cell.c 36.7 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) {
114
115
      struct cell *temp;
      space_getcells(s, 1, &temp);
116
      temp->count = 0;
117
      temp->gcount = 0;
118
      temp->scount = 0;
119
120
121
      temp->loc[0] = c->loc[0];
      temp->loc[1] = c->loc[1];
      temp->loc[2] = c->loc[2];
122
123
124
      temp->width[0] = c->width[0] / 2;
      temp->width[1] = c->width[1] / 2;
      temp->width[2] = c->width[2] / 2;
125
      temp->dmin = c->dmin / 2;
126
127
128
      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];
129
130
      temp->depth = c->depth + 1;
      temp->split = 0;
131
      temp->dx_max = 0.f;
132
133
      temp->nodeID = c->nodeID;
      temp->parent = c;
134
      temp->ti_old = c->ti_old;
135
136
137
      c->progeny[k] = temp;
      c->split = 1;
      count += cell_unpack(&pc[pc->progeny[k]], temp, s);
138
139
    }

140
  /* Return the total number of unpacked cells. */
141
  c->pcell_size = count;
142
  return count;
143
144
145
146
147

#else
  error("SWIFT was not compiled with MPI support.");
  return 0;
#endif
148
}
149

150
/**
151
 * @brief Link the cells recursively to the given #part array.
152
153
154
155
156
157
 *
 * @param c The #cell.
 * @param parts The #part array.
 *
 * @return The number of particles linked.
 */
158
int cell_link_parts(struct cell *c, struct part *parts) {
159

160
161
162
  c->parts = parts;

  /* Fill the progeny recursively, depth-first. */
Pedro Gonnet's avatar
Pedro Gonnet committed
163
164
165
166
  if (c->split) {
    int offset = 0;
    for (int k = 0; k < 8; k++) {
      if (c->progeny[k] != NULL)
167
        offset += cell_link_parts(c->progeny[k], &parts[offset]);
Pedro Gonnet's avatar
Pedro Gonnet committed
168
169
    }
  }
170

171
  /* Return the total number of linked particles. */
172
173
  return c->count;
}
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
199
/**
 * @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;
}

200
201
202
203
204
205
206
207
208
/**
 * @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.
 */
209
210
int cell_pack(struct cell *c, struct pcell *pc) {

211
212
#ifdef WITH_MPI

213
214
  /* Start by packing the data of the current cell. */
  pc->h_max = c->h_max;
215
216
  pc->ti_end_min = c->ti_end_min;
  pc->ti_end_max = c->ti_end_max;
217
  pc->ti_old = c->ti_old;
218
  pc->count = c->count;
219
  pc->gcount = c->gcount;
220
221
222
  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
223
224
  int count = 1;
  for (int k = 0; k < 8; k++)
225
226
227
228
229
230
231
    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. */
232
233
  c->pcell_size = count;
  return count;
234
235
236
237
238

#else
  error("SWIFT was not compiled with MPI support.");
  return 0;
#endif
239
240
}

241
242
243
244
245
246
247
248
/**
 * @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.
 */
249
250
int cell_pack_ti_ends(struct cell *c, int *ti_ends) {

251
252
#ifdef WITH_MPI

253
254
  /* Pack this cell's data. */
  ti_ends[0] = c->ti_end_min;
255

256
257
258
259
260
261
262
263
264
  /* 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;
265
266
267
268
269

#else
  error("SWIFT was not compiled with MPI support.");
  return 0;
#endif
270
271
}

272
273
274
275
/**
 * @brief Unpack the time information of a given cell and its sub-cells.
 *
 * @param c The #cell
276
 * @param ti_ends The time information to unpack
277
278
279
 *
 * @return The number of cells created.
 */
280
281
int cell_unpack_ti_ends(struct cell *c, int *ti_ends) {

282
283
#ifdef WITH_MPI

284
285
  /* Unpack this cell's data. */
  c->ti_end_min = ti_ends[0];
286

287
288
289
290
291
292
293
294
  /* 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. */
295
  return count;
296
297
298
299
300

#else
  error("SWIFT was not compiled with MPI support.");
  return 0;
#endif
301
}
302

303
/**
304
 * @brief Lock a cell for access to its array of #part and hold its parents.
305
306
 *
 * @param c The #cell.
307
 * @return 0 on success, 1 on failure
308
 */
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
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
331
  struct cell *finger;
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
  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
354
355
    for (struct cell *finger2 = c->parent; finger2 != finger;
         finger2 = finger2->parent)
356
      atomic_dec(&finger2->hold);
357
358
359
360
361
362
363
364
365
366

