space.c 39.9 KB
Newer Older
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
/*******************************************************************************
* This file is part of SWIFT.
* Copyright (c) 2012 Pedro Gonnet (pedro.gonnet@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
19
20
21
22
23
24
25
26

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

/* Some standard headers. */
#include <float.h>
#include <limits.h>
#include <math.h>
27
#include <string.h>
28
#include <stdlib.h>
Pedro Gonnet's avatar
Pedro Gonnet committed
29

30
31
/* MPI headers. */
#ifdef WITH_MPI
32
#include <mpi.h>
33
34
#endif

35
36
37
/* This object's header. */
#include "space.h"

Pedro Gonnet's avatar
Pedro Gonnet committed
38
/* Local headers. */
39
#include "atomic.h"
40
#include "engine.h"
41
#include "error.h"
42
43
#include "kernel.h"
#include "lock.h"
44
#include "minmax.h"
45
#include "runner.h"
Pedro Gonnet's avatar
Pedro Gonnet committed
46

47
48
49
/* Shared sort structure. */
struct parallel_sort space_sort_struct;

Pedro Gonnet's avatar
Pedro Gonnet committed
50
51
/* Split size. */
int space_splitsize = space_splitsize_default;
52
int space_subsize = space_subsize_default;
53
int space_maxsize = space_maxsize_default;
Pedro Gonnet's avatar
Pedro Gonnet committed
54
55
56

/* Map shift vector to sortlist. */
const int sortlistID[27] = {
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
    /* ( -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};

85
86
87
88
89
90
91
92
93
94
95
96
/**
 * @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.
 */

97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
int space_getsid(struct space *s, struct cell **ci, struct cell **cj,
                 double *shift) {

  int k, sid = 0, periodic = s->periodic;
  struct cell *temp;
  double dx[3];

  /* Get the relative distance between the pairs, wrapping. */
  for (k = 0; k < 3; k++) {
    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. */
  for (k = 0; k < 3; k++)
    sid = 3 * sid + ((dx[k] < 0.0) ? 0 : ((dx[k] > 0.0) ? 2 : 1));

  /* Switch the cells around? */
  if (runner_flip[sid]) {
    temp = *ci;
    *ci = *cj;
    *cj = temp;
    for (k = 0; k < 3; k++) shift[k] = -shift[k];
  }
  sid = sortlistID[sid];

  /* Return the sort ID. */
  return sid;
}
132

133
/**
134
 * @brief Recursively dismantle a cell tree.
135
136
 *
 */
137
138
139
140
141
142
143
144
145
146
147
148
149
150

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

  int k;

  if (c->split)
    for (k = 0; k < 8; k++)
      if (c->progeny[k] != NULL) {
        space_rebuild_recycle(s, c->progeny[k]);
        space_recycle(s, c->progeny[k]);
        c->progeny[k] = NULL;
      }
}

151
/**
152
 * @brief Re-build the cell grid.
153
 *
154
155
 * @param s The #space.
 * @param cell_max Maximum cell edge length.
156
 * @param verbose Print messages to stdout or not.
157
 */
158

159
void space_regrid(struct space *s, double cell_max, int verbose) {
160
161
162
163

  float h_max = s->cell_min / kernel_gamma / space_stretch, dmin;
  int i, j, k, cdim[3], nr_parts = s->nr_parts;
  struct cell *restrict c;
164
  ticks tic = getticks();
165
166
167
168
169
170

  /* Run through the parts and get the current h_max. */
  // tic = getticks();
  if (s->cells != NULL) {
    for (k = 0; k < s->nr_cells; k++) {
      if (s->cells[k].h_max > h_max) h_max = s->cells[k].h_max;
171
    }
172
173
174
175
176
177
178
179
180
181
182
183
184
185
  } else {
    for (k = 0; k < nr_parts; k++) {
      if (s->parts[k].h > h_max) h_max = s->parts[k].h;
    }
    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)
186
      error("Failed to aggregate the rebuild flag across nodes.");
187
188
189
    h_max = buff;
  }
#endif
190
  if (verbose) message("h_max is %.3e (cell_max=%.3e).", h_max, cell_max);
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
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
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257

  /* Get the new putative cell dimensions. */
  for (k = 0; k < 3; k++)
    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

