scheduler.c 47.1 KB
Newer Older
1
2
/*******************************************************************************
 * This file is part of SWIFT.
3
 * Copyright (c) 2012 Pedro Gonnet (pedro.gonnet@durham.ac.uk)
4
 *                    Matthieu Schaller (matthieu.schaller@durham.ac.uk)
5
 *               2016 Peter W. Draper (p.w.draper@durham.ac.uk)
6
 *
7
8
9
10
 * 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.
11
 *
12
13
14
15
 * 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.
16
 *
17
18
 * 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/>.
19
 *
20
21
22
23
24
25
 ******************************************************************************/

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

/* Some standard headers. */
26
27
28
#include <limits.h>
#include <math.h>
#include <pthread.h>
29
30
31
32
#include <stdio.h>
#include <stdlib.h>
#include <string.h>

33
34
/* MPI headers. */
#ifdef WITH_MPI
35
#include <mpi.h>
36
37
#endif

38
39
40
/* This object's header. */
#include "scheduler.h"

41
42
43
/* Local headers. */
#include "atomic.h"
#include "const.h"
44
45
#include "cycle.h"
#include "error.h"
46
#include "intrinsics.h"
47
#include "kernel_hydro.h"
48
49
50
#include "queue.h"
#include "space.h"
#include "task.h"
51
#include "timers.h"
52

53
54
55
56
57
58
59
60
/**
 * @brief Add an unlock_task to the given task.
 *
 * @param s The #scheduler.
 * @param ta The unlocking #task.
 * @param tb The #task that will be unlocked.
 */

61
62
63
64
65
66
67
void scheduler_addunlock(struct scheduler *s, struct task *ta,
                         struct task *tb) {
  /* Get an index at which to store this unlock. */
  const int ind = atomic_inc(&s->nr_unlocks);

  /* Does the buffer need to be grown? */
  if (ind == s->size_unlocks) {
68
    /* Allocate the new buffer. */
69
70
71
    struct task **unlocks_new;
    int *unlock_ind_new;
    const int size_unlocks_new = s->size_unlocks * 2;
72
    if ((unlocks_new = (struct task **)malloc(sizeof(struct task *) *
73
                                              size_unlocks_new)) == NULL ||
74
75
        (unlock_ind_new = (int *)malloc(sizeof(int) * size_unlocks_new)) ==
            NULL)
76
      error("Failed to re-allocate unlocks.");
77

78
    /* Wait for all writes to the old buffer to complete. */
79
80
81
    while (s->completed_unlock_writes < ind)
      ;

82
    /* Copy the buffers. */
83
84
85
86
87
88
    memcpy(unlocks_new, s->unlocks, sizeof(struct task *) * ind);
    memcpy(unlock_ind_new, s->unlock_ind, sizeof(int) * ind);
    free(s->unlocks);
    free(s->unlock_ind);
    s->unlocks = unlocks_new;
    s->unlock_ind = unlock_ind_new;
89

90
    /* Publish the new buffer size. */
91
92
    s->size_unlocks = size_unlocks_new;
  }
93

94
  /* Wait for there to actually be space at my index. */
95
96
  while (ind > s->size_unlocks)
    ;
97
98
99
100

  /* Write the unlock to the scheduler. */
  s->unlocks[ind] = tb;
  s->unlock_ind[ind] = ta - s->tasks;
101
  atomic_inc(&s->completed_unlock_writes);
102
103
}

104
/**
105
 * @brief Split a task if too large.
106
 *
107
108
 * @param t The #task
 * @param s The #scheduler we are working in.
109
 */
110

111
static void scheduler_splittask(struct task *t, struct scheduler *s) {
112
113

  /* Static constants. */
114
  static const int pts[7][8] = {
Peter W. Draper's avatar
Peter W. Draper committed
115
116
117
118
      {-1, 12, 10, 9, 4, 3, 1, 0},     {-1, -1, 11, 10, 5, 4, 2, 1},
      {-1, -1, -1, 12, 7, 6, 4, 3},    {-1, -1, -1, -1, 8, 7, 5, 4},
      {-1, -1, -1, -1, -1, 12, 10, 9}, {-1, -1, -1, -1, -1, -1, 11, 10},
      {-1, -1, -1, -1, -1, -1, -1, 12}};
Matthieu Schaller's avatar
Matthieu Schaller committed
119
120
121
  static const float sid_scale[13] = {
      0.1897f, 0.4025f, 0.1897f, 0.4025f, 0.5788f, 0.4025f, 0.1897f,
      0.4025f, 0.1897f, 0.4025f, 0.5788f, 0.4025f, 0.5788f};
122

123
124
125
  /* Iterate on this task until we're done with it. */
  int redo = 1;
  while (redo) {
126

