1 /* smpi_mpi_dt.c -- MPI primitives to handle datatypes */
2 /* FIXME: a very incomplete implementation */
4 /* Copyright (c) 2009-2014. The SimGrid Team.
5 * All rights reserved. */
7 /* This program is free software; you can redistribute it and/or modify it
8 * under the terms of the license (GNU LGPL) which comes with this package. */
15 #include "smpi_mpi_dt_private.h"
17 #include "xbt/replay.h"
18 #include "simgrid/modelchecker.h"
20 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_mpi_dt, smpi,
21 "Logging specific to SMPI (datatype)");
23 #define CREATE_MPI_DATATYPE(name, type) \
24 static s_smpi_mpi_datatype_t mpi_##name = { \
25 sizeof(type), /* size */ \
26 0, /*was 1 has_subtype*/ \
28 sizeof(type), /* ub = lb + size */ \
29 DT_FLAG_BASIC, /* flags */ \
30 NULL /* pointer on extended struct*/ \
32 MPI_Datatype name = &mpi_##name;
34 #define CREATE_MPI_DATATYPE_NULL(name) \
35 static s_smpi_mpi_datatype_t mpi_##name = { \
37 0, /*was 1 has_subtype*/ \
39 0, /* ub = lb + size */ \
40 DT_FLAG_BASIC, /* flags */ \
41 NULL /* pointer on extended struct*/ \
43 MPI_Datatype name = &mpi_##name;
45 //The following are datatypes for the MPI functions MPI_MAXLOC and MPI_MINLOC.
79 // Predefined data types
80 CREATE_MPI_DATATYPE(MPI_CHAR, char);
81 CREATE_MPI_DATATYPE(MPI_SHORT, short);
82 CREATE_MPI_DATATYPE(MPI_INT, int);
83 CREATE_MPI_DATATYPE(MPI_LONG, long);
84 CREATE_MPI_DATATYPE(MPI_LONG_LONG, long long);
85 CREATE_MPI_DATATYPE(MPI_SIGNED_CHAR, signed char);
86 CREATE_MPI_DATATYPE(MPI_UNSIGNED_CHAR, unsigned char);
87 CREATE_MPI_DATATYPE(MPI_UNSIGNED_SHORT, unsigned short);
88 CREATE_MPI_DATATYPE(MPI_UNSIGNED, unsigned int);
89 CREATE_MPI_DATATYPE(MPI_UNSIGNED_LONG, unsigned long);
90 CREATE_MPI_DATATYPE(MPI_UNSIGNED_LONG_LONG, unsigned long long);
91 CREATE_MPI_DATATYPE(MPI_FLOAT, float);
92 CREATE_MPI_DATATYPE(MPI_DOUBLE, double);
93 CREATE_MPI_DATATYPE(MPI_LONG_DOUBLE, long double);
94 CREATE_MPI_DATATYPE(MPI_WCHAR, wchar_t);
95 CREATE_MPI_DATATYPE(MPI_C_BOOL, _Bool);
96 CREATE_MPI_DATATYPE(MPI_INT8_T, int8_t);
97 CREATE_MPI_DATATYPE(MPI_INT16_T, int16_t);
98 CREATE_MPI_DATATYPE(MPI_INT32_T, int32_t);
99 CREATE_MPI_DATATYPE(MPI_INT64_T, int64_t);
100 CREATE_MPI_DATATYPE(MPI_UINT8_T, uint8_t);
101 CREATE_MPI_DATATYPE(MPI_UINT16_T, uint16_t);
102 CREATE_MPI_DATATYPE(MPI_UINT32_T, uint32_t);
103 CREATE_MPI_DATATYPE(MPI_UINT64_T, uint64_t);
104 CREATE_MPI_DATATYPE(MPI_C_FLOAT_COMPLEX, float _Complex);
105 CREATE_MPI_DATATYPE(MPI_C_DOUBLE_COMPLEX, double _Complex);
106 CREATE_MPI_DATATYPE(MPI_C_LONG_DOUBLE_COMPLEX, long double _Complex);
107 CREATE_MPI_DATATYPE(MPI_AINT, MPI_Aint);
108 CREATE_MPI_DATATYPE(MPI_OFFSET, MPI_Offset);
110 CREATE_MPI_DATATYPE(MPI_FLOAT_INT, float_int);
111 CREATE_MPI_DATATYPE(MPI_LONG_INT, long_int);
112 CREATE_MPI_DATATYPE(MPI_DOUBLE_INT, double_int);
113 CREATE_MPI_DATATYPE(MPI_SHORT_INT, short_int);
114 CREATE_MPI_DATATYPE(MPI_2INT, int_int);
115 CREATE_MPI_DATATYPE(MPI_2FLOAT, float_float);
116 CREATE_MPI_DATATYPE(MPI_2DOUBLE, double_double);
118 CREATE_MPI_DATATYPE(MPI_LONG_DOUBLE_INT, long_double_int);
120 CREATE_MPI_DATATYPE_NULL(MPI_UB);
121 CREATE_MPI_DATATYPE_NULL(MPI_LB);
122 CREATE_MPI_DATATYPE_NULL(MPI_PACKED);
124 CREATE_MPI_DATATYPE(MPI_PTR, void*);
126 /** Check if the datatype is usable for communications
128 int is_datatype_valid(MPI_Datatype datatype) {
129 return datatype != MPI_DATATYPE_NULL
130 && (datatype->flags & DT_FLAG_COMMITED);
133 size_t smpi_datatype_size(MPI_Datatype datatype)
135 return datatype->size;
138 MPI_Aint smpi_datatype_lb(MPI_Datatype datatype)
143 MPI_Aint smpi_datatype_ub(MPI_Datatype datatype)
148 MPI_Datatype smpi_datatype_dup(MPI_Datatype datatype)
150 MPI_Datatype new_t= xbt_new(s_smpi_mpi_datatype_t,1);
151 memcpy(new_t, datatype, sizeof(s_smpi_mpi_datatype_t));
152 if (datatype->has_subtype)
153 memcpy(new_t->substruct, datatype->substruct, sizeof(s_smpi_subtype_t));
157 int smpi_datatype_extent(MPI_Datatype datatype, MPI_Aint * lb,
161 *extent = datatype->ub - datatype->lb;
165 MPI_Aint smpi_datatype_get_extent(MPI_Datatype datatype){
166 return datatype->ub - datatype->lb;
169 int smpi_datatype_copy(void *sendbuf, int sendcount, MPI_Datatype sendtype,
170 void *recvbuf, int recvcount, MPI_Datatype recvtype)
173 if(smpi_privatize_global_variables){
174 switch_data_segment(smpi_process_index());
176 /* First check if we really have something to do */
177 if (recvcount > 0 && recvbuf != sendbuf) {
178 /* FIXME: treat packed cases */
179 sendcount *= smpi_datatype_size(sendtype);
180 recvcount *= smpi_datatype_size(recvtype);
181 count = sendcount < recvcount ? sendcount : recvcount;
183 if(sendtype->has_subtype == 0 && recvtype->has_subtype == 0) {
184 if(!_xbt_replay_is_active()) memcpy(recvbuf, sendbuf, count);
186 else if (sendtype->has_subtype == 0)
188 s_smpi_subtype_t *subtype = recvtype->substruct;
189 subtype->unserialize( sendbuf, recvbuf,1, subtype);
191 else if (recvtype->has_subtype == 0)
193 s_smpi_subtype_t *subtype = sendtype->substruct;
194 subtype->serialize(sendbuf, recvbuf,1, subtype);
196 s_smpi_subtype_t *subtype = sendtype->substruct;
199 void * buf_tmp = xbt_malloc(count);
201 subtype->serialize( sendbuf, buf_tmp,count/smpi_datatype_size(sendtype), subtype);
202 subtype = recvtype->substruct;
203 subtype->unserialize( buf_tmp, recvbuf,count/smpi_datatype_size(recvtype), subtype);
209 return sendcount > recvcount ? MPI_ERR_TRUNCATE : MPI_SUCCESS;
213 * Copies noncontiguous data into contiguous memory.
