space.c 50.9 KB
Newer Older
1
/*******************************************************************************
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
 * This file is part of SWIFT.
 * Copyright (c) 2012 Pedro Gonnet (pedro.gonnet@durham.ac.uk)
 *                    Matthieu Schaller (matthieu.schaller@durham.ac.uk)
 *               2015 Peter W. Draper (p.w.draper@durham.ac.uk)
 *               2016 John A. Regan (john.a.regan@durham.ac.uk)
 *                    Tom Theuns (tom.theuns@durham.ac.uk)
 *
 * This program is free software: you can redistribute it and/or modify
 * it under the terms of the GNU Lesser General Public License as published
 * by the Free Software Foundation, either version 3 of the License, or
 * (at your option) any later version.
 *
 * This program is distributed in the hope that it will be useful,
 * but WITHOUT ANY WARRANTY; without even the implied warranty of
 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
 * GNU General Public License for more details.
 *
 * You should have received a copy of the GNU Lesser General Public License
 * along with this program.  If not, see <http://www.gnu.org/licenses/>.
 *
 ******************************************************************************/
Pedro Gonnet's avatar
Pedro Gonnet committed
23
24
25
26
27
28
29
30

/* Config parameters. */
#include "../config.h"

/* Some standard headers. */
#include <float.h>
#include <limits.h>
#include <math.h>
31
#include <stdlib.h>
32
#include <string.h>
Pedro Gonnet's avatar
Pedro Gonnet committed
33

34
35
/* MPI headers. */
#ifdef WITH_MPI
36
#include <mpi.h>
37
38
#endif

39
40
41
/* This object's header. */
#include "space.h"

Pedro Gonnet's avatar
Pedro Gonnet committed
42
/* Local headers. */
43
#include "atomic.h"
44
#include "const.h"
45
#include "engine.h"
46
#include "error.h"
47
48
#include "gravity.h"
#include "hydro.h"
49
#include "kernel_hydro.h"
50
#include "lock.h"
51
#include "minmax.h"
52
#include "runner.h"
53
#include "threadpool.h"
54
#include "tools.h"
Pedro Gonnet's avatar
Pedro Gonnet committed
55
56
57

/* Split size. */
int space_splitsize = space_splitsize_default;
58
int space_subsize = space_subsize_default;
59
int space_maxsize = space_maxsize_default;
Pedro Gonnet's avatar
Pedro Gonnet committed
60
61
62

/* Map shift vector to sortlist. */
const int sortlistID[27] = {
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
90
    /* ( -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};

91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
/**
 * @brief Interval stack necessary for parallel particle sorting.
 */
struct qstack {
  volatile ptrdiff_t i, j;
  volatile int min, max;
  volatile int ready;
};

/**
 * @brief Parallel particle-sorting stack
 */
struct parallel_sort {
  struct part *parts;
  struct gpart *gparts;
  struct xpart *xparts;
  int *ind;
  struct qstack *stack;
  unsigned int stack_size;
  volatile unsigned int first, last, waiting;
};

113
114
115
116
117
118
119
120
121
122
123
/**
 * @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.
 */
124
125
126
127
int space_getsid(struct space *s, struct cell **ci, struct cell **cj,
                 double *shift) {

  /* Get the relative distance between the pairs, wrapping. */
128
129
130
  const int periodic = s->periodic;
  double dx[3];
  for (int k = 0; k < 3; k++) {
131
132
133
134
135
136
137
138
139
140
141
    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. */
142
  int sid = 0;
143
  for (int k = 0; k < 3; k++)
144
145
146
147
    sid = 3 * sid + ((dx[k] < 0.0) ? 0 : ((dx[k] > 0.0) ? 2 : 1));

  /* Switch the cells around? */
  if (runner_flip[sid]) {
148
    struct cell *temp = *ci;
149
150
    *ci = *cj;
    *cj = temp;
151
    for (int k = 0; k < 3; k++) shift[k] = -shift[k];
152
153
154
155
156
157
  }
  sid = sortlistID[sid];

  /* Return the sort ID. */
  return sid;
}
158

159
/**
160
 * @brief Recursively dismantle a cell tree.
161
 *
162
163
 * @param s The #space.
 * @param c The #cell to recycle.
164
 */
165
166
167
void space_rebuild_recycle(struct space *s, struct cell *c) {

  if (c->split)
168
    for (int k = 0; k < 8; k++)
169
170
171
172
173
174
175
      if (c->progeny[k] != NULL) {
        space_rebuild_recycle(s, c->progeny[k]);
        space_recycle(s, c->progeny[k]);
        c->progeny[k] = NULL;
      }
}

