space.c 41.6 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"
Pedro Gonnet's avatar
Pedro Gonnet committed
50

51
52
53
/* Shared sort structure. */
struct parallel_sort space_sort_struct;

Pedro Gonnet's avatar
Pedro Gonnet committed
54
55
/* Split size. */
int space_splitsize = space_splitsize_default;
56
int space_subsize = space_subsize_default;
57
int space_maxsize = space_maxsize_default;
Pedro Gonnet's avatar
Pedro Gonnet committed
58
59
60

/* Map shift vector to sortlist. */
const int sortlistID[27] = {
61
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
    /* ( -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};

89
90
91
92
93
94
95
96
97
98
99
100
/**
 * @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.
 */

101
102
103
104
int space_getsid(struct space *s, struct cell **ci, struct cell **cj,
                 double *shift) {

  /* Get the relative distance between the pairs, wrapping. */
105
106
107
  const int periodic = s->periodic;
  double dx[3];
  for (int k = 0; k < 3; k++) {
108
109
110
111
112
113
114
115
116
117
118
    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. */
119
  int sid = 0;
120
  for (int k = 0; k < 3; k++)
121
122
123
124
    sid = 3 * sid + ((dx[k] < 0.0) ? 0 : ((dx[k] > 0.0) ? 2 : 1));

  /* Switch the cells around? */
  if (runner_flip[sid]) {
125
    struct cell *temp = *ci;
126
127
    *ci = *cj;
    *cj = temp;
128
    for (int k = 0; k < 3; k++) shift[k] = -shift[k];
129
130
131
132
133
134
  }
  sid = sortlistID[sid];

  /* Return the sort ID. */
  return sid;
}
135

136
/**
137
 * @brief Recursively dismantle a cell tree.
138
139
 *
 */
140
141
142
143

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

  if (c->split)
144
    for (int k = 0; k < 8; k++)
145
146
147
148
149
150
151
      if (c->progeny[k] != NULL) {
        space_rebuild_recycle(s, c->progeny[k]);
        space_recycle(s, c->progeny[k]);
        c->progeny[k] = NULL;
      }
}

152
/**
153
 * @brief Re-build the cell grid.
154
 *
155
156
 * @param s The #space.
 * @param cell_max Maximum cell edge length.
157
 * @param verbose Print messages to stdout or not.
158
 */
159

160
void space_regrid(struct space *s, double cell_max, int verbose) {
161

162
163
  float h_max = s->cell_min / kernel_gamma / space_stretch;
  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
  if (nr_parts > 0) {
170
    if (s->cells != NULL) {
Tom Theuns's avatar
Tom Theuns committed
171
      for (int k = 0; k < s->nr_cells; k++) {
Matthieu Schaller's avatar
Matthieu Schaller committed
172
        if (s->cells[k].h_max > h_max) h_max = s->cells[k].h_max;
173
174
      }
    } else {
Tom Theuns's avatar
Tom Theuns committed
175
      for (int k = 0; k < nr_parts; k++) {
Matthieu Schaller's avatar
Matthieu Schaller committed
176
        if (s->parts[k].h > h_max) h_max = s->parts[k].h;
177
178
      }
      s->h_max = h_max;
179
    }
Matthieu Schaller's avatar
Matthieu Schaller committed
180
181
182
  } else {
    /* It would be nice to replace this with something more physical or
     * meaningful */
183
    h_max = s->dim[0] / 16.f;
184
185
186
187
188
189
190
191
192
193
    s->h_max = h_max;
  }

/* 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)
194
      error("Failed to aggregate the rebuild flag across nodes.");
195
196
197
    h_max = buff;
  }
#endif
198
  if (verbose) message("h_max is %.3e (cell_max=%.3e).", h_max, cell_max);
199
200

  /* Get the new putative cell dimensions. */
201
  int cdim[3];
202
  for (int k = 0; k < 3; k++)
203
204
205
206
207
208
209
210
211
212
213
214
215
216
    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.");

/* In MPI-Land, we're not allowed to change the top-level cell size. */
#ifdef WITH_MPI
  if (cdim[0] < s->cdim[0] || cdim[1] < s->cdim[1] || cdim[2] < s->cdim[2])
    error("Root-level change of cell size not allowed.");
#endif
217