  /* 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) {
      for (k = 0; k < s->nr_cells; k++) {
        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. */
    for (k = 0; k < 3; k++) {
      s->cdim[k] = cdim[k];
      s->h[k] = s->dim[k] / cdim[k];
      s->ih[k] = 1.0 / s->h[k];
    }
    dmin = fminf(s->h[0], fminf(s->h[1], s->h[2]));

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

    /* Set the cell location and sizes. */
    for (i = 0; i < cdim[0]; i++)
      for (j = 0; j < cdim[1]; j++)
        for (k = 0; k < cdim[2]; k++) {
          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
258
        }
259
260

    /* Be verbose about the change. */
261
262
263
    if (verbose)
      message("set cell dimensions to [ %i %i %i ].", cdim[0], cdim[1],
              cdim[2]);
264
265
266
    fflush(stdout);

  } /* re-build upper-level cells? */
267
268
  // message( "rebuilding upper-level cells took %.3f %s." ,
  // clocks_from_ticks(double)(getticks() - tic), clocks_getunit());
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285

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

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

  if (verbose)
    message("took %.3f %s.", clocks_from_ticks(getticks() - tic),
            clocks_getunit());
298
}
299
300
301
302
303
304

/**
 * @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.
305
 * @param verbose Print messages to stdout or not
306
307
 *
 */
308

309
void space_rebuild(struct space *s, double cell_max, int verbose) {
310

311
  ticks tic = getticks();
312
313
314
315
316

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

  /* Re-grid if necessary, or just re-set the cell data. */
317
  space_regrid(s, cell_max, verbose);
318

319
320
321
322
323
324
  int nr_parts = s->nr_parts;
  int nr_gparts = s->nr_gparts;
  struct cell *restrict cells = s->cells;

  double ih[3], dim[3];
  int cdim[3];
325
326
327
328
329
330
331
332
333
  ih[0] = s->ih[0];
  ih[1] = s->ih[1];
  ih[2] = s->ih[2];
  dim[0] = s->dim[0];
  dim[1] = s->dim[1];
  dim[2] = s->dim[2];
  cdim[0] = s->cdim[0];
  cdim[1] = s->cdim[1];
  cdim[2] = s->cdim[2];
334
335
336
337
338
339
340
341
342
343

  /* Run through the particles and get their cell index. */
  // tic = getticks();
  const size_t ind_size = s->size_parts;
  size_t *ind;
  if ((ind = (size_t *)malloc(sizeof(size_t) * ind_size)) == NULL)
    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++)
344
345
346
347
      if (p->x[j] < 0.0)
        p->x[j] += dim[j];
      else if (p->x[j] >= dim[j])
        p->x[j] -= dim[j];
348
    ind[k] =
349
        cell_getid(cdim, p->x[0] * ih[0], p->x[1] * ih[1], p->x[2] * ih[2]);
350
    cells[ind[k]].count++;
351
  }
Matthieu Schaller's avatar
Matthieu Schaller committed
352
353
// message( "getting particle indices took %.3f %s." ,
// clocks_from_ticks(getticks() - tic), clocks_getunit()):
354
355
356

#ifdef WITH_MPI
  /* Move non-local parts to the end of the list. */
357
358
  const int nodeID = s->e->nodeID;
  for (int k = 0; k < nr_parts; k++)
359
360
361
    if (cells[ind[k]].nodeID != nodeID) {
      cells[ind[k]].count -= 1;
      nr_parts -= 1;
362
363
364
365
366
367
      struct part tp = s->parts[k];
      s->parts[k] = s->parts[nr_parts];
      s->parts[nr_parts] = tp;
      struct xpart txp = s->xparts[k];
      s->xparts[k] = s->xparts[nr_parts];
      s->xparts[nr_parts] = txp;
368
369
370
      int t = ind[k];
      ind[k] = ind[nr_parts];
      ind[nr_parts] = t;
371
372
    }

373
374
375
  /* Exchange the strays, note that this potentially re-allocates
     the parts arrays. */
  s->nr_parts =
376
377
      nr_parts + engine_exchange_strays(s->e, nr_parts, &ind[nr_parts],
                                        s->nr_parts - nr_parts);
378
379

