space.c 56.4 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
/**
 * @brief Information required to compute the particle cell indices.
 */
struct index_data {
  struct space *s;
  struct cell *cells;
  int *ind;
};

124
125
126
127
128
129
130
131
132
133
134
/**
 * @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.
 */
135
136
137
138
int space_getsid(struct space *s, struct cell **ci, struct cell **cj,
                 double *shift) {

  /* Get the relative distance between the pairs, wrapping. */
139
140
141
  const int periodic = s->periodic;
  double dx[3];
  for (int k = 0; k < 3; k++) {
142
143
144
145
146
147
148
149
150
151
152
    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. */
153
  int sid = 0;
154
  for (int k = 0; k < 3; k++)
155
156
157
158
    sid = 3 * sid + ((dx[k] < 0.0) ? 0 : ((dx[k] > 0.0) ? 2 : 1));

  /* Switch the cells around? */
  if (runner_flip[sid]) {
159
    struct cell *temp = *ci;
160
161
    *ci = *cj;
    *cj = temp;
162
    for (int k = 0; k < 3; k++) shift[k] = -shift[k];
163
164
165
166
167
168
  }
  sid = sortlistID[sid];

  /* Return the sort ID. */
  return sid;
}
169

170
/**
171
 * @brief Recursively dismantle a cell tree.
172
 *
173
174
 * @param s The #space.
 * @param c The #cell to recycle.
175
 */
176
177
178
void space_rebuild_recycle(struct space *s, struct cell *c) {

  if (c->split)
179
    for (int k = 0; k < 8; k++)
180
181
182
183
184
185
186
      if (c->progeny[k] != NULL) {
        space_rebuild_recycle(s, c->progeny[k]);
        space_recycle(s, c->progeny[k]);
        c->progeny[k] = NULL;
      }
}

187
/**
188
 * @brief Re-build the top-level cell grid.
189
 *
190
191
 * @param s The #space.
 * @param cell_max Maximum cell edge length.
192
 * @param verbose Print messages to stdout or not.
193
 */
194
void space_regrid(struct space *s, double cell_max, int verbose) {
195

196
  const size_t nr_parts = s->nr_parts;
197
  const ticks tic = getticks();
198
  const int ti_current = (s->e != NULL) ? s->e->ti_current : 0;
199

200
  /* Run through the cells and get the current h_max. */
201
  // tic = getticks();
202
  float h_max = s->cell_min / kernel_gamma / space_stretch;
203
  if (nr_parts > 0) {
204
    if (s->cells_top != NULL) {
Tom Theuns's avatar
Tom Theuns committed
205
      for (int k = 0; k < s->nr_cells; k++) {
206
207
208
        if (s->cells_top[k].nodeID == engine_rank &&
            s->cells_top[k].h_max > h_max) {
          h_max = s->cells_top[k].h_max;
209
        }
210
211
      }
    } else {
212
      for (size_t k = 0; k < nr_parts; k++) {
Matthieu Schaller's avatar
Matthieu Schaller committed
213
        if (s->parts[k].h > h_max) h_max = s->parts[k].h;
214
      }
215
216
217
218
219
220
221
222
223
224
    }
  }

/* 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)
225
      error("Failed to aggregate the rebuild flag across nodes.");
226
227
228
    h_max = buff;
  }
#endif
229
  if (verbose) message("h_max is %.3e (cell_max=%.3e).", h_max, cell_max);
230
231

  /* Get the new putative cell dimensions. */
232
233
234
235
  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))};
236
237
238
239
240

  /* 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 "
241
242
243
244
        "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"
245
        " - the (minimal) time-step is too large leading to particles with "
246
247
248
        "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");
249

250
251
252
253
254
255
  /* 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.");

256
257
258
/* 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. */
259
#ifdef WITH_MPI
Matthieu Schaller's avatar
Matthieu Schaller committed
260
  double oldwidth[3];
261
262
263
264
265
266
267
268
  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];
269
270
271
    oldwidth[0] = s->width[0];
    oldwidth[1] = s->width[1];
    oldwidth[2] = s->width[2];
272
273
274
275
276
277
278
279
280

