red_scat_block.c 48.4 KB
Newer Older
1
/* -*- Mode: C; c-basic-offset:4 ; indent-tabs-mode:nil ; -*- */
2
3
/*
 *
Pavan Balaji's avatar
Pavan Balaji committed
4
 *  (C) 2009 by Argonne National Laboratory.
5
6
7
 *      See COPYRIGHT in top-level directory.
 */

8
9
10
11
12
13

/* This implementation of MPI_Reduce_scatter_block was obtained by taking
   the implementation of MPI_Reduce_scatter from red_scat.c and replacing 
   recvcnts[i] with recvcount everywhere. */


14
#include "mpiimpl.h"
15
#include "collutil.h"
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31

/* -- Begin Profiling Symbol Block for routine MPI_Reduce_scatter_block */
#if defined(HAVE_PRAGMA_WEAK)
#pragma weak MPI_Reduce_scatter_block = PMPI_Reduce_scatter_block
#elif defined(HAVE_PRAGMA_HP_SEC_DEF)
#pragma _HP_SECONDARY_DEF PMPI_Reduce_scatter_block  MPI_Reduce_scatter_block
#elif defined(HAVE_PRAGMA_CRI_DUP)
#pragma _CRI duplicate MPI_Reduce_scatter_block as PMPI_Reduce_scatter_block
#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_Reduce_scatter_block
#define MPI_Reduce_scatter_block PMPI_Reduce_scatter_block
32
33
34
35
36


/* FIXME should we be checking the op_errno here? */

/* Implements the reduce-scatter butterfly algorithm described in J. L. Traff's
37
38
 * "An Improved Algorithm for (Non-commutative) Reduce-Scatter with an 
 * Application"
39
40
41
42
43
44
45
 * from EuroPVM/MPI 2005.  This function currently only implements support for
 * the power-of-2 case. */
#undef FUNCNAME
#define FUNCNAME MPIR_Reduce_scatter_block_noncomm
#undef FCNAME
#define FCNAME MPIU_QUOTE(FUNCNAME)
static int MPIR_Reduce_scatter_block_noncomm (
46
    const void *sendbuf,
47
48
49
50
    void *recvbuf,
    int recvcount,
    MPI_Datatype datatype,
    MPI_Op op,
51
52
    MPID_Comm *comm_ptr,
    int *errflag )
53
54
{
    int mpi_errno = MPI_SUCCESS;
55
    int mpi_errno_ret = MPI_SUCCESS;
56
57
58
59
60
61
62
    int comm_size = comm_ptr->local_size;
    int rank = comm_ptr->rank;
    int pof2;
    int log2_comm_size;
    int i, k;
    int recv_offset, send_offset;
    int block_size, total_count, size;
63
    MPI_Aint true_extent, true_lb;
64
65
66
67
68
69
70
    int buf0_was_inout;
    void *tmp_buf0;
    void *tmp_buf1;
    void *result_ptr;
    MPI_Comm comm = comm_ptr->handle;
    MPIU_CHKLMEM_DECL(3);

71
    MPIR_Type_get_true_extent_impl(datatype, &true_lb, &true_extent);
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

    pof2 = 1;
    log2_comm_size = 0;
    while (pof2 < comm_size) {
        pof2 <<= 1;
        ++log2_comm_size;
    }

    /* begin error checking */
    MPIU_Assert(pof2 == comm_size); /* FIXME this version only works for power of 2 procs */
    /* end error checking */

    /* size of a block (count of datatype per block, NOT bytes per block) */
    block_size = recvcount;
    total_count = block_size * comm_size;

    MPIU_CHKLMEM_MALLOC(tmp_buf0, void *, true_extent * total_count, mpi_errno, "tmp_buf0");
    MPIU_CHKLMEM_MALLOC(tmp_buf1, void *, true_extent * total_count, mpi_errno, "tmp_buf1");
    /* adjust for potential negative lower bound in datatype */
    tmp_buf0 = (void *)((char*)tmp_buf0 - true_lb);
    tmp_buf1 = (void *)((char*)tmp_buf1 - true_lb);

    /* Copy our send data to tmp_buf0.  We do this one block at a time and
       permute the blocks as we go according to the mirror permutation. */
    for (i = 0; i < comm_size; ++i) {
97
98
        mpi_errno = MPIR_Localcopy((char *)(sendbuf == MPI_IN_PLACE ? (const void *)recvbuf : sendbuf) + (i * true_extent * block_size),
                                   block_size, datatype,
99
                                   (char *)tmp_buf0 + (MPIU_Mirror_permutation(i, log2_comm_size) * true_extent * block_size), block_size, datatype);
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
        if (mpi_errno) MPIU_ERR_POP(mpi_errno);
    }
    buf0_was_inout = 1;

    send_offset = 0;
    recv_offset = 0;
    size = total_count;
    for (k = 0; k < log2_comm_size; ++k) {
        /* use a double-buffering scheme to avoid local copies */
        char *incoming_data = (buf0_was_inout ? tmp_buf1 : tmp_buf0);
        char *outgoing_data = (buf0_was_inout ? tmp_buf0 : tmp_buf1);
        int peer = rank ^ (0x1 << k);
        size /= 2;

        if (rank > peer) {
            /* we have the higher rank: send top half, recv bottom half */
            recv_offset += size;
        }
        else {
            /* we have the lower rank: recv top half, send bottom half */
            send_offset += size;
        }

123
        mpi_errno = MPIC_Sendrecv(outgoing_data + send_offset*true_extent,
124
125
126
127
                                     size, datatype, peer, MPIR_REDUCE_SCATTER_BLOCK_TAG,
                                     incoming_data + recv_offset*true_extent,
                                     size, datatype, peer, MPIR_REDUCE_SCATTER_BLOCK_TAG,
                                     comm, MPI_STATUS_IGNORE, errflag);
128
129
        if (mpi_errno) {
            /* for communication errors, just record the error but continue */
130
            *errflag = TRUE;
131
132
133
            MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
            MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
        }
134
135
136
137
        /* always perform the reduction at recv_offset, the data at send_offset
           is now our peer's responsibility */
        if (rank > peer) {
            /* higher ranked value so need to call op(received_data, my_data) */
138
139
            mpi_errno = MPIR_Reduce_local_impl(
                     incoming_data + recv_offset*true_extent,
140
                     outgoing_data + recv_offset*true_extent,
141
                     size, datatype, op);
142
            if (mpi_errno) MPIU_ERR_POP(mpi_errno);
143
144
145
        }
        else {
            /* lower ranked value so need to call op(my_data, received_data) */
146
147
            mpi_errno = MPIR_Reduce_local_impl(
                     outgoing_data + recv_offset*true_extent,
148
                     incoming_data + recv_offset*true_extent,
149
                     size, datatype, op);
150
            if (mpi_errno) MPIU_ERR_POP(mpi_errno);
151
152
153
154
155
156
157
158
159
160
161
162
163
164
            buf0_was_inout = !buf0_was_inout;
        }

