

AI Technical Summary

Problems solved by technology

[0005] At present, it is difficult to find a method suitable for solving structured triangular sparse matrix linear equations with better performance, and the existing high-performance sparse solution schemes cannot have good performance on structured grid problems. Therefore, it is urgent There is a need for a parallel computing method that relies on a multi-core heterogeneous platform to directly solve triangular sparse linear equations

Images