cell.c 34.7 KB
Newer Older
1
/*******************************************************************************
2
 * This file is part of SWIFT.
3
 * Copyright (c) 2012 Pedro Gonnet (pedro.gonnet@durham.ac.uk)
4
5
6
7
 *                    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)
8
 *
9
10
11
12
 * 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.
13
 *
14
15
16
17
 * 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.
18
 *
19
20
 * 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/>.
21
 *
22
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
32
33
34
#include <pthread.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
35

36
37
/* MPI headers. */
#ifdef WITH_MPI
38
#include <mpi.h>
39
40
#endif

41
42
/* Switch off timers. */
#ifdef TIMER
43
#undef TIMER
44
45
#endif

46
47
48
/* This object's header. */
#include "cell.h"

49
/* Local headers. */
50
#include "active.h"
51
#include "atomic.h"
52
#include "drift.h"
53
#include "error.h"
54
#include "gravity.h"
55
#include "hydro.h"
Matthieu Schaller's avatar
Matthieu Schaller committed
56
#include "hydro_properties.h"
Pedro Gonnet's avatar
Pedro Gonnet committed
57
#include "memswap.h"
58
#include "minmax.h"
59
#include "scheduler.h"
60
61
#include "space.h"
#include "timers.h"
62

63
64
65
/* Global variables. */
int cell_next_tag = 0;

66
67
68
69
70
/**
 * @brief Get the size of the cell subtree.
 *
 * @param c The #cell.
 */
71
int cell_getsize(struct cell *c) {
72

Pedro Gonnet's avatar
Pedro Gonnet committed
73
74
  /* Number of cells in this subtree. */
  int count = 1;
75

76
77
  /* Sum up the progeny if split. */
  if (c->split)
Pedro Gonnet's avatar
Pedro Gonnet committed
78
    for (int k = 0; k < 8; k++)
79
80
81
82
83
84
85
      if (c->progeny[k] != NULL) count += cell_getsize(c->progeny[k]);

  /* Return the final count. */
  return count;
}

/**
86
87
88
89
90
91
92
93
 * @brief Unpack the data of a given cell and its sub-cells.
 *
 * @param pc An array of packed #pcell.
 * @param c The #cell in which to unpack the #pcell.
 * @param s The #space in which the cells are created.
 *
 * @return The number of cells created.
 */
94
95
int cell_unpack(struct pcell *pc, struct cell *c, struct space *s) {

96
97
#ifdef WITH_MPI

98
99
  /* Unpack the current pcell. */
  c->h_max = pc->h_max;
100
101
  c->ti_end_min = pc->ti_end_min;
  c->ti_end_max = pc->ti_end_max;
102
  c->ti_old = pc->ti_old;
103
  c->count = pc->count;
104
  c->gcount = pc->gcount;
105
  c->scount = pc->scount;
106
  c->tag = pc->tag;
Matthieu Schaller's avatar
Matthieu Schaller committed
107

108
109
  /* Number of new cells created. */
  int count = 1;
110
111

  /* Fill the progeny recursively, depth-first. */
Pedro Gonnet's avatar
Pedro Gonnet committed
112
  for (int k = 0; k < 8; k++)
113
    if (pc->progeny[k] >= 0) {
Pedro Gonnet's avatar
Pedro Gonnet committed
114
      struct cell *temp = space_getcell(s);
115
      temp->count = 0;
116
      temp->gcount = 0;
117
      temp->scount = 0;
118
119
120
      temp->loc[0] = c->loc[0];
      temp->loc[1] = c->loc[1];
      temp->loc[2] = c->loc[2];
121
122
123
      temp->width[0] = c->width[0] / 2;
      temp->width[1] = c->width[1] / 2;
      temp->width[2] = c->width[2] / 2;
124
      temp->dmin = c->dmin / 2;
125
126
127
      if (k & 4) temp->loc[0] += temp->width[0];
      if (k & 2) temp->loc[1] += temp->width[1];
      if (k & 1) temp->loc[2] += temp->width[2];
128
129
      temp->depth = c->depth + 1;
      temp->split = 0;
130
      temp->dx_max = 0.f;
131
132
      temp->nodeID = c->nodeID;
      temp->parent = c;
133
      temp->ti_old = c->ti_old;
134
135
136
      c->progeny[k] = temp;
      c->split = 1;
      count += cell_unpack(&pc[pc->progeny[k]], temp, s);
137
138
    }

139
  /* Return the total number of unpacked cells. */
140
  c->pcell_size = count;
141
  return count;
142
143
144
145
146

#else
  error("SWIFT was not compiled with MPI support.");
  return 0;
#endif
147
}
148

149
/**
150
 * @brief Link the cells recursively to the given #part array.
151
152
153
154
155
156
 *
 * @param c The #cell.
 * @param parts The #part array.
 *
 * @return The number of particles linked.
 */
157
int cell_link_parts(struct cell *c, struct part *parts) {
158

159
160
161
  c->parts = parts;

  /* Fill the progeny recursively, depth-first. */
Pedro Gonnet's avatar
Pedro Gonnet committed
162
163
164
165
  if (c->split) {
    int offset = 0;
    for (int k = 0; k < 8; k++) {
      if (c->progeny[k] != NULL)
166
        offset += cell_link_parts(c->progeny[k], &parts[offset]);
Pedro Gonnet's avatar
Pedro Gonnet committed
167
168
    }
  }
169

170
  /* Return the total number of linked particles. */
171
172
  return c->count;
}
173

174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
/**
 * @brief Link the cells recursively to the given #gpart array.
 *
 * @param c The #cell.
 * @param gparts The #gpart array.
 *
 * @return The number of particles linked.
 */
int cell_link_gparts(struct cell *c, struct gpart *gparts) {

  c->gparts = gparts;

