space.c 43.6 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
  const size_t nr_parts = s->nr_parts;
163
  struct cell *restrict c;
164
  ticks tic = getticks();
165
166
167

  /* Run through the parts and get the current h_max. */
  // tic = getticks();
168
  float h_max = s->cell_min / kernel_gamma / space_stretch;
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 {
175
      for (size_t 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
180
181
182
183
184
185
186
187
188
    }
  }

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

  /* Get the new putative cell dimensions. */
196
  int cdim[3];
197
  for (int k = 0; k < 3; k++)
198
199
200
201
202
203
204
205
206
207
208
209
210
211
    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
212

213
214
215
216
217
218
219
  /* 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) {
220
      for (int k = 0; k < s->nr_cells; k++) {
221
222
223
224
225
226
227
228
        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. */
229
    for (int k = 0; k < 3; k++) {
230
231
232
233
      s->cdim[k] = cdim[k];
      s->h[k] = s->dim[k] / cdim[k];
      s->ih[k] = 1.0 / s->h[k];
    }
234
    const float dmin = fminf(s->h[0], fminf(s->h[1], s->h[2]));
235
236
237
238
239
240
241

    /* 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));
242
    for (int k = 0; k < s->nr_cells; k++)
243
244
245
      if (lock_init(&s->cells[k].lock) != 0) error("Failed to init spinlock.");

    /* Set the cell location and sizes. */
246
247
248
    for (int i = 0; i < cdim[0]; i++)
      for (int j = 0; j < cdim[1]; j++)
        for (int k = 0; k < cdim[2]; k++) {
249
250
251
252
253
254
255
256
257
258
259
260
261
          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
262
        }
263
264

    /* Be verbose about the change. */
265
266
267
    if (verbose)
      message("set cell dimensions to [ %i %i %i ].", cdim[0], cdim[1],
              cdim[2]);
268
269
270
    fflush(stdout);

  } /* re-build upper-level cells? */
271
272
  // message( "rebuilding upper-level cells took %.3f %s." ,
  // clocks_from_ticks(double)(getticks() - tic), clocks_getunit());
273
274
275
276
277

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

    /* Free the old cells, if they were allocated. */
278
    for (int k = 0; k < s->nr_cells; k++) {
279
280
281
282
283
284
285
286
287
288
289
      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
290
      s->cells[k].init = NULL;
Matthieu Schaller's avatar
Matthieu Schaller committed
291
      s->cells[k].ghost = NULL;
Matthieu Schaller's avatar
Matthieu Schaller committed
292
293
      s->cells[k].drift = NULL;
      s->cells[k].kick = NULL;
294
      s->cells[k].super = &s->cells[k];
295
    }
296
297
    s->maxdepth = 0;
  }
298
299
300
301

  if (verbose)
    message("took %.3f %s.", clocks_from_ticks(getticks() - tic),
            clocks_getunit());
302
}
303
304
305
306
307
308

/**
 * @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.
309
 * @param verbose Print messages to stdout or not
310
311
 *
 */
312

313
void space_rebuild(struct space *s, double cell_max, int verbose) {
314

Matthieu Schaller's avatar
Matthieu Schaller committed
315
  const ticks tic = getticks();
316
317

  /* Be verbose about this. */
318
  // message("re)building space..."); fflush(stdout);
319
320

  /* Re-grid if necessary, or just re-set the cell data. */
321
  space_regrid(s, cell_max, verbose);
322

Pedro Gonnet's avatar
Pedro Gonnet committed
323
324
  size_t nr_parts = s->nr_parts;
  size_t nr_gparts = s->nr_gparts;
325
326
  struct cell *restrict cells = s->cells;

Matthieu Schaller's avatar
Matthieu Schaller committed
327
328
329
  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]};
330
331
332
333

  /* Run through the particles and get their cell index. */
  // tic = getticks();
  const size_t ind_size = s->size_parts;
334
335
  int *ind;
  if ((ind = (int *)malloc(sizeof(int) * ind_size)) == NULL)
336
    error("Failed to allocate temporary particle indices.");
Pedro Gonnet's avatar
Pedro Gonnet committed
337
  for (size_t k = 0; k < nr_parts; k++) {
338
339
    struct part *restrict p = &s->parts[k];
    for (int j = 0; j < 3; j++)
340
341
342
343
      if (p->x[j] < 0.0)
        p->x[j] += dim[j];
      else if (p->x[j] >= dim[j])
        p->x[j] -= dim[j];
344
    ind[k] =
345
        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
346
    cells[ind[k]].count++;
347
  }
Pedro Gonnet's avatar
Pedro Gonnet committed
348
349
  // message( "getting particle indices took %.3f %s." ,
  // clocks_from_ticks(getticks() - tic), clocks_getunit()):
350

