space.c 52.8 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 "const.h"
45
#include "cooling.h"
46
#include "engine.h"
47
#include "error.h"
48
49
#include "gravity.h"
#include "hydro.h"
50
#include "kernel_hydro.h"
51
#include "lock.h"
52
#include "minmax.h"
53
#include "runner.h"
54
#include "threadpool.h"
55
#include "tools.h"
Pedro Gonnet's avatar
Pedro Gonnet committed
56
57
58

/* Split size. */
int space_splitsize = space_splitsize_default;
59
int space_subsize = space_subsize_default;
60
int space_maxsize = space_maxsize_default;
61
int space_maxcount = space_maxcount_default;
Pedro Gonnet's avatar
Pedro Gonnet committed
62
63
64

/* Map shift vector to sortlist. */
const int sortlistID[27] = {
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
    /* ( -1 , -1 , -1 ) */ 0,
    /* ( -1 , -1 ,  0 ) */ 1,
    /* ( -1 , -1 ,  1 ) */ 2,
    /* ( -1 ,  0 , -1 ) */ 3,
    /* ( -1 ,  0 ,  0 ) */ 4,
    /* ( -1 ,  0 ,  1 ) */ 5,
    /* ( -1 ,  1 , -1 ) */ 6,
    /* ( -1 ,  1 ,  0 ) */ 7,
    /* ( -1 ,  1 ,  1 ) */ 8,
    /* (  0 , -1 , -1 ) */ 9,
    /* (  0 , -1 ,  0 ) */ 10,
    /* (  0 , -1 ,  1 ) */ 11,
    /* (  0 ,  0 , -1 ) */ 12,
    /* (  0 ,  0 ,  0 ) */ 0,
    /* (  0 ,  0 ,  1 ) */ 12,
    /* (  0 ,  1 , -1 ) */ 11,
    /* (  0 ,  1 ,  0 ) */ 10,
    /* (  0 ,  1 ,  1 ) */ 9,
    /* (  1 , -1 , -1 ) */ 8,
    /* (  1 , -1 ,  0 ) */ 7,
    /* (  1 , -1 ,  1 ) */ 6,
    /* (  1 ,  0 , -1 ) */ 5,
    /* (  1 ,  0 ,  0 ) */ 4,
    /* (  1 ,  0 ,  1 ) */ 3,
    /* (  1 ,  1 , -1 ) */ 2,
    /* (  1 ,  1 ,  0 ) */ 1,
    /* (  1 ,  1 ,  1 ) */ 0};

93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
/**
 * @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;
  int *ind;
  struct qstack *stack;
  unsigned int stack_size;
  volatile unsigned int first, last, waiting;
};

115
116
117
118
119
120
121
122
123
124
125
/**
 * @brief Get the shift-id of the given pair of cells, swapping them
 *      if need be.
 *
 * @param s The space
 * @param ci Pointer to first #cell.
 * @param cj Pointer second #cell.
 * @param shift Vector from ci to cj.
 *
 * @return The shift ID and set shift, may or may not swap ci and cj.
 */
126
127
128
129
int space_getsid(struct space *s, struct cell **ci, struct cell **cj,
                 double *shift) {

  /* Get the relative distance between the pairs, wrapping. */
130
131
132
  const int periodic = s->periodic;
  double dx[3];
  for (int k = 0; k < 3; k++) {
133
134
135
136
137
138
139
140
141
142
143
    dx[k] = (*cj)->loc[k] - (*ci)->loc[k];
    if (periodic && dx[k] < -s->dim[k] / 2)
      shift[k] = s->dim[k];
    else if (periodic && dx[k] > s->dim[k] / 2)
      shift[k] = -s->dim[k];
    else
      shift[k] = 0.0;
    dx[k] += shift[k];
  }

  /* Get the sorting index. */
144
  int sid = 0;
145
  for (int k = 0; k < 3; k++)
146
147
148
149
    sid = 3 * sid + ((dx[k] < 0.0) ? 0 : ((dx[k] > 0.0) ? 2 : 1));

  /* Switch the cells around? */
  if (runner_flip[sid]) {
150
    struct cell *temp = *ci;
151
152
    *ci = *cj;
    *cj = temp;
153
    for (int k = 0; k < 3; k++) shift[k] = -shift[k];
154
155
156
157
158
159
  }
  sid = sortlistID[sid];

