engine.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)
 *                    Angus Lepper (angus.lepper@ed.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 <sched.h>
#include <stdbool.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <unistd.h>

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

#ifdef HAVE_LIBNUMA
#include <numa.h>
#endif

/* Load the profiler header, if needed. */
#ifdef WITH_PROFILER
#include <gperftools/profiler.h>
#endif

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

/* Local headers. */
#include "active.h"
#include "atomic.h"
#include "cell.h"
#include "chemistry.h"
#include "clocks.h"
#include "cooling.h"
#include "cosmology.h"
#include "cycle.h"
#include "debug.h"
#include "equation_of_state.h"
#include "error.h"
#include "gravity.h"
#include "gravity_cache.h"
#include "hydro.h"
#include "logger.h"
#include "logger_io.h"
#include "map.h"
#include "memswap.h"
#include "minmax.h"
#include "outputlist.h"
#include "parallel_io.h"
#include "part.h"
#include "partition.h"
#include "profiler.h"
#include "proxy.h"
#include "restart.h"
#include "runner.h"
#include "serial_io.h"
#include "single_io.h"
#include "sort_part.h"
#include "sourceterms.h"
#include "stars_io.h"
#include "statistics.h"
#include "timers.h"
#include "tools.h"
#include "units.h"
#include "velociraptor_interface.h"
#include "version.h"

/* Particle cache size. */
#define CACHE_SIZE 512

const char *engine_policy_names[] = {"none",
                                     "rand",
                                     "steal",
                                     "keep",
                                     "block",
                                     "cpu tight",
                                     "mpi",
                                     "numa affinity",
                                     "hydro",
                                     "self gravity",
                                     "external gravity",
                                     "cosmological integration",
                                     "drift everything",
                                     "reconstruct multi-poles",
                                     "cooling",
                                     "sourceterms",
                                     "stars",
                                     "structure finding",
                                     "star formation",
                                     "feedback"};

/** The rank of the engine as a global variable (for messages). */
int engine_rank;

/**
 * @brief Data collected from the cells at the end of a time-step
 */
struct end_of_step_data {

  size_t updated, g_updated, s_updated;
  size_t inhibited, g_inhibited, s_inhibited;
  integertime_t ti_hydro_end_min, ti_hydro_end_max, ti_hydro_beg_max;
  integertime_t ti_gravity_end_min, ti_gravity_end_max, ti_gravity_beg_max;
  struct engine *e;
};

/**
 * @brief Link a density/force task to a cell.
 *
 * @param e The #engine.
 * @param l A pointer to the #link, will be modified atomically.
 * @param t The #task.
 *
 * @return The new #link pointer.
 */
void engine_addlink(struct engine *e, struct link **l, struct task *t) {

  /* Get the next free link. */
  const size_t ind = atomic_inc(&e->nr_links);
  if (ind >= e->size_links) {
    error("Link table overflow.");
  }
  struct link *res = &e->links[ind];

  /* Set it atomically. */
  res->t = t;
  res->next = atomic_swap(l, res);
}

/**
 * @brief Recursively add non-implicit star ghost tasks to a cell hierarchy.
 */
void engine_add_stars_ghosts(struct engine *e, struct cell *c,
                             struct task *stars_ghost_in,
                             struct task *stars_ghost_out) {

  /* If we have reached the leaf OR have to few particles to play with*/
  if (!c->split || c->stars.count < engine_max_sparts_per_ghost) {

    /* Add the ghost task and its dependencies */
    struct scheduler *s = &e->sched;
    c->stars.ghost = scheduler_addtask(s, task_type_stars_ghost,
                                       task_subtype_none, 0, 0, c, NULL);
    scheduler_addunlock(s, stars_ghost_in, c->stars.ghost);
    scheduler_addunlock(s, c->stars.ghost, stars_ghost_out);
  } else {
    /* Keep recursing */
    for (int k = 0; k < 8; k++)
      if (c->progeny[k] != NULL)
        engine_add_stars_ghosts(e, c->progeny[k], stars_ghost_in,
                                stars_ghost_out);
  }
}