218
219
220
221
222
223
224
  /* 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) {
225
      for (int k = 0; k < s->nr_cells; k++) {
226
227
228
229
230
231
232
233
        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. */
234
    for (int k = 0; k < 3; k++) {
235
236
237
238
      s->cdim[k] = cdim[k];
      s->h[k] = s->dim[k] / cdim[k];
      s->ih[k] = 1.0 / s->h[k];
    }
239
    const float dmin = fminf(s->h[0], fminf(s->h[1], s->h[2]));
240
241
242
243
244
245
246

    /* 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));
247
    for (int k = 0; k < s->nr_cells; k++)
248
249
250
      if (lock_init(&s->cells[k].lock) != 0) error("Failed to init spinlock.");

    /* Set the cell location and sizes. */
251
252
253
    for (int i = 0; i < cdim[0]; i++)
      for (int j = 0; j < cdim[1]; j++)
        for (int k = 0; k < cdim[2]; k++) {
254
255
256
257
258
259
260
261
262
263
264
265
266
          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
267
        }
268
269

    /* Be verbose about the change. */
270
271
272
    if (verbose)
      message("set cell dimensions to [ %i %i %i ].", cdim[0], cdim[1],
              cdim[2]);
273
274
275
    fflush(stdout);

  } /* re-build upper-level cells? */
276
277
  // message( "rebuilding upper-level cells took %.3f %s." ,
  // clocks_from_ticks(double)(getticks() - tic), clocks_getunit());
278
279
280
281
282

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

    /* Free the old cells, if they were allocated. */
283
    for (int k = 0; k < s->nr_cells; k++) {
284
285
286
287
288
289
290
291
292
293
294
      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
295
      s->cells[k].init = NULL;
Matthieu Schaller's avatar
Matthieu Schaller committed
296
      s->cells[k].ghost = NULL;
Matthieu Schaller's avatar
Matthieu Schaller committed
297
298
      s->cells[k].drift = NULL;
      s->cells[k].kick = NULL;
299
      s->cells[k].super = &s->cells[k];
300
    }
301
302
    s->maxdepth = 0;
  }
303
304
305
306

  if (verbose)
    message("took %.3f %s.", clocks_from_ticks(getticks() - tic),
            clocks_getunit());
307
}
308
309
310
311
312
313

/**
 * @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.
314
 * @param verbose Print messages to stdout or not
315
316
 *
 */
317

318
void space_rebuild(struct space *s, double cell_max, int verbose) {
319

Matthieu Schaller's avatar
Matthieu Schaller committed
320
  const ticks tic = getticks();
321
322

  /* Be verbose about this. */
323
  // message("re)building space..."); fflush(stdout);
324
325

  /* Re-grid if necessary, or just re-set the cell data. */
326
  space_regrid(s, cell_max, verbose);
327

328
329
330
331
  int nr_parts = s->nr_parts;
  int nr_gparts = s->nr_gparts;
  struct cell *restrict cells = s->cells;

Matthieu Schaller's avatar
Matthieu Schaller committed
332
333
334
  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]};
335
336
337
338

  /* Run through the particles and get their cell index. */
  // tic = getticks();
  const size_t ind_size = s->size_parts;
339
340
  int *ind;
  if ((ind = (int *)malloc(sizeof(int) * ind_size)) == NULL)
341
342
343
344
    error("Failed to allocate temporary particle indices.");
  for (int k = 0; k < nr_parts; k++) {
    struct part *restrict p = &s->parts[k];
    for (int j = 0; j < 3; j++)
345
346
347
348
      if (p->x[j] < 0.0)
        p->x[j] += dim[j];
      else if (p->x[j] >= dim[j])
        p->x[j] -= dim[j];
349
    ind[k] =
350
        cell_getid(cdim, p->x[0] * ih[0], p->x[1] * ih[1], p->x[2] * ih[2]);
Matthieu Schaller's avatar
Matthieu Schaller committed
351
    cells[ind[k]].count++;
352
  }
Pedro Gonnet's avatar
Pedro Gonnet committed
353
354
  // message( "getting particle indices took %.3f %s." ,
  // clocks_from_ticks(getticks() - tic), clocks_getunit()):
355

