allreduce.c 28.5 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
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
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
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
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
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
188
189
190
191
192
/* -*- Mode: C; c-basic-offset:4 ; -*- */
/*  $Id: allreduce.c,v 1.68 2006/12/09 16:42:23 gropp Exp $
 *
 *  (C) 2001 by Argonne National Laboratory.
 *      See COPYRIGHT in top-level directory.
 */

#include "mpiimpl.h"

/* -- Begin Profiling Symbol Block for routine MPI_Allreduce */
#if defined(HAVE_PRAGMA_WEAK)
#pragma weak MPI_Allreduce = PMPI_Allreduce
#elif defined(HAVE_PRAGMA_HP_SEC_DEF)
#pragma _HP_SECONDARY_DEF PMPI_Allreduce  MPI_Allreduce
#elif defined(HAVE_PRAGMA_CRI_DUP)
#pragma _CRI duplicate MPI_Allreduce as PMPI_Allreduce
#endif
/* -- End Profiling Symbol Block */

/* Define MPICH_MPI_FROM_PMPI if weak symbols are not supported to build
   the MPI routines */
#ifndef MPICH_MPI_FROM_PMPI
#undef MPI_Allreduce
#define MPI_Allreduce PMPI_Allreduce

MPI_User_function *MPIR_Op_table[] = { MPIR_MAXF, MPIR_MINF, MPIR_SUM,
                                       MPIR_PROD, MPIR_LAND,
                                       MPIR_BAND, MPIR_LOR, MPIR_BOR,
                                       MPIR_LXOR, MPIR_BXOR,
                                       MPIR_MINLOC, MPIR_MAXLOC, };

MPIR_Op_check_dtype_fn *MPIR_Op_check_dtype_table[] = {
    MPIR_MAXF_check_dtype, MPIR_MINF_check_dtype,
    MPIR_SUM_check_dtype,
    MPIR_PROD_check_dtype, MPIR_LAND_check_dtype,
    MPIR_BAND_check_dtype, MPIR_LOR_check_dtype, MPIR_BOR_check_dtype,
    MPIR_LXOR_check_dtype, MPIR_BXOR_check_dtype,
    MPIR_MINLOC_check_dtype, MPIR_MAXLOC_check_dtype, }; 


/* This is the default implementation of allreduce. The algorithm is:
   
   Algorithm: MPI_Allreduce

   For the heterogeneous case, we call MPI_Reduce followed by MPI_Bcast
   in order to meet the requirement that all processes must have the
   same result. For the homogeneous case, we use the following algorithms.


   For long messages and for builtin ops and if count >= pof2 (where
   pof2 is the nearest power-of-two less than or equal to the number
   of processes), we use Rabenseifner's algorithm (see 
   http://www.hlrs.de/organization/par/services/models/mpi/myreduce.html ).
   This algorithm implements the allreduce in two steps: first a
   reduce-scatter, followed by an allgather. A recursive-halving
   algorithm (beginning with processes that are distance 1 apart) is
   used for the reduce-scatter, and a recursive doubling 
   algorithm is used for the allgather. The non-power-of-two case is
   handled by dropping to the nearest lower power-of-two: the first
   few even-numbered processes send their data to their right neighbors
   (rank+1), and the reduce-scatter and allgather happen among the remaining
   power-of-two processes. At the end, the first few even-numbered
   processes get the result from their right neighbors.

   For the power-of-two case, the cost for the reduce-scatter is 
   lgp.alpha + n.((p-1)/p).beta + n.((p-1)/p).gamma. The cost for the
   allgather lgp.alpha + n.((p-1)/p).beta. Therefore, the
   total cost is:
   Cost = 2.lgp.alpha + 2.n.((p-1)/p).beta + n.((p-1)/p).gamma

   For the non-power-of-two case, 
   Cost = (2.floor(lgp)+2).alpha + (2.((p-1)/p) + 2).n.beta + n.(1+(p-1)/p).gamma

   
   For short messages, for user-defined ops, and for count < pof2 
   we use a recursive doubling algorithm (similar to the one in
   MPI_Allgather). We use this algorithm in the case of user-defined ops
   because in this case derived datatypes are allowed, and the user
   could pass basic datatypes on one process and derived on another as
   long as the type maps are the same. Breaking up derived datatypes
   to do the reduce-scatter is tricky. 