127
128
    /* Reset the redo flag. */
    redo = 0;
129

130
131
132
133
134
135
136
137
    /* Non-splittable task? */
    if ((t->ci == NULL || (t->type == task_type_pair && t->cj == NULL)) ||
        ((t->type == task_type_kick) && t->ci->nodeID != s->nodeID) ||
        ((t->type == task_type_init) && t->ci->nodeID != s->nodeID)) {
      t->type = task_type_none;
      t->skip = 1;
      break;
    }
138

139
140
    /* Self-interaction? */
    if (t->type == task_type_self) {
141

142
143
144
145
146
      /* Get a handle on the cell involved. */
      struct cell *ci = t->ci;

      /* Foreign task? */
      if (ci->nodeID != s->nodeID) {
147
        t->skip = 1;
148
        break;
149
150
      }

151
152
      /* Is this cell even split? */
      if (ci->split) {
153

154
155
156
        /* Make a sub? */
        if (scheduler_dosub && (ci->count * ci->count < space_subsize ||
                                ci->gcount * ci->gcount < space_subsize)) {
157

158
159
160
161
162
          /* convert to a self-subtask. */
          t->type = task_type_sub_self;

          /* Otherwise, make tasks explicitly. */
        } else {
163

164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
          /* Take a step back (we're going to recycle the current task)... */
          redo = 1;

          /* Add the self tasks. */
          int first_child = 0;
          while (ci->progeny[first_child] == NULL) first_child++;
          t->ci = ci->progeny[first_child];
          for (int k = first_child + 1; k < 8; k++)
            if (ci->progeny[k] != NULL)
              scheduler_splittask(
                  scheduler_addtask(s, task_type_self, t->subtype, 0, 0,
                                    ci->progeny[k], NULL, 0),
                  s);

          /* Make a task for each pair of progeny. */
          for (int j = 0; j < 8; j++)
            if (ci->progeny[j] != NULL)
              for (int k = j + 1; k < 8; k++)
                if (ci->progeny[k] != NULL)
                  scheduler_splittask(
                      scheduler_addtask(s, task_type_pair, t->subtype,
                                        pts[j][k], 0, ci->progeny[j],
                                        ci->progeny[k], 0),
                      s);
188
        }
189
      }
190

191
192
      /* Pair interaction? */
    } else if (t->type == task_type_pair && t->subtype != task_subtype_grav) {
193

194
195
196
197
198
      /* Get a handle on the cells involved. */
      struct cell *ci = t->ci;
      struct cell *cj = t->cj;
      const double hi = ci->dmin;
      const double hj = cj->dmin;
199

200
201
202
203
204
      /* Foreign task? */
      if (ci->nodeID != s->nodeID && cj->nodeID != s->nodeID) {
        t->skip = 1;
        break;
      }
205

206
207
208
209
      /* Get the sort ID, use space_getsid and not t->flags
         to make sure we get ci and cj swapped if needed. */
      double shift[3];
      int sid = space_getsid(s->space, &ci, &cj, shift);
210

211
212
213
214
215
216
217
      /* Should this task be split-up? */
      if (ci->split && cj->split &&
          ci->h_max * kernel_gamma * space_stretch < hi / 2 &&
          cj->h_max * kernel_gamma * space_stretch < hj / 2) {

        /* Replace by a single sub-task? */
        if (scheduler_dosub &&
218
            ci->count * sid_scale[sid] < space_subsize / cj->count &&
219
220
221
222
223
224
            sid != 0 && sid != 2 && sid != 6 && sid != 8) {

          /* Make this task a sub task. */
          t->type = task_type_sub_pair;

          /* Otherwise, split it. */
225
226
        } else {

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
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
          /* Take a step back (we're going to recycle the current task)... */
          redo = 1;

          /* For each different sorting type... */
          switch (sid) {

            case 0: /* (  1 ,  1 ,  1 ) */
              t->ci = ci->progeny[7];
              t->cj = cj->progeny[0];
              t->flags = 0;
              break;

            case 1: /* (  1 ,  1 ,  0 ) */
              t->ci = ci->progeny[6];
              t->cj = cj->progeny[0];
              t->flags = 1;
              t->tight = 1;
              scheduler_splittask(
                  scheduler_addtask(s, task_type_pair, t->subtype, 1, 0,
                                    ci->progeny[7], cj->progeny[1], 1),
                  s);
              scheduler_splittask(
                  scheduler_addtask(s, task_type_pair, t->subtype, 0, 0,
                                    ci->progeny[6], cj->progeny[1], 1),
                  s);
              scheduler_splittask(
                  scheduler_addtask(s, task_type_pair, t->subtype, 2, 0,
                                    ci->progeny[7], cj->progeny[0], 1),
                  s);
              break;

            case 2: /* (  1 ,  1 , -1 ) */
              t->ci = ci->progeny[6];
              t->cj = cj->progeny[1];
              t->flags = 2;
              t->tight = 1;
              break;