        /* the next round of send/recv needs to happen within the block (of size
           "size") that we just received and reduced */
        send_offset = recv_offset;
    }

    MPIU_Assert(size == recvcount);

    /* copy the reduced data to the recvbuf */
    result_ptr = (char *)(buf0_was_inout ? tmp_buf0 : tmp_buf1) + recv_offset * true_extent;
    mpi_errno = MPIR_Localcopy(result_ptr, size, datatype,
                               recvbuf, size, datatype);
165
166
    if (mpi_errno) MPIU_ERR_POP(mpi_errno);
    
167
168
fn_exit:
    MPIU_CHKLMEM_FREEALL();
169
    /* --BEGIN ERROR HANDLING-- */
170
171
    if (mpi_errno_ret)
        mpi_errno = mpi_errno_ret;
172
173
    else if (*errflag)
        MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**coll_fail");
174
    /* --END ERROR HANDLING-- */
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
    return mpi_errno;
fn_fail:
    goto fn_exit;
}

/* This is the default implementation of reduce_scatter. The algorithm is:

   Algorithm: MPI_Reduce_scatter

   If the operation is commutative, for short and medium-size
   messages, we use a recursive-halving
   algorithm in which the first p/2 processes send the second n/2 data
   to their counterparts in the other half and receive the first n/2
   data from them. This procedure continues recursively, halving the
   data communicated at each step, for a total of lgp steps. If the
   number of processes is not a power-of-two, we convert it to the
   nearest lower power-of-two by having the first few even-numbered
   processes send their data to the neighboring odd-numbered process
   at (rank+1). Those odd-numbered processes compute the result for
   their left neighbor as well in the recursive halving algorithm, and
   then at  the end send the result back to the processes that didn't
   participate.
   Therefore, if p is a power-of-two,
   Cost = lgp.alpha + n.((p-1)/p).beta + n.((p-1)/p).gamma
   If p is not a power-of-two,
   Cost = (floor(lgp)+2).alpha + n.(1+(p-1+n)/p).beta + n.(1+(p-1)/p).gamma
   The above cost in the non power-of-two case is approximate because
   there is some imbalance in the amount of work each process does
   because some processes do the work of their neighbors as well.

   For commutative operations and very long messages we use
   we use a pairwise exchange algorithm similar to
   the one used in MPI_Alltoall. At step i, each process sends n/p
   amount of data to (rank+i) and receives n/p amount of data from
   (rank-i).
   Cost = (p-1).alpha + n.((p-1)/p).beta + n.((p-1)/p).gamma


   If the operation is not commutative, we do the following:

   We use a recursive doubling algorithm, which
   takes lgp steps. At step 1, processes exchange (n-n/p) amount of
   data; at step 2, (n-2n/p) amount of data; at step 3, (n-4n/p)
   amount of data, and so forth.

   Cost = lgp.alpha + n.(lgp-(p-1)/p).beta + n.(lgp-(p-1)/p).gamma

   Possible improvements:

   End Algorithm: MPI_Reduce_scatter
*/

#undef FUNCNAME
228
#define FUNCNAME MPIR_Reduce_scatter_block_intra
229
230
#undef FCNAME
#define FCNAME MPIU_QUOTE(FUNCNAME)
231

232
/* not declared static because a machine-specific function may call this one in some cases */
233
int MPIR_Reduce_scatter_block_intra ( 
234
    const void *sendbuf, 
235
236
237
238
    void *recvbuf, 
    int recvcount, 
    MPI_Datatype datatype, 
    MPI_Op op, 
239
240
    MPID_Comm *comm_ptr,
    int *errflag )
241
242
243
244
245
{
    int   rank, comm_size, i;
    MPI_Aint extent, true_extent, true_lb; 
    int  *disps;
    void *tmp_recvbuf, *tmp_results;
246
247
    int mpi_errno = MPI_SUCCESS;
    int mpi_errno_ret = MPI_SUCCESS;
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
    int type_size, dis[2], blklens[2], total_count, nbytes, src, dst;
    int mask, dst_tree_root, my_tree_root, j, k;
    int *newcnts, *newdisps, rem, newdst, send_idx, recv_idx,
        last_idx, send_cnt, recv_cnt;
    int pof2, old_i, newrank, received;
    MPI_Datatype sendtype, recvtype;
    int nprocs_completed, tmp_mask, tree_root, is_commutative;
    MPID_Op *op_ptr;
    MPI_Comm comm;
    MPIU_THREADPRIV_DECL;
    MPIU_CHKLMEM_DECL(5);

    comm = comm_ptr->handle;
    comm_size = comm_ptr->local_size;
    rank = comm_ptr->rank;

    /* set op_errno to 0. stored in perthread structure */
    MPIU_THREADPRIV_GET;
    MPIU_THREADPRIV_FIELD(op_errno) = 0;

    if (recvcount == 0) {
        goto fn_exit;
    }

    MPID_Datatype_get_extent_macro(datatype, extent);
273
    MPIR_Type_get_true_extent_impl(datatype, &true_lb, &true_extent);
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
    
    if (HANDLE_GET_KIND(op) == HANDLE_KIND_BUILTIN) {
        is_commutative = 1;
    }
    else {
        MPID_Op_get_ptr(op, op_ptr);
        if (op_ptr->kind == MPID_OP_USER_NONCOMMUTE)
            is_commutative = 0;
        else
            is_commutative = 1;
    }

    MPIU_CHKLMEM_MALLOC(disps, int *, comm_size * sizeof(int), mpi_errno, "disps");

    total_count = comm_size*recvcount;
    for (i=0; i<comm_size; i++) {
        disps[i] = i*recvcount;
    }
    
    MPID_Datatype_get_size_macro(datatype, type_size);
    nbytes = total_count * type_size;
    
