space.c 33.1 KB
Newer Older
Pedro Gonnet's avatar
Pedro Gonnet committed
1
/*******************************************************************************
2
 * This file is part of SWIFT.
Pedro Gonnet's avatar
Pedro Gonnet committed
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
 * Coypright (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/>.
 * 
 ******************************************************************************/

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

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

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

Pedro Gonnet's avatar
Pedro Gonnet committed
38
/* Local headers. */
39
#include "const.h"
Pedro Gonnet's avatar
Pedro Gonnet committed
40
#include "cycle.h"
41
#include "atomic.h"
Pedro Gonnet's avatar
Pedro Gonnet committed
42
#include "lock.h"
43
#include "task.h"
44
#include "kernel.h"
45
#include "part.h"
Pedro Gonnet's avatar
Pedro Gonnet committed
46
#include "space.h"
47
48
49
#include "cell.h"
#include "scheduler.h"
#include "engine.h"
Pedro Gonnet's avatar
Pedro Gonnet committed
50
#include "runner.h"
51
#include "error.h"
Pedro Gonnet's avatar
Pedro Gonnet committed
52
53
54

/* Split size. */
int space_splitsize = space_splitsize_default;
55
int space_subsize = space_subsize_default;
56
int space_maxsize = space_maxsize_default;
Pedro Gonnet's avatar
Pedro Gonnet committed
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
85
86
87

/* Map shift vector to sortlist. */
const int sortlistID[27] = {
    /* ( -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 
    };
88
89
    
    
90
91
92
93
94
95
96
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
132
133
134
135
136
137
/**
 * @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.
 */
 
int space_getsid ( struct space *s , struct cell **ci , struct cell **cj , double *shift ) {

    int k, sid = 0;
    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 ( dx[k] < -s->dim[k]/2 )
            shift[k] = s->dim[k];
        else if ( 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;

    }


138
/**
139
 * @brief Recursively dismantle a cell tree.
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
 *
 */
 
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;
                }
    
    }
156
157
    
    
158
/**
159
 * @brief Re-build the cell grid.
160
 *
161
162
 * @param s The #space.
 * @param cell_max Maximum cell edge length.
163
164
 */
 
165
void space_regrid ( struct space *s , double cell_max ) {
166

167
    float h_max = s->cell_min / kernel_gamma / space_stretch, dmin;
168
    int i, j, k, cdim[3], nr_parts = s->nr_parts;
169
    struct cell *restrict c;
170
    // ticks tic;
171
172
    
    /* Run through the parts and get the current h_max. */
173
    // tic = getticks();
174
175
176
177
178
179
180
181
182
183
184
185
    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;
            }
        }
    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;
186
        }
187
188
    // message( "h_max is %.3e (cell_max=%.3e)." , h_max , cell_max );
    // message( "getting h_min and h_max took %.3f ms." , (double)(getticks() - tic) / CPU_TPS * 1000 );
189
190
191
    
    /* Get the new putative cell dimensions. */
    for ( k = 0 ; k < 3 ; k++ )
192
        cdim[k] = floor( s->dim[k] / fmax( h_max*kernel_gamma*space_stretch , cell_max ) );
193
        
194
195
196
197
198
199
    /* 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
        
200
    /* Do we need to re-build the upper-level cells? */
201
    // tic = getticks();
Pedro Gonnet's avatar
Pedro Gonnet committed
202
    if ( s->cells == NULL ||
203
         cdim[0] < s->cdim[0] || cdim[1] < s->cdim[1] || cdim[2] < s->cdim[2] ) {
204
205
206
    
        /* Free the old cells, if they were allocated. */
        if ( s->cells != NULL ) {
207
            for ( k = 0 ; k < s->nr_cells ; k++ ) {
208
                space_rebuild_recycle( s , &s->cells[k] );
209
210
211
                if ( s->cells[k].sort != NULL )
                    free( s->cells[k].sort );
                }
212
213
214
215
            free( s->cells );
            s->maxdepth = 0;
            }
            
216
        /* Set the new cell dimensions only if smaller. */
217
218
219
220
221
        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];
            }
222
        dmin = fminf( s->h[0] , fminf( s->h[1] , s->h[2] ) );
223