    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);
281
          oldnodeIDs[cid] = s->cells_top[cid].nodeID;
282
283
284
285
286
        }
      }
    }
  }

287
288
289
290
#endif

  /* Do we need to re-build the upper-level cells? */
  // tic = getticks();
291
  if (s->cells_top == NULL || cdim[0] < s->cdim[0] || cdim[1] < s->cdim[1] ||
292
293
294
      cdim[2] < s->cdim[2]) {

    /* Free the old cells, if they were allocated. */
295
    if (s->cells_top != NULL) {
296
      for (int k = 0; k < s->nr_cells; k++) {
297
298
        space_rebuild_recycle(s, &s->cells_top[k]);
        if (s->cells_top[k].sort != NULL) free(s->cells_top[k].sort);
299
      }
300
      free(s->cells_top);
301
302
303
304
      s->maxdepth = 0;
    }

    /* Set the new cell dimensions only if smaller. */
305
    for (int k = 0; k < 3; k++) {
306
      s->cdim[k] = cdim[k];
307
308
      s->width[k] = s->dim[k] / cdim[k];
      s->iwidth[k] = 1.0 / s->width[k];
309
    }
310
    const float dmin = min(s->width[0], min(s->width[1], s->width[2]));
311
312
313

    /* Allocate the highest level of cells. */
    s->tot_cells = s->nr_cells = cdim[0] * cdim[1] * cdim[2];
314
    if (posix_memalign((void *)&s->cells_top, cell_align,
315
316
                       s->nr_cells * sizeof(struct cell)) != 0)
      error("Failed to allocate cells.");
317
    bzero(s->cells_top, s->nr_cells * sizeof(struct cell));
318
    for (int k = 0; k < s->nr_cells; k++)
319
320
      if (lock_init(&s->cells_top[k].lock) != 0)
        error("Failed to init spinlock.");
321
322

    /* Set the cell location and sizes. */
323
324
325
    for (int i = 0; i < cdim[0]; i++)
      for (int j = 0; j < cdim[1]; j++)
        for (int k = 0; k < cdim[2]; k++) {
326
          struct cell *restrict c = &s->cells_top[cell_getid(cdim, i, j, k)];
327
328
329
330
331
332
          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];
333
334
335
336
337
          c->dmin = dmin;
          c->depth = 0;
          c->count = 0;
          c->gcount = 0;
          c->super = c;
338
          c->gsuper = c;
339
          c->ti_old = ti_current;
340
          lock_init(&c->lock);
Pedro Gonnet's avatar
Pedro Gonnet committed
341
        }
342
343

    /* Be verbose about the change. */
344
345
346
    if (verbose)
      message("set cell dimensions to [ %i %i %i ].", cdim[0], cdim[1],
              cdim[2]);
347
348
    fflush(stdout);

349
#ifdef WITH_MPI
350
351
352
353
354
    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)
355
356
357
        message(
            "basic cell dimensions have increased - recalculating the "
            "global partition.");
358

Matthieu Schaller's avatar
Matthieu Schaller committed
359
      if (!partition_space_to_space(oldwidth, oldcdim, oldnodeIDs, s)) {
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377

        /* 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);
378

379
380
      /* Finished with these. */
      free(oldnodeIDs);
381
382
    }
#endif
383
384
385
386

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

387
388
  } /* re-build upper-level cells? */

389
  else { /* Otherwise, just clean up the cells. */
390
391

    /* Free the old cells, if they were allocated. */
392
    for (int k = 0; k < s->nr_cells; k++) {
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
      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];
411
      s->cells_top[k].gsuper = &s->cells_top[k];
412
    }
413
414
    s->maxdepth = 0;
  }
415
416
417
418

  if (verbose)
    message("took %.3f %s.", clocks_from_ticks(getticks() - tic),
            clocks_getunit());
419
}
420
421
422
423
424
425