176
/**
177
 * @brief Re-build the top-level cell grid.
178
 *
179
180
 * @param s The #space.
 * @param cell_max Maximum cell edge length.
181
 * @param verbose Print messages to stdout or not.
182
 */
183
void space_regrid(struct space *s, double cell_max, int verbose) {
184

185
  const size_t nr_parts = s->nr_parts;
186
  const ticks tic = getticks();
187
  const int ti_current = (s->e != NULL) ? s->e->ti_current : 0;
188

189
  /* Run through the cells and get the current h_max. */
190
  // tic = getticks();
191
  float h_max = s->cell_min / kernel_gamma / space_stretch;
192
  if (nr_parts > 0) {
193
    if (s->cells_top != NULL) {
Tom Theuns's avatar
Tom Theuns committed
194
      for (int k = 0; k < s->nr_cells; k++) {
195
196
197
        if (s->cells_top[k].nodeID == engine_rank &&
            s->cells_top[k].h_max > h_max) {
          h_max = s->cells_top[k].h_max;
198
        }
199
200
      }
    } else {
201
      for (size_t k = 0; k < nr_parts; k++) {
Matthieu Schaller's avatar
Matthieu Schaller committed
202
        if (s->parts[k].h > h_max) h_max = s->parts[k].h;
203
      }
204
205
206
207
208
209
210
211
212
213
    }
  }

/* 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)
214
      error("Failed to aggregate the rebuild flag across nodes.");
215
216
217
    h_max = buff;
  }
#endif
218
  if (verbose) message("h_max is %.3e (cell_max=%.3e).", h_max, cell_max);
219
220

  /* Get the new putative cell dimensions. */
221
222
223
224
  const int cdim[3] = {
      floor(s->dim[0] / fmax(h_max * kernel_gamma * space_stretch, cell_max)),
      floor(s->dim[1] / fmax(h_max * kernel_gamma * space_stretch, cell_max)),
      floor(s->dim[2] / fmax(h_max * kernel_gamma * space_stretch, cell_max))};
225
226
227
228
229
230
231

  /* 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.");

232
233
234
235
236
237
  /* Check if we have enough cells for gravity. */
  if (s->gravity && (cdim[0] < 8 || cdim[1] < 8 || cdim[2] < 8))
    error(
        "Must have at least 8 cells in each spatial dimension when gravity "
        "is switched on.");

238
239
240
/* 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. */
241
#ifdef WITH_MPI
Matthieu Schaller's avatar
Matthieu Schaller committed
242
  double oldwidth[3];
243
244
245
246
247
248
249
250
  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];
251
252
253
    oldwidth[0] = s->width[0];
    oldwidth[1] = s->width[1];
    oldwidth[2] = s->width[2];
254
255
256
257
258
259
260
261
262

    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);
263
          oldnodeIDs[cid] = s->cells_top[cid].nodeID;
264
265
266
267
268
        }
      }
    }
  }

269
270
271
272
#endif

  /* Do we need to re-build the upper-level cells? */
  // tic = getticks();
273
  if (s->cells_top == NULL || cdim[0] < s->cdim[0] || cdim[1] < s->cdim[1] ||
274
275
276
      cdim[2] < s->cdim[2]) {

    /* Free the old cells, if they were allocated. */
277
    if (s->cells_top != NULL) {
278
      for (int k = 0; k < s->nr_cells; k++) {
279
280
        space_rebuild_recycle(s, &s->cells_top[k]);
        if (s->cells_top[k].sort != NULL) free(s->cells_top[k].sort);
281
      }
282
      free(s->cells_top);
283
284
285
286
      s->maxdepth = 0;
    }

    /* Set the new cell dimensions only if smaller. */
287
    for (int k = 0; k < 3; k++) {
288
      s->cdim[k] = cdim[k];
289
290
      s->width[k] = s->dim[k] / cdim[k];
      s->iwidth[k] = 1.0 / s->width[k];
291
    }
292
    const float dmin = fminf(s->width[0], fminf(s->width[1], s->width[2]));
293
294
295

