Comment 1 for bug 1176805

Revision history for this message
Patrick Farrell (pefarrell) wrote :

One of the most important/difficult/complicated features of libadjoint is the handling of forward state: you want to forget as much as you can as soon as you can, but no earlier. The algorithm that's there does a pretty good job, but I added that warning message to trap a case that I hadn't dealt with: where the computation of the adjoint solution doesn't require a particular forward variable (and so it's forgotten), but the computation of the \partial F/\partial m term does. If not solved, it will lead to a significant inefficiency, unfortunately, so I'd really like to deal with it. Dealing with it involves careful tracing of exactly when things are computed, stored, and forgotten, and requires incense, self-flagellation and the occasional sacrificial goat.

I'll be at Simula from 13 May so maybe you can show me then?