space.c 148 KB
Newer Older
1
/*******************************************************************************
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
 * 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/>.
 *
 ******************************************************************************/
Pedro Gonnet's avatar
Pedro Gonnet committed
23
24
25
26
27
28
29
30

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

/* Some standard headers. */
#include <float.h>
#include <limits.h>
#include <math.h>
31
#include <stdlib.h>
32
#include <string.h>
Pedro Gonnet's avatar
Pedro Gonnet committed
33

34
35
/* MPI headers. */
#ifdef WITH_MPI
36
#include <mpi.h>
37
38
#endif

39
40
41
/* This object's header. */
#include "space.h"

Pedro Gonnet's avatar
Pedro Gonnet committed
42
/* Local headers. */
43
#include "atomic.h"
44
#include "chemistry.h"
45
#include "const.h"
46
#include "cooling.h"
47
#include "debug.h"
48
#include "engine.h"
49
#include "error.h"
50
51
#include "gravity.h"
#include "hydro.h"
52
#include "kernel_hydro.h"
53
#include "lock.h"
54
#include "memswap.h"
55
#include "minmax.h"
56
#include "multipole.h"
57
#include "restart.h"
58
#include "sort_part.h"
59
#include "stars.h"
60
#include "threadpool.h"
61
#include "tools.h"
Pedro Gonnet's avatar
Pedro Gonnet committed
62
63
64

/* Split size. */
int space_splitsize = space_splitsize_default;
65
66
67
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;
68
int space_subsize_self_grav = space_subsize_self_grav_default;
69
70
int space_subsize_pair_stars = space_subsize_pair_stars_default;
int space_subsize_self_stars = space_subsize_self_stars_default;
71
int space_subdepth_diff_grav = space_subdepth_diff_grav_default;
72
int space_maxsize = space_maxsize_default;
73

74
75
76
/*! Number of extra #part we allocate memory for per top-level cell */
int space_extra_parts = space_extra_parts_default;

77
78
79
/*! Number of extra #spart we allocate memory for per top-level cell */
int space_extra_sparts = space_extra_sparts_default;

80
81
82
/*! Number of extra #gpart we allocate memory for per top-level cell */
int space_extra_gparts = space_extra_gparts_default;

83
84
/*! Expected maximal number of strays received at a rebuild */
int space_expected_max_nr_strays = space_expected_max_nr_strays_default;
85
86
87
#ifdef SWIFT_DEBUG_CHECKS
int last_cell_id;
#endif
Pedro Gonnet's avatar
Pedro Gonnet committed
88

89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
/**
 * @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;
105
  struct spart *sparts;
106
107
108
109
110
111
  int *ind;
  struct qstack *stack;
  unsigned int stack_size;
  volatile unsigned int first, last, waiting;
};

112
113
114
115
116
117
/**
 * @brief Information required to compute the particle cell indices.
 */
struct index_data {
  struct space *s;
  int *ind;
118
  int *cell_counts;
119
120
121
122
123
124
  size_t count_inhibited_part;
  size_t count_inhibited_gpart;
  size_t count_inhibited_spart;
  size_t count_extra_part;
  size_t count_extra_gpart;
  size_t count_extra_spart;
125
126
};

127
/**
128
 * @brief Recursively dismantle a cell tree.
129
 *
130
131
 * @param s The #space.
 * @param c The #cell to recycle.
Matthieu Schaller's avatar
Matthieu Schaller committed
132
133
134
135
136
137
 * @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.
138
 */
139
void space_rebuild_recycle_rec(struct space *s, struct cell *c,
140
141
                               struct cell **cell_rec_begin,
                               struct cell **cell_rec_end,
142
143
                               struct gravity_tensors **multipole_rec_begin,
                               struct gravity_tensors **multipole_rec_end) {
144
  if (c->split)
145
    for (int k = 0; k < 8; k++)
146
      if (c->progeny[k] != NULL) {
147
148
149
150
151
152
        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];
153

154
        if (s->with_self_gravity) {
155
156
          c->progeny[k]->grav.multipole->next = *multipole_rec_begin;
          *multipole_rec_begin = c->progeny[k]->grav.multipole;
157
        }
158
159

        if (*cell_rec_end == NULL) *cell_rec_end = *cell_rec_begin;
160
        if (s->with_self_gravity && *multipole_rec_end == NULL)
161
162
          *multipole_rec_end = *multipole_rec_begin;

163
        c->progeny[k]->grav.multipole = NULL;
164
165
166
167
        c->progeny[k] = NULL;
      }
}