            case 3: /* (  1 ,  0 ,  1 ) */
              t->ci = ci->progeny[5];
              t->cj = cj->progeny[0];
              t->flags = 3;
              t->tight = 1;
              scheduler_splittask(
                  scheduler_addtask(s, task_type_pair, t->subtype, 3, 0,
                                    ci->progeny[7], cj->progeny[2], 1),
                  s);
              scheduler_splittask(
                  scheduler_addtask(s, task_type_pair, t->subtype, 0, 0,
                                    ci->progeny[5], cj->progeny[2], 1),
                  s);
              scheduler_splittask(
                  scheduler_addtask(s, task_type_pair, t->subtype, 6, 0,
                                    ci->progeny[7], cj->progeny[0], 1),
                  s);
              break;

            case 4: /* (  1 ,  0 ,  0 ) */
              t->ci = ci->progeny[4];
              t->cj = cj->progeny[0];
              t->flags = 4;
              t->tight = 1;
              scheduler_splittask(
                  scheduler_addtask(s, task_type_pair, t->subtype, 5, 0,
                                    ci->progeny[5], cj->progeny[0], 1),
                  s);
              scheduler_splittask(
                  scheduler_addtask(s, task_type_pair, t->subtype, 7, 0,
                                    ci->progeny[6], cj->progeny[0], 1),
                  s);
              scheduler_splittask(
                  scheduler_addtask(s, task_type_pair, t->subtype, 8, 0,
                                    ci->progeny[7], cj->progeny[0], 1),
                  s);
              scheduler_splittask(
                  scheduler_addtask(s, task_type_pair, t->subtype, 3, 0,
                                    ci->progeny[4], cj->progeny[1], 1),
                  s);
              scheduler_splittask(
                  scheduler_addtask(s, task_type_pair, t->subtype, 4, 0,
                                    ci->progeny[5], cj->progeny[1], 1),
                  s);
              scheduler_splittask(
                  scheduler_addtask(s, task_type_pair, t->subtype, 6, 0,
                                    ci->progeny[6], cj->progeny[1], 1),
                  s);
              scheduler_splittask(
                  scheduler_addtask(s, task_type_pair, t->subtype, 7, 0,
                                    ci->progeny[7], cj->progeny[1], 1),
                  s);
              scheduler_splittask(
                  scheduler_addtask(s, task_type_pair, t->subtype, 1, 0,
                                    ci->progeny[4], cj->progeny[2], 1),
                  s);
              scheduler_splittask(
                  scheduler_addtask(s, task_type_pair, t->subtype, 2, 0,
                                    ci->progeny[5], cj->progeny[2], 1),
                  s);
              scheduler_splittask(
                  scheduler_addtask(s, task_type_pair, t->subtype, 4, 0,
                                    ci->progeny[6], cj->progeny[2], 1),
                  s);
              scheduler_splittask(
                  scheduler_addtask(s, task_type_pair, t->subtype, 5, 0,
                                    ci->progeny[7], cj->progeny[2], 1),
                  s);
              scheduler_splittask(
                  scheduler_addtask(s, task_type_pair, t->subtype, 0, 0,
                                    ci->progeny[4], cj->progeny[3], 1),
                  s);
              scheduler_splittask(
                  scheduler_addtask(s, task_type_pair, t->subtype, 1, 0,
                                    ci->progeny[5], cj->progeny[3], 1),
                  s);
              scheduler_splittask(
                  scheduler_addtask(s, task_type_pair, t->subtype, 3, 0,
                                    ci->progeny[6], cj->progeny[3], 1),
                  s);
              scheduler_splittask(
                  scheduler_addtask(s, task_type_pair, t->subtype, 4, 0,
                                    ci->progeny[7], cj->progeny[3], 1),
                  s);
              break;

            case 5: /* (  1 ,  0 , -1 ) */
              t->ci = ci->progeny[4];
              t->cj = cj->progeny[1];
              t->flags = 5;
              t->tight = 1;
              scheduler_splittask(
                  scheduler_addtask(s, task_type_pair, t->subtype, 5, 0,
                                    ci->progeny[6], cj->progeny[3], 1),
                  s);
              scheduler_splittask(
                  scheduler_addtask(s, task_type_pair, t->subtype, 2, 0,
                                    ci->progeny[4], cj->progeny[3], 1),
                  s);
              scheduler_splittask(
                  scheduler_addtask(s, task_type_pair, t->subtype, 8, 0,
                                    ci->progeny[6], cj->progeny[1], 1),
                  s);
              break;

            case 6: /* (  1 , -1 ,  1 ) */
              t->ci = ci->progeny[5];
              t->cj = cj->progeny[2];
              t->flags = 6;
              t->tight = 1;
              break;

            case 7: /* (  1 , -1 ,  0 ) */
              t->ci = ci->progeny[4];
              t->cj = cj->progeny[3];
              t->flags = 6;
              t->tight = 1;
              scheduler_splittask(
                  scheduler_addtask(s, task_type_pair, t->subtype, 8, 0,
                                    ci->progeny[5], cj->progeny[2], 1),
                  s);
              scheduler_splittask(
                  scheduler_addtask(s, task_type_pair, t->subtype, 7, 0,
                                    ci->progeny[4], cj->progeny[2], 1),
                  s);
              scheduler_splittask(
                  scheduler_addtask(s, task_type_pair, t->subtype, 7, 0,
                                    ci->progeny[5], cj->progeny[3], 1),
                  s);
              break;