    /* check if multiple threads are calling this collective function */
    MPIDU_ERR_CHECK_MULTIPLE_THREADS_ENTER( comm_ptr );

    /* total_count*extent eventually gets malloced. it isn't added to
     * a user-passed in buffer */
    MPID_Ensure_Aint_fits_in_pointer(total_count * MPIR_MAX(true_extent, extent));

303
    if ((is_commutative) && (nbytes < MPIR_CVAR_REDSCAT_COMMUTATIVE_LONG_MSG_SIZE)) {
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
        /* commutative and short. use recursive halving algorithm */

        /* allocate temp. buffer to receive incoming data */
        MPIU_CHKLMEM_MALLOC(tmp_recvbuf, void *, total_count*(MPIR_MAX(true_extent,extent)), mpi_errno, "tmp_recvbuf");
        /* adjust for potential negative lower bound in datatype */
        tmp_recvbuf = (void *)((char*)tmp_recvbuf - true_lb);
            
        /* need to allocate another temporary buffer to accumulate
           results because recvbuf may not be big enough */
        MPIU_CHKLMEM_MALLOC(tmp_results, void *, total_count*(MPIR_MAX(true_extent,extent)), mpi_errno, "tmp_results");
        /* adjust for potential negative lower bound in datatype */
        tmp_results = (void *)((char*)tmp_results - true_lb);
        
        /* copy sendbuf into tmp_results */
        if (sendbuf != MPI_IN_PLACE)
            mpi_errno = MPIR_Localcopy(sendbuf, total_count, datatype,
                                       tmp_results, total_count, datatype);
        else
            mpi_errno = MPIR_Localcopy(recvbuf, total_count, datatype,
                                       tmp_results, total_count, datatype);
        
        if (mpi_errno) MPIU_ERR_POP(mpi_errno);

        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. */

        if (rank < 2*rem) {
            if (rank % 2 == 0) { /* even */
341
                mpi_errno = MPIC_Send(tmp_results, total_count,
342
343
                                         datatype, rank+1,
                                         MPIR_REDUCE_SCATTER_BLOCK_TAG, comm, errflag);
344
345
                if (mpi_errno) {
                    /* for communication errors, just record the error but continue */
346
                    *errflag = TRUE;
347
348
349
                    MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
                    MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
                }
350
351
352
353
354
355
356
                
                /* temporarily set the rank to -1 so that this
                   process does not pariticipate in recursive
                   doubling */
                newrank = -1; 
            }
            else { /* odd */
357
                mpi_errno = MPIC_Recv(tmp_recvbuf, total_count,
358
359
360
                                         datatype, rank-1,
                                         MPIR_REDUCE_SCATTER_BLOCK_TAG, comm,
                                         MPI_STATUS_IGNORE, errflag);
361
362
                if (mpi_errno) {
                    /* for communication errors, just record the error but continue */
363
                    *errflag = TRUE;
364
365
366
                    MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
                    MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
                }
367
368
369
370
                
                /* do the reduction on received data. since the
                   ordering is right, it doesn't matter whether
                   the operation is commutative or not. */
371
372
                mpi_errno = MPIR_Reduce_local_impl( tmp_recvbuf, tmp_results, 
                                                    total_count, datatype, op);
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
                
                /* change the rank */
                newrank = rank / 2;
            }
        }
        else  /* rank >= 2*rem */
            newrank = rank - rem;

        if (newrank != -1) {
            /* recalculate the recvcnts and disps arrays because the
               even-numbered processes who no longer participate will
               have their result calculated by the process to their
               right (rank+1). */

            MPIU_CHKLMEM_MALLOC(newcnts, int *, pof2*sizeof(int), mpi_errno, "newcnts");
            MPIU_CHKLMEM_MALLOC(newdisps, int *, pof2*sizeof(int), mpi_errno, "newdisps");
            
            for (i=0; i<pof2; i++) {
                /* what does i map to in the old ranking? */
                old_i = (i < rem) ? i*2 + 1 : i + rem;
                if (old_i < 2*rem) {
                    /* This process has to also do its left neighbor's
                       work */
                    newcnts[i] = 2 * recvcount;
                }
                else
                    newcnts[i] = recvcount;
            }
            
            newdisps[0] = 0;
            for (i=1; i<pof2; i++)
                newdisps[i] = newdisps[i-1] + newcnts[i-1];

            mask = pof2 >> 1;
            send_idx = recv_idx = 0;
            last_idx = pof2;
            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) {
                    send_idx = recv_idx + mask;
                    for (i=send_idx; i<last_idx; i++)
                        send_cnt += newcnts[i];
                    for (i=recv_idx; i<send_idx; i++)
                        recv_cnt += newcnts[i];
                }
                else {
                    recv_idx = send_idx + mask;
                    for (i=send_idx; i<recv_idx; i++)
                        send_cnt += newcnts[i];
                    for (i=recv_idx; i<last_idx; i++)
                        recv_cnt += newcnts[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 tmp_results. Recv into tmp_recvbuf */ 
                if ((send_cnt != 0) && (recv_cnt != 0)) 
435
                    mpi_errno = MPIC_Sendrecv((char *) tmp_results +
436
437
438
439
440
441
442
443
                                                 newdisps[send_idx]*extent,
                                                 send_cnt, datatype,
                                                 dst, MPIR_REDUCE_SCATTER_BLOCK_TAG,
                                                 (char *) tmp_recvbuf +
                                                 newdisps[recv_idx]*extent,
                                                 recv_cnt, datatype, dst,
                                                 MPIR_REDUCE_SCATTER_BLOCK_TAG, comm,
                                                 MPI_STATUS_IGNORE, errflag);
444
                else if ((send_cnt == 0) && (recv_cnt != 0))
445
                    mpi_errno = MPIC_Recv((char *) tmp_recvbuf +
446
447
448
449
                                             newdisps[recv_idx]*extent,
                                             recv_cnt, datatype, dst,
                                             MPIR_REDUCE_SCATTER_BLOCK_TAG, comm,
                                             MPI_STATUS_IGNORE, errflag);
450
                else if ((recv_cnt == 0) && (send_cnt != 0))
451
                    mpi_errno = MPIC_Send((char *) tmp_results +
452
453
454
455
                                             newdisps[send_idx]*extent,
                                             send_cnt, datatype,
                                             dst, MPIR_REDUCE_SCATTER_BLOCK_TAG,
                                             comm, errflag);
456

