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*);
127 size_t smpi_datatype_size(MPI_Datatype datatype)
129 return datatype->size;
134 MPI_Aint smpi_datatype_lb(MPI_Datatype datatype)
139 MPI_Aint smpi_datatype_ub(MPI_Datatype datatype)
144 int smpi_datatype_extent(MPI_Datatype datatype, MPI_Aint * lb,
148 *extent = datatype->ub - datatype->lb;
152 MPI_Aint smpi_datatype_get_extent(MPI_Datatype datatype){
153 return datatype->ub - datatype->lb;
156 int smpi_datatype_copy(void *sendbuf, int sendcount, MPI_Datatype sendtype,
157 void *recvbuf, int recvcount, MPI_Datatype recvtype)
160 if(smpi_privatize_global_variables){
161 switch_data_segment(smpi_process_index());
163 /* First check if we really have something to do */
164 if (recvcount > 0 && recvbuf != sendbuf) {
165 /* FIXME: treat packed cases */
166 sendcount *= smpi_datatype_size(sendtype);
167 recvcount *= smpi_datatype_size(recvtype);
168 count = sendcount < recvcount ? sendcount : recvcount;
170 if(sendtype->has_subtype == 0 && recvtype->has_subtype == 0) {
171 if(!_xbt_replay_is_active()) memcpy(recvbuf, sendbuf, count);
173 else if (sendtype->has_subtype == 0)
175 s_smpi_subtype_t *subtype = recvtype->substruct;
176 subtype->unserialize( sendbuf, recvbuf,1, subtype);
178 else if (recvtype->has_subtype == 0)
180 s_smpi_subtype_t *subtype = sendtype->substruct;
181 subtype->serialize(sendbuf, recvbuf,1, subtype);
183 s_smpi_subtype_t *subtype = sendtype->substruct;
186 void * buf_tmp = xbt_malloc(count);
188 subtype->serialize( sendbuf, buf_tmp,count/smpi_datatype_size(sendtype), subtype);
189 subtype = recvtype->substruct;
190 subtype->unserialize( buf_tmp, recvbuf,count/smpi_datatype_size(recvtype), subtype);
196 return sendcount > recvcount ? MPI_ERR_TRUNCATE : MPI_SUCCESS;
200 * Copies noncontiguous data into contiguous memory.
201 * @param contiguous_vector - output vector
202 * @param noncontiguous_vector - input vector
203 * @param type - pointer contening :
204 * - stride - stride of between noncontiguous data
205 * - block_length - the width or height of blocked matrix
206 * - count - the number of rows of matrix
208 void serialize_vector( const void *noncontiguous_vector,
209 void *contiguous_vector,
213 s_smpi_mpi_vector_t* type_c = (s_smpi_mpi_vector_t*)type;
215 char* contiguous_vector_char = (char*)contiguous_vector;
216 char* noncontiguous_vector_char = (char*)noncontiguous_vector;
218 for (i = 0; i < type_c->block_count * count; i++) {
219 if (type_c->old_type->has_subtype == 0)
220 memcpy(contiguous_vector_char,
221 noncontiguous_vector_char, type_c->block_length * type_c->size_oldtype);
223 ((s_smpi_subtype_t*)type_c->old_type->substruct)->serialize( noncontiguous_vector_char,
224 contiguous_vector_char,
225 type_c->block_length,
226 type_c->old_type->substruct);
228 contiguous_vector_char += type_c->block_length*type_c->size_oldtype;
229 if((i+1)%type_c->block_count ==0)
230 noncontiguous_vector_char += type_c->block_length*smpi_datatype_get_extent(type_c->old_type);
232 noncontiguous_vector_char += type_c->block_stride*smpi_datatype_get_extent(type_c->old_type);
237 * Copies contiguous data into noncontiguous memory.
238 * @param noncontiguous_vector - output vector
239 * @param contiguous_vector - input vector
240 * @param type - pointer contening :
241 * - stride - stride of between noncontiguous data
242 * - block_length - the width or height of blocked matrix
243 * - count - the number of rows of matrix
245 void unserialize_vector( const void *contiguous_vector,
246 void *noncontiguous_vector,
250 s_smpi_mpi_vector_t* type_c = (s_smpi_mpi_vector_t*)type;
253 char* contiguous_vector_char = (char*)contiguous_vector;
254 char* noncontiguous_vector_char = (char*)noncontiguous_vector;
256 for (i = 0; i < type_c->block_count * count; i++) {
257 if (type_c->old_type->has_subtype == 0)
258 memcpy(noncontiguous_vector_char,
259 contiguous_vector_char, type_c->block_length * type_c->size_oldtype);
261 ((s_smpi_subtype_t*)type_c->old_type->substruct)->unserialize( contiguous_vector_char,
262 noncontiguous_vector_char,
263 type_c->block_length,
264 type_c->old_type->substruct);
265 contiguous_vector_char += type_c->block_length*type_c->size_oldtype;
266 if((i+1)%type_c->block_count ==0)
267 noncontiguous_vector_char += type_c->block_length*smpi_datatype_get_extent(type_c->old_type);
269 noncontiguous_vector_char += type_c->block_stride*smpi_datatype_get_extent(type_c->old_type);
274 * Create a Sub type vector to be able to serialize and unserialize it
275 * the structure s_smpi_mpi_vector_t is derived from s_smpi_subtype which
276 * required the functions unserialize and serialize
279 s_smpi_mpi_vector_t* smpi_datatype_vector_create( int block_stride,
282 MPI_Datatype old_type,
284 s_smpi_mpi_vector_t *new_t= xbt_new(s_smpi_mpi_vector_t,1);
285 new_t->base.serialize = &serialize_vector;
286 new_t->base.unserialize = &unserialize_vector;
287 new_t->base.subtype_free = &free_vector;
288 new_t->block_stride = block_stride;
289 new_t->block_length = block_length;
290 new_t->block_count = block_count;
291 new_t->old_type = old_type;
292 new_t->size_oldtype = size_oldtype;
296 void smpi_datatype_create(MPI_Datatype* new_type, int size,int lb, int ub, int has_subtype,
297 void *struct_type, int flags){
298 MPI_Datatype new_t= xbt_new(s_smpi_mpi_datatype_t,1);
300 new_t->has_subtype = size>0? has_subtype:0;
303 new_t->flags = flags;
304 new_t->substruct = struct_type;
310 MC_ignore(&(new_t->in_use), sizeof(new_t->in_use));
314 void smpi_datatype_free(MPI_Datatype* type){
316 if((*type)->flags & DT_FLAG_PREDEFINED)return;
318 //if still used, mark for deletion
319 if((*type)->in_use!=0){
320 (*type)->flags |=DT_FLAG_DESTROYED;
324 if ((*type)->has_subtype == 1){
325 ((s_smpi_subtype_t *)(*type)->substruct)->subtype_free(type);
326 xbt_free((*type)->substruct);
332 void smpi_datatype_use(MPI_Datatype type){
333 if(type)type->in_use++;
337 MC_ignore(&(type->in_use), sizeof(type->in_use));
342 void smpi_datatype_unuse(MPI_Datatype type){
343 if(type && type->in_use-- == 0 && (type->flags & DT_FLAG_DESTROYED))
344 smpi_datatype_free(&type);
348 MC_ignore(&(type->in_use), sizeof(type->in_use));
356 Contiguous Implementation
361 * Copies noncontiguous data into contiguous memory.