   Cost = lgp.alpha + n.lgp.beta + n.lgp.gamma

   Possible improvements: 

   End Algorithm: MPI_Allreduce
*/


/* not declared static because a machine-specific function may call this one 
   in some cases */
#undef FCNAME 
#define FCNAME "MPIR_Allreduce"

int MPIR_Allreduce ( 
    void *sendbuf, 
    void *recvbuf, 
    int count, 
    MPI_Datatype datatype, 
    MPI_Op op, 
    MPID_Comm *comm_ptr )
{
    int is_homogeneous;
#ifdef MPID_HAS_HETERO
    int rc;
#endif
    int        comm_size, rank, type_size;
    int        mpi_errno = MPI_SUCCESS;
    int mask, dst, is_commutative, pof2, newrank, rem, newdst, i,
        send_idx, recv_idx, last_idx, send_cnt, recv_cnt, *cnts, *disps; 
    MPI_Aint true_extent, true_lb, extent;
    void *tmp_buf;
    MPI_User_function *uop;
    MPID_Op *op_ptr;
    MPI_Comm comm;
    MPIU_THREADPRIV_DECL;
#ifdef HAVE_CXX_BINDING
    int is_cxx_uop = 0;
#endif
    MPIU_CHKLMEM_DECL(3);
    
    if (count == 0) return MPI_SUCCESS;
    comm = comm_ptr->handle;

    MPIU_THREADPRIV_GET;
    MPIR_Nest_incr();
    
    is_homogeneous = 1;
#ifdef MPID_HAS_HETERO
    if (comm_ptr->is_hetero)
        is_homogeneous = 0;
#endif
    
#ifdef MPID_HAS_HETERO
    if (!is_homogeneous) {
        /* heterogeneous. To get the same result on all processes, we
           do a reduce to 0 and then broadcast. */
        mpi_errno = NMPI_Reduce ( sendbuf, recvbuf, count, datatype,
                                  op, 0, comm );
	/* FIXME: mpi_errno is error CODE, not necessarily the error
	   class MPI_ERR_OP.  In MPICH2, we can get the error class 
	   with 
	       errorclass = mpi_errno & ERROR_CLASS_MASK;
	*/
        if (mpi_errno == MPI_ERR_OP || mpi_errno == MPI_SUCCESS) {
	    /* Allow MPI_ERR_OP since we can continue from this error */
            rc = NMPI_Bcast  ( recvbuf, count, datatype, 0, comm );
            if (rc) mpi_errno = rc;
        }
    }
    else 
#endif /* MPID_HAS_HETERO */
    {
        /* homogeneous */
        
        /* set op_errno to 0. stored in perthread structure */
        MPIU_THREADPRIV_FIELD(op_errno) = 0;

        comm_size = comm_ptr->local_size;
        rank = comm_ptr->rank;
        
        if (HANDLE_GET_KIND(op) == HANDLE_KIND_BUILTIN) {
            is_commutative = 1;
            /* get the function by indexing into the op table */
            uop = MPIR_Op_table[op%16 - 1];
        }
        else {
            MPID_Op_get_ptr(op, op_ptr);
            if (op_ptr->kind == MPID_OP_USER_NONCOMMUTE)
                is_commutative = 0;
            else
                is_commutative = 1;
#ifdef HAVE_CXX_BINDING            
            if (op_ptr->language == MPID_LANG_CXX) {
                uop = (MPI_User_function *) op_ptr->function.c_function;
		is_cxx_uop = 1;
	    }
	    else
#endif
            if ((op_ptr->language == MPID_LANG_C))
                uop = (MPI_User_function *) op_ptr->function.c_function;
            else
                uop = (MPI_User_function *) op_ptr->function.f77_function;
        }
        
