space.c 45.6 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
185
186
187
188
189
    }
  }

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

  /* Get the new putative cell dimensions. */
197
  int cdim[3];
198
  for (int k = 0; k < 3; k++)
199
200
201
202
203
204
205
206
207
    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.");

208
209
210
/* 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. */
211
#ifdef WITH_MPI
212
213
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
  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;
        }
      }
    }
  }

239
240
241
242
243
244
245
246
247
#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) {
248
      for (int k = 0; k < s->nr_cells; k++) {
249
250
251
252
253
254
255
256
        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. */
257
    for (int k = 0; k < 3; k++) {
258
259
260
261
      s->cdim[k] = cdim[k];
      s->h[k] = s->dim[k] / cdim[k];
      s->ih[k] = 1.0 / s->h[k];
    }
262
    const float dmin = fminf(s->h[0], fminf(s->h[1], s->h[2]));
263
264
265
266
267
268
269

    /* 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));
270
    for (int k = 0; k < s->nr_cells; k++)
271
272
273
      if (lock_init(&s->cells[k].lock) != 0) error("Failed to init spinlock.");

    /* Set the cell location and sizes. */
274
275
276
    for (int i = 0; i < cdim[0]; i++)
      for (int j = 0; j < cdim[1]; j++)
        for (int k = 0; k < cdim[2]; k++) {
277
278
279
280
281
282
283
284
285
286
287
288
289
          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
290
        }
291
292

    /* Be verbose about the change. */
293
294
295
    if (verbose)
      message("set cell dimensions to [ %i %i %i ].", cdim[0], cdim[1],
              cdim[2]);
296
297
    fflush(stdout);

298
#ifdef WITH_MPI
299
300
301
302
303
    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)
304
305
306
        message(
            "basic cell dimensions have increased - recalculating the "
            "global partition.");
307

308
      if (!partition_space_to_space(oldh, oldcdim, oldnodeIDs, s)) {
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326

        /* 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);
327

328
329
      /* Finished with these. */
      free(oldnodeIDs);
330
331
    }
#endif
332
  } /* re-build upper-level cells? */
333
334
  // message( "rebuilding upper-level cells took %.3f %s." ,
  // clocks_from_ticks(double)(getticks() - tic), clocks_getunit());
335
336
337
338
339

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

    /* Free the old cells, if they were allocated. */
340
    for (int k = 0; k < s->nr_cells; k++) {
341
342
343
344
345
346
347
348
349
350
351
      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
352
      s->cells[k].init = NULL;
Matthieu Schaller's avatar
Matthieu Schaller committed
353
      s->cells[k].ghost = NULL;
Matthieu Schaller's avatar
Matthieu Schaller committed
354
355
      s->cells[k].drift = NULL;
      s->cells[k].kick = NULL;
356
      s->cells[k].super = &s->cells[k];
357
    }
358
359
    s->maxdepth = 0;
  }
360
361
362
363

  if (verbose)
    message("took %.3f %s.", clocks_from_ticks(getticks() - tic),
            clocks_getunit());
364
}
365
366
367
368
369
370

/**
 * @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.
371
 * @param verbose Print messages to stdout or not
372
373
 *
 */
374

375
void space_rebuild(struct space *s, double cell_max, int verbose) {
376

Matthieu Schaller's avatar
Matthieu Schaller committed
377
  const ticks tic = getticks();
378
379

  /* Be verbose about this. */
380
  // message("re)building space..."); fflush(stdout);
381
382

  /* Re-grid if necessary, or just re-set the cell data. */
383
  space_regrid(s, cell_max, verbose);
384

Pedro Gonnet's avatar
Pedro Gonnet committed
385
386
  size_t nr_parts = s->nr_parts;
  size_t nr_gparts = s->nr_gparts;
387
388
  struct cell *restrict cells = s->cells;

Matthieu Schaller's avatar
Matthieu Schaller committed
389
390
391
  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]};
392
393
394
395

