red_scat_block.c 49.2 KB
Newer Older
1
/* -*- Mode: C; c-basic-offset:4 ; indent-tabs-mode:nil ; -*- */
2
3
/*
 *
4
 *  (C) 2010 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
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
#include "mpiimpl.h"

/* -- 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
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
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
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
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
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
683
684
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
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
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
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000


/* Implements the "mirror permutation" of "bits" bits of an integer "x".

   positions 76543210, bits==3 yields 76543012.

   This function could/should be moved to a common utility location for use in
   other collectives as well. */
ATTRIBUTE((const)) /* tells the compiler that this func only depends on its args
                      and may be optimized much more aggressively, similar to "pure" */
static inline int mirror_permutation(unsigned int x, int bits)
{
    /* a mask for the high order bits that should be copied as-is */
    int high_mask = ~((0x1 << bits) - 1);
    int retval = x & high_mask;
    int i;

    for (i = 0; i < bits; ++i) {
        unsigned int bitval = (x & (0x1 << i)) >> i; /* 0x1 or 0x0 */
        retval |= bitval << ((bits - i) - 1);
    }

    return retval;
}

/* FIXME should we be checking the op_errno here? */
#ifdef HAVE_CXX_BINDING
/* NOTE: assumes 'uop' is the operator function pointer and
   that 'is_cxx_uop' is is a boolean indicating the obvious */
#define call_uop(in_, inout_, count_, datatype_)                                     \
do {                                                                                 \
    if (is_cxx_uop) {                                                                \
        (*MPIR_Process.cxx_call_op_fn)((in_), (inout_), (count_), (datatype_), uop); \
    }                                                                                \
    else {                                                                           \
        (*uop)((in_), (inout_), &(count_), &(datatype_));                            \
    }                                                                                \
} while (0)

#else
#define call_uop(in_, inout_, count_, datatype_)      \
    (*uop)((in_), (inout_), &(count_), &(datatype_))
#endif

/* Implements the reduce-scatter butterfly algorithm described in J. L. Traff's
 * "An Improved Algorithm for (Non-commutative) Reduce-Scatter with an Application"
 * 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 (
    void *sendbuf,
    void *recvbuf,
    int recvcount,
    MPI_Datatype datatype,
    MPI_Op op,
    MPID_Comm *comm_ptr )
{
    int mpi_errno = MPI_SUCCESS;
    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;
    MPI_Aint extent, true_extent, true_lb;
    int is_commutative;
    int buf0_was_inout;
    void *tmp_buf0;
    void *tmp_buf1;
    void *result_ptr;
    MPI_Comm comm = comm_ptr->handle;
    MPI_User_function *uop;
    MPID_Op *op_ptr;
#ifdef HAVE_CXX_BINDING
    int is_cxx_uop = 0;
#endif
    MPIU_CHKLMEM_DECL(3);

    MPID_Datatype_get_extent_macro(datatype, extent);
    /* assumes nesting is handled by the caller right now, may not be true in the future */
    mpi_errno = NMPI_Type_get_true_extent(datatype, &true_lb, &true_extent);

    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;
    }

    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) {
        mpi_errno = MPIR_Localcopy((char *)(sendbuf == MPI_IN_PLACE ? recvbuf : sendbuf) + (i * true_extent * block_size), block_size, datatype,
                                   (char *)tmp_buf0 + (mirror_permutation(i, log2_comm_size) * true_extent * block_size), block_size, datatype);
        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;
        }

        mpi_errno = MPIC_Sendrecv(outgoing_data + send_offset*true_extent,
                                  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);
        /* 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) */
            call_uop(incoming_data + recv_offset*true_extent,
                     outgoing_data + recv_offset*true_extent,
                     size, datatype);
            buf0_was_inout = buf0_was_inout;
        }
        else {
            /* lower ranked value so need to call op(my_data, received_data) */
            call_uop(outgoing_data + recv_offset*true_extent,
                     incoming_data + recv_offset*true_extent,
                     size, datatype);
            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);
fn_exit:
    MPIU_CHKLMEM_FREEALL();
    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
#define FUNCNAME MPIR_Reduce_scatter_block
#undef FCNAME
#define FCNAME MPIU_QUOTE(FUNCNAME)
/* begin:nested */
/* not declared static because a machine-specific function may call this one in some cases */
int MPIR_Reduce_scatter_block ( 
    void *sendbuf, 
    void *recvbuf, 
    int recvcount, 
    MPI_Datatype datatype, 
    MPI_Op op, 
    MPID_Comm *comm_ptr )
{
    int   rank, comm_size, i;
    MPI_Aint extent, true_extent, true_lb; 
    int  *disps;
    void *tmp_recvbuf, *tmp_results;
    int   mpi_errno = MPI_SUCCESS;
    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;
    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(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;

