space.c

/*******************************************************************************
 * This file is part of SWIFT.
 * Copyright (c) 2012 Pedro Gonnet (pedro.gonnet@durham.ac.uk)
 *                    Matthieu Schaller (matthieu.schaller@durham.ac.uk)
 *               2015 Peter W. Draper (p.w.draper@durham.ac.uk)
 *               2016 John A. Regan (john.a.regan@durham.ac.uk)
 *                    Tom Theuns (tom.theuns@durham.ac.uk)
 *
 * This program is free software: you can redistribute it and/or modify
 * it under the terms of the GNU Lesser General Public License as published
 * by the Free Software Foundation, either version 3 of the License, or
 * (at your option) any later version.
 *
 * This program is distributed in the hope that it will be useful,
 * but WITHOUT ANY WARRANTY; without even the implied warranty of
 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
 * GNU General Public License for more details.
 *
 * You should have received a copy of the GNU Lesser General Public License
 * along with this program.  If not, see <http://www.gnu.org/licenses/>.
 *
 ******************************************************************************/

/* Config parameters. */
#include "../config.h"

/* Some standard headers. */
#include <float.h>
#include <limits.h>
#include <math.h>
#include <stdlib.h>
#include <string.h>

/* MPI headers. */
#ifdef WITH_MPI
#include <mpi.h>
#endif

/* This object's header. */
#include "space.h"

/* Local headers. */
#include "atomic.h"
#include "chemistry.h"
#include "const.h"
#include "cooling.h"
#include "debug.h"
#include "engine.h"
#include "error.h"
#include "gravity.h"
#include "hydro.h"
#include "kernel_hydro.h"
#include "lock.h"
#include "memswap.h"
#include "minmax.h"
#include "multipole.h"
#include "restart.h"
#include "sort_part.h"
#include "stars.h"
#include "threadpool.h"
#include "tools.h"

/* Split size. */
int space_splitsize = space_splitsize_default;
int space_subsize_pair_hydro = space_subsize_pair_hydro_default;
int space_subsize_self_hydro = space_subsize_self_hydro_default;
int space_subsize_pair_grav = space_subsize_pair_grav_default;
int space_subsize_self_grav = space_subsize_self_grav_default;
int space_subsize_pair_stars = space_subsize_pair_stars_default;
int space_subsize_self_stars = space_subsize_self_stars_default;
int space_subdepth_grav = space_subdepth_grav_default;
int space_maxsize = space_maxsize_default;
#ifdef SWIFT_DEBUG_CHECKS
int last_cell_id;
#endif

/**
 * @brief Interval stack necessary for parallel particle sorting.
 */
struct qstack {
  volatile ptrdiff_t i, j;
  volatile int min, max;
  volatile int ready;
};

/**
 * @brief Parallel particle-sorting stack
 */
struct parallel_sort {
  struct part *parts;
  struct gpart *gparts;
  struct xpart *xparts;
  struct spart *sparts;
  int *ind;
  struct qstack *stack;
  unsigned int stack_size;
  volatile unsigned int first, last, waiting;
};

/**
 * @brief Information required to compute the particle cell indices.
 */
struct index_data {
  struct space *s;
  int *ind;
  int *cell_counts;
  int count_inhibited_part;
  int count_inhibited_gpart;
  int count_inhibited_spart;
};

/**
 * @brief Recursively dismantle a cell tree.
 *
 * @param s The #space.
 * @param c The #cell to recycle.
 * @param cell_rec_begin Pointer to the start of the list of cells to recycle.
 * @param cell_rec_end Pointer to the end of the list of cells to recycle.
 * @param multipole_rec_begin Pointer to the start of the list of multipoles to
 * recycle.
 * @param multipole_rec_end Pointer to the end of the list of multipoles to
 * recycle.
 */
void space_rebuild_recycle_rec(struct space *s, struct cell *c,
                               struct cell **cell_rec_begin,
                               struct cell **cell_rec_end,
                               struct gravity_tensors **multipole_rec_begin,
                               struct gravity_tensors **multipole_rec_end) {
  if (c->split)
    for (int k = 0; k < 8; k++)
      if (c->progeny[k] != NULL) {
        space_rebuild_recycle_rec(s, c->progeny[k], cell_rec_begin,
                                  cell_rec_end, multipole_rec_begin,
                                  multipole_rec_end);