214 * @param contiguous_vector - output vector
215 * @param noncontiguous_vector - input vector
216 * @param type - pointer contening :
217 * - stride - stride of between noncontiguous data
218 * - block_length - the width or height of blocked matrix
219 * - count - the number of rows of matrix
221 void serialize_vector( const void *noncontiguous_vector,
222 void *contiguous_vector,
226 s_smpi_mpi_vector_t* type_c = (s_smpi_mpi_vector_t*)type;
228 char* contiguous_vector_char = (char*)contiguous_vector;
229 char* noncontiguous_vector_char = (char*)noncontiguous_vector;
231 for (i = 0; i < type_c->block_count * count; i++) {
232 if (type_c->old_type->has_subtype == 0)
233 memcpy(contiguous_vector_char,
234 noncontiguous_vector_char, type_c->block_length * type_c->size_oldtype);
236 ((s_smpi_subtype_t*)type_c->old_type->substruct)->serialize( noncontiguous_vector_char,
237 contiguous_vector_char,
238 type_c->block_length,
239 type_c->old_type->substruct);
241 contiguous_vector_char += type_c->block_length*type_c->size_oldtype;
242 if((i+1)%type_c->block_count ==0)
243 noncontiguous_vector_char += type_c->block_length*smpi_datatype_get_extent(type_c->old_type);
245 noncontiguous_vector_char += type_c->block_stride*smpi_datatype_get_extent(type_c->old_type);
250 * Copies contiguous data into noncontiguous memory.
251 * @param noncontiguous_vector - output vector
252 * @param contiguous_vector - input vector
253 * @param type - pointer contening :
254 * - stride - stride of between noncontiguous data
255 * - block_length - the width or height of blocked matrix
256 * - count - the number of rows of matrix
258 void unserialize_vector( const void *contiguous_vector,
259 void *noncontiguous_vector,
263 s_smpi_mpi_vector_t* type_c = (s_smpi_mpi_vector_t*)type;
266 char* contiguous_vector_char = (char*)contiguous_vector;
267 char* noncontiguous_vector_char = (char*)noncontiguous_vector;
269 for (i = 0; i < type_c->block_count * count; i++) {
270 if (type_c->old_type->has_subtype == 0)
271 memcpy(noncontiguous_vector_char,
272 contiguous_vector_char, type_c->block_length * type_c->size_oldtype);
274 ((s_smpi_subtype_t*)type_c->old_type->substruct)->unserialize( contiguous_vector_char,
275 noncontiguous_vector_char,
276 type_c->block_length,
277 type_c->old_type->substruct);
278 contiguous_vector_char += type_c->block_length*type_c->size_oldtype;
279 if((i+1)%type_c->block_count ==0)
280 noncontiguous_vector_char += type_c->block_length*smpi_datatype_get_extent(type_c->old_type);
282 noncontiguous_vector_char += type_c->block_stride*smpi_datatype_get_extent(type_c->old_type);
287 * Create a Sub type vector to be able to serialize and unserialize it
288 * the structure s_smpi_mpi_vector_t is derived from s_smpi_subtype which
289 * required the functions unserialize and serialize
292 s_smpi_mpi_vector_t* smpi_datatype_vector_create( int block_stride,
295 MPI_Datatype old_type,
297 s_smpi_mpi_vector_t *new_t= xbt_new(s_smpi_mpi_vector_t,1);
298 new_t->base.serialize = &serialize_vector;
299 new_t->base.unserialize = &unserialize_vector;
300 new_t->base.subtype_free = &free_vector;
301 new_t->block_stride = block_stride;
302 new_t->block_length = block_length;
303 new_t->block_count = block_count;
304 smpi_datatype_use(old_type);
305 new_t->old_type = old_type;
306 new_t->size_oldtype = size_oldtype;
310 void smpi_datatype_create(MPI_Datatype* new_type, int size,int lb, int ub, int has_subtype,
311 void *struct_type, int flags){
312 MPI_Datatype new_t= xbt_new(s_smpi_mpi_datatype_t,1);
314 new_t->has_subtype = size>0? has_subtype:0;
317 new_t->flags = flags;
318 new_t->substruct = struct_type;
324 MC_ignore(&(new_t->in_use), sizeof(new_t->in_use));
328 void smpi_datatype_free(MPI_Datatype* type){
330 if((*type)->flags & DT_FLAG_PREDEFINED)return;
332 //if still used, mark for deletion
333 if((*type)->in_use!=0){
334 (*type)->flags |=DT_FLAG_DESTROYED;
338 if ((*type)->has_subtype == 1){
339 ((s_smpi_subtype_t *)(*type)->substruct)->subtype_free(type);
340 xbt_free((*type)->substruct);
343 *type = MPI_DATATYPE_NULL;
346 void smpi_datatype_use(MPI_Datatype type){
347 if(type)type->in_use++;
351 MC_ignore(&(type->in_use), sizeof(type->in_use));
356 void smpi_datatype_unuse(MPI_Datatype type){
357 if(type && type->in_use-- == 0 && (type->flags & DT_FLAG_DESTROYED))
358 smpi_datatype_free(&type);
362 MC_ignore(&(type->in_use), sizeof(type->in_use));
370 Contiguous Implementation
375 * Copies noncontiguous data into contiguous memory.
376 * @param contiguous_hvector - output hvector
377 * @param noncontiguous_hvector - input hvector
378 * @param type - pointer contening :
379 * - stride - stride of between noncontiguous data, in bytes
380 * - block_length - the width or height of blocked matrix
381 * - count - the number of rows of matrix
383 void serialize_contiguous( const void *noncontiguous_hvector,
384 void *contiguous_hvector,
388 s_smpi_mpi_contiguous_t* type_c = (s_smpi_mpi_contiguous_t*)type;
389 char* contiguous_vector_char = (char*)contiguous_hvector;
390 char* noncontiguous_vector_char = (char*)noncontiguous_hvector+type_c->lb;
391 memcpy(contiguous_vector_char,
392 noncontiguous_vector_char, count* type_c->block_count * type_c->size_oldtype);
395 * Copies contiguous data into noncontiguous memory.
396 * @param noncontiguous_vector - output hvector
397 * @param contiguous_vector - input hvector
398 * @param type - pointer contening :
399 * - stride - stride of between noncontiguous data, in bytes
400 * - block_length - the width or height of blocked matrix
401 * - count - the number of rows of matrix
403 void unserialize_contiguous( const void *contiguous_vector,
404 void *noncontiguous_vector,
408 s_smpi_mpi_contiguous_t* type_c = (s_smpi_mpi_contiguous_t*)type;
409 char* contiguous_vector_char = (char*)contiguous_vector;
410 char* noncontiguous_vector_char = (char*)noncontiguous_vector+type_c->lb;
412 memcpy(noncontiguous_vector_char,
413 contiguous_vector_char, count* type_c->block_count * type_c->size_oldtype);
416 void free_contiguous(MPI_Datatype* d){
417 smpi_datatype_unuse(((s_smpi_mpi_indexed_t *)(*d)->substruct)->old_type);
421 * Create a Sub type contiguous to be able to serialize and unserialize it
422 * the structure s_smpi_mpi_contiguous_t is derived from s_smpi_subtype which
423 * required the functions unserialize and serialize
426 s_smpi_mpi_contiguous_t* smpi_datatype_contiguous_create( MPI_Aint lb,
428 MPI_Datatype old_type,
430 s_smpi_mpi_contiguous_t *new_t= xbt_new(s_smpi_mpi_contiguous_t,1);
431 new_t->base.serialize = &serialize_contiguous;
432 new_t->base.unserialize = &unserialize_contiguous;
433 new_t->base.subtype_free = &free_contiguous;
435 new_t->block_count = block_count;
436 new_t->old_type = old_type;
437 new_t->size_oldtype = size_oldtype;
438 smpi_datatype_use(old_type);
445 int smpi_datatype_contiguous(int count, MPI_Datatype old_type, MPI_Datatype* new_type, MPI_Aint lb)
448 if(old_type->has_subtype){
449 //handle this case as a hvector with stride equals to the extent of the datatype
450 return smpi_datatype_hvector(count, 1, smpi_datatype_get_extent(old_type), old_type, new_type);
453 s_smpi_mpi_contiguous_t* subtype = smpi_datatype_contiguous_create( lb,
456 smpi_datatype_size(old_type));
458 smpi_datatype_create(new_type,
459 count * smpi_datatype_size(old_type),
460 lb,lb + count * smpi_datatype_size(old_type),
461 1,subtype, DT_FLAG_CONTIGUOUS);
466 int smpi_datatype_vector(int count, int blocklen, int stride, MPI_Datatype old_type, MPI_Datatype* new_type)
469 if (blocklen<0) return MPI_ERR_ARG;
473 lb=smpi_datatype_lb(old_type);
474 ub=((count-1)*stride+blocklen-1)*smpi_datatype_get_extent(old_type)+smpi_datatype_ub(old_type);
476 if(old_type->has_subtype || stride != blocklen){
479 s_smpi_mpi_vector_t* subtype = smpi_datatype_vector_create( stride,
483 smpi_datatype_size(old_type));
484 smpi_datatype_create(new_type,
485 count * (blocklen) * smpi_datatype_size(old_type), lb,
492 /* in this situation the data are contignous thus it's not
493 * required to serialize and unserialize it*/
494 smpi_datatype_create(new_type, count * blocklen *
495 smpi_datatype_size(old_type), 0, ((count -1) * stride + blocklen)*
496 smpi_datatype_size(old_type),
499 DT_FLAG_VECTOR|DT_FLAG_CONTIGUOUS);
505 void free_vector(MPI_Datatype* d){
506 smpi_datatype_unuse(((s_smpi_mpi_indexed_t *)(*d)->substruct)->old_type);
510 Hvector Implementation - Vector with stride in bytes
515 * Copies noncontiguous data into contiguous memory.