  /* Run through the particles and get their cell index. */
  // tic = getticks();
  const size_t ind_size = s->size_parts;
396
397
  int *ind;
  if ((ind = (int *)malloc(sizeof(int) * ind_size)) == NULL)
398
    error("Failed to allocate temporary particle indices.");
Pedro Gonnet's avatar
Pedro Gonnet committed
399
  for (size_t k = 0; k < nr_parts; k++) {
400
401
    struct part *restrict p = &s->parts[k];
    for (int j = 0; j < 3; j++)
402
403
404
405
      if (p->x[j] < 0.0)
        p->x[j] += dim[j];
      else if (p->x[j] >= dim[j])
        p->x[j] -= dim[j];
406
    ind[k] =
407
        cell_getid(cdim, p->x[0] * ih[0], p->x[1] * ih[1], p->x[2] * ih[2]);
408
    cells[ind[k]].count++;
409
  }
Pedro Gonnet's avatar
Pedro Gonnet committed
410
411
  // message( "getting particle indices took %.3f %s." ,
  // clocks_from_ticks(getticks() - tic), clocks_getunit()):
412

413
414
415
416
417
418
419
  /* 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
420
    struct gpart *restrict gp = &s->gparts[k];
421
422
423
424
425
426
427
428
429
430
431
432
    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());

433
434
#ifdef WITH_MPI
  /* Move non-local parts to the end of the list. */
435
  const int local_nodeID = s->e->nodeID;
436
  for (size_t k = 0; k < nr_parts;) {
437
    if (cells[ind[k]].nodeID != local_nodeID) {
438
439
      cells[ind[k]].count -= 1;
      nr_parts -= 1;
Matthieu Schaller's avatar
Bug fix    
Matthieu Schaller committed
440
      const struct part tp = s->parts[k];
441
442
      s->parts[k] = s->parts[nr_parts];
      s->parts[nr_parts] = tp;
443
444
445
446
447
448
      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
449
      const struct xpart txp = s->xparts[k];
450
451
      s->xparts[k] = s->xparts[nr_parts];
      s->xparts[nr_parts] = txp;
Matthieu Schaller's avatar
Bug fix    
Matthieu Schaller committed
452
      const int t = ind[k];
453
454
      ind[k] = ind[nr_parts];
      ind[nr_parts] = t;
455
    }
456
457
458
459
460
    else {
      /* Increment when not exchanging otherwise we need to retest "k".*/
      k++;
    }
  }
461

462
  /* Move non-local gparts to the end of the list. */
463
  for (int k = 0; k < nr_gparts;) {
Matthieu Schaller's avatar
Bug fix    
Matthieu Schaller committed
464
465
    if (cells[gind[k]].nodeID != local_nodeID) {
      cells[gind[k]].gcount -= 1;
466
      nr_gparts -= 1;
Matthieu Schaller's avatar
Bug fix    
Matthieu Schaller committed
467
      const struct gpart tp = s->gparts[k];
468
469
      s->gparts[k] = s->gparts[nr_gparts];
      s->gparts[nr_gparts] = tp;
470
471
472
473
474
475
      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
476
477
478
      const int t = gind[k];
      gind[k] = gind[nr_gparts];
      gind[nr_gparts] = t;
479
    }
480
481
482
483
484
    else {
      /* Increment when not exchanging otherwise we need to retest "k".*/
      k++;
    }
  }
485

486
487
  /* Exchange the strays, note that this potentially re-allocates
     the parts arrays. */
488
  size_t nr_parts_exchanged = s->nr_parts - nr_parts;
489
  size_t nr_gparts_exchanged = s->nr_gparts - nr_gparts;
Pedro Gonnet's avatar
Pedro Gonnet committed
490
491
492
493
  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. */
494
  s->nr_parts = nr_parts + nr_parts_exchanged;
495
  s->nr_gparts = nr_gparts + nr_gparts_exchanged;
496
497

