Logo AND Algorithmique Numérique Distribuée

Public GIT Repository
Fix signedness errors in format strings.
[simgrid.git] / src / smpi / colls / reduce / reduce-ompi.cpp
1 /* Copyright (c) 2013-2017. The SimGrid Team.
2  * All rights reserved.                                                     */
3
4 /* This program is free software; you can redistribute it and/or modify it
5  * under the terms of the license (GNU LGPL) which comes with this package. */
6
7 /*
8  * Copyright (c) 2004-2005 The Trustees of Indiana University and Indiana
9  *                         University Research and Technology
10  *                         Corporation.  All rights reserved.
11  * Copyright (c) 2004-2009 The University of Tennessee and The University
12  *                         of Tennessee Research Foundation.  All rights
13  *                         reserved.
14  * Copyright (c) 2004-2005 High Performance Computing Center Stuttgart,
15  *                         University of Stuttgart.  All rights reserved.
16  * Copyright (c) 2004-2005 The Regents of the University of California.
17  *                         All rights reserved.
18  *
19  * Additional copyrights may follow
20  */
21
22 #include "../colls_private.h"
23 #include "../coll_tuned_topo.h"
24
25 namespace simgrid{
26 namespace smpi{
27
28 int smpi_coll_tuned_ompi_reduce_generic( void* sendbuf, void* recvbuf, int original_count,
29                                     MPI_Datatype datatype, MPI_Op  op,
30                                     int root, MPI_Comm comm,
31                                     ompi_coll_tree_t* tree, int count_by_segment,
32                                     int max_outstanding_reqs );
33 /**
34  * This is a generic implementation of the reduce protocol. It used the tree
35  * provided as an argument and execute all operations using a segment of
36  * count times a datatype.
37  * For the last communication it will update the count in order to limit
38  * the number of datatype to the original count (original_count)
39  *
40  * Note that for non-commutative operations we cannot save memory copy
41  * for the first block: thus we must copy sendbuf to accumbuf on intermediate
42  * to keep the optimized loop happy.
43  */
44 int smpi_coll_tuned_ompi_reduce_generic( void* sendbuf, void* recvbuf, int original_count,
45                                     MPI_Datatype datatype, MPI_Op  op,
46                                     int root, MPI_Comm comm,
47                                     ompi_coll_tree_t* tree, int count_by_segment,
48                                     int max_outstanding_reqs )
49 {
50     char *inbuf[2] = {NULL, NULL}, *inbuf_free[2] = {NULL, NULL};
51     char *accumbuf = NULL, *accumbuf_free = NULL;
52     char *local_op_buffer = NULL, *sendtmpbuf = NULL;
53     ptrdiff_t extent, lower_bound, segment_increment;
54     MPI_Request  reqs[2] = {MPI_REQUEST_NULL, MPI_REQUEST_NULL};
55     int num_segments, line, ret, segindex, i, rank;
56     int recvcount, prevcount, inbi;
57
58     /**
59      * Determine number of segments and number of elements
60      * sent per operation
61      */
62     datatype->extent(&lower_bound, &extent);
63     num_segments = (original_count + count_by_segment - 1) / count_by_segment;
64     segment_increment = count_by_segment * extent;
65
66     sendtmpbuf = (char*) sendbuf;
67     if( sendbuf == MPI_IN_PLACE ) {
68         sendtmpbuf = (char *)recvbuf;
69     }
70
71     XBT_DEBUG("coll:tuned:reduce_generic count %d, msg size %lu, segsize %lu, max_requests %d", original_count,
72               (unsigned long)(num_segments * segment_increment), (unsigned long)segment_increment,
73               max_outstanding_reqs);
74
75     rank = comm->rank();
76
77     /* non-leaf nodes - wait for children to send me data & forward up
78        (if needed) */
79     if( tree->tree_nextsize > 0 ) {
80         ptrdiff_t true_extent, real_segment_size;
81         true_extent=datatype->get_extent();
82
83         /* handle non existant recv buffer (i.e. its NULL) and