516 * @param contiguous_hvector - output hvector
517 * @param noncontiguous_hvector - input hvector
518 * @param type - pointer contening :
519 * - stride - stride of between noncontiguous data, in bytes
520 * - block_length - the width or height of blocked matrix
521 * - count - the number of rows of matrix
523 void serialize_hvector( const void *noncontiguous_hvector,
524 void *contiguous_hvector,
528 s_smpi_mpi_hvector_t* type_c = (s_smpi_mpi_hvector_t*)type;
530 char* contiguous_vector_char = (char*)contiguous_hvector;
531 char* noncontiguous_vector_char = (char*)noncontiguous_hvector;
533 for (i = 0; i < type_c->block_count * count; i++) {
534 if (type_c->old_type->has_subtype == 0)
535 memcpy(contiguous_vector_char,
536 noncontiguous_vector_char, type_c->block_length * type_c->size_oldtype);
538 ((s_smpi_subtype_t*)type_c->old_type->substruct)->serialize( noncontiguous_vector_char,
539 contiguous_vector_char,
540 type_c->block_length,
541 type_c->old_type->substruct);
543 contiguous_vector_char += type_c->block_length*type_c->size_oldtype;
544 if((i+1)%type_c->block_count ==0)
545 noncontiguous_vector_char += type_c->block_length*type_c->size_oldtype;
547 noncontiguous_vector_char += type_c->block_stride;
551 * Copies contiguous data into noncontiguous memory.
552 * @param noncontiguous_vector - output hvector
553 * @param contiguous_vector - input hvector
554 * @param type - pointer contening :
555 * - stride - stride of between noncontiguous data, in bytes
556 * - block_length - the width or height of blocked matrix
557 * - count - the number of rows of matrix
559 void unserialize_hvector( const void *contiguous_vector,
560 void *noncontiguous_vector,
564 s_smpi_mpi_hvector_t* type_c = (s_smpi_mpi_hvector_t*)type;
567 char* contiguous_vector_char = (char*)contiguous_vector;
568 char* noncontiguous_vector_char = (char*)noncontiguous_vector;
570 for (i = 0; i < type_c->block_count * count; i++) {
571 if (type_c->old_type->has_subtype == 0)
572 memcpy(noncontiguous_vector_char,
573 contiguous_vector_char, type_c->block_length * type_c->size_oldtype);
575 ((s_smpi_subtype_t*)type_c->old_type->substruct)->unserialize( contiguous_vector_char,
576 noncontiguous_vector_char,
577 type_c->block_length,
578 type_c->old_type->substruct);
579 contiguous_vector_char += type_c->block_length*type_c->size_oldtype;
580 if((i+1)%type_c->block_count ==0)
581 noncontiguous_vector_char += type_c->block_length*type_c->size_oldtype;
583 noncontiguous_vector_char += type_c->block_stride;
588 * Create a Sub type vector to be able to serialize and unserialize it
589 * the structure s_smpi_mpi_vector_t is derived from s_smpi_subtype which
590 * required the functions unserialize and serialize
593 s_smpi_mpi_hvector_t* smpi_datatype_hvector_create( MPI_Aint block_stride,
596 MPI_Datatype old_type,
598 s_smpi_mpi_hvector_t *new_t= xbt_new(s_smpi_mpi_hvector_t,1);
599 new_t->base.serialize = &serialize_hvector;
600 new_t->base.unserialize = &unserialize_hvector;
601 new_t->base.subtype_free = &free_hvector;
602 new_t->block_stride = block_stride;
603 new_t->block_length = block_length;
604 new_t->block_count = block_count;
605 new_t->old_type = old_type;
606 new_t->size_oldtype = size_oldtype;
607 smpi_datatype_use(old_type);
611 //do nothing for vector types
612 void free_hvector(MPI_Datatype* d){
613 smpi_datatype_unuse(((s_smpi_mpi_indexed_t *)(*d)->substruct)->old_type);
616 int smpi_datatype_hvector(int count, int blocklen, MPI_Aint stride, MPI_Datatype old_type, MPI_Datatype* new_type)
619 if (blocklen<0) return MPI_ERR_ARG;
623 lb=smpi_datatype_lb(old_type);
624 ub=((count-1)*stride)+(blocklen-1)*smpi_datatype_get_extent(old_type)+smpi_datatype_ub(old_type);
626 if(old_type->has_subtype || stride != blocklen*smpi_datatype_get_extent(old_type)){
627 s_smpi_mpi_hvector_t* subtype = smpi_datatype_hvector_create( stride,
631 smpi_datatype_size(old_type));
633 smpi_datatype_create(new_type, count * blocklen * smpi_datatype_size(old_type),
640 smpi_datatype_create(new_type, count * blocklen *
641 smpi_datatype_size(old_type),0,count * blocklen *
642 smpi_datatype_size(old_type),
645 DT_FLAG_VECTOR|DT_FLAG_CONTIGUOUS);
653 Indexed Implementation
657 * Copies noncontiguous data into contiguous memory.
658 * @param contiguous_indexed - output indexed
659 * @param noncontiguous_indexed - input indexed
660 * @param type - pointer contening :
661 * - block_lengths - the width or height of blocked matrix
662 * - block_indices - indices of each data, in element
663 * - count - the number of rows of matrix
665 void serialize_indexed( const void *noncontiguous_indexed,
666 void *contiguous_indexed,
670 s_smpi_mpi_indexed_t* type_c = (s_smpi_mpi_indexed_t*)type;
672 char* contiguous_indexed_char = (char*)contiguous_indexed;
673 char* noncontiguous_indexed_char = (char*)noncontiguous_indexed+type_c->block_indices[0] * type_c->size_oldtype;
674 for(j=0; j<count;j++){
675 for (i = 0; i < type_c->block_count; i++) {
676 if (type_c->old_type->has_subtype == 0)
677 memcpy(contiguous_indexed_char,
678 noncontiguous_indexed_char, type_c->block_lengths[i] * type_c->size_oldtype);
680 ((s_smpi_subtype_t*)type_c->old_type->substruct)->serialize( noncontiguous_indexed_char,
681 contiguous_indexed_char,
682 type_c->block_lengths[i],
683 type_c->old_type->substruct);
686 contiguous_indexed_char += type_c->block_lengths[i]*type_c->size_oldtype;
687 if (i<type_c->block_count-1)noncontiguous_indexed_char = (char*)noncontiguous_indexed + type_c->block_indices[i+1]*smpi_datatype_get_extent(type_c->old_type);
688 else noncontiguous_indexed_char += type_c->block_lengths[i]*smpi_datatype_get_extent(type_c->old_type);
690 noncontiguous_indexed=(void*)noncontiguous_indexed_char;
694 * Copies contiguous data into noncontiguous memory.