  /* Re-allocate the index array if needed.. */
498
  if (s->nr_parts > ind_size) {
499
500
    int *ind_new;
    if ((ind_new = (int *)malloc(sizeof(int) * s->nr_parts)) == NULL)
501
      error("Failed to allocate temporary particle indices.");
502
    memcpy(ind_new, ind, sizeof(int) * nr_parts);
503
504
    free(ind);
    ind = ind_new;
505
506
507
  }

  /* Assign each particle to its cell. */
Pedro Gonnet's avatar
Pedro Gonnet committed
508
  for (size_t k = nr_parts; k < s->nr_parts; k++) {
Matthieu Schaller's avatar
Matthieu Schaller committed
509
    const struct part *const p = &s->parts[k];
510
    ind[k] =
511
        cell_getid(cdim, p->x[0] * ih[0], p->x[1] * ih[1], p->x[2] * ih[2]);
512
513
514
515
    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 ); */
516
  }
517
  nr_parts = s->nr_parts;
518
519
520
#endif

  /* Sort the parts according to their cells. */
521
  space_parts_sort(s, ind, nr_parts, 0, s->nr_cells - 1, verbose);
522
523

  /* Re-link the gparts. */
Pedro Gonnet's avatar
Pedro Gonnet committed
524
  for (size_t k = 0; k < nr_parts; k++)
525
    if (s->parts[k].gpart != NULL) s->parts[k].gpart->part = &s->parts[k];
526

527
  /* Verify space_sort_struct. */
528
  /* for ( k = 1 ; k < nr_parts ; k++ ) {
529
      if ( ind[k-1] > ind[k] ) {
530
531
          error( "Sort failed!" );
          }
532
      else if ( ind[k] != cell_getid( cdim , parts[k].x[0]*ih[0] ,
533
534
535
536
537
     parts[k].x[1]*ih[1] , parts[k].x[2]*ih[2] ) )
          error( "Incorrect indices!" );
      } */

  /* We no longer need the indices as of here. */
538
  free(ind);
539

540
541
542
543
#ifdef WITH_MPI

  /* Re-allocate the index array if needed.. */
  if (s->nr_gparts > gind_size) {
544
545
    int *gind_new;
    if ((gind_new = (int *)malloc(sizeof(int) * s->nr_gparts)) == NULL)
546
      error("Failed to allocate temporary g-particle indices.");
547
    memcpy(gind_new, gind, sizeof(int) * nr_gparts);
548
549
550
551
552
    free(gind);
    gind = gind_new;
  }

  /* Assign each particle to its cell. */
553
  for (int k = nr_gparts; k < s->nr_gparts; k++) {
Matthieu Schaller's avatar
Matthieu Schaller committed
554
    const struct gpart *const p = &s->gparts[k];
555
556
    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
557
    cells[gind[k]].gcount += 1;
558
559
560
561
562
    /* 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;
563

564
#endif
565
566

  /* Sort the parts according to their cells. */
Matthieu Schaller's avatar
Matthieu Schaller committed
567
  space_gparts_sort(s, gind, nr_gparts, 0, s->nr_cells - 1, verbose);
568
569

  /* Re-link the parts. */
570
  for (int k = 0; k < nr_gparts; k++)
571
    if (s->gparts[k].id > 0) s->gparts[k].part->gpart = &s->gparts[k];
572
573

  /* We no longer need the indices as of here. */
574
  free(gind);
575

576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
  /* 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 !");
    }
  }

598
599
  /* Hook the cells up to the parts. */
  // tic = getticks();
600
601
602
  struct part *finger = s->parts;
  struct xpart *xfinger = s->xparts;
  struct gpart *gfinger = s->gparts;
603
604
  for (int k = 0; k < s->nr_cells; k++) {
    struct cell *restrict c = &cells[k];
605
606
607
608
609
610
611
    c->parts = finger;
    c->xparts = xfinger;
    c->gparts = gfinger;
    finger = &finger[c->count];
    xfinger = &xfinger[c->count];
    gfinger = &gfinger[c->gcount];
  }
612
  // message( "hooking up cells took %.3f %s." ,
Matthieu Schaller's avatar
Matthieu Schaller committed
613
  // clocks_from_ticks(getticks() - tic), clocks_getunit());
614
615
616