362 * @param contiguous_hvector - output hvector
363 * @param noncontiguous_hvector - input hvector
364 * @param type - pointer contening :
365 * - stride - stride of between noncontiguous data, in bytes
366 * - block_length - the width or height of blocked matrix
367 * - count - the number of rows of matrix
369 void serialize_contiguous( const void *noncontiguous_hvector,
370 void *contiguous_hvector,
374 s_smpi_mpi_contiguous_t* type_c = (s_smpi_mpi_contiguous_t*)type;
375 char* contiguous_vector_char = (char*)contiguous_hvector;
376 char* noncontiguous_vector_char = (char*)noncontiguous_hvector+type_c->lb;
377 memcpy(contiguous_vector_char,
378 noncontiguous_vector_char, count* type_c->block_count * type_c->size_oldtype);
381 * Copies contiguous data into noncontiguous memory.
382 * @param noncontiguous_vector - output hvector
383 * @param contiguous_vector - input hvector
384 * @param type - pointer contening :
385 * - stride - stride of between noncontiguous data, in bytes
386 * - block_length - the width or height of blocked matrix
387 * - count - the number of rows of matrix
389 void unserialize_contiguous( const void *contiguous_vector,
390 void *noncontiguous_vector,
394 s_smpi_mpi_contiguous_t* type_c = (s_smpi_mpi_contiguous_t*)type;
395 char* contiguous_vector_char = (char*)contiguous_vector;
396 char* noncontiguous_vector_char = (char*)noncontiguous_vector+type_c->lb;
398 memcpy(noncontiguous_vector_char,
399 contiguous_vector_char, count* type_c->block_count * type_c->size_oldtype);
402 void free_contiguous(MPI_Datatype* d){
406 * Create a Sub type contiguous to be able to serialize and unserialize it
407 * the structure s_smpi_mpi_contiguous_t is derived from s_smpi_subtype which
408 * required the functions unserialize and serialize
411 s_smpi_mpi_contiguous_t* smpi_datatype_contiguous_create( MPI_Aint lb,
413 MPI_Datatype old_type,
415 s_smpi_mpi_contiguous_t *new_t= xbt_new(s_smpi_mpi_contiguous_t,1);
416 new_t->base.serialize = &serialize_contiguous;
417 new_t->base.unserialize = &unserialize_contiguous;
418 new_t->base.subtype_free = &free_contiguous;
420 new_t->block_count = block_count;
421 new_t->old_type = old_type;
422 new_t->size_oldtype = size_oldtype;
429 int smpi_datatype_contiguous(int count, MPI_Datatype old_type, MPI_Datatype* new_type, MPI_Aint lb)
432 if(old_type->has_subtype){
433 //handle this case as a hvector with stride equals to the extent of the datatype
434 return smpi_datatype_hvector(count, 1, smpi_datatype_get_extent(old_type), old_type, new_type);
437 s_smpi_mpi_contiguous_t* subtype = smpi_datatype_contiguous_create( lb,
440 smpi_datatype_size(old_type));
442 smpi_datatype_create(new_type,
443 count * smpi_datatype_size(old_type),
444 lb,lb + count * smpi_datatype_size(old_type),
445 1,subtype, DT_FLAG_CONTIGUOUS);
450 int smpi_datatype_vector(int count, int blocklen, int stride, MPI_Datatype old_type, MPI_Datatype* new_type)
453 if (blocklen<0) return MPI_ERR_ARG;
457 lb=smpi_datatype_lb(old_type);
458 ub=((count-1)*stride+blocklen-1)*smpi_datatype_get_extent(old_type)+smpi_datatype_ub(old_type);
460 if(old_type->has_subtype || stride != blocklen){
463 s_smpi_mpi_vector_t* subtype = smpi_datatype_vector_create( stride,
467 smpi_datatype_size(old_type));
468 smpi_datatype_create(new_type,
469 count * (blocklen) * smpi_datatype_size(old_type), lb,
476 /* in this situation the data are contignous thus it's not
477 * required to serialize and unserialize it*/
478 smpi_datatype_create(new_type, count * blocklen *
479 smpi_datatype_size(old_type), 0, ((count -1) * stride + blocklen)*
480 smpi_datatype_size(old_type),
483 DT_FLAG_VECTOR|DT_FLAG_CONTIGUOUS);
489 void free_vector(MPI_Datatype* d){
493 Hvector Implementation - Vector with stride in bytes
498 * Copies noncontiguous data into contiguous memory.
499 * @param contiguous_hvector - output hvector
500 * @param noncontiguous_hvector - input hvector
501 * @param type - pointer contening :
502 * - stride - stride of between noncontiguous data, in bytes
503 * - block_length - the width or height of blocked matrix
504 * - count - the number of rows of matrix
506 void serialize_hvector( const void *noncontiguous_hvector,
507 void *contiguous_hvector,
511 s_smpi_mpi_hvector_t* type_c = (s_smpi_mpi_hvector_t*)type;
513 char* contiguous_vector_char = (char*)contiguous_hvector;
514 char* noncontiguous_vector_char = (char*)noncontiguous_hvector;
516 for (i = 0; i < type_c->block_count * count; i++) {
517 if (type_c->old_type->has_subtype == 0)
518 memcpy(contiguous_vector_char,
519 noncontiguous_vector_char, type_c->block_length * type_c->size_oldtype);
521 ((s_smpi_subtype_t*)type_c->old_type->substruct)->serialize( noncontiguous_vector_char,
522 contiguous_vector_char,
523 type_c->block_length,
524 type_c->old_type->substruct);
526 contiguous_vector_char += type_c->block_length*type_c->size_oldtype;
527 if((i+1)%type_c->block_count ==0)
528 noncontiguous_vector_char += type_c->block_length*type_c->size_oldtype;
530 noncontiguous_vector_char += type_c->block_stride;
534 * Copies contiguous data into noncontiguous memory.