/**
 * @brief Recursively add non-implicit ghost tasks to a cell hierarchy.
 */
void engine_add_ghosts(struct engine *e, struct cell *c, struct task *ghost_in,
                       struct task *ghost_out) {

  /* If we have reached the leaf OR have to few particles to play with*/
  if (!c->split || c->hydro.count < engine_max_parts_per_ghost) {

    /* Add the ghost task and its dependencies */
    struct scheduler *s = &e->sched;
    c->hydro.ghost =
        scheduler_addtask(s, task_type_ghost, task_subtype_none, 0, 0, c, NULL);
    scheduler_addunlock(s, ghost_in, c->hydro.ghost);
    scheduler_addunlock(s, c->hydro.ghost, ghost_out);
  } else {
    /* Keep recursing */
    for (int k = 0; k < 8; k++)
      if (c->progeny[k] != NULL)
        engine_add_ghosts(e, c->progeny[k], ghost_in, ghost_out);
  }
}

/**
 * @brief Generate the hydro hierarchical tasks for a hierarchy of cells -
 * i.e. all the O(Npart) tasks -- timestep version
 *
 * Tasks are only created here. The dependencies will be added later on.
 *
 * Note that there is no need to recurse below the super-cell. Note also
 * that we only add tasks if the relevant particles are present in the cell.
 *
 * @param e The #engine.
 * @param c The #cell.
 */
void engine_make_hierarchical_tasks_common(struct engine *e, struct cell *c) {

  struct scheduler *s = &e->sched;
  const int is_with_cooling = (e->policy & engine_policy_cooling);
  const int is_with_star_formation = (e->policy & engine_policy_star_formation);

  /* Are we in a super-cell ? */
  if (c->super == c) {

    /* Local tasks only... */
    if (c->nodeID == e->nodeID) {

      /* Add the two half kicks */
      c->kick1 = scheduler_addtask(s, task_type_kick1, task_subtype_none, 0, 0,
                                   c, NULL);

#if defined(WITH_LOGGER)
      c->logger = scheduler_addtask(s, task_type_logger, task_subtype_none, 0,
                                    0, c, NULL);
#endif

      c->kick2 = scheduler_addtask(s, task_type_kick2, task_subtype_none, 0, 0,
                                   c, NULL);

      /* Add the time-step calculation task and its dependency */
      c->timestep = scheduler_addtask(s, task_type_timestep, task_subtype_none,
                                      0, 0, c, NULL);

      /* Add the task finishing the force calculation */
      c->end_force = scheduler_addtask(s, task_type_end_force,
                                       task_subtype_none, 0, 0, c, NULL);

      if (is_with_cooling) {

        c->hydro.cooling = scheduler_addtask(s, task_type_cooling,
                                             task_subtype_none, 0, 0, c, NULL);

        scheduler_addunlock(s, c->end_force, c->hydro.cooling);
        scheduler_addunlock(s, c->hydro.cooling, c->kick2);

      } else {
        scheduler_addunlock(s, c->end_force, c->kick2);
      }

      if (is_with_star_formation) {

        c->hydro.star_formation = scheduler_addtask(
            s, task_type_star_formation, task_subtype_none, 0, 0, c, NULL);

        scheduler_addunlock(s, c->kick2, c->hydro.star_formation);
        scheduler_addunlock(s, c->hydro.star_formation, c->timestep);

      } else {
        scheduler_addunlock(s, c->kick2, c->timestep);
      }
      scheduler_addunlock(s, c->timestep, c->kick1);

#if defined(WITH_LOGGER)
      scheduler_addunlock(s, c->kick1, c->logger);
#endif
    }

  } else { /* We are above the super-cell so need to go deeper */

    /* Recurse. */
    if (c->split)
      for (int k = 0; k < 8; k++)
        if (c->progeny[k] != NULL)
          engine_make_hierarchical_tasks_common(e, c->progeny[k]);
  }
}