84            protect the recv buffer on non-root nodes */
85         accumbuf = (char*)recvbuf;
86         if( (NULL == accumbuf) || (root != rank) ) {
87             /* Allocate temporary accumulator buffer. */
88             accumbuf_free = (char*)smpi_get_tmp_sendbuffer(true_extent +
89                                           (original_count - 1) * extent);
90             if (accumbuf_free == NULL) {
91                 line = __LINE__; ret = -1; goto error_hndl;
92             }
93             accumbuf = accumbuf_free - lower_bound;
94         }
95
96         /* If this is a non-commutative operation we must copy
97            sendbuf to the accumbuf, in order to simplfy the loops */
98         if ((op != MPI_OP_NULL && not op->is_commutative())) {
99           Datatype::copy((char*)sendtmpbuf, original_count, datatype, (char*)accumbuf, original_count, datatype);
100         }
101         /* Allocate two buffers for incoming segments */
102         real_segment_size = true_extent + (count_by_segment - 1) * extent;
103         inbuf_free[0] = (char*) smpi_get_tmp_recvbuffer(real_segment_size);
104         if( inbuf_free[0] == NULL ) {
105             line = __LINE__; ret = -1; goto error_hndl;
106         }
107         inbuf[0] = inbuf_free[0] - lower_bound;
108         /* if there is chance to overlap communication -
109            allocate second buffer */
110         if( (num_segments > 1) || (tree->tree_nextsize > 1) ) {
111             inbuf_free[1] = (char*) smpi_get_tmp_recvbuffer(real_segment_size);
112             if( inbuf_free[1] == NULL ) {
113                 line = __LINE__; ret = -1; goto error_hndl;
114             }
115             inbuf[1] = inbuf_free[1] - lower_bound;
116         }
117
118         /* reset input buffer index and receive count */
119         inbi = 0;
120         recvcount = 0;
121         /* for each segment */
122         for( segindex = 0; segindex <= num_segments; segindex++ ) {
123             prevcount = recvcount;
124             /* recvcount - number of elements in current segment */
125             recvcount = count_by_segment;
126             if( segindex == (num_segments-1) )
127                 recvcount = original_count - count_by_segment * segindex;
128
129             /* for each child */
130             for( i = 0; i < tree->tree_nextsize; i++ ) {
131                 /**
132                  * We try to overlap communication:
133                  * either with next segment or with the next child
134                  */
135                 /* post irecv for current segindex on current child */
136                 if( segindex < num_segments ) {
137                     void* local_recvbuf = inbuf[inbi];
138                     if( 0 == i ) {
139                         /* for the first step (1st child per segment) and
140                          * commutative operations we might be able to irecv
141                          * directly into the accumulate buffer so that we can
142                          * reduce(op) this with our sendbuf in one step as
143                          * ompi_op_reduce only has two buffer pointers,
144                          * this avoids an extra memory copy.
145                          *
146                          * BUT if the operation is non-commutative or
147                          * we are root and are USING MPI_IN_PLACE this is wrong!
148                          */
149                         if(  (op==MPI_OP_NULL || op->is_commutative()) &&
150                             !((MPI_IN_PLACE == sendbuf) && (rank == tree->tree_root)) ) {
151                             local_recvbuf = accumbuf + segindex * segment_increment;
152                         }
153                     }
154
155                     reqs[inbi]=Request::irecv(local_recvbuf, recvcount, datatype,
156                                              tree->tree_next[i],
157                                              COLL_TAG_REDUCE, comm
158                                              );
159                 }
160                 /* wait for previous req to complete, if any.