535 * @param noncontiguous_vector - output hvector
536 * @param contiguous_vector - input hvector
537 * @param type - pointer contening :
538 * - stride - stride of between noncontiguous data, in bytes
539 * - block_length - the width or height of blocked matrix
540 * - count - the number of rows of matrix
542 void unserialize_hvector( const void *contiguous_vector,
543 void *noncontiguous_vector,
547 s_smpi_mpi_hvector_t* type_c = (s_smpi_mpi_hvector_t*)type;
550 char* contiguous_vector_char = (char*)contiguous_vector;
551 char* noncontiguous_vector_char = (char*)noncontiguous_vector;
553 for (i = 0; i < type_c->block_count * count; i++) {
554 if (type_c->old_type->has_subtype == 0)
555 memcpy(noncontiguous_vector_char,
556 contiguous_vector_char, type_c->block_length * type_c->size_oldtype);
558 ((s_smpi_subtype_t*)type_c->old_type->substruct)->unserialize( contiguous_vector_char,
559 noncontiguous_vector_char,
560 type_c->block_length,
561 type_c->old_type->substruct);
562 contiguous_vector_char += type_c->block_length*type_c->size_oldtype;
563 if((i+1)%type_c->block_count ==0)
564 noncontiguous_vector_char += type_c->block_length*type_c->size_oldtype;
566 noncontiguous_vector_char += type_c->block_stride;
571 * Create a Sub type vector to be able to serialize and unserialize it
572 * the structure s_smpi_mpi_vector_t is derived from s_smpi_subtype which
573 * required the functions unserialize and serialize
576 s_smpi_mpi_hvector_t* smpi_datatype_hvector_create( MPI_Aint block_stride,
579 MPI_Datatype old_type,
581 s_smpi_mpi_hvector_t *new_t= xbt_new(s_smpi_mpi_hvector_t,1);
582 new_t->base.serialize = &serialize_hvector;
583 new_t->base.unserialize = &unserialize_hvector;
584 new_t->base.subtype_free = &free_hvector;
585 new_t->block_stride = block_stride;
586 new_t->block_length = block_length;
587 new_t->block_count = block_count;
588 new_t->old_type = old_type;
589 new_t->size_oldtype = size_oldtype;
593 //do nothing for vector types
594 void free_hvector(MPI_Datatype* d){
597 int smpi_datatype_hvector(int count, int blocklen, MPI_Aint stride, MPI_Datatype old_type, MPI_Datatype* new_type)
600 if (blocklen<0) return MPI_ERR_ARG;
604 lb=smpi_datatype_lb(old_type);
605 ub=((count-1)*stride)+(blocklen-1)*smpi_datatype_get_extent(old_type)+smpi_datatype_ub(old_type);
607 if(old_type->has_subtype || stride != blocklen*smpi_datatype_get_extent(old_type)){
608 s_smpi_mpi_hvector_t* subtype = smpi_datatype_hvector_create( stride,
612 smpi_datatype_size(old_type));
614 smpi_datatype_create(new_type, count * blocklen * smpi_datatype_size(old_type),
621 smpi_datatype_create(new_type, count * blocklen *
622 smpi_datatype_size(old_type),0,count * blocklen *
623 smpi_datatype_size(old_type),
626 DT_FLAG_VECTOR|DT_FLAG_CONTIGUOUS);
634 Indexed Implementation
638 * Copies noncontiguous data into contiguous memory.
639 * @param contiguous_indexed - output indexed
640 * @param noncontiguous_indexed - input indexed
641 * @param type - pointer contening :
642 * - block_lengths - the width or height of blocked matrix
643 * - block_indices - indices of each data, in element
644 * - count - the number of rows of matrix
646 void serialize_indexed( const void *noncontiguous_indexed,
647 void *contiguous_indexed,
651 s_smpi_mpi_indexed_t* type_c = (s_smpi_mpi_indexed_t*)type;
653 char* contiguous_indexed_char = (char*)contiguous_indexed;
654 char* noncontiguous_indexed_char = (char*)noncontiguous_indexed+type_c->block_indices[0] * type_c->size_oldtype;
655 for(j=0; j<count;j++){
656 for (i = 0; i < type_c->block_count; i++) {
657 if (type_c->old_type->has_subtype == 0)
658 memcpy(contiguous_indexed_char,
659 noncontiguous_indexed_char, type_c->block_lengths[i] * type_c->size_oldtype);
661 ((s_smpi_subtype_t*)type_c->old_type->substruct)->serialize( noncontiguous_indexed_char,
662 contiguous_indexed_char,
663 type_c->block_lengths[i],
664 type_c->old_type->substruct);
667 contiguous_indexed_char += type_c->block_lengths[i]*type_c->size_oldtype;
668 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);
669 else noncontiguous_indexed_char += type_c->block_lengths[i]*smpi_datatype_get_extent(type_c->old_type);
671 noncontiguous_indexed=(void*)noncontiguous_indexed_char;
675 * Copies contiguous data into noncontiguous memory.