            case 8: /* (  1 , -1 , -1 ) */
              t->ci = ci->progeny[4];
              t->cj = cj->progeny[3];
              t->flags = 8;
              t->tight = 1;
              break;

            case 9: /* (  0 ,  1 ,  1 ) */
              t->ci = ci->progeny[3];
              t->cj = cj->progeny[0];
              t->flags = 9;
              t->tight = 1;
              scheduler_splittask(
                  scheduler_addtask(s, task_type_pair, t->subtype, 9, 0,
                                    ci->progeny[7], cj->progeny[4], 1),
                  s);
              scheduler_splittask(
                  scheduler_addtask(s, task_type_pair, t->subtype, 0, 0,
                                    ci->progeny[3], cj->progeny[4], 1),
                  s);
              scheduler_splittask(
                  scheduler_addtask(s, task_type_pair, t->subtype, 8, 0,
                                    ci->progeny[7], cj->progeny[0], 1),
                  s);
              break;

            case 10: /* (  0 ,  1 ,  0 ) */
              t->ci = ci->progeny[2];
              t->cj = cj->progeny[0];
              t->flags = 10;
              t->tight = 1;
              scheduler_splittask(
                  scheduler_addtask(s, task_type_pair, t->subtype, 11, 0,
                                    ci->progeny[3], cj->progeny[0], 1),
                  s);
              scheduler_splittask(
                  scheduler_addtask(s, task_type_pair, t->subtype, 7, 0,
                                    ci->progeny[6], cj->progeny[0], 1),
                  s);
              scheduler_splittask(
                  scheduler_addtask(s, task_type_pair, t->subtype, 6, 0,
                                    ci->progeny[7], cj->progeny[0], 1),
                  s);
              scheduler_splittask(
                  scheduler_addtask(s, task_type_pair, t->subtype, 9, 0,
                                    ci->progeny[2], cj->progeny[1], 1),
                  s);
              scheduler_splittask(
                  scheduler_addtask(s, task_type_pair, t->subtype, 10, 0,
                                    ci->progeny[3], cj->progeny[1], 1),
                  s);
              scheduler_splittask(
                  scheduler_addtask(s, task_type_pair, t->subtype, 8, 0,
                                    ci->progeny[6], cj->progeny[1], 1),
                  s);
              scheduler_splittask(
                  scheduler_addtask(s, task_type_pair, t->subtype, 7, 0,
                                    ci->progeny[7], cj->progeny[1], 1),
                  s);
              scheduler_splittask(
                  scheduler_addtask(s, task_type_pair, t->subtype, 1, 0,
                                    ci->progeny[2], cj->progeny[4], 1),
                  s);
              scheduler_splittask(
                  scheduler_addtask(s, task_type_pair, t->subtype, 2, 0,
                                    ci->progeny[3], cj->progeny[4], 1),
                  s);
              scheduler_splittask(
                  scheduler_addtask(s, task_type_pair, t->subtype, 10, 0,
                                    ci->progeny[6], cj->progeny[4], 1),
                  s);
              scheduler_splittask(
                  scheduler_addtask(s, task_type_pair, t->subtype, 11, 0,
                                    ci->progeny[7], cj->progeny[4], 1),
                  s);
              scheduler_splittask(
                  scheduler_addtask(s, task_type_pair, t->subtype, 0, 0,
                                    ci->progeny[2], cj->progeny[5], 1),
                  s);
              scheduler_splittask(
                  scheduler_addtask(s, task_type_pair, t->subtype, 1, 0,
                                    ci->progeny[3], cj->progeny[5], 1),
                  s);
              scheduler_splittask(
                  scheduler_addtask(s, task_type_pair, t->subtype, 9, 0,
                                    ci->progeny[6], cj->progeny[5], 1),
                  s);
              scheduler_splittask(
                  scheduler_addtask(s, task_type_pair, t->subtype, 10, 0,
                                    ci->progeny[7], cj->progeny[5], 1),
                  s);
              break;

            case 11: /* (  0 ,  1 , -1 ) */
              t->ci = ci->progeny[2];
              t->cj = cj->progeny[1];
              t->flags = 11;
              t->tight = 1;
              scheduler_splittask(
                  scheduler_addtask(s, task_type_pair, t->subtype, 11, 0,
                                    ci->progeny[6], cj->progeny[5], 1),
                  s);
              scheduler_splittask(
                  scheduler_addtask(s, task_type_pair, t->subtype, 2, 0,
                                    ci->progeny[2], cj->progeny[5], 1),
                  s);
              scheduler_splittask(
                  scheduler_addtask(s, task_type_pair, t->subtype, 6, 0,
                                    ci->progeny[6], cj->progeny[1], 1),
                  s);
              break;

