space.c 46.5 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 <string.h>
32
#include <stdlib.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 "engine.h"
45
#include "error.h"
46
#include "kernel_hydro.h"
47
#include "lock.h"
48
#include "minmax.h"
49
#include "runner.h"
50
#include "tools.h"
Pedro Gonnet's avatar
Pedro Gonnet committed
51

52
53
54
/* Shared sort structure. */
struct parallel_sort space_sort_struct;

Pedro Gonnet's avatar
Pedro Gonnet committed
55
56
/* Split size. */
int space_splitsize = space_splitsize_default;
57
int space_subsize = space_subsize_default;
58
int space_maxsize = space_maxsize_default;
Pedro Gonnet's avatar
Pedro Gonnet committed
59
60
61

/* Map shift vector to sortlist. */
const int sortlistID[27] = {
62
63
64
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
    /* ( -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};

90
91
92
93
94
95
96
97
98
99
100
101
/**
 * @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.
 */

102
103
104
105
int space_getsid(struct space *s, struct cell **ci, struct cell **cj,
                 double *shift) {

  /* Get the relative distance between the pairs, wrapping. */
106
107
108
  const int periodic = s->periodic;
  double dx[3];
  for (int k = 0; k < 3; k++) {
109
110
111
112
113
114
115
116
117
118
119
    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. */
120
  int sid = 0;
121
  for (int k = 0; k < 3; k++)
122
123
124
125
    sid = 3 * sid + ((dx[k] < 0.0) ? 0 : ((dx[k] > 0.0) ? 2 : 1));

  /* Switch the cells around? */
  if (runner_flip[sid]) {
126
    struct cell *temp = *ci;
127
128
    *ci = *cj;
    *cj = temp;
129
    for (int k = 0; k < 3; k++) shift[k] = -shift[k];
130
131
132
133
134
135
  }
  sid = sortlistID[sid];

  /* Return the sort ID. */
  return sid;
}
136

137
/**
138
 * @brief Recursively dismantle a cell tree.
139
140
 *
 */
141
142
143
144

void space_rebuild_recycle(struct space *s, struct cell *c) {

  if (c->split)
145
    for (int k = 0; k < 8; k++)
146
147
148
149
150
151
152
      if (c->progeny[k] != NULL) {
        space_rebuild_recycle(s, c->progeny[k]);
        space_recycle(s, c->progeny[k]);
        c->progeny[k] = NULL;
      }
}

153
/**
154
 * @brief Re-build the cell grid.
155
 *
156
157
 * @param s The #space.
 * @param cell_max Maximum cell edge length.
158
 * @param verbose Print messages to stdout or not.
159
 */
160

161
void space_regrid(struct space *s, double cell_max, int verbose) {
162

163
  const size_t nr_parts = s->nr_parts;
164
  struct cell *restrict c;
165
  ticks tic = getticks();
166
167
168

  /* Run through the parts and get the current h_max. */
  // tic = getticks();
169
  float h_max = s->cell_min / kernel_gamma / space_stretch;
170
  if (nr_parts > 0) {
171
    if (s->cells != NULL) {
Tom Theuns's avatar
Tom Theuns committed
172
      for (int k = 0; k < s->nr_cells; k++) {
Matthieu Schaller's avatar
Matthieu Schaller committed
173
        if (s->cells[k].h_max > h_max) h_max = s->cells[k].h_max;
174
175
      }
    } else {
176
      for (size_t k = 0; k < nr_parts; k++) {
Matthieu Schaller's avatar
Matthieu Schaller committed
177
        if (s->parts[k].h > h_max) h_max = s->parts[k].h;
178
179
      }
      s->h_max = h_max;
180
181
182
    }
  }

183
184
  message("h_max: %f", s->h_max);

185
186
187
188
189
190
191
/* 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)
192
      error("Failed to aggregate the rebuild flag across nodes.");
193
194
195
    h_max = buff;
  }
#endif
196
  if (verbose) message("h_max is %.3e (cell_max=%.3e).", h_max, cell_max);
197
198

  /* Get the new putative cell dimensions. */
199
  int cdim[3];
200
  for (int k = 0; k < 3; k++)
201
202
203
204
205
206
207
208
209
    cdim[k] =
        floor(s->dim[k] / fmax(h_max * kernel_gamma * space_stretch, cell_max));

  /* 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 "
        "is switched on.");

210
211
212
/* 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. */
213
#ifdef WITH_MPI
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
  double oldh[3];
  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];
    oldh[0] = s->h[0];
    oldh[1] = s->h[1];
    oldh[2] = s->h[2];