  /* Fill the progeny recursively, depth-first. */
  if (c->split) {
    int offset = 0;
    for (int k = 0; k < 8; k++) {
      if (c->progeny[k] != NULL)
        offset += cell_link_gparts(c->progeny[k], &gparts[offset]);
    }
  }

  /* Return the total number of linked particles. */
  return c->gcount;
}

199
200
201
202
203
204
205
206
207
/**
 * @brief Pack the data of the given cell and all it's sub-cells.
 *
 * @param c The #cell.
 * @param pc Pointer to an array of packed cells in which the
 *      cells will be packed.
 *
 * @return The number of packed cells.
 */
208
209
int cell_pack(struct cell *c, struct pcell *pc) {

210
211
#ifdef WITH_MPI

212
213
  /* Start by packing the data of the current cell. */
  pc->h_max = c->h_max;
214
215
  pc->ti_end_min = c->ti_end_min;
  pc->ti_end_max = c->ti_end_max;
216
  pc->count = c->count;
217
  pc->gcount = c->gcount;
218
219
220
  c->tag = pc->tag = atomic_inc(&cell_next_tag) % cell_max_tag;

  /* Fill in the progeny, depth-first recursion. */
Pedro Gonnet's avatar
Pedro Gonnet committed
221
222
  int count = 1;
  for (int k = 0; k < 8; k++)
223
224
225
226
227
228
229
    if (c->progeny[k] != NULL) {
      pc->progeny[k] = count;
      count += cell_pack(c->progeny[k], &pc[count]);
    } else
      pc->progeny[k] = -1;

  /* Return the number of packed cells used. */
230
231
  c->pcell_size = count;
  return count;
232
233
234
235
236

#else
  error("SWIFT was not compiled with MPI support.");
  return 0;
#endif
237
238
}

239
240
241
242
243
244
245
246
/**
 * @brief Pack the time information of the given cell and all it's sub-cells.
 *
 * @param c The #cell.
 * @param ti_ends (output) The time information we pack into
 *
 * @return The number of packed cells.
 */
247
248
int cell_pack_ti_ends(struct cell *c, int *ti_ends) {

249
250
#ifdef WITH_MPI

251
252
  /* Pack this cell's data. */
  ti_ends[0] = c->ti_end_min;
253

254
255
256
257
258
259
260
261
262
  /* Fill in the progeny, depth-first recursion. */
  int count = 1;
  for (int k = 0; k < 8; k++)
    if (c->progeny[k] != NULL) {
      count += cell_pack_ti_ends(c->progeny[k], &ti_ends[count]);
    }

  /* Return the number of packed values. */
  return count;
263
264
265
266
267

#else
  error("SWIFT was not compiled with MPI support.");
  return 0;
#endif
268
269
}

270
271
272
273
/**
 * @brief Unpack the time information of a given cell and its sub-cells.
 *
 * @param c The #cell
274
 * @param ti_ends The time information to unpack
275
276
277
 *
 * @return The number of cells created.
 */
278
279
int cell_unpack_ti_ends(struct cell *c, int *ti_ends) {

280
281
#ifdef WITH_MPI

282
283
  /* Unpack this cell's data. */
  c->ti_end_min = ti_ends[0];
284

285
286
287
288
289
290
291
292
  /* Fill in the progeny, depth-first recursion. */
  int count = 1;
  for (int k = 0; k < 8; k++)
    if (c->progeny[k] != NULL) {
      count += cell_unpack_ti_ends(c->progeny[k], &ti_ends[count]);
    }

  /* Return the number of packed values. */
293
  return count;
294
295
296
297
298

#else
  error("SWIFT was not compiled with MPI support.");
  return 0;
#endif
299
}
300

301
/**
302
 * @brief Lock a cell for access to its array of #part and hold its parents.
303
304
 *
 * @param c The #cell.
305
 * @return 0 on success, 1 on failure
306
 */
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
int cell_locktree(struct cell *c) {

  TIMER_TIC

  /* First of all, try to lock this cell. */
  if (c->hold || lock_trylock(&c->lock) != 0) {
    TIMER_TOC(timer_locktree);
    return 1;
  }

  /* Did somebody hold this cell in the meantime? */
  if (c->hold) {

    /* Unlock this cell. */
    if (lock_unlock(&c->lock) != 0) error("Failed to unlock cell.");

    /* Admit defeat. */
    TIMER_TOC(timer_locktree);
    return 1;
  }

  /* Climb up the tree and lock/hold/unlock. */
Pedro Gonnet's avatar
Pedro Gonnet committed
329
  struct cell *finger;
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
  for (finger = c->parent; finger != NULL; finger = finger->parent) {

    /* Lock this cell. */
    if (lock_trylock(&finger->lock) != 0) break;

    /* Increment the hold. */
    atomic_inc(&finger->hold);

    /* Unlock the cell. */
    if (lock_unlock(&finger->lock) != 0) error("Failed to unlock cell.");
  }

  /* If we reached the top of the tree, we're done. */
  if (finger == NULL) {
    TIMER_TOC(timer_locktree);
    return 0;
  }

  /* Otherwise, we hit a snag. */
  else {

    /* Undo the holds up to finger. */
Pedro Gonnet's avatar
Pedro Gonnet committed
352
353
    for (struct cell *finger2 = c->parent; finger2 != finger;
         finger2 = finger2->parent)
354
      atomic_dec(&finger2->hold);
355
356
357
358
359
360
361
362
363
364

    /* Unlock this cell. */
    if (lock_unlock(&c->lock) != 0) error("Failed to unlock cell.");

    /* Admit defeat. */
    TIMER_TOC(timer_locktree);
    return 1;
  }
}

365
366
367
368
369
370
/**
 * @brief Lock a cell for access to its array of #gpart and hold its parents.
 *
 * @param c The #cell.
 * @return 0 on success, 1 on failure
 */
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
int cell_glocktree(struct cell *c) {

