Adjoining Independent Computations

Laurent Hascoët
Stefka Fidanova
Christophe Held
(INRIA, BP93, 06902 Sophia-Antipolis, France)

Proceedings of the AD2000 Conference, Nice, France, june 2000 (7 pages)

Abstract: The reverse or adjoint mode of automatic differentiation is a software engineering technique that permits efficient computation of gradients. However, this technique requires a lot of temporary memory. In this paper, we present a refinement that reduces memory consumption in the case of parallel loops, and we give a proof of its correctness based on properties of the data-dependence graph of adjoint programs and parallel loops. This technique is particularly suitable for assembly loops that dominate in mesh-based computations. Application is done on the kernel of a realistic Navier-Stokes solver.

Keywords:Automatic Differentiation, reverse mode, adjoints, parallel loops, data-dependence, Navier-Stokes

Full text (pdf)

  author = {Hasco\"et, L. and Fidanova, S. and Held, C.},
  title = {Adjoining Independent Computations},
  booktitle = {Automatic Differentiation of Algorithms, from Simulation to Optimization},
  series = "Computer and Information Science",
  publisher = "Springer",
  note = {selected papers from the AD2000 conference, Nice, France},
  pages = "299-304",
  year = 2001