            case 12: /* (  0 ,  0 ,  1 ) */
              t->ci = ci->progeny[1];
              t->cj = cj->progeny[0];
              t->flags = 12;
              t->tight = 1;
              scheduler_splittask(
                  scheduler_addtask(s, task_type_pair, t->subtype, 11, 0,
                                    ci->progeny[3], cj->progeny[0], 1),
                  s);
              scheduler_splittask(
                  scheduler_addtask(s, task_type_pair, t->subtype, 5, 0,
                                    ci->progeny[5], cj->progeny[0], 1),
                  s);
              scheduler_splittask(
                  scheduler_addtask(s, task_type_pair, t->subtype, 2, 0,
                                    ci->progeny[7], cj->progeny[0], 1),
                  s);
              scheduler_splittask(
                  scheduler_addtask(s, task_type_pair, t->subtype, 9, 0,
                                    ci->progeny[1], cj->progeny[2], 1),
                  s);
              scheduler_splittask(
                  scheduler_addtask(s, task_type_pair, t->subtype, 12, 0,
                                    ci->progeny[3], cj->progeny[2], 1),
                  s);
              scheduler_splittask(
                  scheduler_addtask(s, task_type_pair, t->subtype, 8, 0,
                                    ci->progeny[5], cj->progeny[2], 1),
                  s);
              scheduler_splittask(
                  scheduler_addtask(s, task_type_pair, t->subtype, 5, 0,
                                    ci->progeny[7], cj->progeny[2], 1),
                  s);
              scheduler_splittask(
                  scheduler_addtask(s, task_type_pair, t->subtype, 3, 0,
                                    ci->progeny[1], cj->progeny[4], 1),
                  s);
              scheduler_splittask(
                  scheduler_addtask(s, task_type_pair, t->subtype, 6, 0,
                                    ci->progeny[3], cj->progeny[4], 1),
                  s);
              scheduler_splittask(
                  scheduler_addtask(s, task_type_pair, t->subtype, 12, 0,
                                    ci->progeny[5], cj->progeny[4], 1),
                  s);
              scheduler_splittask(
                  scheduler_addtask(s, task_type_pair, t->subtype, 11, 0,
                                    ci->progeny[7], cj->progeny[4], 1),
                  s);
              scheduler_splittask(
                  scheduler_addtask(s, task_type_pair, t->subtype, 0, 0,
                                    ci->progeny[1], cj->progeny[6], 1),
                  s);
              scheduler_splittask(
                  scheduler_addtask(s, task_type_pair, t->subtype, 3, 0,
                                    ci->progeny[3], cj->progeny[6], 1),
                  s);
              scheduler_splittask(
                  scheduler_addtask(s, task_type_pair, t->subtype, 9, 0,
                                    ci->progeny[5], cj->progeny[6], 1),
                  s);
              scheduler_splittask(
                  scheduler_addtask(s, task_type_pair, t->subtype, 12, 0,
                                    ci->progeny[7], cj->progeny[6], 1),
                  s);
              break;
          } /* switch(sid) */
575
576
        }

577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
        /* Otherwise, break it up if it is too large? */
      } else if (scheduler_doforcesplit && ci->split && cj->split &&
                 (ci->count > space_maxsize / cj->count)) {

        // message( "force splitting pair with %i and %i parts." , ci->count ,
        // cj->count );

        /* Replace the current task. */
        t->type = task_type_none;

        for (int j = 0; j < 8; j++)
          if (ci->progeny[j] != NULL)
            for (int k = 0; k < 8; k++)
              if (cj->progeny[k] != NULL) {
                struct task *tl =
                    scheduler_addtask(s, task_type_pair, t->subtype, 0, 0,
                                      ci->progeny[j], cj->progeny[k], 0);
                scheduler_splittask(tl, s);
                tl->flags = space_getsid(s->space, &t->ci, &t->cj, shift);
              }

        /* Otherwise, if not spilt, stitch-up the sorting. */
      } else {

        /* Create the sort for ci. */
        lock_lock(&ci->lock);
        if (ci->sorts == NULL)
          ci->sorts = scheduler_addtask(s, task_type_sort, task_subtype_none,
                                        1 << sid, 0, ci, NULL, 0);
        else
          ci->sorts->flags |= (1 << sid);
        lock_unlock_blind(&ci->lock);
        scheduler_addunlock(s, ci->sorts, t);

        /* Create the sort for cj. */
        lock_lock(&cj->lock);
        if (cj->sorts == NULL)
          cj->sorts = scheduler_addtask(s, task_type_sort, task_subtype_none,
                                        1 << sid, 0, cj, NULL, 0);
        else
          cj->sorts->flags |= (1 << sid);
        lock_unlock_blind(&cj->lock);
        scheduler_addunlock(s, cj->sorts, t);
      }

    } /* pair interaction? */

    /* Long-range gravity interaction ? */
    else if (t->type == task_type_grav_mm) {

      /* Get a handle on the cells involved. */
      struct cell *ci = t->ci;