        c->progeny[k]->next = *cell_rec_begin;
        *cell_rec_begin = c->progeny[k];

        if (s->gravity) {
          c->progeny[k]->grav.multipole->next = *multipole_rec_begin;
          *multipole_rec_begin = c->progeny[k]->grav.multipole;
        }

        if (*cell_rec_end == NULL) *cell_rec_end = *cell_rec_begin;
        if (s->gravity && *multipole_rec_end == NULL)
          *multipole_rec_end = *multipole_rec_begin;

        c->progeny[k]->grav.multipole = NULL;
        c->progeny[k] = NULL;
      }
}

void space_rebuild_recycle_mapper(void *map_data, int num_elements,
                                  void *extra_data) {

  struct space *s = (struct space *)extra_data;
  struct cell *cells = (struct cell *)map_data;

  for (int k = 0; k < num_elements; k++) {
    struct cell *c = &cells[k];
    struct cell *cell_rec_begin = NULL, *cell_rec_end = NULL;
    struct gravity_tensors *multipole_rec_begin = NULL,
                           *multipole_rec_end = NULL;
    space_rebuild_recycle_rec(s, c, &cell_rec_begin, &cell_rec_end,
                              &multipole_rec_begin, &multipole_rec_end);
    if (cell_rec_begin != NULL)
      space_recycle_list(s, cell_rec_begin, cell_rec_end, multipole_rec_begin,
                         multipole_rec_end);
    c->hydro.sorts = NULL;
    c->nr_tasks = 0;
    c->grav.nr_mm_tasks = 0;
    c->hydro.density = NULL;
    c->hydro.gradient = NULL;
    c->hydro.force = NULL;
    c->grav.grav = NULL;
    c->grav.mm = NULL;
    c->hydro.dx_max_part = 0.0f;
    c->hydro.dx_max_sort = 0.0f;
    c->stars.dx_max_part = 0.f;
    c->hydro.sorted = 0;
    c->hydro.count = 0;
    c->hydro.updated = 0;
    c->hydro.inhibited = 0;
    c->grav.count = 0;
    c->grav.updated = 0;
    c->grav.inhibited = 0;
    c->stars.count = 0;
    c->stars.updated = 0;
    c->stars.inhibited = 0;
    c->grav.init = NULL;
    c->grav.init_out = NULL;
    c->hydro.extra_ghost = NULL;
    c->hydro.ghost_in = NULL;
    c->hydro.ghost_out = NULL;
    c->hydro.ghost = NULL;
    c->stars.ghost_in = NULL;
    c->stars.ghost_out = NULL;
    c->stars.ghost = NULL;
    c->stars.density = NULL;
    c->kick1 = NULL;
    c->kick2 = NULL;
    c->timestep = NULL;
    c->end_force = NULL;
    c->hydro.drift = NULL;
    c->grav.drift = NULL;
    c->hydro.cooling = NULL;
    c->sourceterms = NULL;
    c->grav.long_range = NULL;
    c->grav.down_in = NULL;
    c->grav.down = NULL;
    c->grav.mesh = NULL;
    c->super = c;
    c->hydro.super = c;
    c->grav.super = c;
    c->hydro.parts = NULL;
    c->hydro.xparts = NULL;
    c->grav.parts = NULL;
    c->stars.parts = NULL;
    c->hydro.do_sub_sort = 0;
    c->grav.do_sub_drift = 0;
    c->hydro.do_sub_drift = 0;
    c->hydro.ti_end_min = -1;
    c->hydro.ti_end_max = -1;
    c->grav.ti_end_min = -1;
    c->grav.ti_end_max = -1;
#ifdef SWIFT_DEBUG_CHECKS
    c->cellID = 0;
#endif
    if (s->gravity) bzero(c->grav.multipole, sizeof(struct gravity_tensors));
    for (int i = 0; i < 13; i++)
      if (c->hydro.sort[i] != NULL) {
        free(c->hydro.sort[i]);
        c->hydro.sort[i] = NULL;
      }
#if WITH_MPI
    c->mpi.tag = -1;

    c->mpi.hydro.recv_xv = NULL;
    c->mpi.hydro.recv_rho = NULL;
    c->mpi.hydro.recv_gradient = NULL;
    c->mpi.grav.recv = NULL;
    c->mpi.recv_ti = NULL;

    c->mpi.hydro.send_xv = NULL;
    c->mpi.hydro.send_rho = NULL;
    c->mpi.hydro.send_gradient = NULL;
    c->mpi.grav.send = NULL;
    c->mpi.send_ti = NULL;
#endif
  }
}

