An L-Attributed Grammar for Adjoint Code

Abstract

Gradients of high-dimensional functions can be computed efficiently and with machine accuracy by so-called adjoint codes. We present an L-attributed grammar for the single-pass generation of intraprocedural adjoint code for a simple imperative language (a subset of C). Our ideas can easily be applied to any programming language that is suitable for syntax-directed translation.Moreover the conceptual insights are useful in the context of multi-pass generation of adjoint code. Our focus is on correctness. The necessary domain-specific code optimizations are beyond the scope of this paper. We give references to corresponding work in this area