  TIMER_TIC

  /* First of all, try to lock this cell. */
  if (c->ghold || lock_trylock(&c->glock) != 0) {
    TIMER_TOC(timer_locktree);
    return 1;
  }

  /* Did somebody hold this cell in the meantime? */
  if (c->ghold) {

    /* Unlock this cell. */
    if (lock_unlock(&c->glock) != 0) error("Failed to unlock cell.");

    /* Admit defeat. */
    TIMER_TOC(timer_locktree);
    return 1;
  }

  /* Climb up the tree and lock/hold/unlock. */
Pedro Gonnet's avatar
Pedro Gonnet committed
393
  struct cell *finger;
394
395
396
397
398
399
  for (finger = c->parent; finger != NULL; finger = finger->parent) {

    /* Lock this cell. */
    if (lock_trylock(&finger->glock) != 0) break;

    /* Increment the hold. */
400
    atomic_inc(&finger->ghold);
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415

    /* Unlock the cell. */
    if (lock_unlock(&finger->glock) != 0) error("Failed to unlock cell.");
  }

  /* If we reached the top of the tree, we're done. */
  if (finger == NULL) {
    TIMER_TOC(timer_locktree);
    return 0;
  }

  /* Otherwise, we hit a snag. */
  else {

    /* Undo the holds up to finger. */
Pedro Gonnet's avatar
Pedro Gonnet committed
416
417
    for (struct cell *finger2 = c->parent; finger2 != finger;
         finger2 = finger2->parent)
418
      atomic_dec(&finger2->ghold);
419
420
421
422
423
424
425
426
427

    /* Unlock this cell. */
    if (lock_unlock(&c->glock) != 0) error("Failed to unlock cell.");

    /* Admit defeat. */
    TIMER_TOC(timer_locktree);
    return 1;
  }
}
428

429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
/**
 * @brief Lock a cell for access to its array of #spart and hold its parents.
 *
 * @param c The #cell.
 * @return 0 on success, 1 on failure
 */
int cell_slocktree(struct cell *c) {

  TIMER_TIC

  /* First of all, try to lock this cell. */
  if (c->shold || lock_trylock(&c->slock) != 0) {
    TIMER_TOC(timer_locktree);
    return 1;
  }

  /* Did somebody hold this cell in the meantime? */
  if (c->shold) {

    /* Unlock this cell. */
    if (lock_unlock(&c->slock) != 0) error("Failed to unlock cell.");

    /* Admit defeat. */
    TIMER_TOC(timer_locktree);
    return 1;
  }

  /* Climb up the tree and lock/hold/unlock. */
  struct cell *finger;
  for (finger = c->parent; finger != NULL; finger = finger->parent) {

    /* Lock this cell. */
    if (lock_trylock(&finger->slock) != 0) break;

    /* Increment the hold. */
    atomic_inc(&finger->shold);

    /* Unlock the cell. */
    if (lock_unlock(&finger->slock) != 0) error("Failed to unlock cell.");
  }

  /* If we reached the top of the tree, we're done. */
  if (finger == NULL) {
    TIMER_TOC(timer_locktree);
    return 0;
  }

  /* Otherwise, we hit a snag. */
  else {

    /* Undo the holds up to finger. */
    for (struct cell *finger2 = c->parent; finger2 != finger;
         finger2 = finger2->parent)
      atomic_dec(&finger2->shold);

    /* Unlock this cell. */
    if (lock_unlock(&c->slock) != 0) error("Failed to unlock cell.");

    /* Admit defeat. */
    TIMER_TOC(timer_locktree);
    return 1;
  }
}

493
/**
494
 * @brief Unlock a cell's parents for access to #part array.
495
496
497
 *
 * @param c The #cell.
 */
498
499
500
501
502
503
504
505
void cell_unlocktree(struct cell *c) {

  TIMER_TIC

  /* First of all, try to unlock this cell. */
  if (lock_unlock(&c->lock) != 0) error("Failed to unlock cell.");

  /* Climb up the tree and unhold the parents. */
Pedro Gonnet's avatar
Pedro Gonnet committed
506
  for (struct cell *finger = c->parent; finger != NULL; finger = finger->parent)
507
    atomic_dec(&finger->hold);
508
509
510
511

  TIMER_TOC(timer_locktree);
}

512
513
514
515
516
/**
 * @brief Unlock a cell's parents for access to #gpart array.
 *
 * @param c The #cell.
 */
517
518
519
520
521
522
523
524
void cell_gunlocktree(struct cell *c) {

  TIMER_TIC

  /* First of all, try to unlock this cell. */
  if (lock_unlock(&c->glock) != 0) error("Failed to unlock cell.");

  /* Climb up the tree and unhold the parents. */
Pedro Gonnet's avatar
Pedro Gonnet committed
525
  for (struct cell *finger = c->parent; finger != NULL; finger = finger->parent)
526
    atomic_dec(&finger->ghold);
527
528
529
530

  TIMER_TOC(timer_locktree);
}

531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
/**
 * @brief Unlock a cell's parents for access to #spart array.
 *
 * @param c The #cell.
 */
void cell_sunlocktree(struct cell *c) {

  TIMER_TIC

  /* First of all, try to unlock this cell. */
  if (lock_unlock(&c->slock) != 0) error("Failed to unlock cell.");

  /* Climb up the tree and unhold the parents. */
  for (struct cell *finger = c->parent; finger != NULL; finger = finger->parent)
    atomic_dec(&finger->shold);