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

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

367
368
369
370
371
372
/**
 * @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
 */
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
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
395
  struct cell *finger;
396
397
398
399
400
401
  for (finger = c->parent; finger != NULL; finger = finger->parent) {

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

    /* Increment the hold. */
402
    atomic_inc(&finger->ghold);
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417

    /* 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
418
419
    for (struct cell *finger2 = c->parent; finger2 != finger;
         finger2 = finger2->parent)
420
      atomic_dec(&finger2->ghold);
421
422
423
424
425
426
427
428
429

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

    /* Admit defeat. */
    TIMER_TOC(timer_locktree);
    return 1;
  }
}
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
493
494
/**
 * @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;
  }
}

495
/**
496
 * @brief Unlock a cell's parents for access to #part array.
497
498
499
 *
 * @param c The #cell.
 */
500
501
502
503
504
505
506
507
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
508
  for (struct cell *finger = c->parent; finger != NULL; finger = finger->parent)
509
    atomic_dec(&finger->hold);
510
511
512
513

  TIMER_TOC(timer_locktree);
}

514
515
516
517
518
/**
 * @brief Unlock a cell's parents for access to #gpart array.
 *
 * @param c The #cell.
 */
519
520
521
522
523
524
525
526
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
527
  for (struct cell *finger = c->parent; finger != NULL; finger = finger->parent)
528
    atomic_dec(&finger->ghold);
529
530
531
532

  TIMER_TOC(timer_locktree);
}

533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
/**
 * @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);
}

552
553
554
555
/**
 * @brief Sort the parts into eight bins along the given pivots.
 *
 * @param c The #cell array to be sorted.
556
557
 * @param parts_offset Offset of the cell parts array relative to the
 *        space's parts array, i.e. c->parts - s->parts.
558
559
 * @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
560
561
 * @param gbuff A buffer with at least max(c->count, c->gcount) entries,
 *        used for sorting indices for the gparts.
562
 */
563
564
void cell_split(struct cell *c, ptrdiff_t parts_offset, ptrdiff_t sparts_offset,
                struct cell_buff *buff, struct cell_buff *sbuff,
565
                struct cell_buff *gbuff) {
566

567
  const int count = c->count, gcount = c->gcount, scount = c->scount;
568
569
570
  struct part *parts = c->parts;
  struct xpart *xparts = c->xparts;
  struct gpart *gparts = c->gparts;
Matthieu Schaller's avatar
Matthieu Schaller committed
571
  struct spart *sparts = c->sparts;
572
573
574
575
576
577
  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];

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

  /* Fill the buffer with the indices. */
  for (int k = 0; k < count; k++) {
599
600
    const int bid = (buff[k].x[0] > pivot[0]) * 4 +
                    (buff[k].x[1] > pivot[1]) * 2 + (buff[k].x[2] > pivot[2]);
601
    bucket_count[bid]++;
Matthieu Schaller's avatar
Matthieu Schaller committed
602
    buff[k].ind = bid;
603
  }
604

605
606
607
608
609
  /* 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;
610
611
  }

612
613
614
615
  /* 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++) {
Matthieu Schaller's avatar
Matthieu Schaller committed
616
      int bid = buff[k].ind;
617
618
619
      if (bid != bucket) {
        struct part part = parts[k];
        struct xpart xpart = xparts[k];
620
        struct cell_buff temp_buff = buff[k];
621
622
        while (bid != bucket) {
          int j = bucket_offset[bid] + bucket_count[bid]++;
Matthieu Schaller's avatar
Matthieu Schaller committed
623
          while (buff[j].ind == bid) {
624
625
626
            j++;
            bucket_count[bid]++;
          }
Pedro Gonnet's avatar
Pedro Gonnet committed
627
628
          memswap(&parts[j], &part, sizeof(struct part));
          memswap(&xparts[j], &xpart, sizeof(struct xpart));
629
630
          memswap(&buff[j], &temp_buff, sizeof(struct cell_buff));
          bid = temp_buff.ind;
631
632
633
        }
        parts[k] = part;
        xparts[k] = xpart;
634
        buff[k] = temp_buff;
635
      }
636
      bucket_count[bid]++;
637
638
639
640
    }
  }