224
        /* Allocate the highest level of cells. */
Pedro Gonnet's avatar
Pedro Gonnet committed
225
        s->tot_cells = s->nr_cells = cdim[0] * cdim[1] * cdim[2];
226
227
228
229
230
231
232
233
234
235
236
237
238
239
        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];
240
                    c->dmin = dmin;
241
                    c->depth = 0;
Pedro Gonnet's avatar
Pedro Gonnet committed
242
                    c->count = 0;
243
                    lock_init( &c->lock );
244
                    }
245
246
           
        /* Be verbose about the change. */         
247
        message( "set cell dimensions to [ %i %i %i ]." , cdim[0] , cdim[1] , cdim[2] ); fflush(stdout);
248
249
                    
        } /* re-build upper-level cells? */
250
    // message( "rebuilding upper-level cells took %.3f ms." , (double)(getticks() - tic) / CPU_TPS * 1000 );
251
        
Pedro Gonnet's avatar
Pedro Gonnet committed
252
253
254
255
256
257
258
259
260
    /* 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;
261
            s->cells[k].nr_force = 0;
Pedro Gonnet's avatar
Pedro Gonnet committed
262
            s->cells[k].dx_max = 0.0f;
263
            s->cells[k].sorted = 0;
Pedro Gonnet's avatar
Pedro Gonnet committed
264
            s->cells[k].count = 0;
265
            s->cells[k].kick1 = NULL;
Pedro Gonnet's avatar
Pedro Gonnet committed
266
            s->cells[k].kick2 = NULL;
Pedro Gonnet's avatar
Pedro Gonnet committed
267
268
269
270
            }
        s->maxdepth = 0;
    
        }
271
        
272
273
274
275
276
277
278
279
280
281
282
283
284
    }
    

/**
 * @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.
 *
 */
 
void space_rebuild ( struct space *s , double cell_max ) {

285
    float h_max = s->cell_min / kernel_gamma / space_stretch, dmin;
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
    int i, j, k, cdim[3], nr_parts = s->nr_parts;
    struct cell *restrict c, *restrict cells = s->cells;
    struct part *restrict finger, *restrict p, *parts = s->parts;
    struct xpart *xfinger, *xparts = s->xparts;
    int *ind;
    double ih[3], dim[3];
    // ticks tic;
    
    /* Be verbose about this. */
    // message( "re)building space..." ); fflush(stdout);
    
    /* Run through the parts and get the current h_max. */
    // tic = getticks();
    if ( cells != NULL ) {
        for ( k = 0 ; k < s->nr_cells ; k++ ) {
            if ( cells[k].h_max > h_max )
                h_max = cells[k].h_max;
            }
        }
    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;
        }
    // message( "h_max is %.3e (cell_max=%.3e)." , h_max , cell_max );
    // message( "getting h_min and h_max took %.3f ms." , (double)(getticks() - tic) / CPU_TPS * 1000 );
    
    /* 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 ) );
        
    /* 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 ( 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 ( cells != NULL ) {
            for ( k = 0 ; k < s->nr_cells ; k++ ) {
                space_rebuild_recycle( s , &cells[k] );
                if ( cells[k].sort != NULL )
                    free( cells[k].sort );
                }
            free( 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 *)&cells , 64 , s->nr_cells * sizeof(struct cell) ) != 0 )
            error( "Failed to allocate cells." );
        s->cells = cells;
        bzero( cells , s->nr_cells * sizeof(struct cell) );
        for ( k = 0 ; k < s->nr_cells ; k++ )
            if ( lock_init( &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 = &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;
                    lock_init( &c->lock );
                    }
           
        /* Be verbose about the change. */         
        message( "set cell dimensions to [ %i %i %i ]." , cdim[0] , cdim[1] , cdim[2] ); fflush(stdout);
                    
        } /* re-build upper-level cells? */
    // message( "rebuilding upper-level cells took %.3f ms." , (double)(getticks() - tic) / CPU_TPS * 1000 );
        
    /* 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 , &cells[k] );
            cells[k].sorts = NULL;
            cells[k].nr_tasks = 0;
            cells[k].nr_density = 0;
            cells[k].nr_force = 0;
            cells[k].dx_max = 0.0f;
            cells[k].sorted = 0;
            cells[k].count = 0;
            cells[k].kick1 = NULL;
            cells[k].kick2 = NULL;
            }
        s->maxdepth = 0;
    