/**
 * @brief Generate the hydro hierarchical tasks for a hierarchy of cells -
 * i.e. all the O(Npart) tasks -- hydro version
 *
 * Tasks are only created here. The dependencies will be added later on.
 *
 * Note that there is no need to recurse below the super-cell. Note also
 * that we only add tasks if the relevant particles are present in the cell.
 *
 * @param e The #engine.
 * @param c The #cell.
 */
void engine_make_hierarchical_tasks_hydro(struct engine *e, struct cell *c) {

  struct scheduler *s = &e->sched;
  const int is_with_sourceterms = (e->policy & engine_policy_sourceterms);

  /* Are we in a super-cell ? */
  if (c->hydro.super == c) {

    /* Add the sort task. */
    c->hydro.sorts =
        scheduler_addtask(s, task_type_sort, task_subtype_none, 0, 0, c, NULL);

    /* Local tasks only... */
    if (c->nodeID == e->nodeID) {

      /* Add the drift task. */
      c->hydro.drift = scheduler_addtask(s, task_type_drift_part,
                                         task_subtype_none, 0, 0, c, NULL);

      /* Generate the ghost tasks. */
      c->hydro.ghost_in =
          scheduler_addtask(s, task_type_ghost_in, task_subtype_none, 0,
                            /* implicit = */ 1, c, NULL);
      c->hydro.ghost_out =
          scheduler_addtask(s, task_type_ghost_out, task_subtype_none, 0,
                            /* implicit = */ 1, c, NULL);
      engine_add_ghosts(e, c, c->hydro.ghost_in, c->hydro.ghost_out);

#ifdef EXTRA_HYDRO_LOOP
      /* Generate the extra ghost task. */
      c->hydro.extra_ghost = scheduler_addtask(
          s, task_type_extra_ghost, task_subtype_none, 0, 0, c, NULL);
#endif

      /* add source terms */
      if (is_with_sourceterms) {
        c->sourceterms = scheduler_addtask(s, task_type_sourceterms,
                                           task_subtype_none, 0, 0, c, NULL);
      }
    }

  } else { /* We are above the super-cell so need to go deeper */

    /* Recurse. */
    if (c->split)
      for (int k = 0; k < 8; k++)
        if (c->progeny[k] != NULL)
          engine_make_hierarchical_tasks_hydro(e, c->progeny[k]);
  }
}

/**
 * @brief Generate the hydro hierarchical tasks for a hierarchy of cells -
 * i.e. all the O(Npart) tasks -- gravity version
 *
 * Tasks are only created here. The dependencies will be added later on.
 *
 * Note that there is no need to recurse below the super-cell. Note also
 * that we only add tasks if the relevant particles are present in the cell.
 *
 * @param e The #engine.
 * @param c The #cell.
 */
void engine_make_hierarchical_tasks_gravity(struct engine *e, struct cell *c) {

  struct scheduler *s = &e->sched;
  const int periodic = e->s->periodic;
  const int is_self_gravity = (e->policy & engine_policy_self_gravity);

  /* Are we in a super-cell ? */
  if (c->grav.super == c) {

    /* Local tasks only... */
    if (c->nodeID == e->nodeID) {

      c->grav.drift = scheduler_addtask(s, task_type_drift_gpart,
                                        task_subtype_none, 0, 0, c, NULL);

      if (is_self_gravity) {

        /* Initialisation of the multipoles */
        c->grav.init = scheduler_addtask(s, task_type_init_grav,
                                         task_subtype_none, 0, 0, c, NULL);

        /* Gravity non-neighbouring pm calculations */
        c->grav.long_range = scheduler_addtask(
            s, task_type_grav_long_range, task_subtype_none, 0, 0, c, NULL);

        /* Gravity recursive down-pass */
        c->grav.down = scheduler_addtask(s, task_type_grav_down,
                                         task_subtype_none, 0, 0, c, NULL);

        /* Implicit tasks for the up and down passes */
        c->grav.init_out = scheduler_addtask(s, task_type_init_grav_out,
                                             task_subtype_none, 0, 1, c, NULL);
        c->grav.down_in = scheduler_addtask(s, task_type_grav_down_in,
                                            task_subtype_none, 0, 1, c, NULL);