/**
 * @brief Free up any allocated cells.
 */
void space_free_cells(struct space *s) {

  ticks tic = getticks();

  threadpool_map(&s->e->threadpool, space_rebuild_recycle_mapper, s->cells_top,
                 s->nr_cells, sizeof(struct cell), 0, s);
  s->maxdepth = 0;

  if (s->e->verbose)
    message("took %.3f %s.", clocks_from_ticks(getticks() - tic),
            clocks_getunit());
}

/**
 * @brief Re-build the top-level cell grid.
 *
 * @param s The #space.
 * @param verbose Print messages to stdout or not.
 */
void space_regrid(struct space *s, int verbose) {

  const size_t nr_parts = s->nr_parts;
  const size_t nr_sparts = s->nr_sparts;
  const ticks tic = getticks();
  const integertime_t ti_current = (s->e != NULL) ? s->e->ti_current : 0;

  /* Run through the cells and get the current h_max. */
  // tic = getticks();
  float h_max = s->cell_min / kernel_gamma / space_stretch;
  if (nr_parts > 0) {

    /* Can we use the list of local non-empty top-level cells? */
    if (s->local_cells_with_particles_top != NULL) {
      for (int k = 0; k < s->nr_local_cells_with_particles; ++k) {
        const struct cell *c =
            &s->cells_top[s->local_cells_with_particles_top[k]];
        if (c->hydro.h_max > h_max) {
          h_max = c->hydro.h_max;
        }
        if (c->stars.h_max > h_max) {
          h_max = c->stars.h_max;
        }
      }

      /* Can we instead use all the top-level cells? */
    } else if (s->cells_top != NULL) {
      for (int k = 0; k < s->nr_cells; k++) {
        const struct cell *c = &s->cells_top[k];
        if (c->nodeID == engine_rank && c->hydro.h_max > h_max) {
          h_max = c->hydro.h_max;
        }
        if (c->nodeID == engine_rank && c->stars.h_max > h_max) {
          h_max = c->stars.h_max;
        }
      }

      /* Last option: run through the particles */
    } else {
      for (size_t k = 0; k < nr_parts; k++) {
        if (s->parts[k].h > h_max) h_max = s->parts[k].h;
      }
      for (size_t k = 0; k < nr_sparts; k++) {
        if (s->sparts[k].h > h_max) h_max = s->sparts[k].h;
      }
    }
  }

/* If we are running in parallel, make sure everybody agrees on
   how large the largest cell should be. */
#ifdef WITH_MPI
  {
    float buff;
    if (MPI_Allreduce(&h_max, &buff, 1, MPI_FLOAT, MPI_MAX, MPI_COMM_WORLD) !=
        MPI_SUCCESS)
      error("Failed to aggregate the rebuild flag across nodes.");
    h_max = buff;
  }
#endif
  if (verbose) message("h_max is %.3e (cell_min=%.3e).", h_max, s->cell_min);

  /* Get the new putative cell dimensions. */
  const int cdim[3] = {
      (int)floor(s->dim[0] /
                 fmax(h_max * kernel_gamma * space_stretch, s->cell_min)),
      (int)floor(s->dim[1] /
                 fmax(h_max * kernel_gamma * space_stretch, s->cell_min)),
      (int)floor(s->dim[2] /
                 fmax(h_max * kernel_gamma * space_stretch, s->cell_min))};

  /* Check if we have enough cells for periodicity. */
  if (s->periodic && (cdim[0] < 3 || cdim[1] < 3 || cdim[2] < 3))
    error(
        "Must have at least 3 cells in each spatial dimension when periodicity "
        "is switched on.\nThis error is often caused by any of the "
        "followings:\n"
        " - too few particles to generate a sensible grid,\n"
        " - the initial value of 'Scheduler:max_top_level_cells' is too "
        "small,\n"
        " - the (minimal) time-step is too large leading to particles with "
        "predicted smoothing lengths too large for the box size,\n"
        " - particles with velocities so large that they move by more than two "
        "box sizes per time-step.\n");