  /* At this point, we have the upper-level cells, old or new. Now make
     sure that the parts in each cell are ok. */
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
  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
633
  const ticks tic = getticks();
634
635

  for (int k = 0; k < s->nr_cells; k++)
636
637
    scheduler_addtask(&s->e->sched, task_type_split_cell, task_subtype_none, k,
                      0, &cells[k], NULL, 0);
638
  engine_launch(s->e, s->e->nr_threads, 1 << task_type_split_cell, 0);
639

640
641
642
  if (verbose)
    message("took %.3f %s.", clocks_from_ticks(getticks() - tic),
            clocks_getunit());
643
}
644

645
/**
646
647
 * @brief Sort the particles and condensed particles according to the given
 *indices.
648
 *
Matthieu Schaller's avatar
Matthieu Schaller committed
649
 * @param s The #space.
650
651
652
653
 * @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.
654
 * @param verbose Are we talkative ?
655
 */
656

657
void space_parts_sort(struct space *s, int *ind, size_t N, int min, int max,
658
659
                      int verbose) {

Matthieu Schaller's avatar
Matthieu Schaller committed
660
  const ticks tic = getticks();
661
662

  /*Populate the global parallel_sort structure with the input data */
663
664
665
  space_sort_struct.parts = s->parts;
  space_sort_struct.xparts = s->xparts;
  space_sort_struct.ind = ind;
666
  space_sort_struct.stack_size = 2 * (max - min + 1) + 10 + s->e->nr_threads;
667
668
669
670
671
672
  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;

673
  /* Add the first interval. */
674
675
676
677
678
679
680
681
682
  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;

683
  /* Launch the sorting tasks. */
684
  engine_launch(s->e, s->e->nr_threads, (1 << task_type_part_sort), 0);
685
686

  /* Verify space_sort_struct. */
687
  /* for (int i = 1; i < N; i++)
688
    if (ind[i - 1] > ind[i])
689
690
      error("Sorting failed (ind[%i]=%i,ind[%i]=%i), min=%i, max=%i.", i - 1,
  ind[i - 1], i,
691
692
            ind[i], min, max);
  message("Sorting succeeded."); */
693

694
  /* Clean up. */
695
  free(space_sort_struct.stack);
696
697
698
699

  if (verbose)
    message("took %.3f %s.", clocks_from_ticks(getticks() - tic),
            clocks_getunit());
700
}
701

702
void space_do_parts_sort() {
703

704
  /* Pointers to the sorting data. */
705
  int *ind = space_sort_struct.ind;
706
707
  struct part *parts = space_sort_struct.parts;
  struct xpart *xparts = space_sort_struct.xparts;
708

709
  /* Main loop. */
710
  while (space_sort_struct.waiting) {
711

712
    /* Grab an interval off the queue. */
713
714
    int qid =
        atomic_inc(&space_sort_struct.first) % space_sort_struct.stack_size;
715

716
    /* Wait for the entry to be ready, or for the sorting do be done. */
717
718
    while (!space_sort_struct.stack[qid].ready)
      if (!space_sort_struct.waiting) return;
719

720
    /* Get the stack entry. */
721
722
    ptrdiff_t i = space_sort_struct.stack[qid].i;
    ptrdiff_t j = space_sort_struct.stack[qid].j;
723
724
    int min = space_sort_struct.stack[qid].min;
    int max = space_sort_struct.stack[qid].max;
725
    space_sort_struct.stack[qid].ready = 0;
726

727
728
    /* Loop over sub-intervals. */
    while (1) {
729

730
      /* Bring beer. */
731
      const int pivot = (min + max) / 2;
732
733
      /* message("Working on interval [%i,%i] with min=%i, max=%i, pivot=%i.",
              i, j, min, max, pivot); */
734
735