168
169
170
171
172
173
174
175
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];
176
    struct cell *cell_rec_begin = NULL, *cell_rec_end = NULL;
177
178
    struct gravity_tensors *multipole_rec_begin = NULL,
                           *multipole_rec_end = NULL;
179
180
181
182
183
    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);
184
    c->hydro.sorts = NULL;
Loic Hausammann's avatar
Loic Hausammann committed
185
    c->stars.sorts = NULL;
186
    c->nr_tasks = 0;
187
188
189
190
191
192
    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;
193
194
    c->hydro.dx_max_part = 0.0f;
    c->hydro.dx_max_sort = 0.0f;
Loic Hausammann's avatar
Loic Hausammann committed
195
    c->stars.dx_max_part = 0.f;
Loic Hausammann's avatar
Loic Hausammann committed
196
    c->stars.dx_max_sort = 0.f;
197
    c->hydro.sorted = 0;
Loic Hausammann's avatar
Loic Hausammann committed
198
    c->stars.sorted = 0;
199
    c->hydro.count = 0;
200
    c->hydro.count_total = 0;
201
202
    c->hydro.updated = 0;
    c->hydro.inhibited = 0;
203
    c->grav.count = 0;
204
    c->grav.count_total = 0;
205
206
    c->grav.updated = 0;
    c->grav.inhibited = 0;
207
    c->stars.count = 0;
208
    c->stars.count_total = 0;
209
210
    c->stars.updated = 0;
    c->stars.inhibited = 0;
211
212
213
214
215
216
    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;
217
218
219
220
    c->stars.ghost_in = NULL;
    c->stars.ghost_out = NULL;
    c->stars.ghost = NULL;
    c->stars.density = NULL;
Alexei Borissov's avatar
Alexei Borissov committed
221
    c->stars.feedback = NULL;
222
223
    c->kick1 = NULL;
    c->kick2 = NULL;
224
    c->timestep = NULL;
225
    c->end_force = NULL;
226
    c->hydro.drift = NULL;
227
    c->grav.drift = NULL;
228
    c->grav.drift_out = NULL;
229
    c->hydro.cooling = NULL;
230
    c->sourceterms = NULL;
231
232
233
234
    c->grav.long_range = NULL;
    c->grav.down_in = NULL;
    c->grav.down = NULL;
    c->grav.mesh = NULL;
235
    c->super = c;
236
237
238
239
    c->hydro.super = c;
    c->grav.super = c;
    c->hydro.parts = NULL;
    c->hydro.xparts = NULL;
240
241
    c->grav.parts = NULL;
    c->stars.parts = NULL;
242
    c->hydro.do_sub_sort = 0;
Loic Hausammann's avatar
Loic Hausammann committed
243
    c->stars.do_sub_sort = 0;
244
245
246
247
248
249
    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;
Loic Hausammann's avatar
Loic Hausammann committed
250
    c->stars.ti_end_min = -1;
251
252
253
#ifdef SWIFT_DEBUG_CHECKS
    c->cellID = 0;
#endif
254
255
    if (s->with_self_gravity)
      bzero(c->grav.multipole, sizeof(struct gravity_tensors));
Loic Hausammann's avatar
Loic Hausammann committed
256
    for (int i = 0; i < 13; i++) {
257
258
259
      if (c->hydro.sort[i] != NULL) {
        free(c->hydro.sort[i]);
        c->hydro.sort[i] = NULL;
260
      }
Loic Hausammann's avatar
Loic Hausammann committed
261
262
263
264
265
      if (c->stars.sort[i] != NULL) {
        free(c->stars.sort[i]);
        c->stars.sort[i] = NULL;
      }
    }
266
#if WITH_MPI
267
268
    c->mpi.tag = -1;

269
270
271
272
    c->mpi.hydro.recv_xv = NULL;
    c->mpi.hydro.recv_rho = NULL;
    c->mpi.hydro.recv_gradient = NULL;
    c->mpi.grav.recv = NULL;
273
274
    c->mpi.recv_ti = NULL;

275
276
277
278
    c->mpi.hydro.send_xv = NULL;
    c->mpi.hydro.send_rho = NULL;
    c->mpi.hydro.send_gradient = NULL;
    c->mpi.grav.send = NULL;
279
    c->mpi.send_ti = NULL;
280
281
282
283
#endif
  }
}

284
285
286
287
/**
 * @brief Free up any allocated cells.
 */
void space_free_cells(struct space *s) {
288
289
290

  ticks tic = getticks();

Matthieu Schaller's avatar
Matthieu Schaller committed
291
292
  threadpool_map(&s->e->threadpool, space_rebuild_recycle_mapper, s->cells_top,
                 s->nr_cells, sizeof(struct cell), 0, s);
293
  s->maxdepth = 0;
294
295
296
297

  if (s->e->verbose)
    message("took %.3f %s.", clocks_from_ticks(getticks() - tic),
            clocks_getunit());
298
299
}