457
458
                if (mpi_errno) {
                    /* for communication errors, just record the error but continue */
459
                    *errflag = TRUE;
460
461
462
                    MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
                    MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
                }
463
464
465
466
467
                
                /* tmp_recvbuf contains data received in this step.
                   tmp_results contains data accumulated so far */
                
                if (recv_cnt) {
468
469
                    mpi_errno = MPIR_Reduce_local_impl( 
                             (char *) tmp_recvbuf + newdisps[recv_idx]*extent,
470
                             (char *) tmp_results + newdisps[recv_idx]*extent, 
471
                             recv_cnt, datatype, op);
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
                }

                /* update send_idx for next iteration */
                send_idx = recv_idx;
                last_idx = recv_idx + mask;
                mask >>= 1;
            }

            /* copy this process's result from tmp_results to recvbuf */
            mpi_errno = MPIR_Localcopy((char *)tmp_results +
                                       disps[rank]*extent, 
                                       recvcount, datatype, recvbuf,
                                       recvcount, datatype);
            if (mpi_errno) MPIU_ERR_POP(mpi_errno);
        }

        /* In the non-power-of-two case, all odd-numbered
           processes of rank < 2*rem send to (rank-1) the result they
           calculated for that process */
        if (rank < 2*rem) {
            if (rank % 2) { /* odd */
493
                mpi_errno = MPIC_Send((char *) tmp_results +
494
495
496
                                         disps[rank-1]*extent, recvcount,
                                         datatype, rank-1,
                                         MPIR_REDUCE_SCATTER_BLOCK_TAG, comm, errflag);
497
498
            }
            else  {   /* even */
499
                mpi_errno = MPIC_Recv(recvbuf, recvcount,
500
501
502
                                         datatype, rank+1,
                                         MPIR_REDUCE_SCATTER_BLOCK_TAG, comm,
                                         MPI_STATUS_IGNORE, errflag);
503
            }
504
505
            if (mpi_errno) {
                /* for communication errors, just record the error but continue */
506
                *errflag = TRUE;
507
508
509
                MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
                MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
            }
510
511
512
        }
    }
    
513
    if (is_commutative && (nbytes >= MPIR_CVAR_REDSCAT_COMMUTATIVE_LONG_MSG_SIZE)) {
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537

        /* commutative and long message, or noncommutative and long message.
           use (p-1) pairwise exchanges */ 
        
        if (sendbuf != MPI_IN_PLACE) {
            /* copy local data into recvbuf */
            mpi_errno = MPIR_Localcopy(((char *)sendbuf+disps[rank]*extent),
                                       recvcount, datatype, recvbuf,
                                       recvcount, datatype);
            if (mpi_errno) MPIU_ERR_POP(mpi_errno);
        }
        
        /* allocate temporary buffer to store incoming data */
        MPIU_CHKLMEM_MALLOC(tmp_recvbuf, void *, recvcount*(MPIR_MAX(true_extent,extent))+1, mpi_errno, "tmp_recvbuf");
        /* adjust for potential negative lower bound in datatype */
        tmp_recvbuf = (void *)((char*)tmp_recvbuf - true_lb);
        
        for (i=1; i<comm_size; i++) {
            src = (rank - i + comm_size) % comm_size;
            dst = (rank + i) % comm_size;
            
            /* send the data that dst needs. recv data that this process
               needs from src into tmp_recvbuf */
            if (sendbuf != MPI_IN_PLACE) 
538
                mpi_errno = MPIC_Sendrecv(((char *)sendbuf+disps[dst]*extent),
539
540
541
542
543
                                             recvcount, datatype, dst,
                                             MPIR_REDUCE_SCATTER_BLOCK_TAG, tmp_recvbuf,
                                             recvcount, datatype, src,
                                             MPIR_REDUCE_SCATTER_BLOCK_TAG, comm,
                                             MPI_STATUS_IGNORE, errflag);
544
            else
545
                mpi_errno = MPIC_Sendrecv(((char *)recvbuf+disps[dst]*extent),
546
547
548
549
550
                                             recvcount, datatype, dst,
                                             MPIR_REDUCE_SCATTER_BLOCK_TAG, tmp_recvbuf,
                                             recvcount, datatype, src,
                                             MPIR_REDUCE_SCATTER_BLOCK_TAG, comm,
                                             MPI_STATUS_IGNORE, errflag);
551
            
552
553
            if (mpi_errno) {
                /* for communication errors, just record the error but continue */
554
                *errflag = TRUE;
555
556
557
                MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
                MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
            }
558
559
560
            
            if (is_commutative || (src < rank)) {
                if (sendbuf != MPI_IN_PLACE) {
561
562
                    mpi_errno = MPIR_Reduce_local_impl( tmp_recvbuf, recvbuf, 
                                                     recvcount, datatype, op); 
563
564
                }
                else {
565
566
567
                    mpi_errno = MPIR_Reduce_local_impl( 
                          tmp_recvbuf, ((char *)recvbuf+disps[rank]*extent), 
                          recvcount, datatype, op ); 
568
569
570
571
572
573
574
575
576
                    /* we can't store the result at the beginning of
                       recvbuf right here because there is useful data
                       there that other process/processes need. at the
                       end, we will copy back the result to the
                       beginning of recvbuf. */
                }
            }
            else {
                if (sendbuf != MPI_IN_PLACE) {
577
578
                    mpi_errno = MPIR_Reduce_local_impl(recvbuf, tmp_recvbuf, 
                                                       recvcount, datatype,op); 
579
580
581
                    /* copy result back into recvbuf */
                    mpi_errno = MPIR_Localcopy(tmp_recvbuf, recvcount, 
                                               datatype, recvbuf,
582
583
                                               recvcount, datatype);
                    if (mpi_errno) MPIU_ERR_POP(mpi_errno);
584
585
                }
                else {
586
587
588
                    mpi_errno = MPIR_Reduce_local_impl( 
                               ((char *)recvbuf+disps[rank]*extent),
                               tmp_recvbuf, recvcount, datatype, op);
589
590
591
592
593
                    /* copy result back into recvbuf */
                    mpi_errno = MPIR_Localcopy(tmp_recvbuf, recvcount, 
                                               datatype, 
                                               ((char *)recvbuf +
                                                disps[rank]*extent), 
594
595
                                               recvcount, datatype);
                    if (mpi_errno) MPIU_ERR_POP(mpi_errno);
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
                }
            }
        }
        
