Gradual Intelligent Backtracking Vectorized Code Tuning Method

Abstract

The invention relates to a progressive intelligent backtracking vectorization code tuning method. The method mainly comprises the following steps: (1) static tuning, namely adding a compiler directing statement into a source program to generate a vectorization program; (2) dynamic tuning, namely carrying out dynamic instrumentation to obtain program section information and dynamic instrumentation information, and generating optimized vectorization program according to a feedback information file; and (3) configuration of vectorization basic options and optimized options, namely adding the optimized options one by one in a vectorization process, comparing the tuning result with the tuning result before the adding, if the current tuning result is better than the previous one, reserving the current one, otherwise, backtracking. By adopting the steps, the vectorization statement in the generating code can be optimized and the execution efficiency of the generating code is improved so as to relieve the pressure of a programmer in designing and writing parallel programs, and the computation capacity of the current high-performance supercomputer system can be exerted sufficiently.

Description

technical field [0001] The invention relates to a progressive intelligent backtracking vectorized code tuning method. Background technique [0002] In the construction of current high-performance computers, the integrated SIMD (Single Instruction Multiple Data, Single Instruction Multiple Data) short-vector functional parts in the CPU chip used can effectively improve the overall computing power of the computer system, which can significantly expand the instruction system through moderate expansion. It is of great significance to increase the width of data execution and improve the performance of multimedia processing. [0003] But correspondingly, in practice, manually writing or rewriting high-quality vectorized code is a great challenge for programmers. On the one hand, because the discovery of SIMD parallelism often requires a series of code transformations and optimizations, the manual vector recognition process has high requirements for programmers. For the implementa...

AI Technical Summary

Problems solved by technology

On the one hand, because the discovery of SIMD parallelism often requires a series of code transformations and optimizations, the manual vector recognition process has high requirements for programmers. For the implementation of various program transformations and optimization measures, programmers need to be familiar with the compilation technology. On the other hand, the rapid development of computer technology for more than half a century has accumulated a lot of valuable experience and wealth for R&D personnel. In the existing scalar computer application process, a large number of excellent software that have played an important role are urgently needed Efficiently transferred to high-performance computer systems, in order to make full use of the SIMD short-vector functional parts provided in a given CPU, programmers need to do a lot of manual transformation or re-write parallel programs, which is a It is time-consuming and labor-intensive, which causes great work pressure for programmers, and is not conducive to fully utilizing the computing power of high-performance computers

Images