/**
 * @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.
426
 * @param verbose Print messages to stdout or not
427
428
 *
 */
429
void space_rebuild(struct space *s, double cell_max, int verbose) {
430

Matthieu Schaller's avatar
Matthieu Schaller committed
431
  const ticks tic = getticks();
432
433

  /* Be verbose about this. */
434
  // message("re)building space..."); fflush(stdout);
435
436

  /* Re-grid if necessary, or just re-set the cell data. */
437
  space_regrid(s, cell_max, verbose);
438

Pedro Gonnet's avatar
Pedro Gonnet committed
439
440
  size_t nr_parts = s->nr_parts;
  size_t nr_gparts = s->nr_gparts;
441
  struct cell *restrict cells_top = s->cells_top;
442
  const int ti_current = (s->e != NULL) ? s->e->ti_current : 0;
443
444
445

  /* Run through the particles and get their cell index. */
  const size_t ind_size = s->size_parts;
446
447
  int *ind;
  if ((ind = (int *)malloc(sizeof(int) * ind_size)) == NULL)
448
    error("Failed to allocate temporary particle indices.");
449
  if (ind_size > 0) space_parts_get_cell_index(s, ind, cells_top, verbose);
450

451
452
453
454
455
  /* Run through the gravity particles and get their cell index. */
  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.");
456
  if (gind_size > 0) space_gparts_get_cell_index(s, gind, cells_top, verbose);
457
458

#ifdef WITH_MPI
459

460
  /* Move non-local parts to the end of the list. */
461
  const int local_nodeID = s->e->nodeID;
462
  for (size_t k = 0; k < nr_parts;) {
463
464
    if (cells_top[ind[k]].nodeID != local_nodeID) {
      cells_top[ind[k]].count -= 1;
465
      nr_parts -= 1;
Matthieu Schaller's avatar
Bug fix    
Matthieu Schaller committed
466
      const struct part tp = s->parts[k];
467
468
      s->parts[k] = s->parts[nr_parts];
      s->parts[nr_parts] = tp;
469
      if (s->parts[k].gpart != NULL) {
470
        s->parts[k].gpart->id_or_neg_offset = -k;
471
472
      }
      if (s->parts[nr_parts].gpart != NULL) {
473
        s->parts[nr_parts].gpart->id_or_neg_offset = -nr_parts;
474
      }
Matthieu Schaller's avatar
Bug fix    
Matthieu Schaller committed
475
      const struct xpart txp = s->xparts[k];
476
477
      s->xparts[k] = s->xparts[nr_parts];
      s->xparts[nr_parts] = txp;
Matthieu Schaller's avatar
Bug fix    
Matthieu Schaller committed
478
      const int t = ind[k];
479
480
      ind[k] = ind[nr_parts];
      ind[nr_parts] = t;
Matthieu Schaller's avatar
Matthieu Schaller committed
481
    } else {
482
483
484
485
      /* Increment when not exchanging otherwise we need to retest "k".*/
      k++;
    }
  }
486

487
#ifdef SWIFT_DEBUG_CHECKS
Peter W. Draper's avatar
Peter W. Draper committed
488
  /* Check that all parts are in the correct places. */
489
  for (size_t k = 0; k < nr_parts; k++) {
490
    if (cells_top[ind[k]].nodeID != local_nodeID) {
491
492
493
494
      error("Failed to move all non-local parts to send list");
    }
  }
  for (size_t k = nr_parts; k < s->nr_parts; k++) {
495
    if (cells_top[ind[k]].nodeID == local_nodeID) {
496
      error("Failed to remove local parts from send list");
497
    }
498
499
  }
#endif
500