  /* Re-allocate the index array if needed.. */
380
  if (s->nr_parts > ind_size) {
381
382
    size_t *ind_new;
    if ((ind_new = (size_t *)malloc(sizeof(size_t) * s->nr_parts)) == NULL)
383
      error("Failed to allocate temporary particle indices.");
384
    memcpy(ind_new, ind, sizeof(size_t) * nr_parts);
385
386
    free(ind);
    ind = ind_new;
387
388
389
  }

  /* Assign each particle to its cell. */
390
391
  for (int k = nr_parts; k < s->nr_parts; k++) {
    struct part *p = &s->parts[k];
392
    ind[k] =
393
        cell_getid(cdim, p->x[0] * ih[0], p->x[1] * ih[1], p->x[2] * ih[2]);
394
395
396
397
    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 ); */
398
  }
399
  nr_parts = s->nr_parts;
400
401
402
#endif

  /* Sort the parts according to their cells. */
403
  space_parts_sort(s, ind, nr_parts, 0, s->nr_cells - 1, verbose);
404
405

  /* Re-link the gparts. */
406
  for (int k = 0; k < nr_parts; k++)
407
    if (s->parts[k].gpart != NULL) s->parts[k].gpart->part = &s->parts[k];
408

409
  /* Verify space_sort_struct. */
410
  /* for ( k = 1 ; k < nr_parts ; k++ ) {
411
      if ( ind[k-1] > ind[k] ) {
412
413
          error( "Sort failed!" );
          }
414
      else if ( ind[k] != cell_getid( cdim , parts[k].x[0]*ih[0] ,
415
416
417
418
419
     parts[k].x[1]*ih[1] , parts[k].x[2]*ih[2] ) )
          error( "Incorrect indices!" );
      } */

  /* We no longer need the indices as of here. */
420
  free(ind);
421
422
423

  /* Run through the gravity particles and get their cell index. */
  // tic = getticks();
424
425
426
427
428
  const size_t gind_size = s->size_gparts;
  size_t *gind;
  if ((gind = (size_t *)malloc(sizeof(size_t) * gind_size)) == NULL)
    error("Failed to allocate temporary g-particle indices.");
  for (int k = 0; k < nr_gparts; k++) {
429
    struct gpart *gp = &s->gparts[k];
430
    for (int j = 0; j < 3; j++)
431
432
433
434
      if (gp->x[j] < 0.0)
        gp->x[j] += dim[j];
      else if (gp->x[j] >= dim[j])
        gp->x[j] -= dim[j];
435
    gind[k] =
436
        cell_getid(cdim, gp->x[0] * ih[0], gp->x[1] * ih[1], gp->x[2] * ih[2]);
437
    cells[gind[k]].gcount++;
438
  }
439
440
// message( "getting particle indices took %.3f %s." ,
// clocks_from_ticks(getticks() - tic), clocks_getunit());
441

442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
#ifdef WITH_MPI

  /* Move non-local gparts to the end of the list. */
  for (int k = 0; k < nr_gparts; k++)
    if (cells[ind[k]].nodeID != nodeID) {
      cells[ind[k]].gcount -= 1;
      nr_gparts -= 1;
      struct gpart tp = s->gparts[k];
      s->gparts[k] = s->gparts[nr_gparts];
      s->gparts[nr_gparts] = tp;
      int t = ind[k];
      ind[k] = ind[nr_gparts];
      ind[nr_gparts] = t;
    }

  /* Exchange the strays, note that this potentially re-allocates
     the parts arrays. */
Matthieu Schaller's avatar
Matthieu Schaller committed
459
  // s->nr_gparts =
460
461
  //    nr_gparts + engine_exchange_strays(s->e, nr_gparts, &ind[nr_gparts],
  //                                        s->nr_gparts - nr_gparts);
Matthieu Schaller's avatar
Matthieu Schaller committed
462
  if (nr_gparts > 0)
463
464
    error("Need to implement the exchange of strays for the gparts");

465
466
467
468
469
470
471
472
473
474
475
  /* Re-allocate the index array if needed.. */
  if (s->nr_gparts > gind_size) {
    size_t *gind_new;
    if ((gind_new = (size_t *)malloc(sizeof(size_t) * s->nr_gparts)) == NULL)
      error("Failed to allocate temporary g-particle indices.");
    memcpy(gind_new, gind, sizeof(size_t) * nr_gparts);
    free(gind);
    gind = gind_new;
  }