    /* Allocate the highest level of cells. */
    s->tot_cells = s->nr_cells = cdim[0] * cdim[1] * cdim[2];
296
    if (posix_memalign((void *)&s->cells_top, cell_align,
297
298
                       s->nr_cells * sizeof(struct cell)) != 0)
      error("Failed to allocate cells.");
299
    bzero(s->cells_top, s->nr_cells * sizeof(struct cell));
300
    for (int k = 0; k < s->nr_cells; k++)
301
302
      if (lock_init(&s->cells_top[k].lock) != 0)
        error("Failed to init spinlock.");
303
304

    /* Set the cell location and sizes. */
305
306
307
    for (int i = 0; i < cdim[0]; i++)
      for (int j = 0; j < cdim[1]; j++)
        for (int k = 0; k < cdim[2]; k++) {
308
          struct cell *restrict c = &s->cells_top[cell_getid(cdim, i, j, k)];
309
310
311
312
313
314
          c->loc[0] = i * s->width[0];
          c->loc[1] = j * s->width[1];
          c->loc[2] = k * s->width[2];
          c->width[0] = s->width[0];
          c->width[1] = s->width[1];
          c->width[2] = s->width[2];
315
316
317
318
319
          c->dmin = dmin;
          c->depth = 0;
          c->count = 0;
          c->gcount = 0;
          c->super = c;
320
          c->gsuper = c;
321
          c->ti_old = ti_current;
322
          lock_init(&c->lock);
Pedro Gonnet's avatar
Pedro Gonnet committed
323
        }
324
325

    /* Be verbose about the change. */
326
327
328
    if (verbose)
      message("set cell dimensions to [ %i %i %i ].", cdim[0], cdim[1],
              cdim[2]);
329
330
    fflush(stdout);

331
#ifdef WITH_MPI
332
333
334
335
336
    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)
337
338
339
        message(
            "basic cell dimensions have increased - recalculating the "
            "global partition.");
340

Matthieu Schaller's avatar
Matthieu Schaller committed
341
      if (!partition_space_to_space(oldwidth, oldcdim, oldnodeIDs, s)) {
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359

        /* 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);
360

361
362
      /* Finished with these. */
      free(oldnodeIDs);
363
364
    }
#endif
365
366
367
368

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

369
370
  } /* re-build upper-level cells? */

371
  else { /* Otherwise, just clean up the cells. */
372
373

    /* Free the old cells, if they were allocated. */
374
    for (int k = 0; k < s->nr_cells; k++) {
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
      space_rebuild_recycle(s, &s->cells_top[k]);
      s->cells_top[k].sorts = NULL;
      s->cells_top[k].nr_tasks = 0;
      s->cells_top[k].nr_density = 0;
      s->cells_top[k].nr_gradient = 0;
      s->cells_top[k].nr_force = 0;
      s->cells_top[k].density = NULL;
      s->cells_top[k].gradient = NULL;
      s->cells_top[k].force = NULL;
      s->cells_top[k].dx_max = 0.0f;
      s->cells_top[k].sorted = 0;
      s->cells_top[k].count = 0;
      s->cells_top[k].gcount = 0;
      s->cells_top[k].init = NULL;
      s->cells_top[k].extra_ghost = NULL;
      s->cells_top[k].ghost = NULL;
      s->cells_top[k].kick = NULL;
      s->cells_top[k].super = &s->cells_top[k];
393
      s->cells_top[k].gsuper = &s->cells[k];
394
    }
395
396
    s->maxdepth = 0;
  }
397
398
399
400

  if (verbose)
    message("took %.3f %s.", clocks_from_ticks(getticks() - tic),
            clocks_getunit());
401
}
402
403
404
405
406
407

/**
 * @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.
408
 * @param verbose Print messages to stdout or not
409
410
 *
 */
411
void space_rebuild(struct space *s, double cell_max, int verbose) {
412

Matthieu Schaller's avatar
Matthieu Schaller committed
413
  const ticks tic = getticks();
414
415

  /* Be verbose about this. */
416
  // message("re)building space..."); fflush(stdout);
417
418

  /* Re-grid if necessary, or just re-set the cell data. */
419
  space_regrid(s, cell_max, verbose);
420