        /* need to allocate temporary buffer to store incoming data*/
        mpi_errno = NMPI_Type_get_true_extent(datatype, &true_lb,
                                              &true_extent);
	MPIU_ERR_CHKANDJUMP((mpi_errno), mpi_errno, MPI_ERR_OTHER, "**fail");
        MPID_Datatype_get_extent_macro(datatype, extent);

193
        MPID_Ensure_Aint_fits_in_pointer(count * MPIR_MAX(extent, true_extent));
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
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
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
        MPIU_CHKLMEM_MALLOC(tmp_buf, void *, count*(MPIR_MAX(extent,true_extent)), mpi_errno, "temporary buffer");
	
        /* adjust for potential negative lower bound in datatype */
        tmp_buf = (void *)((char*)tmp_buf - true_lb);
        
        /* copy local data into recvbuf */
        if (sendbuf != MPI_IN_PLACE) {
            mpi_errno = MPIR_Localcopy(sendbuf, count, datatype, recvbuf,
                                       count, datatype);
            MPIU_ERR_CHKANDJUMP((mpi_errno), mpi_errno, MPI_ERR_OTHER, "**fail");
        }

        MPID_Datatype_get_size_macro(datatype, type_size);

        /* find nearest power-of-two less than or equal to comm_size */
        pof2 = 1;
        while (pof2 <= comm_size) pof2 <<= 1;
        pof2 >>=1;

        rem = comm_size - pof2;

        /* In the non-power-of-two case, all even-numbered
           processes of rank < 2*rem send their data to
           (rank+1). These even-numbered processes no longer
           participate in the algorithm until the very end. The
           remaining processes form a nice power-of-two. */
        
        /* check if multiple threads are calling this collective function */
        MPIDU_ERR_CHECK_MULTIPLE_THREADS_ENTER( comm_ptr );

        if (rank < 2*rem) {
            if (rank % 2 == 0) { /* even */
                mpi_errno = MPIC_Send(recvbuf, count, 
                                      datatype, rank+1,
                                      MPIR_ALLREDUCE_TAG, comm);
		MPIU_ERR_CHKANDJUMP((mpi_errno), mpi_errno, MPI_ERR_OTHER, "**fail");
                
                /* temporarily set the rank to -1 so that this
                   process does not pariticipate in recursive
                   doubling */
                newrank = -1; 
            }
            else { /* odd */
                mpi_errno = MPIC_Recv(tmp_buf, count, 
                                      datatype, rank-1,
                                      MPIR_ALLREDUCE_TAG, comm,
                                      MPI_STATUS_IGNORE);
		MPIU_ERR_CHKANDJUMP((mpi_errno), mpi_errno, MPI_ERR_OTHER, "**fail");
                
                /* do the reduction on received data. since the
                   ordering is right, it doesn't matter whether
                   the operation is commutative or not. */
#ifdef HAVE_CXX_BINDING
                if (is_cxx_uop) {
                    (*MPIR_Process.cxx_call_op_fn)( tmp_buf, recvbuf, 
                                                    count,
                                                    datatype,
                                                    uop ); 
                }
                else 
#endif
                    (*uop)(tmp_buf, recvbuf, &count, &datatype);
                
                /* change the rank */
                newrank = rank / 2;
            }
        }
        else  /* rank >= 2*rem */
            newrank = rank - rem;
        
        /* If op is user-defined or count is less than pof2, use
           recursive doubling algorithm. Otherwise do a reduce-scatter
           followed by allgather. (If op is user-defined,
           derived datatypes are allowed and the user could pass basic
           datatypes on one process and derived on another as long as
           the type maps are the same. Breaking up derived
           datatypes to do the reduce-scatter is tricky, therefore
           using recursive doubling in that case.) */

        if (newrank != -1) {
            if ((count*type_size <= MPIR_ALLREDUCE_SHORT_MSG) ||
                (HANDLE_GET_KIND(op) != HANDLE_KIND_BUILTIN) ||  
                (count < pof2)) { /* use recursive doubling */
                mask = 0x1;
                while (mask < pof2) {
                    newdst = newrank ^ mask;
                    /* find real rank of dest */
                    dst = (newdst < rem) ? newdst*2 + 1 : newdst + rem;