676 * @param noncontiguous_indexed - output indexed
677 * @param contiguous_indexed - input indexed
678 * @param type - pointer contening :
679 * - block_lengths - the width or height of blocked matrix
680 * - block_indices - indices of each data, in element
681 * - count - the number of rows of matrix
683 void unserialize_indexed( const void *contiguous_indexed,
684 void *noncontiguous_indexed,
689 s_smpi_mpi_indexed_t* type_c = (s_smpi_mpi_indexed_t*)type;
691 char* contiguous_indexed_char = (char*)contiguous_indexed;
692 char* noncontiguous_indexed_char = (char*)noncontiguous_indexed+type_c->block_indices[0]*smpi_datatype_get_extent(type_c->old_type);
693 for(j=0; j<count;j++){
694 for (i = 0; i < type_c->block_count; i++) {
695 if (type_c->old_type->has_subtype == 0)
696 memcpy(noncontiguous_indexed_char ,
697 contiguous_indexed_char, type_c->block_lengths[i] * type_c->size_oldtype);
699 ((s_smpi_subtype_t*)type_c->old_type->substruct)->unserialize( contiguous_indexed_char,
700 noncontiguous_indexed_char,
701 type_c->block_lengths[i],
702 type_c->old_type->substruct);
704 contiguous_indexed_char += type_c->block_lengths[i]*type_c->size_oldtype;
705 if (i<type_c->block_count-1)
706 noncontiguous_indexed_char = (char*)noncontiguous_indexed + type_c->block_indices[i+1]*smpi_datatype_get_extent(type_c->old_type);
707 else noncontiguous_indexed_char += type_c->block_lengths[i]*smpi_datatype_get_extent(type_c->old_type);
709 noncontiguous_indexed=(void*)noncontiguous_indexed_char;
713 void free_indexed(MPI_Datatype* type){
714 xbt_free(((s_smpi_mpi_indexed_t *)(*type)->substruct)->block_lengths);
715 xbt_free(((s_smpi_mpi_indexed_t *)(*type)->substruct)->block_indices);
719 * Create a Sub type indexed to be able to serialize and unserialize it
720 * the structure s_smpi_mpi_indexed_t is derived from s_smpi_subtype which
721 * required the functions unserialize and serialize
723 s_smpi_mpi_indexed_t* smpi_datatype_indexed_create( int* block_lengths,
726 MPI_Datatype old_type,
728 s_smpi_mpi_indexed_t *new_t= xbt_new(s_smpi_mpi_indexed_t,1);
729 new_t->base.serialize = &serialize_indexed;
730 new_t->base.unserialize = &unserialize_indexed;
731 new_t->base.subtype_free = &free_indexed;
732 //TODO : add a custom function for each time to clean these
733 new_t->block_lengths= xbt_new(int, block_count);
734 new_t->block_indices= xbt_new(int, block_count);
736 for(i=0;i<block_count;i++){
737 new_t->block_lengths[i]=block_lengths[i];
738 new_t->block_indices[i]=block_indices[i];
740 new_t->block_count = block_count;
741 new_t->old_type = old_type;
742 new_t->size_oldtype = size_oldtype;
747 int smpi_datatype_indexed(int count, int* blocklens, int* indices, MPI_Datatype old_type, MPI_Datatype* new_type)
756 lb=indices[0]*smpi_datatype_get_extent(old_type);
757 ub=indices[0]*smpi_datatype_get_extent(old_type) + blocklens[0]*smpi_datatype_ub(old_type);
760 for(i=0; i< count; i++){
763 size += blocklens[i];
765 if(indices[i]*smpi_datatype_get_extent(old_type)+smpi_datatype_lb(old_type)<lb)
766 lb = indices[i]*smpi_datatype_get_extent(old_type)+smpi_datatype_lb(old_type);
767 if(indices[i]*smpi_datatype_get_extent(old_type)+blocklens[i]*smpi_datatype_ub(old_type)>ub)
768 ub = indices[i]*smpi_datatype_get_extent(old_type)+blocklens[i]*smpi_datatype_ub(old_type);
770 if ( (i< count -1) && (indices[i]+blocklens[i] != indices[i+1]) )contiguous=0;
772 if (old_type->has_subtype == 1)
776 s_smpi_mpi_indexed_t* subtype = smpi_datatype_indexed_create( blocklens,
780 smpi_datatype_size(old_type));
781 smpi_datatype_create(new_type, size *
782 smpi_datatype_size(old_type),lb,ub,1, subtype, DT_FLAG_DATA);
784 s_smpi_mpi_contiguous_t* subtype = smpi_datatype_contiguous_create( lb,
787 smpi_datatype_size(old_type));
788 smpi_datatype_create(new_type, size *
789 smpi_datatype_size(old_type),lb,ub,1, subtype, DT_FLAG_DATA|DT_FLAG_CONTIGUOUS);
797 Hindexed Implementation - Indexed with indices in bytes
801 * Copies noncontiguous data into contiguous memory.
802 * @param contiguous_hindexed - output hindexed
803 * @param noncontiguous_hindexed - input hindexed
804 * @param type - pointer contening :
805 * - block_lengths - the width or height of blocked matrix
806 * - block_indices - indices of each data, in bytes
807 * - count - the number of rows of matrix
809 void serialize_hindexed( const void *noncontiguous_hindexed,
810 void *contiguous_hindexed,
814 s_smpi_mpi_hindexed_t* type_c = (s_smpi_mpi_hindexed_t*)type;
816 char* contiguous_hindexed_char = (char*)contiguous_hindexed;
817 char* noncontiguous_hindexed_char = (char*)noncontiguous_hindexed+ type_c->block_indices[0];
818 for(j=0; j<count;j++){
819 for (i = 0; i < type_c->block_count; i++) {
820 if (type_c->old_type->has_subtype == 0)
821 memcpy(contiguous_hindexed_char,
822 noncontiguous_hindexed_char, type_c->block_lengths[i] * type_c->size_oldtype);
824 ((s_smpi_subtype_t*)type_c->old_type->substruct)->serialize( noncontiguous_hindexed_char,
825 contiguous_hindexed_char,
826 type_c->block_lengths[i],
827 type_c->old_type->substruct);
829 contiguous_hindexed_char += type_c->block_lengths[i]*type_c->size_oldtype;
830 if (i<type_c->block_count-1)noncontiguous_hindexed_char = (char*)noncontiguous_hindexed + type_c->block_indices[i+1];
831 else noncontiguous_hindexed_char += type_c->block_lengths[i]*smpi_datatype_get_extent(type_c->old_type);
833 noncontiguous_hindexed=(void*)noncontiguous_hindexed_char;
837 * Copies contiguous data into noncontiguous memory.