Pedro Gonnet's avatar
Pedro Gonnet committed
421
422
  size_t nr_parts = s->nr_parts;
  size_t nr_gparts = s->nr_gparts;
423
  struct cell *restrict cells_top = s->cells_top;
424
  const int ti_current = (s->e != NULL) ? s->e->ti_current : 0;
425

426
  const double ih[3] = {s->iwidth[0], s->iwidth[1], s->iwidth[2]};
Matthieu Schaller's avatar
Matthieu Schaller committed
427
428
  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]};
429
430
431
432

  /* Run through the particles and get their cell index. */
  // tic = getticks();
  const size_t ind_size = s->size_parts;
433
434
  int *ind;
  if ((ind = (int *)malloc(sizeof(int) * ind_size)) == NULL)
435
    error("Failed to allocate temporary particle indices.");
Pedro Gonnet's avatar
Pedro Gonnet committed
436
  for (size_t k = 0; k < nr_parts; k++) {
437
438
    struct part *restrict p = &s->parts[k];
    for (int j = 0; j < 3; j++)
439
440
441
442
      if (p->x[j] < 0.0)
        p->x[j] += dim[j];
      else if (p->x[j] >= dim[j])
        p->x[j] -= dim[j];
443
    ind[k] =
444
        cell_getid(cdim, p->x[0] * ih[0], p->x[1] * ih[1], p->x[2] * ih[2]);
445
    cells_top[ind[k]].count++;
446
  }
Pedro Gonnet's avatar
Pedro Gonnet committed
447
448
  // message( "getting particle indices took %.3f %s." ,
  // clocks_from_ticks(getticks() - tic), clocks_getunit()):
449

450
451
452
453
454
455
  /* 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.");
456
  for (size_t k = 0; k < nr_gparts; k++) {
Matthieu Schaller's avatar
Matthieu Schaller committed
457
    struct gpart *restrict gp = &s->gparts[k];
458
459
460
461
462
463
464
    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]);
465
    cells_top[gind[k]].gcount++;
466
  }
467
// message( "getting g-particle indices took %.3f %s." ,
468
// clocks_from_ticks(getticks() - tic), clocks_getunit());
469
470

#ifdef WITH_MPI
471

472
  /* Move non-local parts to the end of the list. */
473
  const int local_nodeID = s->e->nodeID;
474
  for (size_t k = 0; k < nr_parts;) {
475
476
    if (cells_top[ind[k]].nodeID != local_nodeID) {
      cells_top[ind[k]].count -= 1;
477
      nr_parts -= 1;
Matthieu Schaller's avatar
Bug fix    
Matthieu Schaller committed
478
      const struct part tp = s->parts[k];
479
480
      s->parts[k] = s->parts[nr_parts];
      s->parts[nr_parts] = tp;
481
      if (s->parts[k].gpart != NULL) {
482
        s->parts[k].gpart->id_or_neg_offset = -k;
483
484
      }
      if (s->parts[nr_parts].gpart != NULL) {
485
        s->parts[nr_parts].gpart->id_or_neg_offset = -nr_parts;
486
      }
Matthieu Schaller's avatar
Bug fix    
Matthieu Schaller committed
487
      const struct xpart txp = s->xparts[k];
488
489
      s->xparts[k] = s->xparts[nr_parts];
      s->xparts[nr_parts] = txp;
Matthieu Schaller's avatar
Bug fix    
Matthieu Schaller committed
490
      const int t = ind[k];
491
492
      ind[k] = ind[nr_parts];
      ind[nr_parts] = t;
Matthieu Schaller's avatar
Matthieu Schaller committed
493
    } else {
494
495
496
497
      /* Increment when not exchanging otherwise we need to retest "k".*/
      k++;
    }
  }
498

499
#ifdef SWIFT_DEBUG_CHECKS
Peter W. Draper's avatar
Peter W. Draper committed
500
  /* Check that all parts are in the correct places. */
501
  for (size_t k = 0; k < nr_parts; k++) {
502
    if (cells_top[ind[k]].nodeID != local_nodeID) {
503
504
505
506
      error("Failed to move all non-local parts to send list");
    }
  }
  for (size_t k = nr_parts; k < s->nr_parts; k++) {
507
    if (cells_top[ind[k]].nodeID == local_nodeID) {
508
      error("Failed to remove local parts from send list");
509
    }
510
511
  }
#endif
512