356
357
358
359
360
361
362
  /* 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
363
    struct gpart *restrict gp = &s->gparts[k];
364
365
366
367
368
369
370
371
372
    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++;
  }
Matthieu Schaller's avatar
Matthieu Schaller committed
373
// message( "getting particle indices took %.3f %s." ,
374
// clocks_from_ticks(getticks() - tic), clocks_getunit());
375
376
377

#ifdef WITH_MPI
  /* Move non-local parts to the end of the list. */
378
  const int local_nodeID = s->e->nodeID;
379
  for (int k = 0; k < nr_parts; k++)
380
    if (cells[ind[k]].nodeID != local_nodeID) {
381
382
      cells[ind[k]].count -= 1;
      nr_parts -= 1;
Matthieu Schaller's avatar
Bug fix    
Matthieu Schaller committed
383
      const struct part tp = s->parts[k];
384
385
      s->parts[k] = s->parts[nr_parts];
      s->parts[nr_parts] = tp;
386
387
388
389
390
391
      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
392
      const struct xpart txp = s->xparts[k];
393
394
      s->xparts[k] = s->xparts[nr_parts];
      s->xparts[nr_parts] = txp;
Matthieu Schaller's avatar
Bug fix    
Matthieu Schaller committed
395
      const int t = ind[k];
396
397
      ind[k] = ind[nr_parts];
      ind[nr_parts] = t;
398
399
    }

400
401
  /* Move non-local gparts to the end of the list. */
  for (int k = 0; k < nr_gparts; k++)
Matthieu Schaller's avatar
Bug fix    
Matthieu Schaller committed
402
403
    if (cells[gind[k]].nodeID != local_nodeID) {
      cells[gind[k]].gcount -= 1;
404
      nr_gparts -= 1;
Matthieu Schaller's avatar
Bug fix    
Matthieu Schaller committed
405
      const struct gpart tp = s->gparts[k];
406
407
      s->gparts[k] = s->gparts[nr_gparts];
      s->gparts[nr_gparts] = tp;
408
409
410
411
412
413
      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
414
415
416
      const int t = gind[k];
      gind[k] = gind[nr_gparts];
      gind[nr_gparts] = t;
417
418
    }

419
420
  /* Exchange the strays, note that this potentially re-allocates
     the parts arrays. */
421
422
423
  /* TODO: This function also exchanges gparts, but this is shorted-out
     until they are fully implemented. */
  size_t nr_parts_exchanged = s->nr_parts - nr_parts;
424
  size_t nr_gparts_exchanged = s->nr_gparts - nr_gparts;
Pedro Gonnet's avatar
Pedro Gonnet committed
425
426
427
  engine_exchange_strays(s->e, nr_parts, &ind[nr_parts], &nr_parts_exchanged,
                         nr_gparts, &gind[nr_gparts], &nr_gparts_exchanged);

428
  /* Add post-processing, i.e. re-linking/creating of gparts here. */
Pedro Gonnet's avatar
Pedro Gonnet committed
429
430

  /* Set the new particle counts. */
431
  s->nr_parts = nr_parts + nr_parts_exchanged;
432
  s->nr_gparts = nr_gparts + nr_gparts_exchanged;
433
434

  /* Re-allocate the index array if needed.. */
435
  if (s->nr_parts > ind_size) {
436
437
    int *ind_new;
    if ((ind_new = (int *)malloc(sizeof(int) * s->nr_parts)) == NULL)
438
      error("Failed to allocate temporary particle indices.");
439
    memcpy(ind_new, ind, sizeof(size_t) * nr_parts);
440
441
    free(ind);
    ind = ind_new;
442
443
444
  }

  /* Assign each particle to its cell. */
445
  for (int k = nr_parts; k < s->nr_parts; k++) {
Matthieu Schaller's avatar
Matthieu Schaller committed
446
    const struct part *const p = &s->parts[k];
447
    ind[k] =
448
        cell_getid(cdim, p->x[0] * ih[0], p->x[1] * ih[1], p->x[2] * ih[2]);
449
450
451
452
    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 ); */
453
  }
454
  nr_parts = s->nr_parts;
455
456
457
#endif