        /* Gravity mesh force propagation */
        if (periodic)
          c->grav.mesh = scheduler_addtask(s, task_type_grav_mesh,
                                           task_subtype_none, 0, 0, c, NULL);

        if (periodic) scheduler_addunlock(s, c->grav.drift, c->grav.mesh);
        if (periodic) scheduler_addunlock(s, c->grav.mesh, c->grav.down);
        scheduler_addunlock(s, c->grav.init, c->grav.long_range);
        scheduler_addunlock(s, c->grav.long_range, c->grav.down);
        scheduler_addunlock(s, c->grav.down, c->super->end_force);

        /* Link in the implicit tasks */
        scheduler_addunlock(s, c->grav.init, c->grav.init_out);
        scheduler_addunlock(s, c->grav.down_in, c->grav.down);
      }
    }
  }

  /* We are below the super-cell but not below the maximal splitting depth */
  else if (c->grav.super != NULL && c->depth < space_subdepth_grav) {

    /* Local tasks only... */
    if (c->nodeID == e->nodeID) {

      if (is_self_gravity) {

        c->grav.init_out = scheduler_addtask(s, task_type_init_grav_out,
                                             task_subtype_none, 0, 1, c, NULL);

        c->grav.down_in = scheduler_addtask(s, task_type_grav_down_in,
                                            task_subtype_none, 0, 1, c, NULL);

        scheduler_addunlock(s, c->parent->grav.init_out, c->grav.init_out);
        scheduler_addunlock(s, c->grav.down_in, c->parent->grav.down_in);
      }
    }
  }

  /* Recurse but not below the maximal splitting depth */
  if (c->split && c->depth <= space_subdepth_grav)
    for (int k = 0; k < 8; k++)
      if (c->progeny[k] != NULL)
        engine_make_hierarchical_tasks_gravity(e, c->progeny[k]);
}

/**
 * @brief Generate the stars hierarchical tasks for a hierarchy of cells -
 * i.e. all the O(Npart) tasks -- star version
 *
 * Tasks are only created here. The dependencies will be added later on.
 *
 * Note that there is no need to recurse below the super-cell. Note also
 * that we only add tasks if the relevant particles are present in the cell.
 *
 * @param e The #engine.
 * @param c The #cell.
 */
void engine_make_hierarchical_tasks_stars(struct engine *e, struct cell *c) {

  struct scheduler *s = &e->sched;

  /* Are we in a super-cell ? */
  if (c->super == c) {

    /* Local tasks only... */
    if (c->nodeID == e->nodeID) {

      /* Generate the ghost tasks. */
      c->stars.ghost_in =
          scheduler_addtask(s, task_type_stars_ghost_in, task_subtype_none, 0,
                            /* implicit = */ 1, c, NULL);
      c->stars.ghost_out =
          scheduler_addtask(s, task_type_stars_ghost_out, task_subtype_none, 0,
                            /* implicit = */ 1, c, NULL);
      engine_add_stars_ghosts(e, c, c->stars.ghost_in, c->stars.ghost_out);
    }
  } else { /* We are above the super-cell so need to go deeper */

    /* Recurse. */
    if (c->split)
      for (int k = 0; k < 8; k++)
        if (c->progeny[k] != NULL)
          engine_make_hierarchical_tasks_stars(e, c->progeny[k]);
  }
}

void engine_make_hierarchical_tasks_mapper(void *map_data, int num_elements,
                                           void *extra_data) {
  struct engine *e = (struct engine *)extra_data;
  const int is_with_hydro = (e->policy & engine_policy_hydro);
  const int is_with_self_gravity = (e->policy & engine_policy_self_gravity);
  const int is_with_external_gravity =
      (e->policy & engine_policy_external_gravity);
  const int is_with_feedback = (e->policy & engine_policy_feedback);

  for (int ind = 0; ind < num_elements; ind++) {
    struct cell *c = &((struct cell *)map_data)[ind];
    /* Make the common tasks (time integration) */
    engine_make_hierarchical_tasks_common(e, c);
    /* Add the hydro stuff */
    if (is_with_hydro) engine_make_hierarchical_tasks_hydro(e, c);
    /* And the gravity stuff */
    if (is_with_self_gravity || is_with_external_gravity)
      engine_make_hierarchical_tasks_gravity(e, c);
    if (is_with_feedback) engine_make_hierarchical_tasks_stars(e, c);
  }
}