351
352
353
354
355
356
357
  /* 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
358
    struct gpart *restrict gp = &s->gparts[k];
359
360
361
362
363
364
365
366
367
    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
368
// message( "getting particle indices took %.3f %s." ,
369
// clocks_from_ticks(getticks() - tic), clocks_getunit());
370
371
372

#ifdef WITH_MPI
  /* Move non-local parts to the end of the list. */
373
  const int local_nodeID = s->e->nodeID;
Pedro Gonnet's avatar
Pedro Gonnet committed
374
  for (size_t k = 0; k < nr_parts; k++)
375
    if (cells[ind[k]].nodeID != local_nodeID) {
376
377
      cells[ind[k]].count -= 1;
      nr_parts -= 1;
Matthieu Schaller's avatar
Bug fix    
Matthieu Schaller committed
378
      const struct part tp = s->parts[k];
379
380
      s->parts[k] = s->parts[nr_parts];
      s->parts[nr_parts] = tp;
381
382
383
384
385
386
      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
387
      const struct xpart txp = s->xparts[k];
388
389
      s->xparts[k] = s->xparts[nr_parts];
      s->xparts[nr_parts] = txp;
Matthieu Schaller's avatar
Bug fix    
Matthieu Schaller committed
390
      const int t = ind[k];
391
392
      ind[k] = ind[nr_parts];
      ind[nr_parts] = t;
393
394
    }

395
396
  /* 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
397
398
    if (cells[gind[k]].nodeID != local_nodeID) {
      cells[gind[k]].gcount -= 1;
399
      nr_gparts -= 1;
Matthieu Schaller's avatar
Bug fix    
Matthieu Schaller committed
400
      const struct gpart tp = s->gparts[k];
401
402
      s->gparts[k] = s->gparts[nr_gparts];
      s->gparts[nr_gparts] = tp;
403
404
405
406
407
408
      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
409
410
411
      const int t = gind[k];
      gind[k] = gind[nr_gparts];
      gind[nr_gparts] = t;
412
413
    }

414
415
  /* Exchange the strays, note that this potentially re-allocates
     the parts arrays. */
416
417
418
  /* 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;
419
  size_t nr_gparts_exchanged = s->nr_gparts - nr_gparts;
Pedro Gonnet's avatar
Pedro Gonnet committed
420
421
422
  engine_exchange_strays(s->e, nr_parts, &ind[nr_parts], &nr_parts_exchanged,
                         nr_gparts, &gind[nr_gparts], &nr_gparts_exchanged);

423
  /* Add post-processing, i.e. re-linking/creating of gparts here. */
Pedro Gonnet's avatar
Pedro Gonnet committed
424
425

  /* Set the new particle counts. */
426
  s->nr_parts = nr_parts + nr_parts_exchanged;
427
  s->nr_gparts = nr_gparts + nr_gparts_exchanged;
428
429

  /* Re-allocate the index array if needed.. */
430
  if (s->nr_parts > ind_size) {
431
432
    int *ind_new;
    if ((ind_new = (int *)malloc(sizeof(int) * s->nr_parts)) == NULL)
433
      error("Failed to allocate temporary particle indices.");
434
    memcpy(ind_new, ind, sizeof(size_t) * nr_parts);
435
436
    free(ind);
    ind = ind_new;
437
438
439
  }

  /* Assign each particle to its cell. */
Pedro Gonnet's avatar
Pedro Gonnet committed
440
  for (size_t k = nr_parts; k < s->nr_parts; k++) {
Matthieu Schaller's avatar
Matthieu Schaller committed
441
    const struct part *const p = &s->parts[k];
442
    ind[k] =
443
        cell_getid(cdim, p->x[0] * ih[0], p->x[1] * ih[1], p->x[2] * ih[2]);
444
445
446
447
    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 ); */
448
  }
449
  nr_parts = s->nr_parts;
450
451
452
#endif

  /* Sort the parts according to their cells. */
453
  space_parts_sort(s, ind, nr_parts, 0, s->nr_cells - 1, verbose);
454

455
  /* Re-link the gparts. */
Pedro Gonnet's avatar
Pedro Gonnet committed
456
  for (size_t k = 0; k < nr_parts; k++)
457
    if (s->parts[k].gpart != NULL) s->parts[k].gpart->part = &s->parts[k];
458

