A polyhedral model nested loop transformation dynamic solving acceleration method

Abstract

The application provides a polyhedral model nested loop transformation dynamic solving acceleration method, comprising the following steps: obtaining a nested loop in a compiling process, and creating a directed acyclic graph (DAG) according to the nested loop; performing topological sorting on the DAG, and converting the DAG into a linear DAG structure; starting searching a loop fusion / partition combination scheme from a first node of the linear DAG structure by using a DP algorithm, and obtaining a loop transformation corresponding to an optimal fusion / partition scheme of the linear DAG; and performing running and mapping attempts on a loop transformation back end, and completing the whole compilation. The method provided by the application obtains the optimal scheme through dynamic solving of the loop fusion / partition scheme, accelerates the solving process of the nested loop transformation, minimizes the number of back end attempts to find the optimized mapping result, and improves the execution efficiency of the nested loop.

Images