  /* Assign each particle to its cell. */
476
  for (int k = nr_gparts; k < s->nr_gparts; k++) {
477
478
479
480
481
482
483
484
485
486
487
    struct gpart *p = &s->gparts[k];
    gind[k] =
        cell_getid(cdim, p->x[0] * ih[0], p->x[1] * ih[1], p->x[2] * ih[2]);
    cells[gind[k]].count += 1;
    /* 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;

#endif
488
489

  /* Sort the parts according to their cells. */
490
  space_parts_sort(s, ind, nr_gparts, 0, s->nr_cells - 1, verbose);
491
492

  /* Re-link the parts. */
493
  for (int k = 0; k < nr_gparts; k++)
494
    if (s->gparts[k].id > 0) s->gparts[k].part->gpart = &s->gparts[k];
495
496

  /* We no longer need the indices as of here. */
497
  free(gind);
498
499
500

  /* Hook the cells up to the parts. */
  // tic = getticks();
501
502
503
  struct part *finger = s->parts;
  struct xpart *xfinger = s->xparts;
  struct gpart *gfinger = s->gparts;
504
505
  for (int k = 0; k < s->nr_cells; k++) {
    struct cell *restrict c = &cells[k];
506
507
508
509
510
511
512
    c->parts = finger;
    c->xparts = xfinger;
    c->gparts = gfinger;
    finger = &finger[c->count];
    xfinger = &xfinger[c->count];
    gfinger = &gfinger[c->gcount];
  }
513
  // message( "hooking up cells took %.3f %s." ,
Matthieu Schaller's avatar
Matthieu Schaller committed
514
  // clocks_from_ticks(getticks() - tic), clocks_getunit());
515
516
517

  /* At this point, we have the upper-level cells, old or new. Now make
     sure that the parts in each cell are ok. */
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
  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) {

  ticks tic = getticks();

  for (int k = 0; k < s->nr_cells; k++)
537
538
    scheduler_addtask(&s->e->sched, task_type_split_cell, task_subtype_none, k,
                      0, &cells[k], NULL, 0);
539
  engine_launch(s->e, s->e->nr_threads, 1 << task_type_split_cell, 0);
540

541
542
543
  if (verbose)
    message("took %.3f %s.", clocks_from_ticks(getticks() - tic),
            clocks_getunit());
544
}
545

546
/**
547
548
 * @brief Sort the particles and condensed particles according to the given
 *indices.
549
 *
Matthieu Schaller's avatar
Matthieu Schaller committed
550
 * @param s The #space.
551
552
553
554
 * @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.
555
 * @param verbose Are we talkative ?
556
 */
557
void space_parts_sort(struct space *s, size_t *ind, size_t N, int min, int max,
558
559
560
561
562
                      int verbose) {

  ticks tic = getticks();

  /*Populate the global parallel_sort structure with the input data */
563
564
565
  space_sort_struct.parts = s->parts;
  space_sort_struct.xparts = s->xparts;
  space_sort_struct.ind = ind;
566
  space_sort_struct.stack_size = 2 * (max - min + 1) + 10 + s->e->nr_threads;
567
568
569
570
571
572
  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;

573
  /* Add the first interval. */
574
575
576
577
578
579
580
581
582
  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;

583
  /* Launch the sorting tasks. */
584
  engine_launch(s->e, s->e->nr_threads, (1 << task_type_part_sort), 0);
585
586

  /* Verify space_sort_struct. */
587
  /* for (int i = 1; i < N; i++)
588
    if (ind[i - 1] > ind[i])
589
590
      error("Sorting failed (ind[%i]=%i,ind[%i]=%i), min=%i, max=%i.", i - 1,
  ind[i - 1], i,
591
592
            ind[i], min, max);
  message("Sorting succeeded."); */
593

594
  /* Clean up. */
595
  free(space_sort_struct.stack);
596
597
598
599

  if (verbose)
    message("took %.3f %s.", clocks_from_ticks(getticks() - tic),
            clocks_getunit());
600
}
601