  TIMER_TOC(timer_locktree);
}

550
551
552
553
/**
 * @brief Sort the parts into eight bins along the given pivots.
 *
 * @param c The #cell array to be sorted.
554
555
 * @param parts_offset Offset of the cell parts array relative to the
 *        space's parts array, i.e. c->parts - s->parts.
556
557
 * @param buff A buffer with at least max(c->count, c->gcount) entries,
 *        used for sorting indices.
Peter W. Draper's avatar
Peter W. Draper committed
558
559
 * @param gbuff A buffer with at least max(c->count, c->gcount) entries,
 *        used for sorting indices for the gparts.
560
 */
561
562
void cell_split(struct cell *c, ptrdiff_t parts_offset, ptrdiff_t sparts_offset,
                struct cell_buff *buff, struct cell_buff *sbuff,
563
                struct cell_buff *gbuff) {
564

565
  const int count = c->count, gcount = c->gcount, scount = c->scount;
566
567
568
  struct part *parts = c->parts;
  struct xpart *xparts = c->xparts;
  struct gpart *gparts = c->gparts;
569
570
571
572
573
574
  const double pivot[3] = {c->loc[0] + c->width[0] / 2,
                           c->loc[1] + c->width[1] / 2,
                           c->loc[2] + c->width[2] / 2};
  int bucket_count[8] = {0, 0, 0, 0, 0, 0, 0, 0};
  int bucket_offset[9];

575
576
577
#ifdef SWIFT_DEBUG_CHECKS
  /* Check that the buffs are OK. */
  for (int k = 0; k < count; k++) {
578
    if (buff[k].x[0] != parts[k].x[0] || buff[k].x[1] != parts[k].x[1] ||
579
        buff[k].x[2] != parts[k].x[2])
580
581
      error("Inconsistent buff contents.");
  }
582
583
584
585
586
587
588
589
590
591
  for (int k = 0; k < gcount; k++) {
    if (gbuff[k].x[0] != gparts[k].x[0] || gbuff[k].x[1] != gparts[k].x[1] ||
        gbuff[k].x[2] != gparts[k].x[2])
      error("Inconsistent gbuff contents.");
  }
  for (int k = 0; k < scount; k++) {
    if (sbuff[k].x[0] != sparts[k].x[0] || sbuff[k].x[1] != sparts[k].x[1] ||
        sbuff[k].x[2] != sparts[k].x[2])
      error("Inconsistent sbuff contents.");
  }
592
#endif /* SWIFT_DEBUG_CHECKS */
593
594
595
596
597
598
599

  /* Fill the buffer with the indices. */
  for (int k = 0; k < count; k++) {
    const int bid = (parts[k].x[0] > pivot[0]) * 4 +
                    (parts[k].x[1] > pivot[1]) * 2 + (parts[k].x[2] > pivot[2]);
    bucket_count[bid]++;
    buff[k] = bid;
600
  }
601

602
603
604
605
606
  /* Set the buffer offsets. */
  bucket_offset[0] = 0;
  for (int k = 1; k <= 8; k++) {
    bucket_offset[k] = bucket_offset[k - 1] + bucket_count[k - 1];
    bucket_count[k - 1] = 0;
607
608
  }

609
610
611
612
613
614
615
616
617
618
619
620
621
622
  /* Run through the buckets, and swap particles to their correct spot. */
  for (int bucket = 0; bucket < 8; bucket++) {
    for (int k = bucket_offset[bucket] + bucket_count[bucket];
         k < bucket_offset[bucket + 1]; k++) {
      int bid = buff[k];
      if (bid != bucket) {
        struct part part = parts[k];
        struct xpart xpart = xparts[k];
        while (bid != bucket) {
          int j = bucket_offset[bid] + bucket_count[bid]++;
          while (buff[j] == bid) {
            j++;
            bucket_count[bid]++;
          }
Pedro Gonnet's avatar
Pedro Gonnet committed
623
624
625
          memswap(&parts[j], &part, sizeof(struct part));
          memswap(&xparts[j], &xpart, sizeof(struct xpart));
          memswap(&buff[j], &bid, sizeof(int));
626
627
628
629
        }
        parts[k] = part;
        xparts[k] = xpart;
        buff[k] = bid;
630
      }
631
      bucket_count[bid]++;
632
633
634
635
    }
  }

  /* Store the counts and offsets. */
Pedro Gonnet's avatar
Pedro Gonnet committed
636
  for (int k = 0; k < 8; k++) {
637
638
639
    c->progeny[k]->count = bucket_count[k];
    c->progeny[k]->parts = &c->parts[bucket_offset[k]];
    c->progeny[k]->xparts = &c->xparts[bucket_offset[k]];
640
641
642
  }

  /* Re-link the gparts. */
643
644
  if (count > 0 && gcount > 0)
    part_relink_gparts_to_parts(parts, count, parts_offset);
645

646
#ifdef SWIFT_DEBUG_CHECKS
647
  /* Verify that _all_ the parts have been assigned to a cell. */
648
649
650
651
652
653
654
655
  for (int k = 1; k < 8; k++)
    if (&c->progeny[k - 1]->parts[c->progeny[k - 1]->count] !=
        c->progeny[k]->parts)
      error("Particle sorting failed (internal consistency).");
  if (c->progeny[0]->parts != c->parts)
    error("Particle sorting failed (left edge).");
  if (&c->progeny[7]->parts[c->progeny[7]->count] != &c->parts[count])
    error("Particle sorting failed (right edge).");
656
657