                    /* Send the most current data, which is in recvbuf. Recv
                       into tmp_buf */ 
                    mpi_errno = MPIC_Sendrecv(recvbuf, count, datatype, 
                                              dst, MPIR_ALLREDUCE_TAG, tmp_buf,
                                              count, datatype, dst,
                                              MPIR_ALLREDUCE_TAG, comm,
                                              MPI_STATUS_IGNORE);
		    MPIU_ERR_CHKANDJUMP((mpi_errno), mpi_errno, MPI_ERR_OTHER, "**fail");
                    
                    /* tmp_buf contains data received in this step.
                       recvbuf contains data accumulated so far */
                    
                    if (is_commutative  || (dst < rank)) {
                        /* op is commutative OR the order is already right */
#ifdef HAVE_CXX_BINDING
                        if (is_cxx_uop) {
                            (*MPIR_Process.cxx_call_op_fn)( tmp_buf, recvbuf, 
                                                            count,
                                                            datatype,
                                                            uop ); 
                        }
                        else 
#endif
                            (*uop)(tmp_buf, recvbuf, &count, &datatype);
                    }
                    else {
                        /* op is noncommutative and the order is not right */
#ifdef HAVE_CXX_BINDING
                        if (is_cxx_uop) {
                            (*MPIR_Process.cxx_call_op_fn)( recvbuf, tmp_buf, 
                                                            count,
                                                            datatype,
                                                            uop ); 
                        }
                        else 
#endif
                            (*uop)(recvbuf, tmp_buf, &count, &datatype);
                        
                        /* copy result back into recvbuf */
                        mpi_errno = MPIR_Localcopy(tmp_buf, count, datatype,
                                                   recvbuf, count, datatype);
			MPIU_ERR_CHKANDJUMP((mpi_errno), mpi_errno, MPI_ERR_OTHER, "**fail");
                    }
                    mask <<= 1;
                }
            }

            else {

                /* do a reduce-scatter followed by allgather */

                /* for the reduce-scatter, calculate the count that
                   each process receives and the displacement within
                   the buffer */

		MPIU_CHKLMEM_MALLOC(cnts, int *, pof2*sizeof(int), mpi_errno, "counts");
		MPIU_CHKLMEM_MALLOC(disps, int *, pof2*sizeof(int), mpi_errno, "displacements");

                for (i=0; i<(pof2-1); i++) 
                    cnts[i] = count/pof2;
                cnts[pof2-1] = count - (count/pof2)*(pof2-1);

                disps[0] = 0;
                for (i=1; i<pof2; i++)
                    disps[i] = disps[i-1] + cnts[i-1];

                mask = 0x1;
                send_idx = recv_idx = 0;
                last_idx = pof2;
                while (mask < pof2) {
                    newdst = newrank ^ mask;
                    /* find real rank of dest */
                    dst = (newdst < rem) ? newdst*2 + 1 : newdst + rem;

                    send_cnt = recv_cnt = 0;
                    if (newrank < newdst) {
                        send_idx = recv_idx + pof2/(mask*2);
                        for (i=send_idx; i<last_idx; i++)
                            send_cnt += cnts[i];
                        for (i=recv_idx; i<send_idx; i++)
                            recv_cnt += cnts[i];
                    }
                    else {
                        recv_idx = send_idx + pof2/(mask*2);
                        for (i=send_idx; i<recv_idx; i++)
                            send_cnt += cnts[i];
                        for (i=recv_idx; i<last_idx; i++)
                            recv_cnt += cnts[i];
                    }