513
  /* Move non-local gparts to the end of the list. */
514
  for (size_t k = 0; k < nr_gparts;) {
515
516
    if (cells_top[gind[k]].nodeID != local_nodeID) {
      cells_top[gind[k]].gcount -= 1;
517
      nr_gparts -= 1;
Matthieu Schaller's avatar
Bug fix    
Matthieu Schaller committed
518
      const struct gpart tp = s->gparts[k];
519
520
      s->gparts[k] = s->gparts[nr_gparts];
      s->gparts[nr_gparts] = tp;
521
522
      if (s->gparts[k].id_or_neg_offset <= 0) {
        s->parts[-s->gparts[k].id_or_neg_offset].gpart = &s->gparts[k];
523
      }
524
525
526
      if (s->gparts[nr_gparts].id_or_neg_offset <= 0) {
        s->parts[-s->gparts[nr_gparts].id_or_neg_offset].gpart =
            &s->gparts[nr_gparts];
527
      }
Matthieu Schaller's avatar
Bug fix    
Matthieu Schaller committed
528
529
530
      const int t = gind[k];
      gind[k] = gind[nr_gparts];
      gind[nr_gparts] = t;
Matthieu Schaller's avatar
Matthieu Schaller committed
531
    } else {
532
533
534
535
      /* Increment when not exchanging otherwise we need to retest "k".*/
      k++;
    }
  }
536

537
#ifdef SWIFT_DEBUG_CHECKS
538
539
  /* Check that all gparts are in the correct place (untested). */
  for (size_t k = 0; k < nr_gparts; k++) {
540
    if (cells_top[gind[k]].nodeID != local_nodeID) {
541
542
543
544
      error("Failed to move all non-local gparts to send list");
    }
  }
  for (size_t k = nr_gparts; k < s->nr_gparts; k++) {
545
    if (cells_top[gind[k]].nodeID == local_nodeID) {
546
      error("Failed to remove local gparts from send list");
547
    }
548
549
  }
#endif
550

551
552
  /* Exchange the strays, note that this potentially re-allocates
     the parts arrays. */
553
  size_t nr_parts_exchanged = s->nr_parts - nr_parts;
554
  size_t nr_gparts_exchanged = s->nr_gparts - nr_gparts;
Pedro Gonnet's avatar
Pedro Gonnet committed
555
556
557
558
  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. */
559
  s->nr_parts = nr_parts + nr_parts_exchanged;
560
  s->nr_gparts = nr_gparts + nr_gparts_exchanged;
561
562

  /* Re-allocate the index array if needed.. */
563
  if (s->nr_parts > ind_size) {
564
565
    int *ind_new;
    if ((ind_new = (int *)malloc(sizeof(int) * s->nr_parts)) == NULL)
566
      error("Failed to allocate temporary particle indices.");
567
    memcpy(ind_new, ind, sizeof(int) * nr_parts);
568
569
    free(ind);
    ind = ind_new;
570
571
572
  }

  /* Assign each particle to its cell. */
Pedro Gonnet's avatar
Pedro Gonnet committed
573
  for (size_t k = nr_parts; k < s->nr_parts; k++) {
Matthieu Schaller's avatar
Matthieu Schaller committed
574
    const struct part *const p = &s->parts[k];
575
    ind[k] =
576
        cell_getid(cdim, p->x[0] * ih[0], p->x[1] * ih[1], p->x[2] * ih[2]);
577
    cells_top[ind[k]].count += 1;
578
#ifdef SWIFT_DEBUG_CHECKS
579
    if (cells_top[ind[k]].nodeID != local_nodeID)
580
      error("Received part that does not belong to me (nodeID=%i).",
581
            cells_top[ind[k]].nodeID);
582
#endif
583
  }
584
  nr_parts = s->nr_parts;
585
586

#endif /* WITH_MPI */
587
588

  /* Sort the parts according to their cells. */
589
  space_parts_sort(s, ind, nr_parts, 0, s->nr_cells - 1, verbose);
590
591

  /* Re-link the gparts. */
592
  part_relink_gparts(s->parts, nr_parts, 0);
593

594
#ifdef SWIFT_DEBUG_CHECKS
595
  /* Verify space_sort_struct. */
596
597
598
599
600
601
602
603
604
605
  for (size_t k = 1; k < nr_parts; k++) {
    if (ind[k - 1] > ind[k]) {
      error("Sort failed!");
    } else if (ind[k] != cell_getid(cdim, s->parts[k].x[0] * ih[0],
                                    s->parts[k].x[1] * ih[1],
                                    s->parts[k].x[2] * ih[2])) {
      error("Incorrect indices!");
    }
  }
#endif
606
607