    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);
          oldnodeIDs[cid] = s->cells[cid].nodeID;
        }
      }
    }
  }

241
242
243
244
245
246
247
248
249
#endif

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

    /* Free the old cells, if they were allocated. */
    if (s->cells != NULL) {
250
      for (int k = 0; k < s->nr_cells; k++) {
251
252
253
254
255
256
257
258
        space_rebuild_recycle(s, &s->cells[k]);
        if (s->cells[k].sort != NULL) free(s->cells[k].sort);
      }
      free(s->cells);
      s->maxdepth = 0;
    }

    /* Set the new cell dimensions only if smaller. */
259
    for (int k = 0; k < 3; k++) {
260
261
262
263
      s->cdim[k] = cdim[k];
      s->h[k] = s->dim[k] / cdim[k];
      s->ih[k] = 1.0 / s->h[k];
    }
264
    const float dmin = fminf(s->h[0], fminf(s->h[1], s->h[2]));
265
266
267
268
269
270
271

    /* Allocate the highest level of cells. */
    s->tot_cells = s->nr_cells = cdim[0] * cdim[1] * cdim[2];
    if (posix_memalign((void *)&s->cells, 64,
                       s->nr_cells * sizeof(struct cell)) != 0)
      error("Failed to allocate cells.");
    bzero(s->cells, s->nr_cells * sizeof(struct cell));
272
    for (int k = 0; k < s->nr_cells; k++)
273
274
275
      if (lock_init(&s->cells[k].lock) != 0) error("Failed to init spinlock.");

    /* Set the cell location and sizes. */
276
277
278
    for (int i = 0; i < cdim[0]; i++)
      for (int j = 0; j < cdim[1]; j++)
        for (int k = 0; k < cdim[2]; k++) {
279
280
281
282
283
284
285
286
287
288
289
290
291
          c = &s->cells[cell_getid(cdim, i, j, k)];
          c->loc[0] = i * s->h[0];
          c->loc[1] = j * s->h[1];
          c->loc[2] = k * s->h[2];
          c->h[0] = s->h[0];
          c->h[1] = s->h[1];
          c->h[2] = s->h[2];
          c->dmin = dmin;
          c->depth = 0;
          c->count = 0;
          c->gcount = 0;
          c->super = c;
          lock_init(&c->lock);
Pedro Gonnet's avatar
Pedro Gonnet committed
292
        }
293
294

    /* Be verbose about the change. */
295
296
297
    if (verbose)
      message("set cell dimensions to [ %i %i %i ].", cdim[0], cdim[1],
              cdim[2]);
298
299
    fflush(stdout);

300
#ifdef WITH_MPI
301
302
303
304
305
    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)
306
307
308
        message(
            "basic cell dimensions have increased - recalculating the "
            "global partition.");
309

310
      if (!partition_space_to_space(oldh, oldcdim, oldnodeIDs, s)) {
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328

        /* 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);
329

330
331
      /* Finished with these. */
      free(oldnodeIDs);
332
333
    }
#endif
334
  } /* re-build upper-level cells? */
335
336
  // message( "rebuilding upper-level cells took %.3f %s." ,
  // clocks_from_ticks(double)(getticks() - tic), clocks_getunit());
337
338
339
340
341

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

    /* Free the old cells, if they were allocated. */
342
    for (int k = 0; k < s->nr_cells; k++) {
343
344
345
346
347
348
349
350
351
352
353
      space_rebuild_recycle(s, &s->cells[k]);
      s->cells[k].sorts = NULL;
      s->cells[k].nr_tasks = 0;
      s->cells[k].nr_density = 0;
      s->cells[k].nr_force = 0;
      s->cells[k].density = NULL;
      s->cells[k].force = NULL;
      s->cells[k].dx_max = 0.0f;
      s->cells[k].sorted = 0;
      s->cells[k].count = 0;
      s->cells[k].gcount = 0;
Matthieu Schaller's avatar
Matthieu Schaller committed
354
      s->cells[k].init = NULL;
Matthieu Schaller's avatar
Matthieu Schaller committed
355
      s->cells[k].ghost = NULL;
Matthieu Schaller's avatar
Matthieu Schaller committed
356
357
      s->cells[k].drift = NULL;
      s->cells[k].kick = NULL;
358
      s->cells[k].super = &s->cells[k];
359
    }
360
361
    s->maxdepth = 0;
  }
362
363
364
365

  if (verbose)
    message("took %.3f %s.", clocks_from_ticks(getticks() - tic),
            clocks_getunit());
366
}
367
368
369
370
371
372