#ifdef WITH_MPI
/**
 * Do the exchange of one type of particles with all the other nodes.
 *
 * @param counts 2D array with the counts of particles to exchange with
 *               each other node.
 * @param parts the particle data to exchange
 * @param new_nr_parts the number of particles this node will have after all
 *                     exchanges have completed.
 * @param sizeofparts sizeof the particle struct.
 * @param alignsize the memory alignment required for this particle type.
 * @param mpi_type the MPI_Datatype for these particles.
 * @param nr_nodes the number of nodes to exchange with.
 * @param nodeID the id of this node.
 *
 * @result new particle data constructed from all the exchanges with the
 *         given alignment.
 */
static void *engine_do_redistribute(int *counts, char *parts,
                                    size_t new_nr_parts, size_t sizeofparts,
                                    size_t alignsize, MPI_Datatype mpi_type,
                                    int nr_nodes, int nodeID) {

  /* Allocate a new particle array with some extra margin */
  char *parts_new = NULL;
  if (posix_memalign(
          (void **)&parts_new, alignsize,
          sizeofparts * new_nr_parts * engine_redistribute_alloc_margin) != 0)
    error("Failed to allocate new particle data.");

  /* Prepare MPI requests for the asynchronous communications */
  MPI_Request *reqs;
  if ((reqs = (MPI_Request *)malloc(sizeof(MPI_Request) * 2 * nr_nodes)) ==
      NULL)
    error("Failed to allocate MPI request list.");

  /* Only send and receive only "chunk" particles per request. So we need to
   * loop as many times as necessary here. Make 2Gb/sizeofparts so we only
   * send 2Gb packets. */
  const int chunk = INT_MAX / sizeofparts;
  int sent = 0;
  int recvd = 0;

  int activenodes = 1;
  while (activenodes) {

    for (int k = 0; k < 2 * nr_nodes; k++) reqs[k] = MPI_REQUEST_NULL;

    /* Emit the sends and recvs for the data. */
    size_t offset_send = sent;
    size_t offset_recv = recvd;
    activenodes = 0;

    for (int k = 0; k < nr_nodes; k++) {

      /* Indices in the count arrays of the node of interest */
      const int ind_send = nodeID * nr_nodes + k;
      const int ind_recv = k * nr_nodes + nodeID;

      /* Are we sending any data this loop? */
      int sending = counts[ind_send] - sent;
      if (sending > 0) {
        activenodes++;
        if (sending > chunk) sending = chunk;

        /* If the send and receive is local then just copy. */
        if (k == nodeID) {
          int receiving = counts[ind_recv] - recvd;
          if (receiving > chunk) receiving = chunk;
          memcpy(&parts_new[offset_recv * sizeofparts],
                 &parts[offset_send * sizeofparts], sizeofparts * receiving);
        } else {
          /* Otherwise send it. */
          int res =
              MPI_Isend(&parts[offset_send * sizeofparts], sending, mpi_type, k,
                        ind_send, MPI_COMM_WORLD, &reqs[2 * k + 0]);
          if (res != MPI_SUCCESS)
            mpi_error(res, "Failed to isend parts to node %i.", k);
        }
      }

      /* If we're sending to this node, then move past it to next. */
      if (counts[ind_send] > 0) offset_send += counts[ind_send];

      /* Are we receiving any data from this node? Note already done if coming
       * from this node. */
      if (k != nodeID) {
        int receiving = counts[ind_recv] - recvd;
        if (receiving > 0) {
          activenodes++;
          if (receiving > chunk) receiving = chunk;
          int res = MPI_Irecv(&parts_new[offset_recv * sizeofparts], receiving,
                              mpi_type, k, ind_recv, MPI_COMM_WORLD,
                              &reqs[2 * k + 1]);
          if (res != MPI_SUCCESS)
            mpi_error(res, "Failed to emit irecv of parts from node %i.", k);
        }
      }

      /* If we're receiving from this node, then move past it to next. */
      if (counts[ind_recv] > 0) offset_recv += counts[ind_recv];
    }