/*                    printf("Rank %d, send_idx %d, recv_idx %d, send_cnt %d, recv_cnt %d, last_idx %d\n", newrank, send_idx, recv_idx,
                           send_cnt, recv_cnt, last_idx);
                           */
                    /* Send data from recvbuf. Recv into tmp_buf */ 
                    mpi_errno = MPIC_Sendrecv((char *) recvbuf +
                                              disps[send_idx]*extent,
                                              send_cnt, datatype,  
                                              dst, MPIR_ALLREDUCE_TAG, 
                                              (char *) tmp_buf +
                                              disps[recv_idx]*extent,
                                              recv_cnt, datatype, dst,
                                              MPIR_ALLREDUCE_TAG, comm,
                                              MPI_STATUS_IGNORE);
		    MPIU_ERR_CHKANDJUMP((mpi_errno), mpi_errno, MPI_ERR_OTHER, "**fail");
                    
                    /* tmp_buf contains data received in this step.
                       recvbuf contains data accumulated so far */
                    
                    /* This algorithm is used only for predefined ops
                       and predefined ops are always commutative. */

		    (*uop)((char *) tmp_buf + disps[recv_idx]*extent,
			   (char *) recvbuf + disps[recv_idx]*extent, 
			   &recv_cnt, &datatype);
                    
                    /* update send_idx for next iteration */
                    send_idx = recv_idx;
                    mask <<= 1;

                    /* update last_idx, but not in last iteration
                       because the value is needed in the allgather
                       step below. */
                    if (mask < pof2)
                        last_idx = recv_idx + pof2/mask;
                }

                /* now do the allgather */

                mask >>= 1;
                while (mask > 0) {
                    newdst = newrank ^ mask;
                    /* find real rank of dest */
                    dst = (newdst < rem) ? newdst*2 + 1 : newdst + rem;

                    send_cnt = recv_cnt = 0;
                    if (newrank < newdst) {
                        /* update last_idx except on first iteration */
                        if (mask != pof2/2)
                            last_idx = last_idx + pof2/(mask*2);

                        recv_idx = send_idx + pof2/(mask*2);
                        for (i=send_idx; i<recv_idx; i++)
                            send_cnt += cnts[i];
                        for (i=recv_idx; i<last_idx; i++)
                            recv_cnt += cnts[i];
                    }
                    else {
                        recv_idx = send_idx - pof2/(mask*2);
                        for (i=send_idx; i<last_idx; i++)
                            send_cnt += cnts[i];
                        for (i=recv_idx; i<send_idx; i++)
                            recv_cnt += cnts[i];
                    }

                    mpi_errno = MPIC_Sendrecv((char *) recvbuf +
                                              disps[send_idx]*extent,
                                              send_cnt, datatype,  
                                              dst, MPIR_ALLREDUCE_TAG, 
                                              (char *) recvbuf +
                                              disps[recv_idx]*extent,
                                              recv_cnt, datatype, dst,
                                              MPIR_ALLREDUCE_TAG, comm,
                                              MPI_STATUS_IGNORE);
		    MPIU_ERR_CHKANDJUMP((mpi_errno), mpi_errno, MPI_ERR_OTHER, "**fail");

                    if (newrank > newdst) send_idx = recv_idx;

                    mask >>= 1;
                }
            }
        }

        /* In the non-power-of-two case, all odd-numbered
           processes of rank < 2*rem send the result to
           (rank-1), the ranks who didn't participate above. */
        if (rank < 2*rem) {
            if (rank % 2)  /* odd */
                mpi_errno = MPIC_Send(recvbuf, count, 
                                      datatype, rank-1,
                                      MPIR_ALLREDUCE_TAG, comm);
            else  /* even */
                mpi_errno = MPIC_Recv(recvbuf, count,
                                      datatype, rank+1,
                                      MPIR_ALLREDUCE_TAG, comm,
                                      MPI_STATUS_IGNORE); 
            MPIU_ERR_CHKANDJUMP((mpi_errno), mpi_errno, MPI_ERR_OTHER, "**fail");
        }