    MPIR_Nest_incr();

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

    MPID_Datatype_get_extent_macro(datatype, extent);
    mpi_errno = NMPI_Type_get_true_extent(datatype, &true_lb,
                                          &true_extent);  
    if (mpi_errno) MPIU_ERR_POP(mpi_errno);
    
    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;
    }

    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));

    if ((is_commutative) && (nbytes < MPIR_REDSCAT_COMMUTATIVE_LONG_MSG)) {
        /* 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 */
                mpi_errno = MPIC_Send(tmp_results, total_count, 
                                      datatype, rank+1,
                                      MPIR_REDUCE_SCATTER_BLOCK_TAG, comm);
		/* --BEGIN ERROR HANDLING-- */
                if (mpi_errno)
		{
		    mpi_errno = MPIR_Err_create_code(mpi_errno, MPIR_ERR_RECOVERABLE, FCNAME, __LINE__, MPI_ERR_OTHER, "**fail", 0);
		    return mpi_errno;
		}
		/* --END ERROR HANDLING-- */
                
                /* 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_recvbuf, total_count, 
                                      datatype, rank-1,
                                      MPIR_REDUCE_SCATTER_BLOCK_TAG, comm,
                                      MPI_STATUS_IGNORE);
		/* --BEGIN ERROR HANDLING-- */
                if (mpi_errno)
		{
		    mpi_errno = MPIR_Err_create_code(mpi_errno, MPIR_ERR_RECOVERABLE, FCNAME, __LINE__, MPI_ERR_OTHER, "**fail", 0);
		    return mpi_errno;
		}
		/* --END ERROR HANDLING-- */
                
                /* 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_recvbuf, tmp_results, 
                                                    total_count,
                                                    datatype,
                                                    uop ); 
                }
                else 
#endif
                    (*uop)(tmp_recvbuf, tmp_results, &total_count, &datatype);
                
                /* 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)) 
                    mpi_errno = MPIC_Sendrecv((char *) tmp_results +
                                          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); 
                else if ((send_cnt == 0) && (recv_cnt != 0))
                    mpi_errno = MPIC_Recv((char *) tmp_recvbuf +
                                          newdisps[recv_idx]*extent,
                                          recv_cnt, datatype, dst,
                                          MPIR_REDUCE_SCATTER_BLOCK_TAG, comm,
                                          MPI_STATUS_IGNORE);
                else if ((recv_cnt == 0) && (send_cnt != 0))
                    mpi_errno = MPIC_Send((char *) tmp_results +
                                          newdisps[send_idx]*extent,
                                          send_cnt, datatype,  
                                          dst, MPIR_REDUCE_SCATTER_BLOCK_TAG,
                                          comm);  

                if (mpi_errno) MPIU_ERR_POP(mpi_errno);
                
                /* tmp_recvbuf contains data received in this step.
                   tmp_results contains data accumulated so far */
                
                if (recv_cnt) {
#ifdef HAVE_CXX_BINDING
                    if (is_cxx_uop) {
                        (*MPIR_Process.cxx_call_op_fn)((char *) tmp_recvbuf +
                                                   newdisps[recv_idx]*extent,
                                                   (char *) tmp_results + 
                                                   newdisps[recv_idx]*extent, 
                                                   recv_cnt, datatype, uop);
                    }
                    else 
#endif
                        (*uop)((char *) tmp_recvbuf + newdisps[recv_idx]*extent,
                             (char *) tmp_results + newdisps[recv_idx]*extent, 
                               &recv_cnt, &datatype);
                }