        }
        
398
    /* Run through the particles and get their cell index. */
399
    // tic = getticks();
400
    if ( ( ind = (int *)malloc( sizeof(int) * s->size_parts ) ) == NULL )
401
        error( "Failed to allocate temporary particle indices." );
Pedro Gonnet's avatar
Pedro Gonnet committed
402
403
    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];
Pedro Gonnet's avatar
Pedro Gonnet committed
404
    cdim[0] = s->cdim[0]; cdim[1] = s->cdim[1]; cdim[2] = s->cdim[2];
405
    #pragma omp parallel for private(p,j)
406
407
    for ( k = 0 ; k < nr_parts ; k++ )  {
        p = &parts[k];
408
409
        for ( j = 0 ; j < 3 ; j++ )
            if ( p->x[j] < 0.0 )
Pedro Gonnet's avatar
Pedro Gonnet committed
410
411
412
413
                p->x[j] += dim[j];
            else if ( p->x[j] >= dim[j] )
                p->x[j] -= dim[j];
        ind[k] = cell_getid( cdim , p->x[0]*ih[0] , p->x[1]*ih[1] , p->x[2]*ih[2] );
414
        atomic_inc( &cells[ ind[k] ].count );
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
    // message( "getting particle indices took %.3f ms." , (double)(getticks() - tic) / CPU_TPS * 1000 );


    /* Move non-local parts to the end of the list. */
    #ifdef WITH_MPI
        int nodeID = s->e->nodeID;
        for ( k = 0 ; k < nr_parts ; k++ )
            if ( cells[ ind[k] ].nodeID != nodeID ) {
                cells[ ind[k] ].count -= 1;
                nr_parts -= 1;
                struct part tp = parts[k];
                parts[k] = parts[ nr_parts ];
                parts[ nr_parts ] = tp;
                struct xpart txp = xparts[k];
                xparts[k] = xparts[ nr_parts ];
                xparts[ nr_parts ] = txp;
                int t = ind[k];
                ind[k] = ind[ nr_parts ];
                ind[ nr_parts ] = t;
                }
        s->nr_parts = nr_parts + engine_exchange_strays( s->e , &parts[nr_parts] , &xparts[nr_parts] , &ind[nr_parts] , s->nr_parts - nr_parts );
        for ( k = nr_parts ; k < s->nr_parts ; k++ ) {
            p = &parts[k];
            ind[k] = cell_getid( cdim , p->x[0]*ih[0] , p->x[1]*ih[1] , p->x[2]*ih[2] );
            cells[ ind[k] ].count += 1;
441
442
            /* if ( cells[ ind[k] ].nodeID != nodeID )
                error( "Received part that does not belong to me (nodeID=%i)." , cells[ ind[k] ].nodeID ); */
443
444
445
446
            }
        nr_parts = s->nr_parts;
    #endif
    
447
448

    /* Sort the parts according to their cells. */
449
    // tic = getticks();
450
451
    parts_sort( parts , xparts , ind , nr_parts , 0 , s->nr_cells-1 );
    // message( "parts_sort took %.3f ms." , (double)(getticks() - tic) / CPU_TPS * 1000 );
452
    
Pedro Gonnet's avatar
Pedro Gonnet committed
453
454
455
456
457
458
459
460
461
    /* Verify sort. */
    /* for ( k = 1 ; k < nr_parts ; k++ ) {
        if ( ind[k-1] > ind[k] ) {
            error( "Sort failed!" );
            }
        else if ( ind[k] != cell_getid( cdim , parts[k].x[0]*ih[0] , parts[k].x[1]*ih[1] , parts[k].x[2]*ih[2] ) )
            error( "Incorrect indices!" );
        } */
    
462
463
    /* We no longer need the indices as of here. */
    free( ind );    
464
465

    /* Hook the cells up to the parts. */
466
    // tic = getticks();
467
468
    finger = parts;
    xfinger = xparts;
469
    for ( k = 0 ; k < s->nr_cells ; k++ ) {
470
        c = &cells[ k ];
471
        c->parts = finger;
472
        c->xparts = xfinger;
473
        finger = &finger[ c->count ];
474
        xfinger = &xfinger[ c->count ];
475
        }
476
    // message( "hooking up cells took %.3f ms." , (double)(getticks() - tic) / CPU_TPS * 1000 );
477
        
