NaNs on face-centered values when using grid coarsening in 2.1

[ThomasJannaudCAM]

Hi all,

I recently moved from 2.0 to 2.1 and everything works great, except grid coarsening.
When grid coarsening is on, I get NaNs on face-centered values.

The setup I am using is 3D spherical MHD, going from $0$ to $\pi$ in X2 and from $0$ to $\pi/2$ in X3 on 128 cells, and I do static coarsening along X3.
I use the simple following coarsening function

void CoarsenFunction(DataBlock &data) {. 
  IdefixArray2D<int> coarseningLevel = data.coarseningLevel[KDIR];
  IdefixArray1D<real> th = data.x[JDIR];
  idefix_for("set_coarsening", 0, data.np_tot[JDIR], 0, data.np_tot[IDIR],
      KOKKOS_LAMBDA(int j,int i) {
        int c = 1.0/sin(th(j));
        if(c>6) c = 6;
        coarseningLevel(j,i) = c;
      });
}

that I activate via

if(data.haveGridCoarsening) {
    data.EnrollGridCoarseningLevels(&CoarsenFunction);
  }

Unfortunately my supercomputer is down for the week so I won't be able to do any tests.
I am just wondering if it's an issue some of you have come accross. It should be linked to change #284 (Reduce roundoff error on div(B) when using grid coarsening)

Thanks !

[glesur]

I've seen this before when the grid coarsening level becomes negative (i.e. sin(th)<0). A way to cure this is to use std::abs(sin(th(j))). We should have a way to check for this though...

[ThomasJannaudCAM]

Thanks ! I'll check this when I can.
For my setup that would be strange since $\theta \in [0; \pi]$ in the .ini but I'll look it up.

[glesur]

your domain extends beyond that for the ghost zones, and these zones are also coarsened (and needed for some terms, like resistivity)

[ThomasJannaudCAM]

Ah I see, and then the abs(sin(theta)) makes sense, thanks !