      /* Safety thing */
      if (ci->gcount == 0) t->type = task_type_none;

    } /* gravity interaction? */
  }   /* iterate over the current task. */
}

/**
 * @brief Mapper function to split tasks that may be too large.
 *
 * @param map_data the tasks to process
 * @param num_elements the number of tasks.
 * @param extra_data The #scheduler we are working in.
 */

void scheduler_splittasks_mapper(void *map_data, int num_elements,
                                 void *extra_data) {

  /* Extract the parameters. */
  struct scheduler *s = (struct scheduler *)extra_data;
  struct task *tasks = (struct task *)map_data;
651

652
653
654
  for (int ind = 0; ind < num_elements; ind++) {
    struct task *t = &tasks[ind];
    scheduler_splittask(t, s);
655
  }
656
}
657

658
void scheduler_splittasks(struct scheduler *s) {
659

660
661
  /* Call the mapper on each current task. */
  threadpool_map(s->threadpool, scheduler_splittasks_mapper, s->tasks,
662
                 s->nr_tasks, sizeof(struct task), 1000, s);
663
664
}

665
666
667
668
669
670
671
/**
 * @brief Add a #task to the #scheduler.
 *
 * @param s The #scheduler we are working in.
 * @param type The type of the task.
 * @param subtype The sub-type of the task.
 * @param flags The flags of the task.
672
 * @param wait
673
674
675
676
677
 * @param ci The first cell to interact.
 * @param cj The second cell to interact.
 * @param tight
 */

678
679
680
struct task *scheduler_addtask(struct scheduler *s, enum task_types type,
                               enum task_subtypes subtype, int flags, int wait,
                               struct cell *ci, struct cell *cj, int tight) {
681
682

  /* Get the next free task. */
Pedro Gonnet's avatar
Pedro Gonnet committed
683
  const int ind = atomic_inc(&s->tasks_next);
Matthieu Schaller's avatar
Matthieu Schaller committed
684

685
686
687
688
  /* Overflow? */
  if (ind >= s->size) error("Task list overflow.");

  /* Get a pointer to the new task. */
Pedro Gonnet's avatar
Pedro Gonnet committed
689
  struct task *t = &s->tasks[ind];
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705

  /* Copy the data. */
  t->type = type;
  t->subtype = subtype;
  t->flags = flags;
  t->wait = wait;
  t->ci = ci;
  t->cj = cj;
  t->skip = 0;
  t->tight = tight;
  t->implicit = 0;
  t->weight = 0;
  t->rank = 0;
  t->tic = 0;
  t->toc = 0;
  t->nr_unlock_tasks = 0;
706
  t->rid = -1;
707
708
709
710
711
712
713
714
715

  /* Add an index for it. */
  // lock_lock( &s->lock );
  s->tasks_ind[atomic_inc(&s->nr_tasks)] = ind;
  // lock_unlock_blind( &s->lock );

  /* Return a pointer to the new task. */
  return t;
}
716

717
718
719
720
721
722
723
724
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
/**
 * @brief Set the unlock pointers in each task.
 *
 * @param s The #scheduler.
 */

void scheduler_set_unlocks(struct scheduler *s) {

  /* Store the counts for each task. */
  int *counts;
  if ((counts = (int *)malloc(sizeof(int) * s->nr_tasks)) == NULL)
    error("Failed to allocate temporary counts array.");
  bzero(counts, sizeof(int) * s->nr_tasks);
  for (int k = 0; k < s->nr_unlocks; k++) counts[s->unlock_ind[k]] += 1;

  /* Compute the offset for each unlock block. */
  int *offsets;
  if ((offsets = (int *)malloc(sizeof(int) * (s->nr_tasks + 1))) == NULL)
    error("Failed to allocate temporary offsets array.");
  offsets[0] = 0;
  for (int k = 0; k < s->nr_tasks; k++) offsets[k + 1] = offsets[k] + counts[k];

  /* Create and fill a temporary array with the sorted unlocks. */
  struct task **unlocks;
  if ((unlocks = (struct task **)malloc(sizeof(struct task *) *
                                        s->size_unlocks)) == NULL)
    error("Failed to allocate temporary unlocks array.");
  for (int k = 0; k < s->nr_unlocks; k++) {
    const int ind = s->unlock_ind[k];
    unlocks[offsets[ind]] = s->unlocks[k];
    offsets[ind] += 1;
  }

  /* Swap the unlocks. */
  free(s->unlocks);
  s->unlocks = unlocks;

  /* Re-set the offsets. */
  offsets[0] = 0;
  for (int k = 1; k < s->nr_tasks; k++)
    offsets[k] = offsets[k - 1] + counts[k - 1];

  /* Set the unlocks in the tasks. */
  for (int k = 0; k < s->nr_tasks; k++) {
    struct task *t = &s->tasks[k];
    t->nr_unlock_tasks = counts[k];
    t->unlock_tasks = &s->unlocks[offsets[k]];
  }
765