/* In MPI-Land, changing the top-level cell size requires that the
 * global partition is recomputed and the particles redistributed.
 * Be prepared to do that. */
#ifdef WITH_MPI
  double oldwidth[3];
  double oldcdim[3];
  int *oldnodeIDs = NULL;
  if (cdim[0] < s->cdim[0] || cdim[1] < s->cdim[1] || cdim[2] < s->cdim[2]) {

    /* Capture state of current space. */
    oldcdim[0] = s->cdim[0];
    oldcdim[1] = s->cdim[1];
    oldcdim[2] = s->cdim[2];
    oldwidth[0] = s->width[0];
    oldwidth[1] = s->width[1];
    oldwidth[2] = s->width[2];

    if ((oldnodeIDs = (int *)malloc(sizeof(int) * s->nr_cells)) == NULL)
      error("Failed to allocate temporary nodeIDs.");

    int cid = 0;
    for (int i = 0; i < s->cdim[0]; i++) {
      for (int j = 0; j < s->cdim[1]; j++) {
        for (int k = 0; k < s->cdim[2]; k++) {
          cid = cell_getid(oldcdim, i, j, k);
          oldnodeIDs[cid] = s->cells_top[cid].nodeID;
        }
      }
    }
  }

  /* Are we about to allocate new top level cells without a regrid?
   * Can happen when restarting the application. */
  const int no_regrid = (s->cells_top == NULL && oldnodeIDs == NULL);
#endif

  /* Do we need to re-build the upper-level cells? */
  // tic = getticks();
  if (s->cells_top == NULL || cdim[0] < s->cdim[0] || cdim[1] < s->cdim[1] ||
      cdim[2] < s->cdim[2]) {

/* Be verbose about this. */
#ifdef SWIFT_DEBUG_CHECKS
    message("(re)griding space cdim=(%d %d %d)", cdim[0], cdim[1], cdim[2]);
    fflush(stdout);
#endif

    /* Free the old cells, if they were allocated. */
    if (s->cells_top != NULL) {
      space_free_cells(s);
      free(s->local_cells_with_tasks_top);
      free(s->local_cells_top);
      free(s->cells_with_particles_top);
      free(s->local_cells_with_particles_top);
      free(s->cells_top);
      free(s->multipoles_top);
    }

    /* Also free the task arrays, these will be regenerated and we can use the
     * memory while copying the particle arrays. */
    if (s->e != NULL) scheduler_free_tasks(&s->e->sched);

    /* Set the new cell dimensions only if smaller. */
    for (int k = 0; k < 3; k++) {
      s->cdim[k] = cdim[k];
      s->width[k] = s->dim[k] / cdim[k];
      s->iwidth[k] = 1.0 / s->width[k];
    }
    const float dmin = min3(s->width[0], s->width[1], s->width[2]);

    /* Allocate the highest level of cells. */
    s->tot_cells = s->nr_cells = cdim[0] * cdim[1] * cdim[2];
    if (posix_memalign((void **)&s->cells_top, cell_align,
                       s->nr_cells * sizeof(struct cell)) != 0)
      error("Failed to allocate top-level cells.");
    bzero(s->cells_top, s->nr_cells * sizeof(struct cell));

    /* Allocate the multipoles for the top-level cells. */
    if (s->gravity) {
      if (posix_memalign((void **)&s->multipoles_top, multipole_align,
                         s->nr_cells * sizeof(struct gravity_tensors)) != 0)
        error("Failed to allocate top-level multipoles.");
      bzero(s->multipoles_top, s->nr_cells * sizeof(struct gravity_tensors));
    }

    /* Allocate the indices of local cells */
    if (posix_memalign((void **)&s->local_cells_top, SWIFT_STRUCT_ALIGNMENT,
                       s->nr_cells * sizeof(int)) != 0)
      error("Failed to allocate indices of local top-level cells.");
    bzero(s->local_cells_top, s->nr_cells * sizeof(int));

    /* Allocate the indices of local cells with tasks */
    if (posix_memalign((void **)&s->local_cells_with_tasks_top,
                       SWIFT_STRUCT_ALIGNMENT, s->nr_cells * sizeof(int)) != 0)
      error("Failed to allocate indices of local top-level cells with tasks.");
    bzero(s->local_cells_with_tasks_top, s->nr_cells * sizeof(int));