478
479
    /* At this point, we have the upper-level cells, old or new. Now make
       sure that the parts in each cell are ok. */
480
    // tic = getticks();
Pedro Gonnet's avatar
Pedro Gonnet committed
481
    k = 0;
482
    #pragma omp parallel shared(s,k)
Pedro Gonnet's avatar
Pedro Gonnet committed
483
    {
484
485
486
487
        if ( omp_get_thread_num() < 8 )
            while ( 1 ) {
                int myk = atomic_inc( &k );
                if ( myk < s->nr_cells )
488
                    space_split( s , &cells[myk] );
489
490
491
                else
                    break;
                }
Pedro Gonnet's avatar
Pedro Gonnet committed
492
        }
493
494
    // message( "space_split took %.3f ms." , (double)(getticks() - tic) / CPU_TPS * 1000 );
    
495
496
497
    }


498
/**
499
 * @brief Sort the particles and condensed particles according to the given indices.
500
501
502
503
504
505
506
507
 *
 * @param parts The list of #part
 * @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.
 */
 
508
void parts_sort ( struct part *parts , struct xpart *xparts , int *ind , int N , int min , int max ) {
509

510
    struct qstack {
511
        volatile int i, j, min, max;
512
        volatile int ready;
513
514
        };
    struct qstack *qstack;
515
    int qstack_size = (max-min)/2 + 10;
516
    volatile unsigned int first, last, waiting;
517
518
519
    
    int pivot;
    int i, ii, j, jj, temp_i, qid;
520
    struct part temp_p;
521
    struct xpart temp_xp;
522
    
523
524
525
526
    /* Allocate the stack. */
    if ( ( qstack = malloc( sizeof(struct qstack) * qstack_size ) ) == NULL )
        error( "Failed to allocate qstack." );
    
527
528
529
530
531
    /* Init the interval stack. */
    qstack[0].i = 0;
    qstack[0].j = N-1;
    qstack[0].min = min;
    qstack[0].max = max;
532
    qstack[0].ready = 1;
533
    for ( i = 1 ; i < qstack_size ; i++ )
534
535
        qstack[i].ready = 0;
    first = 0; last = 1; waiting = 1;
536
537
    
    /* Parallel bit. */
538
    #pragma omp parallel default(none) shared(first,last,waiting,qstack,parts,xparts,ind,qstack_size,stderr,engine_rank) private(pivot,i,ii,j,jj,min,max,temp_i,qid,temp_xp,temp_p)
539
    {
540
    
541
        /* Main loop. */
542
        if ( omp_get_thread_num() < 8 )
543
544
        while ( waiting > 0 ) {
        
545
            /* Grab an interval off the queue. */
546
            qid = atomic_inc( &first ) % qstack_size;
547
            
548
            /* Wait for the interval to be ready. */
549
            while ( waiting > 0 && atomic_cas( &qstack[qid].ready , 1 , 1 ) != 1 );
550
551
552
553
554
            
            /* Broke loop for all the wrong reasons? */
            if ( waiting == 0 )
                break;
        
555
556
557
558
559
            /* Get the stack entry. */
            i = qstack[qid].i;
            j = qstack[qid].j;
            min = qstack[qid].min;
            max = qstack[qid].max;
560
            qstack[qid].ready = 0;
561
            // message( "thread %i got interval [%i,%i] with values in [%i,%i]." , omp_get_thread_num() , i , j , min , max );
562
            
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
            /* Loop over sub-intervals. */
            while ( 1 ) {
            
                /* Bring beer. */
                pivot = (min + max) / 2;
                
                /* One pass of QuickSort's partitioning. */
                ii = i; jj = j;
                while ( ii < jj ) {
                    while ( ii <= j && ind[ii] <= pivot )
                        ii++;
                    while ( jj >= i && ind[jj] > pivot )
                        jj--;
                    if ( ii < jj ) {
                        temp_i = ind[ii]; ind[ii] = ind[jj]; ind[jj] = temp_i;
                        temp_p = parts[ii]; parts[ii] = parts[jj]; parts[jj] = temp_p;
579
                        temp_xp = xparts[ii]; xparts[ii] = xparts[jj]; xparts[jj] = temp_xp;
580
                        }
581
                    }
582

583
584
585
                /* Verify sort. */
                /* for ( int k = i ; k <= jj ; k++ )
                    if ( ind[k] > pivot ) {
586
                        message( "sorting failed at k=%i, ind[k]=%i, pivot=%i, i=%i, j=%i, N=%i." , k , ind[k] , pivot , i , j , N );
587
588
589
590
                        error( "Partition failed (<=pivot)." );
                        }
                for ( int k = jj+1 ; k <= j ; k++ )
                    if ( ind[k] <= pivot ) {
591
                        message( "sorting failed at k=%i, ind[k]=%i, pivot=%i, i=%i, j=%i, N=%i." , k , ind[k] , pivot , i , j , N );
592
593
594
595
596
597
598
599
                        error( "Partition failed (>pivot)." );
                        } */
                        