695 * @param noncontiguous_indexed - output indexed
696 * @param contiguous_indexed - input indexed
697 * @param type - pointer contening :
698 * - block_lengths - the width or height of blocked matrix
699 * - block_indices - indices of each data, in element
700 * - count - the number of rows of matrix
702 void unserialize_indexed( const void *contiguous_indexed,
703 void *noncontiguous_indexed,
708 s_smpi_mpi_indexed_t* type_c = (s_smpi_mpi_indexed_t*)type;
710 char* contiguous_indexed_char = (char*)contiguous_indexed;
711 char* noncontiguous_indexed_char = (char*)noncontiguous_indexed+type_c->block_indices[0]*smpi_datatype_get_extent(type_c->old_type);
712 for(j=0; j<count;j++){
713 for (i = 0; i < type_c->block_count; i++) {
714 if (type_c->old_type->has_subtype == 0)
715 memcpy(noncontiguous_indexed_char ,
716 contiguous_indexed_char, type_c->block_lengths[i] * type_c->size_oldtype);
718 ((s_smpi_subtype_t*)type_c->old_type->substruct)->unserialize( contiguous_indexed_char,
719 noncontiguous_indexed_char,
720 type_c->block_lengths[i],
721 type_c->old_type->substruct);
723 contiguous_indexed_char += type_c->block_lengths[i]*type_c->size_oldtype;
724 if (i<type_c->block_count-1)
725 noncontiguous_indexed_char = (char*)noncontiguous_indexed + type_c->block_indices[i+1]*smpi_datatype_get_extent(type_c->old_type);
726 else noncontiguous_indexed_char += type_c->block_lengths[i]*smpi_datatype_get_extent(type_c->old_type);
728 noncontiguous_indexed=(void*)noncontiguous_indexed_char;
732 void free_indexed(MPI_Datatype* type){
733 xbt_free(((s_smpi_mpi_indexed_t *)(*type)->substruct)->block_lengths);
734 xbt_free(((s_smpi_mpi_indexed_t *)(*type)->substruct)->block_indices);
735 smpi_datatype_unuse(((s_smpi_mpi_indexed_t *)(*type)->substruct)->old_type);
739 * Create a Sub type indexed to be able to serialize and unserialize it
740 * the structure s_smpi_mpi_indexed_t is derived from s_smpi_subtype which
741 * required the functions unserialize and serialize
743 s_smpi_mpi_indexed_t* smpi_datatype_indexed_create( int* block_lengths,
746 MPI_Datatype old_type,
748 s_smpi_mpi_indexed_t *new_t= xbt_new(s_smpi_mpi_indexed_t,1);
749 new_t->base.serialize = &serialize_indexed;
750 new_t->base.unserialize = &unserialize_indexed;
751 new_t->base.subtype_free = &free_indexed;
752 //TODO : add a custom function for each time to clean these
753 new_t->block_lengths= xbt_new(int, block_count);
754 new_t->block_indices= xbt_new(int, block_count);
756 for(i=0;i<block_count;i++){
757 new_t->block_lengths[i]=block_lengths[i];
758 new_t->block_indices[i]=block_indices[i];
760 new_t->block_count = block_count;
761 smpi_datatype_use(old_type);
762 new_t->old_type = old_type;
763 new_t->size_oldtype = size_oldtype;
768 int smpi_datatype_indexed(int count, int* blocklens, int* indices, MPI_Datatype old_type, MPI_Datatype* new_type)
777 lb=indices[0]*smpi_datatype_get_extent(old_type);
778 ub=indices[0]*smpi_datatype_get_extent(old_type) + blocklens[0]*smpi_datatype_ub(old_type);
781 for(i=0; i< count; i++){
784 size += blocklens[i];
786 if(indices[i]*smpi_datatype_get_extent(old_type)+smpi_datatype_lb(old_type)<lb)
787 lb = indices[i]*smpi_datatype_get_extent(old_type)+smpi_datatype_lb(old_type);
788 if(indices[i]*smpi_datatype_get_extent(old_type)+blocklens[i]*smpi_datatype_ub(old_type)>ub)
789 ub = indices[i]*smpi_datatype_get_extent(old_type)+blocklens[i]*smpi_datatype_ub(old_type);
791 if ( (i< count -1) && (indices[i]+blocklens[i] != indices[i+1]) )contiguous=0;
793 if (old_type->has_subtype == 1)
797 s_smpi_mpi_indexed_t* subtype = smpi_datatype_indexed_create( blocklens,
801 smpi_datatype_size(old_type));
802 smpi_datatype_create(new_type, size *
803 smpi_datatype_size(old_type),lb,ub,1, subtype, DT_FLAG_DATA);
805 s_smpi_mpi_contiguous_t* subtype = smpi_datatype_contiguous_create( lb,
808 smpi_datatype_size(old_type));
809 smpi_datatype_create(new_type, size *
810 smpi_datatype_size(old_type),lb,ub,1, subtype, DT_FLAG_DATA|DT_FLAG_CONTIGUOUS);
818 Hindexed Implementation - Indexed with indices in bytes
822 * Copies noncontiguous data into contiguous memory.
823 * @param contiguous_hindexed - output hindexed
824 * @param noncontiguous_hindexed - input hindexed
825 * @param type - pointer contening :
826 * - block_lengths - the width or height of blocked matrix
827 * - block_indices - indices of each data, in bytes
828 * - count - the number of rows of matrix
830 void serialize_hindexed( const void *noncontiguous_hindexed,
831 void *contiguous_hindexed,
835 s_smpi_mpi_hindexed_t* type_c = (s_smpi_mpi_hindexed_t*)type;
837 char* contiguous_hindexed_char = (char*)contiguous_hindexed;
838 char* noncontiguous_hindexed_char = (char*)noncontiguous_hindexed+ type_c->block_indices[0];
839 for(j=0; j<count;j++){
840 for (i = 0; i < type_c->block_count; i++) {
841 if (type_c->old_type->has_subtype == 0)
842 memcpy(contiguous_hindexed_char,
843 noncontiguous_hindexed_char, type_c->block_lengths[i] * type_c->size_oldtype);
845 ((s_smpi_subtype_t*)type_c->old_type->substruct)->serialize( noncontiguous_hindexed_char,
846 contiguous_hindexed_char,
847 type_c->block_lengths[i],
848 type_c->old_type->substruct);
850 contiguous_hindexed_char += type_c->block_lengths[i]*type_c->size_oldtype;
851 if (i<type_c->block_count-1)noncontiguous_hindexed_char = (char*)noncontiguous_hindexed + type_c->block_indices[i+1];
852 else noncontiguous_hindexed_char += type_c->block_lengths[i]*smpi_datatype_get_extent(type_c->old_type);
854 noncontiguous_hindexed=(void*)noncontiguous_hindexed_char;
858 * Copies contiguous data into noncontiguous memory.