    /* Allocate the indices of cells with particles */
    if (posix_memalign((void **)&s->cells_with_particles_top,
                       SWIFT_STRUCT_ALIGNMENT, s->nr_cells * sizeof(int)) != 0)
      error("Failed to allocate indices of top-level cells with particles.");
    bzero(s->cells_with_particles_top, s->nr_cells * sizeof(int));

    /* Allocate the indices of local cells with particles */
    if (posix_memalign((void **)&s->local_cells_with_particles_top,
                       SWIFT_STRUCT_ALIGNMENT, s->nr_cells * sizeof(int)) != 0)
      error(
          "Failed to allocate indices of local top-level cells with "
          "particles.");
    bzero(s->local_cells_with_particles_top, s->nr_cells * sizeof(int));

    /* Set the cells' locks */
    for (int k = 0; k < s->nr_cells; k++) {
      if (lock_init(&s->cells_top[k].hydro.lock) != 0)
        error("Failed to init spinlock for hydro.");
      if (lock_init(&s->cells_top[k].grav.plock) != 0)
        error("Failed to init spinlock for gravity.");
      if (lock_init(&s->cells_top[k].grav.mlock) != 0)
        error("Failed to init spinlock for multipoles.");
      if (lock_init(&s->cells_top[k].stars.lock) != 0)
        error("Failed to init spinlock for stars.");
    }

    /* Set the cell location and sizes. */
    for (int i = 0; i < cdim[0]; i++)
      for (int j = 0; j < cdim[1]; j++)
        for (int k = 0; k < cdim[2]; k++) {
          const size_t cid = cell_getid(cdim, i, j, k);
          struct cell *restrict c = &s->cells_top[cid];
          c->loc[0] = i * s->width[0];
          c->loc[1] = j * s->width[1];
          c->loc[2] = k * s->width[2];
          c->width[0] = s->width[0];
          c->width[1] = s->width[1];
          c->width[2] = s->width[2];
          c->dmin = dmin;
          c->depth = 0;
          c->hydro.count = 0;
          c->grav.count = 0;
          c->stars.count = 0;
          c->super = c;
          c->hydro.super = c;
          c->grav.super = c;
          c->hydro.ti_old_part = ti_current;
          c->grav.ti_old_part = ti_current;
          c->grav.ti_old_multipole = ti_current;
#ifdef WITH_MPI
          c->mpi.tag = -1;
#endif  // WITH_MPI
          if (s->gravity) c->grav.multipole = &s->multipoles_top[cid];
#ifdef SWIFT_DEBUG_CHECKS
          c->cellID = -last_cell_id;
          last_cell_id++;
#endif
        }

    /* Be verbose about the change. */
    if (verbose)
      message("set cell dimensions to [ %i %i %i ].", cdim[0], cdim[1],
              cdim[2]);

#ifdef WITH_MPI
    if (oldnodeIDs != NULL) {
      /* We have changed the top-level cell dimension, so need to redistribute
       * cells around the nodes. We repartition using the old space node
       * positions as a grid to resample. */
      if (s->e->nodeID == 0)
        message(
            "basic cell dimensions have increased - recalculating the "
            "global partition.");

      if (!partition_space_to_space(oldwidth, oldcdim, oldnodeIDs, s)) {

        /* Failed, try another technique that requires no settings. */
        message("Failed to get a new partition, trying less optimal method");
        struct partition initial_partition;
#if defined(HAVE_PARMETIS) || defined(HAVE_METIS)
        initial_partition.type = INITPART_METIS_NOWEIGHT;
#else
        initial_partition.type = INITPART_VECTORIZE;
#endif
        partition_initial_partition(&initial_partition, s->e->nodeID,
                                    s->e->nr_nodes, s);
      }

      /* Re-distribute the particles to their new nodes. */
      engine_redistribute(s->e);

      /* Make the proxies. */
      engine_makeproxies(s->e);

      /* Finished with these. */
      free(oldnodeIDs);

    } else if (no_regrid && s->e != NULL) {
      /* If we have created the top-levels cells and not done an initial
       * partition (can happen when restarting), then the top-level cells
       * are not assigned to a node, we must do that and then associate the
       * particles with the cells. Note requires that
       * partition_store_celllist() was called once before, or just before
       * dumping the restart files.*/
      partition_restore_celllist(s, s->e->reparttype);

      /* Now re-distribute the particles, should just add to cells? */
      engine_redistribute(s->e);