501
  /* Move non-local gparts to the end of the list. */
502
  for (size_t k = 0; k < nr_gparts;) {
503
504
    if (cells_top[gind[k]].nodeID != local_nodeID) {
      cells_top[gind[k]].gcount -= 1;
505
      nr_gparts -= 1;
Matthieu Schaller's avatar
Bug fix    
Matthieu Schaller committed
506
      const struct gpart tp = s->gparts[k];
507
508
      s->gparts[k] = s->gparts[nr_gparts];
      s->gparts[nr_gparts] = tp;
509
510
      if (s->gparts[k].id_or_neg_offset <= 0) {
        s->parts[-s->gparts[k].id_or_neg_offset].gpart = &s->gparts[k];
511
      }
512
513
514
      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];
515
      }
Matthieu Schaller's avatar
Bug fix    
Matthieu Schaller committed
516
517
518
      const int t = gind[k];
      gind[k] = gind[nr_gparts];
      gind[nr_gparts] = t;
Matthieu Schaller's avatar
Matthieu Schaller committed
519
    } else {
520
521
522
523
      /* Increment when not exchanging otherwise we need to retest "k".*/
      k++;
    }
  }
524

525
#ifdef SWIFT_DEBUG_CHECKS
526
527
  /* Check that all gparts are in the correct place (untested). */
  for (size_t k = 0; k < nr_gparts; k++) {
528
    if (cells_top[gind[k]].nodeID != local_nodeID) {
529
530
531
532
      error("Failed to move all non-local gparts to send list");
    }
  }
  for (size_t k = nr_gparts; k < s->nr_gparts; k++) {
533
    if (cells_top[gind[k]].nodeID == local_nodeID) {
534
      error("Failed to remove local gparts from send list");
535
    }
536
537
  }
#endif
538

539
540
  /* Exchange the strays, note that this potentially re-allocates
     the parts arrays. */
541
  size_t nr_parts_exchanged = s->nr_parts - nr_parts;
542
  size_t nr_gparts_exchanged = s->nr_gparts - nr_gparts;
Pedro Gonnet's avatar
Pedro Gonnet committed
543
544
545
546
  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. */
547
  s->nr_parts = nr_parts + nr_parts_exchanged;
548
  s->nr_gparts = nr_gparts + nr_gparts_exchanged;
549
550

  /* Re-allocate the index array if needed.. */
551
  if (s->nr_parts > ind_size) {
552
553
    int *ind_new;
    if ((ind_new = (int *)malloc(sizeof(int) * s->nr_parts)) == NULL)
554
      error("Failed to allocate temporary particle indices.");
555
    memcpy(ind_new, ind, sizeof(int) * nr_parts);
556
557
    free(ind);
    ind = ind_new;
558
559
  }

560
561
562
  const int cdim[3] = {s->cdim[0], s->cdim[1], s->cdim[2]};
  const double ih[3] = {s->iwidth[0], s->iwidth[1], s->iwidth[2]};

563
  /* Assign each particle to its cell. */
Pedro Gonnet's avatar
Pedro Gonnet committed
564
  for (size_t k = nr_parts; k < s->nr_parts; k++) {
Matthieu Schaller's avatar
Matthieu Schaller committed
565
    const struct part *const p = &s->parts[k];
566
    ind[k] =
567
        cell_getid(cdim, p->x[0] * ih[0], p->x[1] * ih[1], p->x[2] * ih[2]);
568
    cells_top[ind[k]].count += 1;
569
#ifdef SWIFT_DEBUG_CHECKS
570
    if (cells_top[ind[k]].nodeID != local_nodeID)
571
      error("Received part that does not belong to me (nodeID=%i).",
572
            cells_top[ind[k]].nodeID);
573
#endif
574
  }
575
  nr_parts = s->nr_parts;
576
577

#endif /* WITH_MPI */
578
579

  /* Sort the parts according to their cells. */
580
  space_parts_sort(s, ind, nr_parts, 0, s->nr_cells - 1, verbose);
581
582

  /* Re-link the gparts. */
583
  if (nr_parts > 0 && nr_gparts > 0) part_relink_gparts(s->parts, nr_parts, 0);
584

585
#ifdef SWIFT_DEBUG_CHECKS
586
  /* Verify space_sort_struct. */
587
588
589
590
591
592
593
594
595
596
  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
597
598