                /* Split-off largest interval. */
                if ( jj - i > j - jj+1 ) {

                    /* Recurse on the left? */
                    if ( jj > i  && pivot > min ) {
600
                        qid = atomic_inc( &last ) % qstack_size;
601
                        while ( atomic_cas( &qstack[qid].ready , 0 , 0 ) != 0 );
602
603
604
605
606
                        qstack[qid].i = i;
                        qstack[qid].j = jj;
                        qstack[qid].min = min;
                        qstack[qid].max = pivot;
                        qstack[qid].ready = 1;
607
608
                        if ( atomic_inc( &waiting ) >= qstack_size )
                            error( "Qstack overflow." );
609
610
611
612
613
614
615
616
617
618
                        }

                    /* Recurse on the right? */
                    if ( jj+1 < j && pivot+1 < max ) {
                        i = jj+1;
                        min = pivot+1;
                        }
                    else
                        break;
                        
619
                    }
620
621
622
623
624
                    
                else {
                
                    /* Recurse on the right? */
                    if ( jj+1 < j && pivot+1 < max ) {
625
                        qid = atomic_inc( &last ) % qstack_size;
626
                        while ( atomic_cas( &qstack[qid].ready , 0 , 0 ) != 0 );
627
628
629
630
631
                        qstack[qid].i = jj+1;
                        qstack[qid].j = j;
                        qstack[qid].min = pivot+1;
                        qstack[qid].max = max;
                        qstack[qid].ready = 1;
632
633
                        if ( atomic_inc( &waiting ) >= qstack_size )
                            error( "Qstack overflow." );
634
635
636
637
638
639
640
641
642
                        }
                        
                    /* Recurse on the left? */
                    if ( jj > i  && pivot > min ) {
                        j = jj;
                        max = pivot;
                        }
                    else
                        break;
643

644
645
646
                    }
                    
                } /* loop over sub-intervals. */
647
    
648
            atomic_dec( &waiting );
649

650
651
652
            } /* main loop. */
    
        } /* parallel bit. */
653
    
654
    /* Verify sort. */
655
    /* for ( i = 1 ; i < N ; i++ )
656
        if ( ind[i-1] > ind[i] )
657
            error( "Sorting failed!" ); */
658
659
660
            
    /* Clean up. */
    free( qstack );
661

662
663
664
    }


Pedro Gonnet's avatar
Pedro Gonnet committed
665
/**
666
 * @brief Mapping function to free the sorted indices buffers.
Pedro Gonnet's avatar
Pedro Gonnet committed
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
 */

void space_map_clearsort ( struct cell *c , void *data ) {

    if ( c->sort != NULL ) {
        free( c->sort );
        c->sort = NULL;
        }

    }


/**
 * @brief Map a function to all particles in a aspace.
 *
 * @param s The #space we are working in.
683
684
 * @param fun Function pointer to apply on the cells.
 * @param data Data passed to the function fun.
Pedro Gonnet's avatar
Pedro Gonnet committed
685
686
687
688
 */
 
void space_map_parts ( struct space *s , void (*fun)( struct part *p , struct cell *c , void *data ) , void *data ) {

Pedro Gonnet's avatar
bug.    
Pedro Gonnet committed
689
    int cid = 0;
Pedro Gonnet's avatar
Pedro Gonnet committed
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708

    void rec_map ( struct cell *c ) {
    
        int k;
        
        /* No progeny? */
        if ( !c->split )
            for ( k = 0 ; k < c->count ; k++ )
                fun( &c->parts[k] , c , data );
                
        /* Otherwise, recurse. */
        else
            for ( k = 0 ; k < 8 ; k++ )
                if ( c->progeny[k] != NULL )
                    rec_map( c->progeny[k] );
                
        }
        
    /* Call the recursive function on all higher-level cells. */
709
710
711
712
713
714
715
    #pragma omp parallel shared(cid)
    {
        int mycid;
        while ( 1 ) {
            #pragma omp critical
            mycid = cid++;
            if ( mycid < s->nr_cells )
Pedro Gonnet's avatar
bug.    
Pedro Gonnet committed
716
                rec_map( &s->cells[mycid] );
717
718
719
720
            else
                break;
            }
        }
Pedro Gonnet's avatar
Pedro Gonnet committed
721
722
723
724
725
726
727
728