161                    if there are no requests reqs[inbi ^1] will be
162                    MPI_REQUEST_NULL. */
163                 /* wait on data from last child for previous segment */
164                 Request::waitall( 1, &reqs[inbi ^ 1],
165                                              MPI_STATUSES_IGNORE );
166                 local_op_buffer = inbuf[inbi ^ 1];
167                 if( i > 0 ) {
168                     /* our first operation is to combine our own [sendbuf] data
169                      * with the data we recvd from down stream (but only
170                      * the operation is commutative and if we are not root and
171                      * not using MPI_IN_PLACE)
172                      */
173                     if( 1 == i ) {
174                         if( (op==MPI_OP_NULL || op->is_commutative())&&
175                             !((MPI_IN_PLACE == sendbuf) && (rank == tree->tree_root)) ) {
176                             local_op_buffer = sendtmpbuf + segindex * segment_increment;
177                         }
178                     }
179                     /* apply operation */
180                     if(op!=MPI_OP_NULL) op->apply( local_op_buffer,
181                                    accumbuf + segindex * segment_increment,
182                                    &recvcount, datatype );
183                 } else if ( segindex > 0 ) {
184                     void* accumulator = accumbuf + (segindex-1) * segment_increment;
185                     if( tree->tree_nextsize <= 1 ) {
186                         if(  (op==MPI_OP_NULL || op->is_commutative()) &&
187                             !((MPI_IN_PLACE == sendbuf) && (rank == tree->tree_root)) ) {
188                             local_op_buffer = sendtmpbuf + (segindex-1) * segment_increment;
189                         }
190                     }
191                     if(op!=MPI_OP_NULL) op->apply( local_op_buffer, accumulator, &prevcount,
192                                    datatype );
193
194                     /* all reduced on available data this step (i) complete,
195                      * pass to the next process unless you are the root.
196                      */
197                     if (rank != tree->tree_root) {
198                         /* send combined/accumulated data to parent */
199                         Request::send( accumulator, prevcount,
200                                                   datatype, tree->tree_prev,
201                                                   COLL_TAG_REDUCE,
202                                                   comm);
203                     }
204
205                     /* we stop when segindex = number of segments
206                        (i.e. we do num_segment+1 steps for pipelining */
207                     if (segindex == num_segments) break;
208                 }
209
210                 /* update input buffer index */
211                 inbi = inbi ^ 1;
212             } /* end of for each child */
213         } /* end of for each segment */
214
215         /* clean up */
216         smpi_free_tmp_buffer(inbuf_free[0]);
217         smpi_free_tmp_buffer(inbuf_free[1]);
218         smpi_free_tmp_buffer(accumbuf_free);
219     }
220
221     /* leaf nodes
222        Depending on the value of max_outstanding_reqs and
223        the number of segments we have two options:
224        - send all segments using blocking send to the parent, or
225        - avoid overflooding the parent nodes by limiting the number of
226        outstanding requests to max_oustanding_reqs.
227        TODO/POSSIBLE IMPROVEMENT: If there is a way to determine the eager size
228        for the current communication, synchronization should be used only
229        when the message/segment size is smaller than the eager size.
230     */
231     else {
232
233         /* If the number of segments is less than a maximum number of oustanding
234            requests or there is no limit on the maximum number of outstanding
235            requests, we send data to the parent using blocking send */
236         if ((0 == max_outstanding_reqs) ||
237             (num_segments <= max_outstanding_reqs)) {
238
239             segindex = 0;
240             while ( original_count > 0) {
241                 if (original_count < count_by_segment) {
242                     count_by_segment = original_count;
243                 }
244                 Request::send((char*)sendbuf +
245                                          segindex * segment_increment,
246                                          count_by_segment, datatype,
247                                          tree->tree_prev,
248                                          COLL_TAG_REDUCE,
249                                          comm) ;
250                 segindex++;
251                 original_count -= count_by_segment;
252             }
253         }
254
255         /* Otherwise, introduce flow control:
256            - post max_outstanding_reqs non-blocking synchronous send,
257            - for remaining segments
258            - wait for a ssend to complete, and post the next one.
259            - wait for all outstanding sends to complete.