                /* 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 */
                mpi_errno = MPIC_Send((char *) tmp_results +
                                      disps[rank-1]*extent, recvcount,
                                      datatype, rank-1,
                                      MPIR_REDUCE_SCATTER_BLOCK_TAG, comm);
            }
            else  {   /* even */
                mpi_errno = MPIC_Recv(recvbuf, recvcount,
                                      datatype, rank+1,
                                      MPIR_REDUCE_SCATTER_BLOCK_TAG, comm,
                                      MPI_STATUS_IGNORE); 
            }
            if (mpi_errno) MPIU_ERR_POP(mpi_errno);
        }
    }
    
    if (is_commutative && (nbytes >= MPIR_REDSCAT_COMMUTATIVE_LONG_MSG)) {

        /* 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) 
                mpi_errno = MPIC_Sendrecv(((char *)sendbuf+disps[dst]*extent), 
                                          recvcount, datatype, dst,
                                          MPIR_REDUCE_SCATTER_BLOCK_TAG, tmp_recvbuf,
                                          recvcount, datatype, src,
                                          MPIR_REDUCE_SCATTER_BLOCK_TAG, comm,
                                          MPI_STATUS_IGNORE);
            else
                mpi_errno = MPIC_Sendrecv(((char *)recvbuf+disps[dst]*extent), 
                                          recvcount, datatype, dst,
                                          MPIR_REDUCE_SCATTER_BLOCK_TAG, tmp_recvbuf,
                                          recvcount, datatype, src,
                                          MPIR_REDUCE_SCATTER_BLOCK_TAG, comm,
                                          MPI_STATUS_IGNORE);
            
            if (mpi_errno) MPIU_ERR_POP(mpi_errno);
            
            if (is_commutative || (src < rank)) {
                if (sendbuf != MPI_IN_PLACE) {
#ifdef HAVE_CXX_BINDING
                    if (is_cxx_uop) {
                        (*MPIR_Process.cxx_call_op_fn)(tmp_recvbuf, 
                                                       recvbuf, 
                                                       recvcount, 
                                                       datatype, uop );
                    }
                    else 
#endif
                        (*uop)(tmp_recvbuf, recvbuf, &recvcount, 
                               &datatype); 
                }
                else {
#ifdef HAVE_CXX_BINDING
                    if (is_cxx_uop) {
                        (*MPIR_Process.cxx_call_op_fn)( tmp_recvbuf, 
                                                        ((char *)recvbuf+disps[rank]*extent), 
                                                        recvcount, datatype, uop ); 
                    }
                    else 
#endif
                        (*uop)(tmp_recvbuf, ((char *)recvbuf+disps[rank]*extent), 
                               &recvcount, &datatype); 
                    /* 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) {
#ifdef HAVE_CXX_BINDING
                    if (is_cxx_uop) {
                        (*MPIR_Process.cxx_call_op_fn)( recvbuf, 
                                                        tmp_recvbuf, 
                                                        recvcount, 
                                                        datatype, uop );
                    }
                    else 
#endif
                        (*uop)(recvbuf, tmp_recvbuf, &recvcount, &datatype); 
                    /* copy result back into recvbuf */
                    mpi_errno = MPIR_Localcopy(tmp_recvbuf, recvcount, 
                                               datatype, recvbuf,
                                               recvcount, datatype); 
                }
                else {
#ifdef HAVE_CXX_BINDING
                    if (is_cxx_uop) {
                        (*MPIR_Process.cxx_call_op_fn)( 
                            ((char *)recvbuf+disps[rank]*extent),
                            tmp_recvbuf, recvcount, datatype, uop );   
                        
                    }
                    else 
#endif
                        (*uop)(((char *)recvbuf+disps[rank]*extent),
                               tmp_recvbuf, &recvcount, &datatype);   
                    /* copy result back into recvbuf */
                    mpi_errno = MPIR_Localcopy(tmp_recvbuf, recvcount, 
                                               datatype, 
                                               ((char *)recvbuf +
                                                disps[rank]*extent), 
                                               recvcount, datatype); 
                }
                if (mpi_errno) MPIU_ERR_POP(mpi_errno);
            }
        }
        
        /* 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 */
            mpi_errno = MPIR_Reduce_scatter_block_noncomm(sendbuf, recvbuf, recvcount, datatype, op, comm_ptr);
            if (mpi_errno) MPIU_ERR_POP(mpi_errno);
        }
        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;

                NMPI_Type_indexed(2, blklens, dis, datatype, &sendtype);
                NMPI_Type_commit(&sendtype);

                /* 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;

                NMPI_Type_indexed(2, blklens, dis, datatype, &recvtype);
                NMPI_Type_commit(&recvtype);

