You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
firedrake-adjoint seems to use the (very conservative) default quadrature degree estimation for adjoint solves, even if you specify a much lower quadrature degree for both the forward solver and the assembly of the quantity of interest. This causes some distracting warnings and reduced performance.
It would be nice if firedrake-adjoint could apply the same form compiler from the adjoint problem to the adjoint problem. If there's no way to recover the form compiler, there should at least be a way to specify form compiler parameters for the adjoint problem.
The code below demonstrates the problem; it's reduced from a demo I was making with icepack. I can make a more reduced example that doesn't use icepack if that helps with testing.
I'm closing this issue because I managed to fix it on the icepack side. The change was to attach the quadrature degree to dx, ds, etc. This seems to propagate through to the adjoint.
firedrake-adjoint seems to use the (very conservative) default quadrature degree estimation for adjoint solves, even if you specify a much lower quadrature degree for both the forward solver and the assembly of the quantity of interest. This causes some distracting warnings and reduced performance.
It would be nice if firedrake-adjoint could apply the same form compiler from the adjoint problem to the adjoint problem. If there's no way to recover the form compiler, there should at least be a way to specify form compiler parameters for the adjoint problem.
The code below demonstrates the problem; it's reduced from a demo I was making with icepack. I can make a more reduced example that doesn't use icepack if that helps with testing.
The text was updated successfully, but these errors were encountered: