REMORA.cpp

Go to the documentation of this file.

/**
 * \file REMORA.cpp
 */
 
#include <REMORA_prob_common.H>
#include <REMORA.H>
#include <REMORA_MaskedAverageDown.H>
 
#ifdef REMORA_USE_NETCDF
#include "REMORA_NCFile.H"
#endif
 
#include <AMReX_buildInfo.H>
 
using namespace amrex;
 
amrex::Real REMORA::startCPUTime        = zero;
amrex::Real REMORA::previousCPUTimeUsed = zero;
 
Vector<REMORAErrorTag> REMORA::ref_tags;
 
SolverChoice REMORA::solverChoice;
 
// Time step control
amrex::Real REMORA::cfl           =  Real(0.8);
amrex::Real REMORA::fixed_dt      = -one;
amrex::Real REMORA::change_max    =  Real(1.1);
 
int   REMORA::ndtfast             = 0;
 
// Dictate verbosity in screen output
int         REMORA::verbose       = 0;
 
// Frequency of diagnostic output
int         REMORA::sum_interval  = -1;
amrex::Real REMORA::sum_per       = -one;
 
// Minimum number of digits in plotfile name
int         REMORA::file_min_digits = 5;
 
// Do we include staggered velocities in the plotfile?
bool        REMORA::plot_staggered_vels = false;
 
// Do we include nodal data (Nu_nd) in the plotfile?
bool        REMORA::plot_nodal_data = true;
 
// Native AMReX vs NetCDF
PlotfileType REMORA::plotfile_type    = PlotfileType::amrex;
 
#ifdef REMORA_USE_NETCDF
 
int   REMORA::total_nc_plot_file_step = 1;
 
// Do we write one file per timestep (false) or one file for all timesteps (true)
bool  REMORA::write_history_file      = true;
 
// NetCDF initialization file
amrex::Vector<std::string> REMORA::nc_bdry_file = {""}; // Must provide via input
amrex::Vector<amrex::Vector<std::string>> REMORA::nc_init_file = {{""}}; // Must provide via input
amrex::Vector<amrex::Vector<std::string>> REMORA::nc_grid_file = {{""}}; // Must provide via input
#endif
 
/**
 * constructor:
 * - reads in parameters from inputs file
 * - sizes multilevel arrays and data structures
 * - initializes BCRec boundary condition object
 */
REMORA::REMORA ()
{
    BL_PROFILE("REMORA::REMORA()");
 
    if (ParallelDescriptor::IOProcessor()) {
        const char* remora_hash = amrex::buildInfoGetGitHash(1);
        const char* amrex_hash = amrex::buildInfoGetGitHash(2);
        const char* buildgithash = amrex::buildInfoGetBuildGitHash();
        const char* buildgitname = amrex::buildInfoGetBuildGitName();
 
        if (strlen(remora_hash) > 0) {
          amrex::Print() << "\n"
                         << "REMORA git hash: " << remora_hash << "\n";
        }
        if (strlen(amrex_hash) > 0) {
          amrex::Print() << "AMReX git hash: " << amrex_hash << "\n";
        }
        if (strlen(buildgithash) > 0) {
          amrex::Print() << buildgitname << " git hash: " << buildgithash << "\n";
        }
 
        amrex::Print() << "\n";
    }
 
    ReadParameters();
 
    // Blocking factor in z set to very large value to be > nz
    // This guarantees that there will be no domain decomposition in the z-direction
    // We have to set this by hand here because setting it in the input file will
    // cause checks in the AmrCore constructor to fail.
    Vector<IntVect> blocking_factor_vec = Vector<IntVect>();
    blocking_factor_vec.resize(max_level+1);
    for (int lev = 0; lev <= max_level; ++lev) {
        blocking_factor_vec[lev] = blockingFactor(lev);
        blocking_factor_vec[lev][2] = 4096;
    }
    SetBlockingFactor(blocking_factor_vec);
 
    const std::string& pv3d = "plot_vars_3d"; set3DPlotVariables(pv3d);
    const std::string& pv2d = "plot_vars_2d"; set2DPlotVariables(pv2d);
 
    prob = amrex_probinit(geom[0].ProbLo(),geom[0].ProbHi());
 
    // Geometry on all levels has been defined already.
 
    // No valid BoxArray and DistributionMapping have been defined.
    // But the arrays for them have been resized.
 
    int nlevs_max = max_level + 1;
 
    istep.resize(nlevs_max, 0);
    set_nsubsteps(nlevs_max);
 
    physbcs.resize(nlevs_max);
 
    t_new.resize(nlevs_max, zero);
    t_old.resize(nlevs_max, -bogus_large_value);
    dt.resize(nlevs_max, bogus_large_value);
 
    cons_new.resize(nlevs_max);
    cons_old.resize(nlevs_max);
    xvel_new.resize(nlevs_max);
    xvel_old.resize(nlevs_max);
    yvel_new.resize(nlevs_max);
    yvel_old.resize(nlevs_max);
    zvel_new.resize(nlevs_max);
    zvel_old.resize(nlevs_max);
 
    advflux_reg.resize(nlevs_max);
 
    // Initialize tagging criteria for mesh refinement
    refinement_criteria_setup();
 
    init_ref_ratios();
}
 
REMORA::REMORA (const amrex::RealBox& rb, int max_level_in, const amrex::Vector<int>& n_cell_in, int coord, const amrex::Vector<amrex::IntVect>& ref_ratio_in, const amrex::Array<int,AMREX_SPACEDIM>& is_per, std::string prefix)
    : amrex::AmrCore (rb, max_level_in, n_cell_in, coord, ref_ratio_in, is_per)
{
    BL_PROFILE("REMORA::REMORA(explicit)");
    pp_prefix = prefix;
 
    if (ParallelDescriptor::IOProcessor()) {
        const char* remora_hash = amrex::buildInfoGetGitHash(1);
        const char* amrex_hash = amrex::buildInfoGetGitHash(2);
        const char* buildgithash = amrex::buildInfoGetBuildGitHash();
        const char* buildgitname = amrex::buildInfoGetBuildGitName();
 
        if (strlen(remora_hash) > 0) {
          amrex::Print() << "\n"
                         << "REMORA git hash: " << remora_hash << "\n";
        }
        if (strlen(amrex_hash) > 0) {
          amrex::Print() << "AMReX git hash: " << amrex_hash << "\n";
        }
        if (strlen(buildgithash) > 0) {
          amrex::Print() << buildgitname << " git hash: " << buildgithash << "\n";
        }
 
        amrex::Print() << "\n";
    }
 
    ReadParameters();
 
    const std::string& pv3d = "plot_vars_3d"; set3DPlotVariables(pv3d);
    const std::string& pv2d = "plot_vars_2d"; set2DPlotVariables(pv2d);
 
    prob = amrex_probinit(geom[0].ProbLo(),geom[0].ProbHi());
 
    int nlevs_max = max_level + 1;
 
    istep.resize(nlevs_max, 0);
    set_nsubsteps(nlevs_max);
 
    physbcs.resize(nlevs_max);
 
    t_new.resize(nlevs_max, zero);
    t_old.resize(nlevs_max, -bogus_large_value);
    dt.resize(nlevs_max, bogus_large_value);
 
    cons_new.resize(nlevs_max);
    cons_old.resize(nlevs_max);
    xvel_new.resize(nlevs_max);
    xvel_old.resize(nlevs_max);
    yvel_new.resize(nlevs_max);
    yvel_old.resize(nlevs_max);
    zvel_new.resize(nlevs_max);
    zvel_old.resize(nlevs_max);
 
    advflux_reg.resize(nlevs_max);
 
    refinement_criteria_setup();
 
    init_ref_ratios();
}
 
REMORA::~REMORA ()
{
}
 
/**
 * Reject refinement in the vertical, and accumulate the refinement ratios.
 *
 * Shared by both constructors. It used to be written out in each of them, and the explicit one
 * had been left without the cum_ref_ratios half -- so the vector stayed empty, and every
 * full-domain hires array and every mask coarsening that indexes it read out of bounds.
 */
void
REMORA::init_ref_ratios ()
{
    AMREX_ALWAYS_ASSERT(cum_ref_ratios.empty());
 
    IntVect cum_ref_ratio = IntVect(1,1,0);
    cum_ref_ratios.push_back(cum_ref_ratio);
    // We have already read in the ref_ratio (via amr.ref_ratio =) but we need to enforce
    //     that there is no refinement in the vertical so we test on that here.
    for (int lev = 0; lev < max_level; ++lev)
    {
       amrex::Print() << "Refinement ratio at level " << lev << " set to be " <<
          ref_ratio[lev][0]  << " " << ref_ratio[lev][1]  <<  " " << ref_ratio[lev][2] << std::endl;
 
       if (ref_ratio[lev][2] != 1)
       {
           amrex::Print() << "********************************************************************************" << std::endl;
           amrex::Print() << "We don't allow refinement in the vertical -- make sure to set ref_ratio = 1 in z" << std::endl;
           amrex::Print() << "It's possible you set amr.ref_ratio when you meant to set amr.ref_ratio_vect    " << std::endl;
           amrex::Print() << "********************************************************************************" << std::endl;
           amrex::Abort();
       }
 
       cum_ref_ratio[0] *= ref_ratio[lev][0];
       cum_ref_ratio[1] *= ref_ratio[lev][1];
       cum_ref_ratios.push_back(cum_ref_ratio);
    }
}
 
void
REMORA::init_scalar_metadata ()
{
    cons_names.clear();
    cons_names.reserve(ncons);
    cons_names.emplace_back("temp");
    cons_names.emplace_back("salt");
 
    // Passive (dye) scalars come first, then the biology block, matching the component
    // layout: temp, salt, tracer, tracer_1, ..., NO3, NH4, ...
    if (nscalar > 0) {
        cons_names.emplace_back("tracer");
        for (int i = 1; i < nscalar; ++i) {
            cons_names.emplace_back("tracer_" + std::to_string(i));
        }
    }
 
    if (REMORABiology::has_biology(biology_model)) {
        const auto bio_names = REMORABiology::tracer_names(biology_model, fennel_params);
        for (const auto& name : bio_names) {
            cons_names.emplace_back(name);
        }
    }
 
    AMREX_ALWAYS_ASSERT(static_cast<int>(cons_names.size()) == ncons);
}
 
void
REMORA::Evolve ()
{
    BL_PROFILE_VAR("REMORA::Evolve()",evolve);
    Real cur_time = t_new[0];
    const Real stop_elapsed = elapsed_time(stop_time);
 
    // istep[0] advances inside the loop, so keep the value it started at.
    const int first_step = istep[0];
 
    // Levels appear as tagging occurs, so reprint the hierarchy when finest_level changes.
    int reported_finest = -1;
 
    // Take one coarse timestep by calling timeStep -- which recursively calls timeStep
    //      for finer levels (with or without subcycling)
    for (int step = istep[0]; step < max_step && cur_time < stop_elapsed; ++step)
    {
        amrex::Print() << "\nCoarse STEP " << step+1 << " starts ..." << std::endl;
 
        ComputeDt();
 
        // dt is only populated once ComputeDt has run.
        if (step == first_step || finest_level != reported_finest) {
            print_timestep_hierarchy();
            reported_finest = finest_level;
        }
 
        int lev = 0;
        int iteration = 1;
        auto dEvolveTime0 = amrex::second();
 
        // timeStep recurses into finer levels nsubsteps[lev+1] times; timeStepML advances
        // every level once through one shared barotropic loop.
        if (max_level == 0 || do_substep) {
            timeStep(lev, cur_time, iteration);
        }
        else {
            timeStepML(cur_time, iteration);
        }
 
        cur_time  += dt[0];
 
        amrex::Print() << "Coarse STEP " << step+1 << " ends." << " TIME = " << cur_time
                       << " DT = " << dt[0]  << std::endl;
 
        if (verbose > 0)
        {
            auto dEvolveTime = amrex::second() - dEvolveTime0;
            ParallelDescriptor::ReduceRealMax(dEvolveTime,ParallelDescriptor::IOProcessorNumber());
            amrex::Print() << "Timestep time = " << dEvolveTime << " seconds." << '\n';
        }
 
        WriteAtIntermediateTime(step, cur_time);
 
        post_timestep(step, cur_time, dt[0]);
 
#ifdef AMREX_MEM_PROFILING
        {
            std::ostringstream ss;
            ss << "[STEP " << step+1 << "]";
            MemProfiler::report(ss.str());
        }
#endif
 
        if (cur_time >= stop_elapsed - 1.e-6*dt[0]) break;
    }
 
    BL_PROFILE_VAR_STOP(evolve);
 
    WriteAtFinalTime();
}
 
void
REMORA::WriteAtFinalTime()
{
 
    if ( (plot_int > 0 || plot_int_time > zero) && istep[0] > last_plot_file_step)
    {
        WritePlotFile(istep[0]);
        history_count++;
    }
 
    if ((check_int > 0 || check_int_time > zero) && istep[0] > last_check_file_step) {
        WriteCheckpointFile();
    }
}
 
void
REMORA::WriteAtIntermediateTime(int step, amrex::Real cur_time)
{
    if ( (plot_int > 0      && (step+1 - last_plot_file_step) == plot_int         ) ||
         (plot_int_time > 0 && (cur_time >= (last_plot_file_time + plot_int_time))) )
    {
        last_plot_file_step = step+1;
        last_plot_file_time = cur_time;
        WritePlotFile(step+1);
        history_count++;
    }
 
    if ((check_int > 0 && (step+1 - last_check_file_step) == check_int)
            || (check_int_time > 0 && cur_time >= (last_check_file_time + check_int_time))) {
        last_check_file_step = step+1;
        last_check_file_time = cur_time;
        WriteCheckpointFile();
    }
}
 
/**
 * Apply the tracer flux correction accumulated at the lev/lev+1 interface onto lev.
 *
 * Once per step of lev, pairing the reset at the top of Advance(lev), so the accumulate and
 * apply window sits inside one step of lev. Applying it per step of level 0 instead would
 * keep only the last correction whenever lev was itself substepped, and would let the regrid
 * of lev+1 rebuild the register mid-accumulation.
 *
 * @param[in] lev            coarse level of the interface
 */
void
REMORA::reflux_to (int lev)
{
    if (!(do_reflux && do_substep) || lev >= finest_level ||
        solverChoice.coupling_type != CouplingType::two_way) {
        return;
    }
 
    BL_PROFILE("REMORA::reflux_to()");
 
    // The register holds the correction in Hz*t units, the form the tracer is advanced in.
    // Reflux into a scratch fab and divide that down, rather than scaling the level into
    // those units and back: that round trip is inexact for about a tenth of cells, which
    // perturbs cells the correction never reached and leaves the clamp below unable to tell
    // which ones it did.
    MultiFab dcons(cons_new[lev]->boxArray(), cons_new[lev]->DistributionMap(), ncons, 0);
    dcons.setVal(zero);
    getAdvFluxReg(lev+1)->Reflux(dcons, 0, 0, ncons);
 
    const bool clamp = reflux_clamp;
 
#ifdef _OPENMP
#pragma omp parallel if (Gpu::notInLaunchRegion())
#endif
    for (MFIter mfi(*cons_new[lev], TilingIfNotGPU()); mfi.isValid(); ++mfi)
    {
        Array4<Real      > const& c  = cons_new[lev]->array(mfi);
        Array4<Real const> const& dc = dcons.const_array(mfi);
        Array4<Real const> const& hz = vec_Hz[lev]->const_array(mfi);
 
        ParallelFor(mfi.tilebox(), ncons, [=] AMREX_GPU_DEVICE (int i, int j, int k, int n)
        {
            // No thickness is land or dry: leave it rather than divide by zero.
            if (dc(i,j,k,n) == zero || hz(i,j,k) <= zero) { return; }
 
            c(i,j,k,n) += dc(i,j,k,n) / hz(i,j,k);
 
            // Clamping restores the mass the correction removed, so a step that clamps is
            // not conservative: ROMS's trade, and why it is an option. Zero suits a
            // concentration, less so temperature in Celsius.
            if (clamp && c(i,j,k,n) < zero) { c(i,j,k,n) = zero; }
        });
    }
}
 
/**
 * @param[in   ] nstep    which step we're on
 * @param[in   ] time     current time
 * @param[in   ] dt_lev0  time step on level 0
 */
void
REMORA::post_timestep (int nstep, Real time, Real dt_lev0)
{
    BL_PROFILE("REMORA::post_timestep()");
 
#ifdef REMORA_USE_PARTICLES
    particleData.Redistribute();
#endif
 
    if (solverChoice.coupling_type == CouplingType::two_way)
    {
        for (int lev = finest_level-1; lev >= 0; lev--)
        {
            AverageDownTo(lev);
        }
    }
 
    if (is_it_time_for_action(nstep, time, dt_lev0, sum_interval, sum_per)) {
        sum_integrated_quantities(time);
    }
}
 
/**
 * This is called from main.cpp and handles all initialization, whether from start or restart
 */
void
REMORA::InitData ()
{
    BL_PROFILE("REMORA::InitData()");
    if (running_with_coupling_driver) {
        amrex::Print() << "REMORA InitData: driver-managed atm2ocn coupling enabled"
                       << " two_way=" << (driver_uses_two_way_coupling ? 1 : 0)
                       << " active_contract="
                       << (DriverUsesStateForcing(driver_atmos_forcing_mode) ? "state" : "flux")
                       << "\n";
    }
    // Initialize the start time for our CPU-time tracker
    startCPUTime = Real(ParallelDescriptor::second());
 
    // Map the words in the inputs file to BC types, then translate
    //     those types into what they mean for each variable
    init_bcs();
 
    // Init vertical stretching coeffs
    init_stretch_coeffs();
 
    last_plot_file_step = -1;
    last_check_file_step = -1;
    last_plot_file_time = -one;
    last_check_file_time = -one;
 
    if (restart_chkfile == "") {
        // start simulation from the beginning
 
        // t_new counts from start_time, so a fresh run starts at zero.
        InitFromScratch(zero);
 
        if (solverChoice.coupling_type == CouplingType::two_way) {
            AverageDown();
        }
 
    } else { // Restart from a checkpoint
 
        restart();
 
    }
 
    // Every level's mask exists by now, whether built from scratch or read back
    check_mask_consistency();
 
#ifdef REMORA_USE_MOAB
    InitMOABMesh();
#endif
    // Levels that appear later, or are regridded, define their own from make_new_level.
    if (solverChoice.coupling_type == CouplingType::two_way) {
        advflux_reg[0] = nullptr;
        for (int lev = 1; lev <= finest_level; lev++) {
            define_flux_register(lev);
        }
    }
 
    // Fill ghost cells/faces
    for (int lev = 0; lev <= finest_level; ++lev)
    {
        if (lev > 0 && cf_width >= 0) {
            Construct_REMORAFillPatchers(lev);
        }
 
        if (restart_chkfile == "") {
            FillPatch(lev, t_new[lev], *cons_new[lev], cons_new, BCVars::cons_bc, BdyVars::t, 0, true, false,0,0,zero,*cons_new[lev]);
            FillPatch(lev, t_new[lev], *xvel_new[lev], xvel_new, xvel_bc(), BdyVars::u, 0, true, false,0,0,zero,*xvel_new[lev]);
            FillPatch(lev, t_new[lev], *yvel_new[lev], yvel_new, yvel_bc(), BdyVars::v, 0, true, false,0,0,zero,*yvel_new[lev]);
            FillPatch(lev, t_new[lev], *zvel_new[lev], zvel_new, zvel_bc(), BdyVars::null, 0, true, false);
 
            // Copy from new into old just in case when initializing from scratch
            int ngs   = cons_new[lev]->nGrow();
            int ngvel = xvel_new[lev]->nGrow();
            MultiFab::Copy(*cons_old[lev],*cons_new[lev],0,0,ncons,ngs);
            MultiFab::Copy(*xvel_old[lev],*xvel_new[lev],0,0,1,ngvel);
            MultiFab::Copy(*yvel_old[lev],*yvel_new[lev],0,0,1,ngvel);
            MultiFab::Copy(*zvel_old[lev],*zvel_new[lev],0,0,1,IntVect(ngvel,ngvel,0));
        }
    } // lev
 
    // Check for additional plotting variables that are available after
    // particle containers are setup.
    const std::string& pv3d = "plot_vars_3d"; append3DPlotVariables(pv3d);
    const std::string& pv2d = "plot_vars_2d"; append2DPlotVariables(pv2d);
 
    if (restart_chkfile == "" && (check_int > 0 || check_int_time > zero))
    {
        WriteCheckpointFile();
        last_check_file_step = 0;
    }
 
    // plot_file_on_restart currently always 1
    if ( (restart_chkfile == "") ||
         (restart_chkfile != "" && plot_file_on_restart) )
    {
        if (plot_int > 0 || plot_int_time > zero)
        {
            int step0 = 0;
            WritePlotFile(step0);
            history_count++;
            last_plot_file_step = istep[0];
        }
    }
 
    if (is_it_time_for_action(istep[0], t_new[0], dt[0], sum_interval, sum_per)) {
        sum_integrated_quantities(t_new[0]);
    }
 
    // dt is read from checkpoint on restart so it only needs to be computed if
    // not restarting
    if (restart_chkfile == "") {
        ComputeDt();
    }
 
}
 
/**
 * @param[in   ] lev    level to operate on
 */
void
REMORA::Construct_REMORAFillPatchers (int lev)
{
    BL_PROFILE("REMORA::Construct_REMORAFillPatchers()");
    amrex::Print() << ":::Construct_REMORAFillPatchers " << lev << std::endl;
 
    auto& ba_fine  = cons_new[lev  ]->boxArray();
    auto& ba_crse  = cons_new[lev-1]->boxArray();
    auto& dm_fine  = cons_new[lev  ]->DistributionMap();
    auto& dm_crse  = cons_new[lev-1]->DistributionMap();
 
    BoxList bl2d_fine = ba_fine.boxList();
    for (auto& b : bl2d_fine) {
        b.setRange(2,0);
    }
    BoxArray ba2d_fine(std::move(bl2d_fine));
 
    BoxList bl2d_crse = ba_crse.boxList();
    for (auto& b : bl2d_crse) {
        b.setRange(2,0);
    }
    BoxArray ba2d_crse(std::move(bl2d_crse));
 
    int ncomp = cons_new[lev]->nComp();
 
    FPr_c.emplace_back(ba_fine, dm_fine, geom[lev]  ,
                       ba_crse, dm_crse, geom[lev-1],
                       -cf_width, -cf_set_width, ncomp, &cell_cons_interp);
    FPr_u.emplace_back(convert(ba_fine, IntVect(1,0,0)), dm_fine, geom[lev]  ,
                       convert(ba_crse, IntVect(1,0,0)), dm_crse, geom[lev-1],
                       -cf_width, -cf_set_width, 1, &face_cons_linear_interp);
    FPr_v.emplace_back(convert(ba_fine, IntVect(0,1,0)), dm_fine, geom[lev]  ,
                       convert(ba_crse, IntVect(0,1,0)), dm_crse, geom[lev-1],
                       -cf_width, -cf_set_width, 1, &face_cons_linear_interp);
    FPr_w.emplace_back(convert(ba_fine, IntVect(0,0,1)), dm_fine, geom[lev]  ,
                       convert(ba_crse, IntVect(0,0,1)), dm_crse, geom[lev-1],
                       -cf_width, -cf_set_width, 1, &face_cons_linear_interp);
 
    FPr_ubar.emplace_back(convert(ba2d_fine, IntVect(1,0,0)), dm_fine, geom[lev]  ,
                       convert(ba2d_crse, IntVect(1,0,0)), dm_crse, geom[lev-1],
                       -cf_width, -cf_set_width, 3, &face_cons_linear_interp);
    FPr_vbar.emplace_back(convert(ba2d_fine, IntVect(0,1,0)), dm_fine, geom[lev]  ,
                       convert(ba2d_crse, IntVect(0,1,0)), dm_crse, geom[lev-1],
                       -cf_width, -cf_set_width, 3, &face_cons_linear_interp);
 
    FPr_Dubar.emplace_back(convert(ba2d_fine, IntVect(1,0,0)), dm_fine, geom[lev]  ,
                       convert(ba2d_crse, IntVect(1,0,0)), dm_crse, geom[lev-1],
                       -cf_width, -cf_set_width, 1, &face_cons_linear_interp);
    FPr_Dvbar.emplace_back(convert(ba2d_fine, IntVect(0,1,0)), dm_fine, geom[lev]  ,
                       convert(ba2d_crse, IntVect(0,1,0)), dm_crse, geom[lev-1],
                       -cf_width, -cf_set_width, 1, &face_cons_linear_interp);
}
 
/**
 * @param[in   ] lev    level to operate on
 */
void
REMORA::Define_REMORAFillPatchers (int lev)
{
    BL_PROFILE("REMORA::Define_REMORAFillPatchers()");
    if (verbose > 0) {
        amrex::Print() << ":::Define_REMORAFillPatchers " << lev << std::endl;
    }
 
    auto& ba_fine  = cons_new[lev  ]->boxArray();
    auto& ba_crse  = cons_new[lev-1]->boxArray();
    auto& dm_fine  = cons_new[lev  ]->DistributionMap();
    auto& dm_crse  = cons_new[lev-1]->DistributionMap();
 
    BoxList bl2d_fine = ba_fine.boxList();
    for (auto& b : bl2d_fine) {
        b.setRange(2,0);
    }
    BoxArray ba2d_fine(std::move(bl2d_fine));
 
    BoxList bl2d_crse = ba_crse.boxList();
    for (auto& b : bl2d_crse) {
        b.setRange(2,0);
    }
    BoxArray ba2d_crse(std::move(bl2d_crse));
 
 
    int ncomp = cons_new[lev]->nComp();
 
    FPr_c[lev-1].Define(ba_fine, dm_fine, geom[lev]  ,
                        ba_crse, dm_crse, geom[lev-1],
                        -cf_width, -cf_set_width, ncomp, &cell_cons_interp);
    FPr_u[lev-1].Define(convert(ba_fine, IntVect(1,0,0)), dm_fine, geom[lev]  ,
                        convert(ba_crse, IntVect(1,0,0)), dm_crse, geom[lev-1],
                        -cf_width, -cf_set_width, 1, &face_cons_linear_interp);
    FPr_v[lev-1].Define(convert(ba_fine, IntVect(0,1,0)), dm_fine, geom[lev]  ,
                        convert(ba_crse, IntVect(0,1,0)), dm_crse, geom[lev-1],
                        -cf_width, -cf_set_width, 1, &face_cons_linear_interp);
    FPr_w[lev-1].Define(convert(ba_fine, IntVect(0,0,1)), dm_fine, geom[lev]  ,
                        convert(ba_crse, IntVect(0,0,1)), dm_crse, geom[lev-1],
                        -cf_width, -cf_set_width, 1, &face_cons_linear_interp);
 
    FPr_ubar[lev-1].Define(convert(ba2d_fine, IntVect(1,0,0)), dm_fine, geom[lev]  ,
                        convert(ba2d_crse, IntVect(1,0,0)), dm_crse, geom[lev-1],
                        -cf_width, -cf_set_width, 3, &face_cons_linear_interp);
    FPr_vbar[lev-1].Define(convert(ba2d_fine, IntVect(0,1,0)), dm_fine, geom[lev]  ,
                        convert(ba2d_crse, IntVect(0,1,0)), dm_crse, geom[lev-1],
                        -cf_width, -cf_set_width, 3, &face_cons_linear_interp);
 
    FPr_Dubar[lev-1].Define(convert(ba2d_fine, IntVect(1,0,0)), dm_fine, geom[lev]  ,
                        convert(ba2d_crse, IntVect(1,0,0)), dm_crse, geom[lev-1],
                        -cf_width, -cf_set_width, 1, &face_cons_linear_interp);
    FPr_Dvbar[lev-1].Define(convert(ba2d_fine, IntVect(0,1,0)), dm_fine, geom[lev]  ,
                        convert(ba2d_crse, IntVect(0,1,0)), dm_crse, geom[lev-1],
                        -cf_width, -cf_set_width, 1, &face_cons_linear_interp);
}
 
void
REMORA::restart ()
{
    BL_PROFILE("REMORA::restart()");
    ReadCheckpointFile();
 
    // We set this here so that we don't over-write the checkpoint file we just started from
    last_check_file_step = istep[0];
    // last_plot_file_step will be updated when plotfile is unconditionally written after restart
 
    last_plot_file_time  = t_new[0];
    last_check_file_time = t_new[0];
}
 
/**
 * @param[in   ] lev    level to operate on
 */
void
REMORA::set_zeta (int lev)
{
    BL_PROFILE("REMORA::set_zeta()");
    if (lev==0) {
        if (hires_init_level < 0) {
            if (solverChoice.ic_type == IC_Type::analytic) {
                prob->init_analytic_zeta(lev, geom[lev], solverChoice, *this, prob_coords(lev), *vec_zeta[lev]);
            } else if (solverChoice.ic_type == IC_Type::netcdf) {
#ifdef REMORA_USE_NETCDF
                amrex::Print() << "Calling init_zeta_from_netcdf on level " << lev << std::endl;
                init_zeta_from_netcdf(lev);
                amrex::Print() << "Sea surface height loaded from netcdf file \n " << std::endl;
#endif
            } else {
                amrex::Abort("Unknown IC_Type");
            }
        } else {
            set_zeta_averaged_down(lev);
        }
        vec_zeta[lev]->FillBoundary(geom[lev].periodicity());
    } else {
        // If our level is higher than the high resolution grid or initialization
        // is analytic, interpolate from level below. Otherwise, copy over the bathymetry
        // data that has been averaged down
        if (lev > hires_init_level) {
            Real dummy_time = zero;
            FillCoarsePatch(lev,dummy_time,vec_zeta[lev].get(), vec_zeta[lev-1].get(),BCVars::cons_bc);
        } else {
            set_zeta_averaged_down(lev);
            vec_zeta[lev]->FillBoundary(geom[lev].periodicity());
        }
    }
    set_zeta_average(lev);
}
 
/**
 * @param[in   ] lev    level to operate on
 */
void
REMORA::set_bathymetry (int lev)
{
    BL_PROFILE("REMORA::bathymetry()");
    // Only set bathymetry on level 0, and interpolate for finer levels
    if (lev==0) {
        // If grid data is not defined on a level > 0 (negative level) then
        // initialize from low-resolution grid normally. Otherwise use high-resolution
        // grid data averaged down to level 0
        if (hires_grid_level < 0) {
            if (solverChoice.ic_type == IC_Type::analytic) {
                prob->init_analytic_bathymetry(lev, geom[lev], solverChoice, *this, *vec_h[lev]);
            } else if (solverChoice.ic_type == IC_Type::netcdf) {
#ifdef REMORA_USE_NETCDF
                amrex::Print() << "Calling init_bathymetry_from_netcdf " << std::endl;
                init_bathymetry_from_netcdf(lev);
                amrex::Print() << "Bathymetry loaded from netcdf file \n " << std::endl;
                amrex::Print() << "Calling init_grid_vars_from_netcdf " << std::endl;
                init_grid_vars_from_netcdf(lev);
                amrex::Print() << "Grid variables loaded from netcdf file \n " << std::endl;
#endif
            } else {
                amrex::Abort("Unknown IC_Type");
            }
        } else {
            set_bathymetry_averaged_down(lev);
            // Only the netcdf path fills vec_pm/pn_full_domain; with analytic initialization
            // init_bathymetry_full_domain_from_analytic fills h alone, and set_grid_scale
            // below derives pm/pn from the geometry.
            if (solverChoice.ic_type == IC_Type::netcdf) {
                set_grid_vars_averaged_down(lev);
            }
        }
        // Need FillBoundary to fill at grid-grid boundaries, and EnforcePeriodicity
        // to make sure ghost cells in the domain corners are consistent.
        vec_h[lev]->FillBoundary(geom[lev].periodicity());
        vec_h[lev]->EnforcePeriodicity(geom[lev].periodicity());
    } else {
        // If our level is higher than the high resolution grid or initialization
        // is analytic, interpolate from level below. Otherwise, copy over the bathymetry
        // data that has been averaged down
        if (lev > hires_grid_level) {
            Real dummy_time = zero;
            FillCoarsePatch(lev,dummy_time,vec_h[lev].get(), vec_h[lev-1].get(),
                    foextrap_periodic_bc(),BdyVars::null,0,false);
            FillCoarsePatch(lev,dummy_time,vec_h[lev].get(), vec_h[lev-1].get(),
                    foextrap_periodic_bc(),BdyVars::null,1,false);
        } else {
            set_bathymetry_averaged_down(lev);
            vec_h[lev]->FillBoundary(geom[lev].periodicity());
            vec_h[lev]->EnforcePeriodicity(geom[lev].periodicity());
        }
    }
    set_grid_scale(lev);
}
 
/**
 * @param[in   ] lev   level to operate on
 */
void
REMORA::set_bathymetry_averaged_down (int lev) {
    Real dummy_time = zero;
    // Note: don't understand why the grow vector args aren't vec_h and then vec_h_full_domain
    ParallelCopy(*vec_h[lev].get(), *vec_h_full_domain[lev].get(), 0, 0, 1,vec_h_full_domain[lev]->nGrowVect(),vec_h[lev]->nGrowVect());
    ParallelCopy(*vec_h[lev].get(), *vec_h_full_domain[lev].get(), 0, 1, 1,vec_h_full_domain[lev]->nGrowVect(),vec_h[lev]->nGrowVect());
    FillPatch(lev,dummy_time,*vec_h[lev],GetVecOfPtrs(vec_h),
            foextrap_periodic_bc(),
            BdyVars::null,0,false,false,1);
    FillPatch(lev,dummy_time,*vec_h[lev],GetVecOfPtrs(vec_h),
            foextrap_periodic_bc(),
            BdyVars::null,1,false,false,1);
}
 
/**
 * @param[in   ] lev   level to operate on
 */
void
REMORA::set_grid_vars_averaged_down (int lev) {
    Real dummy_time = zero;
    ParallelCopy(*vec_pm[lev].get(), *vec_pm_full_domain[lev].get(), 0, 0, 1,
            vec_pm_full_domain[lev]->nGrowVect(),vec_pm[lev]->nGrowVect());
    ParallelCopy(*vec_pn[lev].get(), *vec_pn_full_domain[lev].get(), 0, 0, 1,
            vec_pn_full_domain[lev]->nGrowVect(),vec_pn[lev]->nGrowVect());
    FillPatch(lev,dummy_time,*vec_pm[lev],GetVecOfPtrs(vec_pm),
            foextrap_periodic_bc(),
            BdyVars::null,0,false);
    FillPatch(lev,dummy_time,*vec_pn[lev],GetVecOfPtrs(vec_pn),
            foextrap_periodic_bc(),
            BdyVars::null,0,false);
}
 
/**
 * @param[in   ] lev   level to operate on
 */
void
REMORA::set_zeta_averaged_down (int lev) {
    ParallelCopy(*vec_zeta[lev].get(), *vec_zeta_full_domain[lev].get(), 0, 0, 1,
            vec_zeta_full_domain[lev]->nGrowVect(),vec_zeta[lev]->nGrowVect());
    FillPatch(lev, t_new[lev], *vec_zeta[lev], GetVecOfPtrs(vec_zeta), zeta_bc(), bdy_zeta(),
                  0, false,false,0,0,zero,*vec_zeta[lev]);
}
 
/**
 * @param[in   ] lev   level to operate on
 */
void
REMORA::set_init_data_averaged_down (int lev) {
    ParallelCopy(*cons_new[lev], *vec_cons_full_domain[lev], 0, 0, ncons,
            vec_cons_full_domain[lev]->nGrowVect(),cons_new[lev]->nGrowVect());
    ParallelCopy(*xvel_new[lev], *vec_xvel_full_domain[lev], 0, 0, 1,
            vec_xvel_full_domain[lev]->nGrowVect(),xvel_new[lev]->nGrowVect());
    ParallelCopy(*yvel_new[lev], *vec_yvel_full_domain[lev], 0, 0, 1,
            vec_yvel_full_domain[lev]->nGrowVect(),yvel_new[lev]->nGrowVect());
 
    FillPatch(lev, t_new[lev], *cons_new[lev], cons_new, BCVars::cons_bc, BdyVars::t, 0, true, false,0,0,zero,*cons_new[lev]);
    FillPatch(lev, t_new[lev], *xvel_new[lev], xvel_new, xvel_bc(), BdyVars::u, 0, true, false,0,0,zero,*xvel_new[lev]);
    FillPatch(lev, t_new[lev], *yvel_new[lev], yvel_new, yvel_bc(), BdyVars::v, 0, true, false,0,0,zero,*yvel_new[lev]);
}
 
/**
 * @param[in   ] lev    level to operate on
 */
void
REMORA::set_coriolis(int lev) {
    BL_PROFILE("REMORA::set_coriolis()");
    if (solverChoice.use_coriolis) {
        if (solverChoice.coriolis_type == Cor_Type::analytic) {
            prob->init_analytic_coriolis(lev, geom[lev], solverChoice, *this, *vec_fcor[lev]);
        } else if (solverChoice.coriolis_type == Cor_Type::beta_plane) {
            init_beta_plane_coriolis(lev);
#ifdef REMORA_USE_NETCDF
        } else if (solverChoice.coriolis_type == Cor_Type::netcdf) {
            if (lev == 0) {
                amrex::Print() << "Calling init_coriolis_from_netcdf " << std::endl;
                init_coriolis_from_netcdf(lev);
                amrex::Print() << "Coriolis loaded from netcdf file \n" << std::endl;
            } else {
                Real dummy_time = zero;
                FillCoarsePatch(lev,dummy_time,vec_fcor[lev].get(), vec_fcor[lev-1].get(),BCVars::cons_bc);
            }
#endif
        } else {
            Abort("Don't know this coriolis_type!");
        }
 
        Real time = zero;
        FillPatch(lev, time, *vec_fcor[lev], GetVecOfPtrs(vec_fcor), foextrap_bc());
        vec_fcor[lev]->EnforcePeriodicity(geom[lev].periodicity());
    }
}
 
void
REMORA::init_set_vmix(int lev) {
    BL_PROFILE("REMORA::init_set_vmix()");
    if (solverChoice.vert_mixing_type == VertMixingType::analytic) {
        set_analytic_vmix(lev);
    } else if (solverChoice.vert_mixing_type == VertMixingType::GLS) {
        init_gls_vmix(lev, solverChoice);
        // The GLS initialization just sets the multifab to a value, so there's
        // no need to call FillPatch here
    } else {
        Abort("Don't know this vertical mixing type");
    }
}
 
/**
 * @param[in   ] lev    level to operate on
 */
void
REMORA::set_analytic_vmix(int lev) {
    BL_PROFILE("REMORA::set_analytic_vmix()");
    Real time = zero;
    vec_Akv[lev]->setVal(solverChoice.Akv_bak);
    for (int n = 0; n < NAT; n++) {
        vec_Akt[lev]->setVal(solverChoice.Akt_bak[n], n, 1);
    }
    prob->init_analytic_vmix(lev, geom[lev], solverChoice, *this,*vec_Akv[lev], *vec_Akt[lev]);
    FillPatch(lev, time, *vec_Akv[lev], GetVecOfPtrs(vec_Akv), zvel_bc(), BdyVars::null,0,true,false);
    for (int n = 0; n < NAT; n++) {
        FillPatch(lev, time, *vec_Akt[lev], GetVecOfPtrs(vec_Akt), zvel_bc(), BdyVars::null,n,false,false);
    }
}
 
/**
 * Initialize the land-sea mask on this level.
 *
 * Mirrors set_bathymetry: the mask is specified once, on level 0 or at hires_grid_level, and
 * every other level derived from it, so the levels cannot disagree about the coastline.
 *
 * @param[in   ] lev    level to operate on
 */
void
REMORA::set_masks (int lev)
{
    // Ahead of the mask_type == none return as well: that branch still fills the masks, and
    // AverageDownTo still reads them, so its cached copies go stale here too.
    clear_avgdown_masks(lev);
 
    if (solverChoice.mask_type == MaskType::none) {
        fill_3d_masks(lev);
        return;
    }
 
    if (lev == 0) {
        // If grid data is not defined on a level > 0 (negative level) then initialize from
        // the low-resolution grid normally. Otherwise use high-resolution grid data
        // coarsened down to level 0.
        if (hires_grid_level < 0) {
            if (solverChoice.mask_type == MaskType::analytic) {
                prob->init_analytic_masks(lev,geom[lev], solverChoice, *this, *vec_mskr[lev]);
                // The analytic hook writes each grid's own cells only, so this is what makes
                // the mask agree across grid-grid and periodic boundaries.
                vec_mskr[lev]->FillBoundary(geom[lev].periodicity());
                update_nodal_masks(lev);
            } else if (solverChoice.mask_type == MaskType::netcdf) {
#ifdef REMORA_USE_NETCDF
                amrex::Print() << "Calling init_masks_from_netcdf level " << lev << std::endl;
                init_masks_from_netcdf(lev);
                amrex::Print() << "Masks loaded from netcdf file \n " << std::endl;
#endif
            }
        } else {
            set_masks_averaged_down(lev);
        }
    } else {
        // If our level is higher than the high resolution grid, interpolate from the level
        // below. Otherwise, copy over the mask that has been coarsened down.
        if (lev > hires_grid_level) {
            Real dummy_time = zero;
            FillCoarsePatchPC(lev, dummy_time, vec_mskr[lev].get(), vec_mskr[lev-1].get(),
                    foextrap_bc());
            update_nodal_masks(lev);
        } else {
            set_masks_averaged_down(lev);
        }
    }
    fill_3d_masks(lev);
}
 
/**
 * @param[in   ] lev   level to operate on
 */
void
REMORA::set_masks_averaged_down (int lev) {
    ParallelCopy(*vec_mskr[lev].get(), *vec_mskr_full_domain[lev].get(), 0, 0, 1,
            vec_mskr_full_domain[lev]->nGrowVect(),vec_mskr[lev]->nGrowVect());
    // Not a FillPatch, unlike the bathymetry analogue: its interpolation from the coarser
    // level is not piecewise constant, so it would put fractional values in a mask the rest
    // of the code compares against 0 and 1 exactly.
    vec_mskr[lev]->FillBoundary(geom[lev].periodicity());
    update_nodal_masks(lev);
}
 
/**
 * Coarsen the full-domain rho-mask from crse_lev+1 onto crse_lev, grow cells included, so a
 * coarse cell is land only if every one of its fine cells is land.
 *
 * average_down_with_grow_cells cannot be used: an arithmetic mean over a partly wet group of
 * fine cells gives a fractional value, and the mask has to stay exactly 0 or 1. Taking the
 * cell as wet also means the coarse level never declares land where the fine grid found water.
 *
 * @param[in   ] crse_lev   level to coarsen onto
 */
void
REMORA::coarsen_masks_with_grow_cells (int crse_lev)
{
    auto const& crsema = vec_mskr_full_domain[crse_lev]->arrays();
    auto const& finema = vec_mskr_full_domain[crse_lev+1]->const_arrays();
    auto ratio = refRatio(crse_lev);
    // As in average_down_with_grow_cells, but cell-centered, so no index-type correction.
    auto nghost_crse = cum_ref_ratios[crse_lev];
    ParallelFor(*vec_mskr_full_domain[crse_lev], nghost_crse, 1,
            [=] AMREX_GPU_DEVICE (int box_no, int i, int j, int k, int n) noexcept
    {
        const int ii = i * ratio[0];
        const int jj = j * ratio[1];
        Real wet = zero;
        for (int jref = 0; jref < ratio[1]; ++jref) {
            for (int iref = 0; iref < ratio[0]; ++iref) {
                wet += amrex::min(Real(1.0), finema[box_no](ii+iref, jj+jref, k, n));
            }
        }
        crsema[box_no](i,j,k,n) = (wet > zero) ? one : zero;
    });
    Gpu::streamSynchronize();
}
 
 
/**
 * Check the land-sea masks for what the rest of the code relies on. Only runs when
 * remora.check_mask_consistency is set; remora.mask_consistency picks abort or warn.
 *
 * Per level, that the masks hold only the values they are meant to and that no water cell has
 * a non-positive depth. Per level pair, over the region the finer level covers, that no coarse
 * water point sits over fine points that are all land.
 */
void
REMORA::check_mask_consistency ()
{
    BL_PROFILE("REMORA::check_mask_consistency()");
    if (!solverChoice.do_check_mask_consistency || solverChoice.mask_type == MaskType::none) {
        return;
    }
 
    Long nbad_val = 0, nbad_h = 0, ndry_r = 0, ndry_u = 0, ndry_v = 0, nmissed = 0;
 
    // Per-level checks. Mask values matter because the plotfile writer decides what to blank
    // by comparing them against 0 exactly, so a fractional mask stops masking; a water cell
    // with h <= 0 matters because stretch_transform divides by hc + h, giving quiet garbage
    // rather than a crash.
    for (int lev = 0; lev <= finest_level; ++lev)
    {
        ReduceOps<ReduceOpSum, ReduceOpSum> reduce_op;
        ReduceData<Long, Long> reduce_data(reduce_op);
        using ReduceTuple = typename decltype(reduce_data)::Type;
 
        for ( MFIter mfi(*vec_mskr[lev], TilingIfNotGPU()); mfi.isValid(); ++mfi )
        {
            Array4<const Real> const& mskr = vec_mskr[lev]->const_array(mfi);
            Array4<const Real> const& msku = vec_msku[lev]->const_array(mfi);
            Array4<const Real> const& mskv = vec_mskv[lev]->const_array(mfi);
            Array4<const Real> const& mskp = vec_mskp[lev]->const_array(mfi);
            Array4<const Real> const& h    = vec_h[lev]->const_array(mfi);
 
            Box bx = mfi.tilebox(); bx.makeSlab(2,0);
 
            reduce_op.eval(bx, reduce_data, [=] AMREX_GPU_DEVICE (int i, int j, int k)
                -> ReduceTuple
            {
                auto is_01 = [] (Real v) {
                    return v == Real(0.0) || v == Real(1.0);
                };
                const bool bad = !is_01(mskr(i,j,k)) || !is_01(msku(i,j,k)) ||
                                 !is_01(mskv(i,j,k)) ||
                                 !(is_01(mskp(i,j,k)) || mskp(i,j,k) == Real(2.0));
                const bool bad_h = (mskr(i,j,k) > Real(0.5)) && (h(i,j,k) <= Real(0.0));
                return {static_cast<Long>(bad), static_cast<Long>(bad_h)};
            });
        }
        ReduceTuple hv = reduce_data.value(reduce_op);
        nbad_val += amrex::get<0>(hv);
        nbad_h   += amrex::get<1>(hv);
    }
 
    // Level-pair checks, over the region the finer level covers. A coarse water point over
    // nothing but land would leave the average-down nothing to divide by. Coarsening makes
    // that unreachable for cell centers -- a wet coarse cell is wet because some fine cell in
    // its own block is -- but not for faces: a coarse u-face is open whenever both its cells
    // are wet, only the fine faces in its own plane count, and the wet fine cells that made
    // those coarse cells wet may all lie elsewhere in their blocks.
    for (int crse_lev = 0; crse_lev < finest_level; ++crse_lev)
    {
        const int flev = crse_lev + 1;
        const IntVect ratio = refRatio(crse_lev);
        const BoxArray cba = amrex::coarsen(vec_mskr[flev]->boxArray(), ratio);
        const DistributionMapping& dmf = vec_mskr[flev]->DistributionMap();
 
        // Sentinel, so an incomplete copy shows up as itself rather than as a coarse land
        // point that the checks below would quietly pass over.
        MultiFab cmskr(cba, dmf, 1, 0);
        MultiFab cmsku(amrex::convert(cba, IntVect(1,0,0)), dmf, 1, 0);
        MultiFab cmskv(amrex::convert(cba, IntVect(0,1,0)), dmf, 1, 0);
        cmskr.setVal(-one); cmsku.setVal(-one); cmskv.setVal(-one);
        cmskr.ParallelCopy(*vec_mskr[crse_lev], 0, 0, 1);
        cmsku.ParallelCopy(*vec_msku[crse_lev], 0, 0, 1);
        cmskv.ParallelCopy(*vec_mskv[crse_lev], 0, 0, 1);
 
        ReduceOps<ReduceOpSum, ReduceOpSum, ReduceOpSum, ReduceOpSum> reduce_op;
        ReduceData<Long, Long, Long, Long> reduce_data(reduce_op);
        using ReduceTuple = typename decltype(reduce_data)::Type;
 
        for ( MFIter mfi(cmskr, TilingIfNotGPU()); mfi.isValid(); ++mfi )
        {
            Array4<const Real> const& cr = cmskr.const_array(mfi);
            Array4<const Real> const& cu = cmsku.const_array(mfi);
            Array4<const Real> const& cv = cmskv.const_array(mfi);
            Array4<const Real> const& fr = vec_mskr[flev]->const_array(mfi);
            Array4<const Real> const& fu = vec_msku[flev]->const_array(mfi);
            Array4<const Real> const& fv = vec_mskv[flev]->const_array(mfi);
 
            Box bx = mfi.tilebox(); bx.makeSlab(2,0);
            const int rx = ratio[0];
            const int ry = ratio[1];
 
            // Three passes: a cell-centered box stops at hi, so it would miss the high-side
            // face that average_down_masked does iterate. nodaltilebox partitions the nodal
            // range; faces shared between boxes are still counted twice, which can only
            // inflate a diagnostic.
            const Long lzero = 0;
 
            // Cell centers: the whole rx by ry block.
            reduce_op.eval(bx, reduce_data, [=] AMREX_GPU_DEVICE (int i, int j, int k)
                -> ReduceTuple
            {
                const int ii = i * rx;
                const int jj = j * ry;
 
                Real wet_r = zero;
                for (int jref = 0; jref < ry; ++jref) {
                    for (int iref = 0; iref < rx; ++iref) {
                        wet_r += amrex::min(Real(1.0), fr(ii+iref, jj+jref, k));
                    }
                }
 
                const bool missed = cr(i,j,k) < zero;
                return {static_cast<Long>(missed),
                        static_cast<Long>(!missed && cr(i,j,k) > Real(0.5) && wet_r == zero),
                        lzero, lzero};
            });
 
            // u-faces: only the fine faces in the coarse face's plane. Tests the sentinel too,
            // so an incomplete copy reports itself instead of failing the > 0.5 test.
            Box ubx = mfi.nodaltilebox(0); ubx.makeSlab(2,0);
            reduce_op.eval(ubx, reduce_data, [=] AMREX_GPU_DEVICE (int i, int j, int k)
                -> ReduceTuple
            {
                const int ii = i * rx;
                const int jj = j * ry;
 
                Real wet_u = zero;
                for (int jref = 0; jref < ry; ++jref) {
                    wet_u += amrex::min(Real(1.0), fu(ii, jj+jref, k));
                }
 
                const bool missed = cu(i,j,k) < zero;
                return {static_cast<Long>(missed), lzero,
                        static_cast<Long>(!missed && cu(i,j,k) > Real(0.5) && wet_u == zero),
                        lzero};
            });
 
            // v-faces, likewise.
            Box vbx = mfi.nodaltilebox(1); vbx.makeSlab(2,0);
            reduce_op.eval(vbx, reduce_data, [=] AMREX_GPU_DEVICE (int i, int j, int k)
                -> ReduceTuple
            {
                const int ii = i * rx;
                const int jj = j * ry;
 
                Real wet_v = zero;
                for (int iref = 0; iref < rx; ++iref) {
                    wet_v += amrex::min(Real(1.0), fv(ii+iref, jj, k));
                }
 
                const bool missed = cv(i,j,k) < zero;
                return {static_cast<Long>(missed), lzero, lzero,
                        static_cast<Long>(!missed && cv(i,j,k) > Real(0.5) && wet_v == zero)};
            });
        }
        ReduceTuple hv = reduce_data.value(reduce_op);
        nmissed += amrex::get<0>(hv);
        ndry_r  += amrex::get<1>(hv);
        ndry_u  += amrex::get<2>(hv);
        ndry_v  += amrex::get<3>(hv);
    }
 
    ParallelDescriptor::ReduceLongSum(nbad_val);
    ParallelDescriptor::ReduceLongSum(nbad_h);
    ParallelDescriptor::ReduceLongSum(nmissed);
    ParallelDescriptor::ReduceLongSum(ndry_r);
    ParallelDescriptor::ReduceLongSum(ndry_u);
    ParallelDescriptor::ReduceLongSum(ndry_v);
 
    if (nbad_val == 0 && nbad_h == 0 && nmissed == 0 &&
        ndry_r == 0 && ndry_u == 0 && ndry_v == 0) {
        if (verbose > 0) {
            amrex::Print() << "Land-sea masks are consistent across " << finest_level+1
                           << " level(s)" << std::endl;
        }
        return;
    }
 
    std::string msg = "Land-sea mask problems:";
    if (nbad_val > 0) {
        msg += "\n  " + std::to_string(nbad_val) + " point(s) where a mask is neither 0 nor 1"
               " (psi may also be 2). The plotfile writer decides what to blank by comparing"
               " masks against 0 exactly, so a fractional mask silently stops masking.";
    }
    if (nbad_h > 0) {
        msg += "\n  " + std::to_string(nbad_h) + " water point(s) with depth <= 0."
               " stretch_transform divides by hc + h, so this is quiet garbage rather than a"
               " crash.";
    }
    if (nmissed > 0) {
        msg += "\n  " + std::to_string(nmissed) + " refined point(s) with no coarse point"
               " beneath them, which should be impossible under proper nesting.";
    }
    if (ndry_r > 0) {
        msg += "\n  " + std::to_string(ndry_r) + " coarse water cell(s) with only land"
               " beneath them.";
    }
    if (ndry_u > 0 || ndry_v > 0) {
        msg += "\n  " + std::to_string(ndry_u) + " coarse u-face(s) and " +
               std::to_string(ndry_v) + " v-face(s) that are open with no open fine face"
               " beneath them. The coarse grid cannot see a barrier the fine grid resolves."
               " Closing the coarse face would contradict the coarsening rule, so move the"
               " refined grids off it or coarsen the mask by hand.";
    }
    if (ndry_r > 0 || ndry_u > 0 || ndry_v > 0) {
        msg += "\nThe two-way average-down divides by the number of wet fine points, so these"
               " have no value to take.";
    }
    msg += "\nSet remora.mask_consistency = warn to continue anyway.";
 
    if (solverChoice.mask_consistency == MaskConsistency::abort) {
        amrex::Abort(msg);
    } else {
        amrex::Print() << "WARNING: " << msg << std::endl;
    }
}
 
/**
 * Coarsen the full-domain bathymetry from crse_lev+1 onto crse_lev, grow cells included,
 * weighted by the land/sea mask.
 *
 * Averaging every fine cell would mix in whatever the grid file holds under land, which on a
 * ROMS grid is a fill value with no physical meaning. A wet coarse cell should take the depth
 * of the water under it. See REMORA_MaskedAverageDown.H for why the arithmetic is written the
 * way it is.
 *
 * @param[in   ] crse_lev   level to coarsen onto
 */
void
REMORA::coarsen_bathymetry_with_grow_cells (int crse_lev)
{
    // The mask is indexed by the same local box number as the bathymetry, so the two have to
    // be distributed alike. full_domain_dmap is what makes that true; assert it here rather
    // than read another rank's box.
    AMREX_ALWAYS_ASSERT(vec_mskr_full_domain[crse_lev+1]->DistributionMap() ==
                        vec_h_full_domain[crse_lev+1]->DistributionMap());
 
    auto const& crsema = vec_h_full_domain[crse_lev]->arrays();
    auto const& finema = vec_h_full_domain[crse_lev+1]->const_arrays();
    auto const& fmskma = vec_mskr_full_domain[crse_lev+1]->const_arrays();
    auto ratio = refRatio(crse_lev);
    auto nghost_crse = cum_ref_ratios[crse_lev];
    const int ncomp = vec_h_full_domain[crse_lev]->nComp();
    ParallelFor(*vec_h_full_domain[crse_lev], nghost_crse, ncomp,
            [=] AMREX_GPU_DEVICE (int box_no, int i, int j, int k, int n) noexcept
    {
        const int ii = i * ratio[0];
        const int jj = j * ratio[1];
        Real num = zero, den = zero, sum_all = zero;
        for (int jref = 0; jref < ratio[1]; ++jref) {
            for (int iref = 0; iref < ratio[0]; ++iref) {
                const Real hf = finema[box_no](ii+iref, jj+jref, k, n);
                const Real m = amrex::min(Real(1.0), fmskma[box_no](ii+iref, jj+jref, k));
                num += hf * m;
                den += m;
                sum_all += hf;
            }
        }
        // All-land: no water to average, but h still has to hold something, so fall back to
        // the plain mean. That also makes an all-wet or all-land group reproduce
        // average_down_with_grow_cells bit for bit. This is where the bathymetry parts company
        // with avgdown_masked, which multiplies by the coarse mask and so leaves an all-land
        // point at zero: a free surface under land need not hold anything, a depth does.
        crsema[box_no](i,j,k,n) = (den > zero)
                                ? num * (one/den)
                                : sum_all * (one/Real(ratio[0]*ratio[1]));
    });
    Gpu::streamSynchronize();
}
 
/**
 * @param[in   ] lev    level to operate on
 */
void
REMORA::set_hmixcoef(int lev)
{
    BL_PROFILE("REMORA::set_hmixcoef()");
 
    // Optional AMR scaling: decrease coefficients on refined levels linearly
    // with grid size (i.e., proportional to sqrt(cell area)). For a horizontal
    // refinement ratio rx x ry, the effective scale factor is 1/sqrt(rx*ry).
    Real lev_scale = one;
    if ((solverChoice.scaled_to_grid_amr_scaling == ScaledToGridAMRScaling::linear) && (lev > 0)) {
        Real rf = one;
        for (int l = 0; l < lev; ++l) {
            rf *= std::sqrt(static_cast<Real>(ref_ratio[l][0]) * static_cast<Real>(ref_ratio[l][1]));
        }
        lev_scale = one / rf;
    }
 
    if (solverChoice.horiz_mixing_type == HorizMixingType::analytic) {
        prob->init_analytic_hmix(lev, geom[lev], solverChoice,
                                 *this, *vec_visc2_p[lev], *vec_visc2_r[lev], *vec_diff2[lev]);
 
    } else if (solverChoice.horiz_mixing_type == HorizMixingType::constant) {
        vec_visc2_p[lev]->setVal(solverChoice.visc2 * lev_scale);
        vec_visc2_r[lev]->setVal(solverChoice.visc2 * lev_scale);
        for (int n = 0; n < ncons; n++) {
            vec_diff2[lev]->setVal(solverChoice.tnu2[n] * lev_scale, n, 1);
        }
 
    // Scale harmonic viscosity and diffusivity by the grid size as ROMS
    // does in Utility/ini_hmixcoef.F. Intended for curvilinear grids.
    //
    // Define the ROMS grid factor (grdscl):
    //     G(i,j) = sqrt( 1 / (pm(i,j) * pn(i,j)) )
    //            = sqrt(cell area)
    //     Gmax   = max over grid of G(i,j)
    //
    // Then horizontal harmonic mixing coefficients are scaled as:
    //     nu(i,j)       = nu0    * G(i,j) / Gmax
    //     kappa_n(i,j)  = kappa0 * G(i,j) / Gmax
    //
    // where:
    //     nu0     = solverChoice.visc2
    //     kappa0  = solverChoice.tnu2[n]
    //
    // This makes mixing strongest where grid spacing is largest.
    //
    // NOTE: The normalization (Gmax) is computed over the entire grid (ignoring masks).
    // Therefore, if the largest cell area occurs over land, the maximum over *wet* cells
    // (or in masked output files) may be smaller than the user-specified value.
 
    } else if (solverChoice.horiz_mixing_type == HorizMixingType::scaled_to_grid) {
 
        // ------------------------------------------------------------
        // Step 1: Compute grdmax over entire grid
        // ------------------------------------------------------------
        vec_visc2_r[lev]->setVal(solverChoice.visc2);
        vec_visc2_p[lev]->setVal(solverChoice.visc2);
        for (int n = 0; n < ncons; n++) {
            vec_diff2[lev]->setVal(solverChoice.tnu2[n], n, 1);
        }
 
        // NOTE: This must be GPU-safe. Do not dereference MultiFab data on host.
        // Force the reduction to run in the GPU launch region if GPUs are enabled.
        // (If the launch region is disabled at runtime, ReduceMax may fall back to
        // a host path that can try to read device-only data.)
        amrex::Gpu::LaunchSafeGuard lsg(true);
        Real denom_min = amrex::ReduceMin(*vec_pm[lev], *vec_pn[lev], 0,
            [=] AMREX_GPU_HOST_DEVICE (Box const& bx,
                                      Array4<Real const> const& pm,
                                      Array4<Real const> const& pn) -> Real
            {
                Real local_min = bogus_large_value;
                amrex::Loop(bx, [=,&local_min] (int i, int j, int) noexcept
                {
                    local_min = amrex::min(local_min, pm(i,j,0) * pn(i,j,0));
                });
                return local_min;
            });
 
        ParallelDescriptor::ReduceRealMin(denom_min);
        if (denom_min <= zero) {
            Abort("scaled_to_grid: found non-positive pm*pn (grid metrics must be > 0)");
        }
 
        Real grdmax = amrex::ReduceMax(*vec_pm[lev], *vec_pn[lev], 0,
            [=] AMREX_GPU_HOST_DEVICE (Box const& bx,
                                      Array4<Real const> const& pm,
                                      Array4<Real const> const& pn) -> Real
            {
                Real local_max = zero;
                amrex::Loop(bx, [=,&local_max] (int i, int j, int) noexcept
                {
                    Real denom = pm(i,j,0) * pn(i,j,0);
                    if (denom > zero) {
                        Real G = std::sqrt(one / denom);
                        local_max = amrex::max(local_max, G);
                    }
                });
                return local_max;
            });
 
        ParallelDescriptor::ReduceRealMax(grdmax);
        if (grdmax <= zero) {
            Abort("scaled_to_grid: grdmax <= 0");
        }
 
        // Optional AMR scaling: decrease coefficients on refined levels linearly
        // with grid size (i.e., proportional to sqrt(cell area)). For a horizontal
        // refinement ratio rx x ry, the effective scale factor is 1/sqrt(rx*ry).
        lev_scale = one;
        if ((solverChoice.scaled_to_grid_amr_scaling == ScaledToGridAMRScaling::linear) && (lev > 0)) {
            Real rf = one;
            for (int l = 0; l < lev; ++l) {
                rf *= std::sqrt(static_cast<Real>(ref_ratio[l][0]) * static_cast<Real>(ref_ratio[l][1]));
            }
            lev_scale = one / rf;
        }
 
        Real visc0 = solverChoice.visc2 * lev_scale;
        Real cff   = visc0 / grdmax;
 
        // ------------------------------------------------------------
        // Step 2: Set rho coefficients everywhere
        // ------------------------------------------------------------
        amrex::Gpu::DeviceVector<Real> diff0_d(ncons);
        amrex::Gpu::copy(amrex::Gpu::hostToDevice,
                         solverChoice.tnu2.begin(), solverChoice.tnu2.begin() + ncons,
                         diff0_d.begin());
        Real const* diff0_ptr = diff0_d.data();
 
        for (MFIter mfi(*vec_visc2_r[lev], TilingIfNotGPU()); mfi.isValid(); ++mfi)
        {
            const Box& bx = mfi.validbox();
            auto pm    = vec_pm[lev]->const_array(mfi);
            auto pn    = vec_pn[lev]->const_array(mfi);
            auto visc2_r = vec_visc2_r[lev]->array(mfi);
            auto diff2   = vec_diff2[lev]->array(mfi);
 
            int ncons_local = ncons;
            ParallelFor(makeSlab(bx,2,0), [=] AMREX_GPU_DEVICE (int i, int j, int) noexcept
            {
                Real denom  = pm(i,j,0) * pn(i,j,0);
                Real grdscl = (denom > zero) ? std::sqrt(one / denom) : zero;
                visc2_r(i,j,0) = cff * grdscl;
 
                for (int n = 0; n < ncons_local; n++) {
                    diff2(i,j,0,n) = ((diff0_ptr[n] * lev_scale) / grdmax) * grdscl;
                }
            });
        }
 
        // Fill ghost cells for rho coefficients BEFORE psi averaging
        Real time = zero;
        FillPatch(lev, time, *vec_visc2_r[lev], GetVecOfPtrs(vec_visc2_r), foextrap_periodic_bc());
 
        // ------------------------------------------------------------
        // Step 3: Psi coefficients = average of 4 surrounding rho
        // ------------------------------------------------------------
        for (MFIter mfi(*vec_visc2_p[lev], TilingIfNotGPU()); mfi.isValid(); ++mfi)
        {
            const Box& bx = mfi.validbox();
            auto visc2_p = vec_visc2_p[lev]->array(mfi);
            auto visc2_r = vec_visc2_r[lev]->const_array(mfi);
 
            ParallelFor(makeSlab(bx,2,0), [=] AMREX_GPU_DEVICE (int i, int j, int) noexcept
            {
                visc2_p(i,j,0) = fourth * (
                    visc2_r(i-1,j-1,0) +
                    visc2_r(i  ,j-1,0) +
                    visc2_r(i-1,j  ,0) +
                    visc2_r(i  ,j  ,0)
                );
            });
        }
 
        FillPatch(lev, time, *vec_visc2_p[lev], GetVecOfPtrs(vec_visc2_p), foextrap_periodic_bc());
 
        // Diagnostics
        // NOTE: coefficients are computed everywhere (including land). Output routines may later
        // mask land points (e.g., to FillValue in NetCDF/plotfiles), and analysis tools may
        // additionally apply mask_rho (setting land to 0). Report both conventions.
        //
        // Global (MPI-reduced) extrema over all valid cells (no ghost).
        Real visc_min_all = vec_visc2_r[lev]->min(0,0,false);
        Real visc_max_all = vec_visc2_r[lev]->max(0,0,false);
 
        // Global extrema over *wet* rho points only, k=0.
        amrex::Gpu::LaunchSafeGuard lsg_diag(true);
        Real visc_min_wet = amrex::ReduceMin(*vec_visc2_r[lev], *vec_mskr[lev], 0,
            [=] AMREX_GPU_HOST_DEVICE (Box const& bx,
                                      Array4<Real const> const& visc2,
                                      Array4<Real const> const& mskr) -> Real
            {
                Real local_min = bogus_large_value;
                amrex::Loop(bx, [=,&local_min] (int i, int j, int) noexcept
                {
                    if (mskr(i,j,0) > zero) {
                        local_min = amrex::min(local_min, visc2(i,j,0));
                    }
                });
                return local_min;
            });
        ParallelDescriptor::ReduceRealMin(visc_min_wet);
 
        Real visc_max_wet = amrex::ReduceMax(*vec_visc2_r[lev], *vec_mskr[lev], 0,
            [=] AMREX_GPU_HOST_DEVICE (Box const& bx,
                                      Array4<Real const> const& visc2,
                                      Array4<Real const> const& mskr) -> Real
            {
                Real local_max = -bogus_large_value;
                amrex::Loop(bx, [=,&local_max] (int i, int j, int) noexcept
                {
                    if (mskr(i,j,0) > zero) {
                        local_max = amrex::max(local_max, visc2(i,j,0));
                    }
                });
                return local_max;
            });
        ParallelDescriptor::ReduceRealMax(visc_max_wet);
 
        // Mimic "apply mask_rho" convention (dry -> 0).
        Real visc_min_mask0 = amrex::ReduceMin(*vec_visc2_r[lev], *vec_mskr[lev], 0,
            [=] AMREX_GPU_HOST_DEVICE (Box const& bx,
                                      Array4<Real const> const& visc2,
                                      Array4<Real const> const& mskr) -> Real
            {
                Real local_min = bogus_large_value;
                amrex::Loop(bx, [=,&local_min] (int i, int j, int) noexcept
                {
                    const Real v = (mskr(i,j,0) > zero) ? visc2(i,j,0) : zero;
                    local_min = amrex::min(local_min, v);
                });
                return local_min;
            });
        ParallelDescriptor::ReduceRealMin(visc_min_mask0);
 
        Real visc_max_mask0 = amrex::ReduceMax(*vec_visc2_r[lev], *vec_mskr[lev], 0,
            [=] AMREX_GPU_HOST_DEVICE (Box const& bx,
                                      Array4<Real const> const& visc2,
                                      Array4<Real const> const& mskr) -> Real
            {
                Real local_max = -bogus_large_value;
                amrex::Loop(bx, [=,&local_max] (int i, int j, int) noexcept
                {
                    const Real v = (mskr(i,j,0) > zero) ? visc2(i,j,0) : zero;
                    local_max = amrex::max(local_max, v);
                });
                return local_max;
            });
        ParallelDescriptor::ReduceRealMax(visc_max_mask0);
        if (ParallelDescriptor::IOProcessor() && lev == 0)
        {
            Print() << "\nHorizontal mixing scaled by grid metric\n";
            Print() << "grdmax = " << grdmax << "\n";
            if (solverChoice.scaled_to_grid_amr_scaling == ScaledToGridAMRScaling::linear) {
                Print() << "AMR scaling (linear) lev_scale = " << lev_scale << "\n";
            }
            Print() << "visc2(all)      min/max = "
                    << visc_min_all << " / "
                    << visc_max_all << "\n";
            Print() << "visc2(wet,k=0)  min/max = "
                    << visc_min_wet << " / "
                    << visc_max_wet << "\n";
            Print() << "visc2(mask->0)  min/max = "
                    << visc_min_mask0 << " / "
                    << visc_max_mask0 << "\n";
        }
 
    } else {
        Abort("Don't know this horizontal mixing type");
    }
 
    // Final FillPatch for all fields
    Real time = zero;
    FillPatch(lev, time, *vec_visc2_p[lev], GetVecOfPtrs(vec_visc2_p), foextrap_periodic_bc());
    FillPatch(lev, time, *vec_visc2_r[lev], GetVecOfPtrs(vec_visc2_r), foextrap_periodic_bc());
    for (int n = 0; n < ncons; n++) {
        FillPatch(lev, time, *vec_diff2[lev], GetVecOfPtrs(vec_diff2),
                  foextrap_periodic_bc(), BdyVars::null, n, false);
    }
}
 
/**
 * @param[in   ] lev    level to operate on
 */
void
REMORA::set_smflux(int lev)
{
    BL_PROFILE("REMORA::set_smflux()");
    if (solverChoice.smflux_type == SMFluxType::analytic) {
        prob->init_analytic_smflux(lev, geom[lev], solverChoice, *this,*vec_sustr[lev], *vec_svstr[lev]);
    } else if (solverChoice.smflux_type == SMFluxType::netcdf) {
#ifdef REMORA_USE_NETCDF
        sustr_data_from_file->update_interpolated_to_time(model_time(t_old[lev]), lev, vec_sustr[lev].get(), geom, ref_ratio);
        svstr_data_from_file->update_interpolated_to_time(model_time(t_old[lev]), lev, vec_svstr[lev].get(), geom, ref_ratio);
        FillPatch(lev, t_old[lev], *vec_sustr[lev], GetVecOfPtrs(vec_sustr), foextrap_periodic_bc(), BdyVars::null,0,false,false);
        FillPatch(lev, t_old[lev], *vec_svstr[lev], GetVecOfPtrs(vec_svstr), foextrap_periodic_bc(), BdyVars::null,0,false,false);
#endif
    }
}
 
/**
 * @param[in   ] lev    level to operate on
 */
void
REMORA::set_surface_state (int lev)
{
    BL_PROFILE("REMORA::set_surface_state()");
 
    auto& bulk_flux_type = solverChoice.bulk_flux_type;
 
    // Every update below skips driver-supplied lanes individually, on
    // !driver_atmos_state_from_driver[...]. This used to abort outright if any
    // lane was driver-supplied, which made the function unreachable in a coupled
    // run and so denied the *withheld* lanes the fallback those guards provide.
 
#ifdef REMORA_USE_NETCDF
    auto update_from_netcdf = [&](std::unique_ptr<NCTimeSeries>& data_from_file,
                                  Vector<std::unique_ptr<MultiFab>>& mf_vec) {
        data_from_file->update_interpolated_to_time(model_time(t_old[lev]), lev, mf_vec[lev].get(), geom, ref_ratio);
        FillPatch(lev, t_old[lev], *mf_vec[lev], GetVecOfPtrs(mf_vec),
                  foextrap_periodic_bc(), BdyVars::null, 0, false);
    };
 
    if (bulk_flux_type[BulkFlux::Uwind] == BulkForcingType::netcdf && !driver_atmos_state_from_driver[AtmosState::Uwind]) {
        update_from_netcdf(Uwind_data_from_file, vec_uwind);
    }
    if (bulk_flux_type[BulkFlux::Vwind] == BulkForcingType::netcdf && !driver_atmos_state_from_driver[AtmosState::Vwind]) {
        update_from_netcdf(Vwind_data_from_file, vec_vwind);
    }
 
    if (bulk_flux_type[BulkFlux::Tair] == BulkForcingType::netcdf && !driver_atmos_state_from_driver[AtmosState::Tair]) {
        update_from_netcdf(Tair_data_from_file, vec_Tair);
    }
    if (bulk_flux_type[BulkFlux::Qair] == BulkForcingType::netcdf && !driver_atmos_state_from_driver[AtmosState::Qair]) {
        update_from_netcdf(qair_data_from_file, vec_qair);
        if (solverChoice.qair_is_percent) {
            vec_qair[lev]->mult(amrex::Real(0.01));
        }
    }
    if (bulk_flux_type[BulkFlux::Pair] == BulkForcingType::netcdf && !driver_atmos_state_from_driver[AtmosState::Pair]) {
        update_from_netcdf(Pair_data_from_file, vec_Pair);
    }
    if (bulk_flux_type[BulkFlux::SWrad] == BulkForcingType::netcdf && !driver_atmos_state_from_driver[AtmosState::SWrad]) {
        update_from_netcdf(srflx_data_from_file, vec_srflx);
    }
    if (bulk_flux_type[BulkFlux::LWrad] == BulkForcingType::netcdf && !driver_atmos_state_from_driver[AtmosState::LWrad]) {
        update_from_netcdf(longwave_down_data_from_file, vec_longwave_down);
    }
    if (bulk_flux_type[BulkFlux::Rain] == BulkForcingType::netcdf && !driver_atmos_state_from_driver[AtmosState::Rain]) {
        update_from_netcdf(rain_data_from_file, vec_rain);
    }
    if (bulk_flux_type[BulkFlux::Cloud] == BulkForcingType::netcdf && !driver_atmos_state_from_driver[AtmosState::Cloud]) {
        update_from_netcdf(cloud_data_from_file, vec_cloud);
    }
    if (bulk_flux_type[BulkFlux::EminusP] == BulkForcingType::netcdf) {
        update_from_netcdf(EminusP_data_from_file, vec_EminusP);
    }
#else
    for (int idx = 0; idx < BulkFlux::NumTypes; ++idx) {
        if (bulk_flux_type[idx] == BulkForcingType::netcdf) {
            amrex::Abort("NetCDF bulk-flux forcing requires building with NetCDF");
        }
    }
#endif
 
    MultiFab* analytic_uwind = (bulk_flux_type[BulkFlux::Uwind] == BulkForcingType::analytic &&
                               !driver_atmos_state_from_driver[AtmosState::Uwind]) ? vec_uwind[lev].get() : nullptr;
    MultiFab* analytic_vwind = (bulk_flux_type[BulkFlux::Vwind] == BulkForcingType::analytic &&
                               !driver_atmos_state_from_driver[AtmosState::Vwind]) ? vec_vwind[lev].get() : nullptr;
    MultiFab* analytic_Tair = (bulk_flux_type[BulkFlux::Tair] == BulkForcingType::analytic &&
                               !driver_atmos_state_from_driver[AtmosState::Tair]) ? vec_Tair[lev].get() : nullptr;
    MultiFab* analytic_qair = (bulk_flux_type[BulkFlux::Qair] == BulkForcingType::analytic &&
                               !driver_atmos_state_from_driver[AtmosState::Qair]) ? vec_qair[lev].get() : nullptr;
    MultiFab* analytic_Pair = (bulk_flux_type[BulkFlux::Pair] == BulkForcingType::analytic &&
                               !driver_atmos_state_from_driver[AtmosState::Pair]) ? vec_Pair[lev].get() : nullptr;
    MultiFab* analytic_srflx = (bulk_flux_type[BulkFlux::SWrad] == BulkForcingType::analytic &&
                                !driver_atmos_state_from_driver[AtmosState::SWrad]) ? vec_srflx[lev].get() : nullptr;
    MultiFab* analytic_lwrad = (bulk_flux_type[BulkFlux::LWrad] == BulkForcingType::analytic &&
                                !driver_atmos_state_from_driver[AtmosState::LWrad]) ? vec_longwave_down[lev].get() : nullptr;
    MultiFab* analytic_rain = (bulk_flux_type[BulkFlux::Rain] == BulkForcingType::analytic &&
                               !driver_atmos_state_from_driver[AtmosState::Rain]) ? vec_rain[lev].get() : nullptr;
    MultiFab* analytic_cloud = (bulk_flux_type[BulkFlux::Cloud] == BulkForcingType::analytic &&
                                !driver_atmos_state_from_driver[AtmosState::Cloud]) ? vec_cloud[lev].get() : nullptr;
    MultiFab* analytic_EminusP = bulk_flux_type[BulkFlux::EminusP] == BulkForcingType::analytic ? vec_EminusP[lev].get() : nullptr;
 
    if (analytic_uwind != nullptr || analytic_vwind != nullptr ||
        analytic_Tair != nullptr || analytic_qair != nullptr || analytic_Pair != nullptr ||
        analytic_srflx != nullptr || analytic_lwrad != nullptr || analytic_rain != nullptr ||
        analytic_cloud != nullptr || analytic_EminusP != nullptr) {
        // Every field has to be passed to init_analytic_surface_var, but only the
        // analytic ones may be modified: the others hold constant or NetCDF data that is
        // set once at level creation or interpolated just above. Hand the non-analytic
        // slots scratch data that is thrown away on return, so the problem code can write
        // to all ten references unconditionally without clobbering anything.
        Vector<std::unique_ptr<MultiFab>> scratch_mf;
        auto analytic_or_scratch = [&] (MultiFab* mf_analytic,
                                        const std::unique_ptr<MultiFab>& mf_lev) -> MultiFab&
        {
            if (mf_analytic != nullptr) { return *mf_analytic; }
            scratch_mf.emplace_back(new MultiFab(mf_lev->boxArray(), mf_lev->DistributionMap(),
                                                 mf_lev->nComp(), mf_lev->nGrowVect()));
            return *scratch_mf.back();
        };
 
        prob->init_analytic_surface_var(lev, geom[lev], solverChoice, *this,
                                        analytic_or_scratch(analytic_uwind, vec_uwind[lev]),
                                        analytic_or_scratch(analytic_vwind, vec_vwind[lev]),
                                        analytic_or_scratch(analytic_Tair, vec_Tair[lev]),
                                        analytic_or_scratch(analytic_qair, vec_qair[lev]),
                                        analytic_or_scratch(analytic_Pair, vec_Pair[lev]),
                                        analytic_or_scratch(analytic_srflx, vec_srflx[lev]),
                                        analytic_or_scratch(analytic_lwrad, vec_longwave_down[lev]),
                                        analytic_or_scratch(analytic_rain, vec_rain[lev]),
                                        analytic_or_scratch(analytic_cloud, vec_cloud[lev]),
                                        analytic_or_scratch(analytic_EminusP, vec_EminusP[lev]));
    }
 
    if (vec_uwind[lev] != nullptr) { vec_uwind[lev]->FillBoundary(geom[lev].periodicity()); }
    if (vec_vwind[lev] != nullptr) { vec_vwind[lev]->FillBoundary(geom[lev].periodicity()); }
    if (vec_Tair[lev] != nullptr) { vec_Tair[lev]->FillBoundary(geom[lev].periodicity()); }
    if (vec_qair[lev] != nullptr) { vec_qair[lev]->FillBoundary(geom[lev].periodicity()); }
    if (vec_Pair[lev] != nullptr) { vec_Pair[lev]->FillBoundary(geom[lev].periodicity()); }
    if (vec_srflx[lev] != nullptr) { vec_srflx[lev]->FillBoundary(geom[lev].periodicity()); }
    if (vec_longwave_down[lev] != nullptr) { vec_longwave_down[lev]->FillBoundary(geom[lev].periodicity()); }
    if (vec_rain[lev] != nullptr) { vec_rain[lev]->FillBoundary(geom[lev].periodicity()); }
    if (vec_cloud[lev] != nullptr) { vec_cloud[lev]->FillBoundary(geom[lev].periodicity()); }
    if (vec_EminusP[lev] != nullptr) { vec_EminusP[lev]->FillBoundary(geom[lev].periodicity()); }
}
 
/**
 * @param[in   ] lev    level to operate on
 * @param[in   ] time   current time for initialization
 */
void
REMORA::init_only (int lev, Real time)
{
    BL_PROFILE("REMORA::init_only()");
    t_new[lev] = time;
    t_old[lev] = time - bogus_large_value;
 
    cons_new[lev]->setVal(zero);
    xvel_new[lev]->setVal(zero);
    yvel_new[lev]->setVal(zero);
    zvel_new[lev]->setVal(zero);
 
    xvel_old[lev]->setVal(zero);
    yvel_old[lev]->setVal(zero);
    zvel_old[lev]->setVal(zero);
 
    vec_ru[lev]->setVal(zero);
    vec_rv[lev]->setVal(zero);
 
    vec_ru2d[lev]->setVal(zero);
    vec_rv2d[lev]->setVal(zero);
 
    if (solverChoice.ic_type == IC_Type::analytic) {
        set_grid_scale(lev);
    }
 
    // High-resolution grid data, both sources in one place because the order matters: the mask
    // is coarsened first, and the bathymetry is then coarsened with it, so a coarse cell only
    // partly covered by water takes the depth of that water. mask_type and ic_type are set
    // independently, so each is dispatched on its own.
    if (lev==0 and hires_grid_level > 0) {
        allocate_bathymetry_grid_vars_full_domain();
 
        if (solverChoice.mask_type == MaskType::analytic) {
            init_masks_full_domain_from_analytic();
        } else if (solverChoice.mask_type == MaskType::netcdf) {
#ifdef REMORA_USE_NETCDF
            amrex::Print() << "Reading high resolution land-sea mask" << std::endl;
            init_masks_full_domain_from_netcdf();
            amrex::Print() << "Done reading in high resolution land-sea mask" << std::endl;
#endif
        }
 
        if (solverChoice.ic_type == IC_Type::analytic) {
            init_bathymetry_full_domain_from_analytic();
        } else if (solverChoice.ic_type == IC_Type::netcdf) {
#ifdef REMORA_USE_NETCDF
            amrex::Print() << "Reading high resolution bathymetry and grid data" << std::endl;
            init_bathymetry_full_domain_from_netcdf();
            init_grid_vars_full_domain_from_netcdf();
            amrex::Print() << "Done reading in high resolution bathymetry and grid data" << std::endl;
#endif
        }
    }
 
#ifdef REMORA_USE_NETCDF
    if (solverChoice.ic_type == IC_Type::netcdf) {
        init_clim_nudg_coeff(lev);
 
        if (solverChoice.do_any_clim_nudg && lev == 0) {
            if (nc_clim_his_file.empty() || nc_clim_his_file[0].empty()) {
                amrex::Error("NetCDF climatology file name must be provided via input");
            }
            if (solverChoice.do_m2_clim_nudg) {
                ubar_clim_data_from_file.reset(new NCTimeSeries(nc_clim_his_file, "ubar",
                            clim_ubar_time_varname, geom[lev].Domain(),vec_ubar[lev].get(),true,true));
                vbar_clim_data_from_file.reset(new NCTimeSeries(nc_clim_his_file, "vbar",
                            clim_ubar_time_varname, geom[lev].Domain(),vec_vbar[lev].get(),true,true));
                ubar_clim_data_from_file->Initialize();
                vbar_clim_data_from_file->Initialize();
            }
            if (solverChoice.do_m3_clim_nudg) {
                u_clim_data_from_file.reset(new NCTimeSeries(nc_clim_his_file, "u", clim_u_time_varname, geom[lev].Domain(),xvel_new[lev],false,true));
                v_clim_data_from_file.reset(new NCTimeSeries(nc_clim_his_file, "v", clim_v_time_varname, geom[lev].Domain(),yvel_new[lev],false,true));
                u_clim_data_from_file->Initialize();
                v_clim_data_from_file->Initialize();
            }
            // Since the NCTimeSeries object isn't filling the cons_new MultiFab directly, we don't have to specify a component.
            // It just needs to know the shape of the MultiFab
            cons_clim_data_from_file.resize(ncons);
            for (int icomp = 0; icomp < ncons; ++icomp) {
                if (!solverChoice.do_cons_clim_nudg[icomp]) { continue; }
                // A tracer's climatology is stored in the file under the tracer's own
                // name, following the same convention ROMS uses. Check up front rather
                // than letting the read fail deep inside NCTimeSeries.
                for (const auto& fname : nc_clim_his_file) {
                    if (!QueryNetCDFHasVars(fname, {cons_names[icomp]})) {
                        amrex::Abort("Climatology file " + fname + " does not contain '" +
                                     cons_names[icomp] + "', which is required by remora.do_" +
                                     cons_names[icomp] + "_clim_nudg. Either add it to the file "
                                     "or turn that flag off.");
                    }
                }
                cons_clim_data_from_file[icomp].reset(new NCTimeSeries(nc_clim_his_file, cons_names[icomp],
                            clim_cons_time_varname[icomp],geom[lev].Domain(),cons_new[lev],false,true));
                cons_clim_data_from_file[icomp]->Initialize();
            }
        }
    }
 
    if (solverChoice.boundary_from_netcdf) {
        amrex::Print() << "Calling init_bdry_from_netcdf at level " << lev << std::endl;
        init_bdry_from_netcdf(lev);
        amrex::Print() << "Boundary data loaded from netcdf file \n " << std::endl;
    }
 
    // This will be a non-op if forcings specified analytically
    if (solverChoice.smflux_type == SMFluxType::netcdf) {
        if (lev==0) {
            if (nc_frc_file.empty() || nc_frc_file[0].empty()) {
                amrex::Error("NetCDF forcing file name must be provided via input for surface momentum fluxes");
            }
            sustr_data_from_file.reset(new NCTimeSeries(nc_frc_file, "sustr", frc_time_varname, geom[lev].Domain(),vec_sustr[lev].get(), true, false));
            svstr_data_from_file.reset(new NCTimeSeries(nc_frc_file, "svstr", frc_time_varname, geom[lev].Domain(),vec_svstr[lev].get(), true, false));
            sustr_data_from_file->Initialize();
            svstr_data_from_file->Initialize();
        } else {
            FillCoarsePatch(lev, time, vec_sustr[lev].get(), vec_sustr[lev-1].get(), foextrap_bc());
            FillCoarsePatch(lev, time, vec_svstr[lev].get(), vec_svstr[lev-1].get(), foextrap_bc());
        }
    }
 
    // Conditionally load atmospheric forcing fields from NetCDF based on source type.
    const auto& bulk_flux_type = solverChoice.bulk_flux_type;
    bool any_bulk_netcdf = false;
    for (int idx = 0; idx < BulkFlux::NumTypes; ++idx) {
        any_bulk_netcdf = any_bulk_netcdf || bulk_flux_type[idx] == BulkForcingType::netcdf;
    }
    if (lev == 0 && any_bulk_netcdf && (nc_frc_file.empty() || nc_frc_file[0].empty())) {
        amrex::Error("NetCDF forcing file name must be provided via input for bulk-flux atmospheric forcing");
    }
 
    if (lev==0) {
        if (bulk_flux_type[BulkFlux::Uwind] == BulkForcingType::netcdf) {
            Uwind_data_from_file.reset(new NCTimeSeries(nc_frc_file, "Uwind", frc_time_varname, geom[lev].Domain(),vec_uwind[lev].get(), true, false));
            Uwind_data_from_file->Initialize();
        }
        if (bulk_flux_type[BulkFlux::Vwind] == BulkForcingType::netcdf) {
            Vwind_data_from_file.reset(new NCTimeSeries(nc_frc_file, "Vwind", frc_time_varname, geom[lev].Domain(),vec_vwind[lev].get(), true, false));
            Vwind_data_from_file->Initialize();
        }
        if (bulk_flux_type[BulkFlux::Tair] == BulkForcingType::netcdf) {
            Tair_data_from_file.reset(new NCTimeSeries(nc_frc_file, "Tair", frc_time_varname, geom[lev].Domain(),vec_Tair[lev].get(), true, false));
            Tair_data_from_file->Initialize();
        }
        if (bulk_flux_type[BulkFlux::Qair] == BulkForcingType::netcdf) {
            qair_data_from_file.reset(new NCTimeSeries(nc_frc_file, "qair", frc_time_varname, geom[lev].Domain(),vec_qair[lev].get(), true, false));
            qair_data_from_file->Initialize();
        }
        if (bulk_flux_type[BulkFlux::Pair] == BulkForcingType::netcdf) {
            Pair_data_from_file.reset(new NCTimeSeries(nc_frc_file, "Pair", frc_time_varname, geom[lev].Domain(),vec_Pair[lev].get(), true, false));
            Pair_data_from_file->Initialize();
        }
        if (bulk_flux_type[BulkFlux::SWrad] == BulkForcingType::netcdf) {
            srflx_data_from_file.reset(new NCTimeSeries(nc_frc_file, "swrad", frc_time_varname, geom[lev].Domain(),vec_srflx[lev].get(), true, false));
            srflx_data_from_file->Initialize();
        }
        if (bulk_flux_type[BulkFlux::Rain] == BulkForcingType::netcdf) {
            rain_data_from_file.reset(new NCTimeSeries(nc_frc_file, "rain", frc_time_varname, geom[lev].Domain(),vec_rain[lev].get(), true, false));
            rain_data_from_file->Initialize();
        }
        if (bulk_flux_type[BulkFlux::Cloud] == BulkForcingType::netcdf) {
            cloud_data_from_file.reset(new NCTimeSeries(nc_frc_file, "cloud", frc_time_varname, geom[lev].Domain(),vec_cloud[lev].get(), true, false));
            cloud_data_from_file->Initialize();
        }
        if (bulk_flux_type[BulkFlux::EminusP] == BulkForcingType::netcdf) {
            EminusP_data_from_file.reset(new NCTimeSeries(nc_frc_file, "EminusP", frc_time_varname, geom[lev].Domain(),vec_EminusP[lev].get(), true, false));
            EminusP_data_from_file->Initialize();
        }
        if (bulk_flux_type[BulkFlux::LWrad] == BulkForcingType::netcdf) {
            longwave_down_data_from_file.reset(new NCTimeSeries(nc_frc_file, solverChoice.longwave_netcdf_varname, frc_time_varname,
                                                                geom[lev].Domain(), vec_longwave_down[lev].get(), true, false));
            longwave_down_data_from_file->Initialize();
        }
    } else {
        if (bulk_flux_type[BulkFlux::Uwind] == BulkForcingType::netcdf) {
            FillCoarsePatch(lev, time, vec_uwind[lev].get(), vec_uwind[lev-1].get(), foextrap_bc());
        }
        if (bulk_flux_type[BulkFlux::Vwind] == BulkForcingType::netcdf) {
            FillCoarsePatch(lev, time, vec_vwind[lev].get(), vec_vwind[lev-1].get(), foextrap_bc());
        }
        if (bulk_flux_type[BulkFlux::Tair] == BulkForcingType::netcdf) {
            FillCoarsePatch(lev, time, vec_Tair[lev].get(), vec_Tair[lev-1].get(), foextrap_bc());
        }
        if (bulk_flux_type[BulkFlux::Qair] == BulkForcingType::netcdf) {
            FillCoarsePatch(lev, time, vec_qair[lev].get(), vec_qair[lev-1].get(), foextrap_bc());
        }
        if (bulk_flux_type[BulkFlux::Pair] == BulkForcingType::netcdf) {
            FillCoarsePatch(lev, time, vec_Pair[lev].get(), vec_Pair[lev-1].get(), foextrap_bc());
        }
        if (bulk_flux_type[BulkFlux::SWrad] == BulkForcingType::netcdf) {
            FillCoarsePatch(lev, time, vec_srflx[lev].get(), vec_srflx[lev-1].get(), foextrap_bc());
        }
        if (bulk_flux_type[BulkFlux::Rain] == BulkForcingType::netcdf) {
            FillCoarsePatch(lev, time, vec_rain[lev].get(), vec_rain[lev-1].get(), foextrap_bc());
        }
        if (bulk_flux_type[BulkFlux::Cloud] == BulkForcingType::netcdf) {
            FillCoarsePatch(lev, time, vec_cloud[lev].get(), vec_cloud[lev-1].get(), foextrap_bc());
        }
        if (bulk_flux_type[BulkFlux::EminusP] == BulkForcingType::netcdf) {
            FillCoarsePatch(lev, time, vec_EminusP[lev].get(), vec_EminusP[lev-1].get(), foextrap_bc());
        }
        if (bulk_flux_type[BulkFlux::LWrad] == BulkForcingType::netcdf) {
            FillCoarsePatch(lev, time, vec_longwave_down[lev].get(), vec_longwave_down[lev-1].get(), foextrap_bc());
        }
    }
 
    // Only need to read in rivers on level 0
    // Will need to be on higher levels eventually
    if (solverChoice.do_rivers) {
        if (nc_riv_file.empty() || nc_riv_file[0].empty()) {
            amrex::Error("NetCDF river file name must be provided via input for rivers");
        }
        auto dom = geom[0].Domain();
        int nz = dom.length(2);
        // Every cell-centered tracer can take river input. The field is named for the
        // tracer, as in ROMS: river_temp, river_salt, river_tracer, river_NO3, ...
        river_source_cons.resize(ncons);
        for (int icomp = 0; icomp < ncons; ++icomp) {
            if (!solverChoice.do_rivers_cons[icomp]) { continue; }
 
            const std::string field = "river_" + cons_names[icomp];
            for (const auto& fname : nc_riv_file) {
                if (!QueryNetCDFHasVars(fname, {field})) {
                    // The flag may have come from remora.do_rivers_scalar rather than the
                    // per-tracer key, so name the tracer and the key that switches it off.
                    amrex::Abort("River file " + fname + " does not contain '" + field +
                                 "', but river input is enabled for tracer '" +
                                 cons_names[icomp] + "'. Either add that variable to the "
                                 "file, or set remora.do_rivers_" + cons_names[icomp] +
                                 " = false.");
                }
            }
 
            river_source_cons[icomp].reset(new NCTimeSeriesRiver(nc_riv_file, field, riv_time_varname, nz));
            river_source_cons[icomp]->Initialize();
        }
        river_source_transport.reset(new NCTimeSeriesRiver(nc_riv_file, "river_transport", riv_time_varname, nz, 0, 1));
        river_source_transport->Initialize();
        river_source_transportbar.reset(new NCTimeSeriesRiver(nc_riv_file, "river_transport", riv_time_varname, nz, 1, 1));
        river_source_transportbar->Initialize();
        init_riv_pos_from_netcdf(lev);
    }
 
#else
    if (solverChoice.ic_type == IC_Type::netcdf) {
        Abort("Not compiled with NetCDF, but remora.ic_type = netcdf reads initial and grid data from file");
    }
    // hires_grid_level and hires_init_level need no guard: with analytic initialization
    // neither reads NetCDF, and the netcdf case is caught above.
    if (solverChoice.boundary_from_netcdf) {
        Abort("Not compiled with NetCDF, but selected boundary conditions require NetCDF");
    }
    if (solverChoice.do_rivers) {
        Abort("Not compiled with NetCDF, but using river sources requires NetCDF");
    }
#endif
 
    set_bathymetry(lev);
    // Has to follow set_bathymetry, not precede it as it used to: the mask now has a
    // hires_grid_level path of its own, which needs the full-domain data read just above,
    // and an analytic mask needs the grid coordinates that set_bathymetry -> set_grid_scale
    // fills on the netcdf path.
    set_masks(lev);
 
    // Has to sit between set_masks and set_zeta. After set_masks, since
    // ensure_full_domain_masks seeds level 0 from vec_mskr[0], and init_masks' all-water
    // placeholder would reduce the weighting below to a plain mean. Before set_zeta, which
    // reads the vec_zeta_full_domain this fills.
    if (lev==0 and hires_init_level > 0) {
        if (solverChoice.ic_type == IC_Type::netcdf) {
#ifdef REMORA_USE_NETCDF
            amrex::Print() << "Reading high resolution initial data" << std::endl;
            allocate_init_full_domain();
            // The initial-state cascade is mask-weighted, so a mask has to exist this high up.
            ensure_full_domain_masks(hires_init_level);
            init_data_full_domain_from_netcdf();
            // Biology source is chosen by remora.biology_ic_type, not by ic_type,
            // so this goes through the same dispatcher as the per-level path.
            // Must follow init_data_full_domain_from_netcdf: the analytic biology
            // profiles read temperature.
            init_biology_ic_full_domain();
            init_zeta_full_domain_from_netcdf();
            amrex::Print() << "Done reading in high resolution initial data" << std::endl;
#endif
        } else if (solverChoice.ic_type == IC_Type::analytic) {
            allocate_init_full_domain();
            ensure_full_domain_masks(hires_init_level);
            init_full_domain_from_analytic();
        }
    }
 
    set_zeta(lev);
    stretch_transform(lev);
 
    if (lev==0) {
        if (hires_init_level < 0) {
            if (solverChoice.ic_type == IC_Type::analytic) {
                init_analytic(lev);
            } else if (solverChoice.ic_type == IC_Type::netcdf) {
#ifdef REMORA_USE_NETCDF
                amrex::Print() << "Calling init_data_from_netcdf " << std::endl;
                init_data_from_netcdf(lev);
                bool apply_eminusp = false;
                set_zeta_to_Ztavg(lev, apply_eminusp);
                amrex::Print() << "Initial data loaded from netcdf file \n " << std::endl;
#endif
            } else {
                amrex::Abort("Unknown IC_Type");
            }
            // Biology last, and outside the ic_type branches: its source is
            // chosen independently by remora.biology_ic_type, and the analytic
            // profiles need the physical fields already in place.
            init_biology_ic(lev);
        } else {
            set_init_data_averaged_down(lev); // also sets biology data
            bool apply_eminusp = false;
            set_zeta_to_Ztavg(lev, apply_eminusp);
            // Since set_grid_scale is usually called from init_analytic for analytic problems
            if (solverChoice.ic_type == IC_Type::analytic) {
                set_grid_scale(lev);
            }
        }
    } else {
        if (lev > hires_init_level) {
            FillCoarsePatch(lev, time, cons_new[lev], cons_new[lev-1],BCVars::Temp_bc_comp,BdyVars::t);
            FillCoarsePatch(lev, time, xvel_new[lev], xvel_new[lev-1], xvel_bc(), BdyVars::u);
            FillCoarsePatch(lev, time, yvel_new[lev], yvel_new[lev-1], yvel_bc(), BdyVars::v);
            FillCoarsePatch(lev, time, zvel_new[lev], zvel_new[lev-1], zvel_bc(), BdyVars::null);
        } else {
            set_init_data_averaged_down(lev); // also sets biology data
            bool apply_eminusp = false;
            set_zeta_to_Ztavg(lev, apply_eminusp);
            if (solverChoice.ic_type == IC_Type::analytic) {
                // Since set_grid_scale is usually called from init_analytic for analytic problems
                set_grid_scale(lev);
            }
        }
    }
 
    // Ensure that the face-based data are the same on both sides of a periodic domain.
    // The data associated with the lower grid ID is considered the correct value.
    xvel_new[lev]->OverrideSync(geom[lev].periodicity());
    yvel_new[lev]->OverrideSync(geom[lev].periodicity());
    zvel_new[lev]->OverrideSync(geom[lev].periodicity());
 
    set_2darrays(lev);
 
    init_set_vmix(lev);
    set_hmixcoef(lev);
    set_coriolis(lev);
 
    // Previously set smflux here with OverrideSync:
//    set_smflux(lev);
//    prob->init_analytic_smflux(lev, geom[lev], solverChoice, *this, *vec_sustr[lev], *vec_svstr[lev]);
//    vec_sustr[lev]->OverrideSync(geom[lev].periodicity());
//    vec_svstr[lev]->OverrideSync(geom[lev].periodicity());
 
}
 
void
REMORA::ReadParameters ()
{
    BL_PROFILE("REMORA::ReadParameters()");
    {
        ParmParse pp;  // Traditionally, max_step and stop_time do not have prefix, so allow it for now.
        bool noprefix_max_step = pp.queryAdd("max_step", max_step);
        bool noprefix_stop_time = pp.queryAdd("stop_time", stop_time);
        bool remora_max_step = pp.queryAdd("remora.max_step", max_step);
        bool remora_stop_time = pp.queryAdd("remora.stop_time", stop_time);
        if (remora_max_step and noprefix_max_step) {
            Abort("remora.max_step and max_step are both specified. Please use only one!");
        }
        if (remora_stop_time and noprefix_stop_time) {
            Abort("remora.stop_time and stop_time are both specified. Please use only one!");
        }
    }
 
    ParmParse pp(pp_prefix);
 
    // Common physics and simulation parameters
    std::string biology_model_string = REMORABiology::biology_model_name(biology_model);
    pp.queryAdd("biology_model", biology_model_string);
    biology_model = REMORABiology::parse_biology_model(biology_model_string);
 
    if (REMORABiology::has_biology(biology_model)) {
        fennel_params.init_params(pp_prefix);
 
        // Source of the biology initial condition, independent of ic_type.
        // Default "follow" reproduces the previous behaviour exactly.
        std::string biology_ic_string = REMORABiology::biology_ic_type_name(biology_ic_type);
        pp.queryAdd("biology_ic_type", biology_ic_string);
        biology_ic_type = REMORABiology::parse_biology_ic_type(biology_ic_string);
 
        // Bridge-vs-native selection and diagnostic verbosity are runtime
        // controls so a parity comparison never requires a rebuild. Both
        // parse unconditionally; without USE_FENNEL_FORT there is no bridge
        // to select, so asking for it is an error rather than a silent
        // fallback to the path being validated.
        pp.queryAdd("use_biology_cpp_answer", use_biology_cpp_answer);
        pp.queryAdd("biology_debug", biology_debug);
        pp.queryAdd("biology_debug_i", biology_debug_i);
        pp.queryAdd("biology_debug_j", biology_debug_j);
#ifndef REMORA_USE_FENNEL_FORT
        if (use_biology_cpp_answer == 0) {
            amrex::Abort("remora.use_biology_cpp_answer = 0 selects the ROMS "
                         "Fennel Fortran bridge, which is not compiled in. "
                         "Rebuild with USE_FENNEL_FORT=TRUE (GNUmake) or "
                         "-DREMORA_ENABLE_FENNEL_FORT=ON (CMake).");
        }
#endif
#ifndef REMORA_USE_BIOLOGY_DIAG
        if (biology_debug > 0) {
            amrex::Abort("remora.biology_debug > 0 requests the Fennel parity "
                         "diagnostics, which are not compiled in. Rebuild with "
                         "USE_BIOLOGY_DIAG=TRUE (GNUmake) or "
                         "-DREMORA_ENABLE_BIOLOGY_DIAG=ON (CMake).");
        }
#endif
 
        // Biology tracers are counted separately from the passive scalars, so a run can
        // carry dye and biology at once.
        nbio = static_cast<int>(REMORABiology::tracer_names(biology_model, fennel_params).size());
    } else {
        nbio = 0;
    }
    // Dye is opt-in, biology or not: a component nothing asked for is one more thing to
    // advect, diffuse, and explain in every plotfile and boundary file.
    pp.queryAdd("nscalar", nscalar);
    if (nscalar < 0) {
        amrex::Abort("remora.nscalar must be non-negative");
    }
    Bio_comp = Tracer_comp + nscalar;
    ncons = Tracer_comp + nscalar + nbio;
    init_scalar_metadata();
 
    // remora.nscalar used to be required to equal the biology tracer count; it now counts
    // dye only and adds to it. Print the layout so a run carrying both is unmistakable,
    // and an input written against the old meaning is caught by eye rather than by a
    // surprising component count much later.
    if (nbio > 0 && nscalar > 0) {
        amrex::Print() << "Carrying " << nscalar << " passive scalar(s) and " << nbio
                       << " biology tracer(s), for " << ncons << " cell-centered components: ";
        for (int icomp = 0; icomp < ncons; ++icomp) {
            amrex::Print() << cons_names[icomp] << (icomp + 1 < ncons ? " " : "\n");
        }
    }
 
    pp.queryAdd("check_file", check_file);
    pp.queryAdd("check_int", check_int);
    pp.queryAdd("check_int_time", check_int_time);
    pp.queryAdd("expand_plotvars_to_unif_rr", expand_plotvars_to_unif_rr);
    pp.query("plotfile_fill_value", plotfile_fill_value);
    pp.query("netcdf_fill_value", netcdf_fill_value);
    pp.queryAdd("restart", restart_chkfile);
    pp.queryAdd("start_time", start_time);
 
    num_boxes_at_level.resize(max_level + 1, 0);
    boxes_at_level.resize(max_level + 1);
    num_boxes_at_level[0] = 1;
    boxes_at_level[0].resize(1);
    boxes_at_level[0][0] = geom[0].Domain();
 
    if (pp.contains("data_log")) {
        int num_datalogs = pp.countval("data_log");
        datalog.resize(num_datalogs);
        datalogname.resize(num_datalogs);
        pp.queryarr("data_log", datalogname, 0, num_datalogs);
        for (int i = 0; i < num_datalogs; i++)
            setRecordDataInfo(i, datalogname[i]);
    }
 
    pp.queryAdd("v", verbose);
    pp.queryAdd("sum_interval", sum_interval);
    pp.queryAdd("sum_period", sum_per);
    pp.queryAdd("file_min_digits", file_min_digits);
 
    if (file_min_digits < 0) {
        amrex::Abort("remora.file_min_digits must be non-negative");
    }
 
    pp.queryAdd("cfl", cfl);
    pp.queryAdd("change_max", change_max);
    pp.queryAdd("fixed_dt", fixed_dt);
 
    // remora.fixed_fast_dt has been removed. It only ever served to infer the number of
    // barotropic substeps, and only when remora.fixed_dt was also given -- which left the
    // ratio at zero on every other path, including a CFL-driven run. amrex does not abort
    // on unused inputs by default, so catch it here rather than letting a stale input file
    // silently fall back to whatever remora.ndtfast happens to be.
    if (pp.contains("fixed_fast_dt")) {
        amrex::Abort("remora.fixed_fast_dt has been removed. Set remora.ndtfast (the "
                     "number of barotropic steps per baroclinic step) instead; it is what "
                     "fixed_fast_dt was used to infer, as remora.fixed_dt / "
                     "remora.fixed_fast_dt");
    }
 
    // remora.ndtfast is the preferred name; remora.fixed_ndtfast_ratio is kept as a
    // deprecated alias so existing input files keep working. Read the alias first, so
    // that the queryAdd below records the resulting value under the preferred name.
    if (pp.contains("fixed_ndtfast_ratio")) {
        if (pp.contains("ndtfast")) {
            amrex::Abort("remora.ndtfast and remora.fixed_ndtfast_ratio are both "
                         "specified. Please use only remora.ndtfast");
        }
        amrex::Print() << "WARNING: remora.fixed_ndtfast_ratio is deprecated. "
                       << "Please use remora.ndtfast instead." << std::endl;
        // Deprecated alias for remora.ndtfast.
        pp.queryAdd("fixed_ndtfast_ratio", ndtfast);
    }
    // Number of barotropic (fast) steps taken per baroclinic (slow) step.
    pp.queryAdd("ndtfast", ndtfast);
 
    // 0 selects timeStepML, kept as a comparison path. amr.do_substep is the original
    // spelling; read it first so the queryAdd below records the value under the new name.
    {
        ParmParse pp_amr("amr");
        if (pp_amr.contains("do_substep")) {
            if (pp.contains("do_substep")) {
                amrex::Abort("remora.do_substep and amr.do_substep are both specified. "
                             "Please use only remora.do_substep");
            }
            amrex::Print() << "WARNING: amr.do_substep is deprecated. "
                           << "Please use remora.do_substep instead." << std::endl;
            pp_amr.queryAdd("do_substep", do_substep);
        }
    }
    pp.queryAdd("do_substep", do_substep);
 
    if (max_level > 0) {
        if (do_substep) {
            amrex::Print() << "WARNING: subcycling refined levels in time is EXPERIMENTAL.\n"
                           << "         Multi-level answers are not yet considered production\n"
                           << "         quality; check them against a single-level run before\n"
                           << "         relying on them. remora.do_substep = 0 selects the older\n"
                           << "         lockstep driver instead.\n";
        } else {
            amrex::Print() << "NOTE: remora.do_substep = 0 selects the lockstep driver. It cannot\n"
                           << "      impose the parent's mass flux at a coarse-fine interface, so\n"
                           << "      it conserves volume less well: 2.0e-6 against 2.9e-08 on\n"
                           << "      Dogbone. Refinement is EXPERIMENTAL on either driver.\n";
        }
    }
 
    // Write the parent's flux straight onto DUon/DVom instead of letting the solver rebuild it
    // from the imposed velocity. See set_2d_cf_flux.
    pp.queryAdd("cf_impose_flux", cf_impose_flux);
 
    // See set_2d_cf_bcs. Only has an effect when remora.do_substep = 1.
    pp.queryAdd("time_interp_flux", time_interp_flux);
 
    // Tracer flux correction at the coarse-fine interface. Needs remora.do_substep and
    // two-way coupling to do anything.
    pp.queryAdd("do_reflux", do_reflux);
 
    // Mirror ROMS's fine2coarse, which returns every covered cell but not the perimeter's
    // normal-velocity faces. 0 restores the behaviour from before that was matched.
    pp.queryAdd("cf_avgdown_perimeter", cf_avgdown_perimeter);
 
    // Diagnostics; see the declarations in REMORA.H.
    pp.queryAdd("cf_fill_vel_after", cf_fill_vel_after);
    pp.queryAdd("cf_set_2d_bcs", cf_set_2d_bcs);
    pp.queryAdd("cf_avgdown_bar", cf_avgdown_bar);
    pp.queryAdd("cf_fill_all_kcomp", cf_fill_all_kcomp);
    pp.queryAdd("cf_time_interp_zeta", cf_time_interp_zeta);
    pp.queryAdd("cf_flux_pc", cf_flux_pc);
    pp.queryAdd("cf_avgdown_stencil", cf_avgdown_stencil);
    pp.queryAdd("cf_print_iface", cf_print_iface);
 
    // Whether to zero a tracer the correction drives negative, as ROMS does.
    pp.queryAdd("reflux_clamp", reflux_clamp);
 
    // See check_cf_metrics. Off by default: it is a property of the grid, so one run says as
    // much as every run.
    pp.queryAdd("check_cf_metrics", check_cf_metrics_flag);
    pp.queryAdd("check_cf_tol", check_cf_tol);
 
    // Advance and timeStepML form the fast step as dt / ndtfast, and set_weights sizes
    // the barotropic filter with the same number, so a non-positive value divides by zero
    // at all three sites. Nothing can infer it: dt is not known until run time on a
    // CFL-driven run.
    if (ndtfast <= 0) {
        amrex::Abort("remora.ndtfast must be a positive integer: it is the number of "
                     "barotropic steps taken per baroclinic step");
    }
 
    // remora.use_barotropic has been removed -- the barotropic mode is always on. Reject
    // it on presence rather than on value: amrex does not abort on unused inputs by
    // default, so a stale "= false" would otherwise silently run different physics than
    // the input file asks for. Tested with contains() rather than query() so the schema
    // scraper behind Exec/Generic does not re-advertise a parameter that no longer exists.
    if (pp.contains("use_barotropic")) {
        amrex::Abort("remora.use_barotropic has been removed. The barotropic (2D) mode is "
                     "always active; please delete this line from your inputs file");
    }
 
    AMREX_ASSERT(cfl > zero || fixed_dt > zero);
 
    num_files_at_level.resize(max_level + 1, 0);
    num_boxes_at_level.resize(max_level + 1, 0);
    boxes_at_level.resize(max_level + 1);
    num_boxes_at_level[0] = 1;
    boxes_at_level[0].resize(1);
    boxes_at_level[0][0] = geom[0].Domain();
 
    pp.queryAdd("plot_file", plot_file_name);
    pp.queryAdd("plot_int", plot_int);
    pp.queryAdd("plot_int_time", plot_int_time);
    pp.query("plot_staggered_vels", plot_staggered_vels);
    pp.query("plot_nodal_data", plot_nodal_data);
 
    std::string plotfile_type_str = "amrex";
    pp.queryAdd("plotfile_type", plotfile_type_str);
    if (plotfile_type_str == "amrex") {
        plotfile_type = PlotfileType::amrex;
    } else if (plotfile_type_str == "netcdf" || plotfile_type_str == "NetCDF") {
        plotfile_type = PlotfileType::netcdf;
#ifdef REMORA_USE_NETCDF
        pp.queryAdd("write_history_file",write_history_file);
        pp.queryAdd("chunk_history_file",chunk_history_file);
        pp.queryAdd("steps_per_history_file",steps_per_history_file);
        // CDF-5 output has no practical size limit, so REMORA doesn't size history
        // files. Chunking is opt-in; the writer divides by steps_per_history_file.
        if (write_history_file and chunk_history_file) {
            if (steps_per_history_file <= 0) {
                amrex::Abort("remora.chunk_history_file requires remora.steps_per_history_file > 0");
            }
            Print() << "NetCDF history files will have " << steps_per_history_file << " steps per file." << std::endl;
        }
#endif
    } else {
        amrex::Print() << "User selected plotfile_type = " << plotfile_type_str << std::endl;
        amrex::Abort("Dont know this plotfile_type");
    }
#ifndef REMORA_USE_NETCDF
    if (plotfile_type == PlotfileType::netcdf)
    {
        amrex::Abort("Please compile with NetCDF in order to enable NetCDF plotfiles");
    }
 
#endif
#ifdef REMORA_USE_NETCDF
    nc_init_file.resize(max_level+1);
    nc_grid_file.resize(max_level+1);
    num_files_at_level.resize(max_level + 1, 0);
 
    boundary_series.resize(max_level+1);
 
 
    // NetCDF initialization and grid files, read independently of each other: with
    // hires_grid_level or hires_init_level set, the corresponding level 0 file is not
    // needed. Whether a level 0 file is required is checked once the solver choices
    // are known, below.
    for (int lev = 0; lev <= max_level; lev++)
    {
        const std::string nc_file_names = amrex::Concatenate("nc_init_file_",lev,1);
        const std::string nc_bathy_file_names = amrex::Concatenate("nc_grid_file_",lev,1);
 
        if (pp.contains(nc_file_names.c_str()))
        {
            int num_files = pp.countval(nc_file_names.c_str());
            num_files_at_level[lev] = num_files;
            nc_init_file[lev].resize(num_files);
            pp.queryarr(nc_file_names.c_str(), nc_init_file[lev], 0, num_files);
        }
        if (pp.contains(nc_bathy_file_names.c_str()))
        {
            int num_bathy_files = pp.countval(nc_bathy_file_names.c_str());
            nc_grid_file[lev].resize(num_bathy_files);
            pp.queryarr(nc_bathy_file_names.c_str(), nc_grid_file[lev], 0, num_bathy_files);
        }
    }
 
    pp.queryAdd("nc_grid_file_hires", nc_grid_file_hires);
    pp.queryAdd("nc_init_file_hires", nc_init_file_hires);
 
    // We only read boundary data at level 0
    pp.queryarr("nc_bdry_file", nc_bdry_file);
 
    // Also only read forcings at level 0 (for now)
    if (pp.contains("nc_frc_file")) {
        int num_files = pp.countval("nc_frc_file");
        nc_frc_file.resize(num_files);
        pp.queryarr("nc_frc_file", nc_frc_file, 0, num_files);
    }
 
    // Get river file
    if (pp.contains("nc_river_file")) {
        int num_files = pp.countval("nc_river_file");
        nc_riv_file.resize(num_files);
        pp.queryarr("nc_river_file", nc_riv_file, 0, num_files);
    }
 
    // Read in file names for climatology history and nudging weights
    if (pp.contains("nc_clim_his_file")) {
        int num_files = pp.countval("nc_clim_his_file");
        nc_clim_his_file.resize(num_files);
        pp.queryarr("nc_clim_his_file", nc_clim_his_file, 0, num_files);
    }
    pp.queryAdd("nc_clim_coeff_file", nc_clim_coeff_file);
 
    for (int i=0; i<BdyVars::NumTypes(ncons); i++) {
        bdry_time_name_byvar.push_back("");
    }
    pp.queryAdd("bdy_time_varname",bdry_time_varname);
    // Every tracer takes its time-axis name from its own variable name, so temp and salt
    // keep bdy_temp_time_varname / bdy_salt_time_varname and a biology tracer uses e.g.
    // bdy_NO3_time_varname
    for (int icomp = 0; icomp < ncons; ++icomp) {
        pp.queryAdd(("bdy_"+cons_names[icomp]+"_time_varname").c_str(),
                    bdry_time_name_byvar[BdyVars::cons(icomp)]);
    }
    pp.queryAdd("bdy_u_time_varname",bdry_time_name_byvar[BdyVars::u]);
    pp.queryAdd("bdy_v_time_varname",bdry_time_name_byvar[BdyVars::v]);
    pp.queryAdd("bdy_ubar_time_varname",bdry_time_name_byvar[BdyVars::ubar(ncons)]);
    pp.queryAdd("bdy_vbar_time_varname",bdry_time_name_byvar[BdyVars::vbar(ncons)]);
    pp.queryAdd("bdy_zeta_time_varname",bdry_time_name_byvar[BdyVars::zeta(ncons)]);
 
    // If not specified per variable, populate with the default
    for (int i=0; i<BdyVars::NumTypes(ncons); i++) {
        if (bdry_time_name_byvar[i] == "") {
            bdry_time_name_byvar[i] = bdry_time_varname;
        }
    }
 
    pp.queryAdd("frc_time_varname",frc_time_varname);
 
    pp.queryAdd("riv_time_varname",riv_time_varname);
 
    pp.queryAdd("clim_ubar_time_varname",clim_ubar_time_varname);
    pp.queryAdd("clim_vbar_time_varname",clim_vbar_time_varname);
    pp.queryAdd("clim_u_time_varname",clim_u_time_varname);
    pp.queryAdd("clim_v_time_varname",clim_v_time_varname);
    // As for the boundary data, each tracer takes its climatology time-axis name from its
    // own variable name, so temp and salt keep clim_temp_time_varname /
    // clim_salt_time_varname and a biology tracer uses e.g. clim_NO3_time_varname
    clim_cons_time_varname.assign(ncons, "ocean_time");
    for (int icomp = 0; icomp < ncons; ++icomp) {
        pp.queryAdd(("clim_"+cons_names[icomp]+"_time_varname").c_str(),
                    clim_cons_time_varname[icomp]);
    }
 
#endif
    // A hires level of 0 is not "level 0 is the hires level", it is a null pointer: the
    // full-domain arrays are only allocated for lev > 0 (see allocate_init_full_domain and
    // allocate_bathymetry_grid_vars_full_domain), while every consumer branch tests < 0 and
    // so would take the averaged-down path against an unallocated MultiFab. -1 means off.
    pp.queryAdd("hires_grid_level", hires_grid_level);
    if (hires_grid_level > max_level) {
        amrex::Abort("hires_grid_level must be less than or equal to amr.max_level");
    }
    if (hires_grid_level == 0) {
        amrex::Abort("hires_grid_level must be greater than 0; use -1 to specify grid data at level 0");
    }
    pp.queryAdd("hires_init_level", hires_init_level);
    if (hires_init_level > max_level) {
        amrex::Abort("hires_init_level must be less than or equal to amr.max_level");
    }
    if (hires_init_level == 0) {
        amrex::Abort("hires_init_level must be greater than 0; use -1 to specify initial data at level 0");
    }
#ifdef REMORA_USE_PARTICLES
    readTracersParams();
#endif
 
    {
        ParmParse pp_amr("amr");
        pp_amr.queryAdd("regrid_int", regrid_int);
    }
    solverChoice.init_params(ncons, nscalar, cons_names);
 
    // The biology IC source is chosen independently of ic_type, but only one of the two
    // mixed combinations works: NetCDF physics with analytic biology. The reverse has no
    // file to read from -- nc_init_file is only populated on the netcdf path -- so catch
    // it here instead of failing inside PnetCDF on an empty file name.
    if (REMORABiology::has_biology(biology_model) and
        biology_ic_type == REMORABiology::BiologyICType::netcdf and
        solverChoice.ic_type != IC_Type::netcdf) {
        amrex::Abort("remora.biology_ic_type = netcdf requires remora.ic_type = netcdf: the biology "
                     "initial data is read from the same files as the physical initial data, and no "
                     "such file is given for analytic initial conditions. Use "
                     "remora.biology_ic_type = analytic (or follow) instead.");
    }
 
#ifndef REMORA_USE_NETCDF
    if (solverChoice.ic_type == IC_Type::netcdf) {
        amrex::Abort("Please compile with NetCDF in order to use remora.ic_type = netcdf");
    }
#else
    // Level 0 files are read only for the fields not supplied at hires_init_level or
    // hires_grid_level, so require each one only where it will be read.
    {
        const bool have_init_0 = !nc_init_file[0].empty() && !nc_init_file[0][0].empty();
        const bool have_grid_0 = !nc_grid_file[0].empty() && !nc_grid_file[0][0].empty();
        if (solverChoice.ic_type == IC_Type::netcdf && hires_init_level < 0 && !have_init_0) {
            amrex::Abort("remora.ic_type = netcdf requires remora.nc_init_file_0 unless "
                         "remora.hires_init_level is set");
        }
        if (hires_grid_level < 0 && !have_grid_0 &&
            (solverChoice.ic_type == IC_Type::netcdf || solverChoice.mask_type == MaskType::netcdf)) {
            amrex::Abort("remora.ic_type = netcdf or remora.mask_type = netcdf requires "
                         "remora.nc_grid_file_0 unless remora.hires_grid_level is set");
        }
        if (solverChoice.use_coriolis && solverChoice.coriolis_type == Cor_Type::netcdf &&
            !have_grid_0) {
            amrex::Abort("remora.coriolis_type = netcdf requires remora.nc_grid_file_0");
        }
    }
#endif
 
    // init_full_domain_from_analytic builds x_r and y_r from the cell size.
    if (hires_init_level > 0 and solverChoice.ic_type == IC_Type::analytic and
        solverChoice.grid_scale_type != GridScaleType::constant) {
        amrex::Abort("High-resolution initialization with analytic initial conditions requires "
                     "remora.grid_scale_type = constant");
    }
 
}
 
 
void
REMORA::AverageDown ()
{
    BL_PROFILE("REMORA::AverageDown()");
    for (int lev = finest_level-1; lev >= 0; --lev)
    {
        AverageDownTo(lev);
    }
}
 
/**
 * Drop the cached average-down masks of every level pair that involves lev, so the next
 * AverageDownTo rebuilds them.
 *
 * Called from set_masks, which is the one place a level's mask is written, so a regrid cannot
 * leave a cached copy of a mask that no longer exists behind.
 *
 * @param[in   ] lev   level whose mask has just been rebuilt
 */
void
REMORA::clear_avgdown_masks (int lev)
{
    // lev as the coarse half of a pair, and lev as the fine half, whose layout is what the
    // cached arrays are built on.
    for (int crse_lev : {lev-1, lev}) {
        if (crse_lev >= 0 && crse_lev < static_cast<int>(vec_mskr_crse_on_fine.size())) {
            vec_mskr_crse_on_fine[crse_lev].reset();
            vec_msku_crse_on_fine[crse_lev].reset();
            vec_mskv_crse_on_fine[crse_lev].reset();
        }
    }
}
 
/**
 * Measure whether the fine cell edges tile the coarse ones across a coarse-fine interface:
 * the sum of on_u over the r covering fine faces against on_u on the coarse face.
 *
 * Exact where pm and pn are uniform, and not guaranteed otherwise -- on the NetCDF path a
 * finer level interpolates its metrics from the parent's and rescales them, which need not
 * preserve a sum. set_2d_cf_bcs cannot impose a conservative flux without this identity.
 *
 * @param[in   ] crse_lev  coarse side of the interface
 */
void
REMORA::check_cf_metrics (int crse_lev)
{
    const int rrx = ref_ratio[crse_lev][0];
    const int rry = ref_ratio[crse_lev][1];
 
    auto edge_residual = [&] (int dir, int ratio) -> Real
    {
        const IntVect ixt = (dir == 0) ? IntVect(1,0,0) : IntVect(0,1,0);
        const MultiFab& metric_f = (dir == 0) ? *vec_pn[crse_lev+1] : *vec_pm[crse_lev+1];
        const MultiFab& metric_c = (dir == 0) ? *vec_pn[crse_lev  ] : *vec_pm[crse_lev  ];
 
        auto edge_lengths = [&] (const MultiFab& metric, int lev, MultiFab& out)
        {
            // pm and pn already live on the z-flattened BoxArray.
            BoxArray ba = convert(metric.boxArray(), ixt);
            out.define(ba, metric.DistributionMap(), 1, 0);
            for (MFIter mfi(out, TilingIfNotGPU()); mfi.isValid(); ++mfi) {
                Array4<Real      > const& e = out.array(mfi);
                Array4<Real const> const& m = metric.const_array(mfi);
                const int d = dir;
                ParallelFor(mfi.tilebox(), [=] AMREX_GPU_DEVICE (int i, int j, int)
                {
                    const int im = (d == 0) ? i-1 : i;
                    const int jm = (d == 0) ? j   : j-1;
                    e(i,j,0) = two / (m(i,j,0) + m(im,jm,0));
                });
            }
            amrex::ignore_unused(lev);
        };
 
        MultiFab edge_f, edge_c;
        edge_lengths(metric_f, crse_lev+1, edge_f);
        edge_lengths(metric_c, crse_lev  , edge_c);
 
        // average_down_faces leaves faces the finer level does not cover untouched. Seeding
        // with the coarse value over the ratio makes those contribute exactly zero below,
        // rather than whatever the allocation happened to hold.
        MultiFab avg(edge_c.boxArray(), edge_c.DistributionMap(), 1, 0);
        MultiFab::Copy(avg, edge_c, 0, 0, 1, 0);
        avg.mult(one / Real(ratio), 0, 1, 0);
 
        average_down_faces(edge_f, avg, refRatio(crse_lev), 0);
 
        // avg is the mean over the covering fine faces, so ratio*avg is their sum.
        Real worst = zero;
        for (MFIter mfi(avg, TilingIfNotGPU()); mfi.isValid(); ++mfi) {
            Array4<Real const> const& a = avg.const_array(mfi);
            Array4<Real const> const& c = edge_c.const_array(mfi);
            ReduceOps<ReduceOpMax> reduce_op;
            ReduceData<Real> reduce_data(reduce_op);
            reduce_op.eval(mfi.tilebox(), reduce_data,
            [=] AMREX_GPU_DEVICE (int i, int j, int k) -> GpuTuple<Real>
            {
                if (c(i,j,k) == zero) { return {zero}; }
                return {amrex::Math::abs(Real(ratio) * a(i,j,k) - c(i,j,k)) /
                        amrex::Math::abs(c(i,j,k))};
            });
            worst = amrex::max(worst, amrex::get<0>(reduce_data.value(reduce_op)));
        }
        ParallelDescriptor::ReduceRealMax(worst);
        return worst;
    };
 
    const Real res_u = edge_residual(0, rry);
    const Real res_v = edge_residual(1, rrx);
 
    amrex::Print() << "CF edge tiling, levels " << crse_lev << "/" << crse_lev+1
                   << ": max relative residual on_u " << res_u
                   << ", om_v " << res_v << std::endl;
 
    AMREX_ALWAYS_ASSERT_WITH_MESSAGE(
        amrex::max(res_u, res_v) < check_cf_tol,
        "REMORA::check_cf_metrics: fine cell edges do not sum to the coarse edge across the "
        "coarse-fine interface, so the mass flux set_2d_cf_bcs imposes there cannot be "
        "conservative. Raise remora.check_cf_tol only if you know why the grid does this.");
}
 
/**
 * Build the flux register holding the tracer flux mismatch between lev and lev-1. Called
 * whenever a level is created or its grids change, since one appearing mid-run through
 * tagging would otherwise have no register.
 *
 * @param[in   ] lev  level of refinement, > 0
 */
void
REMORA::define_flux_register (int lev)
{
    if (lev == 0 || solverChoice.coupling_type != CouplingType::two_way) { return; }
 
    // Take the layout from the MultiFabs rather than grids/dmap: when a level is first
    // created those have not been set on AmrCore yet, as the fill patchers here also assume.
    advflux_reg[lev].reset( new YAFluxRegister(cons_new[lev  ]->boxArray(),
                                               cons_new[lev-1]->boxArray(),
                                               cons_new[lev  ]->DistributionMap(),
                                               cons_new[lev-1]->DistributionMap(),
                                               geom[lev], geom[lev-1],
                                               ref_ratio[lev-1], lev, ncons));
 
    // The constructor sizes the accumulators without zeroing them, and Reflux adds all of
    // m_crse_data to the state, so a register must be zero before its first use whichever
    // path built it.
    advflux_reg[lev]->reset();
}
 
/**
 * Set how many steps each level takes per parent step, from remora.dt_ref_ratio.
 *
 * @param[in   ] nlevs_max  max_level + 1
 */
void
REMORA::set_nsubsteps (int nlevs_max)
{
    nsubsteps.resize(nlevs_max, 1);
    if (!do_substep) { return; }
 
    for (int lev = 1; lev <= max_level; ++lev) {
        nsubsteps[lev] = MaxRefRatio(lev-1);
    }
 
    // Defaults to the spatial ratio, but the two are independent in ROMS (RefineSteps
    // against RefineScale) and in ERF. One value for all levels or one per level.
    if (max_level > 0) {
        ParmParse pp("remora");
        int count = pp.countval("dt_ref_ratio");
        if (count > 0) {
            Vector<int> nsub(nlevs_max, 0);
            if (count == 1) {
                pp.queryarr("dt_ref_ratio", nsub, 0, 1);
                for (int lev = 1; lev <= max_level; ++lev) { nsubsteps[lev] = nsub[0]; }
            } else {
                pp.queryarr("dt_ref_ratio", nsub, 0, max_level);
                for (int lev = 1; lev <= max_level; ++lev) { nsubsteps[lev] = nsub[lev-1]; }
            }
            for (int lev = 1; lev <= max_level; ++lev) {
                AMREX_ALWAYS_ASSERT_WITH_MESSAGE(nsubsteps[lev] > 0,
                    "remora.dt_ref_ratio must be positive: it divides the parent timestep");
            }
        }
    }
}
 
/**
 * Make sure the coarse rho-, u- and v-masks are defined on the layout average_down_masked
 * computes on: level crse_lev+1's grids coarsened, rather than level crse_lev's own grids.
 *
 * The masks are a function of position alone, so between regrids this is the same answer every
 * step; building it once turns three allocations and three ParallelCopy calls per step into
 * three per regrid.
 *
 * @param[in   ] crse_lev   coarse level of the pair
 */
void
REMORA::update_avgdown_masks (int crse_lev)
{
    BL_PROFILE("REMORA::update_avgdown_masks()");
    const int flev = crse_lev + 1;
    const IntVect ratio = refRatio(crse_lev);
 
    const BoxArray cba = amrex::coarsen(vec_mskr[flev]->boxArray(), ratio);
    const DistributionMapping& dmf = vec_mskr[flev]->DistributionMap();
 
    // clear_avgdown_masks drops the cache whenever a mask is rewritten; this catches anything
    // that gets here without having gone through it, by rebuilding when the layout has moved.
    if (vec_mskr_crse_on_fine[crse_lev] &&
        vec_mskr_crse_on_fine[crse_lev]->boxArray() == cba &&
        vec_mskr_crse_on_fine[crse_lev]->DistributionMap() == dmf) {
        return;
    }
 
    vec_mskr_crse_on_fine[crse_lev].reset(new MultiFab(cba, dmf, 1, 0));
    vec_msku_crse_on_fine[crse_lev].reset(
            new MultiFab(amrex::convert(cba, IntVect(1,0,0)), dmf, 1, 0));
    vec_mskv_crse_on_fine[crse_lev].reset(
            new MultiFab(amrex::convert(cba, IntVect(0,1,0)), dmf, 1, 0));
 
    vec_mskr_crse_on_fine[crse_lev]->ParallelCopy(*vec_mskr[crse_lev], 0, 0, 1);
    vec_msku_crse_on_fine[crse_lev]->ParallelCopy(*vec_msku[crse_lev], 0, 0, 1);
    vec_mskv_crse_on_fine[crse_lev]->ParallelCopy(*vec_mskv[crse_lev], 0, 0, 1);
}
 
/**
 * @param[in   ] crse_lev  level to average down to
 */
 
namespace {
/** \brief Replace each face by the mean of itself and its neighbours along dir.
 *
 * ROMS's fine2coarse2d averages the donor over a (Rscale-1)/2 half-width stencil in both
 * horizontal directions, so at ratio 3 a coarse face takes the mean of nine fine faces;
 * AMReX's average_down_faces takes only the three, all tangential, that tile it. Smoothing
 * along the face normal first supplies the missing direction and reproduces ROMS exactly.
 * Masked as ROMS does it: sum over wet points, divide by the wet count.
 *
 * Deliberately not conservative -- the nine-point mean does not preserve the transport through
 * the interface, where average_down_faces alone does. ROMS accepts that trade.
 */
void smooth_faces_along (amrex::MultiFab& mf, const amrex::MultiFab& msk, int dir,
                         const amrex::Geometry& geom, int ncomp)
{
    mf.FillBoundary(geom.periodicity());
    amrex::MultiFab orig(mf.boxArray(), mf.DistributionMap(), ncomp, mf.nGrowVect());
    amrex::MultiFab::Copy(orig, mf, 0, 0, ncomp, mf.nGrowVect());
    const int di = (dir == 0) ? 1 : 0;
    const int dj = (dir == 1) ? 1 : 0;
    for (amrex::MFIter mfi(mf, amrex::TilingIfNotGPU()); mfi.isValid(); ++mfi) {
        const amrex::Box& bx = mfi.tilebox();
        const auto& a = mf.array(mfi);
        const auto& o = orig.const_array(mfi);
        const auto& m = msk.const_array(mfi);
        amrex::ParallelFor(bx, ncomp,
        [=] AMREX_GPU_DEVICE (int i, int j, int k, int n)
        {
            amrex::Real sum = amrex::Real(0.0), cnt = amrex::Real(0.0);
            for (int t = -1; t <= 1; ++t) {
                const int ii = i + t*di, jj = j + t*dj;
                const amrex::Real w = amrex::min(amrex::Real(1.0), m(ii,jj,0));
                sum += o(ii,jj,k,n) * m(ii,jj,0);
                cnt += w;
            }
            if (cnt > amrex::Real(0.0)) { a(i,j,k,n) = sum / cnt; }
        });
    }
}
} // namespace
 
void
REMORA::AverageDownTo (int crse_lev)
{
    BL_PROFILE("REMORA::AverageDownTo()");
    const int flev = crse_lev + 1;
 
    // average_down_masked indexes the coarse mask with the same MFIter as its coarsened-fine
    // temporary, so the mask has to be defined on that layout. It is the same between regrids,
    // so this builds it once instead of every step.
    update_avgdown_masks(crse_lev);
    const MultiFab& cmskr = *vec_mskr_crse_on_fine[crse_lev];
    const MultiFab& cmsku = *vec_msku_crse_on_fine[crse_lev];
    const MultiFab& cmskv = *vec_mskv_crse_on_fine[crse_lev];
 
    // Which mask goes with which field follows the ROMS fine2coarse call sites: rmask for the
    // free surface and the tracers, umask and vmask for the momenta.
    average_down_masked(crse_lev, *cons_new[flev], *cons_new[crse_lev],
                        *vec_mskr[flev], cmskr, cons_new[crse_lev]->nComp(), -1);
    average_down_masked(crse_lev, *vec_Zt_avg1[flev], *vec_Zt_avg1[crse_lev],
                        *vec_mskr[flev], cmskr, vec_Zt_avg1[crse_lev]->nComp(), -1);
    // ROMS returns every covered cell but not the perimeter's normal-velocity faces, so the
    // transport its nested boundary condition imposes there never comes back to the parent.
    // Without that exclusion set_2d_cf_bcs writes the parent's own flux onto the interface and
    // the average hands it straight back: a factor of 20 to 40 in the barotropic mode against
    // ROMS on the matched dogbone.
    MultiFab covered;
    MultiFab u_save, v_save;
    const bool keep_perimeter = (cf_avgdown_perimeter != 0);
    if (keep_perimeter) {
        build_covered_mask(crse_lev, covered);
        u_save.define(xvel_new[crse_lev]->boxArray(), xvel_new[crse_lev]->DistributionMap(),
                      1, 0, MFInfo());
        v_save.define(yvel_new[crse_lev]->boxArray(), yvel_new[crse_lev]->DistributionMap(),
                      1, 0, MFInfo());
        MultiFab::Copy(u_save, *xvel_new[crse_lev], 0, 0, 1, 0);
        MultiFab::Copy(v_save, *yvel_new[crse_lev], 0, 0, 1, 0);
    }
 
    // See smooth_faces_along: ROMS averages the donor over nine fine faces, AMReX over the
    // three that tile the coarse face, so the normal direction is smoothed in first.
    MultiFab xv_sm, yv_sm;
    const MultiFab* xv_f = xvel_new[flev];
    const MultiFab* yv_f = yvel_new[flev];
    if (cf_avgdown_stencil) {
        xv_sm.define(xvel_new[flev]->boxArray(), xvel_new[flev]->DistributionMap(),
                     1, xvel_new[flev]->nGrowVect());
        MultiFab::Copy(xv_sm, *xvel_new[flev], 0, 0, 1, xvel_new[flev]->nGrowVect());
        smooth_faces_along(xv_sm, *vec_msku[flev], 0, geom[flev], 1);
        xv_f = &xv_sm;
 
        yv_sm.define(yvel_new[flev]->boxArray(), yvel_new[flev]->DistributionMap(),
                     1, yvel_new[flev]->nGrowVect());
        MultiFab::Copy(yv_sm, *yvel_new[flev], 0, 0, 1, yvel_new[flev]->nGrowVect());
        smooth_faces_along(yv_sm, *vec_mskv[flev], 1, geom[flev], 1);
        yv_f = &yv_sm;
    }
 
    average_down_masked(crse_lev, *xv_f, *xvel_new[crse_lev],
                        *vec_msku[flev], cmsku, 1, 0);
    average_down_masked(crse_lev, *yv_f, *yvel_new[crse_lev],
                        *vec_mskv[flev], cmskv, 1, 1);
 
    if (keep_perimeter) {
        restore_perimeter_faces(u_save, *xvel_new[crse_lev], covered, 0, 1);
        restore_perimeter_faces(v_save, *yvel_new[crse_lev], covered, 1, 1);
    }
    average_down_masked(crse_lev, *zvel_new[flev], *zvel_new[crse_lev],
                        *vec_mskr[flev], cmskr, 1, 2);
 
    if (check_cf_metrics_flag) { check_cf_metrics(crse_lev); }
 
    // Hand the child's 2D momentum back, as ROMS's fine2coarse does. The parent's next
    // advance_2d reads ubar(krhs) to form DUon, so dropping this moves Dogbone's x-velocity
    // by 9%. Subcycling only: timeStepML keeps the behaviour its answers were recorded with.
    if (do_substep && cf_avgdown_bar) {
        // Components 0 and 1 only. The three are leapfrog slots rotating per level, so the
        // two levels need not agree on which holds what -- but update_massflux_3d has just
        // set both of these to the same velocity, so averaging them cannot mix time levels.
        MultiFab ub_save, vb_save;
        if (keep_perimeter) {
            ub_save.define(vec_ubar[crse_lev]->boxArray(),
                           vec_ubar[crse_lev]->DistributionMap(), 2, 0, MFInfo());
            vb_save.define(vec_vbar[crse_lev]->boxArray(),
                           vec_vbar[crse_lev]->DistributionMap(), 2, 0, MFInfo());
            MultiFab::Copy(ub_save, *vec_ubar[crse_lev], 0, 0, 2, 0);
            MultiFab::Copy(vb_save, *vec_vbar[crse_lev], 0, 0, 2, 0);
        }
 