  /* Verify a few sub-cells. */
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
  for (int k = 0; k < c->progeny[0]->count; k++)
    if (c->progeny[0]->parts[k].x[0] > pivot[0] ||
        c->progeny[0]->parts[k].x[1] > pivot[1] ||
        c->progeny[0]->parts[k].x[2] > pivot[2])
      error("Sorting failed (progeny=0).");
  for (int k = 0; k < c->progeny[1]->count; k++)
    if (c->progeny[1]->parts[k].x[0] > pivot[0] ||
        c->progeny[1]->parts[k].x[1] > pivot[1] ||
        c->progeny[1]->parts[k].x[2] <= pivot[2])
      error("Sorting failed (progeny=1).");
  for (int k = 0; k < c->progeny[2]->count; k++)
    if (c->progeny[2]->parts[k].x[0] > pivot[0] ||
        c->progeny[2]->parts[k].x[1] <= pivot[1] ||
        c->progeny[2]->parts[k].x[2] > pivot[2])
      error("Sorting failed (progeny=2).");
#endif
674

675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
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
  /* Now do the same song and dance for the sparts. */
  for (int k = 0; k < 8; k++) bucket_count[k] = 0;

  /* Fill the buffer with the indices. */
  for (int k = 0; k < scount; k++) {
    const int bid = (sbuff[k].x[0] > pivot[0]) * 4 +
                    (sbuff[k].x[1] > pivot[1]) * 2 + (sbuff[k].x[2] > pivot[2]);
    bucket_count[bid]++;
    sbuff[k].ind = bid;
  }

  /* Set the buffer offsets. */
  bucket_offset[0] = 0;
  for (int k = 1; k <= 8; k++) {
    bucket_offset[k] = bucket_offset[k - 1] + bucket_count[k - 1];
    bucket_count[k - 1] = 0;
  }

  /* Run through the buckets, and swap particles to their correct spot. */
  for (int bucket = 0; bucket < 8; bucket++) {
    for (int k = bucket_offset[bucket] + bucket_count[bucket];
         k < bucket_offset[bucket + 1]; k++) {
      int bid = sbuff[k].ind;
      if (bid != bucket) {
        struct spart spart = sparts[k];
        struct cell_buff temp_buff = sbuff[k];
        while (bid != bucket) {
          int j = bucket_offset[bid] + bucket_count[bid]++;
          while (sbuff[j].ind == bid) {
            j++;
            bucket_count[bid]++;
          }
          memswap(&sparts[j], &spart, sizeof(struct spart));
          memswap(&sbuff[j], &temp_buff, sizeof(struct cell_buff));
          bid = temp_buff.ind;
        }
        sparts[k] = spart;
        sbuff[k] = temp_buff;
      }
      bucket_count[bid]++;
    }
  }

  /* Store the counts and offsets. */
  for (int k = 0; k < 8; k++) {
    c->progeny[k]->scount = bucket_count[k];
    c->progeny[k]->sparts = &c->sparts[bucket_offset[k]];
  }

  /* Re-link the gparts. */
  if (scount > 0 && gcount > 0)
    part_relink_gparts_to_sparts(sparts, count, sparts_offset);

  /* Finally, do the same song and dance for the gparts. */
729
730
731
732
733
734
735
736
737
  for (int k = 0; k < 8; k++) bucket_count[k] = 0;

  /* Fill the buffer with the indices. */
  for (int k = 0; k < gcount; k++) {
    const int bid = (gparts[k].x[0] > pivot[0]) * 4 +
                    (gparts[k].x[1] > pivot[1]) * 2 +
                    (gparts[k].x[2] > pivot[2]);
    bucket_count[bid]++;
    buff[k] = bid;
738
  }
739
740
741
742
743
744

  /* Set the buffer offsets. */
  bucket_offset[0] = 0;
  for (int k = 1; k <= 8; k++) {
    bucket_offset[k] = bucket_offset[k - 1] + bucket_count[k - 1];
    bucket_count[k - 1] = 0;
745
746
  }

747
748
749
750
751
752
753
754
755
756
757
758
759
  /* Run through the buckets, and swap particles to their correct spot. */
  for (int bucket = 0; bucket < 8; bucket++) {
    for (int k = bucket_offset[bucket] + bucket_count[bucket];
         k < bucket_offset[bucket + 1]; k++) {
      int bid = buff[k];
      if (bid != bucket) {
        struct gpart gpart = gparts[k];
        while (bid != bucket) {
          int j = bucket_offset[bid] + bucket_count[bid]++;
          while (buff[j] == bid) {
            j++;
            bucket_count[bid]++;
          }
Pedro Gonnet's avatar
Pedro Gonnet committed
760
761
          memswap(&gparts[j], &gpart, sizeof(struct gpart));
          memswap(&buff[j], &bid, sizeof(int));
762
763
764
        }
        gparts[k] = gpart;
        buff[k] = bid;
765
      }
766
      bucket_count[bid]++;
767
768
769
770
    }
  }

  /* Store the counts and offsets. */
Pedro Gonnet's avatar
Pedro Gonnet committed
771
  for (int k = 0; k < 8; k++) {
772
773
    c->progeny[k]->gcount = bucket_count[k];
    c->progeny[k]->gparts = &c->gparts[bucket_offset[k]];
774
775
776
  }

  /* Re-link the parts. */
777
  if (count > 0 && gcount > 0)
778
    part_relink_parts_to_gparts(gparts, gcount, parts - parts_offset);
779
780
781
782

  /* Re-link the sparts. */
  if (scount > 0 && gcount > 0)
    part_relink_sparts_to_gparts(gparts, gcount, sparts - sparts_offset);
783
}
784

785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
/**
 * @brief Sanitizes the smoothing length values of cells by setting large
 * outliers to more sensible values.
 *
 * We compute the mean and standard deviation of the smoothing lengths in
 * logarithmic space and limit values to mean + 4 sigma.
 *
 * @param c The cell.
 */
void cell_sanitize(struct cell *c) {

  const int count = c->count;
  struct part *parts = c->parts;

  /* First collect some statistics */
  float h_mean = 0.f, h_mean2 = 0.f;
  float h_min = FLT_MAX, h_max = 0.f;
  for (int i = 0; i < count; ++i) {