260         */
261         else {
262
263             int creq = 0;
264             MPI_Request* sreq = NULL;
265
266             sreq = (MPI_Request*) calloc( max_outstanding_reqs,
267                                               sizeof(MPI_Request ) );
268             if (NULL == sreq) { line = __LINE__; ret = -1; goto error_hndl; }
269
270             /* post first group of requests */
271             for (segindex = 0; segindex < max_outstanding_reqs; segindex++) {
272                 sreq[segindex]=Request::isend((char*)sendbuf +
273                                           segindex * segment_increment,
274                                           count_by_segment, datatype,
275                                           tree->tree_prev,
276                                           COLL_TAG_REDUCE,
277                                           comm);
278                 original_count -= count_by_segment;
279             }
280
281             creq = 0;
282             while ( original_count > 0 ) {
283                 /* wait on a posted request to complete */
284                 Request::wait(&sreq[creq], MPI_STATUS_IGNORE);
285                 sreq[creq] = MPI_REQUEST_NULL;
286
287                 if( original_count < count_by_segment ) {
288                     count_by_segment = original_count;
289                 }
290                 sreq[creq]=Request::isend((char*)sendbuf +
291                                           segindex * segment_increment,
292                                           count_by_segment, datatype,
293                                           tree->tree_prev,
294                                           COLL_TAG_REDUCE,
295                                           comm );
296                 creq = (creq + 1) % max_outstanding_reqs;
297                 segindex++;
298                 original_count -= count_by_segment;
299             }
300
301             /* Wait on the remaining request to complete */
302             Request::waitall( max_outstanding_reqs, sreq,
303                                          MPI_STATUSES_IGNORE );
304
305             /* free requests */
306             free(sreq);
307         }
308     }
309     free(tree);
310     return MPI_SUCCESS;
311
312  error_hndl:  /* error handler */
313     XBT_DEBUG("ERROR_HNDL: node %d file %s line %d error %d\n",
314                    rank, __FILE__, line, ret );
315     if( inbuf_free[0] != NULL ) free(inbuf_free[0]);
316     if( inbuf_free[1] != NULL ) free(inbuf_free[1]);
317     if( accumbuf_free != NULL ) free(accumbuf);
318     return ret;
319 }
320
321 /* Attention: this version of the reduce operations does not
322    work for:
323    - non-commutative operations
324    - segment sizes which are not multiplies of the extent of the datatype
325      meaning that at least one datatype must fit in the segment !
326 */
327
328
329 int Coll_reduce_ompi_chain::reduce( void *sendbuf, void *recvbuf, int count,
330                                         MPI_Datatype datatype,
331                                         MPI_Op  op, int root,
332                                         MPI_Comm  comm
333                                         )
334 {
335     uint32_t segsize=64*1024;
336     int segcount = count;
337     size_t typelng;
338     int fanout = comm->size()/2;
339
340     XBT_DEBUG("coll:tuned:reduce_intra_chain rank %d fo %d ss %5u", comm->rank(), fanout, segsize);
341
342     /**
343      * Determine number of segments and number of elements
344      * sent per operation
345      */
346     typelng = datatype->size();
347
348     COLL_TUNED_COMPUTED_SEGCOUNT( segsize, typelng, segcount );
349
350     return smpi_coll_tuned_ompi_reduce_generic( sendbuf, recvbuf, count, datatype,
351                                            op, root, comm,
352                                            ompi_coll_tuned_topo_build_chain(fanout, comm, root),
353                                            segcount, 0 );
354 }
355
356
357 int Coll_reduce_ompi_pipeline::reduce( void *sendbuf, void *recvbuf,
358                                            int count, MPI_Datatype datatype,
359                                            MPI_Op  op, int root,