        MultiFab ub_sm, vb_sm;
        const MultiFab* ub_src = vec_ubar[crse_lev+1].get();
        const MultiFab* vb_src = vec_vbar[crse_lev+1].get();
        if (cf_avgdown_stencil) {
            ub_sm.define(vec_ubar[flev]->boxArray(), vec_ubar[flev]->DistributionMap(),
                         2, vec_ubar[flev]->nGrowVect());
            MultiFab::Copy(ub_sm, *vec_ubar[flev], 0, 0, 2, vec_ubar[flev]->nGrowVect());
            smooth_faces_along(ub_sm, *vec_msku[flev], 0, geom[flev], 2);
            ub_src = &ub_sm;
 
            vb_sm.define(vec_vbar[flev]->boxArray(), vec_vbar[flev]->DistributionMap(),
                         2, vec_vbar[flev]->nGrowVect());
            MultiFab::Copy(vb_sm, *vec_vbar[flev], 0, 0, 2, vec_vbar[flev]->nGrowVect());
            smooth_faces_along(vb_sm, *vec_mskv[flev], 1, geom[flev], 2);
            vb_src = &vb_sm;
        }
 
        for (int icomp = 0; icomp < 2; ++icomp) {
            MultiFab ubar_f(*ub_src, make_alias, icomp, 1);
            MultiFab ubar_c(*vec_ubar[crse_lev  ], make_alias, icomp, 1);
            average_down_faces(ubar_f, ubar_c, refRatio(crse_lev), geom[crse_lev]);
 
            MultiFab vbar_f(*vb_src, make_alias, icomp, 1);
            MultiFab vbar_c(*vec_vbar[crse_lev  ], make_alias, icomp, 1);
            average_down_faces(vbar_f, vbar_c, refRatio(crse_lev), geom[crse_lev]);
        }
 
        if (keep_perimeter) {
            restore_perimeter_faces(ub_save, *vec_ubar[crse_lev], covered, 0, 2);
            restore_perimeter_faces(vb_save, *vec_vbar[crse_lev], covered, 1, 2);
        }
 
        // The averages above rewrote this level's valid cells under the patch but not the
        // ghosts imaging them across a periodic seam, and nothing else refreshes those before
        // the next barotropic step: FillPatch refills one component, fill_ghost_kcomps runs on
        // fine levels only. Where the patch abuts a periodic boundary the two copies of the
        // coarse periodic u-face then read different vbar for the same physical cell -- freshly
        // averaged against stale -- and the asymmetry grows step by step. On Channel_Test with
        // a half-width patch on the seam: 3.4e-4 by step 20 and NaN by ~300 without this,
        // exactly zero with it, and an interior patch unaffected either way.
        vec_ubar[crse_lev]->FillBoundary(geom[crse_lev].periodicity());
        vec_vbar[crse_lev]->FillBoundary(geom[crse_lev].periodicity());
 
        // zeta is deliberately absent: set_zeta_to_Ztavg overwrites all three of its
        // components from Zt_avg1 next step and stretch_transform reads Zt_avg1 anyway, so
        // averaging it here would write something nothing reads.
    }
 
    stretch_transform(crse_lev);
}
 
/**
 * Coarse-level cell mask: 1 where level crse_lev+1 covers the cell, 0 elsewhere, with one grow
 * cell so a face straddling the patch edge can see the cell outside it.
 *
 * @param[in   ] crse_lev  coarse level
 * @param[out  ] covered   mask on the coarse layout
 */
void
REMORA::build_covered_mask (int crse_lev, MultiFab& covered)
{
    // makeFineMask's (crse_value, fine_value) = (0, 1) puts a 1 exactly on the covered cells.
    iMultiFab imask = makeFineMask(grids[crse_lev], dmap[crse_lev], grids[crse_lev+1],
                                   ref_ratio[crse_lev], 0, 1);
 
    covered.define(grids[crse_lev], dmap[crse_lev], 1, 1, MFInfo());
    covered.setVal(zero);                       // grow cells outside the patch read as uncovered
    const auto cma = covered.arrays();
    const auto ima = imask.const_arrays();
    ParallelFor(covered, [=] AMREX_GPU_DEVICE (int bno, int i, int j, int k) noexcept
    {
        cma[bno](i,j,k) = Real(ima[bno](i,j,k));
    });
    Gpu::streamSynchronize();
    covered.FillBoundary(geom[crse_lev].periodicity());
}
 
/**
 * Put back the values a fine-to-coarse average just wrote onto the fine patch's perimeter
 * normal-velocity faces.
 *
 * ROMS's fine2coarse hands back every covered cell but excludes those faces: for a child
 * spanning coarse cells I_lo..I_hi it updates u-faces I_lo+1..I_hi only, so the face the
 * nested barotropic boundary condition writes is never returned to the parent. A face is
 * interior exactly when both of the cells it separates are covered, which is the test used
 * here.
 *
 * @param[in   ] saved    coarse field as it was before the average
 * @param[inout] mf       coarse field to repair
 * @param[in   ] covered  cell mask from build_covered_mask
 * @param[in   ] dir      face normal direction
 * @param[in   ] ncomp    components to repair
 */
void
REMORA::restore_perimeter_faces (const MultiFab& saved, MultiFab& mf,
                                 const MultiFab& covered, int dir, int ncomp)
{
    const IntVect off = IntVect::TheDimensionVector(dir);
    for (MFIter mfi(mf, TilingIfNotGPU()); mfi.isValid(); ++mfi)
    {
        const Box& bx = mfi.tilebox();
        Array4<Real      > const& a = mf.array(mfi);
        Array4<Real const> const& o = saved.const_array(mfi);
        Array4<Real const> const& c = covered.const_array(mfi);
        ParallelFor(bx, ncomp, [=] AMREX_GPU_DEVICE (int i, int j, int k, int n) noexcept
        {
            const bool interior = (c(i,j,0) > Real(0.5)) &&
                                  (c(i-off[0], j-off[1], 0) > Real(0.5));
            if (!interior) { a(i,j,k,n) = o(i,j,k,n); }
        });
    }
    Gpu::streamSynchronize();
}
 
/**
 * Average one field from crse_lev+1 onto crse_lev, weighting by the land/sea mask.
 *
 * Follows amrex::average_down's non-MFIter-safe branch, since coarsen(grids[flev]) does not
 * match grids[crse_lev] in general: compute onto a temporary on the coarsened-fine layout,
 * then ParallelCopy that onto the coarse level.
 *
 * @param[in   ] crse_lev   level to average down to
 * @param[in   ] S_fine     fine-level field
 * @param[out  ] S_crse     coarse-level field
 * @param[in   ] msk_fine   fine-level mask, on S_fine's layout and nodality
 * @param[in   ] cmsk       coarse-level mask, already on the coarsened-fine layout
 * @param[in   ] ncomp      number of components to average
 * @param[in   ] face_dir   face direction, or -1 for a cell-centered field
 */
void
REMORA::average_down_masked (int crse_lev, const MultiFab& S_fine, MultiFab& S_crse,
                             const MultiFab& msk_fine, const MultiFab& cmsk,
                             int ncomp, int face_dir)
{
    BL_PROFILE("REMORA::average_down_masked()");
    const IntVect ratio = refRatio(crse_lev);
 
    BoxArray cba = amrex::coarsen(S_fine.boxArray(), ratio);
    MultiFab ctmp(cba, S_fine.DistributionMap(), ncomp, 0);
 
    // One MFIter indexes all four arrays in the loop below, by local box index, so the masks
    // have to be distributed exactly as S_fine is. Equal DistributionMappings imply equal box
    // counts, since a ProcessorMap holds one entry per box.
    AMREX_ALWAYS_ASSERT(msk_fine.DistributionMap() == S_fine.DistributionMap());
    AMREX_ALWAYS_ASSERT(cmsk.DistributionMap()     == S_fine.DistributionMap());
 
    for (MFIter mfi(ctmp, TilingIfNotGPU()); mfi.isValid(); ++mfi)
    {
        const Box& bx = mfi.tilebox();
        Array4<      Real> const& c  = ctmp.array(mfi);
        Array4<const Real> const& f  = S_fine.const_array(mfi);
        Array4<const Real> const& fm = msk_fine.const_array(mfi);
        Array4<const Real> const& cm = cmsk.const_array(mfi);
 
        if (face_dir < 0) {
            ParallelFor(bx, ncomp, [=] AMREX_GPU_DEVICE (int i, int j, int k, int n) noexcept
            {
                REMORAMaskedAvgDown::avgdown_masked(i,j,k,n,c,f,fm,cm,0,0,ratio);
            });
        } else {
            ParallelFor(bx, ncomp, [=] AMREX_GPU_DEVICE (int i, int j, int k, int n) noexcept
            {
                REMORAMaskedAvgDown::avgdown_faces_masked(i,j,k,n,c,f,fm,cm,0,0,ratio,face_dir);
            });
        }
    }
    Gpu::streamSynchronize();
 
    // Periodicity arguments as amrex::average_down and average_down_faces pass them: the
    // cell-centered copy takes none, the face copy needs it so shared faces across a periodic
    // boundary agree.
    if (face_dir < 0) {
        S_crse.ParallelCopy(ctmp, 0, 0, ncomp);
    } else {
        S_crse.ParallelCopy(ctmp, 0, 0, ncomp, IntVect(0), IntVect(0),
                            geom[crse_lev].periodicity());
    }
}
 
/**
 * Inject the full-domain rho-mask from fine_lev-1 up onto fine_lev, grow cells included.
 *
 * Piecewise constant, which is the rule set_masks already uses for a level above the one the
 * mask was specified on. Nothing finer is known there, so refining cannot add coastline.
 *
 * @param[in   ] fine_lev   level to inject onto
 */
void
REMORA::refine_masks_with_grow_cells (int fine_lev)
{
    auto const& finema = vec_mskr_full_domain[fine_lev]->arrays();
    auto const& crsema = vec_mskr_full_domain[fine_lev-1]->const_arrays();
    auto ratio = refRatio(fine_lev-1);
    auto nghost_fine = cum_ref_ratios[fine_lev];
    ParallelFor(*vec_mskr_full_domain[fine_lev], nghost_fine, 1,
            [=] AMREX_GPU_DEVICE (int box_no, int i, int j, int k, int n) noexcept
    {
        // amrex::coarsen floors rather than truncating, which is what the negative indices of
        // the grow cells need.
        finema[box_no](i,j,k,n) = crsema[box_no](amrex::coarsen(i, ratio[0]),
                                                 amrex::coarsen(j, ratio[1]), k, n);
    });
    Gpu::streamSynchronize();
}
 
/**
 * Make sure the full-domain rho-mask exists on levels 0 through top_lev.
 *
 * Only fills what the hires_grid_level coarsening has not: levels above it, by injection, and
 * level 0 itself when there is no high-resolution grid and the mask was given per level.
 *
 * @param[in   ] top_lev   highest level that needs a mask
 */
void
REMORA::ensure_full_domain_masks (int top_lev)
{
    if (solverChoice.mask_type == MaskType::none || top_lev <= 0) { return; }
 
    // Levels up to hires_grid_level were coarsened down from it already.
    const int have = (hires_grid_level > 0) ? hires_grid_level : 0;
    if (top_lev <= have) { return; }
 
    BoxArray ba;
    ba.define(makeSlab(geom[0].Domain(),2,0));
    const DistributionMapping& dm = full_domain_dmap();
    auto mskr_growvect = vec_mskr[0]->nGrowVect();
 
    if (hires_grid_level < 0) {
        // No high-resolution grid, so the specification lives on level 0. Seed from it.
        vec_mskr_full_domain[0].reset(new MultiFab(ba, dm, 1, IntVect(1,1,0)));
        vec_mskr_full_domain[0]->setVal(one);
        ParallelCopy(*vec_mskr_full_domain[0].get(), *vec_mskr[0].get(), 0, 0, 1,
                vec_mskr[0]->nGrowVect(), vec_mskr_full_domain[0]->nGrowVect());
    }
 
    for (int lev = 1; lev <= top_lev; lev++) {
        ba = ba.refine(refRatio(lev-1));
        if (lev <= have) { continue; }
        vec_mskr_full_domain[lev].reset(new MultiFab(ba, dm, 1,
                    max(cum_ref_ratios[lev], mskr_growvect)));
        vec_mskr_full_domain[lev]->setVal(one);
        refine_masks_with_grow_cells(lev);
    }
}
 
namespace {
/**
 * Build a face-centered mask from a rho-point one, following ROMS set_masks.F:
 * msku = mskr(i-1,j)*mskr(i,j) and mskv = mskr(i,j-1)*mskr(i,j).
 *
 * Derived where it is needed rather than stored. Only the full-domain average-down wants
 * these, at most once per level during initialization, so a stored pair would be two more
 * arrays to keep in step with the rho mask for no measurable saving.
 */
void derive_face_mask (const MultiFab& mskr, MultiFab& mskf, int idir)
{
    const IntVect ndir = (idir == 0) ? IntVect(1,0,0) : IntVect(0,1,0);
    // One ring narrower than the rho mask in the normal direction, where the stencil reaches.
    const IntVect ng = max(mskr.nGrowVect() - ndir, IntVect(0));
    mskf.define(convert(mskr.boxArray(), ndir), mskr.DistributionMap(), 1, ng);
    mskf.setVal(one);
 
    for (MFIter mfi(mskf, TilingIfNotGPU()); mfi.isValid(); ++mfi) {
        const Box& bx = mfi.growntilebox();
        Array4<      Real> const& mf = mskf.array(mfi);
        Array4<const Real> const& mr = mskr.const_array(mfi);
        const int l_idir = idir;
        ParallelFor(bx, [=] AMREX_GPU_DEVICE (int i, int j, int k) noexcept
        {
            mf(i,j,k) = (l_idir == 0) ? mr(i-1,j,0) * mr(i,j,0)
                                      : mr(i,j-1,0) * mr(i,j,0);
        });
    }
}
} // namespace
 
/**
 * Average a full-domain field from crse_lev+1 onto crse_lev, grow cells included.
 *
 * With use_mask this takes REMORAMaskedAvgDown's mask-weighted mean instead of the plain one,
 * so the initial state on a coarse cell only partly covered by water comes from that water.
 * It is the formula AverageDownTo applies every step, so the initial and the running state
 * agree on what a land point holds. Leave use_mask off for grid metrics: a cell size is well
 * defined under land, and masking pm/pn would corrupt it.
 *
 * @param[in   ] crse_lev   level to average data down to
 * @param[inout] vec_mf     vector over levels of multifabs containing data to average
 * @param[in   ] use_mask   weight by the land/sea mask rather than averaging every fine cell
 */
void
REMORA::average_down_with_grow_cells (int crse_lev, Vector<std::unique_ptr<MultiFab>>& vec_mf,
                                      bool use_mask)
{
    auto const& crsema = vec_mf[crse_lev]->arrays();
    auto const& finema = vec_mf[crse_lev+1]->const_arrays();
    auto ref_ratio_crse = refRatio(crse_lev);
    auto index_type = (vec_mf[crse_lev]->boxArray().ixType()).toIntVect();
    auto nghost_crse = cum_ref_ratios[crse_lev] - index_type;
 
    const bool masked = use_mask && (solverChoice.mask_type != MaskType::none);
    if (masked) {
        // One box per level on a matching DistributionMapping, so an MFIter over one array
        // indexes the other. Assert it rather than assume it.
        AMREX_ALWAYS_ASSERT(vec_mskr_full_domain[crse_lev] && vec_mskr_full_domain[crse_lev+1]);
        AMREX_ALWAYS_ASSERT(vec_mskr_full_domain[crse_lev]->boxArray().size() ==
                            vec_mf[crse_lev]->boxArray().size());
        AMREX_ALWAYS_ASSERT(vec_mskr_full_domain[crse_lev]->DistributionMap() ==
                            vec_mf[crse_lev]->DistributionMap());
 
        const int idir = (index_type[0]==1) ? 0 : ((index_type[1]==1) ? 1 : -1);
        MultiFab fmsk, cmsk;
        if (idir >= 0) {
            derive_face_mask(*vec_mskr_full_domain[crse_lev+1], fmsk, idir);
            derive_face_mask(*vec_mskr_full_domain[crse_lev  ], cmsk, idir);
        }
        auto const& fmskma = (idir >= 0) ? fmsk.const_arrays()
                                         : vec_mskr_full_domain[crse_lev+1]->const_arrays();
        auto const& cmskma = (idir >= 0) ? cmsk.const_arrays()
                                         : vec_mskr_full_domain[crse_lev]->const_arrays();
        ParallelFor(*vec_mf[crse_lev], nghost_crse, vec_mf[crse_lev]->nComp(),
                [=] AMREX_GPU_DEVICE (int box_no, int i, int j, int k, int n) noexcept
        {
            if (idir < 0) {
                REMORAMaskedAvgDown::avgdown_masked(i,j,k,n,crsema[box_no],finema[box_no],
                        fmskma[box_no],cmskma[box_no],0,0,ref_ratio_crse);
            } else {
                REMORAMaskedAvgDown::avgdown_faces_masked(i,j,k,n,crsema[box_no],finema[box_no],
                        fmskma[box_no],cmskma[box_no],0,0,ref_ratio_crse,idir);
            }
        });
        Gpu::streamSynchronize();
        return;
    }
 
    if (index_type[0]==0 and index_type[1]==0) {
        ParallelFor(*vec_mf[crse_lev], nghost_crse, vec_mf[crse_lev]->nComp(),
                [=] AMREX_GPU_DEVICE (int box_no, int i, int j, int k, int n) noexcept
        {
            amrex_avgdown(i,j,k,n,crsema[box_no],finema[box_no],0,0,ref_ratio_crse);
        });
    } else if (index_type[0]==1 and index_type[1]==0) {
        ParallelFor(*vec_mf[crse_lev], nghost_crse, vec_mf[crse_lev]->nComp(),
                [=] AMREX_GPU_DEVICE (int box_no, int i, int j, int k, int n) noexcept
        {
            amrex_avgdown_faces(i,j,k,n,crsema[box_no],finema[box_no],0,0,ref_ratio_crse,0);
        });
    } else if (index_type[0]==0 and index_type[1]==1) {
        ParallelFor(*vec_mf[crse_lev], nghost_crse, vec_mf[crse_lev]->nComp(),
                [=] AMREX_GPU_DEVICE (int box_no, int i, int j, int k, int n) noexcept
        {
            amrex_avgdown_faces(i,j,k,n,crsema[box_no],finema[box_no],0,0,ref_ratio_crse,1);
        });
    } else {
        amrex::Abort("Unexpected nodality in average_down_with_grow_cells");
    }
    Gpu::streamSynchronize();
}
 
/**
 * @param[in   ] lev    level at which to get time
 */
amrex::Real REMORA::get_t_old(int lev) const
{
    return t_old[lev];
}