        /* check if multiple threads are calling this collective function */
        MPIDU_ERR_CHECK_MULTIPLE_THREADS_EXIT( comm_ptr );

        if (MPIU_THREADPRIV_FIELD(op_errno)) 
	    mpi_errno = MPIU_THREADPRIV_FIELD(op_errno);
    }

  fn_exit:
    MPIU_CHKLMEM_FREEALL();
    MPIR_Nest_decr();
    return (mpi_errno);

  fn_fail:
    goto fn_exit;
}


/* not declared static because a machine-specific function may call this one 
   in some cases */
#undef FCNAME
#define FCNAME "MPIR_Allreduce_inter"
int MPIR_Allreduce_inter ( 
    void *sendbuf, 
    void *recvbuf, 
    int count, 
    MPI_Datatype datatype, 
    MPI_Op op, 
    MPID_Comm *comm_ptr )
{
/* Intercommunicator Allreduce.
   We first do an intercommunicator reduce to rank 0 on left group,
   then an intercommunicator reduce to rank 0 on right group, followed
   by local intracommunicator broadcasts in each group.

   We don't do local reduces first and then intercommunicator
   broadcasts because it would require allocation of a temporary buffer. 
*/
    int rank, mpi_errno, root;
    MPID_Comm *newcomm_ptr = NULL;
    MPIU_THREADPRIV_DECL;

    MPIU_THREADPRIV_GET;

    MPIR_Nest_incr();
    
    rank = comm_ptr->rank;

    /* first do a reduce from right group to rank 0 in left group,
       then from left group to rank 0 in right group*/
    if (comm_ptr->is_low_group) {
        /* reduce from right group to rank 0*/
        root = (rank == 0) ? MPI_ROOT : MPI_PROC_NULL;
        mpi_errno = MPIR_Reduce_inter(sendbuf, recvbuf, count, datatype, op,
				      root, comm_ptr);
	MPIU_ERR_CHKANDJUMP((mpi_errno), mpi_errno, MPI_ERR_OTHER, "**fail");

        /* reduce to rank 0 of right group */
        root = 0;
        mpi_errno = MPIR_Reduce_inter(sendbuf, recvbuf, count, datatype, op,
				      root, comm_ptr);
	MPIU_ERR_CHKANDJUMP((mpi_errno), mpi_errno, MPI_ERR_OTHER, "**fail");
    }
    else {
        /* reduce to rank 0 of left group */
        root = 0;
        mpi_errno = MPIR_Reduce_inter(sendbuf, recvbuf, count, datatype, op,
				      root, comm_ptr);
	MPIU_ERR_CHKANDJUMP((mpi_errno), mpi_errno, MPI_ERR_OTHER, "**fail");

        /* reduce from right group to rank 0 */
        root = (rank == 0) ? MPI_ROOT : MPI_PROC_NULL;
        mpi_errno = MPIR_Reduce_inter(sendbuf, recvbuf, count, datatype, op,
				      root, comm_ptr);
	MPIU_ERR_CHKANDJUMP((mpi_errno), mpi_errno, MPI_ERR_OTHER, "**fail");
    }

    /* Get the local intracommunicator */
    if (!comm_ptr->local_comm)
	MPIR_Setup_intercomm_localcomm( comm_ptr );

    newcomm_ptr = comm_ptr->local_comm;

    mpi_errno = MPIR_Bcast(recvbuf, count, datatype, 0, newcomm_ptr);
    MPIU_ERR_CHKANDJUMP((mpi_errno), mpi_errno, MPI_ERR_OTHER, "**fail");

  fn_exit:
    MPIR_Nest_decr();
    return mpi_errno;

  fn_fail:
    goto fn_exit;
}

#endif

#undef FUNCNAME
#define FUNCNAME MPI_Allreduce
#undef FCNAME

/*@
MPI_Allreduce - Combines values from all processes and distributes the result
                back to all processes

Input Parameters:
+ sendbuf - starting address of send buffer (choice) 
. count - number of elements in send buffer (integer) 
. datatype - data type of elements of send buffer (handle) 
. op - operation (handle) 
- comm - communicator (handle) 

Output Parameter:
. recvbuf - starting address of receive buffer (choice) 