838 * @param noncontiguous_hindexed - output hindexed
839 * @param contiguous_hindexed - input hindexed
840 * @param type - pointer contening :
841 * - block_lengths - the width or height of blocked matrix
842 * - block_indices - indices of each data, in bytes
843 * - count - the number of rows of matrix
845 void unserialize_hindexed( const void *contiguous_hindexed,
846 void *noncontiguous_hindexed,
850 s_smpi_mpi_hindexed_t* type_c = (s_smpi_mpi_hindexed_t*)type;
853 char* contiguous_hindexed_char = (char*)contiguous_hindexed;
854 char* noncontiguous_hindexed_char = (char*)noncontiguous_hindexed+ type_c->block_indices[0];
855 for(j=0; j<count;j++){
856 for (i = 0; i < type_c->block_count; i++) {
857 if (type_c->old_type->has_subtype == 0)
858 memcpy(noncontiguous_hindexed_char,
859 contiguous_hindexed_char, type_c->block_lengths[i] * type_c->size_oldtype);
861 ((s_smpi_subtype_t*)type_c->old_type->substruct)->unserialize( contiguous_hindexed_char,
862 noncontiguous_hindexed_char,
863 type_c->block_lengths[i],
864 type_c->old_type->substruct);
866 contiguous_hindexed_char += type_c->block_lengths[i]*type_c->size_oldtype;
867 if (i<type_c->block_count-1)noncontiguous_hindexed_char = (char*)noncontiguous_hindexed + type_c->block_indices[i+1];
868 else noncontiguous_hindexed_char += type_c->block_lengths[i]*smpi_datatype_get_extent(type_c->old_type);
870 noncontiguous_hindexed=(void*)noncontiguous_hindexed_char;
874 void free_hindexed(MPI_Datatype* type){
875 xbt_free(((s_smpi_mpi_hindexed_t *)(*type)->substruct)->block_lengths);
876 xbt_free(((s_smpi_mpi_hindexed_t *)(*type)->substruct)->block_indices);
880 * Create a Sub type hindexed to be able to serialize and unserialize it
881 * the structure s_smpi_mpi_hindexed_t is derived from s_smpi_subtype which
882 * required the functions unserialize and serialize
884 s_smpi_mpi_hindexed_t* smpi_datatype_hindexed_create( int* block_lengths,
885 MPI_Aint* block_indices,
887 MPI_Datatype old_type,
889 s_smpi_mpi_hindexed_t *new_t= xbt_new(s_smpi_mpi_hindexed_t,1);
890 new_t->base.serialize = &serialize_hindexed;
891 new_t->base.unserialize = &unserialize_hindexed;
892 new_t->base.subtype_free = &free_hindexed;
893 //TODO : add a custom function for each time to clean these
894 new_t->block_lengths= xbt_new(int, block_count);
895 new_t->block_indices= xbt_new(MPI_Aint, block_count);
897 for(i=0;i<block_count;i++){
898 new_t->block_lengths[i]=block_lengths[i];
899 new_t->block_indices[i]=block_indices[i];
901 new_t->block_count = block_count;
902 new_t->old_type = old_type;
903 new_t->size_oldtype = size_oldtype;
908 int smpi_datatype_hindexed(int count, int* blocklens, MPI_Aint* indices, MPI_Datatype old_type, MPI_Datatype* new_type)
917 lb=indices[0] + smpi_datatype_lb(old_type);
918 ub=indices[0] + blocklens[0]*smpi_datatype_ub(old_type);
920 for(i=0; i< count; i++){
923 size += blocklens[i];
925 if(indices[i]+smpi_datatype_lb(old_type)<lb) lb = indices[i]+smpi_datatype_lb(old_type);
926 if(indices[i]+blocklens[i]*smpi_datatype_ub(old_type)>ub) ub = indices[i]+blocklens[i]*smpi_datatype_ub(old_type);
928 if ( (i< count -1) && (indices[i]+blocklens[i]*smpi_datatype_size(old_type) != indices[i+1]) )contiguous=0;
930 if (old_type->has_subtype == 1 || lb!=0)
934 s_smpi_mpi_hindexed_t* subtype = smpi_datatype_hindexed_create( blocklens,
938 smpi_datatype_size(old_type));
939 smpi_datatype_create(new_type, size * smpi_datatype_size(old_type),
942 ,1, subtype, DT_FLAG_DATA);
944 s_smpi_mpi_contiguous_t* subtype = smpi_datatype_contiguous_create( lb,
947 smpi_datatype_size(old_type));
948 smpi_datatype_create(new_type, size * smpi_datatype_size(old_type),
949 0,size * smpi_datatype_size(old_type),
950 1, subtype, DT_FLAG_DATA|DT_FLAG_CONTIGUOUS);
958 struct Implementation - Indexed with indices in bytes
962 * Copies noncontiguous data into contiguous memory.
963 * @param contiguous_struct - output struct
964 * @param noncontiguous_struct - input struct
965 * @param type - pointer contening :
966 * - stride - stride of between noncontiguous data
967 * - block_length - the width or height of blocked matrix
968 * - count - the number of rows of matrix
970 void serialize_struct( const void *noncontiguous_struct,
971 void *contiguous_struct,
975 s_smpi_mpi_struct_t* type_c = (s_smpi_mpi_struct_t*)type;
977 char* contiguous_struct_char = (char*)contiguous_struct;
978 char* noncontiguous_struct_char = (char*)noncontiguous_struct+ type_c->block_indices[0];
979 for(j=0; j<count;j++){
980 for (i = 0; i < type_c->block_count; i++) {
981 if (type_c->old_types[i]->has_subtype == 0)
982 memcpy(contiguous_struct_char,
983 noncontiguous_struct_char, type_c->block_lengths[i] * smpi_datatype_size(type_c->old_types[i]));
985 ((s_smpi_subtype_t*)type_c->old_types[i]->substruct)->serialize( noncontiguous_struct_char,
986 contiguous_struct_char,
987 type_c->block_lengths[i],
988 type_c->old_types[i]->substruct);
991 contiguous_struct_char += type_c->block_lengths[i]*smpi_datatype_size(type_c->old_types[i]);
992 if (i<type_c->block_count-1)noncontiguous_struct_char = (char*)noncontiguous_struct + type_c->block_indices[i+1];
993 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 ?
995 noncontiguous_struct=(void*)noncontiguous_struct_char;
999 * Copies contiguous data into noncontiguous memory.