    }


/**
 * @brief Map a function to all particles in a aspace.
 *
 * @param s The #space we are working in.
729
 * @param full Map to all cells, including cells with sub-cells.
730
731
 * @param fun Function pointer to apply on the cells.
 * @param data Data passed to the function fun.
Pedro Gonnet's avatar
Pedro Gonnet committed
732
733
 */
 
734
735
void space_map_cells_post ( struct space *s , int full , void (*fun)( struct cell *c , void *data ) , void *data ) {

Pedro Gonnet's avatar
bug.    
Pedro Gonnet committed
736
    int cid = 0;
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754

    void rec_map ( struct cell *c ) {
    
        int k;
        
        /* Recurse. */
        if ( c->split )
            for ( k = 0 ; k < 8 ; k++ )
                if ( c->progeny[k] != NULL )
                    rec_map( c->progeny[k] );
                
        /* No progeny? */
        if ( full || !c->split )
            fun( c , data );
                
        }
        
    /* Call the recursive function on all higher-level cells. */
755
    // #pragma omp parallel shared(s,cid)
756
757
758
    {
        int mycid;
        while ( 1 ) {
759
            // #pragma omp critical
760
761
            mycid = cid++;
            if ( mycid < s->nr_cells )
Pedro Gonnet's avatar
bug.    
Pedro Gonnet committed
762
                rec_map( &s->cells[mycid] );
763
764
765
766
            else
                break;
            }
        }
767
768
769
770
771

    }


void space_map_cells_pre ( struct space *s , int full , void (*fun)( struct cell *c , void *data ) , void *data ) {
Pedro Gonnet's avatar
Pedro Gonnet committed
772

Pedro Gonnet's avatar
bug.    
Pedro Gonnet committed
773
    int cid = 0;
Pedro Gonnet's avatar
Pedro Gonnet committed
774
775
776
777
778
779

    void rec_map ( struct cell *c ) {
    
        int k;
        
        /* No progeny? */
780
        if ( full || !c->split )
Pedro Gonnet's avatar
Pedro Gonnet committed
781
782
            fun( c , data );
                
783
784
        /* Recurse. */
        if ( c->split )
Pedro Gonnet's avatar
Pedro Gonnet committed
785
786
787
788
789
790
791
            for ( k = 0 ; k < 8 ; k++ )
                if ( c->progeny[k] != NULL )
                    rec_map( c->progeny[k] );
                
        }
        
    /* Call the recursive function on all higher-level cells. */
792
    // #pragma omp parallel shared(s,cid)
Pedro Gonnet's avatar
Pedro Gonnet committed
793
794
795
    {
        int mycid;
        while ( 1 ) {
796
            // #pragma omp critical
Pedro Gonnet's avatar
Pedro Gonnet committed
797
798
            mycid = cid++;
            if ( mycid < s->nr_cells )
Pedro Gonnet's avatar
bug.    
Pedro Gonnet committed
799
                rec_map( &s->cells[mycid] );
Pedro Gonnet's avatar
Pedro Gonnet committed
800
801
802
803
            else
                break;
            }
        }
Pedro Gonnet's avatar
Pedro Gonnet committed
804
805
806
807
808
809
810
811
812
813
814
815
816

    }


/**
 * @brief Split cells that contain too many particles.
 *
 * @param s The #space we are working in.
 * @param c The #cell under consideration.
 */
 
void space_split ( struct space *s , struct cell *c ) {

Pedro Gonnet's avatar
Pedro Gonnet committed
817
    int k, count = c->count, maxdepth = 0;
818
819
820
    float h, h_max = 0.0f, dt, dt_min = c->parts[0].dt, dt_max = dt_min;
    struct cell *temp;
    struct part *p, *parts = c->parts;
821
    struct xpart *xp, *xparts = c->xparts;
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
    
    /* Check the depth. */
    if ( c->depth > s->maxdepth )
        s->maxdepth = c->depth;
    
    /* Split or let it be? */
    if ( count > space_splitsize ) {
    
        /* No longer just a leaf. */
        c->split = 1;
        