  /* We no longer need the indices as of here. */
608
  free(ind);
609

610
611
612
613
#ifdef WITH_MPI

  /* Re-allocate the index array if needed.. */
  if (s->nr_gparts > gind_size) {
614
615
    int *gind_new;
    if ((gind_new = (int *)malloc(sizeof(int) * s->nr_gparts)) == NULL)
616
      error("Failed to allocate temporary g-particle indices.");
617
    memcpy(gind_new, gind, sizeof(int) * nr_gparts);
618
619
620
621
622
    free(gind);
    gind = gind_new;
  }

  /* Assign each particle to its cell. */
623
  for (size_t k = nr_gparts; k < s->nr_gparts; k++) {
Matthieu Schaller's avatar
Matthieu Schaller committed
624
    const struct gpart *const p = &s->gparts[k];
625
626
    gind[k] =
        cell_getid(cdim, p->x[0] * ih[0], p->x[1] * ih[1], p->x[2] * ih[2]);
627
628
629
630
631
632
633
    cells_top[gind[k]].gcount += 1;

#ifdef SWIFT_DEBUG_CHECKS
    if (cells_top[ind[k]].nodeID != s->e->nodeID)
      error("Received part that does not belong to me (nodeID=%i).",
            cells_top[ind[k]].nodeID);
#endif
634
635
636
637
  }
  nr_gparts = s->nr_gparts;

#endif
638

639
  /* Sort the gparts according to their cells. */
Matthieu Schaller's avatar
Matthieu Schaller committed
640
  space_gparts_sort(s, gind, nr_gparts, 0, s->nr_cells - 1, verbose);
641
642

  /* Re-link the parts. */
643
  part_relink_parts(s->gparts, nr_gparts, s->parts);
644
645

  /* We no longer need the indices as of here. */
646
  free(gind);
647

648
#ifdef SWIFT_DEBUG_CHECKS
649
650
651
  /* Verify that the links are correct */
  for (size_t k = 0; k < nr_gparts; ++k) {

652
    if (s->gparts[k].id_or_neg_offset < 0) {
653

654
      const struct part *part = &s->parts[-s->gparts[k].id_or_neg_offset];
655

656
657
658
659
      if (part->gpart != &s->gparts[k]) error("Linking problem !");

      if (s->gparts[k].x[0] != part->x[0] || s->gparts[k].x[1] != part->x[1] ||
          s->gparts[k].x[2] != part->x[2])
660
661
662
663
664
        error("Linked particles are not at the same position !");
    }
  }
  for (size_t k = 0; k < nr_parts; ++k) {

665
    if (s->parts[k].gpart != NULL &&
666
        s->parts[k].gpart->id_or_neg_offset != -(ptrdiff_t)k) {
667
      error("Linking problem !");
668
669
    }
  }
670
#endif
671

672
673
  /* Hook the cells up to the parts. */
  // tic = getticks();
674
675
676
  struct part *finger = s->parts;
  struct xpart *xfinger = s->xparts;
  struct gpart *gfinger = s->gparts;
677
  for (int k = 0; k < s->nr_cells; k++) {
678
    struct cell *restrict c = &cells_top[k];
679
    c->ti_old = ti_current;
680
681
682
683
684
685
686
    c->parts = finger;
    c->xparts = xfinger;
    c->gparts = gfinger;
    finger = &finger[c->count];
    xfinger = &xfinger[c->count];
    gfinger = &gfinger[c->gcount];
  }
687
  // message( "hooking up cells took %.3f %s." ,
Matthieu Schaller's avatar
Matthieu Schaller committed
688
  // clocks_from_ticks(getticks() - tic), clocks_getunit());
689
690
691

  /* At this point, we have the upper-level cells, old or new. Now make
     sure that the parts in each cell are ok. */
692
  space_split(s, cells_top, verbose);
693
694
695
696
697
698
699
700
701

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

/**
 * @brief Split particles between cells of a hierarchy
 *
702
703
 * This is done in parallel using threads in the #threadpool.
 *
704
705
706
707
708
709
 * @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
710
  const ticks tic = getticks();
711