766
767
768
769
770
771
772
773
774
775
  /* Verify that there are no duplicate unlocks. */
  /* for (int k = 0; k < s->nr_tasks; k++) {
    struct task *t = &s->tasks[k];
    for (int i = 0; i < t->nr_unlock_tasks; i++) {
      for (int j = i + 1; j < t->nr_unlock_tasks; j++) {
        if (t->unlock_tasks[i] == t->unlock_tasks[j])
          error("duplicate unlock!");
      }
    }
  } */
776
777
778
779
780
781

  /* Clean up. */
  free(counts);
  free(offsets);
}

782
/**
783
784
785
786
 * @brief Sort the tasks in topological order over all queues.
 *
 * @param s The #scheduler.
 */
787
788
789

void scheduler_ranktasks(struct scheduler *s) {

Pedro Gonnet's avatar
Pedro Gonnet committed
790
791
792
  struct task *tasks = s->tasks;
  int *tid = s->tasks_ind;
  const int nr_tasks = s->nr_tasks;
793

794
  /* Run through the tasks and get all the waits right. */
795
796
797
798
799
800
801
802
  for (int i = 0; i < nr_tasks; i++) {
    struct task *t = &tasks[i];

    // Increment the waits of the dependances
    for (int k = 0; k < t->nr_unlock_tasks; k++) {
      t->unlock_tasks[k]->wait++;
    }
  }
803

804
805
806
807
808
809
810
811
  /* Load the tids of tasks with no waits. */
  int left = 0;
  for (int k = 0; k < nr_tasks; k++)
    if (tasks[k].wait == 0) {
      tid[left] = k;
      left += 1;
    }

812
  /* Main loop. */
813
  for (int j = 0, rank = 0; left < nr_tasks; rank++) {
814
815
816

    /* Did we get anything? */
    if (j == left) error("Unsatisfiable task dependencies detected.");
817
    const int left_old = left;
818
819

    /* Unlock the next layer of tasks. */
820
821
    for (; j < left_old; j++) {
      struct task *t = &tasks[tid[j]];
822
823
824
825
      t->rank = rank;
      /* message( "task %i of type %s has rank %i." , i ,
          (t->type == task_type_self) ? "self" : (t->type == task_type_pair) ?
         "pair" : "sort" , rank ); */
826
827
828
829
830
831
832
      for (int k = 0; k < t->nr_unlock_tasks; k++) {
        struct task *u = t->unlock_tasks[k];
        if (--u->wait == 0) {
          tid[left] = u - tasks;
          left += 1;
        }
      }
833
834
    }

835
836
    /* Move back to the old left (like Sanders). */
    j = left_old;
837
838
  }

839
#ifdef SWIFT_DEBUG_CHECKS
840
  /* Verify that the tasks were ranked correctly. */
841
  for (int k = 1; k < s->nr_tasks; k++)
842
    if (tasks[tid[k - 1]].rank > tasks[tid[k - 1]].rank)
843
844
      error("Task ranking failed.");
#endif
845
}
846
847
848
849
850
851
852
853

/**
 * @brief (Re)allocate the task arrays.
 *
 * @param s The #scheduler.
 * @param size The maximum number of tasks in the #scheduler.
 */

854
void scheduler_reset(struct scheduler *s, int size) {
855

856
857
  /* Do we need to re-allocate? */
  if (size > s->size) {
858

859
    /* Free existing task lists if necessary. */
860
861
    if (s->tasks != NULL) free(s->tasks);
    if (s->tasks_ind != NULL) free(s->tasks_ind);
862

863
    /* Allocate the new lists. */
864
865
    if ((s->tasks = (struct task *)malloc(sizeof(struct task) * size)) ==
            NULL ||
866
867
868
        (s->tasks_ind = (int *)malloc(sizeof(int) * size)) == NULL)
      error("Failed to allocate task lists.");
  }
869

870
871
872
873
874
  /* Reset the counters. */
  s->size = size;
  s->nr_tasks = 0;
  s->tasks_next = 0;
  s->waiting = 0;
875
  s->mask = 0;
876
  s->submask = 0;
877
  s->nr_unlocks = 0;
878
  s->completed_unlock_writes = 0;
879
880

  /* Set the task pointers in the queues. */
Pedro Gonnet's avatar
Pedro Gonnet committed
881
  for (int k = 0; k < s->nr_queues; k++) s->queues[k].tasks = s->tasks;
882
}
883
884

/**
885
 * @brief Compute the task weights
886
887
888
 *
 * @param s The #scheduler.
 */
889

890
891
void scheduler_reweight(struct scheduler *s) {

Pedro Gonnet's avatar
Pedro Gonnet committed
892
893
894
895
  const int nr_tasks = s->nr_tasks;
  int *tid = s->tasks_ind;
  struct task *tasks = s->tasks;
  const int nodeID = s->nodeID;
Pedro Gonnet's avatar
Pedro Gonnet committed
896
897
898
  const float sid_scale[13] = {0.1897, 0.4025, 0.1897, 0.4025, 0.5788,
                               0.4025, 0.1897, 0.4025, 0.1897, 0.4025,
                               0.5788, 0.4025, 0.5788};
Pedro Gonnet's avatar
Pedro Gonnet committed
899
  const float wscale = 0.001;
900
901
  // ticks tic;