459
  /* Verify space_sort_struct. */
460
  /* for ( k = 1 ; k < nr_parts ; k++ ) {
461
      if ( ind[k-1] > ind[k] ) {
462
463
          error( "Sort failed!" );
          }
464
      else if ( ind[k] != cell_getid( cdim , parts[k].x[0]*ih[0] ,
465
466
467
468
469
     parts[k].x[1]*ih[1] , parts[k].x[2]*ih[2] ) )
          error( "Incorrect indices!" );
      } */

  /* We no longer need the indices as of here. */
470
  free(ind);
471

Tom Theuns's avatar
Tom Theuns committed
472
#ifdef WITH_MPI
473
474
475

  /* Re-allocate the index array if needed.. */
  if (s->nr_gparts > gind_size) {
476
477
    int *gind_new;
    if ((gind_new = (int *)malloc(sizeof(int) * s->nr_gparts)) == NULL)
478
      error("Failed to allocate temporary g-particle indices.");
479
    memcpy(gind_new, gind, sizeof(int) * nr_gparts);
480
481
482
483
484
    free(gind);
    gind = gind_new;
  }

  /* Assign each particle to its cell. */
485
  for (int k = nr_gparts; k < s->nr_gparts; k++) {
Matthieu Schaller's avatar
Matthieu Schaller committed
486
    const struct gpart *const p = &s->gparts[k];
487
488
    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
489
    cells[gind[k]].gcount += 1;
490
491
492
493
494
495
    /* 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
496
#endif
497
498

  /* Sort the parts according to their cells. */
Matthieu Schaller's avatar
Matthieu Schaller committed
499
  space_gparts_sort(s, gind, nr_gparts, 0, s->nr_cells - 1, verbose);
500
501

  /* Re-link the parts. */
502
  for (int k = 0; k < nr_gparts; k++)
503
    if (s->gparts[k].id > 0) s->gparts[k].part->gpart = &s->gparts[k];
504
505

  /* We no longer need the indices as of here. */
506
  free(gind);
507

508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
  /* 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 !");
    }
  }

530
531
  /* Hook the cells up to the parts. */
  // tic = getticks();
532
533
534
  struct part *finger = s->parts;
  struct xpart *xfinger = s->xparts;
  struct gpart *gfinger = s->gparts;
535
536
  for (int k = 0; k < s->nr_cells; k++) {
    struct cell *restrict c = &cells[k];
537
538
539
540
541
542
543
    c->parts = finger;
    c->xparts = xfinger;
    c->gparts = gfinger;
    finger = &finger[c->count];
    xfinger = &xfinger[c->count];
    gfinger = &gfinger[c->gcount];
  }
544
  // message( "hooking up cells took %.3f %s." ,
Matthieu Schaller's avatar
Matthieu Schaller committed
545
  // clocks_from_ticks(getticks() - tic), clocks_getunit());
546
547
548

  /* At this point, we have the upper-level cells, old or new. Now make
     sure that the parts in each cell are ok. */
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
  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
565
  const ticks tic = getticks();
566
567

  for (int k = 0; k < s->nr_cells; k++)
568
569
    scheduler_addtask(&s->e->sched, task_type_split_cell, task_subtype_none, k,
                      0, &cells[k], NULL, 0);
570
  engine_launch(s->e, s->e->nr_threads, 1 << task_type_split_cell, 0);
571

572
573
574
  if (verbose)
    message("took %.3f %s.", clocks_from_ticks(getticks() - tic),
            clocks_getunit());
575
}
576

577
/**
578
579
 * @brief Sort the particles and condensed particles according to the given
 *indices.
580
 *
Matthieu Schaller's avatar
Matthieu Schaller committed
581
 * @param s The #space.
582
583
584
585
 * @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.
586
 * @param verbose Are we talkative ?
587
 */
588

589
void space_parts_sort(struct space *s, int *ind, size_t N, int min, int max,
590
591
                      int verbose) {

Matthieu Schaller's avatar
Matthieu Schaller committed
592
  const ticks tic = getticks();
593
594

  /*Populate the global parallel_sort structure with the input data */
595
596
597
  space_sort_struct.parts = s->parts;
  space_sort_struct.xparts = s->xparts;
  space_sort_struct.ind = ind;
598
  space_sort_struct.stack_size = 2 * (max - min + 1) + 10 + s->e->nr_threads;
599
600
601
602
603
604
  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;

605
  /* Add the first interval. */
606
607
608
609
610
611
612
613
614
  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;

615
  /* Launch the sorting tasks. */
616
  engine_launch(s->e, s->e->nr_threads, (1 << task_type_part_sort), 0);
617
618

