We present capabilities and implementation details of the open-source hybrid-parallel discrete-adjoint flow solver STAMPS (Source Transformed Adjoint Multi-Physics Solver). The adjoint is produced by Automatic Differentiation with the source- transformation tool Tapenade, which results in adjoint code with an extremely low memory footprint. The solver implements the very stable JT-KIRK implicit GMRES scheme to enable robust convergence of the primal and its discrete adjoint. A two-layer halo partitioning minimises the size of MPI messages and avoids MPI communication inside the iterative fixed-point loop which simplifies the adjoint build. The solver offers a complete reverse-mode computation of surface sensitivities, coupling with a differentiated CAD tool in an optimisation case is demonstrated.
展开▼