      /* Make the proxies. */
      engine_makeproxies(s->e);
    }
#endif /* WITH_MPI */

    // message( "rebuilding upper-level cells took %.3f %s." ,
    // clocks_from_ticks(double)(getticks() - tic), clocks_getunit());

  }      /* re-build upper-level cells? */
  else { /* Otherwise, just clean up the cells. */

    /* Free the old cells, if they were allocated. */
    space_free_cells(s);
  }

  if (verbose)
    message("took %.3f %s.", clocks_from_ticks(getticks() - tic),
            clocks_getunit());
}

/**
 * @brief Re-build the cells as well as the tasks.
 *
 * @param s The #space in which to update the cells.
 * @param repartitioned Did we just repartition?
 * @param verbose Print messages to stdout or not
 */
void space_rebuild(struct space *s, int repartitioned, int verbose) {

  const ticks tic = getticks();

/* Be verbose about this. */
#ifdef SWIFT_DEBUG_CHECKS
  if (s->e->nodeID == 0 || verbose) message("(re)building space");
  fflush(stdout);
#endif

  /* Re-grid if necessary, or just re-set the cell data. */
  space_regrid(s, verbose);

  size_t nr_parts = s->nr_parts;
  size_t nr_gparts = s->nr_gparts;
  size_t nr_sparts = s->nr_sparts;
  int count_inhibited_parts = 0;
  int count_inhibited_gparts = 0;
  int count_inhibited_sparts = 0;
  struct cell *restrict cells_top = s->cells_top;
  const integertime_t ti_current = (s->e != NULL) ? s->e->ti_current : 0;

  /* Run through the particles and get their cell index. Allocates
     an index that is larger than the number of particles to avoid
     re-allocating after shuffling. */
  const size_t ind_size = s->size_parts + 100;
  int *ind = (int *)malloc(sizeof(int) * ind_size);
  if (ind == NULL) error("Failed to allocate temporary particle indices.");
  int *cell_part_counts = (int *)calloc(sizeof(int), s->nr_cells);
  if (cell_part_counts == NULL)
    error("Failed to allocate cell part count buffer.");
  if (s->size_parts > 0)
    space_parts_get_cell_index(s, ind, cell_part_counts, &count_inhibited_parts,
                               verbose);

  /* Run through the gravity particles and get their cell index. */
  const size_t gind_size = s->size_gparts + 100;
  int *gind = (int *)malloc(sizeof(int) * gind_size);
  if (gind == NULL) error("Failed to allocate temporary g-particle indices.");
  int *cell_gpart_counts = (int *)calloc(sizeof(int), s->nr_cells);
  if (cell_gpart_counts == NULL)
    error("Failed to allocate cell gpart count buffer.");
  if (s->size_gparts > 0)
    space_gparts_get_cell_index(s, gind, cell_gpart_counts,
                                &count_inhibited_gparts, verbose);

  /* Run through the star particles and get their cell index. */
  const size_t sind_size = s->size_sparts + 100;
  int *sind = (int *)malloc(sizeof(int) * sind_size);
  if (sind == NULL) error("Failed to allocate temporary s-particle indices.");
  int *cell_spart_counts = (int *)calloc(sizeof(int), s->nr_cells);
  if (cell_spart_counts == NULL)
    error("Failed to allocate cell gpart count buffer.");
  if (s->size_sparts > 0)
    space_sparts_get_cell_index(s, sind, cell_spart_counts,
                                &count_inhibited_sparts, verbose);

#ifdef SWIFT_DEBUG_CHECKS
  if (repartitioned && count_inhibited_parts)
    error("We just repartitioned but still found inhibited parts.");
  if (repartitioned && count_inhibited_sparts)
    error("We just repartitioned but still found inhibited sparts.");
  if (repartitioned && count_inhibited_gparts)
    error("We just repartitioned but still found inhibited gparts.");
#endif

  const int local_nodeID = s->e->nodeID;

  /* Move non-local parts and inhibited parts to the end of the list. */
  if (!repartitioned && (s->e->nr_nodes > 1 || count_inhibited_parts > 0)) {
    for (size_t k = 0; k < nr_parts; /* void */) {

      /* Inhibited particle or foreign particle */
      if (ind[k] == -1 || cells_top[ind[k]].nodeID != local_nodeID) {

        /* One fewer particle */
        nr_parts -= 1;