  /* Verify space_sort_struct. */
619
  /* for (int i = 1; i < N; i++)
620
    if (ind[i - 1] > ind[i])
621
622
      error("Sorting failed (ind[%i]=%i,ind[%i]=%i), min=%i, max=%i.", i - 1,
  ind[i - 1], i,
623
624
            ind[i], min, max);
  message("Sorting succeeded."); */
625

626
  /* Clean up. */
627
  free(space_sort_struct.stack);
628
629
630
631

  if (verbose)
    message("took %.3f %s.", clocks_from_ticks(getticks() - tic),
            clocks_getunit());
632
}
633

634
void space_do_parts_sort() {
635

636
  /* Pointers to the sorting data. */
637
  int *ind = space_sort_struct.ind;
638
639
  struct part *parts = space_sort_struct.parts;
  struct xpart *xparts = space_sort_struct.xparts;
640

641
  /* Main loop. */
642
  while (space_sort_struct.waiting) {
643

644
    /* Grab an interval off the queue. */
645
646
    int qid =
        atomic_inc(&space_sort_struct.first) % space_sort_struct.stack_size;
647

648
    /* Wait for the entry to be ready, or for the sorting do be done. */
649
650
    while (!space_sort_struct.stack[qid].ready)
      if (!space_sort_struct.waiting) return;
651

652
    /* Get the stack entry. */
653
654
    ptrdiff_t i = space_sort_struct.stack[qid].i;
    ptrdiff_t j = space_sort_struct.stack[qid].j;
655
656
    int min = space_sort_struct.stack[qid].min;
    int max = space_sort_struct.stack[qid].max;
657
    space_sort_struct.stack[qid].ready = 0;
658

659
660
    /* Loop over sub-intervals. */
    while (1) {
661

662
      /* Bring beer. */
663
      const int pivot = (min + max) / 2;
664
665
      /* message("Working on interval [%i,%i] with min=%i, max=%i, pivot=%i.",
              i, j, min, max, pivot); */
666
667

      /* One pass of QuickSort's partitioning. */
668
669
      ptrdiff_t ii = i;
      ptrdiff_t jj = j;
670
671
672
673
      while (ii < jj) {
        while (ii <= j && ind[ii] <= pivot) ii++;
        while (jj >= i && ind[jj] > pivot) jj--;
        if (ii < jj) {
674
          size_t temp_i = ind[ii];
675
676
          ind[ii] = ind[jj];
          ind[jj] = temp_i;
677
          struct part temp_p = parts[ii];
678
679
          parts[ii] = parts[jj];
          parts[jj] = temp_p;
680
          struct xpart temp_xp = xparts[ii];
681
682
683
684
          xparts[ii] = xparts[jj];
          xparts[jj] = temp_xp;
        }
      }
685

686
      /* Verify space_sort_struct. */
687
688
689
690
691
692
693
694
695
696
697
698
      /* 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).");
        } */
699
700
701
702
703
704

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

        /* Recurse on the left? */
        if (jj > i && pivot > min) {
705
706
          qid = atomic_inc(&space_sort_struct.last) %
                space_sort_struct.stack_size;
707
708
          while (space_sort_struct.stack[qid].ready)
            ;
709
710
711
712
713
714
715
          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.");
716
          space_sort_struct.stack[qid].ready = 1;
717
        }
718

719
720
721
722
723
724
725
726
727
728
        /* Recurse on the right? */
        if (jj + 1 < j && pivot + 1 < max) {
          i = jj + 1;
          min = pivot + 1;
        } else
          break;

      } else {

        /* Recurse on the right? */
729
        if (pivot + 1 < max) {
730
731
          qid = atomic_inc(&space_sort_struct.last) %
                space_sort_struct.stack_size;
732
733
          while (space_sort_struct.stack[qid].ready)
            ;
734
735
736
737
738
739
740
          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.");
741
          space_sort_struct.stack[qid].ready = 1;
742
        }
743

744
745
746
747
748
749
750
        /* Recurse on the left? */
        if (jj > i && pivot > min) {
          j = jj;
          max = pivot;
        } else
          break;
      }
751

752
753
    } /* loop over sub-intervals. */

754
    atomic_dec(&space_sort_struct.waiting);
755
756

  } /* main loop. */
757
758
}

759
760
761
762
763
/**
 * @brief Sort the g-particles and condensed particles according to the given
 *indices.
 *
 * @param s The #space.
Matthieu Schaller's avatar
Matthieu Schaller committed
764
765
 * @param ind The indices with respect to which the gparts are sorted.
 * @param N The number of gparts
766
767
768
769
 * @param min Lowest index.
 * @param max highest index.
 * @param verbose Are we talkative ?
 */
Matthieu Schaller's avatar
Matthieu Schaller committed
770
void space_gparts_sort(struct space *s, int *ind, size_t N, int min, int max,
771
772
                       int verbose) {