  /* Store the counts and offsets. */
Pedro Gonnet's avatar
Pedro Gonnet committed
641
  for (int k = 0; k < 8; k++) {
642
643
644
    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]];
645
646
647
  }

  /* Re-link the gparts. */
648
649
  if (count > 0 && gcount > 0)
    part_relink_gparts_to_parts(parts, count, parts_offset);
650

651
#ifdef SWIFT_DEBUG_CHECKS
652
653
654
655
656
657
658
659
  /* Check that the buffs are OK. */
  for (int k = 1; k < count; k++) {
    if (buff[k].ind < buff[k - 1].ind) error("Buff not sorted.");
    if (buff[k].x[0] != parts[k].x[0] || buff[k].x[1] != parts[k].x[1] ||
        buff[k].x[2] != parts[k].x[2])
      error("Inconsistent buff contents (k=%i).", k);
  }

660
  /* Verify that _all_ the parts have been assigned to a cell. */
661
662
663
664
665
666
667
668
  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).");
669
670

  /* Verify a few sub-cells. */
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
  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).");
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
  for (int k = 0; k < c->progeny[3]->count; k++)
    if (c->progeny[3]->parts[k].x[0] > pivot[0] ||
        c->progeny[3]->parts[k].x[1] <= pivot[1] ||
        c->progeny[3]->parts[k].x[2] <= pivot[2])
      error("Sorting failed (progeny=3).");
  for (int k = 0; k < c->progeny[4]->count; k++)
    if (c->progeny[4]->parts[k].x[0] <= pivot[0] ||
        c->progeny[4]->parts[k].x[1] > pivot[1] ||
        c->progeny[4]->parts[k].x[2] > pivot[2])
      error("Sorting failed (progeny=4).");
  for (int k = 0; k < c->progeny[5]->count; k++)
    if (c->progeny[5]->parts[k].x[0] <= pivot[0] ||
        c->progeny[5]->parts[k].x[1] > pivot[1] ||
        c->progeny[5]->parts[k].x[2] <= pivot[2])
      error("Sorting failed (progeny=5).");
  for (int k = 0; k < c->progeny[6]->count; k++)
    if (c->progeny[6]->parts[k].x[0] <= pivot[0] ||
        c->progeny[6]->parts[k].x[1] <= pivot[1] ||
        c->progeny[6]->parts[k].x[2] > pivot[2])
      error("Sorting failed (progeny=6).");
  for (int k = 0; k < c->progeny[7]->count; k++)
    if (c->progeny[7]->parts[k].x[0] <= pivot[0] ||
        c->progeny[7]->parts[k].x[1] <= pivot[1] ||
        c->progeny[7]->parts[k].x[2] <= pivot[2])
      error("Sorting failed (progeny=7).");
711
#endif
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
  /* 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)
764
    part_relink_gparts_to_sparts(sparts, scount, sparts_offset);
765
766

  /* Finally, do the same song and dance for the gparts. */
767
768
769
770
  for (int k = 0; k < 8; k++) bucket_count[k] = 0;

  /* Fill the buffer with the indices. */
  for (int k = 0; k < gcount; k++) {
771
772
    const int bid = (gbuff[k].x[0] > pivot[0]) * 4 +
                    (gbuff[k].x[1] > pivot[1]) * 2 + (gbuff[k].x[2] > pivot[2]);
773
    bucket_count[bid]++;
774
    gbuff[k].ind = bid;
775
  }
776
777
778
779
780
781

  /* 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;
782
783
  }

784
785
786
787
  /* 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++) {
788
      int bid = gbuff[k].ind;
789
790
      if (bid != bucket) {
        struct gpart gpart = gparts[k];
791
        struct cell_buff temp_buff = gbuff[k];
792
793
        while (bid != bucket) {
          int j = bucket_offset[bid] + bucket_count[bid]++;
794
          while (gbuff[j].ind == bid) {
795
796
797
            j++;
            bucket_count[bid]++;
          }
Pedro Gonnet's avatar
Pedro Gonnet committed
798
          memswap(&gparts[j], &gpart, sizeof(struct gpart));
799
800
          memswap(&gbuff[j], &temp_buff, sizeof(struct cell_buff));
          bid = temp_buff.ind;
801
802
        }
        gparts[k] = gpart;
803
        gbuff[k] = temp_buff;
804
      }
805
      bucket_count[bid]++;
806
807
808
809
    }
  }