712
  threadpool_map(&s->e->threadpool, space_split_mapper, cells, s->nr_cells,
713
                 sizeof(struct cell), 1, s);
714

715
716
717
  if (verbose)
    message("took %.3f %s.", clocks_from_ticks(getticks() - tic),
            clocks_getunit());
718
}
719

720
/**
721
 * @brief Sort the particles and condensed particles according to the given
722
 * indices.
723
 *
Matthieu Schaller's avatar
Matthieu Schaller committed
724
 * @param s The #space.
725
726
727
728
 * @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.
729
 * @param verbose Are we talkative ?
730
 */
731
void space_parts_sort(struct space *s, int *ind, size_t N, int min, int max,
732
733
                      int verbose) {

Matthieu Schaller's avatar
Matthieu Schaller committed
734
  const ticks tic = getticks();
735

736
737
738
739
740
741
  /* Populate a parallel_sort structure with the input data */
  struct parallel_sort sort_struct;
  sort_struct.parts = s->parts;
  sort_struct.xparts = s->xparts;
  sort_struct.ind = ind;
  sort_struct.stack_size = 2 * (max - min + 1) + 10 + s->e->nr_threads;
742
743
  if ((sort_struct.stack =
           malloc(sizeof(struct qstack) * sort_struct.stack_size)) == NULL)
744
    error("Failed to allocate sorting stack.");
Matthieu Schaller's avatar
Matthieu Schaller committed
745
  for (unsigned int i = 0; i < sort_struct.stack_size; i++)
746
    sort_struct.stack[i].ready = 0;
747

748
  /* Add the first interval. */
749
750
751
752
753
754
755
756
757
758
759
  sort_struct.stack[0].i = 0;
  sort_struct.stack[0].j = N - 1;
  sort_struct.stack[0].min = min;
  sort_struct.stack[0].max = max;
  sort_struct.stack[0].ready = 1;
  sort_struct.first = 0;
  sort_struct.last = 1;
  sort_struct.waiting = 1;

  /* Launch the sorting tasks with a stride of zero such that the same
     map data is passed to each thread. */
760
761
  threadpool_map(&s->e->threadpool, space_parts_sort_mapper, &sort_struct,
                 s->e->threadpool.num_threads, 0, 1, NULL);
762

763
#ifdef SWIFT_DEBUG_CHECKS
764
  /* Verify space_sort_struct. */
765
  for (size_t i = 1; i < N; i++)
766
    if (ind[i - 1] > ind[i])
767
      error("Sorting failed (ind[%zu]=%i,ind[%zu]=%i), min=%i, max=%i.", i - 1,
768
769
770
            ind[i - 1], i, ind[i], min, max);
  message("Sorting succeeded.");
#endif
771

772
  /* Clean up. */
773
  free(sort_struct.stack);
774
775
776
777

  if (verbose)
    message("took %.3f %s.", clocks_from_ticks(getticks() - tic),
            clocks_getunit());
778
}
779

780
781
void space_parts_sort_mapper(void *map_data, int num_elements,
                             void *extra_data) {
782
783
784

  /* Unpack the mapping data. */
  struct parallel_sort *sort_struct = (struct parallel_sort *)map_data;
785

786
  /* Pointers to the sorting data. */
787
788
789
  int *ind = sort_struct->ind;
  struct part *parts = sort_struct->parts;
  struct xpart *xparts = sort_struct->xparts;
790

791
  /* Main loop. */
792
  while (sort_struct->waiting) {
793

794
    /* Grab an interval off the queue. */
795
    int qid = atomic_inc(&sort_struct->first) % sort_struct->stack_size;
796

797
    /* Wait for the entry to be ready, or for the sorting do be done. */
798
799
    while (!sort_struct->stack[qid].ready)
      if (!sort_struct->waiting) return;
800

801
    /* Get the stack entry. */
802
803
804
805
806
    ptrdiff_t i = sort_struct->stack[qid].i;
    ptrdiff_t j = sort_struct->stack[qid].j;
    int min = sort_struct->stack[qid].min;
    int max = sort_struct->stack[qid].max;
    sort_struct->stack[qid].ready = 0;
807