        /* Swap the particle */
        memswap(&s->parts[k], &s->parts[nr_parts], sizeof(struct part));

        /* Swap the link with the gpart */
        if (s->parts[k].gpart != NULL) {
          s->parts[k].gpart->id_or_neg_offset = -k;
        }
        if (s->parts[nr_parts].gpart != NULL) {
          s->parts[nr_parts].gpart->id_or_neg_offset = -nr_parts;
        }

        /* Swap the xpart */
        memswap(&s->xparts[k], &s->xparts[nr_parts], sizeof(struct xpart));
        /* Swap the index */
        memswap(&ind[k], &ind[nr_parts], sizeof(int));

      } else {
        /* Increment when not exchanging otherwise we need to retest "k".*/
        k++;
      }
    }
  }

#ifdef SWIFT_DEBUG_CHECKS
  /* Check that all parts are in the correct places. */
  int check_count_inhibited_part = 0;
  for (size_t k = 0; k < nr_parts; k++) {
    if (ind[k] == -1 || cells_top[ind[k]].nodeID != local_nodeID) {
      error("Failed to move all non-local parts to send list");
    }
  }
  for (size_t k = nr_parts; k < s->nr_parts; k++) {
    if (ind[k] != -1 && cells_top[ind[k]].nodeID == local_nodeID) {
      error("Failed to remove local parts from send list");
    }
    if (ind[k] == -1) ++check_count_inhibited_part;
  }
  if (check_count_inhibited_part != count_inhibited_parts)
    error("Counts of inhibited particles do not match!");
#endif /* SWIFT_DEBUG_CHECKS */

  /* Move non-local sparts and inhibited sparts to the end of the list. */
  if (!repartitioned && (s->e->nr_nodes > 1 || count_inhibited_sparts > 0)) {
    for (size_t k = 0; k < nr_sparts; /* void */) {

      /* Inhibited particle or foreign particle */
      if (sind[k] == -1 || cells_top[sind[k]].nodeID != local_nodeID) {

        /* One fewer particle */
        nr_sparts -= 1;

        /* Swap the particle */
        memswap(&s->sparts[k], &s->sparts[nr_sparts], sizeof(struct spart));

        /* Swap the link with the gpart */
        if (s->sparts[k].gpart != NULL) {
          s->sparts[k].gpart->id_or_neg_offset = -k;
        }
        if (s->sparts[nr_sparts].gpart != NULL) {
          s->sparts[nr_sparts].gpart->id_or_neg_offset = -nr_sparts;
        }

        /* Swap the index */
        memswap(&sind[k], &sind[nr_sparts], sizeof(int));

      } else {
        /* Increment when not exchanging otherwise we need to retest "k".*/
        k++;
      }
    }
  }

#ifdef SWIFT_DEBUG_CHECKS
  /* Check that all sparts are in the correct place. */
  int check_count_inhibited_spart = 0;
  for (size_t k = 0; k < nr_sparts; k++) {
    if (sind[k] == -1 || cells_top[sind[k]].nodeID != local_nodeID) {
      error("Failed to move all non-local sparts to send list");
    }
  }
  for (size_t k = nr_sparts; k < s->nr_sparts; k++) {
    if (sind[k] != -1 && cells_top[sind[k]].nodeID == local_nodeID) {
      error("Failed to remove local sparts from send list");
    }
    if (sind[k] == -1) ++check_count_inhibited_spart;
  }
  if (check_count_inhibited_spart != count_inhibited_sparts)
    error("Counts of inhibited s-particles do not match!");
#endif /* SWIFT_DEBUG_CHECKS */

  /* Move non-local gparts and inhibited parts to the end of the list. */
  if (!repartitioned && (s->e->nr_nodes > 1 || count_inhibited_gparts > 0)) {
    for (size_t k = 0; k < nr_gparts; /* void */) {

      /* Inhibited particle or foreign particle */
      if (gind[k] == -1 || cells_top[gind[k]].nodeID != local_nodeID) {

        /* One fewer particle */
        nr_gparts -= 1;

        /* Swap the particle */
        memswap(&s->gparts[k], &s->gparts[nr_gparts], sizeof(struct gpart));