  /* Return the sort ID. */
  return sid;
}
160

161
/**
162
 * @brief Recursively dismantle a cell tree.
163
 *
164
165
 * @param s The #space.
 * @param c The #cell to recycle.
166
 */
167
168
169
void space_rebuild_recycle(struct space *s, struct cell *c) {

  if (c->split)
170
    for (int k = 0; k < 8; k++)
171
172
173
174
175
176
177
      if (c->progeny[k] != NULL) {
        space_rebuild_recycle(s, c->progeny[k]);
        space_recycle(s, c->progeny[k]);
        c->progeny[k] = NULL;
      }
}

178
/**
179
 * @brief Re-build the top-level cell grid.
180
 *
181
182
 * @param s The #space.
 * @param cell_max Maximum cell edge length.
183
 * @param verbose Print messages to stdout or not.
184
 */
185
void space_regrid(struct space *s, double cell_max, int verbose) {
186

187
  const size_t nr_parts = s->nr_parts;
188
  const ticks tic = getticks();
189
  const int ti_current = (s->e != NULL) ? s->e->ti_current : 0;
190

191
  /* Run through the cells and get the current h_max. */
192
  // tic = getticks();
193
  float h_max = s->cell_min / kernel_gamma / space_stretch;
194
  if (nr_parts > 0) {
195
    if (s->cells_top != NULL) {
Tom Theuns's avatar
Tom Theuns committed
196
      for (int k = 0; k < s->nr_cells; k++) {
197
198
199
        if (s->cells_top[k].nodeID == engine_rank &&
            s->cells_top[k].h_max > h_max) {
          h_max = s->cells_top[k].h_max;
200
        }
201
202
      }
    } else {
203
      for (size_t k = 0; k < nr_parts; k++) {
Matthieu Schaller's avatar
Matthieu Schaller committed
204
        if (s->parts[k].h > h_max) h_max = s->parts[k].h;
205
      }
206
207
208
209
210
211
212
213
214
215
    }
  }

/* 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)
216
      error("Failed to aggregate the rebuild flag across nodes.");
217
218
219
    h_max = buff;
  }
#endif
220
  if (verbose) message("h_max is %.3e (cell_max=%.3e).", h_max, cell_max);
221
222

  /* Get the new putative cell dimensions. */
223
224
225
226
  const int cdim[3] = {
      floor(s->dim[0] / fmax(h_max * kernel_gamma * space_stretch, cell_max)),
      floor(s->dim[1] / fmax(h_max * kernel_gamma * space_stretch, cell_max)),
      floor(s->dim[2] / fmax(h_max * kernel_gamma * space_stretch, cell_max))};
227
228
229
230
231

  /* 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 "
232
233
234
235
        "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 'SPH:max_smoothing_length' is too large,\n"
236
        " - the (minimal) time-step is too large leading to particles with "
237
238
239
        "predicted smoothing lengths too large for the box size,\n"
        " - particle with velocities so large that they move by more than two "
        "box sizes per time-step.\n");
240

241
242
243
244
245
246
  /* Check if we have enough cells for gravity. */
  if (s->gravity && (cdim[0] < 8 || cdim[1] < 8 || cdim[2] < 8))
    error(
        "Must have at least 8 cells in each spatial dimension when gravity "
        "is switched on.");

247
248
249
/* 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. */
250
#ifdef WITH_MPI
Matthieu Schaller's avatar
Matthieu Schaller committed
251
  double oldwidth[3];
252
253
254
255
256
257
258
259
  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];
260
261
262
    oldwidth[0] = s->width[0];
    oldwidth[1] = s->width[1];
    oldwidth[2] = s->width[2];
263
264
265
266
267
268
269
270
271

    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);
272
          oldnodeIDs[cid] = s->cells_top[cid].nodeID;
273
274
275
276
277
        }
      }
    }
  }

278
279
280
281
#endif

  /* Do we need to re-build the upper-level cells? */
  // tic = getticks();
282
  if (s->cells_top == NULL || cdim[0] < s->cdim[0] || cdim[1] < s->cdim[1] ||
283
284
285
      cdim[2] < s->cdim[2]) {

    /* Free the old cells, if they were allocated. */
286
    if (s->cells_top != NULL) {
287
      for (int k = 0; k < s->nr_cells; k++) {
288
289
        space_rebuild_recycle(s, &s->cells_top[k]);
        if (s->cells_top[k].sort != NULL) free(s->cells_top[k].sort);
290
      }
291
      free(s->cells_top);
292
293
294
295
      s->maxdepth = 0;
    }

    /* Set the new cell dimensions only if smaller. */
296
    for (int k = 0; k < 3; k++) {
297
      s->cdim[k] = cdim[k];
298
299
      s->width[k] = s->dim[k] / cdim[k];
      s->iwidth[k] = 1.0 / s->width[k];
300
    }
301
    const float dmin = min(s->width[0], min(s->width[1], s->width[2]));
302
303
304