804
    const float h = logf(parts[i].h);
805
806
807
808
809
810
811
812
    h_mean += h;
    h_mean2 += h * h;
    h_max = max(h_max, h);
    h_min = min(h_min, h);
  }
  h_mean /= count;
  h_mean2 /= count;
  const float h_var = h_mean2 - h_mean * h_mean;
813
  const float h_std = (h_var > 0.f) ? sqrtf(h_var) : 0.1f * h_mean;
814
815

  /* Choose a cut */
816
  const float h_limit = expf(h_mean + 4.f * h_std);
817
818

  /* Be verbose this is not innocuous */
819
820
  message("Cell properties: h_min= %f h_max= %f geometric mean= %f.",
          expf(h_min), expf(h_max), expf(h_mean));
821
822
823

  if (c->h_max > h_limit) {

824
    message("Smoothing lengths will be limited to (mean + 4sigma)= %f.",
825
826
827
828
829
830
            h_limit);

    /* Apply the cut */
    for (int i = 0; i < count; ++i) parts->h = min(parts[i].h, h_limit);

    c->h_max = h_limit;
831
832
833
834

  } else {

    message("Smoothing lengths will not be limited.");
835
836
837
  }
}

838
/**
839
 * @brief Converts hydro quantities to a valid state after the initial density
840
 * calculation
841
842
843
844
845
846
847
 *
 * @param c Cell to act upon
 * @param data Unused parameter
 */
void cell_convert_hydro(struct cell *c, void *data) {

  struct part *p = c->parts;
848
849

  for (int i = 0; i < c->count; ++i) {
850
851
852
853
    hydro_convert_quantities(&p[i]);
  }
}

Matthieu Schaller's avatar
Matthieu Schaller committed
854
855
856
857
858
859
/**
 * @brief Cleans the links in a given cell.
 *
 * @param c Cell to act upon
 * @param data Unused parameter
 */
860
void cell_clean_links(struct cell *c, void *data) {
Matthieu Schaller's avatar
Matthieu Schaller committed
861
  c->density = NULL;
862
  c->gradient = NULL;
Matthieu Schaller's avatar
Matthieu Schaller committed
863
  c->force = NULL;
864
  c->grav = NULL;
Matthieu Schaller's avatar
Matthieu Schaller committed
865
}
866

867
868
869
870
871
872
873
874
875
876
877
878
/**
 * @brief Checks that a cell is at the current point in time
 *
 * Calls error() if the cell is not at the current time.
 *
 * @param c Cell to act upon
 * @param data The current time on the integer time-line
 */
void cell_check_drift_point(struct cell *c, void *data) {

  const int ti_current = *(int *)data;

879
  if (c->ti_old != ti_current && c->nodeID == engine_rank)
880
881
882
883
    error("Cell in an incorrect time-zone! c->ti_old=%d ti_current=%d",
          c->ti_old, ti_current);
}

884
885
886
887
888
889
890
891
892
893
894
895
/**
 * @brief Checks whether the cells are direct neighbours ot not. Both cells have
 * to be of the same size
 *
 * @param ci First #cell.
 * @param cj Second #cell.
 *
 * @todo Deal with periodicity.
 */
int cell_are_neighbours(const struct cell *restrict ci,
                        const struct cell *restrict cj) {

Matthieu Schaller's avatar
Matthieu Schaller committed
896
#ifdef SWIFT_DEBUG_CHECKS
Matthieu Schaller's avatar
Matthieu Schaller committed
897
  if (ci->width[0] != cj->width[0]) error("Cells of different size !");
898
899
900
#endif

  /* Maximum allowed distance */
901
902
  const double min_dist =
      1.2 * ci->width[0]; /* 1.2 accounts for rounding errors */
903
904
905
906
907

  /* (Manhattan) Distance between the cells */
  for (int k = 0; k < 3; k++) {
    const double center_i = ci->loc[k];
    const double center_j = cj->loc[k];
908
    if (fabs(center_i - center_j) > min_dist) return 0;
909
910
911
912
913
  }

  return 1;
}

914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
/**
 * @brief Computes the multi-pole brutally and compare to the
 * recursively computed one.
 *
 * @param c Cell to act upon
 * @param data Unused parameter
 */
void cell_check_multipole(struct cell *c, void *data) {

  struct multipole ma;

  if (c->gcount > 0) {

    /* Brute-force calculation */
    multipole_init(&ma, c->gparts, c->gcount);

    /* Compare with recursive one */
    struct multipole mb = c->multipole;

    if (fabsf(ma.mass - mb.mass) / fabsf(ma.mass + mb.mass) > 1e-5)
      error("Multipole masses are different (%12.15e vs. %12.15e)", ma.mass,
            mb.mass);

    for (int k = 0; k < 3; ++k)
938
      if (fabs(ma.CoM[k] - mb.CoM[k]) / fabs(ma.CoM[k] + mb.CoM[k]) > 1e-5)
939
940
941
        error("Multipole CoM are different (%12.15e vs. %12.15e", ma.CoM[k],
              mb.CoM[k]);