.N ThreadSafe

.N Fortran

.N collops

.N Errors
.N MPI_ERR_BUFFER
.N MPI_ERR_COUNT
.N MPI_ERR_TYPE
.N MPI_ERR_OP
.N MPI_ERR_COMM
@*/
int MPI_Allreduce ( void *sendbuf, void *recvbuf, int count, 
		    MPI_Datatype datatype, MPI_Op op, MPI_Comm comm )
{
    static const char FCNAME[] = "MPI_Allreduce";
    int mpi_errno = MPI_SUCCESS;
    MPID_Comm *comm_ptr = NULL;
    MPID_MPI_STATE_DECL(MPID_STATE_MPI_ALLREDUCE);

    MPIR_ERRTEST_INITIALIZED_ORDIE();
    
607
    MPIU_THREAD_CS_ENTER(ALLFUNC,);
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
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
    MPID_MPI_COLL_FUNC_ENTER(MPID_STATE_MPI_ALLREDUCE);

    /* Validate parameters, especially handles needing to be converted */
#   ifdef HAVE_ERROR_CHECKING
    {
        MPID_BEGIN_ERROR_CHECKS;
        {
	    MPIR_ERRTEST_COMM(comm, mpi_errno);
            if (mpi_errno != MPI_SUCCESS) goto fn_fail;
	}
        MPID_END_ERROR_CHECKS;
    }
#   endif /* HAVE_ERROR_CHECKING */

    /* Convert MPI object handles to object pointers */
    MPID_Comm_get_ptr( comm, comm_ptr );

    /* Validate parameters and objects (post conversion) */
#   ifdef HAVE_ERROR_CHECKING
    {
        MPID_BEGIN_ERROR_CHECKS;
        {
            MPID_Datatype *datatype_ptr = NULL;
            MPID_Op *op_ptr = NULL;

            MPID_Comm_valid_ptr( comm_ptr, mpi_errno );
            if (mpi_errno != MPI_SUCCESS) goto fn_fail;
	    MPIR_ERRTEST_COUNT(count, mpi_errno);
	    MPIR_ERRTEST_DATATYPE(datatype, "datatype", mpi_errno);
	    MPIR_ERRTEST_OP(op, mpi_errno);
	    
            if (HANDLE_GET_KIND(datatype) != HANDLE_KIND_BUILTIN) {
                MPID_Datatype_get_ptr(datatype, datatype_ptr);
                MPID_Datatype_valid_ptr( datatype_ptr, mpi_errno );
                MPID_Datatype_committed_ptr( datatype_ptr, mpi_errno );
            }

	    if (comm_ptr->comm_kind == MPID_INTERCOMM)
                MPIR_ERRTEST_SENDBUF_INPLACE(sendbuf, count, mpi_errno);
            
            if (sendbuf != MPI_IN_PLACE) 
                MPIR_ERRTEST_USERBUFFER(sendbuf,count,datatype,mpi_errno);

            MPIR_ERRTEST_RECVBUF_INPLACE(recvbuf, count, mpi_errno);
	    MPIR_ERRTEST_USERBUFFER(recvbuf,count,datatype,mpi_errno);

	    if (mpi_errno != MPI_SUCCESS) goto fn_fail;

            if (HANDLE_GET_KIND(op) != HANDLE_KIND_BUILTIN) {
                MPID_Op_get_ptr(op, op_ptr);
                MPID_Op_valid_ptr( op_ptr, mpi_errno );
            }
            if (HANDLE_GET_KIND(op) == HANDLE_KIND_BUILTIN) {
                mpi_errno = 
                    ( * MPIR_Op_check_dtype_table[op%16 - 1] )(datatype); 
            }
	    if (mpi_errno != MPI_SUCCESS) goto fn_fail;
        }

	MPIR_ERRTEST_ALIAS_COLL(sendbuf, recvbuf, mpi_errno);
	if (mpi_errno != MPI_SUCCESS) goto fn_fail;

        MPID_END_ERROR_CHECKS;
    }
#   endif /* HAVE_ERROR_CHECKING */

    /* ... body of routine ...  */

    if (comm_ptr->coll_fns != NULL && comm_ptr->coll_fns->Allreduce != NULL)
    {
	mpi_errno = comm_ptr->coll_fns->Allreduce(sendbuf, recvbuf, count,
                                              datatype, op, comm_ptr);
    }
    else
    {
683
        if (comm_ptr->comm_kind == MPID_INTRACOMM) {
684
            /* intracommunicator */
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
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
#if USE_SMP_COLLECTIVES
	    MPID_Op *op_ptr;
	    int is_commutative; 