902
  /* Run through the tasks backwards and set their waits and
903
904
     weights. */
  // tic = getticks();
Pedro Gonnet's avatar
Pedro Gonnet committed
905
906
  for (int k = nr_tasks - 1; k >= 0; k--) {
    struct task *t = &tasks[tid[k]];
907
    t->weight = 0;
Pedro Gonnet's avatar
Pedro Gonnet committed
908
    for (int j = 0; j < t->nr_unlock_tasks; j++)
909
910
911
912
913
914
915
      if (t->unlock_tasks[j]->weight > t->weight)
        t->weight = t->unlock_tasks[j]->weight;
    if (!t->implicit && t->tic > 0)
      t->weight += wscale * (t->toc - t->tic);
    else
      switch (t->type) {
        case task_type_sort:
916
917
          t->weight += wscale * intrinsics_popcount(t->flags) * t->ci->count *
                       (sizeof(int) * 8 - intrinsics_clz(t->ci->count));
918
919
920
921
922
923
924
925
926
927
928
929
          break;
        case task_type_self:
          t->weight += 1 * t->ci->count * t->ci->count;
          break;
        case task_type_pair:
          if (t->ci->nodeID != nodeID || t->cj->nodeID != nodeID)
            t->weight +=
                3 * wscale * t->ci->count * t->cj->count * sid_scale[t->flags];
          else
            t->weight +=
                2 * wscale * t->ci->count * t->cj->count * sid_scale[t->flags];
          break;
930
        case task_type_sub_pair:
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
          if (t->ci->nodeID != nodeID || t->cj->nodeID != nodeID) {
            if (t->flags < 0)
              t->weight += 3 * wscale * t->ci->count * t->cj->count;
            else
              t->weight += 3 * wscale * t->ci->count * t->cj->count *
                           sid_scale[t->flags];
          } else {
            if (t->flags < 0)
              t->weight += 2 * wscale * t->ci->count * t->cj->count;
            else
              t->weight += 2 * wscale * t->ci->count * t->cj->count *
                           sid_scale[t->flags];
          }
          break;
        case task_type_sub_self:
          t->weight += 1 * wscale * t->ci->count * t->ci->count;
947
948
949
950
          break;
        case task_type_ghost:
          if (t->ci == t->ci->super) t->weight += wscale * t->ci->count;
          break;
Matthieu Schaller's avatar
Matthieu Schaller committed
951
952
953
954
        case task_type_kick:
          t->weight += wscale * t->ci->count;
          break;
        case task_type_init:
955
956
957
958
959
960
          t->weight += wscale * t->ci->count;
          break;
        default:
          break;
      }
  }
961
  // message( "weighting tasks took %.3f %s." ,
Matthieu Schaller's avatar
Matthieu Schaller committed
962
  // clocks_from_ticks( getticks() - tic ), clocks_getunit());
963
964
965
966
967
968
969
970
971

  /* int min = tasks[0].weight, max = tasks[0].weight;
  for ( k = 1 ; k < nr_tasks ; k++ )
      if ( tasks[k].weight < min )
          min = tasks[k].weight;
      else if ( tasks[k].weight > max )
          max = tasks[k].weight;
  message( "task weights are in [ %i , %i ]." , min , max ); */
}
972

Pedro Gonnet's avatar
Pedro Gonnet committed
973
974
975
976
/**
 * @brief #threadpool_map function which runs through the task
 *        graph and re-computes the task wait counters.
 */
977

978
979
void scheduler_rewait_mapper(void *map_data, int num_elements,
                             void *extra_data) {
Pedro Gonnet's avatar
Pedro Gonnet committed
980
981

  struct scheduler *s = (struct scheduler *)extra_data;
982
  struct task *tasks = (struct task *)map_data;
Pedro Gonnet's avatar
Pedro Gonnet committed
983

984
985
  for (int ind = 0; ind < num_elements; ind++) {
    struct task *t = &tasks[ind];
Pedro Gonnet's avatar
Pedro Gonnet committed
986

987
988
989
    if (t->skip || !((1 << t->type) & s->mask) ||
        !((1 << t->subtype) & s->submask))
      continue;
Pedro Gonnet's avatar
Pedro Gonnet committed
990

991
992
    /* Skip sort tasks that have already been performed */
    if (t->type == task_type_sort && t->flags == 0) {
993
      error("Empty sort task encountered.");
994
    }
Pedro Gonnet's avatar
Pedro Gonnet committed
995

996
997
998
999
1000
    /* Sets the waits of the dependances */
    for (int k = 0; k < t->nr_unlock_tasks; k++) {
      struct task *u = t->unlock_tasks[k];
      atomic_inc(&u->wait);
    }