    /* Wait for all the sends and recvs to tumble in. */
    MPI_Status stats[2 * nr_nodes];
    int res;
    if ((res = MPI_Waitall(2 * nr_nodes, reqs, stats)) != MPI_SUCCESS) {
      for (int k = 0; k < 2 * nr_nodes; k++) {
        char buff[MPI_MAX_ERROR_STRING];
        MPI_Error_string(stats[k].MPI_ERROR, buff, &res);
        message("request from source %i, tag %i has error '%s'.",
                stats[k].MPI_SOURCE, stats[k].MPI_TAG, buff);
      }
      error("Failed during waitall for part data.");
    }

    /* Move to next chunks. */
    sent += chunk;
    recvd += chunk;
  }

  /* Free temps. */
  free(reqs);

  /* And return new memory. */
  return parts_new;
}
#endif

#ifdef WITH_MPI /* redist_mapper */

/* Support for engine_redistribute threadpool dest mappers. */
struct redist_mapper_data {
  int *counts;
  int *dest;
  int nodeID;
  int nr_nodes;
  struct cell *cells;
  struct space *s;
  void *base;
};

/* Generic function for accumulating counts for TYPE parts. Note
 * we use a local counts array to avoid the atomic_add in the parts
 * loop. */
#define ENGINE_REDISTRIBUTE_DEST_MAPPER(TYPE)                              \
  engine_redistribute_dest_mapper_##TYPE(void *map_data, int num_elements, \
                                         void *extra_data) {               \
    struct TYPE *parts = (struct TYPE *)map_data;                          \
    struct redist_mapper_data *mydata =                                    \
        (struct redist_mapper_data *)extra_data;                           \
    struct space *s = mydata->s;                                           \
    int *dest =                                                            \
        mydata->dest + (ptrdiff_t)(parts - (struct TYPE *)mydata->base);   \
    int *lcounts = NULL;                                                   \
    if ((lcounts = (int *)calloc(                                          \
             sizeof(int), mydata->nr_nodes * mydata->nr_nodes)) == NULL)   \
      error("Failed to allocate counts thread-specific buffer");           \
    for (int k = 0; k < num_elements; k++) {                               \
      for (int j = 0; j < 3; j++) {                                        \
        if (parts[k].x[j] < 0.0)                                           \
          parts[k].x[j] += s->dim[j];                                      \
        else if (parts[k].x[j] >= s->dim[j])                               \
          parts[k].x[j] -= s->dim[j];                                      \
      }                                                                    \
      const int cid = cell_getid(s->cdim, parts[k].x[0] * s->iwidth[0],    \
                                 parts[k].x[1] * s->iwidth[1],             \
                                 parts[k].x[2] * s->iwidth[2]);            \
      dest[k] = s->cells_top[cid].nodeID;                                  \
      size_t ind = mydata->nodeID * mydata->nr_nodes + dest[k];            \
      lcounts[ind] += 1;                                                   \
    }                                                                      \
    for (int k = 0; k < (mydata->nr_nodes * mydata->nr_nodes); k++)        \
      atomic_add(&mydata->counts[k], lcounts[k]);                          \
    free(lcounts);                                                         \
  }

/**
 * @brief Accumulate the counts of particles per cell.
 * Threadpool helper for accumulating the counts of particles per cell.
 *
 * part version.
 */
static void ENGINE_REDISTRIBUTE_DEST_MAPPER(part);

/**
 * @brief Accumulate the counts of star particles per cell.
 * Threadpool helper for accumulating the counts of particles per cell.
 *
 * spart version.
 */
static void ENGINE_REDISTRIBUTE_DEST_MAPPER(spart);

/**
 * @brief Accumulate the counts of gravity particles per cell.
 * Threadpool helper for accumulating the counts of particles per cell.
 *
 * gpart version.
 */
static void ENGINE_REDISTRIBUTE_DEST_MAPPER(gpart);

#endif /* redist_mapper_data */

#ifdef WITH_MPI /* savelink_mapper_data */

/* Support for saving the linkage between gparts and parts/sparts. */
struct savelink_mapper_data {
  int nr_nodes;
  int *counts;
  void *parts;
  int nodeID;
};