1000 * @param noncontiguous_struct - output struct
1001 * @param contiguous_struct - input struct
1002 * @param type - pointer contening :
1003 * - stride - stride of between noncontiguous data
1004 * - block_length - the width or height of blocked matrix
1005 * - count - the number of rows of matrix
1007 void unserialize_struct( const void *contiguous_struct,
1008 void *noncontiguous_struct,
1012 s_smpi_mpi_struct_t* type_c = (s_smpi_mpi_struct_t*)type;
1015 char* contiguous_struct_char = (char*)contiguous_struct;
1016 char* noncontiguous_struct_char = (char*)noncontiguous_struct+ type_c->block_indices[0];
1017 for(j=0; j<count;j++){
1018 for (i = 0; i < type_c->block_count; i++) {
1019 if (type_c->old_types[i]->has_subtype == 0)
1020 memcpy(noncontiguous_struct_char,
1021 contiguous_struct_char, type_c->block_lengths[i] * smpi_datatype_size(type_c->old_types[i]));
1023 ((s_smpi_subtype_t*)type_c->old_types[i]->substruct)->unserialize( contiguous_struct_char,
1024 noncontiguous_struct_char,
1025 type_c->block_lengths[i],
1026 type_c->old_types[i]->substruct);
1028 contiguous_struct_char += type_c->block_lengths[i]*smpi_datatype_size(type_c->old_types[i]);
1029 if (i<type_c->block_count-1)noncontiguous_struct_char = (char*)noncontiguous_struct + type_c->block_indices[i+1];
1030 else noncontiguous_struct_char += type_c->block_lengths[i]*smpi_datatype_get_extent(type_c->old_types[i]);
1032 noncontiguous_struct=(void*)noncontiguous_struct_char;
1037 void free_struct(MPI_Datatype* type){
1038 xbt_free(((s_smpi_mpi_struct_t *)(*type)->substruct)->block_lengths);
1039 xbt_free(((s_smpi_mpi_struct_t *)(*type)->substruct)->block_indices);
1040 xbt_free(((s_smpi_mpi_struct_t *)(*type)->substruct)->old_types);
1044 * Create a Sub type struct to be able to serialize and unserialize it
1045 * the structure s_smpi_mpi_struct_t is derived from s_smpi_subtype which
1046 * required the functions unserialize and serialize
1048 s_smpi_mpi_struct_t* smpi_datatype_struct_create( int* block_lengths,
1049 MPI_Aint* block_indices,
1051 MPI_Datatype* old_types){
1052 s_smpi_mpi_struct_t *new_t= xbt_new(s_smpi_mpi_struct_t,1);
1053 new_t->base.serialize = &serialize_struct;
1054 new_t->base.unserialize = &unserialize_struct;
1055 new_t->base.subtype_free = &free_struct;
1056 //TODO : add a custom function for each time to clean these
1057 new_t->block_lengths= xbt_new(int, block_count);
1058 new_t->block_indices= xbt_new(MPI_Aint, block_count);
1059 new_t->old_types= xbt_new(MPI_Datatype, block_count);
1061 for(i=0;i<block_count;i++){
1062 new_t->block_lengths[i]=block_lengths[i];
1063 new_t->block_indices[i]=block_indices[i];
1064 new_t->old_types[i]=old_types[i];
1066 //new_t->block_lengths = block_lengths;
1067 //new_t->block_indices = block_indices;
1068 new_t->block_count = block_count;
1069 //new_t->old_types = old_types;
1074 int smpi_datatype_struct(int count, int* blocklens, MPI_Aint* indices, MPI_Datatype* old_types, MPI_Datatype* new_type)
1083 lb=indices[0] + smpi_datatype_lb(old_types[0]);
1084 ub=indices[0] + blocklens[0]*smpi_datatype_ub(old_types[0]);
1088 for(i=0; i< count; i++){
1091 if (old_types[i]->has_subtype == 1)
1094 size += blocklens[i]*smpi_datatype_size(old_types[i]);
1095 if (old_types[i]==MPI_LB){
1099 if (old_types[i]==MPI_UB){
1104 if(!forced_lb && indices[i]+smpi_datatype_lb(old_types[i])<lb) lb = indices[i];
1105 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]);
1107 if ( (i< count -1) && (indices[i]+blocklens[i]*smpi_datatype_size(old_types[i]) != indices[i+1]) )contiguous=0;
1111 s_smpi_mpi_struct_t* subtype = smpi_datatype_struct_create( blocklens,
1116 smpi_datatype_create(new_type, size, lb, ub,1, subtype, DT_FLAG_DATA);
1118 s_smpi_mpi_contiguous_t* subtype = smpi_datatype_contiguous_create( lb,
1122 smpi_datatype_create(new_type, size, lb, ub,1, subtype, DT_FLAG_DATA|DT_FLAG_CONTIGUOUS);
1127 void smpi_datatype_commit(MPI_Datatype *datatype)
1129 (*datatype)->flags= ((*datatype)->flags | DT_FLAG_COMMITED);
1132 typedef struct s_smpi_mpi_op {
1133 MPI_User_function *func;
1137 #define MAX_OP(a, b) (b) = (a) < (b) ? (b) : (a)
1138 #define MIN_OP(a, b) (b) = (a) < (b) ? (a) : (b)
1139 #define SUM_OP(a, b) (b) += (a)
1140 #define PROD_OP(a, b) (b) *= (a)
1141 #define LAND_OP(a, b) (b) = (a) && (b)
1142 #define LOR_OP(a, b) (b) = (a) || (b)
1143 #define LXOR_OP(a, b) (b) = (!(a) && (b)) || ((a) && !(b))
1144 #define BAND_OP(a, b) (b) &= (a)
1145 #define BOR_OP(a, b) (b) |= (a)
1146 #define BXOR_OP(a, b) (b) ^= (a)
1147 #define MAXLOC_OP(a, b) (b) = (a.value) < (b.value) ? (b) : (a)
1148 #define MINLOC_OP(a, b) (b) = (a.value) < (b.value) ? (a) : (b)
1149 //TODO : MINLOC & MAXLOC
1151 #define APPLY_FUNC(a, b, length, type, func) \