  /* We no longer need the indices as of here. */
599
  free(ind);
600

601
602
603
604
#ifdef WITH_MPI

  /* Re-allocate the index array if needed.. */
  if (s->nr_gparts > gind_size) {
605
606
    int *gind_new;
    if ((gind_new = (int *)malloc(sizeof(int) * s->nr_gparts)) == NULL)
607
      error("Failed to allocate temporary g-particle indices.");
608
    memcpy(gind_new, gind, sizeof(int) * nr_gparts);
609
610
611
612
613
    free(gind);
    gind = gind_new;
  }

  /* Assign each particle to its cell. */
614
  for (size_t k = nr_gparts; k < s->nr_gparts; k++) {
Matthieu Schaller's avatar
Matthieu Schaller committed
615
    const struct gpart *const p = &s->gparts[k];
616
617
    gind[k] =
        cell_getid(cdim, p->x[0] * ih[0], p->x[1] * ih[1], p->x[2] * ih[2]);
618
619
620
621
622
623
624
    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
625
626
627
628
  }
  nr_gparts = s->nr_gparts;

#endif
629

630
  /* Sort the gparts according to their cells. */
Matthieu Schaller's avatar
Matthieu Schaller committed
631
  space_gparts_sort(s, gind, nr_gparts, 0, s->nr_cells - 1, verbose);
632
633

  /* Re-link the parts. */
634
635
  if (nr_parts > 0 && nr_gparts > 0)
    part_relink_parts(s->gparts, nr_gparts, s->parts);
636
637

  /* We no longer need the indices as of here. */
638
  free(gind);
639

640
#ifdef SWIFT_DEBUG_CHECKS
641
642
643
  /* Verify that the links are correct */
  for (size_t k = 0; k < nr_gparts; ++k) {

644
    if (s->gparts[k].id_or_neg_offset < 0) {
645

646
      const struct part *part = &s->parts[-s->gparts[k].id_or_neg_offset];
647

648
649
650
651
      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])
652
653
654
655
656
        error("Linked particles are not at the same position !");
    }
  }
  for (size_t k = 0; k < nr_parts; ++k) {

657
    if (s->parts[k].gpart != NULL &&
658
        s->parts[k].gpart->id_or_neg_offset != -(ptrdiff_t)k) {
659
      error("Linking problem !");
660
661
    }
  }
662
#endif
663

664
665
  /* Hook the cells up to the parts. */
  // tic = getticks();
666
667
668
  struct part *finger = s->parts;
  struct xpart *xfinger = s->xparts;
  struct gpart *gfinger = s->gparts;
669
  for (int k = 0; k < s->nr_cells; k++) {
670
    struct cell *restrict c = &cells_top[k];
671
    c->ti_old = ti_current;
672
673
674
675
676
677
678
    c->parts = finger;
    c->xparts = xfinger;
    c->gparts = gfinger;
    finger = &finger[c->count];
    xfinger = &xfinger[c->count];
    gfinger = &gfinger[c->gcount];
  }
679
  // message( "hooking up cells took %.3f %s." ,
Matthieu Schaller's avatar
Matthieu Schaller committed
680
  // clocks_from_ticks(getticks() - tic), clocks_getunit());
681
682
683

  /* At this point, we have the upper-level cells, old or new. Now make
     sure that the parts in each cell are ok. */
684
  space_split(s, cells_top, s->nr_cells, verbose);
685
686
687
688
689
690
691
692
693

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

/**
 * @brief Split particles between cells of a hierarchy
 *
694
695
 * This is done in parallel using threads in the #threadpool.
 *
696
 * @param s The #space.
697
698
 * @param cells The cell hierarchy.
 * @param nr_cells The number of cells.
699
700
 * @param verbose Are we talkative ?
 */
701
702
void space_split(struct space *s, struct cell *cells, int nr_cells,
                 int verbose) {
703

Matthieu Schaller's avatar
Matthieu Schaller committed
704
  const ticks tic = getticks();
705

706
  threadpool_map(&s->e->threadpool, space_split_mapper, cells, nr_cells,
707
                 sizeof(struct cell), 1, s);
708