942
#if const_gravity_multipole_order >= 2
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
    if (fabsf(ma.I_xx - mb.I_xx) / fabsf(ma.I_xx + mb.I_xx) > 1e-5 &&
        ma.I_xx > 1e-9)
      error("Multipole I_xx are different (%12.15e vs. %12.15e)", ma.I_xx,
            mb.I_xx);
    if (fabsf(ma.I_yy - mb.I_yy) / fabsf(ma.I_yy + mb.I_yy) > 1e-5 &&
        ma.I_yy > 1e-9)
      error("Multipole I_yy are different (%12.15e vs. %12.15e)", ma.I_yy,
            mb.I_yy);
    if (fabsf(ma.I_zz - mb.I_zz) / fabsf(ma.I_zz + mb.I_zz) > 1e-5 &&
        ma.I_zz > 1e-9)
      error("Multipole I_zz are different (%12.15e vs. %12.15e)", ma.I_zz,
            mb.I_zz);
    if (fabsf(ma.I_xy - mb.I_xy) / fabsf(ma.I_xy + mb.I_xy) > 1e-5 &&
        ma.I_xy > 1e-9)
      error("Multipole I_xy are different (%12.15e vs. %12.15e)", ma.I_xy,
            mb.I_xy);
    if (fabsf(ma.I_xz - mb.I_xz) / fabsf(ma.I_xz + mb.I_xz) > 1e-5 &&
        ma.I_xz > 1e-9)
      error("Multipole I_xz are different (%12.15e vs. %12.15e)", ma.I_xz,
            mb.I_xz);
    if (fabsf(ma.I_yz - mb.I_yz) / fabsf(ma.I_yz + mb.I_yz) > 1e-5 &&
        ma.I_yz > 1e-9)
      error("Multipole I_yz are different (%12.15e vs. %12.15e)", ma.I_yz,
            mb.I_yz);
967
#endif
968
  }
969
970
}

971
/**
972
 * @brief Frees up the memory allocated for this #cell.
973
 *
974
 * @param c The #cell.
975
 */
976
977
978
979
980
981
982
void cell_clean(struct cell *c) {

  free(c->sort);

  /* Recurse */
  for (int k = 0; k < 8; k++)
    if (c->progeny[k]) cell_clean(c->progeny[k]);
983
}
984
985
986
987
988

/**
 * @brief Checks whether a given cell needs drifting or not.
 *
 * @param c the #cell.
989
 * @param e The #engine (holding current time information).
990
991
992
 *
 * @return 1 If the cell needs drifting, 0 otherwise.
 */
993
int cell_is_drift_needed(struct cell *c, const struct engine *e) {
994
995

  /* Do we have at least one active particle in the cell ?*/
996
  if (cell_is_active(c, e)) return 1;
997
998
999
1000
1001
1002
1003

  /* Loop over the pair tasks that involve this cell */
  for (struct link *l = c->density; l != NULL; l = l->next) {

    if (l->t->type != task_type_pair && l->t->type != task_type_sub_pair)
      continue;

1004
1005
1006
    /* Is the other cell in the pair active ? */
    if ((l->t->ci == c && cell_is_active(l->t->cj, e)) ||
        (l->t->cj == c && cell_is_active(l->t->ci, e)))
1007
      return 1;
1008
1009
1010
1011
1012
  }

  /* No neighbouring cell has active particles. Drift not necessary */
  return 0;
}
1013
1014
1015
1016
1017
1018

/**
 * @brief Un-skips all the tasks associated with a given cell and checks
 * if the space needs to be rebuilt.
 *
 * @param c the #cell.
Peter W. Draper's avatar
Peter W. Draper committed
1019
 * @param s the #scheduler.
1020
1021
1022
 *
 * @return 1 If the space needs rebuilding. 0 otherwise.
 */
1023
int cell_unskip_tasks(struct cell *c, struct scheduler *s) {
1024
1025
1026
1027
1028
1029

  /* Un-skip the density tasks involved with this cell. */
  for (struct link *l = c->density; l != NULL; l = l->next) {
    struct task *t = l->t;
    const struct cell *ci = t->ci;
    const struct cell *cj = t->cj;
1030
    scheduler_activate(s, t);
1031
1032
1033
1034
1035

    /* Set the correct sorting flags */
    if (t->type == task_type_pair) {
      if (!(ci->sorted & (1 << t->flags))) {
        atomic_or(&ci->sorts->flags, (1 << t->flags));
1036
        scheduler_activate(s, ci->sorts);
1037
1038
1039
      }
      if (!(cj->sorted & (1 << t->flags))) {
        atomic_or(&cj->sorts->flags, (1 << t->flags));
1040
        scheduler_activate(s, cj->sorts);
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
      }
    }

    /* Check whether there was too much particle motion */
    if (t->type == task_type_pair || t->type == task_type_sub_pair) {
      if (t->tight &&
          (max(ci->h_max, cj->h_max) + ci->dx_max + cj->dx_max > cj->dmin ||
           ci->dx_max > space_maxreldx * ci->h_max ||
           cj->dx_max > space_maxreldx * cj->h_max))
        return 1;

#ifdef WITH_MPI
1053
      /* Activate the send/recv flags. */
1054
      if (ci->nodeID != engine_rank) {
1055
1056

        /* Activate the tasks to recv foreign cell ci's data. */
1057
1058
1059
        scheduler_activate(s, ci->recv_xv);
        scheduler_activate(s, ci->recv_rho);
        scheduler_activate(s, ci->recv_ti);
1060
1061
1062
1063

        /* Look for the local cell cj's send tasks. */
        struct link *l = NULL;
        for (l = cj->send_xv; l != NULL && l->t->cj->nodeID != ci->nodeID;
1064
1065
             l = l->next)
          ;
1066
        if (l == NULL) error("Missing link to send_xv task.");
1067
        scheduler_activate(s, l->t);
1068

Matthieu Schaller's avatar
Matthieu Schaller committed
1069
1070
1071
1072
        if (cj->super->drift)
          scheduler_activate(s, cj->super->drift);
        else
          error("Drift task missing !");
1073
1074

        for (l = cj->send_rho; l != NULL && l->t->cj->nodeID != ci->nodeID;
1075
1076
             l = l->next)
          ;
1077
        if (l == NULL) error("Missing link to send_rho task.");
1078
        scheduler_activate(s, l->t);
1079
1080

        for (l = cj->send_ti; l != NULL && l->t->cj->nodeID != ci->nodeID;
1081
1082
             l = l->next)
          ;
1083
        if (l == NULL) error("Missing link to send_ti task.");
1084
        scheduler_activate(s, l->t);
1085