	    /* is the op commutative? We do SMP optimizations only if it is. */ 
	    if (HANDLE_GET_KIND(op) == HANDLE_KIND_BUILTIN)
		is_commutative = 1;
	    else {
		MPID_Op_get_ptr(op, op_ptr);
		is_commutative = (op_ptr->kind == MPID_OP_USER_NONCOMMUTE) ? 0 : 1;
	    }

            if (MPIR_Comm_is_node_aware(comm_ptr) && is_commutative) {
		/* on each node, do a reduce to the local root */ 
                if (comm_ptr->node_comm != NULL)
                {
		    /* take care of the MPI_IN_PLACE case. For reduce, 
		        MPI_IN_PLACE is specified only on the root; 
                        for allreduce it is specified on all processes. */

		    if ((sendbuf == MPI_IN_PLACE) && (comm_ptr->node_comm->rank != 0)) {
			/* IN_PLACE and not root of reduce. Data supplied to this
			allreduce is in recvbuf. Pass that as the sendbuf to reduce. */
			
			mpi_errno = MPIR_Reduce(recvbuf, NULL, count, datatype,
						op, 0, comm_ptr->node_comm);
		    }
		    else {
			mpi_errno = MPIR_Reduce(sendbuf, recvbuf, count, datatype,
						op, 0, comm_ptr->node_comm);
		    }
		    if (mpi_errno) goto fn_fail;
                }

                /* now do an IN_PLACE allreduce among the local roots of all nodes */
                if (comm_ptr->node_roots_comm != NULL) {
                    mpi_errno = MPIR_Allreduce(MPI_IN_PLACE, recvbuf, count, datatype,
					       op, comm_ptr->node_roots_comm);
                    if (mpi_errno) goto fn_fail;
                }

		/* now broadcast the result among local processes */
                if (comm_ptr->node_comm != NULL) {
		    MPIU_THREADPRIV_DECL;
		    MPIU_THREADPRIV_GET;
		    
		    MPIR_Nest_incr();
                    mpi_errno = MPIR_Bcast(recvbuf, count, datatype, 0, comm_ptr->node_comm);
		    MPIR_Nest_decr();
		}
            }
            else {
                mpi_errno = MPIR_Allreduce(sendbuf, recvbuf, count, datatype,
                                       op, comm_ptr);
            }
#else
741
742
            mpi_errno = MPIR_Allreduce(sendbuf, recvbuf, count, datatype,
                                       op, comm_ptr); 
743
744
#endif
	}
745
746
747
748
749
750
751
752
753
754
755
756
757
        else {
            /* intercommunicator */
            mpi_errno = MPIR_Allreduce_inter(sendbuf, recvbuf, count,
					     datatype, op, comm_ptr);       
        }
    }

    if (mpi_errno != MPI_SUCCESS) goto fn_fail;

    /* ... end of body of routine ... */
    
  fn_exit:
    MPID_MPI_COLL_FUNC_EXIT(MPID_STATE_MPI_ALLREDUCE);
758
    MPIU_THREAD_CS_EXIT(ALLFUNC,);
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
    return mpi_errno;

  fn_fail:
    /* --BEGIN ERROR HANDLING-- */
#   ifdef HAVE_ERROR_CHECKING
    {
	mpi_errno = MPIR_Err_create_code(
	    mpi_errno, MPIR_ERR_RECOVERABLE, FCNAME, __LINE__, MPI_ERR_OTHER, "**mpi_allreduce",
	    "**mpi_allreduce %p %p %d %D %O %C", sendbuf, recvbuf, count, datatype, op, comm);
    }
#   endif
    mpi_errno = MPIR_Err_return_comm( comm_ptr, FCNAME, mpi_errno );
    goto fn_exit;
    /* --END ERROR HANDLING-- */
}