/**
 * @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.
373
 * @param verbose Print messages to stdout or not
374
375
 *
 */
376

377
void space_rebuild(struct space *s, double cell_max, int verbose) {
378

Matthieu Schaller's avatar
Matthieu Schaller committed
379
  const ticks tic = getticks();
380
381

  /* Be verbose about this. */
382
  // message("re)building space..."); fflush(stdout);
383
384

  /* Re-grid if necessary, or just re-set the cell data. */
385
  space_regrid(s, cell_max, verbose);
386

Pedro Gonnet's avatar
Pedro Gonnet committed
387
388
  size_t nr_parts = s->nr_parts;
  size_t nr_gparts = s->nr_gparts;
389
390
  struct cell *restrict cells = s->cells;

Matthieu Schaller's avatar
Matthieu Schaller committed
391
392
393
  const double ih[3] = {s->ih[0], s->ih[1], s->ih[2]};
  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]};
394
395
396
397

  /* Run through the particles and get their cell index. */
  // tic = getticks();
  const size_t ind_size = s->size_parts;
398
399
  int *ind;
  if ((ind = (int *)malloc(sizeof(int) * ind_size)) == NULL)
400
    error("Failed to allocate temporary particle indices.");
Pedro Gonnet's avatar
Pedro Gonnet committed
401
  for (size_t k = 0; k < nr_parts; k++) {
402
403
    struct part *restrict p = &s->parts[k];
    for (int j = 0; j < 3; j++)
404
405
406
407
      if (p->x[j] < 0.0)
        p->x[j] += dim[j];
      else if (p->x[j] >= dim[j])
        p->x[j] -= dim[j];
408
    ind[k] =
409
        cell_getid(cdim, p->x[0] * ih[0], p->x[1] * ih[1], p->x[2] * ih[2]);
410
    cells[ind[k]].count++;
411
  }
Pedro Gonnet's avatar
Pedro Gonnet committed
412
413
  // message( "getting particle indices took %.3f %s." ,
  // clocks_from_ticks(getticks() - tic), clocks_getunit()):
414

415
416
417
418
419
420
421
  /* Run through the gravity particles and get their cell index. */
  // tic = getticks();
  const size_t gind_size = s->size_gparts;
  int *gind;
  if ((gind = (int *)malloc(sizeof(int) * gind_size)) == NULL)
    error("Failed to allocate temporary g-particle indices.");
  for (int k = 0; k < nr_gparts; k++) {
Matthieu Schaller's avatar
Matthieu Schaller committed
422
    struct gpart *restrict gp = &s->gparts[k];
423
424
425
426
427
428
429
430
431
432
433
434
    for (int j = 0; j < 3; j++)
      if (gp->x[j] < 0.0)
        gp->x[j] += dim[j];
      else if (gp->x[j] >= dim[j])
        gp->x[j] -= dim[j];
    gind[k] =
        cell_getid(cdim, gp->x[0] * ih[0], gp->x[1] * ih[1], gp->x[2] * ih[2]);
    cells[gind[k]].gcount++;
  }
// message( "getting particle indices took %.3f %s." ,
// clocks_from_ticks(getticks() - tic), clocks_getunit());

435
436
#ifdef WITH_MPI
  /* Move non-local parts to the end of the list. */
437
  const int local_nodeID = s->e->nodeID;
438
  for (size_t k = 0; k < nr_parts;) {
439
    if (cells[ind[k]].nodeID != local_nodeID) {
440
441
      cells[ind[k]].count -= 1;
      nr_parts -= 1;
Matthieu Schaller's avatar
Bug fix    
Matthieu Schaller committed
442
      const struct part tp = s->parts[k];
443
444
      s->parts[k] = s->parts[nr_parts];
      s->parts[nr_parts] = tp;
445
446
447
448
449
450
      if (s->parts[k].gpart != NULL) {
        s->parts[k].gpart->part = &s->parts[k];
      }
      if (s->parts[nr_parts].gpart != NULL) {
        s->parts[nr_parts].gpart->part = &s->parts[nr_parts];
      }
Matthieu Schaller's avatar
Bug fix    
Matthieu Schaller committed
451
      const struct xpart txp = s->xparts[k];
452
453
      s->xparts[k] = s->xparts[nr_parts];
      s->xparts[nr_parts] = txp;
Matthieu Schaller's avatar
Bug fix    
Matthieu Schaller committed
454
      const int t = ind[k];
455
456
      ind[k] = ind[nr_parts];
      ind[nr_parts] = t;
457
    } else {
458
459
460
461
      /* Increment when not exchanging otherwise we need to retest "k".*/
      k++;
    }
  }
462