  /* Sort the parts according to their cells. */
458
  space_parts_sort(s, ind, nr_parts, 0, s->nr_cells - 1, verbose);
459

460
  /* Re-link the gparts. */
461
  for (int k = 0; k < nr_parts; k++)
462
    if (s->parts[k].gpart != NULL) s->parts[k].gpart->part = &s->parts[k];
463

464
  /* Verify space_sort_struct. */
465
  /* for ( k = 1 ; k < nr_parts ; k++ ) {
466
      if ( ind[k-1] > ind[k] ) {
467
468
          error( "Sort failed!" );
          }
469
      else if ( ind[k] != cell_getid( cdim , parts[k].x[0]*ih[0] ,
470
471
472
473
474
     parts[k].x[1]*ih[1] , parts[k].x[2]*ih[2] ) )
          error( "Incorrect indices!" );
      } */

  /* We no longer need the indices as of here. */
475
  free(ind);
476

Tom Theuns's avatar
Tom Theuns committed
477
#ifdef WITH_MPI
478
479
480

  /* Re-allocate the index array if needed.. */
  if (s->nr_gparts > gind_size) {
481
482
    int *gind_new;
    if ((gind_new = (int *)malloc(sizeof(int) * s->nr_gparts)) == NULL)
483
      error("Failed to allocate temporary g-particle indices.");
484
    memcpy(gind_new, gind, sizeof(int) * nr_gparts);
485
486
487
488
489
    free(gind);
    gind = gind_new;
  }

  /* Assign each particle to its cell. */
490
  for (int k = nr_gparts; k < s->nr_gparts; k++) {
Matthieu Schaller's avatar
Matthieu Schaller committed
491
    const struct gpart *const p = &s->gparts[k];
492
493
    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
494
    cells[gind[k]].gcount += 1;
495
496
497
498
499
500
    /* 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;

Tom Theuns's avatar
Tom Theuns committed
501
#endif
502
503

  /* Sort the parts according to their cells. */
Matthieu Schaller's avatar
Matthieu Schaller committed
504
  space_gparts_sort(s, gind, nr_gparts, 0, s->nr_cells - 1, verbose);
505
506

  /* Re-link the parts. */
507
  for (int k = 0; k < nr_gparts; k++)
508
    if (s->gparts[k].id > 0) s->gparts[k].part->gpart = &s->gparts[k];
509
510

  /* We no longer need the indices as of here. */
511
  free(gind);
512

513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
  /* 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 !");
    }
  }

535
536
  /* Hook the cells up to the parts. */
  // tic = getticks();
537
538
539
  struct part *finger = s->parts;
  struct xpart *xfinger = s->xparts;
  struct gpart *gfinger = s->gparts;
540
541
  for (int k = 0; k < s->nr_cells; k++) {
    struct cell *restrict c = &cells[k];
542
543
544
545
546
547
548
    c->parts = finger;
    c->xparts = xfinger;
    c->gparts = gfinger;
    finger = &finger[c->count];
    xfinger = &xfinger[c->count];
    gfinger = &gfinger[c->gcount];
  }
549
  // message( "hooking up cells took %.3f %s." ,
Matthieu Schaller's avatar
Matthieu Schaller committed
550
  // clocks_from_ticks(getticks() - tic), clocks_getunit());
551
552
553

  /* At this point, we have the upper-level cells, old or new. Now make
     sure that the parts in each cell are ok. */
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
  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
570
  const ticks tic = getticks();
571
572

  for (int k = 0; k < s->nr_cells; k++)
573
574
    scheduler_addtask(&s->e->sched, task_type_split_cell, task_subtype_none, k,
                      0, &cells[k], NULL, 0);
575
  engine_launch(s->e, s->e->nr_threads, 1 << task_type_split_cell, 0);
576

577
578
579
  if (verbose)
    message("took %.3f %s.", clocks_from_ticks(getticks() - tic),
            clocks_getunit());
580
}
581

582
/**
583
584
 * @brief Sort the particles and condensed particles according to the given
 *indices.
585
 *
Matthieu Schaller's avatar
Matthieu Schaller committed
586
 * @param s The #space.
587
588
589
590
 * @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.
591
 * @param verbose Are we talkative ?
592
 */
593