360                                            MPI_Comm  comm  )
361 {
362
363     uint32_t segsize;
364     int segcount = count;
365     size_t typelng;
366 //    COLL_TUNED_UPDATE_PIPELINE( comm, tuned_module, root );
367
368     /**
369      * Determine number of segments and number of elements
370      * sent per operation
371      */
372     const double a2 =  0.0410 / 1024.0; /* [1/B] */
373     const double b2 =  9.7128;
374     const double a4 =  0.0033 / 1024.0; /* [1/B] */
375     const double b4 =  1.6761;
376     typelng= datatype->size();
377     int communicator_size = comm->size();
378     size_t message_size = typelng * count;
379
380     if (communicator_size > (a2 * message_size + b2)) {
381         // Pipeline_1K
382         segsize = 1024;
383     }else if (communicator_size > (a4 * message_size + b4)) {
384         // Pipeline_32K
385         segsize = 32*1024;
386     } else {
387         // Pipeline_64K
388         segsize = 64*1024;
389     }
390
391     XBT_DEBUG("coll:tuned:reduce_intra_pipeline rank %d ss %5u", comm->rank(), segsize);
392
393     COLL_TUNED_COMPUTED_SEGCOUNT( segsize, typelng, segcount );
394
395     return smpi_coll_tuned_ompi_reduce_generic( sendbuf, recvbuf, count, datatype,
396                                            op, root, comm,
397                                            ompi_coll_tuned_topo_build_chain( 1, comm, root),
398                                            segcount, 0);
399 }
400
401 int Coll_reduce_ompi_binary::reduce( void *sendbuf, void *recvbuf,
402                                          int count, MPI_Datatype datatype,
403                                          MPI_Op  op, int root,
404                                          MPI_Comm  comm)
405 {
406     uint32_t segsize;
407     int segcount = count;
408     size_t typelng;
409
410
411
412     /**
413      * Determine number of segments and number of elements
414      * sent per operation
415      */
416     typelng=datatype->size();
417
418         // Binary_32K
419     segsize = 32*1024;
420
421     XBT_DEBUG("coll:tuned:reduce_intra_binary rank %d ss %5u", comm->rank(), segsize);
422
423     COLL_TUNED_COMPUTED_SEGCOUNT( segsize, typelng, segcount );
424
425     return smpi_coll_tuned_ompi_reduce_generic( sendbuf, recvbuf, count, datatype,
426                                            op, root, comm,
427                                            ompi_coll_tuned_topo_build_tree(2, comm, root),
428                                            segcount, 0);
429 }
430
431 int Coll_reduce_ompi_binomial::reduce( void *sendbuf, void *recvbuf,
432                                            int count, MPI_Datatype datatype,
433                                            MPI_Op  op, int root,
434                                            MPI_Comm  comm)
435 {
436
437     uint32_t segsize=0;
438     int segcount = count;
439     size_t typelng;
440
441     const double a1 =  0.6016 / 1024.0; /* [1/B] */
442     const double b1 =  1.3496;
443
444 //    COLL_TUNED_UPDATE_IN_ORDER_BMTREE( comm, tuned_module, root );
445
446     /**
447      * Determine number of segments and number of elements
448      * sent per operation
449      */
450     typelng= datatype->size();
451     int communicator_size = comm->size();
452     size_t message_size = typelng * count;
453     if (((communicator_size < 8) && (message_size < 20480)) ||
454                (message_size < 2048) || (count <= 1)) {
455         /* Binomial_0K */
456         segsize = 0;
457     } else if (communicator_size > (a1 * message_size + b1)) {
458         // Binomial_1K
459         segsize = 1024;
460     }
461
462     XBT_DEBUG("coll:tuned:reduce_intra_binomial rank %d ss %5u", comm->rank(), segsize);
463     COLL_TUNED_COMPUTED_SEGCOUNT( segsize, typelng, segcount );
464
465     return smpi_coll_tuned_ompi_reduce_generic( sendbuf, recvbuf, count, datatype,
466                                            op, root, comm,
467                                            ompi_coll_tuned_topo_build_in_order_bmtree(comm, root),
468                                            segcount, 0);
469 }
470
471 /*
472  * reduce_intra_in_order_binary
473  *
474  * Function:      Logarithmic reduce operation for non-commutative operations.