Peter W. Draper's avatar
Peter W. Draper committed
463
  /* Check that all parts are in the correct places. */
464
465
466
467
468
469
470
471
472
473
474
  /*  for (size_t k = 0; k < nr_parts; k++) {
    if (cells[ind[k]].nodeID != local_nodeID) {
      error("Failed to move all non-local parts to send list");
    }
  }
  for (size_t k = nr_parts; k < s->nr_parts; k++) {
    if (cells[ind[k]].nodeID == local_nodeID) {
      error("Failed to remove local parts from send list");
    }
  }*/

475
  /* Move non-local gparts to the end of the list. */
476
  for (int k = 0; k < nr_gparts;) {
Matthieu Schaller's avatar
Bug fix    
Matthieu Schaller committed
477
478
    if (cells[gind[k]].nodeID != local_nodeID) {
      cells[gind[k]].gcount -= 1;
479
      nr_gparts -= 1;
Matthieu Schaller's avatar
Bug fix    
Matthieu Schaller committed
480
      const struct gpart tp = s->gparts[k];
481
482
      s->gparts[k] = s->gparts[nr_gparts];
      s->gparts[nr_gparts] = tp;
483
484
485
486
487
488
      if (s->gparts[k].id > 0) {
        s->gparts[k].part->gpart = &s->gparts[k];
      }
      if (s->gparts[nr_gparts].id > 0) {
        s->gparts[nr_gparts].part->gpart = &s->gparts[nr_gparts];
      }
Matthieu Schaller's avatar
Bug fix    
Matthieu Schaller committed
489
490
491
      const int t = gind[k];
      gind[k] = gind[nr_gparts];
      gind[nr_gparts] = t;
492
    } else {
493
494
495
496
      /* Increment when not exchanging otherwise we need to retest "k".*/
      k++;
    }
  }
497

498
499
500
501
502
503
504
505
506
507
508
509
510
  /* Check that all gparts are in the correct place (untested). */
  /*
  for (size_t k = 0; k < nr_gparts; k++) {
    if (cells[gind[k]].nodeID != local_nodeID) {
      error("Failed to move all non-local gparts to send list");
    }
  }
  for (size_t k = nr_gparts; k < s->nr_gparts; k++) {
    if (cells[gind[k]].nodeID == local_nodeID) {
      error("Failed to remove local gparts from send list");
    }
  }*/

511
512
  /* Exchange the strays, note that this potentially re-allocates
     the parts arrays. */
513
  size_t nr_parts_exchanged = s->nr_parts - nr_parts;
514
  size_t nr_gparts_exchanged = s->nr_gparts - nr_gparts;
Pedro Gonnet's avatar
Pedro Gonnet committed
515
516
517
518
  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. */
519
  s->nr_parts = nr_parts + nr_parts_exchanged;
520
  s->nr_gparts = nr_gparts + nr_gparts_exchanged;
521
522

  /* Re-allocate the index array if needed.. */
523
  if (s->nr_parts > ind_size) {
524
525
    int *ind_new;
    if ((ind_new = (int *)malloc(sizeof(int) * s->nr_parts)) == NULL)
526
      error("Failed to allocate temporary particle indices.");
527
    memcpy(ind_new, ind, sizeof(int) * nr_parts);
528
529
    free(ind);
    ind = ind_new;
530
531
532
  }

  /* Assign each particle to its cell. */
Pedro Gonnet's avatar
Pedro Gonnet committed
533
  for (size_t k = nr_parts; k < s->nr_parts; k++) {
Matthieu Schaller's avatar
Matthieu Schaller committed
534
    const struct part *const p = &s->parts[k];
535
    ind[k] =
536
        cell_getid(cdim, p->x[0] * ih[0], p->x[1] * ih[1], p->x[2] * ih[2]);
537
538
539
540
    cells[ind[k]].count += 1;
    /* if ( cells[ ind[k] ].nodeID != nodeID )
        error( "Received part that does not belong to me (nodeID=%i)." , cells[
       ind[k] ].nodeID ); */
541
  }
542
  nr_parts = s->nr_parts;
543
544
545
#endif

  /* Sort the parts according to their cells. */
546
  space_parts_sort(s, ind, nr_parts, 0, s->nr_cells - 1, verbose);
547
548