594
void space_parts_sort(struct space *s, int *ind, size_t N, int min, int max,
595
596
                      int verbose) {

Matthieu Schaller's avatar
Matthieu Schaller committed
597
  const ticks tic = getticks();
598
599

  /*Populate the global parallel_sort structure with the input data */
600
601
602
  space_sort_struct.parts = s->parts;
  space_sort_struct.xparts = s->xparts;
  space_sort_struct.ind = ind;
603
  space_sort_struct.stack_size = 2 * (max - min + 1) + 10 + s->e->nr_threads;
604
605
606
607
608
609
  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;

610
  /* Add the first interval. */
611
612
613
614
615
616
617
618
619
  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;

620
  /* Launch the sorting tasks. */
621
  engine_launch(s->e, s->e->nr_threads, (1 << task_type_part_sort), 0);
622
623

  /* Verify space_sort_struct. */
624
  /* for (int i = 1; i < N; i++)
625
    if (ind[i - 1] > ind[i])
626
627
      error("Sorting failed (ind[%i]=%i,ind[%i]=%i), min=%i, max=%i.", i - 1,
  ind[i - 1], i,
628
629
            ind[i], min, max);
  message("Sorting succeeded."); */
630

631
  /* Clean up. */
632
  free(space_sort_struct.stack);
633
634
635
636

  if (verbose)
    message("took %.3f %s.", clocks_from_ticks(getticks() - tic),
            clocks_getunit());
637
}
638

639
void space_do_parts_sort() {
640

641
  /* Pointers to the sorting data. */
642
  int *ind = space_sort_struct.ind;
643
644
  struct part *parts = space_sort_struct.parts;
  struct xpart *xparts = space_sort_struct.xparts;
645

646
  /* Main loop. */
647
  while (space_sort_struct.waiting) {
648

649
    /* Grab an interval off the queue. */
650
651
    int qid =
        atomic_inc(&space_sort_struct.first) % space_sort_struct.stack_size;
652

653
    /* Wait for the entry to be ready, or for the sorting do be done. */
654
655
    while (!space_sort_struct.stack[qid].ready)
      if (!space_sort_struct.waiting) return;
656

657
    /* Get the stack entry. */
658
659
    ptrdiff_t i = space_sort_struct.stack[qid].i;
    ptrdiff_t j = space_sort_struct.stack[qid].j;
660
661
    int min = space_sort_struct.stack[qid].min;
    int max = space_sort_struct.stack[qid].max;
662
    space_sort_struct.stack[qid].ready = 0;
663

664
665
    /* Loop over sub-intervals. */
    while (1) {
666

667
      /* Bring beer. */
668
      const int pivot = (min + max) / 2;
669
670
      /* message("Working on interval [%i,%i] with min=%i, max=%i, pivot=%i.",
              i, j, min, max, pivot); */
671
672

      /* One pass of QuickSort's partitioning. */
673
674
      ptrdiff_t ii = i;
      ptrdiff_t jj = j;
675
676
677
678
      while (ii < jj) {
        while (ii <= j && ind[ii] <= pivot) ii++;
        while (jj >= i && ind[jj] > pivot) jj--;
        if (ii < jj) {
679
          size_t temp_i = ind[ii];
680
681
          ind[ii] = ind[jj];
          ind[jj] = temp_i;
682
          struct part temp_p = parts[ii];
683
684
          parts[ii] = parts[jj];
          parts[jj] = temp_p;
685
          struct xpart temp_xp = xparts[ii];
686
687
688
689
          xparts[ii] = xparts[jj];
          xparts[jj] = temp_xp;
        }
      }
690

691
      /* Verify space_sort_struct. */
692
693
694
695
696
697
698
699
700
701
702
703
      /* 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).");
        } */
704
705
706
707
708
709

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

        /* Recurse on the left? */
        if (jj > i && pivot > min) {
710
711
          qid = atomic_inc(&space_sort_struct.last) %
                space_sort_struct.stack_size;
712
713
          while (space_sort_struct.stack[qid].ready)
            ;
714
715
716
717
718
719
720
          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.");
721
          space_sort_struct.stack[qid].ready = 1;
722
        }
723

