Abstract
A specification-based method for the automatic generation of executable Fortran adjoint code is presented. The method is embedded within the program generation framework for the forward model, and automatic differentiation techniques are applied to the forward-model specifications themselves rather than to its Fortran source. A distinction is made between linearization and stencil processing. For the latter, special adjointing rules for stencil operators are used. The work is discussed in the light of various computational differentiation issues including arithmetic efficiency, forward/reverse hybridization trade-offs, portability to high-performance platforms, and source-to-source adjoint methodologies. Results of a computer algebra-based prototype are illustrated for forward and adjoint code for the dynamics part of a high-resolution, limited-area weather forecasting grid-point model.
| Original language | English |
|---|---|
| Title of host publication | Computational Differentiation, Techniques, Applications and Tools |
| Editors | Martin Berz |
| Place of Publication | Philadelphia, USA |
| Publisher | SIAM |
| Pages | 185-194 |
| Number of pages | 10 |
| ISBN (Print) | 9780898713855 |
| Publication status | Published - 6 Feb 1996 |
| Event | 2nd International Workshop on Computational Differentiation 1996 - Sante Fe, United States Duration: 12 Feb 1996 → 14 Feb 1996 Conference number: 2 |
Publication series
| Name | Siam Proceedings in Applied Mathematics Series |
|---|---|
| Publisher | Siam |
| Volume | 89 |
Workshop
| Workshop | 2nd International Workshop on Computational Differentiation 1996 |
|---|---|
| Country/Territory | United States |
| City | Sante Fe |
| Period | 12/02/96 → 14/02/96 |