  /* Re-link the gparts. */
Pedro Gonnet's avatar
Pedro Gonnet committed
549
  for (size_t k = 0; k < nr_parts; k++)
550
    if (s->parts[k].gpart != NULL) s->parts[k].gpart->part = &s->parts[k];
551

552
  /* Verify space_sort_struct. */
553
  /* for ( k = 1 ; k < nr_parts ; k++ ) {
554
      if ( ind[k-1] > ind[k] ) {
555
556
          error( "Sort failed!" );
          }
557
      else if ( ind[k] != cell_getid( cdim , parts[k].x[0]*ih[0] ,
558
559
560
561
562
     parts[k].x[1]*ih[1] , parts[k].x[2]*ih[2] ) )
          error( "Incorrect indices!" );
      } */

  /* We no longer need the indices as of here. */
563
  free(ind);
564

565
566
567
568
#ifdef WITH_MPI

  /* Re-allocate the index array if needed.. */
  if (s->nr_gparts > gind_size) {
569
570
    int *gind_new;
    if ((gind_new = (int *)malloc(sizeof(int) * s->nr_gparts)) == NULL)
571
      error("Failed to allocate temporary g-particle indices.");
572
    memcpy(gind_new, gind, sizeof(int) * nr_gparts);
573
574
575
576
577
    free(gind);
    gind = gind_new;
  }

  /* Assign each particle to its cell. */
578
  for (int k = nr_gparts; k < s->nr_gparts; k++) {
Matthieu Schaller's avatar
Matthieu Schaller committed
579
    const struct gpart *const p = &s->gparts[k];
580
581
    gind[k] =
        cell_getid(cdim, p->x[0] * ih[0], p->x[1] * ih[1], p->x[2] * ih[2]);
Matthieu Schaller's avatar
Typo    
Matthieu Schaller committed
582
    cells[gind[k]].gcount += 1;
583
584
585
586
587
    /* if ( cells[ ind[k] ].nodeID != nodeID )
        error( "Received part that does not belong to me (nodeID=%i)." , cells[
       ind[k] ].nodeID ); */
  }
  nr_gparts = s->nr_gparts;
588

589
#endif
590
591

  /* Sort the parts according to their cells. */
Matthieu Schaller's avatar
Matthieu Schaller committed
592
  space_gparts_sort(s, gind, nr_gparts, 0, s->nr_cells - 1, verbose);
593
594

  /* Re-link the parts. */
595
  for (int k = 0; k < nr_gparts; k++)
596
    if (s->gparts[k].id > 0) s->gparts[k].part->gpart = &s->gparts[k];
597
598

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

601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
  /* Verify that the links are correct */
  /* MATTHIEU: To be commented out once we are happy */
  for (size_t k = 0; k < nr_gparts; ++k) {

    if (s->gparts[k].id > 0) {

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

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

    if (s->parts[k].gpart != NULL) {

      if (s->parts[k].gpart->part != &s->parts[k]) error("Linking problem !");
    }
  }

623
624
  /* Hook the cells up to the parts. */
  // tic = getticks();
625
626
627
  struct part *finger = s->parts;
  struct xpart *xfinger = s->xparts;
  struct gpart *gfinger = s->gparts;
628
629
  for (int k = 0; k < s->nr_cells; k++) {
    struct cell *restrict c = &cells[k];
630
631
632
633
634
635
636
    c->parts = finger;
    c->xparts = xfinger;
    c->gparts = gfinger;
    finger = &finger[c->count];
    xfinger = &xfinger[c->count];
    gfinger = &gfinger[c->gcount];
  }
637
  // message( "hooking up cells took %.3f %s." ,
Matthieu Schaller's avatar
Matthieu Schaller committed
638
  // clocks_from_ticks(getticks() - tic), clocks_getunit());
639
640
641

  /* At this point, we have the upper-level cells, old or new. Now make
     sure that the parts in each cell are ok. */
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
  space_split(s, cells, verbose);

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

/**
 * @brief Split particles between cells of a hierarchy
 *
 * @param s The #space.
 * @param cells The cell hierarchy
 * @param verbose Are we talkative ?
 */
void space_split(struct space *s, struct cell *cells, int verbose) {

Matthieu Schaller's avatar
Matthieu Schaller committed
658
  const ticks tic = getticks();
659
660

  for (int k = 0; k < s->nr_cells; k++)
661
662
    scheduler_addtask(&s->e->sched, task_type_split_cell, task_subtype_none, k,
                      0, &cells[k], NULL, 0);
663
  engine_launch(s->e, s->e->nr_threads, 1 << task_type_split_cell, 0);
664

