Computing Adjoints by Automatic Differentiation with TAPENADE

Laurent Hascoët
Rose-Marie Greborio
Valérie Pascual
(INRIA, BP93, 06902 Sophia-Antipolis, France)

Book chapter, "Problemes non-lineaires appliques" (21 pages)

Abstract: We present the Automatic Differentiation (AD) tool TAPENADE, with emphasis on the so-called reverse mode which computes gradients. Computing gradients with the reverse mode is the discrete equivalent of writing and then solving the adjoint equations. We present the main usages of AD, and its fundamental model, based on the chain rule. We detail the architecture and algorithms used in TAPENADE, highlighting the strategies used for an efficient reverse mode. We present the user interface of TAPENADE and give pointers to the on-line documentation. We describe the difficulties that AD tools still have to overcome to reach a larger audience, and how we plan to address them.

Keywords: Automatic Differentiation, Algorithmic Differentiation, Adjoint, Gradient, Optimization, Inverse Problems, Static Analysis, Data-Flow Analysis, Compilation

Full text (pdf)

  author = {Hasco\"et, L. and Greborio, R.-M. and Pascual, V.},
  title = {Computing Adjoints by Automatic Differentiation with TAPENADE},
  booktitle = {Ecole INRIA-CEA-EDF "Problemes non-lineaires appliques"},
  editor = {Sportisse, B. and LeDimet, F.-X.},
  publisher = "Springer",
  note = {to appear},