300
/**
301
 * @brief Re-build the top-level cell grid.
302
 *
303
 * @param s The #space.
304
 * @param verbose Print messages to stdout or not.
305
 */
306
void space_regrid(struct space *s, int verbose) {
307

308
  const size_t nr_parts = s->nr_parts;
Loic Hausammann's avatar
Loic Hausammann committed
309
  const size_t nr_sparts = s->nr_sparts;
310
  const ticks tic = getticks();
311
  const integertime_t ti_current = (s->e != NULL) ? s->e->ti_current : 0;
312

313
  /* Run through the cells and get the current h_max. */
314
  // tic = getticks();
315
  float h_max = s->cell_min / kernel_gamma / space_stretch;
316
  if (nr_parts > 0) {
317
318
319
320
321
322

    /* 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]];
323
        if (c->hydro.h_max > h_max) {
324
          h_max = c->hydro.h_max;
325
        }
Loic Hausammann's avatar
Loic Hausammann committed
326
        if (c->stars.h_max > h_max) {
327
          h_max = c->stars.h_max;
Loic Hausammann's avatar
Loic Hausammann committed
328
        }
329
      }
330
331

      /* Can we instead use all the top-level cells? */
332
    } else if (s->cells_top != NULL) {
Tom Theuns's avatar
Tom Theuns committed
333
      for (int k = 0; k < s->nr_cells; k++) {
334
        const struct cell *c = &s->cells_top[k];
335
        if (c->nodeID == engine_rank && c->hydro.h_max > h_max) {
336
          h_max = c->hydro.h_max;
337
        }
Loic Hausammann's avatar
Loic Hausammann committed
338
        if (c->nodeID == engine_rank && c->stars.h_max > h_max) {
339
          h_max = c->stars.h_max;
Loic Hausammann's avatar
Loic Hausammann committed
340
        }
341
      }
342
343

      /* Last option: run through the particles */
344
    } else {
345
      for (size_t k = 0; k < nr_parts; k++) {
Matthieu Schaller's avatar
Matthieu Schaller committed
346
        if (s->parts[k].h > h_max) h_max = s->parts[k].h;
347
      }
Loic Hausammann's avatar
Loic Hausammann committed
348
349
350
      for (size_t k = 0; k < nr_sparts; k++) {
        if (s->sparts[k].h > h_max) h_max = s->sparts[k].h;
      }
351
352
353
354
355
356
357
358
359
360
    }
  }

/* 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)
361
      error("Failed to aggregate the rebuild flag across nodes.");
362
363
364
    h_max = buff;
  }
#endif
365
  if (verbose) message("h_max is %.3e (cell_min=%.3e).", h_max, s->cell_min);
366
367

  /* Get the new putative cell dimensions. */
368
  const int cdim[3] = {
369
370
371
372
373
374
      (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))};
375
376
377
378
379

  /* 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 "
380
381
382
        "is switched on.\nThis error is often caused by any of the "
        "followings:\n"
        " - too few particles to generate a sensible grid,\n"
383
384
        " - the initial value of 'Scheduler:max_top_level_cells' is too "
        "small,\n"
385
        " - the (minimal) time-step is too large leading to particles with "
386
        "predicted smoothing lengths too large for the box size,\n"
387
        " - particles with velocities so large that they move by more than two "
388
        "box sizes per time-step.\n");
389

390
391
392
/* 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. */
393
#ifdef WITH_MPI
Matthieu Schaller's avatar
Matthieu Schaller committed
394
  double oldwidth[3];
395
396
397
398
399
400
401
402
  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];
403
404
405
    oldwidth[0] = s->width[0];
    oldwidth[1] = s->width[1];
    oldwidth[2] = s->width[2];
406
407
408
409
410
411
412
413
414

    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);
415
          oldnodeIDs[cid] = s->cells_top[cid].nodeID;
416
417
418
419
420
        }
      }
    }
  }

Peter W. Draper's avatar
Peter W. Draper committed
421
  /* Are we about to allocate new top level cells without a regrid?
Peter W. Draper's avatar
Peter W. Draper committed
422
   * Can happen when restarting the application. */
423
  const int no_regrid = (s->cells_top == NULL && oldnodeIDs == NULL);
424
425
426
427
#endif

  /* Do we need to re-build the upper-level cells? */
  // tic = getticks();
428
  if (s->cells_top == NULL || cdim[0] < s->cdim[0] || cdim[1] < s->cdim[1] ||
429
430
      cdim[2] < s->cdim[2]) {