665
666
667
  if (verbose)
    message("took %.3f %s.", clocks_from_ticks(getticks() - tic),
            clocks_getunit());
668
}
669

670
/**
671
672
 * @brief Sort the particles and condensed particles according to the given
 *indices.
673
 *
Matthieu Schaller's avatar
Matthieu Schaller committed
674
 * @param s The #space.
675
676
677
678
 * @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.
679
 * @param verbose Are we talkative ?
680
 */
681

682
void space_parts_sort(struct space *s, int *ind, size_t N, int min, int max,
683
684
                      int verbose) {

Matthieu Schaller's avatar
Matthieu Schaller committed
685
  const ticks tic = getticks();
686
687

  /*Populate the global parallel_sort structure with the input data */
688
689
690
  space_sort_struct.parts = s->parts;
  space_sort_struct.xparts = s->xparts;
  space_sort_struct.ind = ind;
691
  space_sort_struct.stack_size = 2 * (max - min + 1) + 10 + s->e->nr_threads;
692
693
694
695
696
697
  if ((space_sort_struct.stack = malloc(sizeof(struct qstack) *
                                        space_sort_struct.stack_size)) == NULL)
    error("Failed to allocate sorting stack.");
  for (int i = 0; i < space_sort_struct.stack_size; i++)
    space_sort_struct.stack[i].ready = 0;

698
  /* Add the first interval. */
699
700
701
702
703
704
705
706
707
  space_sort_struct.stack[0].i = 0;
  space_sort_struct.stack[0].j = N - 1;
  space_sort_struct.stack[0].min = min;
  space_sort_struct.stack[0].max = max;
  space_sort_struct.stack[0].ready = 1;
  space_sort_struct.first = 0;
  space_sort_struct.last = 1;
  space_sort_struct.waiting = 1;

708
  /* Launch the sorting tasks. */
709
  engine_launch(s->e, s->e->nr_threads, (1 << task_type_part_sort), 0);
710
711

  /* Verify space_sort_struct. */
712
  /* for (int i = 1; i < N; i++)
713
    if (ind[i - 1] > ind[i])
714
715
      error("Sorting failed (ind[%i]=%i,ind[%i]=%i), min=%i, max=%i.", i - 1,
  ind[i - 1], i,
716
717
            ind[i], min, max);
  message("Sorting succeeded."); */
718

719
  /* Clean up. */
720
  free(space_sort_struct.stack);
721
722
723
724

  if (verbose)
    message("took %.3f %s.", clocks_from_ticks(getticks() - tic),
            clocks_getunit());
725
}
726

727
void space_do_parts_sort() {
728

729
  /* Pointers to the sorting data. */
730
  int *ind = space_sort_struct.ind;
731
732
  struct part *parts = space_sort_struct.parts;
  struct xpart *xparts = space_sort_struct.xparts;
733

734
  /* Main loop. */
735
  while (space_sort_struct.waiting) {
736

737
    /* Grab an interval off the queue. */
738
739
    int qid =
        atomic_inc(&space_sort_struct.first) % space_sort_struct.stack_size;
740

741
    /* Wait for the entry to be ready, or for the sorting do be done. */
742
743
    while (!space_sort_struct.stack[qid].ready)
      if (!space_sort_struct.waiting) return;
744

745
    /* Get the stack entry. */
746
747
    ptrdiff_t i = space_sort_struct.stack[qid].i;
    ptrdiff_t j = space_sort_struct.stack[qid].j;
748
749
    int min = space_sort_struct.stack[qid].min;
    int max = space_sort_struct.stack[qid].max;
750
    space_sort_struct.stack[qid].ready = 0;
751

752
753
    /* Loop over sub-intervals. */
    while (1) {
754

755
      /* Bring beer. */
756
      const int pivot = (min + max) / 2;
757
758
      /* message("Working on interval [%i,%i] with min=%i, max=%i, pivot=%i.",
              i, j, min, max, pivot); */
759
760

      /* One pass of QuickSort's partitioning. */
761
762
      ptrdiff_t ii = i;
      ptrdiff_t jj = j;
763
764
765
766
      while (ii < jj) {
        while (ii <= j && ind[ii] <= pivot) ii++;
        while (jj >= i && ind[jj] > pivot) jj--;
        if (ii < jj) {
767
          size_t temp_i = ind[ii];
768
769
          ind[ii] = ind[jj];
          ind[jj] = temp_i;
770
          struct part temp_p = parts[ii];
771
772
          parts[ii] = parts[jj];
          parts[jj] = temp_p;
773
          struct xpart temp_xp = xparts[ii];
774
775
776
777
          xparts[ii] = xparts[jj];
          xparts[jj] = temp_xp;
        }
      }
778