859 * @param noncontiguous_hindexed - output hindexed
860 * @param contiguous_hindexed - input hindexed
861 * @param type - pointer contening :
862 * - block_lengths - the width or height of blocked matrix
863 * - block_indices - indices of each data, in bytes
864 * - count - the number of rows of matrix
866 void unserialize_hindexed( const void *contiguous_hindexed,
867 void *noncontiguous_hindexed,
871 s_smpi_mpi_hindexed_t* type_c = (s_smpi_mpi_hindexed_t*)type;
874 char* contiguous_hindexed_char = (char*)contiguous_hindexed;
875 char* noncontiguous_hindexed_char = (char*)noncontiguous_hindexed+ type_c->block_indices[0];
876 for(j=0; j<count;j++){
877 for (i = 0; i < type_c->block_count; i++) {
878 if (type_c->old_type->has_subtype == 0)
879 memcpy(noncontiguous_hindexed_char,
880 contiguous_hindexed_char, type_c->block_lengths[i] * type_c->size_oldtype);
882 ((s_smpi_subtype_t*)type_c->old_type->substruct)->unserialize( contiguous_hindexed_char,
883 noncontiguous_hindexed_char,
884 type_c->block_lengths[i],
885 type_c->old_type->substruct);
887 contiguous_hindexed_char += type_c->block_lengths[i]*type_c->size_oldtype;
888 if (i<type_c->block_count-1)noncontiguous_hindexed_char = (char*)noncontiguous_hindexed + type_c->block_indices[i+1];
889 else noncontiguous_hindexed_char += type_c->block_lengths[i]*smpi_datatype_get_extent(type_c->old_type);
891 noncontiguous_hindexed=(void*)noncontiguous_hindexed_char;
895 void free_hindexed(MPI_Datatype* type){
896 xbt_free(((s_smpi_mpi_hindexed_t *)(*type)->substruct)->block_lengths);
897 xbt_free(((s_smpi_mpi_hindexed_t *)(*type)->substruct)->block_indices);
898 smpi_datatype_unuse(((s_smpi_mpi_indexed_t *)(*type)->substruct)->old_type);
902 * Create a Sub type hindexed to be able to serialize and unserialize it
903 * the structure s_smpi_mpi_hindexed_t is derived from s_smpi_subtype which
904 * required the functions unserialize and serialize
906 s_smpi_mpi_hindexed_t* smpi_datatype_hindexed_create( int* block_lengths,
907 MPI_Aint* block_indices,
909 MPI_Datatype old_type,
911 s_smpi_mpi_hindexed_t *new_t= xbt_new(s_smpi_mpi_hindexed_t,1);
912 new_t->base.serialize = &serialize_hindexed;
913 new_t->base.unserialize = &unserialize_hindexed;
914 new_t->base.subtype_free = &free_hindexed;
915 //TODO : add a custom function for each time to clean these
916 new_t->block_lengths= xbt_new(int, block_count);
917 new_t->block_indices= xbt_new(MPI_Aint, block_count);
919 for(i=0;i<block_count;i++){
920 new_t->block_lengths[i]=block_lengths[i];
921 new_t->block_indices[i]=block_indices[i];
923 new_t->block_count = block_count;
924 new_t->old_type = old_type;
925 new_t->size_oldtype = size_oldtype;
930 int smpi_datatype_hindexed(int count, int* blocklens, MPI_Aint* indices, MPI_Datatype old_type, MPI_Datatype* new_type)
939 lb=indices[0] + smpi_datatype_lb(old_type);
940 ub=indices[0] + blocklens[0]*smpi_datatype_ub(old_type);
942 for(i=0; i< count; i++){
945 size += blocklens[i];
947 if(indices[i]+smpi_datatype_lb(old_type)<lb) lb = indices[i]+smpi_datatype_lb(old_type);
948 if(indices[i]+blocklens[i]*smpi_datatype_ub(old_type)>ub) ub = indices[i]+blocklens[i]*smpi_datatype_ub(old_type);
950 if ( (i< count -1) && (indices[i]+blocklens[i]*smpi_datatype_size(old_type) != indices[i+1]) )contiguous=0;
952 if (old_type->has_subtype == 1 || lb!=0)
956 s_smpi_mpi_hindexed_t* subtype = smpi_datatype_hindexed_create( blocklens,
960 smpi_datatype_size(old_type));
961 smpi_datatype_create(new_type, size * smpi_datatype_size(old_type),
964 ,1, subtype, DT_FLAG_DATA);
966 s_smpi_mpi_contiguous_t* subtype = smpi_datatype_contiguous_create( lb,
969 smpi_datatype_size(old_type));
970 smpi_datatype_create(new_type, size * smpi_datatype_size(old_type),
971 0,size * smpi_datatype_size(old_type),
972 1, subtype, DT_FLAG_DATA|DT_FLAG_CONTIGUOUS);
980 struct Implementation - Indexed with indices in bytes
984 * Copies noncontiguous data into contiguous memory.
985 * @param contiguous_struct - output struct
986 * @param noncontiguous_struct - input struct
987 * @param type - pointer contening :
988 * - stride - stride of between noncontiguous data
989 * - block_length - the width or height of blocked matrix
990 * - count - the number of rows of matrix
992 void serialize_struct( const void *noncontiguous_struct,
993 void *contiguous_struct,
997 s_smpi_mpi_struct_t* type_c = (s_smpi_mpi_struct_t*)type;
999 char* contiguous_struct_char = (char*)contiguous_struct;
1000 char* noncontiguous_struct_char = (char*)noncontiguous_struct+ type_c->block_indices[0];
1001 for(j=0; j<count;j++){
1002 for (i = 0; i < type_c->block_count; i++) {
1003 if (type_c->old_types[i]->has_subtype == 0)
1004 memcpy(contiguous_struct_char,
1005 noncontiguous_struct_char, type_c->block_lengths[i] * smpi_datatype_size(type_c->old_types[i]));
1007 ((s_smpi_subtype_t*)type_c->old_types[i]->substruct)->serialize( noncontiguous_struct_char,
1008 contiguous_struct_char,
1009 type_c->block_lengths[i],
1010 type_c->old_types[i]->substruct);
1013 contiguous_struct_char += type_c->block_lengths[i]*smpi_datatype_size(type_c->old_types[i]);
1014 if (i<type_c->block_count-1)noncontiguous_struct_char = (char*)noncontiguous_struct + type_c->block_indices[i+1];
1015 else noncontiguous_struct_char += type_c->block_lengths[i]*smpi_datatype_get_extent(type_c->old_types[i]);//let's hope this is MPI_UB ?
1017 noncontiguous_struct=(void*)noncontiguous_struct_char;
1021 * Copies contiguous data into noncontiguous memory.