        /* if MPI_IN_PLACE, move output data to the beginning of
           recvbuf. already done for rank 0. */
        if ((sendbuf == MPI_IN_PLACE) && (rank != 0)) {
            mpi_errno = MPIR_Localcopy(((char *)recvbuf +
                                        disps[rank]*extent),  
                                       recvcount, datatype, 
                                       recvbuf, 
                                       recvcount, datatype); 
            if (mpi_errno) MPIU_ERR_POP(mpi_errno);
        }
    }
    
    if (!is_commutative) {

        /* power of two check */
        if (!(comm_size & (comm_size - 1))) {
            /* noncommutative, pof2 size */
617
618
619
620
621
622
623
            mpi_errno = MPIR_Reduce_scatter_block_noncomm(sendbuf, recvbuf, recvcount, datatype, op, comm_ptr, errflag);
            if (mpi_errno) {
                /* for communication errors, just record the error but continue */
                *errflag = TRUE;
                MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
                MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
            }
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
        }
        else {
            /* noncommutative and non-pof2, use recursive doubling. */

            /* need to allocate temporary buffer to receive incoming data*/
            MPIU_CHKLMEM_MALLOC(tmp_recvbuf, void *, total_count*(MPIR_MAX(true_extent,extent)), mpi_errno, "tmp_recvbuf");
            /* adjust for potential negative lower bound in datatype */
            tmp_recvbuf = (void *)((char*)tmp_recvbuf - true_lb);

            /* need to allocate another temporary buffer to accumulate
               results */
            MPIU_CHKLMEM_MALLOC(tmp_results, void *, total_count*(MPIR_MAX(true_extent,extent)), mpi_errno, "tmp_results");
            /* adjust for potential negative lower bound in datatype */
            tmp_results = (void *)((char*)tmp_results - true_lb);

            /* copy sendbuf into tmp_results */
            if (sendbuf != MPI_IN_PLACE)
                mpi_errno = MPIR_Localcopy(sendbuf, total_count, datatype,
                                           tmp_results, total_count, datatype);
            else
                mpi_errno = MPIR_Localcopy(recvbuf, total_count, datatype,
                                           tmp_results, total_count, datatype);

            if (mpi_errno) MPIU_ERR_POP(mpi_errno);

            mask = 0x1;
            i = 0;
            while (mask < comm_size) {
                dst = rank ^ mask;

                dst_tree_root = dst >> i;
                dst_tree_root <<= i;

                my_tree_root = rank >> i;
                my_tree_root <<= i;

                /* At step 1, processes exchange (n-n/p) amount of
                   data; at step 2, (n-2n/p) amount of data; at step 3, (n-4n/p)
                   amount of data, and so forth. We use derived datatypes for this.

                   At each step, a process does not need to send data
                   indexed from my_tree_root to
                   my_tree_root+mask-1. Similarly, a process won't receive
                   data indexed from dst_tree_root to dst_tree_root+mask-1. */

                /* calculate sendtype */
                blklens[0] = blklens[1] = 0;
                for (j=0; j<my_tree_root; j++)
                    blklens[0] += recvcount;
                for (j=my_tree_root+mask; j<comm_size; j++)
                    blklens[1] += recvcount;

                dis[0] = 0;
                dis[1] = blklens[0];
                for (j=my_tree_root; (j<my_tree_root+mask) && (j<comm_size); j++)
                    dis[1] += recvcount;

681
682
683
684
685
                mpi_errno = MPIR_Type_indexed_impl(2, blklens, dis, datatype, &sendtype);
                if (mpi_errno) MPIU_ERR_POP(mpi_errno);
                
                mpi_errno = MPIR_Type_commit_impl(&sendtype);
                if (mpi_errno) MPIU_ERR_POP(mpi_errno);
686
687
688
689
690
691
692
693
694
695
696
697
698

                /* calculate recvtype */
                blklens[0] = blklens[1] = 0;
                for (j=0; j<dst_tree_root && j<comm_size; j++)
                    blklens[0] += recvcount;
                for (j=dst_tree_root+mask; j<comm_size; j++)
                    blklens[1] += recvcount;

                dis[0] = 0;
                dis[1] = blklens[0];
                for (j=dst_tree_root; (j<dst_tree_root+mask) && (j<comm_size); j++)
                    dis[1] += recvcount;

699
700
701
702
703
                mpi_errno = MPIR_Type_indexed_impl(2, blklens, dis, datatype, &recvtype);
                if (mpi_errno) MPIU_ERR_POP(mpi_errno);
                
                mpi_errno = MPIR_Type_commit_impl(&recvtype);
                if (mpi_errno) MPIU_ERR_POP(mpi_errno);
704
705
706
707
708
709
710

                received = 0;
                if (dst < comm_size) {
                    /* tmp_results contains data to be sent in each step. Data is
                       received in tmp_recvbuf and then accumulated into
                       tmp_results. accumulation is done later below.   */ 

711
                    mpi_errno = MPIC_Sendrecv(tmp_results, 1, sendtype, dst,
712
713
714
715
                                                 MPIR_REDUCE_SCATTER_BLOCK_TAG, 
                                                 tmp_recvbuf, 1, recvtype, dst,
                                                 MPIR_REDUCE_SCATTER_BLOCK_TAG, comm,
                                                 MPI_STATUS_IGNORE, errflag);
716
                    received = 1;
717
718
                    if (mpi_errno) {
                        /* for communication errors, just record the error but continue */
719
                        *errflag = TRUE;
720
721
722
                        MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
                        MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
                    }
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
                }

                /* if some processes in this process's subtree in this step
                   did not have any destination process to communicate with
                   because of non-power-of-two, we need to send them the
                   result. We use a logarithmic recursive-halfing algorithm
                   for this. */

                if (dst_tree_root + mask > comm_size) {
                    nprocs_completed = comm_size - my_tree_root - mask;
                    /* nprocs_completed is the number of processes in this
                       subtree that have all the data. Send data to others
                       in a tree fashion. First find root of current tree
                       that is being divided into two. k is the number of
                       least-significant bits in this process's rank that
                       must be zeroed out to find the rank of the root */ 
                    j = mask;
                    k = 0;
                    while (j) {
                        j >>= 1;
                        k++;
                    }
                    k--;

                    tmp_mask = mask >> 1;
                    while (tmp_mask) {
                        dst = rank ^ tmp_mask;

                        tree_root = rank >> k;
                        tree_root <<= k;