1154 type* x = (type*)(a); \
1155 type* y = (type*)(b); \
1156 for(i = 0; i < *(length); i++) { \
1161 static void max_func(void *a, void *b, int *length,
1162 MPI_Datatype * datatype)
1164 if (*datatype == MPI_CHAR) {
1165 APPLY_FUNC(a, b, length, char, MAX_OP);
1166 } else if (*datatype == MPI_SHORT) {
1167 APPLY_FUNC(a, b, length, short, MAX_OP);
1168 } else if (*datatype == MPI_INT) {
1169 APPLY_FUNC(a, b, length, int, MAX_OP);
1170 } else if (*datatype == MPI_LONG) {
1171 APPLY_FUNC(a, b, length, long, MAX_OP);
1172 } else if (*datatype == MPI_UNSIGNED_SHORT) {
1173 APPLY_FUNC(a, b, length, unsigned short, MAX_OP);
1174 } else if (*datatype == MPI_UNSIGNED) {
1175 APPLY_FUNC(a, b, length, unsigned int, MAX_OP);
1176 } else if (*datatype == MPI_UNSIGNED_LONG) {
1177 APPLY_FUNC(a, b, length, unsigned long, MAX_OP);
1178 } else if (*datatype == MPI_FLOAT) {
1179 APPLY_FUNC(a, b, length, float, MAX_OP);
1180 } else if (*datatype == MPI_DOUBLE) {
1181 APPLY_FUNC(a, b, length, double, MAX_OP);
1182 } else if (*datatype == MPI_LONG_DOUBLE) {
1183 APPLY_FUNC(a, b, length, long double, MAX_OP);
1187 static void min_func(void *a, void *b, int *length,
1188 MPI_Datatype * datatype)
1190 if (*datatype == MPI_CHAR) {
1191 APPLY_FUNC(a, b, length, char, MIN_OP);
1192 } else if (*datatype == MPI_SHORT) {
1193 APPLY_FUNC(a, b, length, short, MIN_OP);
1194 } else if (*datatype == MPI_INT) {
1195 APPLY_FUNC(a, b, length, int, MIN_OP);
1196 } else if (*datatype == MPI_LONG) {
1197 APPLY_FUNC(a, b, length, long, MIN_OP);
1198 } else if (*datatype == MPI_UNSIGNED_SHORT) {
1199 APPLY_FUNC(a, b, length, unsigned short, MIN_OP);
1200 } else if (*datatype == MPI_UNSIGNED) {
1201 APPLY_FUNC(a, b, length, unsigned int, MIN_OP);
1202 } else if (*datatype == MPI_UNSIGNED_LONG) {
1203 APPLY_FUNC(a, b, length, unsigned long, MIN_OP);
1204 } else if (*datatype == MPI_FLOAT) {
1205 APPLY_FUNC(a, b, length, float, MIN_OP);
1206 } else if (*datatype == MPI_DOUBLE) {
1207 APPLY_FUNC(a, b, length, double, MIN_OP);
1208 } else if (*datatype == MPI_LONG_DOUBLE) {
1209 APPLY_FUNC(a, b, length, long double, MIN_OP);
1213 static void sum_func(void *a, void *b, int *length,
1214 MPI_Datatype * datatype)
1216 if (*datatype == MPI_CHAR) {
1217 APPLY_FUNC(a, b, length, char, SUM_OP);
1218 } else if (*datatype == MPI_SHORT) {
1219 APPLY_FUNC(a, b, length, short, SUM_OP);
1220 } else if (*datatype == MPI_INT) {
1221 APPLY_FUNC(a, b, length, int, SUM_OP);
1222 } else if (*datatype == MPI_LONG) {
1223 APPLY_FUNC(a, b, length, long, SUM_OP);
1224 } else if (*datatype == MPI_UNSIGNED_SHORT) {
1225 APPLY_FUNC(a, b, length, unsigned short, SUM_OP);
1226 } else if (*datatype == MPI_UNSIGNED) {
1227 APPLY_FUNC(a, b, length, unsigned int, SUM_OP);
1228 } else if (*datatype == MPI_UNSIGNED_LONG) {
1229 APPLY_FUNC(a, b, length, unsigned long, SUM_OP);
1230 } else if (*datatype == MPI_FLOAT) {
1231 APPLY_FUNC(a, b, length, float, SUM_OP);
1232 } else if (*datatype == MPI_DOUBLE) {
1233 APPLY_FUNC(a, b, length, double, SUM_OP);
1234 } else if (*datatype == MPI_LONG_DOUBLE) {
1235 APPLY_FUNC(a, b, length, long double, SUM_OP);
1236 } else if (*datatype == MPI_C_FLOAT_COMPLEX) {
1237 APPLY_FUNC(a, b, length, float _Complex, SUM_OP);
1238 } else if (*datatype == MPI_C_DOUBLE_COMPLEX) {
1239 APPLY_FUNC(a, b, length, double _Complex, SUM_OP);
1240 } else if (*datatype == MPI_C_LONG_DOUBLE_COMPLEX) {
1241 APPLY_FUNC(a, b, length, long double _Complex, SUM_OP);
1245 static void prod_func(void *a, void *b, int *length,
1246 MPI_Datatype * datatype)
1248 if (*datatype == MPI_CHAR) {
1249 APPLY_FUNC(a, b, length, char, PROD_OP);
1250 } else if (*datatype == MPI_SHORT) {
1251 APPLY_FUNC(a, b, length, short, PROD_OP);
1252 } else if (*datatype == MPI_INT) {
1253 APPLY_FUNC(a, b, length, int, PROD_OP);
1254 } else if (*datatype == MPI_LONG) {
1255 APPLY_FUNC(a, b, length, long, PROD_OP);
1256 } else if (*datatype == MPI_UNSIGNED_SHORT) {
1257 APPLY_FUNC(a, b, length, unsigned short, PROD_OP);
1258 } else if (*datatype == MPI_UNSIGNED) {
1259 APPLY_FUNC(a, b, length, unsigned int, PROD_OP);
1260 } else if (*datatype == MPI_UNSIGNED_LONG) {
1261 APPLY_FUNC(a, b, length, unsigned long, PROD_OP);
1262 } else if (*datatype == MPI_FLOAT) {
1263 APPLY_FUNC(a, b, length, float, PROD_OP);
1264 } else if (*datatype == MPI_DOUBLE) {
1265 APPLY_FUNC(a, b, length, double, PROD_OP);
1266 } else if (*datatype == MPI_LONG_DOUBLE) {
1267 APPLY_FUNC(a, b, length, long double, PROD_OP);
1268 } else if (*datatype == MPI_C_FLOAT_COMPLEX) {
1269 APPLY_FUNC(a, b, length, float _Complex, PROD_OP);
1270 } else if (*datatype == MPI_C_DOUBLE_COMPLEX) {
1271 APPLY_FUNC(a, b, length, double _Complex, PROD_OP);