    /* Allocate the highest level of cells. */
    s->tot_cells = s->nr_cells = cdim[0] * cdim[1] * cdim[2];
305
    if (posix_memalign((void *)&s->cells_top, cell_align,
306
307
                       s->nr_cells * sizeof(struct cell)) != 0)
      error("Failed to allocate cells.");
308
    bzero(s->cells_top, s->nr_cells * sizeof(struct cell));
309
    for (int k = 0; k < s->nr_cells; k++)
310
311
      if (lock_init(&s->cells_top[k].lock) != 0)
        error("Failed to init spinlock.");
312
313

    /* Set the cell location and sizes. */
314
315
316
    for (int i = 0; i < cdim[0]; i++)
      for (int j = 0; j < cdim[1]; j++)
        for (int k = 0; k < cdim[2]; k++) {
317
          struct cell *restrict c = &s->cells_top[cell_getid(cdim, i, j, k)];
318
319
320
321
322
323
          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];
324
325
326
327
328
          c->dmin = dmin;
          c->depth = 0;
          c->count = 0;
          c->gcount = 0;
          c->super = c;
329
          c->gsuper = c;
330
          c->ti_old = ti_current;
331
          lock_init(&c->lock);
Pedro Gonnet's avatar
Pedro Gonnet committed
332
        }
333
334

    /* Be verbose about the change. */
335
336
337
    if (verbose)
      message("set cell dimensions to [ %i %i %i ].", cdim[0], cdim[1],
              cdim[2]);
338
339
    fflush(stdout);

340
#ifdef WITH_MPI
341
342
343
344
345
    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)
346
347
348
        message(
            "basic cell dimensions have increased - recalculating the "
            "global partition.");
349

Matthieu Schaller's avatar
Matthieu Schaller committed
350
      if (!partition_space_to_space(oldwidth, oldcdim, oldnodeIDs, s)) {
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368

        /* Failed, try another technique that requires no settings. */
        message("Failed to get a new partition, trying less optimal method");
        struct partition initial_partition;
#ifdef 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);
369

370
371
      /* Finished with these. */
      free(oldnodeIDs);
372
373
    }
#endif
374
375
376
377

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

378
379
  } /* re-build upper-level cells? */

380
  else { /* Otherwise, just clean up the cells. */
381
382

    /* Free the old cells, if they were allocated. */
383
    for (int k = 0; k < s->nr_cells; k++) {
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
      space_rebuild_recycle(s, &s->cells_top[k]);
      s->cells_top[k].sorts = NULL;
      s->cells_top[k].nr_tasks = 0;
      s->cells_top[k].nr_density = 0;
      s->cells_top[k].nr_gradient = 0;
      s->cells_top[k].nr_force = 0;
      s->cells_top[k].density = NULL;
      s->cells_top[k].gradient = NULL;
      s->cells_top[k].force = NULL;
      s->cells_top[k].dx_max = 0.0f;
      s->cells_top[k].sorted = 0;
      s->cells_top[k].count = 0;
      s->cells_top[k].gcount = 0;
      s->cells_top[k].init = NULL;
      s->cells_top[k].extra_ghost = NULL;
      s->cells_top[k].ghost = NULL;
      s->cells_top[k].kick = NULL;
      s->cells_top[k].super = &s->cells_top[k];
402
      s->cells_top[k].gsuper = &s->cells_top[k];
403
    }
404
405
    s->maxdepth = 0;
  }
406
407
408
409

  if (verbose)
    message("took %.3f %s.", clocks_from_ticks(getticks() - tic),
            clocks_getunit());
410
}
411
412
413
414
415
416

/**
 * @brief Re-build the cells as well as the tasks.
 *
 * @param s The #space in which to update the cells.
 * @param cell_max Maximal cell size.
417
 * @param verbose Print messages to stdout or not
418
419
 *
 */
420
void space_rebuild(struct space *s, double cell_max, int verbose) {
421

Matthieu Schaller's avatar
Matthieu Schaller committed
422
  const ticks tic = getticks();
423
424

  /* Be verbose about this. */
425
  // message("re)building space..."); fflush(stdout);
426
427

  /* Re-grid if necessary, or just re-set the cell data. */
428
  space_regrid(s, cell_max, verbose);
429