      /* One pass of QuickSort's partitioning. */
736
737
      ptrdiff_t ii = i;
      ptrdiff_t jj = j;
738
739
740
741
      while (ii < jj) {
        while (ii <= j && ind[ii] <= pivot) ii++;
        while (jj >= i && ind[jj] > pivot) jj--;
        if (ii < jj) {
742
          size_t temp_i = ind[ii];
743
744
          ind[ii] = ind[jj];
          ind[jj] = temp_i;
745
          struct part temp_p = parts[ii];
746
747
          parts[ii] = parts[jj];
          parts[jj] = temp_p;
748
          struct xpart temp_xp = xparts[ii];
749
750
751
752
          xparts[ii] = xparts[jj];
          xparts[jj] = temp_xp;
        }
      }
753

754
      /* Verify space_sort_struct. */
755
756
757
758
759
760
761
762
763
764
765
766
      /* 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).");
        } */
767
768
769
770
771
772

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

        /* Recurse on the left? */
        if (jj > i && pivot > min) {
773
774
          qid = atomic_inc(&space_sort_struct.last) %
                space_sort_struct.stack_size;
775
776
          while (space_sort_struct.stack[qid].ready)
            ;
777
778
779
780
781
782
783
          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.");
784
          space_sort_struct.stack[qid].ready = 1;
785
        }
786

787
788
789
790
791
792
793
794
795
796
        /* Recurse on the right? */
        if (jj + 1 < j && pivot + 1 < max) {
          i = jj + 1;
          min = pivot + 1;
        } else
          break;

      } else {

        /* Recurse on the right? */
797
        if (pivot + 1 < max) {
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 = 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.");
809
          space_sort_struct.stack[qid].ready = 1;
810
        }
811

812
813
814
815
816
817
818
        /* Recurse on the left? */
        if (jj > i && pivot > min) {
          j = jj;
          max = pivot;
        } else
          break;
      }
819

820
821
    } /* loop over sub-intervals. */

822
    atomic_dec(&space_sort_struct.waiting);
823
824

  } /* main loop. */
825
826
}

827
828
829
830
831
/**
 * @brief Sort the g-particles and condensed particles according to the given
 *indices.
 *
 * @param s The #space.
Matthieu Schaller's avatar
Matthieu Schaller committed
832
833
 * @param ind The indices with respect to which the gparts are sorted.
 * @param N The number of gparts
834
835
836
837
 * @param min Lowest index.
 * @param max highest index.
 * @param verbose Are we talkative ?
 */
Matthieu Schaller's avatar
Matthieu Schaller committed
838
void space_gparts_sort(struct space *s, int *ind, size_t N, int min, int max,
839
840
                       int verbose) {

Matthieu Schaller's avatar
Matthieu Schaller committed
841
  const ticks tic = getticks();
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
877
878
879
880
881
882
883
884

  /*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
885
  int *ind = space_sort_struct.ind;
886
  struct gpart *gparts = space_sort_struct.gparts;
887

888
  /* Main loop. */
889
  while (space_sort_struct.waiting) {
890

891
    /* Grab an interval off the queue. */
892
893
894
895
896
897
    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;
898

899
    /* Get the stack entry. */
900
901
902
903
904
    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;
905
906
907

    /* Loop over sub-intervals. */
    while (1) {
908

909
      /* Bring beer. */
910
911
912
      const int pivot = (min + max) / 2;
      /* message("Working on interval [%i,%i] with min=%i, max=%i, pivot=%i.",
              i, j, min, max, pivot); */
913
914

      /* One pass of QuickSort's partitioning. */
915
916
      ptrdiff_t ii = i;
      ptrdiff_t jj = j;
917
918
919
920
      while (ii < jj) {
        while (ii <= j && ind[ii] <= pivot) ii++;
        while (jj >= i && ind[jj] > pivot) jj--;
        if (ii < jj) {
921
          size_t temp_i = ind[ii];
922
923
          ind[ii] = ind[jj];
          ind[jj] = temp_i;
924
          struct gpart temp_p = gparts[ii];
925
926
927
928
          gparts[ii] = gparts[jj];
          gparts[jj] = temp_p;
        }
      }
929