808
809
    /* Loop over sub-intervals. */
    while (1) {
810

811
      /* Bring beer. */
812
      const int pivot = (min + max) / 2;
813
814
      /* message("Working on interval [%i,%i] with min=%i, max=%i, pivot=%i.",
              i, j, min, max, pivot); */
815
816

      /* One pass of QuickSort's partitioning. */
817
818
      ptrdiff_t ii = i;
      ptrdiff_t jj = j;
819
820
821
822
      while (ii < jj) {
        while (ii <= j && ind[ii] <= pivot) ii++;
        while (jj >= i && ind[jj] > pivot) jj--;
        if (ii < jj) {
823
          size_t temp_i = ind[ii];
824
825
          ind[ii] = ind[jj];
          ind[jj] = temp_i;
826
          struct part temp_p = parts[ii];
827
828
          parts[ii] = parts[jj];
          parts[jj] = temp_p;
829
          struct xpart temp_xp = xparts[ii];
830
831
832
833
          xparts[ii] = xparts[jj];
          xparts[jj] = temp_xp;
        }
      }
834

835
#ifdef SWIFT_DEBUG_CHECKS
836
      /* Verify space_sort_struct. */
837
      for (int k = i; k <= jj; k++)
838
        if (ind[k] > pivot) {
839
840
          message("sorting failed at k=%i, ind[k]=%i, pivot=%i, i=%li, j=%li.",
                  k, ind[k], pivot, i, j);
841
842
843
844
          error("Partition failed (<=pivot).");
        }
      for (int k = jj + 1; k <= j; k++)
        if (ind[k] <= pivot) {
845
846
          message("sorting failed at k=%i, ind[k]=%i, pivot=%i, i=%li, j=%li.",
                  k, ind[k], pivot, i, j);
847
          error("Partition failed (>pivot).");
848
849
        }
#endif
850
851
852
853
854
855

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

        /* Recurse on the left? */
        if (jj > i && pivot > min) {
856
          qid = atomic_inc(&sort_struct->last) % sort_struct->stack_size;
857
          while (sort_struct->stack[qid].ready)
858
            ;
859
860
861
862
          sort_struct->stack[qid].i = i;
          sort_struct->stack[qid].j = jj;
          sort_struct->stack[qid].min = min;
          sort_struct->stack[qid].max = pivot;
863
          if (atomic_inc(&sort_struct->waiting) >= sort_struct->stack_size)
864
            error("Qstack overflow.");
865
          sort_struct->stack[qid].ready = 1;
866
        }
867

868
869
870
871
872
873
874
875
876
877
        /* Recurse on the right? */
        if (jj + 1 < j && pivot + 1 < max) {
          i = jj + 1;
          min = pivot + 1;
        } else
          break;

      } else {

        /* Recurse on the right? */
878
        if (pivot + 1 < max) {
879
          qid = atomic_inc(&sort_struct->last) % sort_struct->stack_size;
880
          while (sort_struct->stack[qid].ready)
881
            ;
882
883
884
885
          sort_struct->stack[qid].i = jj + 1;
          sort_struct->stack[qid].j = j;
          sort_struct->stack[qid].min = pivot + 1;
          sort_struct->stack[qid].max = max;
886
          if (atomic_inc(&sort_struct->waiting) >= sort_struct->stack_size)
887
            error("Qstack overflow.");
888
          sort_struct->stack[qid].ready = 1;
889
        }
890

891
892
893
894
895
896
897
        /* Recurse on the left? */
        if (jj > i && pivot > min) {
          j = jj;
          max = pivot;
        } else
          break;
      }
898

899
900
    } /* loop over sub-intervals. */

901
    atomic_dec(&sort_struct->waiting);
902
903

  } /* main loop. */
904
905
}

906
/**
907
 * @brief Sort the g-particles according to the given indices.
908
909
 *
 * @param s The #space.
Matthieu Schaller's avatar
Matthieu Schaller committed
910
911
 * @param ind The indices with respect to which the gparts are sorted.
 * @param N The number of gparts
912
913
914
915
 * @param min Lowest index.
 * @param max highest index.
 * @param verbose Are we talkative ?
 */
Matthieu Schaller's avatar
Matthieu Schaller committed
916
void space_gparts_sort(struct space *s, int *ind, size_t N, int min, int max,
917
918
                       int verbose) {

Matthieu Schaller's avatar
Matthieu Schaller committed
919
  const ticks tic = getticks();