Pedro Gonnet's avatar
Pedro Gonnet committed
430
431
  size_t nr_parts = s->nr_parts;
  size_t nr_gparts = s->nr_gparts;
432
  struct cell *restrict cells_top = s->cells_top;
433
  const int ti_current = (s->e != NULL) ? s->e->ti_current : 0;
434

435
  const double ih[3] = {s->iwidth[0], s->iwidth[1], s->iwidth[2]};
Matthieu Schaller's avatar
Matthieu Schaller committed
436
437
  const double dim[3] = {s->dim[0], s->dim[1], s->dim[2]};
  const int cdim[3] = {s->cdim[0], s->cdim[1], s->cdim[2]};
438
439
440
441

  /* Run through the particles and get their cell index. */
  // tic = getticks();
  const size_t ind_size = s->size_parts;
442
443
  int *ind;
  if ((ind = (int *)malloc(sizeof(int) * ind_size)) == NULL)
444
    error("Failed to allocate temporary particle indices.");
Pedro Gonnet's avatar
Pedro Gonnet committed
445
  for (size_t k = 0; k < nr_parts; k++) {
446
447
    struct part *restrict p = &s->parts[k];
    for (int j = 0; j < 3; j++)
448
449
450
451
      if (p->x[j] < 0.0)
        p->x[j] += dim[j];
      else if (p->x[j] >= dim[j])
        p->x[j] -= dim[j];
452
    ind[k] =
453
        cell_getid(cdim, p->x[0] * ih[0], p->x[1] * ih[1], p->x[2] * ih[2]);
454
    cells_top[ind[k]].count++;
455
  }
Pedro Gonnet's avatar
Pedro Gonnet committed
456
457
  // message( "getting particle indices took %.3f %s." ,
  // clocks_from_ticks(getticks() - tic), clocks_getunit()):
458

459
460
461
462
463
464
  /* Run through the gravity particles and get their cell index. */
  // tic = getticks();
  const size_t gind_size = s->size_gparts;
  int *gind;
  if ((gind = (int *)malloc(sizeof(int) * gind_size)) == NULL)
    error("Failed to allocate temporary g-particle indices.");
465
  for (size_t k = 0; k < nr_gparts; k++) {
Matthieu Schaller's avatar
Matthieu Schaller committed
466
    struct gpart *restrict gp = &s->gparts[k];
467
468
469
470
471
472
473
    for (int j = 0; j < 3; j++)
      if (gp->x[j] < 0.0)
        gp->x[j] += dim[j];
      else if (gp->x[j] >= dim[j])
        gp->x[j] -= dim[j];
    gind[k] =
        cell_getid(cdim, gp->x[0] * ih[0], gp->x[1] * ih[1], gp->x[2] * ih[2]);
474
    cells_top[gind[k]].gcount++;
475
  }
476
// message( "getting g-particle indices took %.3f %s." ,
477
// clocks_from_ticks(getticks() - tic), clocks_getunit());
478
479

#ifdef WITH_MPI
480

481
  /* Move non-local parts to the end of the list. */
482
  const int local_nodeID = s->e->nodeID;
483
  for (size_t k = 0; k < nr_parts;) {
484
485
    if (cells_top[ind[k]].nodeID != local_nodeID) {
      cells_top[ind[k]].count -= 1;
486
      nr_parts -= 1;
Matthieu Schaller's avatar
Bug fix    
Matthieu Schaller committed
487
      const struct part tp = s->parts[k];
488
489
      s->parts[k] = s->parts[nr_parts];
      s->parts[nr_parts] = tp;
490
      if (s->parts[k].gpart != NULL) {
491
        s->parts[k].gpart->id_or_neg_offset = -k;
492
493
      }
      if (s->parts[nr_parts].gpart != NULL) {
494
        s->parts[nr_parts].gpart->id_or_neg_offset = -nr_parts;
495
      }
Matthieu Schaller's avatar
Bug fix    
Matthieu Schaller committed
496
      const struct xpart txp = s->xparts[k];
497
498
      s->xparts[k] = s->xparts[nr_parts];
      s->xparts[nr_parts] = txp;
Matthieu Schaller's avatar
Bug fix    
Matthieu Schaller committed
499
      const int t = ind[k];
500
501
      ind[k] = ind[nr_parts];
      ind[nr_parts] = t;
Matthieu Schaller's avatar
Matthieu Schaller committed
502
    } else {
503
504
505
506
      /* Increment when not exchanging otherwise we need to retest "k".*/
      k++;
    }
  }
507