1022 * @param noncontiguous_struct - output struct
1023 * @param contiguous_struct - input struct
1024 * @param type - pointer contening :
1025 * - stride - stride of between noncontiguous data
1026 * - block_length - the width or height of blocked matrix
1027 * - count - the number of rows of matrix
1029 void unserialize_struct( const void *contiguous_struct,
1030 void *noncontiguous_struct,
1034 s_smpi_mpi_struct_t* type_c = (s_smpi_mpi_struct_t*)type;
1037 char* contiguous_struct_char = (char*)contiguous_struct;
1038 char* noncontiguous_struct_char = (char*)noncontiguous_struct+ type_c->block_indices[0];
1039 for(j=0; j<count;j++){
1040 for (i = 0; i < type_c->block_count; i++) {
1041 if (type_c->old_types[i]->has_subtype == 0)
1042 memcpy(noncontiguous_struct_char,
1043 contiguous_struct_char, type_c->block_lengths[i] * smpi_datatype_size(type_c->old_types[i]));
1045 ((s_smpi_subtype_t*)type_c->old_types[i]->substruct)->unserialize( contiguous_struct_char,
1046 noncontiguous_struct_char,
1047 type_c->block_lengths[i],
1048 type_c->old_types[i]->substruct);
1050 contiguous_struct_char += type_c->block_lengths[i]*smpi_datatype_size(type_c->old_types[i]);
1051 if (i<type_c->block_count-1)noncontiguous_struct_char = (char*)noncontiguous_struct + type_c->block_indices[i+1];
1052 else noncontiguous_struct_char += type_c->block_lengths[i]*smpi_datatype_get_extent(type_c->old_types[i]);
1054 noncontiguous_struct=(void*)noncontiguous_struct_char;
1059 void free_struct(MPI_Datatype* type){
1060 xbt_free(((s_smpi_mpi_struct_t *)(*type)->substruct)->block_lengths);
1061 xbt_free(((s_smpi_mpi_struct_t *)(*type)->substruct)->block_indices);
1063 for (i = 0; i < ((s_smpi_mpi_struct_t *)(*type)->substruct)->block_count; i++)
1064 smpi_datatype_unuse(((s_smpi_mpi_struct_t *)(*type)->substruct)->old_types[i]);
1065 xbt_free(((s_smpi_mpi_struct_t *)(*type)->substruct)->old_types);
1069 * Create a Sub type struct to be able to serialize and unserialize it
1070 * the structure s_smpi_mpi_struct_t is derived from s_smpi_subtype which
1071 * required the functions unserialize and serialize
1073 s_smpi_mpi_struct_t* smpi_datatype_struct_create( int* block_lengths,
1074 MPI_Aint* block_indices,
1076 MPI_Datatype* old_types){
1077 s_smpi_mpi_struct_t *new_t= xbt_new(s_smpi_mpi_struct_t,1);
1078 new_t->base.serialize = &serialize_struct;
1079 new_t->base.unserialize = &unserialize_struct;
1080 new_t->base.subtype_free = &free_struct;
1081 //TODO : add a custom function for each time to clean these
1082 new_t->block_lengths= xbt_new(int, block_count);
1083 new_t->block_indices= xbt_new(MPI_Aint, block_count);
1084 new_t->old_types= xbt_new(MPI_Datatype, block_count);
1086 for(i=0;i<block_count;i++){
1087 new_t->block_lengths[i]=block_lengths[i];
1088 new_t->block_indices[i]=block_indices[i];
1089 new_t->old_types[i]=old_types[i];
1090 smpi_datatype_use(new_t->old_types[i]);
1092 //new_t->block_lengths = block_lengths;
1093 //new_t->block_indices = block_indices;
1094 new_t->block_count = block_count;
1095 //new_t->old_types = old_types;
1100 int smpi_datatype_struct(int count, int* blocklens, MPI_Aint* indices, MPI_Datatype* old_types, MPI_Datatype* new_type)
1109 lb=indices[0] + smpi_datatype_lb(old_types[0]);
1110 ub=indices[0] + blocklens[0]*smpi_datatype_ub(old_types[0]);
1114 for(i=0; i< count; i++){
1117 if (old_types[i]->has_subtype == 1)
1120 size += blocklens[i]*smpi_datatype_size(old_types[i]);
1121 if (old_types[i]==MPI_LB){
1125 if (old_types[i]==MPI_UB){
1130 if(!forced_lb && indices[i]+smpi_datatype_lb(old_types[i])<lb) lb = indices[i];
1131 if(!forced_ub && indices[i]+blocklens[i]*smpi_datatype_ub(old_types[i])>ub) ub = indices[i]+blocklens[i]*smpi_datatype_ub(old_types[i]);
1133 if ( (i< count -1) && (indices[i]+blocklens[i]*smpi_datatype_size(old_types[i]) != indices[i+1]) )contiguous=0;
1137 s_smpi_mpi_struct_t* subtype = smpi_datatype_struct_create( blocklens,
1142 smpi_datatype_create(new_type, size, lb, ub,1, subtype, DT_FLAG_DATA);
1144 s_smpi_mpi_contiguous_t* subtype = smpi_datatype_contiguous_create( lb,
1148 smpi_datatype_create(new_type, size, lb, ub,1, subtype, DT_FLAG_DATA|DT_FLAG_CONTIGUOUS);
1153 void smpi_datatype_commit(MPI_Datatype *datatype)
1155 (*datatype)->flags= ((*datatype)->flags | DT_FLAG_COMMITED);
1158 typedef struct s_smpi_mpi_op {
1159 MPI_User_function *func;
1163 #define MAX_OP(a, b) (b) = (a) < (b) ? (b) : (a)
1164 #define MIN_OP(a, b) (b) = (a) < (b) ? (a) : (b)
1165 #define SUM_OP(a, b) (b) += (a)
1166 #define PROD_OP(a, b) (b) *= (a)
1167 #define LAND_OP(a, b) (b) = (a) && (b)
1168 #define LOR_OP(a, b) (b) = (a) || (b)
1169 #define LXOR_OP(a, b) (b) = (!(a) && (b)) || ((a) && !(b))
1170 #define BAND_OP(a, b) (b) &= (a)
1171 #define BOR_OP(a, b) (b) |= (a)
1172 #define BXOR_OP(a, b) (b) ^= (a)
1173 #define MAXLOC_OP(a, b) (b) = (a.value) < (b.value) ? (b) : (a)
1174 #define MINLOC_OP(a, b) (b) = (a.value) < (b.value) ? (a) : (b)
1175 #define REPLACE_OP(a,b) (b) = (a)
1177 #define APPLY_FUNC(a, b, length, type, func) \
1180 type* x = (type*)(a); \
1181 type* y = (type*)(b); \
1182 for(i = 0; i < *(length); i++) { \
1187 static void max_func(void *a, void *b, int *length,
1188 MPI_Datatype * datatype)
1190 if (*datatype == MPI_CHAR) {
1191 APPLY_FUNC(a, b, length, char, MAX_OP);
1192 } else if (*datatype == MPI_SHORT) {
1193 APPLY_FUNC(a, b, length, short, MAX_OP);
1194 } else if (*datatype == MPI_INT) {
1195 APPLY_FUNC(a, b, length, int, MAX_OP);
1196 } else if (*datatype == MPI_LONG) {
1197 APPLY_FUNC(a, b, length, long, MAX_OP);
1198 } else if (*datatype == MPI_UNSIGNED_SHORT) {
1199 APPLY_FUNC(a, b, length, unsigned short, MAX_OP);
1200 } else if (*datatype == MPI_UNSIGNED) {
1201 APPLY_FUNC(a, b, length, unsigned int, MAX_OP);
1202 } else if (*datatype == MPI_UNSIGNED_LONG) {
1203 APPLY_FUNC(a, b, length, unsigned long, MAX_OP);
1204 } else if (*datatype == MPI_FLOAT) {
1205 APPLY_FUNC(a, b, length, float, MAX_OP);
1206 } else if (*datatype == MPI_DOUBLE) {
1207 APPLY_FUNC(a, b, length, double, MAX_OP);
1208 } else if (*datatype == MPI_LONG_DOUBLE) {
1209 APPLY_FUNC(a, b, length, long double, MAX_OP);
1213 static void min_func(void *a, void *b, int *length,
1214 MPI_Datatype * datatype)
1216 if (*datatype == MPI_CHAR) {