                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.   */ 

                    mpi_errno = MPIC_Sendrecv(tmp_results, 1, sendtype, dst,
                                              MPIR_REDUCE_SCATTER_BLOCK_TAG, 
                                              tmp_recvbuf, 1, recvtype, dst,
                                              MPIR_REDUCE_SCATTER_BLOCK_TAG, comm,
                                              MPI_STATUS_IGNORE); 
                    received = 1;
                    if (mpi_errno) MPIU_ERR_POP(mpi_errno);
                }

                /* 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 */
                            mpi_errno = MPIC_Send(tmp_recvbuf, 1, recvtype,
                                                  dst, MPIR_REDUCE_SCATTER_BLOCK_TAG,
                                                  comm);  
                            if (mpi_errno) MPIU_ERR_POP(mpi_errno);
                        }
                        /* 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)) {
                            mpi_errno = MPIC_Recv(tmp_recvbuf, 1, recvtype, dst,
                                                  MPIR_REDUCE_SCATTER_BLOCK_TAG,
                                                  comm, MPI_STATUS_IGNORE); 
                            received = 1;
                            if (mpi_errno) MPIU_ERR_POP(mpi_errno);
                        }
                        tmp_mask >>= 1;
                        k--;
                    }
                }

                /* The following reduction is done here instead of after 
                   the MPIC_Sendrecv or MPIC_Recv above. This is
                   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)) {
#ifdef HAVE_CXX_BINDING
                        if (is_cxx_uop) {
                            (*MPIR_Process.cxx_call_op_fn)( tmp_recvbuf, 
                                                            tmp_results, blklens[0],
                                                            datatype, uop); 
                            (*MPIR_Process.cxx_call_op_fn)( 
                                ((char *)tmp_recvbuf + dis[1]*extent),
                                ((char *)tmp_results + dis[1]*extent),
                                blklens[1], datatype, uop ); 
                        }
                        else
#endif
                        {
                            (*uop)(tmp_recvbuf, tmp_results, &blklens[0],
                                   &datatype); 
                            (*uop)(((char *)tmp_recvbuf + dis[1]*extent),
                                   ((char *)tmp_results + dis[1]*extent),
                                   &blklens[1], &datatype); 
                        }
                    }
                    else {
#ifdef HAVE_CXX_BINDING
                        if (is_cxx_uop) {
                            (*MPIR_Process.cxx_call_op_fn)( tmp_results, 
                                                            tmp_recvbuf, blklens[0],
                                                            datatype, uop ); 
                            (*MPIR_Process.cxx_call_op_fn)( 
                                ((char *)tmp_results + dis[1]*extent),
                                ((char *)tmp_recvbuf + dis[1]*extent),
                                blklens[1], datatype, uop ); 
                        }
                        else 
#endif
                        {
                            (*uop)(tmp_results, tmp_recvbuf, &blklens[0],
                                   &datatype); 
                            (*uop)(((char *)tmp_results + dis[1]*extent),
                                   ((char *)tmp_recvbuf + dis[1]*extent),
                                   &blklens[1], &datatype); 
                        }
                        /* 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);
                    }
                }

                NMPI_Type_free(&sendtype);
                NMPI_Type_free(&recvtype);

                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();

    MPIR_Nest_decr();
    /* 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);

    return (mpi_errno);
fn_fail:
    goto fn_exit;
}
/* end:nested */

#undef FUNCNAME
#define FUNCNAME MPIR_Reduce_scatter_block
#undef FCNAME
#define FCNAME MPIU_QUOTE(FUNCNAME)
/* begin:nested */
/* not declared static because a machine-specific function may call this one in some cases */
int MPIR_Reduce_scatter_block_inter ( 
    void *sendbuf, 
    void *recvbuf, 
    int recvcount, 
    MPI_Datatype datatype, 
    MPI_Op op, 
    MPID_Comm *comm_ptr )
{
/* 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.
*/
    
    int rank, mpi_errno, root, local_size, total_count, i;
    MPI_Aint true_extent, true_lb = 0, extent;
    void *tmp_buf=NULL;
    MPID_Comm *newcomm_ptr = NULL;

    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 */

        mpi_errno = NMPI_Type_get_true_extent(datatype, &true_lb,
                                              &true_extent);  
	/* --BEGIN ERROR HANDLING-- */
        if (mpi_errno)
	{
	    mpi_errno = MPIR_Err_create_code(mpi_errno, MPIR_ERR_RECOVERABLE, FCNAME, __LINE__, MPI_ERR_OTHER, "**fail", 0);
	    return mpi_errno;
For faster browsing, not all history is shown. View entire blame