        /* Swap the link with part/spart */
        if (s->gparts[k].type == swift_type_gas) {
          s->parts[-s->gparts[k].id_or_neg_offset].gpart = &s->gparts[k];
        } else if (s->gparts[k].type == swift_type_stars) {
          s->sparts[-s->gparts[k].id_or_neg_offset].gpart = &s->gparts[k];
        }
        if (s->gparts[nr_gparts].type == swift_type_gas) {
          s->parts[-s->gparts[nr_gparts].id_or_neg_offset].gpart =
              &s->gparts[nr_gparts];
        } else if (s->gparts[nr_gparts].type == swift_type_stars) {
          s->sparts[-s->gparts[nr_gparts].id_or_neg_offset].gpart =
              &s->gparts[nr_gparts];
        }

        /* Swap the index */
        memswap(&gind[k], &gind[nr_gparts], sizeof(int));
      } else {
        /* Increment when not exchanging otherwise we need to retest "k".*/
        k++;
      }
    }
  }

#ifdef SWIFT_DEBUG_CHECKS
  /* Check that all gparts are in the correct place. */
  int check_count_inhibited_gpart = 0;
  for (size_t k = 0; k < nr_gparts; k++) {
    if (gind[k] == -1 || cells_top[gind[k]].nodeID != local_nodeID) {
      error("Failed to move all non-local gparts to send list");
    }
  }
  for (size_t k = nr_gparts; k < s->nr_gparts; k++) {
    if (gind[k] != -1 && cells_top[gind[k]].nodeID == local_nodeID) {
      error("Failed to remove local gparts from send list");
    }
    if (gind[k] == -1) ++check_count_inhibited_gpart;
  }
  if (check_count_inhibited_gpart != count_inhibited_gparts)
    error("Counts of inhibited g-particles do not match!");
#endif /* SWIFT_DEBUG_CHECKS */

#ifdef WITH_MPI

  /* Exchange the strays, note that this potentially re-allocates
     the parts arrays. This can be skipped if we just repartitioned aspace
     there should be no strays */
  if (!repartitioned) {

    size_t nr_parts_exchanged = s->nr_parts - nr_parts;
    size_t nr_gparts_exchanged = s->nr_gparts - nr_gparts;
    size_t nr_sparts_exchanged = s->nr_sparts - nr_sparts;
    engine_exchange_strays(s->e, nr_parts, &ind[nr_parts], &nr_parts_exchanged,
                           nr_gparts, &gind[nr_gparts], &nr_gparts_exchanged,
                           nr_sparts, &sind[nr_sparts], &nr_sparts_exchanged);

    /* Set the new particle counts. */
    s->nr_parts = nr_parts + nr_parts_exchanged;
    s->nr_gparts = nr_gparts + nr_gparts_exchanged;
    s->nr_sparts = nr_sparts + nr_sparts_exchanged;
  } else {
#ifdef SWIFT_DEBUG_CHECKS
    if (s->nr_parts != nr_parts)
      error("Number of parts changing after repartition");
    if (s->nr_sparts != nr_sparts)
      error("Number of sparts changing after repartition");
    if (s->nr_gparts != nr_gparts)
      error("Number of gparts changing after repartition");
#endif
  }

  /* Clear non-local cell counts. */
  for (int k = 0; k < s->nr_cells; k++) {
    if (s->cells_top[k].nodeID != local_nodeID) {
      cell_part_counts[k] = 0;
      cell_spart_counts[k] = 0;
      cell_gpart_counts[k] = 0;
    }
  }

  /* Re-allocate the index array for the parts if needed.. */
  if (s->nr_parts + 1 > ind_size) {
    int *ind_new;
    if ((ind_new = (int *)malloc(sizeof(int) * (s->nr_parts + 1))) == NULL)
      error("Failed to allocate temporary particle indices.");
    memcpy(ind_new, ind, sizeof(int) * nr_parts);
    free(ind);
    ind = ind_new;
  }

  /* Re-allocate the index array for the sparts if needed.. */
  if (s->nr_sparts + 1 > sind_size) {
    int *sind_new;
    if ((sind_new = (int *)malloc(sizeof(int) * (s->nr_sparts + 1))) == NULL)
      error("Failed to allocate temporary s-particle indices.");
    memcpy(sind_new, sind, sizeof(int) * nr_sparts);
    free(sind);
    sind = sind_new;
  }

  const int cdim[3] = {s->cdim[0], s->cdim[1], s->cdim[2]};
  const double ih[3] = {s->iwidth[0],