1086
      } else if (cj->nodeID != engine_rank) {
1087
1088

        /* Activate the tasks to recv foreign cell cj's data. */
1089
1090
1091
        scheduler_activate(s, cj->recv_xv);
        scheduler_activate(s, cj->recv_rho);
        scheduler_activate(s, cj->recv_ti);
Matthieu Schaller's avatar
Matthieu Schaller committed
1092

1093
1094
1095
        /* Look for the local cell ci's send tasks. */
        struct link *l = NULL;
        for (l = ci->send_xv; l != NULL && l->t->cj->nodeID != cj->nodeID;
1096
1097
             l = l->next)
          ;
1098
        if (l == NULL) error("Missing link to send_xv task.");
1099
        scheduler_activate(s, l->t);
1100

Matthieu Schaller's avatar
Matthieu Schaller committed
1101
1102
1103
1104
        if (ci->super->drift)
          scheduler_activate(s, ci->super->drift);
        else
          error("Drift task missing !");
1105
1106

        for (l = ci->send_rho; l != NULL && l->t->cj->nodeID != cj->nodeID;
1107
1108
             l = l->next)
          ;
1109
        if (l == NULL) error("Missing link to send_rho task.");
1110
        scheduler_activate(s, l->t);
1111
1112

        for (l = ci->send_ti; l != NULL && l->t->cj->nodeID != cj->nodeID;
1113
1114
             l = l->next)
          ;
1115
        if (l == NULL) error("Missing link to send_ti task.");
1116
        scheduler_activate(s, l->t);
1117
1118
1119
1120
1121
1122
      }
#endif
    }
  }

  /* Unskip all the other task types. */
1123
  for (struct link *l = c->gradient; l != NULL; l = l->next)
1124
    scheduler_activate(s, l->t);
1125
  for (struct link *l = c->force; l != NULL; l = l->next)
1126
    scheduler_activate(s, l->t);
1127
  for (struct link *l = c->grav; l != NULL; l = l->next)
1128
1129
1130
1131
    scheduler_activate(s, l->t);
  if (c->extra_ghost != NULL) scheduler_activate(s, c->extra_ghost);
  if (c->ghost != NULL) scheduler_activate(s, c->ghost);
  if (c->init != NULL) scheduler_activate(s, c->init);
Matthieu Schaller's avatar
Matthieu Schaller committed
1132
  if (c->drift != NULL) scheduler_activate(s, c->drift);
1133
1134
1135
  if (c->kick != NULL) scheduler_activate(s, c->kick);
  if (c->cooling != NULL) scheduler_activate(s, c->cooling);
  if (c->sourceterms != NULL) scheduler_activate(s, c->sourceterms);
1136
1137
1138

  return 0;
}
1139

1140
1141
1142
1143
1144
1145
1146
1147
1148
/**
 * @brief Set the super-cell pointers for all cells in a hierarchy.
 *
 * @param c The top-level #cell to play with.
 * @param super Pointer to the deepest cell with tasks in this part of the tree.
 */
void cell_set_super(struct cell *c, struct cell *super) {

  /* Are we in a cell with some kind of self/pair task ? */
1149
  if (super == NULL && c->nr_tasks > 0) super = c;
1150
1151
1152
1153

  /* Set the super-cell */
  c->super = super;

1154
1155
  /* Recurse */
  if (c->split)
1156
1157
1158
    for (int k = 0; k < 8; k++)
      if (c->progeny[k] != NULL) cell_set_super(c->progeny[k], super);
}
1159

1160
1161
1162
1163
1164
1165
/**
 * @brief Recursively drifts all particles and g-particles in a cell hierarchy.
 *
 * @param c The #cell.
 * @param e The #engine (to get ti_current).
 */
1166
void cell_drift(struct cell *c, const struct engine *e) {
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177

  const double timeBase = e->timeBase;
  const int ti_old = c->ti_old;
  const int ti_current = e->ti_current;
  struct part *const parts = c->parts;
  struct xpart *const xparts = c->xparts;
  struct gpart *const gparts = c->gparts;

  /* Drift from the last time the cell was drifted to the current time */
  const double dt = (ti_current - ti_old) * timeBase;
  float dx_max = 0.f, dx2_max = 0.f, h_max = 0.f;
1178

1179
  /* Check that we are actually going to move forward. */
Matthieu Schaller's avatar
Matthieu Schaller committed
1180
  if (ti_current < ti_old) error("Attempt to drift to the past");
1181

1182
  /* Are we not in a leaf ? */
1183
1184
1185
1186
1187
1188
  if (c->split) {

    /* Loop over the progeny and collect their data. */
    for (int k = 0; k < 8; k++)
      if (c->progeny[k] != NULL) {
        struct cell *cp = c->progeny[k];
1189
1190
        cell_drift(cp, e);
        dx_max = max(dx_max, cp->dx_max);
1191
1192
        h_max = max(h_max, cp->h_max);
      }
1193

Matthieu Schaller's avatar
Matthieu Schaller committed
1194
  } else if (ti_current > ti_old) {
1195

1196
1197
1198
    /* Loop over all the g-particles in the cell */
    const size_t nr_gparts = c->gcount;
    for (size_t k = 0; k < nr_gparts; k++) {
1199

1200
1201
      /* Get a handle on the gpart. */
      struct gpart *const gp = &gparts[k];
1202

1203
1204
      /* Drift... */
      drift_gpart(gp, dt, timeBase, ti_old, ti_current);
1205

1206
1207
      /* Compute (square of) motion since last cell construction */
      const float dx2 = gp->x_diff[0] * gp->x_diff[0] +