  /* Store the counts and offsets. */
Pedro Gonnet's avatar
Pedro Gonnet committed
810
  for (int k = 0; k < 8; k++) {
811
812
    c->progeny[k]->gcount = bucket_count[k];
    c->progeny[k]->gparts = &c->gparts[bucket_offset[k]];
813
814
815
  }

  /* Re-link the parts. */
816
  if (count > 0 && gcount > 0)
817
    part_relink_parts_to_gparts(gparts, gcount, parts - parts_offset);
818
819
820
821

  /* Re-link the sparts. */
  if (scount > 0 && gcount > 0)
    part_relink_sparts_to_gparts(gparts, gcount, sparts - sparts_offset);
822
}
823

824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
/**
 * @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) {

843
    const float h = logf(parts[i].h);
844
845
846
847
848
849
850
851
    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;
852
  const float h_std = (h_var > 0.f) ? sqrtf(h_var) : 0.1f * h_mean;
853
854

  /* Choose a cut */
855
  const float h_limit = expf(h_mean + 4.f * h_std);
856
857

  /* Be verbose this is not innocuous */
858
859
  message("Cell properties: h_min= %f h_max= %f geometric mean= %f.",
          expf(h_min), expf(h_max), expf(h_mean));
860
861
862

  if (c->h_max > h_limit) {

863
    message("Smoothing lengths will be limited to (mean + 4sigma)= %f.",
864
865
866
867
868
869
            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;
870
871
872
873

  } else {

    message("Smoothing lengths will not be limited.");
874
875
876
  }
}

877
/**
878
 * @brief Converts hydro quantities to a valid state after the initial density
879
 * calculation
880
881
882
883
884
885
886
 *
 * @param c Cell to act upon
 * @param data Unused parameter
 */
void cell_convert_hydro(struct cell *c, void *data) {

  struct part *p = c->parts;
887
  struct xpart *xp = c->xparts;
888
889

  for (int i = 0; i < c->count; ++i) {
890
    hydro_convert_quantities(&p[i], &xp[i]);
891
892
893
  }
}

Matthieu Schaller's avatar
Matthieu Schaller committed
894
895
896
897
898
899
/**
 * @brief Cleans the links in a given cell.
 *
 * @param c Cell to act upon
 * @param data Unused parameter
 */
900
void cell_clean_links(struct cell *c, void *data) {
Matthieu Schaller's avatar
Matthieu Schaller committed
901
  c->density = NULL;
902
  c->gradient = NULL;
Matthieu Schaller's avatar
Matthieu Schaller committed
903
  c->force = NULL;
904
  c->grav = NULL;
Matthieu Schaller's avatar
Matthieu Schaller committed
905
}
906

907
908
909
910
911
912
913
914
915
916
/**
 * @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) {

917
  integertime_t ti_current = *(integertime_t *)data;
918

919
  if (c->ti_old != ti_current && c->nodeID == engine_rank)
920
    error("Cell in an incorrect time-zone! c->ti_old=%lld ti_current=%lld",
921
922
923
          c->ti_old, ti_current);
}

924
925
926
927
928
929
930
931
932
933
934
935
/**
 * @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
936
#ifdef SWIFT_DEBUG_CHECKS
Matthieu Schaller's avatar
Matthieu Schaller committed
937
  if (ci->width[0] != cj->width[0]) error("Cells of different size !");
938
939
940
#endif

  /* Maximum allowed distance */
941
942
  const double min_dist =
      1.2 * ci->width[0]; /* 1.2 accounts for rounding errors */
943
944
945
946
947

  /* (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];
948
    if (fabs(center_i - center_j) > min_dist) return 0;
949
950
951
952
953
  }

  return 1;
}

954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
/**
 * @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)
978
      if (fabs(ma.CoM[k] - mb.CoM[k]) / fabs(ma.CoM[k] + mb.CoM[k]) > 1e-5)
979
980
981
        error("Multipole CoM are different (%12.15e vs. %12.15e", ma.CoM[k],
              mb.CoM[k]);

982
#if const_gravity_multipole_order >= 2
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
    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);
1007
#endif
1008
  }
1009
1010
}

1011
/**
1012
 * @brief Frees up the memory allocated for this #cell.
1013
 *
1014
 * @param c The #cell.
1015
 */
1016
1017
1018
1019
1020
1021
1022
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]);
1023
}
1024
1025
1026
1027
1028