930
      /* Verify space_sort_struct. */
931
932
933
934
935
936
937
938
939
940
941
942
      /* 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).");
        } */
943
944
945
946
947
948

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

        /* Recurse on the left? */
        if (jj > i && pivot > min) {
949
950
951
952
953
954
955
956
957
958
959
960
          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;
961
        }
962

963
964
965
966
967
968
        /* Recurse on the right? */
        if (jj + 1 < j && pivot + 1 < max) {
          i = jj + 1;
          min = pivot + 1;
        } else
          break;
969

970
971
972
      } else {

        /* Recurse on the right? */
973
        if (pivot + 1 < max) {
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 = 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.");
          space_sort_struct.stack[qid].ready = 1;
986
987
988
989
990
991
992
993
994
995
996
997
        }

        /* Recurse on the left? */
        if (jj > i && pivot > min) {
          j = jj;
          max = pivot;
        } else
          break;
      }

    } /* loop over sub-intervals. */

998
    atomic_dec(&space_sort_struct.waiting);
999
1000

  } /* main loop. */
1001
}
1002

Pedro Gonnet's avatar
Pedro Gonnet committed
1003
/**
1004
 * @brief Mapping function to free the sorted indices buffers.
Pedro Gonnet's avatar
Pedro Gonnet committed
1005
1006
 */

1007
void space_map_clearsort(struct cell *c, void *data) {
Pedro Gonnet's avatar
Pedro Gonnet committed
1008

1009
1010
1011
1012
1013
  if (c->sort != NULL) {
    free(c->sort);
    c->sort = NULL;
  }
}
Pedro Gonnet's avatar
Pedro Gonnet committed
1014

1015
1016
1017
/**
 * @brief Map a function to all particles in a cell recursively.
 *
1018
 * @param c The #cell we are working in.
1019
1020
1021
1022
 * @param fun Function pointer to apply on the cells.
 * @param data Data passed to the function fun.
 */

Pedro Gonnet's avatar
Pedro Gonnet committed
1023
1024
1025
1026
static void rec_map_parts(struct cell *c,
                          void (*fun)(struct part *p, struct cell *c,
                                      void *data),
                          void *data) {
1027
1028
1029
1030
1031
1032

  int k;

  /* No progeny? */
  if (!c->split)
    for (k = 0; k < c->count; k++) fun(&c->parts[k], c, data);
Pedro Gonnet's avatar
Pedro Gonnet committed
1033

1034
1035
1036
1037
1038
1039
  /* Otherwise, recurse. */
  else
    for (k = 0; k < 8; k++)
      if (c->progeny[k] != NULL) rec_map_parts(c->progeny[k], fun, data);
}

Pedro Gonnet's avatar
Pedro Gonnet committed
1040
/**
1041
 * @brief Map a function to all particles in a space.
Pedro Gonnet's avatar
Pedro Gonnet committed
1042
1043
 *
 * @param s The #space we are working in.
1044
1045
 * @param fun Function pointer to apply on the cells.
 * @param data Data passed to the function fun.
Pedro Gonnet's avatar
Pedro Gonnet committed
1046
1047
 */

1048
1049
1050
void space_map_parts(struct space *s,
                     void (*fun)(struct part *p, struct cell *c, void *data),
                     void *data) {
Pedro Gonnet's avatar
Pedro Gonnet committed
1051

1052
1053
  int cid = 0;

1054
  /* Call the recursive function on all higher-level cells. */
Pedro Gonnet's avatar
Pedro Gonnet committed
1055
1056
  for (cid = 0; cid < s->nr_cells; cid++)
    rec_map_parts(&s->cells[cid], fun, data);
1057
}
1058

1059
1060
1061