508
#ifdef SWIFT_DEBUG_CHECKS
Peter W. Draper's avatar
Peter W. Draper committed
509
  /* Check that all parts are in the correct places. */
510
  for (size_t k = 0; k < nr_parts; k++) {
511
    if (cells_top[ind[k]].nodeID != local_nodeID) {
512
513
514
515
      error("Failed to move all non-local parts to send list");
    }
  }
  for (size_t k = nr_parts; k < s->nr_parts; k++) {
516
    if (cells_top[ind[k]].nodeID == local_nodeID) {
517
      error("Failed to remove local parts from send list");
518
    }
519
520
  }
#endif
521

522
  /* Move non-local gparts to the end of the list. */
523
  for (size_t k = 0; k < nr_gparts;) {
524
525
    if (cells_top[gind[k]].nodeID != local_nodeID) {
      cells_top[gind[k]].gcount -= 1;
526
      nr_gparts -= 1;
Matthieu Schaller's avatar
Bug fix    
Matthieu Schaller committed
527
      const struct gpart tp = s->gparts[k];
528
529
      s->gparts[k] = s->gparts[nr_gparts];
      s->gparts[nr_gparts] = tp;
530
531
      if (s->gparts[k].id_or_neg_offset <= 0) {
        s->parts[-s->gparts[k].id_or_neg_offset].gpart = &s->gparts[k];
532
      }
533
534
535
      if (s->gparts[nr_gparts].id_or_neg_offset <= 0) {
        s->parts[-s->gparts[nr_gparts].id_or_neg_offset].gpart =
            &s->gparts[nr_gparts];
536
      }
Matthieu Schaller's avatar
Bug fix    
Matthieu Schaller committed
537
538
539
      const int t = gind[k];
      gind[k] = gind[nr_gparts];
      gind[nr_gparts] = t;
Matthieu Schaller's avatar
Matthieu Schaller committed
540
    } else {
541
542
543
544
      /* Increment when not exchanging otherwise we need to retest "k".*/
      k++;
    }
  }
545

546
#ifdef SWIFT_DEBUG_CHECKS
547
548
  /* Check that all gparts are in the correct place (untested). */
  for (size_t k = 0; k < nr_gparts; k++) {
549
    if (cells_top[gind[k]].nodeID != local_nodeID) {
550
551
552
553
      error("Failed to move all non-local gparts to send list");
    }
  }
  for (size_t k = nr_gparts; k < s->nr_gparts; k++) {
554
    if (cells_top[gind[k]].nodeID == local_nodeID) {
555
      error("Failed to remove local gparts from send list");
556
    }
557
558
  }
#endif
559

560
561
  /* Exchange the strays, note that this potentially re-allocates
     the parts arrays. */
562
  size_t nr_parts_exchanged = s->nr_parts - nr_parts;
563
  size_t nr_gparts_exchanged = s->nr_gparts - nr_gparts;
Pedro Gonnet's avatar
Pedro Gonnet committed
564
565
566
567
  engine_exchange_strays(s->e, nr_parts, &ind[nr_parts], &nr_parts_exchanged,
                         nr_gparts, &gind[nr_gparts], &nr_gparts_exchanged);

  /* Set the new particle counts. */
568
  s->nr_parts = nr_parts + nr_parts_exchanged;
569
  s->nr_gparts = nr_gparts + nr_gparts_exchanged;
570
571

  /* Re-allocate the index array if needed.. */
572
  if (s->nr_parts > ind_size) {
573
574
    int *ind_new;
    if ((ind_new = (int *)malloc(sizeof(int) * s->nr_parts)) == NULL)
575
      error("Failed to allocate temporary particle indices.");
576
    memcpy(ind_new, ind, sizeof(int) * nr_parts);
577
578
    free(ind);
    ind = ind_new;
579
580
581
  }

  /* Assign each particle to its cell. */
Pedro Gonnet's avatar
Pedro Gonnet committed
582
  for (size_t k = nr_parts; k < s->nr_parts; k++) {
Matthieu Schaller's avatar
Matthieu Schaller committed
583
    const struct part *const p = &s->parts[k];
584
    ind[k] =
585
        cell_getid(cdim, p->x[0] * ih[0], p->x[1] * ih[1], p->x[2] * ih[2]);
586
    cells_top[ind[k]].count += 1;
587
#ifdef SWIFT_DEBUG_CHECKS
588
    if (cells_top[ind[k]].nodeID != local_nodeID)
589
      error("Received part that does not belong to me (nodeID=%i).",
590
            cells_top[ind[k]].nodeID);
591
#endif
592
  }
593
  nr_parts = s->nr_parts;
594
595

