Performance improvement of CMS simulation via loop transformation

Abstract

High performance processor can tackle bottleneck issues by increasing vector lengths and leveling effectiveness of memory hierarchies to address these issue. Manual optimization of code is a difficult task when having multiple architecturedependent transformation. Our goal is to develop a tool that performs source code transformation based on loop optimization techniques, since a loop plays an important role in improving of performance in scientific simulation software. We implement an source-to-source transformation tool based libTooling, a Clang’s library, based on polyhedral model to simplify a loop transformation of CMSSW building pipeline. The tool also can be used for automatically transformation. The results show that any simple loop transformations can trigger other optimizations in compilers.