Matthieu Schaller's avatar
Matthieu Schaller committed
773
  const ticks tic = getticks();
774
775
776
777
778
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

  /*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
817
  int *ind = space_sort_struct.ind;
818
  struct gpart *gparts = space_sort_struct.gparts;
819

820
  /* Main loop. */
821
  while (space_sort_struct.waiting) {
822

823
    /* Grab an interval off the queue. */
824
825
826
827
828
829
    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;
830

831
    /* Get the stack entry. */
832
833
834
835
836
    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;
837
838
839

    /* Loop over sub-intervals. */
    while (1) {
840

841
      /* Bring beer. */
842
843
844
      const int pivot = (min + max) / 2;
      /* message("Working on interval [%i,%i] with min=%i, max=%i, pivot=%i.",
              i, j, min, max, pivot); */
845
846

      /* One pass of QuickSort's partitioning. */
847
848
      ptrdiff_t ii = i;
      ptrdiff_t jj = j;
849
850
851
852
      while (ii < jj) {
        while (ii <= j && ind[ii] <= pivot) ii++;
        while (jj >= i && ind[jj] > pivot) jj--;
        if (ii < jj) {
853
          size_t temp_i = ind[ii];
854
855
          ind[ii] = ind[jj];
          ind[jj] = temp_i;
856
          struct gpart temp_p = gparts[ii];
857
858
859
860
          gparts[ii] = gparts[jj];
          gparts[jj] = temp_p;
        }
      }
861

862
      /* Verify space_sort_struct. */
863
864
865
866
867
868
869
870
871
872
873
874
      /* 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).");
        } */
875
876
877
878
879
880

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

        /* Recurse on the left? */
        if (jj > i && pivot > min) {
881
882
883
884
885
886
887
888
889
890
891
892
          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;
893
        }
894

895
896
897
898
899
900
        /* Recurse on the right? */
        if (jj + 1 < j && pivot + 1 < max) {
          i = jj + 1;
          min = pivot + 1;
        } else
          break;
901

902
903
904
      } else {

        /* Recurse on the right? */
905
        if (pivot + 1 < max) {
906
907
908
909
910
911
912
913
914
915
916
917
          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;
918
919
920
921
922
923
924
925
926
927
928
929
        }

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

    } /* loop over sub-intervals. */

930
    atomic_dec(&space_sort_struct.waiting);
931
932

  } /* main loop. */
933
}
934

Pedro Gonnet's avatar
Pedro Gonnet committed
935
/**
936
 * @brief Mapping function to free the sorted indices buffers.
Pedro Gonnet's avatar
Pedro Gonnet committed
937
938
 */

939
void space_map_clearsort(struct cell *c, void *data) {
Pedro Gonnet's avatar
Pedro Gonnet committed
940

941
942
943
944
945
  if (c->sort != NULL) {
    free(c->sort);
    c->sort = NULL;
  }
}
Pedro Gonnet's avatar
Pedro Gonnet committed
946

947
948
949
/**
 * @brief Map a function to all particles in a cell recursively.
 *
950
 * @param c The #cell we are working in.
951
952
953
954
 * @param fun Function pointer to apply on the cells.
 * @param data Data passed to the function fun.
 */

Pedro Gonnet's avatar
Pedro Gonnet committed
955
956
957
958
static void rec_map_parts(struct cell *c,
                          void (*fun)(struct part *p, struct cell *c,
                                      void *data),
                          void *data) {
959
960
961
962
963
964

  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
965

966
967
968
969
970
971
  /* 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
972
/**
973
 * @brief Map a function to all particles in a space.
Pedro Gonnet's avatar
Pedro Gonnet committed
974
975
 *
 * @param s The #space we are working in.
976
977
 * @param fun Function pointer to apply on the cells.
 * @param data Data passed to the function fun.
Pedro Gonnet's avatar
Pedro Gonnet committed
978
979
 */

980
981
982
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
983

984
985
  int cid = 0;

986
  /* Call the recursive function on all higher-level cells. */
Pedro Gonnet's avatar
Pedro Gonnet committed
987
988
  for (cid = 0; cid < s->nr_cells; cid++)
    rec_map_parts(&s->cells[cid], fun, data);
989
}
990

991
992
993
994
995
996
997
998
/**
 * @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,
999
1000
                                 void (*fun)(struct part *p, struct xpart *xp,
                                             struct cell *c)) {
For faster browsing, not all history is shown. View entire blame