Section: New Results

AD-adjoints of Iterative Processes

Participants : Laurent Hascoët, Ala Taftaf, Sri Hari Krishna Narayanan [Argonne National Lab. (Illinois, USA)] , Daniel Goldberg [University of Edinburgh, UK] .

Adjoint codes naturally propagate partial gradients backwards from the result of the simulation. However, this uses the data flow of the simulation in reverse order, at a cost that increases with the length of the simulation. In the special case of iterative Fixed-Point loops, only the final converged result should be used: the “initial guess” and the intermediate non-converged states should not be considered by the adjoint calculation, and this remark brings enormous gain in memory use. We selected the strategy proposed by Bruce Christianson  [22] and this year we continued its application to medium-size testcases provided by Queen Mary University for the AboutFlow project. We also simplified the user interface provided to trigger this special strategy extension in Tapenade. Ala Taftaf presented her results at the ECCOMAS Eurogen conference in Glasgow [15] , september 14-16.

In parallel, we collaborated with Krishna Narayanan from ANL and Dan Goldberg from University of Edinburgh (UK) to implement the same strategy into the OpenAD tool, in view of applying it to a glaciology configuration of the MIT GCM code. A joint article describing the results has been submitted for publication.