475  * Acecpts:       same as MPI_Reduce()
476  * Returns:       MPI_SUCCESS or error code
477  */
478 int Coll_reduce_ompi_in_order_binary::reduce( void *sendbuf, void *recvbuf,
479                                                   int count,
480                                                   MPI_Datatype datatype,
481                                                   MPI_Op  op, int root,
482                                                   MPI_Comm  comm)
483 {
484     uint32_t segsize=0;
485     int ret;
486     int rank, size, io_root;
487     int segcount = count;
488     void *use_this_sendbuf = NULL, *use_this_recvbuf = NULL;
489     size_t typelng;
490
491     rank = comm->rank();
492     size = comm->size();
493     XBT_DEBUG("coll:tuned:reduce_intra_in_order_binary rank %d ss %5u", rank, segsize);
494
495     /**
496      * Determine number of segments and number of elements
497      * sent per operation
498      */
499     typelng=datatype->size();
500     COLL_TUNED_COMPUTED_SEGCOUNT( segsize, typelng, segcount );
501
502     /* An in-order binary tree must use root (size-1) to preserve the order of
503        operations.  Thus, if root is not rank (size - 1), then we must handle
504        1. MPI_IN_PLACE option on real root, and
505        2. we must allocate temporary recvbuf on rank (size - 1).
506        Note that generic function must be careful not to switch order of
507        operations for non-commutative ops.
508     */
509     io_root = size - 1;
510     use_this_sendbuf = sendbuf;
511     use_this_recvbuf = recvbuf;
512     if (io_root != root) {
513         ptrdiff_t text, ext;
514         char *tmpbuf = NULL;
515
516         ext=datatype->get_extent();
517         text=datatype->get_extent();
518
519         if ((root == rank) && (MPI_IN_PLACE == sendbuf)) {
520             tmpbuf = (char *) smpi_get_tmp_sendbuffer(text + (count - 1) * ext);
521             if (NULL == tmpbuf) {
522                 return MPI_ERR_INTERN;
523             }
524             Datatype::copy (
525                                                 (char*)recvbuf, count, datatype,
526                                                 (char*)tmpbuf, count, datatype);
527             use_this_sendbuf = tmpbuf;
528         } else if (io_root == rank) {
529             tmpbuf = (char *) smpi_get_tmp_recvbuffer(text + (count - 1) * ext);
530             if (NULL == tmpbuf) {
531                 return MPI_ERR_INTERN;
532             }
533             use_this_recvbuf = tmpbuf;
534         }
535     }
536
537     /* Use generic reduce with in-order binary tree topology and io_root */
538     ret = smpi_coll_tuned_ompi_reduce_generic( use_this_sendbuf, use_this_recvbuf, count, datatype,
539                                           op, io_root, comm,
540                                           ompi_coll_tuned_topo_build_in_order_bintree(comm),
541                                           segcount, 0 );
542     if (MPI_SUCCESS != ret) { return ret; }
543
544     /* Clean up */
545     if (io_root != root) {
546         if (root == rank) {
547             /* Receive result from rank io_root to recvbuf */
548             Request::recv(recvbuf, count, datatype, io_root,
549                                     COLL_TAG_REDUCE, comm,
550                                     MPI_STATUS_IGNORE);
551             if (MPI_IN_PLACE == sendbuf) {
552               smpi_free_tmp_buffer(use_this_sendbuf);
553             }
554
555         } else if (io_root == rank) {
556             /* Send result from use_this_recvbuf to root */
557             Request::send(use_this_recvbuf, count, datatype, root,
558                                     COLL_TAG_REDUCE,
559                                     comm);
560             smpi_free_tmp_buffer(use_this_recvbuf);
561         }
562     }
563
564     return MPI_SUCCESS;
565 }
566
567 /*
568  * Linear functions are copied from the BASIC coll module
569  * they do not segment the message and are simple implementations
570  * but for some small number of nodes and/or small data sizes they
571  * are just as fast as tuned/tree based segmenting operations
572  * and as such may be selected by the decision functions
573  * These are copied into this module due to the way we select modules
574  * in V1. i.e. in V2 we will handle this differently and so will not
575  * have to duplicate code.