431
432
/* Be verbose about this. */
#ifdef SWIFT_DEBUG_CHECKS
433
    message("(re)griding space cdim=(%d %d %d)", cdim[0], cdim[1], cdim[2]);
434
435
436
    fflush(stdout);
#endif

437
    /* Free the old cells, if they were allocated. */
438
    if (s->cells_top != NULL) {
439
      space_free_cells(s);
440
      free(s->local_cells_with_tasks_top);
441
      free(s->local_cells_top);
442
      free(s->cells_with_particles_top);
443
      free(s->local_cells_with_particles_top);
444
      free(s->cells_top);
445
      free(s->multipoles_top);
446
447
    }

448
449
450
451
    /* 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);

452
    /* Set the new cell dimensions only if smaller. */
453
    for (int k = 0; k < 3; k++) {
454
      s->cdim[k] = cdim[k];
455
456
      s->width[k] = s->dim[k] / cdim[k];
      s->iwidth[k] = 1.0 / s->width[k];
457
    }
458
    const float dmin = min3(s->width[0], s->width[1], s->width[2]);
459
460
461

    /* Allocate the highest level of cells. */
    s->tot_cells = s->nr_cells = cdim[0] * cdim[1] * cdim[2];
462
    if (posix_memalign((void **)&s->cells_top, cell_align,
463
                       s->nr_cells * sizeof(struct cell)) != 0)
464
      error("Failed to allocate top-level cells.");
465
    bzero(s->cells_top, s->nr_cells * sizeof(struct cell));
466

467
    /* Allocate the multipoles for the top-level cells. */
468
    if (s->with_self_gravity) {
469
      if (posix_memalign((void **)&s->multipoles_top, multipole_align,
470
                         s->nr_cells * sizeof(struct gravity_tensors)) != 0)
471
        error("Failed to allocate top-level multipoles.");
472
      bzero(s->multipoles_top, s->nr_cells * sizeof(struct gravity_tensors));
473
474
    }

475
    /* Allocate the indices of local cells */
476
    if (posix_memalign((void **)&s->local_cells_top, SWIFT_STRUCT_ALIGNMENT,
477
478
479
480
                       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));

481
    /* Allocate the indices of local cells with tasks */
482
483
    if (posix_memalign((void **)&s->local_cells_with_tasks_top,
                       SWIFT_STRUCT_ALIGNMENT, s->nr_cells * sizeof(int)) != 0)
484
      error("Failed to allocate indices of local top-level cells with tasks.");
485
486
    bzero(s->local_cells_with_tasks_top, s->nr_cells * sizeof(int));

487
    /* Allocate the indices of cells with particles */
488
    if (posix_memalign((void **)&s->cells_with_particles_top,
489
                       SWIFT_STRUCT_ALIGNMENT, s->nr_cells * sizeof(int)) != 0)
490
491
      error("Failed to allocate indices of top-level cells with particles.");
    bzero(s->cells_with_particles_top, s->nr_cells * sizeof(int));
492

493
494
495
496
497
498
499
500
    /* 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));

501
    /* Set the cells' locks */
502
    for (int k = 0; k < s->nr_cells; k++) {
503
      if (lock_init(&s->cells_top[k].hydro.lock) != 0)
504
        error("Failed to init spinlock for hydro.");
505
      if (lock_init(&s->cells_top[k].grav.plock) != 0)
506
        error("Failed to init spinlock for gravity.");
507
      if (lock_init(&s->cells_top[k].grav.mlock) != 0)
508
        error("Failed to init spinlock for multipoles.");
509
      if (lock_init(&s->cells_top[k].stars.lock) != 0)
510
511
        error("Failed to init spinlock for stars.");
    }
512
513

