Refinement Criterion Near Particles in errorEst()

[Afraid31]

Hi everyone,

I’m working on implementing a refinement criterion in errorEst() that tags only the cells near a set of boundary points, rather than tagging all cells inside a large bounding box.

In my AmrLevel setup, there are four levels of refinement. Previously, I used the following approach:

auto bounds = IBP->ComputeBoundary();
auto pbox_lo = bounds.first;
auto pbox_hi = bounds.second;

for (int d = 0; d < BL_SPACEDIM; ++d) {
    pbox_lo[d] -= 10.0 * dx[d];
    pbox_hi[d] += 10.0 * dx[d];
}

#ifdef AMREX_USE_OMP
#pragma omp parallel if (Gpu::notInLaunchRegion())
#endif
for (MFIter mfi(tags, TilingIfNotGPU()); mfi.isValid(); ++mfi) {
    const Box& tilebx = mfi.tilebox();
    auto const& tagarr = tags.array(mfi);
    auto prob_lo = geom.ProbLoArray();

    amrex::ParallelFor(tilebx,
        [=] AMREX_GPU_DEVICE(int i, int j, int k) noexcept
    {
        particle_box_error(i, j, k, tagarr, dx, prob_lo, pbox_lo, pbox_hi, tagval);
    });
}

Here, IBP is a user-defined class derived from AmrParticleContainer.
Its member function ComputeBoundary() computes the minimum and maximum particle coordinates in each dimension (i.e., the bounding box of all particles).

Now, I would like to modify this so that only the cells close to the boundary points (for example, within a few grid cells from each boundary point) are tagged for refinement.

What’s the best way to achieve this in AMReX?
Should I loop over the boundary point list directly inside the GPU kernel, or is there a more efficient way to mark nearby cells?

Thanks in advance for any suggestions!

[atmyers]

You could use amrex::ParticleToMesh to deposit the particles in your container, and then tag cells where the corresponding density is non-zero. This will do the loop over the boundary points internally in a performance portable way.

This shows how to use ParticleToMesh with linear interpolation, you could also modify the interpolation kernel to include more points if you wish.

-Andrew