576  * GEF Oct05 after asking Jeff.
577  */
578
579 /* copied function (with appropriate renaming) starts here */
580
581 /*
582  *  reduce_lin_intra
583  *
584  *  Function:   - reduction using O(N) algorithm
585  *  Accepts:    - same as MPI_Reduce()
586  *  Returns:    - MPI_SUCCESS or error code
587  */
588
589 int
590 Coll_reduce_ompi_basic_linear::reduce(void *sbuf, void *rbuf, int count,
591                                           MPI_Datatype dtype,
592                                           MPI_Op op,
593                                           int root,
594                                           MPI_Comm comm)
595 {
596     int i, rank, size;
597     ptrdiff_t true_extent, lb, extent;
598     char *free_buffer = NULL;
599     char *pml_buffer = NULL;
600     char *inplace_temp = NULL;
601     char *inbuf;
602
603     /* Initialize */
604
605     rank = comm->rank();
606     size = comm->size();
607
608     XBT_DEBUG("coll:tuned:reduce_intra_basic_linear rank %d", rank);
609
610     /* If not root, send data to the root. */
611
612     if (rank != root) {
613         Request::send(sbuf, count, dtype, root,
614                                 COLL_TAG_REDUCE,
615                                 comm);
616         return MPI_SUCCESS;
617     }
618
619     /* see discussion in ompi_coll_basic_reduce_lin_intra about
620        extent and true extent */
621     /* for reducing buffer allocation lengths.... */
622
623     dtype->extent(&lb, &extent);
624     true_extent = dtype->get_extent();
625
626     if (MPI_IN_PLACE == sbuf) {
627         sbuf = rbuf;
628         inplace_temp = (char*)smpi_get_tmp_recvbuffer(true_extent + (count - 1) * extent);
629         if (NULL == inplace_temp) {
630             return -1;
631         }
632         rbuf = inplace_temp - lb;
633     }
634
635     if (size > 1) {
636         free_buffer = (char*)smpi_get_tmp_recvbuffer(true_extent + (count - 1) * extent);
637         pml_buffer = free_buffer - lb;
638     }
639
640     /* Initialize the receive buffer. */
641
642     if (rank == (size - 1)) {
643         Datatype::copy((char*)sbuf, count, dtype,(char*)rbuf, count, dtype);
644     } else {
645         Request::recv(rbuf, count, dtype, size - 1,
646                                 COLL_TAG_REDUCE, comm,
647                                 MPI_STATUS_IGNORE);
648     }
649
650     /* Loop receiving and calling reduction function (C or Fortran). */
651
652     for (i = size - 2; i >= 0; --i) {
653         if (rank == i) {
654             inbuf = (char*)sbuf;
655         } else {
656             Request::recv(pml_buffer, count, dtype, i,
657                                     COLL_TAG_REDUCE, comm,
658                                     MPI_STATUS_IGNORE);
659             inbuf = pml_buffer;
660         }
661
662         /* Perform the reduction */
663         if(op!=MPI_OP_NULL) op->apply( inbuf, rbuf, &count, dtype);
664     }
665
666     if (NULL != inplace_temp) {
667         Datatype::copy(inplace_temp, count, dtype,(char*)sbuf
668                                                   ,count , dtype);
669         smpi_free_tmp_buffer(inplace_temp);
670     }
671     if (NULL != free_buffer) {
672         smpi_free_tmp_buffer(free_buffer);
673     }
674
675     /* All done */
676     return MPI_SUCCESS;
677 }
678
679 /* copied function (with appropriate renaming) ends here */
680
681
682 }
683 }