                        /* send only if this proc has data and destination
                           doesn't have data. at any step, multiple processes
                           can send if they have the data */
                        if ((dst > rank) && 
                            (rank < tree_root + nprocs_completed)
                            && (dst >= tree_root + nprocs_completed)) {
                            /* send the current result */
761
                            mpi_errno = MPIC_Send(tmp_recvbuf, 1, recvtype,
762
763
                                                     dst, MPIR_REDUCE_SCATTER_BLOCK_TAG,
                                                     comm, errflag);
764
765
                            if (mpi_errno) {
                                /* for communication errors, just record the error but continue */
766
                                *errflag = TRUE;
767
768
769
                                MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
                                MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
                            }
770
771
772
773
774
775
                        }
                        /* recv only if this proc. doesn't have data and sender
                           has data */
                        else if ((dst < rank) && 
                                 (dst < tree_root + nprocs_completed) &&
                                 (rank >= tree_root + nprocs_completed)) {
776
                            mpi_errno = MPIC_Recv(tmp_recvbuf, 1, recvtype, dst,
777
778
                                                     MPIR_REDUCE_SCATTER_BLOCK_TAG,
                                                     comm, MPI_STATUS_IGNORE, errflag);
779
                            received = 1;
780
781
                            if (mpi_errno) {
                                /* for communication errors, just record the error but continue */
782
                                *errflag = TRUE;
783
784
785
                                MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
                                MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
                            }
786
787
788
789
790
791
792
                        }
                        tmp_mask >>= 1;
                        k--;
                    }
                }

                /* The following reduction is done here instead of after 
793
                   the MPIC_Sendrecv or MPIC_Recv above. This is
794
795
796
797
798
799
800
801
                   because to do it above, in the noncommutative 
                   case, we would need an extra temp buffer so as not to
                   overwrite temp_recvbuf, because temp_recvbuf may have
                   to be communicated to other processes in the
                   non-power-of-two case. To avoid that extra allocation,
                   we do the reduce here. */
                if (received) {
                    if (is_commutative || (dst_tree_root < my_tree_root)) {
802
803
804
805
806
807
808
                        mpi_errno = MPIR_Reduce_local_impl(
                                          tmp_recvbuf, tmp_results, blklens[0],
                                          datatype, op); 
                        mpi_errno = MPIR_Reduce_local_impl( 
                                          ((char *)tmp_recvbuf + dis[1]*extent),
                                          ((char *)tmp_results + dis[1]*extent),
                                          blklens[1], datatype, op); 
809
810
                    }
                    else {
811
812
813
814
815
816
817
                        mpi_errno = MPIR_Reduce_local_impl( 
                                        tmp_results, tmp_recvbuf, blklens[0],
                                        datatype, op); 
                        mpi_errno = MPIR_Reduce_local_impl( 
                                        ((char *)tmp_results + dis[1]*extent),
                                        ((char *)tmp_recvbuf + dis[1]*extent),
                                        blklens[1], datatype, op); 
818
819
820
821
822
823
824
                        /* copy result back into tmp_results */
                        mpi_errno = MPIR_Localcopy(tmp_recvbuf, 1, recvtype, 
                                                   tmp_results, 1, recvtype);
                        if (mpi_errno) MPIU_ERR_POP(mpi_errno);
                    }
                }

825
826
                MPIR_Type_free_impl(&sendtype);
                MPIR_Type_free_impl(&recvtype);
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848

                mask <<= 1;
                i++;
            }

            /* now copy final results from tmp_results to recvbuf */
            mpi_errno = MPIR_Localcopy(((char *)tmp_results+disps[rank]*extent),
                                       recvcount, datatype, recvbuf,
                                       recvcount, datatype); 
            if (mpi_errno) MPIU_ERR_POP(mpi_errno);
        }
    }

fn_exit:
    MPIU_CHKLMEM_FREEALL();

    /* 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);

849
    /* --BEGIN ERROR HANDLING-- */
850
851
    if (mpi_errno_ret)
        mpi_errno = mpi_errno_ret;
852
853
    else if (*errflag)
        MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**coll_fail");
854
    /* --END ERROR HANDLING-- */
855
    return mpi_errno;
856
857
858
fn_fail:
    goto fn_exit;
}
859

860
861

#undef FUNCNAME
862
#define FUNCNAME MPIR_Reduce_scatter_block_inter
863
864
#undef FCNAME
#define FCNAME MPIU_QUOTE(FUNCNAME)
865

866
867
/* not declared static because a machine-specific function may call this one in some cases */
int MPIR_Reduce_scatter_block_inter ( 
868
    const void *sendbuf, 
869
870
871
872
    void *recvbuf, 
    int recvcount, 
    MPI_Datatype datatype, 
    MPI_Op op, 
873
    MPID_Comm *comm_ptr, int *errflag )
874
875
876
877
878
879
880
{
/* Intercommunicator Reduce_scatter_block.
   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 scattervs in each group.
*/
    
881
    int rank, mpi_errno, root, local_size, total_count;
882
    int mpi_errno_ret = MPI_SUCCESS;
883
884
885
    MPI_Aint true_extent, true_lb = 0, extent;
    void *tmp_buf=NULL;
    MPID_Comm *newcomm_ptr = NULL;
886
    MPIU_CHKLMEM_DECL(1);
887
888
889
890
891
892
893
894
895
896

    rank = comm_ptr->rank;
    local_size = comm_ptr->local_size;

    total_count = local_size * recvcount;

    if (rank == 0) {
        /* In each group, rank 0 allocates a temp. buffer for the 
           reduce */

897
        MPIR_Type_get_true_extent_impl(datatype, &true_lb, &true_extent);
898
899
        MPID_Datatype_get_extent_macro(datatype, extent);

900
901
        MPIU_CHKLMEM_MALLOC(tmp_buf, void *, total_count*(MPIR_MAX(extent,true_extent)), mpi_errno, "tmp_buf");

902
903
904
905
906
907
908
909
910
911
        /* adjust for potential negative lower bound in datatype */
        tmp_buf = (void *)((char*)tmp_buf - true_lb);
    }

    /* 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, tmp_buf, total_count, datatype, op,
912
                                root, comm_ptr, errflag);
913
914
        if (mpi_errno) {
            /* for communication errors, just record the error but continue */
915
            *errflag = TRUE;
916
917
918
            MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
            MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
        }
919
        
