Automatic Generation of Parallel Code for Hessian Computations

Abstract

Given a program to compute some function, automatic differentiationcan be used to mechanically generate another program capable of evaluatingfirst- and higher-order derivatives of that function. A new strategyfor the computation of Hessians by automatic differentiation is proposedwhere the generated code is automatically parallelized using OpenMP.The approach is applied to compute second-order derivatives of anatmospheric reference model and performance results on a SunFireE6900 system are reported.