724
725
726
727
728
729
730
731
732
733
        /* Recurse on the right? */
        if (jj + 1 < j && pivot + 1 < max) {
          i = jj + 1;
          min = pivot + 1;
        } else
          break;

      } else {

        /* Recurse on the right? */
734
        if (pivot + 1 < max) {
735
736
          qid = atomic_inc(&space_sort_struct.last) %
                space_sort_struct.stack_size;
737
738
          while (space_sort_struct.stack[qid].ready)
            ;
739
740
741
742
743
744
745
          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.");
746
          space_sort_struct.stack[qid].ready = 1;
747
        }
748

749
750
751
752
753
754
755
        /* Recurse on the left? */
        if (jj > i && pivot > min) {
          j = jj;
          max = pivot;
        } else
          break;
      }
756

757
758
    } /* loop over sub-intervals. */

759
    atomic_dec(&space_sort_struct.waiting);
760
761

  } /* main loop. */
762
763
}

764
765
766
767
768
/**
 * @brief Sort the g-particles and condensed particles according to the given
 *indices.
 *
 * @param s The #space.
Matthieu Schaller's avatar
Matthieu Schaller committed
769
770
 * @param ind The indices with respect to which the gparts are sorted.
 * @param N The number of gparts
771
772
773
774
 * @param min Lowest index.
 * @param max highest index.
 * @param verbose Are we talkative ?
 */
Matthieu Schaller's avatar
Matthieu Schaller committed
775
void space_gparts_sort(struct space *s, int *ind, size_t N, int min, int max,
776
777
                       int verbose) {

Matthieu Schaller's avatar
Matthieu Schaller committed
778
  const ticks tic = getticks();
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821

  /*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
822
  int *ind = space_sort_struct.ind;
823
  struct gpart *gparts = space_sort_struct.gparts;
824

825
  /* Main loop. */
826
  while (space_sort_struct.waiting) {
827

828
    /* Grab an interval off the queue. */
829
830
831
832
833
834
    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;
835

836
    /* Get the stack entry. */
837
838
839
840
841
    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;
842
843
844

    /* Loop over sub-intervals. */
    while (1) {
845

846
      /* Bring beer. */
847
848
849
      const int pivot = (min + max) / 2;
      /* message("Working on interval [%i,%i] with min=%i, max=%i, pivot=%i.",
              i, j, min, max, pivot); */
850
851

      /* One pass of QuickSort's partitioning. */
852
853
      ptrdiff_t ii = i;
      ptrdiff_t jj = j;
854
855
856
857
      while (ii < jj) {
        while (ii <= j && ind[ii] <= pivot) ii++;
        while (jj >= i && ind[jj] > pivot) jj--;
        if (ii < jj) {
858
          size_t temp_i = ind[ii];
859
860
          ind[ii] = ind[jj];
          ind[jj] = temp_i;
861
          struct gpart temp_p = gparts[ii];
862
863
864
865
          gparts[ii] = gparts[jj];
          gparts[jj] = temp_p;
        }
      }
866

867
      /* Verify space_sort_struct. */
868
869
870
871
872
873
874
875
876
877
878
879
      /* 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).");
        } */
880
881
882
883
884
885

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

        /* Recurse on the left? */
        if (jj > i && pivot > min) {
886
887
888
889
890
891
892
893
894
895
896
897
          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;
898
        }
899

900
901
902
903
904
905
        /* Recurse on the right? */
        if (jj + 1 < j && pivot + 1 < max) {
          i = jj + 1;
          min = pivot + 1;
        } else
          break;
906

907
908
909
      } else {

        /* Recurse on the right? */
910
        if (pivot + 1 < max) {
911
912
913
914
915
916
917
918
919
920
921
922
          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;
923
924
925
926
927
928
929
930
931
932
933
934
        }

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

    } /* loop over sub-intervals. */

935
    atomic_dec(&space_sort_struct.waiting);
936
937

  } /* main loop. */
938
}
939

Pedro Gonnet's avatar
Pedro Gonnet committed
940
/**
941
 * @brief Mapping function to free the sorted indices buffers.
Pedro Gonnet's avatar
Pedro Gonnet committed
942
943
 */