602
void space_do_parts_sort() {
603

604
  /* Pointers to the sorting data. */
605
  size_t *ind = space_sort_struct.ind;
606
607
  struct part *parts = space_sort_struct.parts;
  struct xpart *xparts = space_sort_struct.xparts;
608

609
  /* Main loop. */
610
  while (space_sort_struct.waiting) {
611

612
    /* Grab an interval off the queue. */
613
614
    int qid =
        atomic_inc(&space_sort_struct.first) % space_sort_struct.stack_size;
615

616
    /* Wait for the entry to be ready, or for the sorting do be done. */
617
618
    while (!space_sort_struct.stack[qid].ready)
      if (!space_sort_struct.waiting) return;
619

620
    /* Get the stack entry. */
621
622
    ptrdiff_t i = space_sort_struct.stack[qid].i;
    ptrdiff_t j = space_sort_struct.stack[qid].j;
623
624
    int min = space_sort_struct.stack[qid].min;
    int max = space_sort_struct.stack[qid].max;
625
    space_sort_struct.stack[qid].ready = 0;
626

627
628
    /* Loop over sub-intervals. */
    while (1) {
629

630
      /* Bring beer. */
631
      const int pivot = (min + max) / 2;
632
633
      /* message("Working on interval [%i,%i] with min=%i, max=%i, pivot=%i.",
              i, j, min, max, pivot); */
634
635

      /* One pass of QuickSort's partitioning. */
636
637
      ptrdiff_t ii = i;
      ptrdiff_t jj = j;
638
639
640
641
      while (ii < jj) {
        while (ii <= j && ind[ii] <= pivot) ii++;
        while (jj >= i && ind[jj] > pivot) jj--;
        if (ii < jj) {
642
          size_t temp_i = ind[ii];
643
644
          ind[ii] = ind[jj];
          ind[jj] = temp_i;
645
          struct part temp_p = parts[ii];
646
647
          parts[ii] = parts[jj];
          parts[jj] = temp_p;
648
          struct xpart temp_xp = xparts[ii];
649
650
651
652
          xparts[ii] = xparts[jj];
          xparts[jj] = temp_xp;
        }
      }
653

654
      /* Verify space_sort_struct. */
655
656
657
658
659
660
661
662
663
664
665
666
      /* 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).");
        } */
667
668
669
670
671
672

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

        /* Recurse on the left? */
        if (jj > i && pivot > min) {
673
674
          qid = atomic_inc(&space_sort_struct.last) %
                space_sort_struct.stack_size;
675
676
          while (space_sort_struct.stack[qid].ready)
            ;
677
678
679
680
681
682
683
          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.");
684
          space_sort_struct.stack[qid].ready = 1;
685
        }
686

687
688
689
690
691
692
693
694
695
696
        /* Recurse on the right? */
        if (jj + 1 < j && pivot + 1 < max) {
          i = jj + 1;
          min = pivot + 1;
        } else
          break;

      } else {

        /* Recurse on the right? */
697
        if (pivot + 1 < max) {
698
699
          qid = atomic_inc(&space_sort_struct.last) %
                space_sort_struct.stack_size;
700
701
          while (space_sort_struct.stack[qid].ready)
            ;
702
703
704
705
706
707
708
          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.");
709
          space_sort_struct.stack[qid].ready = 1;
710
        }
711

712
713
714
715
716
717
718
        /* Recurse on the left? */
        if (jj > i && pivot > min) {
          j = jj;
          max = pivot;
        } else
          break;
      }
719

720
721
    } /* loop over sub-intervals. */

722
    atomic_dec(&space_sort_struct.waiting);
723
724

  } /* main loop. */
725
726
}

727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
/**
 * @brief Sort the g-particles and condensed particles according to the given
 *indices.
 *
 * @param s The #space.
 * @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.
 * @param verbose Are we talkative ?
 */
void space_gparts_sort(struct space *s, size_t *ind, size_t N, int min, int max,
                       int verbose) {

  ticks tic = getticks();

  /*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. */
  size_t *ind = space_sort_struct.ind;
  struct gpart *gparts = space_sort_struct.gparts;
787

