Section: New Results

Application to large industrial codes

Participants : Valérie Pascual, Laurent Hascoët, Nicole Goutal [EDF-LNHE] , Andrea Piacentini [CERFACS-GLOBC] , Charlotte Kotas [Oak Ridge National Lab. (Tennessee, USA)] .

We support industrial users with their first experiments of Algorithmic Differentiation of large in-house codes.

A previous collaboration with EDF and CERFACS has been continued by a new three-months contract, with the objective of improving the AD adjoint of the hydrodynamic code Mascaret. The tangent and adjoint differentiated codes have been built for the calculation of steady subcritical flow (“Sarap” kernel) with the latest Mascaret Version 8.1. The differentiation process has been simplified and it exploits the latest capacities of Tapenade on Fortran 95. In particular, the differentiated code manages Fortran 95 dynamic memory through our library ADMM. Connection with the C-written ADMM uses the Fortran 2003 standardized interoperability primitives. Validation was conducted on two test cases (named “Garonne” and “Oraison”).

We support AD experiments taking place at Oak Ridge National Laboratory, targetted at building the adjoint of a large CFD application called “Rex”. After one year of collaboration, we reached a first milestone with a working tangent differentiation of a sequential (i.e. non MPI) version of the code. Differentiation in tangent mode is significantly easier than in adjoint mode. Therefore it is a good practice to differentiate first in tangent mode even when the final goal is to produce gradients, which require adjoint mode. Moreover, a validated tangent code is very helpful to validate and debug an adjoint code. The next step will be extension to the MPI-parallel version of the code. This will exploit and develop the AMPI library, co-developped with Argonne National Lab, for automated differentiation of MPI communication routines.