Static Performance Prediction of Compiler Optimizations

Content Provider	Semantic Scholar
Author	Dennis, Michael L.
Copyright Year	2015
Abstract	Tunning the interaction of various compiler optimizations to be optimal for a specific platform is a daunting task with complicated interacting considerations. The problem is exacerbated by the fact that what is optimal on one architecture likely performs terribly on another. Since manually tuning a compiler for each new architecture is infeasible, this project will use machine learning to create a model specific to each architecture that will accurately predict the runtime performance of code based on static features to better inform compiler optimizations. Once the predictive model is created, it can be quickly used to determine near optimal settings for compiler tunning parameters specific to each block of code. To give this research direction, we will begin our work with stencil code, moving on to dynamic programing after preliminary results. These classes of algorithms provides more structure to the problem by focusing on programs whose performance is minimally input dependent allowing our methods to achieve more accurate results.
File Format	PDF HTM / HTML
Alternate Webpage(s)	http://cs.uccs.edu/~jkalita/work/reu/REU2015/FinalPapers/09Dennis.pdf
Language	English
Access Restriction	Open
Content Type	Text
Resource Type	Article