#endif /* WITH_MPI */
596
597

  /* Sort the parts according to their cells. */
598
  space_parts_sort(s, ind, nr_parts, 0, s->nr_cells - 1, verbose);
599
600

  /* Re-link the gparts. */
601
  if (nr_parts > 0 && nr_gparts > 0) part_relink_gparts(s->parts, nr_parts, 0);
602

603
#ifdef SWIFT_DEBUG_CHECKS
604
  /* Verify space_sort_struct. */
605
606
607
608
609
610
611
612
613
614
  for (size_t k = 1; k < nr_parts; k++) {
    if (ind[k - 1] > ind[k]) {
      error("Sort failed!");
    } else if (ind[k] != cell_getid(cdim, s->parts[k].x[0] * ih[0],
                                    s->parts[k].x[1] * ih[1],
                                    s->parts[k].x[2] * ih[2])) {
      error("Incorrect indices!");
    }
  }
#endif
615
616

  /* We no longer need the indices as of here. */
617
  free(ind);
618

619
620
621
622
#ifdef WITH_MPI

  /* Re-allocate the index array if needed.. */
  if (s->nr_gparts > gind_size) {
623
624
    int *gind_new;
    if ((gind_new = (int *)malloc(sizeof(int) * s->nr_gparts)) == NULL)
625
      error("Failed to allocate temporary g-particle indices.");
626
    memcpy(gind_new, gind, sizeof(int) * nr_gparts);
627
628
629
630
631
    free(gind);
    gind = gind_new;
  }

  /* Assign each particle to its cell. */
632
  for (size_t k = nr_gparts; k < s->nr_gparts; k++) {
Matthieu Schaller's avatar
Matthieu Schaller committed
633
    const struct gpart *const p = &s->gparts[k];
634
635
    gind[k] =
        cell_getid(cdim, p->x[0] * ih[0], p->x[1] * ih[1], p->x[2] * ih[2]);
636
637
638
639
640
641
642
    cells_top[gind[k]].gcount += 1;

#ifdef SWIFT_DEBUG_CHECKS
    if (cells_top[ind[k]].nodeID != s->e->nodeID)
      error("Received part that does not belong to me (nodeID=%i).",
            cells_top[ind[k]].nodeID);
#endif
643
644
645
646
  }
  nr_gparts = s->nr_gparts;

#endif
647

648
  /* Sort the gparts according to their cells. */
Matthieu Schaller's avatar
Matthieu Schaller committed
649
  space_gparts_sort(s, gind, nr_gparts, 0, s->nr_cells - 1, verbose);
650
651

  /* Re-link the parts. */
652
653
  if (nr_parts > 0 && nr_gparts > 0)
    part_relink_parts(s->gparts, nr_gparts, s->parts);
654
655

  /* We no longer need the indices as of here. */
656
  free(gind);
657

658
#ifdef SWIFT_DEBUG_CHECKS
659
660
661
  /* Verify that the links are correct */
  for (size_t k = 0; k < nr_gparts; ++k) {

662
    if (s->gparts[k].id_or_neg_offset < 0) {
663

664
      const struct part *part = &s->parts[-s->gparts[k].id_or_neg_offset];
665

666
667
668
669
      if (part->gpart != &s->gparts[k]) error("Linking problem !");

      if (s->gparts[k].x[0] != part->x[0] || s->gparts[k].x[1] != part->x[1] ||
          s->gparts[k].x[2] != part->x[2])
670
671
672
673
674
        error("Linked particles are not at the same position !");
    }
  }
  for (size_t k = 0; k < nr_parts; ++k) {

675
    if (s->parts[k].gpart != NULL &&
676
        s->parts[k].gpart->id_or_neg_offset != -(ptrdiff_t)k) {
677
      error("Linking problem !");
678
679
    }
  }
680
#endif
681

682
683
  /* Hook the cells up to the parts. */
  // tic = getticks();
684
685
686
  struct part *finger = s->parts;
  struct xpart *xfinger = s->xparts;
  struct gpart *gfinger = s->gparts;
687
  for (int k = 0; k < s->nr_cells; k++) {
688
    struct cell *restrict c = &cells_top[k];
689
    c->ti_old = ti_current;
690
691
692
693
694
695
696
    c->parts = finger;
    c->xparts = xfinger;
    c->gparts = gfinger;
    finger = &finger[c->count];
    xfinger = &xfinger[c->count];
    gfinger = &gfinger[c->gcount];
  }
697
  // message( "hooking up cells took %.3f %s." ,
Matthieu Schaller's avatar
Matthieu Schaller committed
698
  // clocks_from_ticks(getticks() - tic), clocks_getunit());
699
700
701