1217 APPLY_FUNC(a, b, length, char, MIN_OP);
1218 } else if (*datatype == MPI_SHORT) {
1219 APPLY_FUNC(a, b, length, short, MIN_OP);
1220 } else if (*datatype == MPI_INT) {
1221 APPLY_FUNC(a, b, length, int, MIN_OP);
1222 } else if (*datatype == MPI_LONG) {
1223 APPLY_FUNC(a, b, length, long, MIN_OP);
1224 } else if (*datatype == MPI_UNSIGNED_SHORT) {
1225 APPLY_FUNC(a, b, length, unsigned short, MIN_OP);
1226 } else if (*datatype == MPI_UNSIGNED) {
1227 APPLY_FUNC(a, b, length, unsigned int, MIN_OP);
1228 } else if (*datatype == MPI_UNSIGNED_LONG) {
1229 APPLY_FUNC(a, b, length, unsigned long, MIN_OP);
1230 } else if (*datatype == MPI_FLOAT) {
1231 APPLY_FUNC(a, b, length, float, MIN_OP);
1232 } else if (*datatype == MPI_DOUBLE) {
1233 APPLY_FUNC(a, b, length, double, MIN_OP);
1234 } else if (*datatype == MPI_LONG_DOUBLE) {
1235 APPLY_FUNC(a, b, length, long double, MIN_OP);
1239 static void sum_func(void *a, void *b, int *length,
1240 MPI_Datatype * datatype)
1242 if (*datatype == MPI_CHAR) {
1243 APPLY_FUNC(a, b, length, char, SUM_OP);
1244 } else if (*datatype == MPI_SHORT) {
1245 APPLY_FUNC(a, b, length, short, SUM_OP);
1246 } else if (*datatype == MPI_INT) {
1247 APPLY_FUNC(a, b, length, int, SUM_OP);
1248 } else if (*datatype == MPI_LONG) {
1249 APPLY_FUNC(a, b, length, long, SUM_OP);
1250 } else if (*datatype == MPI_UNSIGNED_SHORT) {
1251 APPLY_FUNC(a, b, length, unsigned short, SUM_OP);
1252 } else if (*datatype == MPI_UNSIGNED) {
1253 APPLY_FUNC(a, b, length, unsigned int, SUM_OP);
1254 } else if (*datatype == MPI_UNSIGNED_LONG) {
1255 APPLY_FUNC(a, b, length, unsigned long, SUM_OP);
1256 } else if (*datatype == MPI_FLOAT) {
1257 APPLY_FUNC(a, b, length, float, SUM_OP);
1258 } else if (*datatype == MPI_DOUBLE) {
1259 APPLY_FUNC(a, b, length, double, SUM_OP);
1260 } else if (*datatype == MPI_LONG_DOUBLE) {
1261 APPLY_FUNC(a, b, length, long double, SUM_OP);
1262 } else if (*datatype == MPI_C_FLOAT_COMPLEX) {
1263 APPLY_FUNC(a, b, length, float _Complex, SUM_OP);
1264 } else if (*datatype == MPI_C_DOUBLE_COMPLEX) {
1265 APPLY_FUNC(a, b, length, double _Complex, SUM_OP);
1266 } else if (*datatype == MPI_C_LONG_DOUBLE_COMPLEX) {
1267 APPLY_FUNC(a, b, length, long double _Complex, SUM_OP);
1271 static void prod_func(void *a, void *b, int *length,
1272 MPI_Datatype * datatype)
1274 if (*datatype == MPI_CHAR) {
1275 APPLY_FUNC(a, b, length, char, PROD_OP);
1276 } else if (*datatype == MPI_SHORT) {
1277 APPLY_FUNC(a, b, length, short, PROD_OP);
1278 } else if (*datatype == MPI_INT) {
1279 APPLY_FUNC(a, b, length, int, PROD_OP);
1280 } else if (*datatype == MPI_LONG) {
1281 APPLY_FUNC(a, b, length, long, PROD_OP);
1282 } else if (*datatype == MPI_UNSIGNED_SHORT) {
1283 APPLY_FUNC(a, b, length, unsigned short, PROD_OP);
1284 } else if (*datatype == MPI_UNSIGNED) {
1285 APPLY_FUNC(a, b, length, unsigned int, PROD_OP);
1286 } else if (*datatype == MPI_UNSIGNED_LONG) {
1287 APPLY_FUNC(a, b, length, unsigned long, PROD_OP);
1288 } else if (*datatype == MPI_FLOAT) {
1289 APPLY_FUNC(a, b, length, float, PROD_OP);
1290 } else if (*datatype == MPI_DOUBLE) {
1291 APPLY_FUNC(a, b, length, double, PROD_OP);
1292 } else if (*datatype == MPI_LONG_DOUBLE) {
1293 APPLY_FUNC(a, b, length, long double, PROD_OP);
1294 } else if (*datatype == MPI_C_FLOAT_COMPLEX) {
1295 APPLY_FUNC(a, b, length, float _Complex, PROD_OP);
1296 } else if (*datatype == MPI_C_DOUBLE_COMPLEX) {
1297 APPLY_FUNC(a, b, length, double _Complex, PROD_OP);
1298 } else if (*datatype == MPI_C_LONG_DOUBLE_COMPLEX) {
1299 APPLY_FUNC(a, b, length, long double _Complex, PROD_OP);
1303 static void land_func(void *a, void *b, int *length,
1304 MPI_Datatype * datatype)
1306 if (*datatype == MPI_CHAR) {
1307 APPLY_FUNC(a, b, length, char, LAND_OP);
1308 } else if (*datatype == MPI_SHORT) {
1309 APPLY_FUNC(a, b, length, short, LAND_OP);
1310 } else if (*datatype == MPI_INT) {
1311 APPLY_FUNC(a, b, length, int, LAND_OP);
1312 } else if (*datatype == MPI_LONG) {
1313 APPLY_FUNC(a, b, length, long, LAND_OP);
1314 } else if (*datatype == MPI_UNSIGNED_SHORT) {
1315 APPLY_FUNC(a, b, length, unsigned short, LAND_OP);
1316 } else if (*datatype == MPI_UNSIGNED) {
1317 APPLY_FUNC(a, b, length, unsigned int, LAND_OP);
1318 } else if (*datatype == MPI_UNSIGNED_LONG) {
1319 APPLY_FUNC(a, b, length, unsigned long, LAND_OP);
1320 } else if (*datatype == MPI_C_BOOL) {
1321 APPLY_FUNC(a, b, length, _Bool, LAND_OP);
1325 static void lor_func(void *a, void *b, int *length,
1326 MPI_Datatype * datatype)
1328 if (*datatype == MPI_CHAR) {
1329 APPLY_FUNC(a, b, length, char, LOR_OP);
1330 } else if (*datatype == MPI_SHORT) {
1331 APPLY_FUNC(a, b, length, short, LOR_OP);
1332 } else if (*datatype == MPI_INT) {
1333 APPLY_FUNC(a, b, length, int, LOR_OP);
1334 } else if (*datatype == MPI_LONG) {
1335 APPLY_FUNC(a, b, length, long, LOR_OP);
1336 } else if (*datatype == MPI_UNSIGNED_SHORT) {
1337 APPLY_FUNC(a, b, length, unsigned short, LOR_OP);
1338 } else if (*datatype == MPI_UNSIGNED) {
1339 APPLY_FUNC(a, b, length, unsigned int, LOR_OP);
1340 } else if (*datatype == MPI_UNSIGNED_LONG) {
1341 APPLY_FUNC(a, b, length, unsigned long, LOR_OP);
1342 } else if (*datatype == MPI_C_BOOL) {
1343 APPLY_FUNC(a, b, length, _Bool, LOR_OP);
1347 static void lxor_func(void *a, void *b, int *length,
1348 MPI_Datatype * datatype)
1350 if (*datatype == MPI_CHAR) {
1351 APPLY_FUNC(a, b, length, char, LXOR_OP);
1352 } else if (*datatype == MPI_SHORT) {
1353 APPLY_FUNC(a, b, length, short, LXOR_OP);
1354 } else if (*datatype == MPI_INT) {
1355 APPLY_FUNC(a, b, length, int, LXOR_OP);
1356 } else if (*datatype == MPI_LONG) {
1357 APPLY_FUNC(a, b, length, long, LXOR_OP);
1358 } else if (*datatype == MPI_UNSIGNED_SHORT) {
1359 APPLY_FUNC(a, b, length, unsigned short, LXOR_OP);
1360 } else if (*datatype == MPI_UNSIGNED) {
1361 APPLY_FUNC(a, b, length, unsigned int, LXOR_OP);
1362 } else if (*datatype == MPI_UNSIGNED_LONG) {
1363 APPLY_FUNC(a, b, length, unsigned long, LXOR_OP);
1364 } else if (*datatype == MPI_C_BOOL) {
1365 APPLY_FUNC(a, b, length, _Bool, LXOR_OP);
1369 static void band_func(void *a, void *b, int *length,
1370 MPI_Datatype * datatype)
1372 if (*datatype == MPI_CHAR) {
1373 APPLY_FUNC(a, b, length, char, BAND_OP);
1375 if (*datatype == MPI_SHORT) {