    /* Set the cell location and sizes. */
514
515
516
    for (int i = 0; i < cdim[0]; i++)
      for (int j = 0; j < cdim[1]; j++)
        for (int k = 0; k < cdim[2]; k++) {
517
518
          const size_t cid = cell_getid(cdim, i, j, k);
          struct cell *restrict c = &s->cells_top[cid];
519
520
521
522
523
524
          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];
525
526
          c->dmin = dmin;
          c->depth = 0;
527
          c->split = 0;
528
          c->hydro.count = 0;
529
530
          c->grav.count = 0;
          c->stars.count = 0;
531
          c->super = c;
532
533
          c->hydro.super = c;
          c->grav.super = c;
534
535
          c->hydro.ti_old_part = ti_current;
          c->grav.ti_old_part = ti_current;
536
          c->grav.ti_old_multipole = ti_current;
Pedro Gonnet's avatar
Pedro Gonnet committed
537
#ifdef WITH_MPI
538
          c->mpi.tag = -1;
539
540
541
542
543
544
545
546
          c->mpi.hydro.recv_xv = NULL;
          c->mpi.hydro.recv_rho = NULL;
          c->mpi.hydro.recv_gradient = NULL;
          c->mpi.hydro.send_xv = NULL;
          c->mpi.hydro.send_rho = NULL;
          c->mpi.hydro.send_gradient = NULL;
          c->mpi.grav.recv = NULL;
          c->mpi.grav.send = NULL;
Pedro Gonnet's avatar
Pedro Gonnet committed
547
#endif  // WITH_MPI
548
          if (s->with_self_gravity) c->grav.multipole = &s->multipoles_top[cid];
549
#ifdef SWIFT_DEBUG_CHECKS
Matthieu Schaller's avatar
Matthieu Schaller committed
550
551
          c->cellID = -last_cell_id;
          last_cell_id++;
552
#endif
Pedro Gonnet's avatar
Pedro Gonnet committed
553
        }
554
555

    /* Be verbose about the change. */
556
557
558
    if (verbose)
      message("set cell dimensions to [ %i %i %i ].", cdim[0], cdim[1],
              cdim[2]);
559

560
#ifdef WITH_MPI
561
562
563
564
565
    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)
566
567
568
        message(
            "basic cell dimensions have increased - recalculating the "
            "global partition.");
569

Matthieu Schaller's avatar
Matthieu Schaller committed
570
      if (!partition_space_to_space(oldwidth, oldcdim, oldnodeIDs, s)) {
571
572
573
574

        /* Failed, try another technique that requires no settings. */
        message("Failed to get a new partition, trying less optimal method");
        struct partition initial_partition;
575
#if defined(HAVE_PARMETIS) || defined(HAVE_METIS)
576
577
578
579
580
581
582
583
584
585
586
587
588
        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);
589

590
591
      /* Finished with these. */
      free(oldnodeIDs);
Peter W. Draper's avatar
Peter W. Draper committed
592
593

    } else if (no_regrid && s->e != NULL) {
Peter W. Draper's avatar
Peter W. Draper committed
594
595
596
597
598
599
600
601
602
603
604
605
606
      /* 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);
607
    }
Pedro Gonnet's avatar
Pedro Gonnet committed
608
#endif /* WITH_MPI */
609
610
611
612

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

613
  }      /* re-build upper-level cells? */
614
  else { /* Otherwise, just clean up the cells. */
615
616

    /* Free the old cells, if they were allocated. */
617
    space_free_cells(s);
618
  }
619
620
621
622

  if (verbose)
    message("took %.3f %s.", clocks_from_ticks(getticks() - tic),
            clocks_getunit());
623
}
624

625
626
627
628
629
630
/**
 * @brief Allocate memory for the extra particles used for on-the-fly creation.
 *
 * This rarely actually allocates memory. Most of the time, we convert
 * pre-allocated memory inot extra particles.
 *
631
632
 * This function also sets the extra particles' location to their top-level
 * cells. They can then be sorted into their correct memory position later on.
633
634
635
636
 *
 * @param s The current #space.
 * @param verbose Are we talkative?
 */
637
638
639
640
void space_allocate_extras(struct space *s, int verbose) {

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

641
642
643
644
645
  /* Anything to do here? (Abort if we don't want extras)*/
  if (space_extra_parts == 0 && space_extra_gparts == 0 &&
      space_extra_sparts == 0)
    return;

646
647
648
649
650
651
  /* The top-level cells */
  const struct cell *cells = s->cells_top;
  const double half_cell_width[3] = {0.5 * cells[0].width[0],
                                     0.5 * cells[0].width[1],
                                     0.5 * cells[0].width[2]};

652
  /* The current number of particles (including spare ones) */
653
654
655
656
  size_t nr_parts = s->nr_parts;
  size_t nr_gparts = s->nr_gparts;
  size_t nr_sparts = s->nr_sparts;

657
658
659
660
661
  /* The current number of actual particles */
  size_t nr_actual_parts = nr_parts - s->nr_extra_parts;
  size_t nr_actual_gparts = nr_gparts - s->nr_extra_gparts;
  size_t nr_actual_sparts = nr_sparts - s->nr_extra_sparts;

662
663
664
665
666
667
668
669
670
  /* The number of particles we allocated memory for (MPI overhead) */
  size_t size_parts = s->size_parts;
  size_t size_gparts = s->size_gparts;
  size_t size_sparts = s->size_sparts;

  int local_cells = 0;
  for (int i = 0; i < s->nr_cells; ++i)
    if (s->cells_top[i].nodeID == local_nodeID) local_cells++;

