Adding tools for DG methods to amrex code-base?

[dunhamsj]

Hello, I've made additions to my fork of amrex that allow a user to call FillPatch and AverageDown on a MultiFab containing a nodal DG solution for each field. Would it be possible to merge these features into the main code-base of amrex? They have not been tested on GPUs, but they pass all of the github actions.

[WeiqunZhang]

Could you explain what's "a MultiFab containing a nodal DG solution for each field" and what you need to do in FillPatch and AverageDown?

[dunhamsj]

Ah, I was confused. I didn't have any of the dg_interp functions in my new class. Now it makes sense.

I'm now wondering though how to use my FillPatch routine that takes in four multifabs: two multifabs contain the fluid (one on the coarse level and one on the fine level) and the other two contain the geometry (one for the coarse level and one for the fine level), which is needed to interpolate while conserving the cell averages. Can that be done, or should I create a new multifab that holds the fluid and the geometry?

[WeiqunZhang]

I see. In this case, I think you can make your interpolater derive from MFInterpolater instead of Interpolater.

DGInterpolater final
    : public MFInterpolater
{
public:
    DGInterpolater(....); // constructor

    Box CoarseBox (Box const& fine, int ratio) override;
    Box CoarseBox (Box const& fine, IntVect const& ratio) override;

    void interp (MultiFab const& crsemf, int ccomp, MultiFab& finemf, int fcomp, int ncomp,
                         IntVect const& ng, Geometry const& cgeom, Geometry const& fgeom,
                         Box const& dest_domain, IntVect const& ratio,
                         Vector<BCRec> const& bcs, int bcomp) override;
}

You can then pass additional information into the constructor and store them. Everytime you call FillPatch, you need to make a local DGInterpolater object that has the information you need.

[dunhamsj]

I see. I would want this to be something useful to other users and it seems like it's becoming a little too specialized for that. Because of that, I think I'll just keep the additional code on our repo. Thank you very much for your help

[BenWibking]

@dunhamsj Just FYI- if it's a consideration as to whether to upstream this, the ability to do fixed-order nodal DG with AMR would indeed be very useful for our code as well.

[dunhamsj]

Hi @BenWibking, that's good to know, I was wondering how useful it would be to others. I'm concerned because the way we handle curvilinear coordinates is different from how amrex does it, and I'm worried that may cause confusion. I will keep thinking about it though