779
      /* Verify space_sort_struct. */
780
781
782
783
784
785
786
787
788
789
790
791
      /* for (int k = i; k <= jj; k++)
        if (ind[k] > pivot) {
          message("sorting failed at k=%i, ind[k]=%i, pivot=%i, i=%i, j=%i.", k,
                  ind[k], pivot, i, j);
          error("Partition failed (<=pivot).");
        }
      for (int k = jj + 1; k <= j; k++)
        if (ind[k] <= pivot) {
          message("sorting failed at k=%i, ind[k]=%i, pivot=%i, i=%i, j=%i.", k,
                  ind[k], pivot, i, j);
          error("Partition failed (>pivot).");
        } */
792
793
794
795
796
797

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

        /* Recurse on the left? */
        if (jj > i && pivot > min) {
798
799
          qid = atomic_inc(&space_sort_struct.last) %
                space_sort_struct.stack_size;
800
801
          while (space_sort_struct.stack[qid].ready)
            ;
802
803
804
805
806
807
808
          space_sort_struct.stack[qid].i = i;
          space_sort_struct.stack[qid].j = jj;
          space_sort_struct.stack[qid].min = min;
          space_sort_struct.stack[qid].max = pivot;
          if (atomic_inc(&space_sort_struct.waiting) >=
              space_sort_struct.stack_size)
            error("Qstack overflow.");
809
          space_sort_struct.stack[qid].ready = 1;
810
        }
811

812
813
814
815
816
817
818
819
820
821
        /* Recurse on the right? */
        if (jj + 1 < j && pivot + 1 < max) {
          i = jj + 1;
          min = pivot + 1;
        } else
          break;

      } else {

        /* Recurse on the right? */
822
        if (pivot + 1 < max) {
823
824
          qid = atomic_inc(&space_sort_struct.last) %
                space_sort_struct.stack_size;
825
826
          while (space_sort_struct.stack[qid].ready)
            ;
827
828
829
830
831
832
833
          space_sort_struct.stack[qid].i = jj + 1;
          space_sort_struct.stack[qid].j = j;
          space_sort_struct.stack[qid].min = pivot + 1;
          space_sort_struct.stack[qid].max = max;
          if (atomic_inc(&space_sort_struct.waiting) >=
              space_sort_struct.stack_size)
            error("Qstack overflow.");
834
          space_sort_struct.stack[qid].ready = 1;
835
        }
836

837
838
839
840
841
842
843
        /* Recurse on the left? */
        if (jj > i && pivot > min) {
          j = jj;
          max = pivot;
        } else
          break;
      }
844

845
846
    } /* loop over sub-intervals. */

847
    atomic_dec(&space_sort_struct.waiting);
848
849

  } /* main loop. */
850
851
}

852
853
854
855
856
/**
 * @brief Sort the g-particles and condensed particles according to the given
 *indices.
 *
 * @param s The #space.
Matthieu Schaller's avatar
Matthieu Schaller committed
857
858
 * @param ind The indices with respect to which the gparts are sorted.
 * @param N The number of gparts
859
860
861
862
 * @param min Lowest index.
 * @param max highest index.
 * @param verbose Are we talkative ?
 */
Matthieu Schaller's avatar
Matthieu Schaller committed
863
void space_gparts_sort(struct space *s, int *ind, size_t N, int min, int max,
864
865
                       int verbose) {

Matthieu Schaller's avatar
Matthieu Schaller committed
866
  const ticks tic = getticks();
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909

  /*Populate the global parallel_sort structure with the input data */
  space_sort_struct.gparts = s->gparts;
  space_sort_struct.ind = ind;
  space_sort_struct.stack_size = 2 * (max - min + 1) + 10 + s->e->nr_threads;
  if ((space_sort_struct.stack = malloc(sizeof(struct qstack) *
                                        space_sort_struct.stack_size)) == NULL)
    error("Failed to allocate sorting stack.");
  for (int i = 0; i < space_sort_struct.stack_size; i++)
    space_sort_struct.stack[i].ready = 0;

  /* Add the first interval. */
  space_sort_struct.stack[0].i = 0;
  space_sort_struct.stack[0].j = N - 1;
  space_sort_struct.stack[0].min = min;
  space_sort_struct.stack[0].max = max;
  space_sort_struct.stack[0].ready = 1;
  space_sort_struct.first = 0;
  space_sort_struct.last = 1;
  space_sort_struct.waiting = 1;