671
672
  /* Number of extra particles we want for each type */
  const size_t expected_num_extra_parts = local_cells * space_extra_parts;
673
674
  const size_t expected_num_extra_gparts = local_cells * space_extra_gparts;
  const size_t expected_num_extra_sparts = local_cells * space_extra_sparts;
675

676
677
678
  if (verbose) {
    message("Currently have %zd/%zd/%zd real particles.", nr_actual_parts,
            nr_actual_gparts, nr_actual_sparts);
679
    message("Currently have %zd/%zd/%zd spaces for extra particles.",
680
            s->nr_extra_parts, s->nr_extra_gparts, s->nr_extra_sparts);
681
    message("Requesting space for future %zd/%zd/%zd part/gpart/sparts.",
682
683
684
            expected_num_extra_parts, expected_num_extra_gparts,
            expected_num_extra_sparts);
  }
685

686
687
688
689
690
691
692
  if (expected_num_extra_parts < s->nr_extra_parts)
    error("Reduction in top-level cells number not handled.");
  if (expected_num_extra_gparts < s->nr_extra_gparts)
    error("Reduction in top-level cells number not handled.");
  if (expected_num_extra_sparts < s->nr_extra_sparts)
    error("Reduction in top-level cells number not handled.");

693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
  /* Do we have enough space for the extra gparts (i.e. we haven't used up any)
   * ? */
  if (nr_gparts + expected_num_extra_gparts > size_gparts) {

    /* Ok... need to put some more in the game */

    /* Do we need to reallocate? */
    if (nr_actual_gparts + expected_num_extra_gparts > size_gparts) {

      size_gparts = (nr_actual_gparts + expected_num_extra_gparts) *
                    engine_redistribute_alloc_margin;

      if (verbose)
        message("Re-allocating gparts array from %zd to %zd", s->size_gparts,
                size_gparts);

      /* Create more space for parts */
      struct gpart *gparts_new = NULL;
      if (posix_memalign((void **)&gparts_new, gpart_align,
                         sizeof(struct gpart) * size_gparts) != 0)
        error("Failed to allocate new gpart data");
      const ptrdiff_t delta = gparts_new - s->gparts;
      memcpy(gparts_new, s->gparts, sizeof(struct gpart) * s->size_gparts);
      free(s->gparts);
      s->gparts = gparts_new;

      /* Update the counter */
      s->size_gparts = size_gparts;

      /* We now need to reset all the part and spart pointers */
      for (size_t i = 0; i < nr_parts; ++i) {
        if (s->parts[i].time_bin != time_bin_not_created)
          s->parts[i].gpart += delta;
      }
      for (size_t i = 0; i < nr_sparts; ++i) {
        if (s->sparts[i].time_bin != time_bin_not_created)
          s->sparts[i].gpart += delta;
      }
    }

    /* Turn some of the allocated spares into particles we can use */
    for (size_t i = nr_gparts; i < nr_actual_gparts + expected_num_extra_gparts;
         ++i) {
      bzero(&s->gparts[i], sizeof(struct gpart));
      s->gparts[i].time_bin = time_bin_not_created;
      s->gparts[i].type = swift_type_dark_matter;
      s->gparts[i].id_or_neg_offset = -1;
    }

      /* Put the spare particles in their correct cell */
#ifdef WITH_MPI
    error("Need to do this correctly over MPI for only the local cells.");
#endif
    int count_in_cell = 0, current_cell = 0;
    size_t count_extra_gparts = 0;
    for (size_t i = 0; i < nr_actual_gparts + expected_num_extra_gparts; ++i) {

#ifdef SWIFT_DEBUG_CHECKS
      if (current_cell == s->nr_cells)
        error("Cell counter beyond the maximal nr. cells.");
#endif

      if (s->gparts[i].time_bin == time_bin_not_created) {

        /* We want the extra particles to be at the centre of their cell */
        s->gparts[i].x[0] = cells[current_cell].loc[0] + half_cell_width[0];
        s->gparts[i].x[1] = cells[current_cell].loc[1] + half_cell_width[1];
        s->gparts[i].x[2] = cells[current_cell].loc[2] + half_cell_width[2];
        ++count_in_cell;
        count_extra_gparts++;
      }

      /* Once we have reached the number of extra gpart per cell, we move to the
       * next */
      if (count_in_cell == space_extra_gparts) {
        ++current_cell;
        count_in_cell = 0;
      }
    }