920
921
922
        /* reduce to rank 0 of right group */
        root = 0;
        mpi_errno = MPIR_Reduce_inter(sendbuf, tmp_buf, total_count, datatype, op,
923
                                root, comm_ptr, errflag);
924
925
        if (mpi_errno) {
            /* for communication errors, just record the error but continue */
926
            *errflag = TRUE;
927
928
929
            MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
            MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
        }
930
931
932
933
934
    }
    else {
        /* reduce to rank 0 of left group */
        root = 0;
        mpi_errno = MPIR_Reduce_inter(sendbuf, tmp_buf, total_count, datatype, op,
935
                                root, comm_ptr, errflag);
936
937
        if (mpi_errno) {
            /* for communication errors, just record the error but continue */
938
            *errflag = TRUE;
939
940
941
            MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
            MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
        }
942
943
944
945

        /* reduce from right group to rank 0 */
        root = (rank == 0) ? MPI_ROOT : MPI_PROC_NULL;
        mpi_errno = MPIR_Reduce_inter(sendbuf, tmp_buf, total_count, datatype, op,
946
                                root, comm_ptr, errflag);
947
948
        if (mpi_errno) {
            /* for communication errors, just record the error but continue */
949
            *errflag = TRUE;
950
951
952
            MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
            MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
        }
953
954
955
956
957
958
959
960
    }

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

    newcomm_ptr = comm_ptr->local_comm;

961
    mpi_errno = MPIR_Scatter_impl(tmp_buf, recvcount, datatype, recvbuf,
962
                                  recvcount, datatype, 0, newcomm_ptr, errflag);
963
964
    if (mpi_errno) {
        /* for communication errors, just record the error but continue */
965
        *errflag = TRUE;
966
967
968
        MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
        MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
    }
969
    
970
971
 fn_exit:
    MPIU_CHKLMEM_FREEALL();
972
    /* --BEGIN ERROR HANDLING-- */
973
974
    if (mpi_errno_ret)
        mpi_errno = mpi_errno_ret;
975
976
    else if (*errflag)
        MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**coll_fail");
977
    /* --END ERROR HANDLING-- */
978
979
980
981
    return mpi_errno;
 fn_fail:
    goto fn_exit;
}
982

983
984
985
986
987
988
989
990
991

/* MPIR_Reduce_scatter_block performs a red_scat_block using
   point-to-point messages.  This is intended to be used by
   device-specific implementations of red_scat_block.  In all other
   cases MPIR_Reduce_scatter_block_impl should be used. */
#undef FUNCNAME
#define FUNCNAME MPIR_Reduce_scatter_block
#undef FCNAME
#define FCNAME MPIU_QUOTE(FUNCNAME)
992
993
int MPIR_Reduce_scatter_block(const void *sendbuf, void *recvbuf, 
                              int recvcount, MPI_Datatype datatype,
994
                              MPI_Op op, MPID_Comm *comm_ptr, int *errflag)
995
996
997
998
999
{
    int mpi_errno = MPI_SUCCESS;
        
    if (comm_ptr->comm_kind == MPID_INTRACOMM) {
        /* intracommunicator */
1000
        mpi_errno = MPIR_Reduce_scatter_block_intra(sendbuf, recvbuf, recvcount, datatype, op, comm_ptr, errflag);
1001
1002
1003
        if (mpi_errno) MPIU_ERR_POP(mpi_errno);
    } else {
        /* intercommunicator */
1004
        mpi_errno = MPIR_Reduce_scatter_block_inter(sendbuf, recvbuf, recvcount, datatype, op, comm_ptr, errflag);
1005
        if (mpi_errno) MPIU_ERR_POP(mpi_errno);
1006
1007
    }

1008
 fn_exit:
1009
    return mpi_errno;
1010
1011
1012
 fn_fail:
    goto fn_exit;
}
1013

1014
1015
1016
1017
1018
1019
1020
1021
1022
/* MPIR_Reduce_scatter_block_impl should be called by any internal
   component that would otherwise call MPI_Reduce_scatter_block.  This
   differs from MPIR_Reduce_scatter_block in that this will call the
   coll_fns version if it exists.  This is a replacement for
   NMPI_Reduce_scatter_block. */
#undef FUNCNAME
#define FUNCNAME MPIR_Reduce_scatter_block_impl
#undef FCNAME
#define FCNAME MPIU_QUOTE(FUNCNAME)
1023
1024
int MPIR_Reduce_scatter_block_impl(const void *sendbuf, void *recvbuf, 
                                   int recvcount, MPI_Datatype datatype,
1025
                                   MPI_Op op, MPID_Comm *comm_ptr, int *errflag)
1026
1027
1028
1029
{
    int mpi_errno = MPI_SUCCESS;
        
    if (comm_ptr->coll_fns != NULL && comm_ptr->coll_fns->Reduce_scatter_block != NULL) {
1030
	/* --BEGIN USEREXTENSION-- */
1031
	mpi_errno = comm_ptr->coll_fns->Reduce_scatter_block(sendbuf, recvbuf, recvcount, datatype, op, comm_ptr, errflag);
1032
        if (mpi_errno) MPIU_ERR_POP(mpi_errno);
1033
	/* --END USEREXTENSION-- */
1034
1035
1036
    } else {
        if (comm_ptr->comm_kind == MPID_INTRACOMM) {
            /* intracommunicator */
1037
            mpi_errno = MPIR_Reduce_scatter_block_intra(sendbuf, recvbuf, recvcount, datatype, op, comm_ptr, errflag);
1038
1039
1040
            if (mpi_errno) MPIU_ERR_POP(mpi_errno);
        } else {
            /* intercommunicator */
1041
            mpi_errno = MPIR_Reduce_scatter_block_inter(sendbuf, recvbuf, recvcount, datatype, op, comm_ptr, errflag);
1042
1043
1044
1045
1046
1047
1048
1049
            if (mpi_errno) MPIU_ERR_POP(mpi_errno);
        }
    }

 fn_exit:
    return mpi_errno;
 fn_fail:
    goto fn_exit;
1050
}
1051
1052


1053
1054
1055
1056
1057
1058
1059
#endif

#undef FUNCNAME
#define FUNCNAME MPI_Reduce_scatter_block
#undef FCNAME
#define FCNAME MPIU_QUOTE(FUNCNAME)
/*@
1060

1061
1062
1063
MPI_Reduce_scatter_block - Combines values and scatters the results