944
void space_map_clearsort(struct cell *c, void *data) {
Pedro Gonnet's avatar
Pedro Gonnet committed
945

946
947
948
949
950
  if (c->sort != NULL) {
    free(c->sort);
    c->sort = NULL;
  }
}
Pedro Gonnet's avatar
Pedro Gonnet committed
951

952
953
954
/**
 * @brief Map a function to all particles in a cell recursively.
 *
955
 * @param c The #cell we are working in.
956
957
958
959
 * @param fun Function pointer to apply on the cells.
 * @param data Data passed to the function fun.
 */

Pedro Gonnet's avatar
Pedro Gonnet committed
960
961
962
963
static void rec_map_parts(struct cell *c,
                          void (*fun)(struct part *p, struct cell *c,
                                      void *data),
                          void *data) {
964
965
966
967
968
969

  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
970

971
972
973
974
975
976
  /* 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
977
/**
978
 * @brief Map a function to all particles in a space.
Pedro Gonnet's avatar
Pedro Gonnet committed
979
980
 *
 * @param s The #space we are working in.
981
982
 * @param fun Function pointer to apply on the cells.
 * @param data Data passed to the function fun.
Pedro Gonnet's avatar
Pedro Gonnet committed
983
984
 */

985
986
987
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
988

989
990
  int cid = 0;

991
  /* Call the recursive function on all higher-level cells. */
Pedro Gonnet's avatar
Pedro Gonnet committed
992
993
  for (cid = 0; cid < s->nr_cells; cid++)
    rec_map_parts(&s->cells[cid], fun, data);
994
}
995

996
997
998
999
1000
1001
1002
1003
/**
 * @brief Map a function to all particles in a cell recursively.
 *
 * @param c The #cell we are working in.
 * @param fun Function pointer to apply on the cells.
 */

static void rec_map_parts_xparts(struct cell *c,
1004
1005
                                 void (*fun)(struct part *p, struct xpart *xp,
                                             struct cell *c)) {
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026

  int k;

  /* No progeny? */
  if (!c->split)
    for (k = 0; k < c->count; k++) fun(&c->parts[k], &c->xparts[k], c);

  /* Otherwise, recurse. */
  else
    for (k = 0; k < 8; k++)
      if (c->progeny[k] != NULL) rec_map_parts_xparts(c->progeny[k], fun);
}

/**
 * @brief Map a function to all particles (#part and #xpart) in a space.
 *
 * @param s The #space we are working in.
 * @param fun Function pointer to apply on the particles in the cells.
 */

void space_map_parts_xparts(struct space *s,
1027
1028
                            void (*fun)(struct part *p, struct xpart *xp,
                                        struct cell *c)) {
1029
1030
1031
1032
1033
1034
1035
1036

  int cid = 0;

  /* Call the recursive function on all higher-level cells. */
  for (cid = 0; cid < s->nr_cells; cid++)
    rec_map_parts_xparts(&s->cells[cid], fun);
}

1037
1038
1039
/**
 * @brief Map a function to all particles in a cell recursively.
 *
1040
 * @param c The #cell we are working in.
1041
1042
1043
1044
 * @param full Map to all cells, including cells with sub-cells.
 * @param fun Function pointer to apply on the cells.
 * @param data Data passed to the function fun.
 */
1045

Pedro Gonnet's avatar
Pedro Gonnet committed
1046
1047
1048
static void rec_map_cells_post(struct cell *c, int full,
                               void (*fun)(struct cell *c, void *data),
                               void *data) {
1049

1050
1051
1052
1053
1054
  int k;

  /* Recurse. */
  if (c->split)
    for (k = 0; k < 8; k++)
Pedro Gonnet's avatar
Pedro Gonnet committed
1055
1056
1057
      if (c->progeny[k] != NULL)
        rec_map_cells_post(c->progeny[k], full, fun, data);

1058
1059
  /* No progeny? */
  if (full || !c->split) fun(c, data);
1060
}
Pedro Gonnet's avatar
Pedro Gonnet committed
1061
1062

/**
1063
 * @brief Map a function to all particles in a aspace.