788
  /* Main loop. */
789
  while (space_sort_struct.waiting) {
790

791
    /* Grab an interval off the queue. */
792
793
794
795
796
797
    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;
798

799
    /* Get the stack entry. */
800
801
802
803
804
    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;
805
806
807

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

809
      /* Bring beer. */
810
811
812
      const int pivot = (min + max) / 2;
      /* message("Working on interval [%i,%i] with min=%i, max=%i, pivot=%i.",
              i, j, min, max, pivot); */
813
814

      /* One pass of QuickSort's partitioning. */
815
816
      ptrdiff_t ii = i;
      ptrdiff_t jj = j;
817
818
819
820
      while (ii < jj) {
        while (ii <= j && ind[ii] <= pivot) ii++;
        while (jj >= i && ind[jj] > pivot) jj--;
        if (ii < jj) {
821
          size_t temp_i = ind[ii];
822
823
          ind[ii] = ind[jj];
          ind[jj] = temp_i;
824
          struct gpart temp_p = gparts[ii];
825
826
827
828
          gparts[ii] = gparts[jj];
          gparts[jj] = temp_p;
        }
      }
829

830
      /* Verify space_sort_struct. */
831
832
833
834
835
836
837
838
839
840
841
842
      /* 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).");
        } */
843
844
845
846
847
848

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

        /* Recurse on the left? */
        if (jj > i && pivot > min) {
849
850
851
852
853
854
855
856
857
858
859
860
          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;
861
        }
862

863
864
865
866
867
868
        /* Recurse on the right? */
        if (jj + 1 < j && pivot + 1 < max) {
          i = jj + 1;
          min = pivot + 1;
        } else
          break;
869

870
871
872
      } else {

        /* Recurse on the right? */
873
        if (pivot + 1 < max) {
874
875
876
877
878
879
880
881
882
883
884
885
          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;
886
887
888
889
890
891
892
893
894
895
896
897
        }

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

    } /* loop over sub-intervals. */

898
    atomic_dec(&space_sort_struct.waiting);
899
900

  } /* main loop. */
901
}
902

Pedro Gonnet's avatar
Pedro Gonnet committed
903
/**
904
 * @brief Mapping function to free the sorted indices buffers.
Pedro Gonnet's avatar
Pedro Gonnet committed
905
906
 */

907
void space_map_clearsort(struct cell *c, void *data) {
Pedro Gonnet's avatar
Pedro Gonnet committed
908

909
910
911
912
913
  if (c->sort != NULL) {
    free(c->sort);
    c->sort = NULL;
  }
}
Pedro Gonnet's avatar
Pedro Gonnet committed
914

915
916
917
/**
 * @brief Map a function to all particles in a cell recursively.
 *
918
 * @param c The #cell we are working in.
919
920
921
922
 * @param fun Function pointer to apply on the cells.
 * @param data Data passed to the function fun.
 */

Pedro Gonnet's avatar
Pedro Gonnet committed
923
924
925
926
static void rec_map_parts(struct cell *c,
                          void (*fun)(struct part *p, struct cell *c,
                                      void *data),
                          void *data) {
927
928
929
930
931
932

  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
933

934
935
936
937
938
939
  /* 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
940
/**
941
 * @brief Map a function to all particles in a space.
Pedro Gonnet's avatar
Pedro Gonnet committed
942
943
 *
 * @param s The #space we are working in.
944
945
 * @param fun Function pointer to apply on the cells.
 * @param data Data passed to the function fun.
Pedro Gonnet's avatar
Pedro Gonnet committed
946
947
 */

948
949
950
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
951

952
953
  int cid = 0;

954
  /* Call the recursive function on all higher-level cells. */
Pedro Gonnet's avatar
Pedro Gonnet committed
955
956
  for (cid = 0; cid < s->nr_cells; cid++)
    rec_map_parts(&s->cells[cid], fun, data);
957
}
958

959
960
961
962
963
964
965
966
/**
 * @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,
967
968
                                 void (*fun)(struct part *p, struct xpart *xp,
                                             struct cell *c)) {
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989

  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,
990
991
                            void (*fun)(struct part *p, struct xpart *xp,
                                        struct cell *c)) {
992
993
994
995
996
997
998
999

  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);
}

1000
/**
For faster browsing, not all history is shown. View entire blame