Input Parameters:
1064
+ sendbuf - starting address of send buffer (choice) 
1065
. recvcount - element count per block (non-negative integer)
1066
1067
1068
. datatype - data type of elements of input buffer (handle) 
. op - operation (handle) 
- comm - communicator (handle) 
1069

1070
Output Parameters:
1071
. recvbuf - starting address of receive buffer (choice) 
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087

.N ThreadSafe

.N Fortran

.N collops

.N Errors
.N MPI_SUCCESS
.N MPI_ERR_COMM
.N MPI_ERR_COUNT
.N MPI_ERR_TYPE
.N MPI_ERR_BUFFER
.N MPI_ERR_OP
.N MPI_ERR_BUFFER_ALIAS
@*/
1088
int MPI_Reduce_scatter_block(const void *sendbuf, void *recvbuf, 
1089
1090
                             int recvcount, 
                             MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
1091
1092
1093
{
    int mpi_errno = MPI_SUCCESS;
    MPID_Comm *comm_ptr = NULL;
1094
    int errflag = FALSE;
1095
1096
1097
    MPID_MPI_STATE_DECL(MPID_STATE_MPI_REDUCE_SCATTER_BLOCK);

    MPIR_ERRTEST_INITIALIZED_ORDIE();
1098
    
1099
1100
1101
1102
1103
1104
1105
1106
    MPIU_THREAD_CS_ENTER(ALLFUNC,);
    MPID_MPI_COLL_FUNC_ENTER(MPID_STATE_MPI_REDUCE_SCATTER_BLOCK);

    /* Validate parameters, especially handles needing to be converted */
#   ifdef HAVE_ERROR_CHECKING
    {
        MPID_BEGIN_ERROR_CHECKS;
        {
1107
1108
	    MPIR_ERRTEST_COMM(comm, mpi_errno);
	}
1109
1110
1111
1112
1113
        MPID_END_ERROR_CHECKS;
    }
#   endif /* HAVE_ERROR_CHECKING */

    /* Convert MPI object handles to object pointers */
1114
    MPID_Comm_get_ptr( comm, comm_ptr );
1115
1116
1117
1118
1119
1120

    /* Validate parameters and objects (post conversion) */
#   ifdef HAVE_ERROR_CHECKING
    {
        MPID_BEGIN_ERROR_CHECKS;
        {
1121
	    MPID_Datatype *datatype_ptr = NULL;
1122
            MPID_Op *op_ptr = NULL;
1123
1124
	    
            MPID_Comm_valid_ptr( comm_ptr, mpi_errno );
1125
1126
            if (mpi_errno != MPI_SUCCESS) goto fn_fail;

1127
            MPIR_ERRTEST_COUNT(recvcount,mpi_errno);
1128

1129
	    MPIR_ERRTEST_DATATYPE(datatype, "datatype", mpi_errno);
1130
1131
            if (HANDLE_GET_KIND(datatype) != HANDLE_KIND_BUILTIN) {
                MPID_Datatype_get_ptr(datatype, datatype_ptr);
1132
                MPID_Datatype_valid_ptr( datatype_ptr, mpi_errno );
1133
                if (mpi_errno != MPI_SUCCESS) goto fn_fail;
1134
                MPID_Datatype_committed_ptr( datatype_ptr, mpi_errno );
1135
                if (mpi_errno != MPI_SUCCESS) goto fn_fail;
1136
1137
1138
            }

            MPIR_ERRTEST_RECVBUF_INPLACE(recvbuf, recvcount, mpi_errno);
1139
            if (comm_ptr->comm_kind == MPID_INTERCOMM) {
1140
                MPIR_ERRTEST_SENDBUF_INPLACE(sendbuf, recvcount, mpi_errno);
1141
            } else if (sendbuf != MPI_IN_PLACE && recvcount != 0)
1142
                MPIR_ERRTEST_ALIAS_COLL(sendbuf, recvbuf, mpi_errno)
1143
1144

            MPIR_ERRTEST_USERBUFFER(recvbuf,recvcount,datatype,mpi_errno);
1145
            MPIR_ERRTEST_USERBUFFER(sendbuf,recvcount,datatype,mpi_errno); 
1146

1147
	    MPIR_ERRTEST_OP(op, mpi_errno);
1148
1149
1150
1151

            if (mpi_errno != MPI_SUCCESS) goto fn_fail;
            if (HANDLE_GET_KIND(op) != HANDLE_KIND_BUILTIN) {
                MPID_Op_get_ptr(op, op_ptr);
1152
                MPID_Op_valid_ptr( op_ptr, mpi_errno );
1153
1154
            }
            if (HANDLE_GET_KIND(op) == HANDLE_KIND_BUILTIN) {
1155
                mpi_errno = 
1156
                    ( * MPIR_OP_HDL_TO_DTYPE_FN(op) )(datatype); 
1157
            }
1158
            if (mpi_errno != MPI_SUCCESS) goto fn_fail;
1159
1160
1161
1162
1163
1164
1165
        }
        MPID_END_ERROR_CHECKS;
    }
#   endif /* HAVE_ERROR_CHECKING */

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

1166
1167
    mpi_errno = MPIR_Reduce_scatter_block_impl(sendbuf, recvbuf, recvcount, 
                                          datatype, op, comm_ptr, &errflag);
1168
    if (mpi_errno) goto fn_fail;
1169
1170

    /* ... end of body of routine ... */
1171
    
1172
1173
1174
1175
1176
1177
1178
1179
1180
  fn_exit:
    MPID_MPI_COLL_FUNC_EXIT(MPID_STATE_MPI_REDUCE_SCATTER_BLOCK);
    MPIU_THREAD_CS_EXIT(ALLFUNC,);
    return mpi_errno;

  fn_fail:
    /* --BEGIN ERROR HANDLING-- */
#   ifdef HAVE_ERROR_CHECKING
    {
1181
1182
1183
	mpi_errno = MPIR_Err_create_code(
	    mpi_errno, MPIR_ERR_RECOVERABLE, FCNAME, __LINE__, MPI_ERR_OTHER, "**mpi_reduce_scatter_block",
	    "**mpi_reduce_scatter_block %p %p %d %D %O %C", sendbuf, recvbuf, recvcount, datatype, op, comm);
1184
1185
1186
1187
1188
1189
    }
#   endif
    mpi_errno = MPIR_Err_return_comm( comm_ptr, FCNAME, mpi_errno );
    goto fn_exit;
    /* --END ERROR HANDLING-- */
}