/**
 * @brief Checks whether a given cell needs drifting or not.
 *
 * @param c the #cell.
1029
 * @param e The #engine (holding current time information).
1030
1031
1032
 *
 * @return 1 If the cell needs drifting, 0 otherwise.
 */
1033
int cell_is_drift_needed(struct cell *c, const struct engine *e) {
1034
1035

  /* Do we have at least one active particle in the cell ?*/
1036
  if (cell_is_active(c, e)) return 1;
1037
1038
1039
1040
1041
1042
1043

  /* 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;

1044
1045
1046
    /* 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)))
1047
      return 1;
1048
1049
1050
1051
1052
  }

  /* No neighbouring cell has active particles. Drift not necessary */
  return 0;
}
1053
1054
1055
1056
1057
1058

/**
 * @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
1059
 * @param s the #scheduler.
1060
1061
1062
 *
 * @return 1 If the space needs rebuilding. 0 otherwise.
 */
1063
int cell_unskip_tasks(struct cell *c, struct scheduler *s) {
1064
1065
1066
1067
1068
1069

  /* 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;
1070
    scheduler_activate(s, t);
1071
1072
1073
1074
1075

    /* 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));
1076
        scheduler_activate(s, ci->sorts);
1077
1078
1079
      }
      if (!(cj->sorted & (1 << t->flags))) {
        atomic_or(&cj->sorts->flags, (1 << t->flags));
1080
        scheduler_activate(s, cj->sorts);
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
      }
    }

    /* 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
1093
      /* Activate the send/recv flags. */
1094
      if (ci->nodeID != engine_rank) {
1095
1096

        /* Activate the tasks to recv foreign cell ci's data. */
1097
1098
1099
        scheduler_activate(s, ci->recv_xv);
        scheduler_activate(s, ci->recv_rho);
        scheduler_activate(s, ci->recv_ti);
1100
1101
1102
1103

        /* 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;
1104
1105
             l = l->next)
          ;
1106
        if (l == NULL) error("Missing link to send_xv task.");
1107
        scheduler_activate(s, l->t);
1108

Matthieu Schaller's avatar
Matthieu Schaller committed
1109
1110
1111
1112
        if (cj->super->drift)
          scheduler_activate(s, cj->super->drift);
        else
          error("Drift task missing !");
1113
1114

        for (l = cj->send_rho; l != NULL && l->t->cj->nodeID != ci->nodeID;
1115
1116
             l = l->next)
          ;
1117
        if (l == NULL) error("Missing link to send_rho task.");
1118
        scheduler_activate(s, l->t);
1119
1120

        for (l = cj->send_ti; l != NULL && l->t->cj->nodeID != ci->nodeID;
1121
1122
             l = l->next)
          ;
1123
        if (l == NULL) error("Missing link to send_ti task.");
1124
        scheduler_activate(s, l->t);
1125

1126
      } else if (cj->nodeID != engine_rank) {
1127
1128

        /* Activate the tasks to recv foreign cell cj's data. */
1129
1130
1131
        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
1132

1133
1134
1135
        /* 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;
1136
1137
             l = l->next)
          ;
1138
        if (l == NULL) error("Missing link to send_xv task.");
1139
        scheduler_activate(s, l->t);
1140

Matthieu Schaller's avatar
Matthieu Schaller committed
1141
1142
1143
1144
        if (ci->super->drift)
          scheduler_activate(s, ci->super->drift);
        else
          error("Drift task missing !");
1145
1146

        for (l = ci->send_rho; l != NULL && l->t->cj->nodeID != cj->nodeID;
1147
1148
             l = l->next)
          ;
1149
        if (l == NULL) error("Missing link to send_rho task.");
1150
        scheduler_activate(s, l->t);
1151
1152

        for (l = ci->send_ti; l != NULL && l->t->cj->nodeID != cj->nodeID;
1153
1154
             l = l->next)
          ;
1155
        if (l == NULL) error("Missing link to send_ti task.");
1156
        scheduler_activate(s, l->t);
1157
1158
1159
1160
1161
1162
      }
#endif
    }
  }

  /* Unskip all the other task types. */
1163
  for (struct link *l = c->gradient; l != NULL; l = l->next)
1164
    scheduler_activate(s, l->t);
1165
  for (struct link *l = c->force; l != NULL; l = l->next)
1166
    scheduler_activate(s, l->t);
1167
  for (struct link *l = c->grav; l != NULL; l = l->next)
1168
1169
1170
1171
    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);