  /* Launch the sorting tasks. */
  engine_launch(s->e, s->e->nr_threads, (1 << task_type_gpart_sort), 0);

  /* Verify space_sort_struct. */
  /* for (int i = 1; i < N; i++)
    if (ind[i - 1] > ind[i])
      error("Sorting failed (ind[%i]=%i,ind[%i]=%i), min=%i, max=%i.", i - 1,
  ind[i - 1], i,
            ind[i], min, max);
  message("Sorting succeeded."); */

  /* Clean up. */
  free(space_sort_struct.stack);

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

void space_do_gparts_sort() {

  /* Pointers to the sorting data. */
Matthieu Schaller's avatar
Matthieu Schaller committed
910
  int *ind = space_sort_struct.ind;
911
  struct gpart *gparts = space_sort_struct.gparts;
912

913
  /* Main loop. */
914
  while (space_sort_struct.waiting) {
915

916
    /* Grab an interval off the queue. */
917
918
919
920
921
922
    int qid =
        atomic_inc(&space_sort_struct.first) % space_sort_struct.stack_size;

    /* Wait for the entry to be ready, or for the sorting do be done. */
    while (!space_sort_struct.stack[qid].ready)
      if (!space_sort_struct.waiting) return;
923

924
    /* Get the stack entry. */
925
926
927
928
929
    ptrdiff_t i = space_sort_struct.stack[qid].i;
    ptrdiff_t j = space_sort_struct.stack[qid].j;
    int min = space_sort_struct.stack[qid].min;
    int max = space_sort_struct.stack[qid].max;
    space_sort_struct.stack[qid].ready = 0;
930
931
932

    /* Loop over sub-intervals. */
    while (1) {
933

934
      /* Bring beer. */
935
936
937
      const int pivot = (min + max) / 2;
      /* message("Working on interval [%i,%i] with min=%i, max=%i, pivot=%i.",
              i, j, min, max, pivot); */
938
939

      /* One pass of QuickSort's partitioning. */
940
941
      ptrdiff_t ii = i;
      ptrdiff_t jj = j;
942
943
944
945
      while (ii < jj) {
        while (ii <= j && ind[ii] <= pivot) ii++;
        while (jj >= i && ind[jj] > pivot) jj--;
        if (ii < jj) {
946
          size_t temp_i = ind[ii];
947
948
          ind[ii] = ind[jj];
          ind[jj] = temp_i;
949
          struct gpart temp_p = gparts[ii];
950
951
952
953
          gparts[ii] = gparts[jj];
          gparts[jj] = temp_p;
        }
      }
954

955
      /* Verify space_sort_struct. */
956
957
958
959
960
961
962
963
964
965
966
967
      /* for (int k = i; k <= jj; k++)
        if (ind[k] > pivot) {
          message("sorting failed at k=%i, ind[k]=%i, pivot=%i, i=%i, j=%i.", k,
                  ind[k], pivot, i, j);
          error("Partition failed (<=pivot).");
        }
      for (int k = jj + 1; k <= j; k++)
        if (ind[k] <= pivot) {
          message("sorting failed at k=%i, ind[k]=%i, pivot=%i, i=%i, j=%i.", k,
                  ind[k], pivot, i, j);
          error("Partition failed (>pivot).");
        } */
968
969
970
971
972
973

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

        /* Recurse on the left? */
        if (jj > i && pivot > min) {
974
975
976
977
978
979
980
981
982
983
984
985
          qid = atomic_inc(&space_sort_struct.last) %
                space_sort_struct.stack_size;
          while (space_sort_struct.stack[qid].ready)
            ;
          space_sort_struct.stack[qid].i = i;
          space_sort_struct.stack[qid].j = jj;
          space_sort_struct.stack[qid].min = min;
          space_sort_struct.stack[qid].max = pivot;
          if (atomic_inc(&space_sort_struct.waiting) >=
              space_sort_struct.stack_size)
            error("Qstack overflow.");
          space_sort_struct.stack[qid].ready = 1;
986
        }
987

988
989
990
991
992
993
        /* Recurse on the right? */
        if (jj + 1 < j && pivot + 1 < max) {
          i = jj + 1;
          min = pivot + 1;
        } else
          break;
994

995
996
997
      } else {

        /* Recurse on the right? */
998
        if (pivot + 1 < max) {
999
1000
          qid = atomic_inc(&space_sort_struct.last) %
                space_sort_struct.stack_size;