        /* Create the cell's progeny. */
        for ( k = 0 ; k < 8 ; k++ ) {
            temp = space_getcell( s );
            temp->count = 0;
            temp->loc[0] = c->loc[0];
            temp->loc[1] = c->loc[1];
            temp->loc[2] = c->loc[2];
            temp->h[0] = c->h[0]/2;
            temp->h[1] = c->h[1]/2;
            temp->h[2] = c->h[2]/2;
            temp->dmin = c->dmin/2;
            if ( k & 4 )
                temp->loc[0] += temp->h[0];
            if ( k & 2 )
                temp->loc[1] += temp->h[1];
            if ( k & 1 )
                temp->loc[2] += temp->h[2];
            temp->depth = c->depth + 1;
            temp->split = 0;
            temp->h_max = 0.0;
            temp->dx_max = 0.0;
854
            temp->nodeID = c->nodeID;
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
            temp->parent = c;
            c->progeny[k] = temp;
            }
            
        /* Split the cell data. */
        cell_split( c );
            
        /* Remove any progeny with zero parts. */
        for ( k = 0 ; k < 8 ; k++ )
            if ( c->progeny[k]->count == 0 ) {
                space_recycle( s , c->progeny[k] );
                c->progeny[k] = NULL;
                }
            else {
                space_split( s , c->progeny[k] );
Pedro Gonnet's avatar
Pedro Gonnet committed
870
871
872
873
874
                h_max = fmaxf( h_max , c->progeny[k]->h_max );
                dt_min = fminf( dt_min , c->progeny[k]->dt_min );
                dt_max = fmaxf( dt_max , c->progeny[k]->dt_max );
                if ( c->progeny[k]->maxdepth > maxdepth )
                    maxdepth = c->progeny[k]->maxdepth;
875
                }
876
877
878
879
880
881
                
        /* Set the values for this cell. */
        c->h_max = h_max;
        c->dt_min = dt_min;
        c->dt_max = dt_max;
        c->maxdepth = maxdepth;
882
883
884
885
886
887
888
889
890
                
        }
        
    /* Otherwise, collect the data for this cell. */
    else {
    
        /* Clear the progeny. */
        bzero( c->progeny , sizeof(struct cell *) * 8 );
        c->split = 0;
Pedro Gonnet's avatar
Pedro Gonnet committed
891
        c->maxdepth = c->depth;
892
893
894
895
896
        
        /* Get dt_min/dt_max. */
        
        for ( k = 0 ; k < count ; k++ ) {
            p = &parts[k];
897
            xp = &xparts[k];
Pedro Gonnet's avatar
Pedro Gonnet committed
898
899
900
            xp->x_old[0] = p->x[0];
            xp->x_old[1] = p->x[1];
            xp->x_old[2] = p->x[2];
Pedro Gonnet's avatar
Pedro Gonnet committed
901
            dt = p->dt;
902
            h = p->h;
903
904
905
906
907
908
909
910
911
912
913
914
            if ( h > h_max )
                h_max = h;
            if ( dt < dt_min )
                dt_min = dt;
            if ( dt > dt_max )
                dt_max = dt;
            }
        c->h_max = h_max;
        c->dt_min = dt_min;
        c->dt_max = dt_max;
            
        }
915
916
        
    /* Set ownership accorind to the start of the parts array. */
917
    c->owner = ( ( c->parts - s->parts ) % s->nr_parts ) * s->nr_queues / s->nr_parts;
918
919
920
921

    }


Pedro Gonnet's avatar
Pedro Gonnet committed
922
923
924
925
926
927
928
929
930
/**
 * @brief Return a used cell to the cell buffer.
 *
 * @param s The #space.
 * @param c The #cell.
 */
 
void space_recycle ( struct space *s , struct cell *c ) {

931
932
933
    /* Lock the space. */
    lock_lock( &s->lock );
    
Pedro Gonnet's avatar
Pedro Gonnet committed
934
935
936
    /* Clear the cell. */
    if ( lock_destroy( &c->lock ) != 0 )
        error( "Failed to destroy spinlock." );
937
938
939
940
        
    /* Clear this cell's sort arrays. */
    if ( c->sort != NULL )
        free( c->sort );
941
942
943
        
    /* Clear the cell data. */
    bzero( c , sizeof(struct cell) );
Pedro Gonnet's avatar
Pedro Gonnet committed
944
945
946
947
948
949
    
    /* Hook this cell into the buffer. */
    c->next = s->cells_new;
    s->cells_new = c;
    s->tot_cells -= 1;
    
