Adjoining Independent Computations
- 319 Downloads
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 chapter, 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.
KeywordsAutomatic Differentiation Parallel Loop Forward Sweep Reverse Sweep Adjoint Mode
Unable to display preview. Download preview PDF.