709
710
711
  if (verbose)
    message("took %.3f %s.", clocks_from_ticks(getticks() - tic),
            clocks_getunit());
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
/**
 * @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);
    }
  }
}

737
/**
738
 * @brief #threadpool mapper function to compute the particle cell indices.
739
 *
740
 * @param map_data Pointer towards the particles.
741
 * @param nr_parts The number of particles to treat.
742
 * @param extra_data Pointers to the space and index list
743
 */
744
745
void space_parts_get_cell_index_mapper(void *map_data, int nr_parts,
                                       void *extra_data) {
746

747
748
749
750
  /* Unpack the data */
  struct part *restrict parts = (struct part *)map_data;
  struct index_data *data = (struct index_data *)extra_data;
  struct space *s = data->s;
751
  int *ind = data->ind + (ptrdiff_t)(parts - s->parts);
752
753
754
  struct cell *cells = data->cells;

  /* Get some constants */
755
756
757
758
  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]};
  const double ih[3] = {s->iwidth[0], s->iwidth[1], s->iwidth[2]};

759
  for (int k = 0; k < nr_parts; k++) {
760
761
762
763
764

    /* Get the particle */
    struct part *restrict p = &parts[k];

    /* Put it back into the simulation volume */
765
766
767
    p->x[0] = box_wrap(p->x[0], 0.0, dim[0]);
    p->x[1] = box_wrap(p->x[1], 0.0, dim[1]);
    p->x[2] = box_wrap(p->x[2], 0.0, dim[2]);
768
769
770
771

    /* Get its cell index */
    const int index =
        cell_getid(cdim, p->x[0] * ih[0], p->x[1] * ih[1], p->x[2] * ih[2]);
772
    ind[k] = index;
773
774

    /* Tell the cell it has a new member */
775
    atomic_inc(&(cells[index].count));
776
777
778
779
  }
}

/**
780
 * @brief #threadpool mapper function to compute the g-particle cell indices.
781
 *
782
 * @param map_data Pointer towards the g-particles.
783
 * @param nr_gparts The number of g-particles to treat.
784
 * @param extra_data Pointers to the space and index list
785
 */
786
787
void space_gparts_get_cell_index_mapper(void *map_data, int nr_gparts,
                                        void *extra_data) {
788

789
790
791
792
  /* Unpack the data */
  struct gpart *restrict gparts = (struct gpart *)map_data;
  struct index_data *data = (struct index_data *)extra_data;
  struct space *s = data->s;
793
  int *ind = data->ind + (ptrdiff_t)(gparts - s->gparts);
794
795
796
  struct cell *cells = data->cells;

  /* Get some constants */
797
798
799
800
  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]};
  const double ih[3] = {s->iwidth[0], s->iwidth[1], s->iwidth[2]};

801
  for (int k = 0; k < nr_gparts; k++) {
802
803
804
805
806

    /* Get the particle */
    struct gpart *restrict gp = &gparts[k];

    /* Put it back into the simulation volume */
807
808
809
    gp->x[0] = box_wrap(gp->x[0], 0.0, dim[0]);
    gp->x[1] = box_wrap(gp->x[1], 0.0, dim[1]);
    gp->x[2] = box_wrap(gp->x[2], 0.0, dim[2]);
810
811
812
813

    /* Get its cell index */
    const int index =
        cell_getid(cdim, gp->x[0] * ih[0], gp->x[1] * ih[1], gp->x[2] * ih[2]);
814
    ind[k] = index;
815
816

    /* Tell the cell it has a new member */
817
    atomic_inc(&(cells[index].gcount));
818
819
820
  }
}

821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
/**
 * @brief Computes the cell index of all the particles and update the cell
 * count.
 *
 * @param s The #space.
 * @param ind The array of indices to fill.
 * @param cells The array of #cell to update.
 * @param verbose Are we talkative ?
 */
void space_parts_get_cell_index(struct space *s, int *ind, struct cell *cells,
                                int verbose) {

  const ticks tic = getticks();

