Conservation in a composite linear solve

[ZiyangHuang]

Greetings,

When I use the composite linear solve (solving all amr levels altogether), do I need to additionally do averagedown for the solution and averagedown_face for the fluxes to ensure conservation?

Thanks!

[WeiqunZhang]

You don't need to call averagdown on the solution. However, you do need to call averagedown_faces on fluxes if that's what you want.

[ZiyangHuang]

Thanks for your reply. I have a follow-up question. In the example code, the composite solver covers all the amr levels (i.e., from lev=0 to lev=max_level). However, in practice, we may only look for a solution from lev=0 to lev=finest_level<=max_level. Does it mean that it is preferable to have a composite solver covering lev=0 to lev=finest_level. If that's the case, the vectors of geom, grids, and dmap for building a MLLinOP should contain data only from lev=0 to lev=finest_level, should they?

[WeiqunZhang]

No code was ever mentioned in this Discussion. So I am not sure what's the example code you are talking about. Anyway, the sizes of the vectors of geom, grids and dmap should be the same. For AmrCore and AmrLevel based codes, there are no grids and dmap for levels above the current finest level (i.e., AmrMesh::finest_level). So it's not possible to solve from level 0 to max_level, if max_level > finest_level.

[ZiyangHuang]

The example code I mentioned was "ABecLaplacian_C". I am trying to adapt it to Advection_AmrCore to build a convection-diffusion solver. Based on what you said, in AmrCore, geom.size() = finest_level, does it?

Also, it seems not allowed to construct a vector of MultiFab with {}, which works for geom, grids, and dmap.

[WeiqunZhang]

For AmrCore and AmrLevel based codes, there are no grids and dmap for levels above the current finest level (i.e., AmrMesh::finest_level).

I intentionally did not mention geom because its size is max_level+1, not necessarily finest_level+1. You can use

amrex/Src/AmrCore/AMReX_AmrMesh.H

Line 106 in 3396b1d

[[nodiscard]] Vector<Geometry> Geom (int a_coarsest_lev, int a_finest_lev) const noexcept {

to a vector of Geometry from level 0 to finest_level. Maybe what I said about the sizes of the vectors of geom, grids, dmap should be the same was confusing. What I meant there was the requirement on the arguments passed to the linear solver. At that time, AmrCore had not been mentioned yet.

Also, it seems not allowed to construct a vector of MultiFab with {}

Not sure what it means.

[ZiyangHuang]

Thanks a lot for your reply. Let me rephase my second question. Let MF is a defined MultiFab, is it possible to pass {MF} as a vector of MultiFab to a function?

[WeiqunZhang]

No, because MultiFab does not have copy constructor. This is intentional because we do not want to unnecessarily deep-copy a MultiFab. That's why the interface of a lot of amrex functions uses vector of multifab pointers.

(Could one store MulitFab in a vector? Yes, because MultiFab has a move constructor and move assignment operator.)

[ZiyangHuang]

Thanks! It is clear to me now.