  /* At this point, we have the upper-level cells, old or new. Now make
     sure that the parts in each cell are ok. */
702
  space_split(s, cells_top, s->nr_cells, verbose);
703
704
705
706
707
708
709
710
711

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

/**
 * @brief Split particles between cells of a hierarchy
 *
712
713
 * This is done in parallel using threads in the #threadpool.
 *
714
 * @param s The #space.
715
716
 * @param cells The cell hierarchy.
 * @param nr_cells The number of cells.
717
718
 * @param verbose Are we talkative ?
 */
719
720
void space_split(struct space *s, struct cell *cells, int nr_cells,
                 int verbose) {
721

Matthieu Schaller's avatar
Matthieu Schaller committed
722
  const ticks tic = getticks();
723

724
  threadpool_map(&s->e->threadpool, space_split_mapper, cells, nr_cells,
725
                 sizeof(struct cell), 1, s);
726

727
728
729
  if (verbose)
    message("took %.3f %s.", clocks_from_ticks(getticks() - tic),
            clocks_getunit());
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
/**
 * @brief Runs through the top-level cells and checks whether tasks associated
 * with them can be split. If not, try to sanitize the cells.
 *
 * @param s The #space to act upon.
 */
void space_sanitize(struct space *s) {

  for (int k = 0; k < s->nr_cells; k++) {

    struct cell *c = &s->cells_top[k];
    const double min_width = c->dmin;

    /* Do we have a problem ? */
    if (c->h_max * kernel_gamma * space_stretch > min_width * 0.5 &&
        c->count > space_maxcount) {

      /* Ok, clean-up the mess */
      cell_sanitize(c);
    }
  }
}

755
/**
756
 * @brief Sort the particles and condensed particles according to the given
757
 * indices.
758
 *
Matthieu Schaller's avatar
Matthieu Schaller committed
759
 * @param s The #space.
760
761
762
763
 * @param ind The indices with respect to which the parts are sorted.
 * @param N The number of parts
 * @param min Lowest index.
 * @param max highest index.
764
 * @param verbose Are we talkative ?
765
 */
766
void space_parts_sort(struct space *s, int *ind, size_t N, int min, int max,
767
768
                      int verbose) {

Matthieu Schaller's avatar
Matthieu Schaller committed
769
  const ticks tic = getticks();
770

771
772
773
774
775
776
  /* Populate a parallel_sort structure with the input data */
  struct parallel_sort sort_struct;
  sort_struct.parts = s->parts;
  sort_struct.xparts = s->xparts;
  sort_struct.ind = ind;
  sort_struct.stack_size = 2 * (max - min + 1) + 10 + s->e->nr_threads;
777
778
  if ((sort_struct.stack =
           malloc(sizeof(struct qstack) * sort_struct.stack_size)) == NULL)
779
    error("Failed to allocate sorting stack.");
Matthieu Schaller's avatar
Matthieu Schaller committed
780
  for (unsigned int i = 0; i < sort_struct.stack_size; i++)
781
    sort_struct.stack[i].ready = 0;
782

783
  /* Add the first interval. */
784
785
786
787
788
789
790
791
792
793
794
  sort_struct.stack[0].i = 0;
  sort_struct.stack[0].j = N - 1;
  sort_struct.stack[0].min = min;
  sort_struct.stack[0].max = max;
  sort_struct.stack[0].ready = 1;
  sort_struct.first = 0;
  sort_struct.last = 1;
  sort_struct.waiting = 1;

  /* Launch the sorting tasks with a stride of zero such that the same
     map data is passed to each thread. */
795
796
  threadpool_map(&s->e->threadpool, space_parts_sort_mapper, &sort_struct,
                 s->e->threadpool.num_threads, 0, 1, NULL);
797

798
#ifdef SWIFT_DEBUG_CHECKS
799
  /* Verify space_sort_struct. */
800
  for (size_t i = 1; i < N; i++)
801
    if (ind[i - 1] > ind[i])
802
      error("Sorting failed (ind[%zu]=%i,ind[%zu]=%i), min=%i, max=%i.", i - 1,
803
804
805
            ind[i - 1], i, ind[i], min, max);
  message("Sorting succeeded.");
#endif
806

807
  /* Clean up. */
808
  free(sort_struct.stack);
809
810
811
812

  if (verbose)
    message("took %.3f %s.", clocks_from_ticks(getticks() - tic),
            clocks_getunit());
813
}
814

815
816
void space_parts_sort_mapper(void *map_data, int num_elements,
                             void *extra_data) {
817
818
819