950
951
952
    /* Unlock the space. */
    lock_unlock_blind( &s->lock );
    
Pedro Gonnet's avatar
Pedro Gonnet committed
953
954
955
956
957
958
959
960
961
962
963
964
965
966
    }


/**
 * @brief Get a new empty cell.
 *
 * @param s The #space.
 */
 
struct cell *space_getcell ( struct space *s ) {

    struct cell *c;
    int k;
    
967
968
969
    /* Lock the space. */
    lock_lock( &s->lock );
    
Pedro Gonnet's avatar
Pedro Gonnet committed
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
    /* Is the buffer empty? */
    if ( s->cells_new == NULL ) {
        if ( posix_memalign( (void *)&s->cells_new , 64 , space_cellallocchunk * sizeof(struct cell) ) != 0 )
            error( "Failed to allocate more cells." );
        bzero( s->cells_new , space_cellallocchunk * sizeof(struct cell) );
        for ( k = 0 ; k < space_cellallocchunk-1 ; k++ )
            s->cells_new[k].next = &s->cells_new[k+1];
        s->cells_new[ space_cellallocchunk-1 ].next = NULL;
        }

    /* Pick off the next cell. */
    c = s->cells_new;
    s->cells_new = c->next;
    s->tot_cells += 1;
    
985
986
987
    /* Unlock the space. */
    lock_unlock_blind( &s->lock );
    
Pedro Gonnet's avatar
Pedro Gonnet committed
988
    /* Init some things in the cell. */
989
990
    bzero( c , sizeof(struct cell) );
    c->nodeID = -1;
991
992
    if ( lock_init( &c->lock ) != 0 )
        error( "Failed to initialize cell spinlock." );
Pedro Gonnet's avatar
Pedro Gonnet committed
993
994
995
996
997
998
999
1000
1001
        
    return c;

    }


/**
 * @brief Split the space into cells given the array of particles.
 *
1002
 * @param s The #space to initialize.
Pedro Gonnet's avatar
Pedro Gonnet committed
1003
1004
1005
1006
 * @param dim Spatial dimensions of the domain.
 * @param parts Pointer to an array of #part.
 * @param N The number of parts in the space.
 * @param periodic flag whether the domain is periodic or not.
1007
 * @param h_max The maximal interaction radius.
Pedro Gonnet's avatar
Pedro Gonnet committed
1008
1009
1010
1011
1012
1013
1014
1015
 *
 * Makes a grid of edge length > r_max and fills the particles
 * into the respective cells. Cells containing more than #space_maxppc
 * parts with a cutoff below half the cell width are then split
 * recursively.
 */


Pedro Gonnet's avatar
Pedro Gonnet committed
1016
void space_init ( struct space *s , double dim[3] , struct part *parts , int N , int periodic , double h_max ) {
Pedro Gonnet's avatar
Pedro Gonnet committed
1017
1018
1019
1020
1021

    /* Store eveything in the space. */
    s->dim[0] = dim[0]; s->dim[1] = dim[1]; s->dim[2] = dim[2];
    s->periodic = periodic;
    s->nr_parts = N;
1022
    s->size_parts = N;
1023
    s->parts = parts;
1024
    s->cell_min = h_max;
1025
    s->nr_queues = 1;
1026
    s->size_parts_foreign = 0;
1027
    
1028
    /* Allocate the xtra parts array. */
1029
1030
    if ( posix_memalign( (void *)&s->xparts , 32 , N * sizeof(struct xpart) ) != 0 )
        error( "Failed to allocate xparts." );
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
    bzero( s->xparts , N * sizeof(struct xpart) );
    
    /* Initialize the velocities and internal energies. */
    for ( int k = 0 ; k < N ; k++ ) {
        struct part *p = &parts[k];
        struct xpart *xp = &s->xparts[k];
        xp->v_hdt[0] = p->v[0];
        xp->v_hdt[1] = p->v[1];
        xp->v_hdt[2] = p->v[2];
        xp->u_hdt = p->u;
        }
1042
        
1043
1044
1045
    /* Init the space lock. */
    if ( lock_init( &s->lock ) != 0 )
        error( "Failed to create space spin-lock." );
Pedro Gonnet's avatar
Pedro Gonnet committed
1046
    
1047
    /* Build the cells and the tasks. */
1048
    space_regrid( s , h_max );
Pedro Gonnet's avatar
Pedro Gonnet committed
1049
1050
1051
        
    }