A fast inference engine for Boosted Decision Trees

Decision trees and derivatives such as Boosted Decision Trees, Adaboost, XG- boost, Random forest are widely used in the world, and are now part of the High Energy Physics toolbox. However, High Energy Physics has some spe- cial requirements, it needs the inference of such tree to be as fast as possible, to be used during reconstruction of events or even trigger. In this work, we design and implement a fast inference engine for decision trees in C++ and benchmark it against XGBoost C API, showing a >15x improvement event-by-event and a 3x improvement for batched versions. The code is freely available on github. This is realized in the context of the CERN summer student program for the ROOT-TMVA project.