/**
 * @brief Save the offset of each gravity partner of a part or spart.
 *
 * The offset is from the start of the sorted particles to be sent to a node.
 * This is possible as parts without gravity partners have a positive id.
 * These offsets are used to restore the pointers on the receiving node.
 *
 * CHECKS should be eliminated as dead code when optimizing.
 */
#define ENGINE_REDISTRIBUTE_SAVELINK_MAPPER(TYPE, CHECKS)                      \
  engine_redistribute_savelink_mapper_##TYPE(void *map_data, int num_elements, \
                                             void *extra_data) {               \
    int *nodes = (int *)map_data;                                              \
    struct savelink_mapper_data *mydata =                                      \
        (struct savelink_mapper_data *)extra_data;                             \
    int nodeID = mydata->nodeID;                                               \
    int nr_nodes = mydata->nr_nodes;                                           \
    int *counts = mydata->counts;                                              \
    struct TYPE *parts = (struct TYPE *)mydata->parts;                         \
                                                                               \
    for (int j = 0; j < num_elements; j++) {                                   \
      int node = nodes[j];                                                     \
      int count = 0;                                                           \
      size_t offset = 0;                                                       \
      for (int i = 0; i < node; i++) offset += counts[nodeID * nr_nodes + i];  \
                                                                               \
      for (int k = 0; k < counts[nodeID * nr_nodes + node]; k++) {             \
        if (parts[k + offset].gpart != NULL) {                                 \
          if (CHECKS)                                                          \
            if (parts[k + offset].gpart->id_or_neg_offset > 0)                 \
              error("Trying to link a partnerless " #TYPE "!");                \
          parts[k + offset].gpart->id_or_neg_offset = -count;                  \
          count++;                                                             \
        }                                                                      \
      }                                                                        \
    }                                                                          \
  }

/**
 * @brief Save position of part-gpart links.
 * Threadpool helper for accumulating the counts of particles per cell.
 */
#ifdef SWIFT_DEBUG_CHECKS
static void ENGINE_REDISTRIBUTE_SAVELINK_MAPPER(part, 1);
#else
static void ENGINE_REDISTRIBUTE_SAVELINK_MAPPER(part, 0);
#endif

/**
 * @brief Save position of spart-gpart links.
 * Threadpool helper for accumulating the counts of particles per cell.
 */
#ifdef SWIFT_DEBUG_CHECKS
static void ENGINE_REDISTRIBUTE_SAVELINK_MAPPER(spart, 1);
#else
static void ENGINE_REDISTRIBUTE_SAVELINK_MAPPER(spart, 0);
#endif

#endif /* savelink_mapper_data */

#ifdef WITH_MPI /* relink_mapper_data */

/* Support for relinking parts, gparts and sparts after moving between nodes. */
struct relink_mapper_data {
  int nodeID;
  int nr_nodes;
  int *counts;
  int *s_counts;
  int *g_counts;
  struct space *s;
};

/**
 * @brief Restore the part/gpart and spart/gpart links for a list of nodes.
 *
 * @param map_data address of nodes to process.
 * @param num_elements the number nodes to process.
 * @param extra_data additional data defining the context (a
 * relink_mapper_data).
 */
static void engine_redistribute_relink_mapper(void *map_data, int num_elements,
                                              void *extra_data) {

  int *nodes = (int *)map_data;
  struct relink_mapper_data *mydata = (struct relink_mapper_data *)extra_data;

  int nodeID = mydata->nodeID;
  int nr_nodes = mydata->nr_nodes;
  int *counts = mydata->counts;
  int *g_counts = mydata->g_counts;
  int *s_counts = mydata->s_counts;
  struct space *s = mydata->s;

  for (int i = 0; i < num_elements; i++) {

    int node = nodes[i];

    /* Get offsets to correct parts of the counts arrays for this node. */
    size_t offset_parts = 0;
    size_t offset_gparts = 0;
    size_t offset_sparts = 0;
    for (int n = 0; n < node; n++) {
      int ind_recv = n * nr_nodes + nodeID;
      offset_parts += counts[ind_recv];
      offset_gparts += g_counts[ind_recv];
      offset_sparts += s_counts[ind_recv];
    }