  /* Pack the extra information */
  struct index_data data;
  data.s = s;
  data.cells = cells;
  data.ind = ind;

  threadpool_map(&s->e->threadpool, space_parts_get_cell_index_mapper, s->parts,
                 s->nr_parts, sizeof(struct part), 1000, &data);

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

/**
 * @brief Computes the cell index of all the g-particles and update the cell
 * gcount.
 *
 * @param s The #space.
 * @param gind The array of indices to fill.
 * @param cells The array of #cell to update.
 * @param verbose Are we talkative ?
 */
void space_gparts_get_cell_index(struct space *s, int *gind, struct cell *cells,
                                 int verbose) {

  const ticks tic = getticks();

  /* Pack the extra information */
  struct index_data data;
  data.s = s;
  data.cells = cells;
  data.ind = gind;

  threadpool_map(&s->e->threadpool, space_gparts_get_cell_index_mapper,
                 s->gparts, s->nr_gparts, sizeof(struct gpart), 1000, &data);

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

877
/**
878
 * @brief Sort the particles and condensed particles according to the given
879
 * indices.
880
 *
Matthieu Schaller's avatar
Matthieu Schaller committed
881
 * @param s The #space.
882
883
884
885
 * @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.
886
 * @param verbose Are we talkative ?
887
 */
888
void space_parts_sort(struct space *s, int *ind, size_t N, int min, int max,
889
890
                      int verbose) {

Matthieu Schaller's avatar
Matthieu Schaller committed
891
  const ticks tic = getticks();
892

893
894
895
896
897
898
  /* 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;
899
900
  if ((sort_struct.stack =
           malloc(sizeof(struct qstack) * sort_struct.stack_size)) == NULL)
901
    error("Failed to allocate sorting stack.");
Matthieu Schaller's avatar
Matthieu Schaller committed
902
  for (unsigned int i = 0; i < sort_struct.stack_size; i++)
903
    sort_struct.stack[i].ready = 0;
904

905
  /* Add the first interval. */
906
907
908
909
910
911
912
913
914
915
916
  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. */
917
918
  threadpool_map(&s->e->threadpool, space_parts_sort_mapper, &sort_struct,
                 s->e->threadpool.num_threads, 0, 1, NULL);
919

920
#ifdef SWIFT_DEBUG_CHECKS
921
  /* Verify space_sort_struct. */
922
  for (size_t i = 1; i < N; i++)
923
    if (ind[i - 1] > ind[i])
924
      error("Sorting failed (ind[%zu]=%i,ind[%zu]=%i), min=%i, max=%i.", i - 1,
925
926
927
            ind[i - 1], i, ind[i], min, max);
  message("Sorting succeeded.");
#endif
928

929
  /* Clean up. */
930
  free(sort_struct.stack);
931
932
933
934

  if (verbose)
    message("took %.3f %s.", clocks_from_ticks(getticks() - tic),
            clocks_getunit());
935
}
936

937
938
void space_parts_sort_mapper(void *map_data, int num_elements,
                             void *extra_data) {
939
940
941

  /* Unpack the mapping data. */
  struct parallel_sort *sort_struct = (struct parallel_sort *)map_data;
942

943
  /* Pointers to the sorting data. */
944
945
946
  int *ind = sort_struct->ind;
  struct part *parts = sort_struct->parts;
  struct xpart *xparts = sort_struct->xparts;
947

948
  /* Main loop. */
949
  while (sort_struct->waiting) {
950

951
    /* Grab an interval off the queue. */
952
    int qid = atomic_inc(&sort_struct->first) % sort_struct->stack_size;
953

954
    /* Wait for the entry to be ready, or for the sorting do be done. */
955
956
    while (!sort_struct->stack[qid].ready)
      if (!sort_struct->waiting) return;
957

958
    /* Get the stack entry. */
959
960
961
962
963
    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;
964

965
966
    /* Loop over sub-intervals. */
    while (1) {
967

968
      /* Bring beer. */
969
      const int pivot = (min + max) / 2;
970
971
      /* message("Working on interval [%i,%i] with min=%i, max=%i, pivot=%i.",
              i, j, min, max, pivot); */