#ifdef SWIFT_DEBUG_CHECKS
    if (count_extra_gparts != expected_num_extra_gparts)
      error("Constructed the wrong number of extra gparts (%zd vs. %zd)",
            count_extra_gparts, expected_num_extra_gparts);
#endif

    /* Update the counters */
    s->nr_gparts = nr_actual_gparts + expected_num_extra_gparts;
    s->nr_extra_gparts = expected_num_extra_gparts;
  }

784
785
  /* Do we have enough space for the extra parts (i.e. we haven't used up any) ?
   */
786
787
  if (expected_num_extra_parts > s->nr_extra_parts) {

788
789
    /* Ok... need to put some more in the game */

790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
    /* Do we need to reallocate? */
    if (nr_actual_parts + expected_num_extra_parts > size_parts) {

      size_parts = (nr_actual_parts + expected_num_extra_parts) *
                   engine_redistribute_alloc_margin;

      if (verbose)
        message("Re-allocating parts array from %zd to %zd", s->size_parts,
                size_parts);

      /* Create more space for parts */
      struct part *parts_new = NULL;
      if (posix_memalign((void **)&parts_new, part_align,
                         sizeof(struct part) * size_parts) != 0)
        error("Failed to allocate new part data");
      memcpy(parts_new, s->parts, sizeof(struct part) * s->size_parts);
      free(s->parts);
      s->parts = parts_new;

      /* Same for xparts */
      struct xpart *xparts_new = NULL;
      if (posix_memalign((void **)&xparts_new, xpart_align,
                         sizeof(struct xpart) * size_parts) != 0)
        error("Failed to allocate new xpart data");
      memcpy(xparts_new, s->xparts, sizeof(struct xpart) * s->size_parts);
      free(s->xparts);
      s->xparts = xparts_new;

      /* Update the counter */
      s->size_parts = size_parts;
    }

822
    /* Turn some of the allocated spares into particles we can use */
823
824
825
826
827
    for (size_t i = nr_parts; i < nr_actual_parts + expected_num_extra_parts;
         ++i) {
      bzero(&s->parts[i], sizeof(struct part));
      bzero(&s->xparts[i], sizeof(struct xpart));
      s->parts[i].time_bin = time_bin_not_created;
828
      s->parts[i].id = -1;
829
830
    }

831
832
833
834
835
836
837
      /* Put the spare particles in their correct cell */
#ifdef WITH_MPI
    error("Need to do this correctly over MPI for only the local cells.");
#endif
    int count_in_cell = 0, current_cell = 0;
    size_t count_extra_parts = 0;
    for (size_t i = 0; i < nr_actual_parts + expected_num_extra_parts; ++i) {
838
839
840
841
842
843

#ifdef SWIFT_DEBUG_CHECKS
      if (current_cell == s->nr_cells)
        error("Cell counter beyond the maximal nr. cells.");
#endif

844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
      if (s->parts[i].time_bin == time_bin_not_created) {

        /* We want the extra particles to be at the centre of their cell */
        s->parts[i].x[0] = cells[current_cell].loc[0] + half_cell_width[0];
        s->parts[i].x[1] = cells[current_cell].loc[1] + half_cell_width[1];
        s->parts[i].x[2] = cells[current_cell].loc[2] + half_cell_width[2];
        ++count_in_cell;
        count_extra_parts++;
      }

      /* Once we have reached the number of extra part per cell, we move to the
       * next */
      if (count_in_cell == space_extra_parts) {
        ++current_cell;
        count_in_cell = 0;
      }
    }

#ifdef SWIFT_DEBUG_CHECKS
    if (count_extra_parts != expected_num_extra_parts)
      error("Constructed the wrong number of extra parts (%zd vs. %zd)",
            count_extra_parts, expected_num_extra_parts);
#endif

868
869
870
    /* Update the counters */
    s->nr_parts = nr_actual_parts + expected_num_extra_parts;
    s->nr_extra_parts = expected_num_extra_parts;
871
  }
872

873
874
  /* Do we have enough space for the extra sparts (i.e. we haven't used up any)
   * ? */
875
  if (nr_actual_sparts + expected_num_extra_sparts > nr_sparts) {
876

877
878
    /* Ok... need to put some more in the game */

879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
    /* Do we need to reallocate? */
    if (nr_actual_sparts + expected_num_extra_sparts > size_sparts) {

      size_sparts = (nr_actual_sparts + expected_num_extra_sparts) *
                    engine_redistribute_alloc_margin;

      if (verbose)
        message("Re-allocating sparts array from %zd to %zd", s->size_sparts,
                size_sparts);

