Hierarchical Automatic Differentiation by Vertex Elimination and Source Transformation

摘要

We present a hierarchical scheme to extend the applicability of automatic differentiation (AD) by vertex elimination from the basic block level to code with branches and subroutine calls. We introduce the ELIAD tool that implements our scheme. Results from computational fluid dynamics (CFD) flux linearisations show runtime speedup by a typical factor of two over both finite-differencing and traditional forward and reverse modes of AD.