    /* Number of gparts sent from this node. */
    int ind_recv = node * nr_nodes + nodeID;
    const size_t count_gparts = g_counts[ind_recv];

    /* Loop over the gparts received from this node */
    for (size_t k = offset_gparts; k < offset_gparts + count_gparts; k++) {

      /* Does this gpart have a gas partner ? */
      if (s->gparts[k].type == swift_type_gas) {

        const ptrdiff_t partner_index =
            offset_parts - s->gparts[k].id_or_neg_offset;

        /* Re-link */
        s->gparts[k].id_or_neg_offset = -partner_index;
        s->parts[partner_index].gpart = &s->gparts[k];
      }

      /* Does this gpart have a star partner ? */
      else if (s->gparts[k].type == swift_type_stars) {

        const ptrdiff_t partner_index =
            offset_sparts - s->gparts[k].id_or_neg_offset;

        /* Re-link */
        s->gparts[k].id_or_neg_offset = -partner_index;
        s->sparts[partner_index].gpart = &s->gparts[k];
      }
    }
  }
}

#endif /* relink_mapper_data */

/**
 * @brief Redistribute the particles amongst the nodes according
 *      to their cell's node IDs.
 *
 * The strategy here is as follows:
 * 1) Each node counts the number of particles it has to send to each other
 * node.
 * 2) The number of particles of each type is then exchanged.
 * 3) The particles to send are placed in a temporary buffer in which the
 * part-gpart links are preserved.
 * 4) Each node allocates enough space for the new particles.
 * 5) (Asynchronous) communications are issued to transfer the data.
 *
 *
 * @param e The #engine.
 */
void engine_redistribute(struct engine *e) {

#ifdef WITH_MPI

  const int nr_nodes = e->nr_nodes;
  const int nodeID = e->nodeID;
  struct space *s = e->s;
  struct cell *cells = s->cells_top;
  const int nr_cells = s->nr_cells;
  struct xpart *xparts = s->xparts;
  struct part *parts = s->parts;
  struct gpart *gparts = s->gparts;
  struct spart *sparts = s->sparts;
  ticks tic = getticks();

  size_t nr_parts = s->nr_parts;
  size_t nr_gparts = s->nr_gparts;
  size_t nr_sparts = s->nr_sparts;

  /* Start by moving inhibited particles to the end of the arrays */
  for (size_t k = 0; k < nr_parts; /* void */) {
    if (parts[k].time_bin == time_bin_inhibited) {
      nr_parts -= 1;

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

      /* Swap the xpart */
      memswap(&xparts[k], &xparts[nr_parts], sizeof(struct xpart));

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

  /* Now move inhibited star particles to the end of the arrays */
  for (size_t k = 0; k < nr_sparts; /* void */) {
    if (sparts[k].time_bin == time_bin_inhibited) {
      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;
      }
    } else {
      k++;
    }
  }

  /* Finally do the same with the gravity particles */
  for (size_t k = 0; k < nr_gparts; /* void */) {
    if (gparts[k].time_bin == time_bin_inhibited) {
      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];
      }
    } else {
      k++;
    }
  }

  /* Now we are ready to deal with real particles and can start the exchange. */

  /* Allocate temporary arrays to store the counts of particles to be sent
   * and the destination of each particle */
  int *counts;
  if ((counts = (int *)calloc(sizeof(int), nr_nodes * nr_nodes)) == NULL)
    error("Failed to allocate counts temporary buffer.");

  int *dest;
  if ((dest = (int *)malloc(sizeof(int) * nr_parts)) == NULL)
    error("Failed to allocate dest temporary buffer.");

  /* Simple index of node IDs, used for mappers over nodes. */
  int *nodes = NULL;
  if ((nodes = (int *)malloc(sizeof(int) * nr_nodes)) == NULL)
    error("Failed to allocate nodes temporary buffer.");
  for (int k = 0; k < nr_nodes; k++) nodes[k] = k;