1376 APPLY_FUNC(a, b, length, short, BAND_OP);
1377 } else if (*datatype == MPI_INT) {
1378 APPLY_FUNC(a, b, length, int, BAND_OP);
1379 } else if (*datatype == MPI_LONG) {
1380 APPLY_FUNC(a, b, length, long, BAND_OP);
1381 } else if (*datatype == MPI_UNSIGNED_SHORT) {
1382 APPLY_FUNC(a, b, length, unsigned short, BAND_OP);
1383 } else if (*datatype == MPI_UNSIGNED) {
1384 APPLY_FUNC(a, b, length, unsigned int, BAND_OP);
1385 } else if (*datatype == MPI_UNSIGNED_LONG) {
1386 APPLY_FUNC(a, b, length, unsigned long, BAND_OP);
1387 } else if (*datatype == MPI_BYTE) {
1388 APPLY_FUNC(a, b, length, uint8_t, BAND_OP);
1392 static void bor_func(void *a, void *b, int *length,
1393 MPI_Datatype * datatype)
1395 if (*datatype == MPI_CHAR) {
1396 APPLY_FUNC(a, b, length, char, BOR_OP);
1397 } else if (*datatype == MPI_SHORT) {
1398 APPLY_FUNC(a, b, length, short, BOR_OP);
1399 } else if (*datatype == MPI_INT) {
1400 APPLY_FUNC(a, b, length, int, BOR_OP);
1401 } else if (*datatype == MPI_LONG) {
1402 APPLY_FUNC(a, b, length, long, BOR_OP);
1403 } else if (*datatype == MPI_UNSIGNED_SHORT) {
1404 APPLY_FUNC(a, b, length, unsigned short, BOR_OP);
1405 } else if (*datatype == MPI_UNSIGNED) {
1406 APPLY_FUNC(a, b, length, unsigned int, BOR_OP);
1407 } else if (*datatype == MPI_UNSIGNED_LONG) {
1408 APPLY_FUNC(a, b, length, unsigned long, BOR_OP);
1409 } else if (*datatype == MPI_BYTE) {
1410 APPLY_FUNC(a, b, length, uint8_t, BOR_OP);
1414 static void bxor_func(void *a, void *b, int *length,
1415 MPI_Datatype * datatype)
1417 if (*datatype == MPI_CHAR) {
1418 APPLY_FUNC(a, b, length, char, BXOR_OP);
1419 } else if (*datatype == MPI_SHORT) {
1420 APPLY_FUNC(a, b, length, short, BXOR_OP);
1421 } else if (*datatype == MPI_INT) {
1422 APPLY_FUNC(a, b, length, int, BXOR_OP);
1423 } else if (*datatype == MPI_LONG) {
1424 APPLY_FUNC(a, b, length, long, BXOR_OP);
1425 } else if (*datatype == MPI_UNSIGNED_SHORT) {
1426 APPLY_FUNC(a, b, length, unsigned short, BXOR_OP);
1427 } else if (*datatype == MPI_UNSIGNED) {
1428 APPLY_FUNC(a, b, length, unsigned int, BXOR_OP);
1429 } else if (*datatype == MPI_UNSIGNED_LONG) {
1430 APPLY_FUNC(a, b, length, unsigned long, BXOR_OP);
1431 } else if (*datatype == MPI_BYTE) {
1432 APPLY_FUNC(a, b, length, uint8_t, BXOR_OP);
1436 static void minloc_func(void *a, void *b, int *length,
1437 MPI_Datatype * datatype)
1439 if (*datatype == MPI_FLOAT_INT) {
1440 APPLY_FUNC(a, b, length, float_int, MINLOC_OP);
1441 } else if (*datatype == MPI_LONG_INT) {
1442 APPLY_FUNC(a, b, length, long_int, MINLOC_OP);
1443 } else if (*datatype == MPI_DOUBLE_INT) {
1444 APPLY_FUNC(a, b, length, double_int, MINLOC_OP);
1445 } else if (*datatype == MPI_SHORT_INT) {
1446 APPLY_FUNC(a, b, length, short_int, MINLOC_OP);
1447 } else if (*datatype == MPI_2INT) {
1448 APPLY_FUNC(a, b, length, int_int, MINLOC_OP);
1449 } else if (*datatype == MPI_LONG_DOUBLE_INT) {
1450 APPLY_FUNC(a, b, length, long_double_int, MINLOC_OP);
1451 } else if (*datatype == MPI_2FLOAT) {
1452 APPLY_FUNC(a, b, length, float_float, MINLOC_OP);
1453 } else if (*datatype == MPI_2DOUBLE) {
1454 APPLY_FUNC(a, b, length, double_double, MINLOC_OP);
1458 static void maxloc_func(void *a, void *b, int *length,
1459 MPI_Datatype * datatype)
1461 if (*datatype == MPI_FLOAT_INT) {
1462 APPLY_FUNC(a, b, length, float_int, MAXLOC_OP);
1463 } else if (*datatype == MPI_LONG_INT) {
1464 APPLY_FUNC(a, b, length, long_int, MAXLOC_OP);
1465 } else if (*datatype == MPI_DOUBLE_INT) {
1466 APPLY_FUNC(a, b, length, double_int, MAXLOC_OP);
1467 } else if (*datatype == MPI_SHORT_INT) {
1468 APPLY_FUNC(a, b, length, short_int, MAXLOC_OP);
1469 } else if (*datatype == MPI_2INT) {
1470 APPLY_FUNC(a, b, length, int_int, MAXLOC_OP);
1471 } else if (*datatype == MPI_LONG_DOUBLE_INT) {
1472 APPLY_FUNC(a, b, length, long_double_int, MAXLOC_OP);
1473 } else if (*datatype == MPI_2FLOAT) {
1474 APPLY_FUNC(a, b, length, float_float, MAXLOC_OP);
1475 } else if (*datatype == MPI_2DOUBLE) {
1476 APPLY_FUNC(a, b, length, double_double, MAXLOC_OP);
1480 static void replace_func(void *a, void *b, int *length,
1481 MPI_Datatype * datatype)
1483 if (*datatype == MPI_CHAR) {
1484 APPLY_FUNC(a, b, length, char, REPLACE_OP);
1485 } else if (*datatype == MPI_SHORT) {
1486 APPLY_FUNC(a, b, length, short, REPLACE_OP);
1487 } else if (*datatype == MPI_INT) {
1488 APPLY_FUNC(a, b, length, int, REPLACE_OP);
1489 } else if (*datatype == MPI_LONG) {
1490 APPLY_FUNC(a, b, length, long, REPLACE_OP);
1491 } else if (*datatype == MPI_UNSIGNED_SHORT) {
1492 APPLY_FUNC(a, b, length, unsigned short, REPLACE_OP);
1493 } else if (*datatype == MPI_UNSIGNED) {
1494 APPLY_FUNC(a, b, length, unsigned int, REPLACE_OP);
1495 } else if (*datatype == MPI_UNSIGNED_LONG) {
1496 APPLY_FUNC(a, b, length, unsigned long, REPLACE_OP);
1497 } else if (*datatype == MPI_BYTE) {
1498 APPLY_FUNC(a, b, length, uint8_t, REPLACE_OP);
1502 #define CREATE_MPI_OP(name, func) \
1503 static s_smpi_mpi_op_t mpi_##name = { &(func) /* func */, TRUE }; \
1504 MPI_Op name = &mpi_##name;
1506 CREATE_MPI_OP(MPI_MAX, max_func);
1507 CREATE_MPI_OP(MPI_MIN, min_func);
1508 CREATE_MPI_OP(MPI_SUM, sum_func);
1509 CREATE_MPI_OP(MPI_PROD, prod_func);
1510 CREATE_MPI_OP(MPI_LAND, land_func);
1511 CREATE_MPI_OP(MPI_LOR, lor_func);
1512 CREATE_MPI_OP(MPI_LXOR, lxor_func);
1513 CREATE_MPI_OP(MPI_BAND, band_func);
1514 CREATE_MPI_OP(MPI_BOR, bor_func);
1515 CREATE_MPI_OP(MPI_BXOR, bxor_func);
1516 CREATE_MPI_OP(MPI_MAXLOC, maxloc_func);
1517 CREATE_MPI_OP(MPI_MINLOC, minloc_func);
1518 CREATE_MPI_OP(MPI_REPLACE, replace_func);
1521 MPI_Op smpi_op_new(MPI_User_function * function, int commute)
1524 op = xbt_new(s_smpi_mpi_op_t, 1);
1525 op->func = function;
1526 op-> is_commute = commute;
1530 int smpi_op_is_commute(MPI_Op op)
1532 return (op==MPI_OP_NULL) ? 1 : op-> is_commute;
1535 void smpi_op_destroy(MPI_Op op)
1540 void smpi_op_apply(MPI_Op op, void *invec, void *inoutvec, int *len,
1541 MPI_Datatype * datatype)
1543 if(smpi_privatize_global_variables){ //we need to switch here, as the called function may silently touch global variables
1544 XBT_VERB("Applying operation, switch to the right data frame ");
1545 switch_data_segment(smpi_process_index());
1548 if(!_xbt_replay_is_active())
1549 op->func(invec, inoutvec, len, datatype);