  /* Unpack the mapping data. */
  struct parallel_sort *sort_struct = (struct parallel_sort *)map_data;
820

821
  /* Pointers to the sorting data. */
822
823
824
  int *ind = sort_struct->ind;
  struct part *parts = sort_struct->parts;
  struct xpart *xparts = sort_struct->xparts;
825

826
  /* Main loop. */
827
  while (sort_struct->waiting) {
828

829
    /* Grab an interval off the queue. */
830
    int qid = atomic_inc(&sort_struct->first) % sort_struct->stack_size;
831

832
    /* Wait for the entry to be ready, or for the sorting do be done. */
833
834
    while (!sort_struct->stack[qid].ready)
      if (!sort_struct->waiting) return;
835

836
    /* Get the stack entry. */
837
838
839
840
841
    ptrdiff_t i = sort_struct->stack[qid].i;
    ptrdiff_t j = sort_struct->stack[qid].j;
    int min = sort_struct->stack[qid].min;
    int max = sort_struct->stack[qid].max;
    sort_struct->stack[qid].ready = 0;
842

843
844
    /* Loop over sub-intervals. */
    while (1) {
845

846
      /* Bring beer. */
847
      const int pivot = (min + max) / 2;
848
849
      /* message("Working on interval [%i,%i] with min=%i, max=%i, pivot=%i.",
              i, j, min, max, pivot); */
850
851

      /* One pass of QuickSort's partitioning. */
852
853
      ptrdiff_t ii = i;
      ptrdiff_t jj = j;
854
855
856
857
      while (ii < jj) {
        while (ii <= j && ind[ii] <= pivot) ii++;
        while (jj >= i && ind[jj] > pivot) jj--;
        if (ii < jj) {
858
          size_t temp_i = ind[ii];
859
860
          ind[ii] = ind[jj];
          ind[jj] = temp_i;
861
          struct part temp_p = parts[ii];
862
863
          parts[ii] = parts[jj];
          parts[jj] = temp_p;
864
          struct xpart temp_xp = xparts[ii];
865
866
867
868
          xparts[ii] = xparts[jj];
          xparts[jj] = temp_xp;
        }
      }
869

870
#ifdef SWIFT_DEBUG_CHECKS
871
      /* Verify space_sort_struct. */
872
      for (int k = i; k <= jj; k++)
873
        if (ind[k] > pivot) {
874
875
          message("sorting failed at k=%i, ind[k]=%i, pivot=%i, i=%li, j=%li.",
                  k, ind[k], pivot, i, j);
876
877
878
879
          error("Partition failed (<=pivot).");
        }
      for (int k = jj + 1; k <= j; k++)
        if (ind[k] <= pivot) {
880
881
          message("sorting failed at k=%i, ind[k]=%i, pivot=%i, i=%li, j=%li.",
                  k, ind[k], pivot, i, j);
882
          error("Partition failed (>pivot).");
883
884
        }
#endif
885
886
887
888
889
890

      /* Split-off largest interval. */
      if (jj - i > j - jj + 1) {

        /* Recurse on the left? */
        if (jj > i && pivot > min) {
891
          qid = atomic_inc(&sort_struct->last) % sort_struct->stack_size;
892
          while (sort_struct->stack[qid].ready)
893
            ;
894
895
896
897
          sort_struct->stack[qid].i = i;
          sort_struct->stack[qid].j = jj;
          sort_struct->stack[qid].min = min;
          sort_struct->stack[qid].max = pivot;
898
          if (atomic_inc(&sort_struct->waiting) >= sort_struct->stack_size)
899
            error("Qstack overflow.");
900
          sort_struct->stack[qid].ready = 1;
901
        }
902

903
904
905
906
907
908
909
910
911
912
        /* Recurse on the right? */
        if (jj + 1 < j && pivot + 1 < max) {
          i = jj + 1;
          min = pivot + 1;
        } else
          break;

      } else {

        /* Recurse on the right? */
913
        if (pivot + 1 < max) {
914
          qid = atomic_inc(&sort_struct->last) % sort_struct->stack_size;
915
          while (sort_struct->stack[qid].ready)
916
            ;
917
918
919
920
          sort_struct->stack[qid].i = jj + 1;
          sort_struct->stack[qid].j = j;
          sort_struct->stack[qid].min = pivot + 1;
          sort_struct->stack[qid].max = max;
921
          if (atomic_inc(&sort_struct->waiting) >= sort_struct->stack_size)
922
            error("Qstack overflow.");
923
          sort_struct->stack[qid].ready = 1;
924
        }