      /* Create more space for parts */
      struct spart *sparts_new = NULL;
      if (posix_memalign((void **)&sparts_new, spart_align,
                         sizeof(struct spart) * size_sparts) != 0)
        error("Failed to allocate new spart data");
      memcpy(sparts_new, s->sparts, sizeof(struct spart) * s->size_sparts);
      free(s->sparts);
      s->sparts = sparts_new;

      /* Update the counter */
      s->size_sparts = size_sparts;
    }

    /* Turn some of the allocated spares into particles we can use */
    for (size_t i = nr_sparts; i < nr_actual_sparts + expected_num_extra_sparts;
         ++i) {
      bzero(&s->sparts[i], sizeof(struct spart));
      s->sparts[i].time_bin = time_bin_not_created;
907
      s->sparts[i].id = -42;
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
    }

      /* Put the spare particles in their correct cell */
#ifdef WITH_MPI
    error("Need to do this correctly over MPI for only the local cells.");
#endif
    int count_in_cell = 0, current_cell = 0;
    size_t count_extra_sparts = 0;
    for (size_t i = 0; i < nr_actual_sparts + expected_num_extra_sparts; ++i) {

#ifdef SWIFT_DEBUG_CHECKS
      if (current_cell == s->nr_cells)
        error("Cell counter beyond the maximal nr. cells.");
#endif

      if (s->sparts[i].time_bin == time_bin_not_created) {

        /* We want the extra particles to be at the centre of their cell */
        s->sparts[i].x[0] = cells[current_cell].loc[0] + half_cell_width[0];
        s->sparts[i].x[1] = cells[current_cell].loc[1] + half_cell_width[1];
        s->sparts[i].x[2] = cells[current_cell].loc[2] + half_cell_width[2];
        ++count_in_cell;
        count_extra_sparts++;
      }

      /* Once we have reached the number of extra spart per cell, we move to the
       * next */
      if (count_in_cell == space_extra_sparts) {
        ++current_cell;
        count_in_cell = 0;
      }
    }

#ifdef SWIFT_DEBUG_CHECKS
    if (count_extra_sparts != expected_num_extra_sparts)
      error("Constructed the wrong number of extra sparts (%zd vs. %zd)",
            count_extra_sparts, expected_num_extra_sparts);
#endif

    /* Update the counters */
    s->nr_sparts = nr_actual_sparts + expected_num_extra_sparts;
    s->nr_extra_sparts = expected_num_extra_sparts;
950
  }
951
952
953
954
955
956
957

#ifdef SWIFT_DEBUG_CHECKS
  /* Verify that the links are correct */
  if ((nr_gparts > 0 && nr_parts > 0) || (nr_gparts > 0 && nr_sparts > 0))
    part_verify_links(s->parts, s->gparts, s->sparts, nr_parts, nr_gparts,
                      nr_sparts, verbose);
#endif
958
959
}

960
961
962
963
/**
 * @brief Re-build the cells as well as the tasks.
 *
 * @param s The #space in which to update the cells.
964
 * @param repartitioned Did we just repartition?
965
 * @param verbose Print messages to stdout or not
966
 */
967
void space_rebuild(struct space *s, int repartitioned, int verbose) {
968

Matthieu Schaller's avatar
Matthieu Schaller committed
969
  const ticks tic = getticks();
970

971
972
/* Be verbose about this. */
#ifdef SWIFT_DEBUG_CHECKS
Pedro Gonnet's avatar
Pedro Gonnet committed
973
  if (s->e->nodeID == 0 || verbose) message("(re)building space");
974
975
  fflush(stdout);
#endif
976
977

  /* Re-grid if necessary, or just re-set the cell data. */
978
  space_regrid(s, verbose);
979

980
  /* Allocate extra space for particles that will be created */
981
  if (s->with_star_formation) space_allocate_extras(s, verbose);
982

983
984
  struct cell *cells_top = s->cells_top;
  const integertime_t ti_current = (s->e != NULL) ? s->e->ti_current : 0;
985
  const int local_nodeID = s->e->nodeID;
986
987

  /* The current number of particles */
Pedro Gonnet's avatar
Pedro Gonnet committed
988
989
  size_t nr_parts = s->nr_parts;
  size_t nr_gparts = s->nr_gparts;
990
  size_t nr_sparts = s->nr_sparts;
991

992
993
994
995
996
  /* The number of particles we allocated memory for */
  size_t size_parts = s->size_parts;
  size_t size_gparts = s->size_gparts;
  size_t size_sparts = s->size_sparts;

997
  /* Counter for the number of inhibited particles found on the node */
998
999
1000
  size_t count_inhibited_parts = 0;
  size_t count_inhibited_gparts = 0;
  size_t count_inhibited_sparts = 0;