1272 } else if (*datatype == MPI_C_LONG_DOUBLE_COMPLEX) {
1273 APPLY_FUNC(a, b, length, long double _Complex, PROD_OP);
1277 static void land_func(void *a, void *b, int *length,
1278 MPI_Datatype * datatype)
1280 if (*datatype == MPI_CHAR) {
1281 APPLY_FUNC(a, b, length, char, LAND_OP);
1282 } else if (*datatype == MPI_SHORT) {
1283 APPLY_FUNC(a, b, length, short, LAND_OP);
1284 } else if (*datatype == MPI_INT) {
1285 APPLY_FUNC(a, b, length, int, LAND_OP);
1286 } else if (*datatype == MPI_LONG) {
1287 APPLY_FUNC(a, b, length, long, LAND_OP);
1288 } else if (*datatype == MPI_UNSIGNED_SHORT) {
1289 APPLY_FUNC(a, b, length, unsigned short, LAND_OP);
1290 } else if (*datatype == MPI_UNSIGNED) {
1291 APPLY_FUNC(a, b, length, unsigned int, LAND_OP);
1292 } else if (*datatype == MPI_UNSIGNED_LONG) {
1293 APPLY_FUNC(a, b, length, unsigned long, LAND_OP);
1294 } else if (*datatype == MPI_C_BOOL) {
1295 APPLY_FUNC(a, b, length, _Bool, LAND_OP);
1299 static void lor_func(void *a, void *b, int *length,
1300 MPI_Datatype * datatype)
1302 if (*datatype == MPI_CHAR) {
1303 APPLY_FUNC(a, b, length, char, LOR_OP);
1304 } else if (*datatype == MPI_SHORT) {
1305 APPLY_FUNC(a, b, length, short, LOR_OP);
1306 } else if (*datatype == MPI_INT) {
1307 APPLY_FUNC(a, b, length, int, LOR_OP);
1308 } else if (*datatype == MPI_LONG) {
1309 APPLY_FUNC(a, b, length, long, LOR_OP);
1310 } else if (*datatype == MPI_UNSIGNED_SHORT) {
1311 APPLY_FUNC(a, b, length, unsigned short, LOR_OP);
1312 } else if (*datatype == MPI_UNSIGNED) {
1313 APPLY_FUNC(a, b, length, unsigned int, LOR_OP);
1314 } else if (*datatype == MPI_UNSIGNED_LONG) {
1315 APPLY_FUNC(a, b, length, unsigned long, LOR_OP);
1316 } else if (*datatype == MPI_C_BOOL) {
1317 APPLY_FUNC(a, b, length, _Bool, LOR_OP);
1321 static void lxor_func(void *a, void *b, int *length,
1322 MPI_Datatype * datatype)
1324 if (*datatype == MPI_CHAR) {
1325 APPLY_FUNC(a, b, length, char, LXOR_OP);
1326 } else if (*datatype == MPI_SHORT) {
1327 APPLY_FUNC(a, b, length, short, LXOR_OP);
1328 } else if (*datatype == MPI_INT) {
1329 APPLY_FUNC(a, b, length, int, LXOR_OP);
1330 } else if (*datatype == MPI_LONG) {
1331 APPLY_FUNC(a, b, length, long, LXOR_OP);
1332 } else if (*datatype == MPI_UNSIGNED_SHORT) {
1333 APPLY_FUNC(a, b, length, unsigned short, LXOR_OP);
1334 } else if (*datatype == MPI_UNSIGNED) {
1335 APPLY_FUNC(a, b, length, unsigned int, LXOR_OP);
1336 } else if (*datatype == MPI_UNSIGNED_LONG) {
1337 APPLY_FUNC(a, b, length, unsigned long, LXOR_OP);
1338 } else if (*datatype == MPI_C_BOOL) {
1339 APPLY_FUNC(a, b, length, _Bool, LXOR_OP);
1343 static void band_func(void *a, void *b, int *length,
1344 MPI_Datatype * datatype)
1346 if (*datatype == MPI_CHAR) {
1347 APPLY_FUNC(a, b, length, char, BAND_OP);
1349 if (*datatype == MPI_SHORT) {
1350 APPLY_FUNC(a, b, length, short, BAND_OP);
1351 } else if (*datatype == MPI_INT) {
1352 APPLY_FUNC(a, b, length, int, BAND_OP);
1353 } else if (*datatype == MPI_LONG) {
1354 APPLY_FUNC(a, b, length, long, BAND_OP);
1355 } else if (*datatype == MPI_UNSIGNED_SHORT) {
1356 APPLY_FUNC(a, b, length, unsigned short, BAND_OP);
1357 } else if (*datatype == MPI_UNSIGNED) {
1358 APPLY_FUNC(a, b, length, unsigned int, BAND_OP);
1359 } else if (*datatype == MPI_UNSIGNED_LONG) {
1360 APPLY_FUNC(a, b, length, unsigned long, BAND_OP);
1361 } else if (*datatype == MPI_BYTE) {
1362 APPLY_FUNC(a, b, length, uint8_t, BAND_OP);
1366 static void bor_func(void *a, void *b, int *length,
1367 MPI_Datatype * datatype)
1369 if (*datatype == MPI_CHAR) {
1370 APPLY_FUNC(a, b, length, char, BOR_OP);
1371 } else if (*datatype == MPI_SHORT) {
1372 APPLY_FUNC(a, b, length, short, BOR_OP);
1373 } else if (*datatype == MPI_INT) {
1374 APPLY_FUNC(a, b, length, int, BOR_OP);
1375 } else if (*datatype == MPI_LONG) {
1376 APPLY_FUNC(a, b, length, long, BOR_OP);
1377 } else if (*datatype == MPI_UNSIGNED_SHORT) {
1378 APPLY_FUNC(a, b, length, unsigned short, BOR_OP);
1379 } else if (*datatype == MPI_UNSIGNED) {
1380 APPLY_FUNC(a, b, length, unsigned int, BOR_OP);
1381 } else if (*datatype == MPI_UNSIGNED_LONG) {
1382 APPLY_FUNC(a, b, length, unsigned long, BOR_OP);
1383 } else if (*datatype == MPI_BYTE) {
1384 APPLY_FUNC(a, b, length, uint8_t, BOR_OP);
1388 static void bxor_func(void *a, void *b, int *length,
1389 MPI_Datatype * datatype)
1391 if (*datatype == MPI_CHAR) {
1392 APPLY_FUNC(a, b, length, char, BXOR_OP);
1393 } else if (*datatype == MPI_SHORT) {
1394 APPLY_FUNC(a, b, length, short, BXOR_OP);
1395 } else if (*datatype == MPI_INT) {
1396 APPLY_FUNC(a, b, length, int, BXOR_OP);
1397 } else if (*datatype == MPI_LONG) {
1398 APPLY_FUNC(a, b, length, long, BXOR_OP);
1399 } else if (*datatype == MPI_UNSIGNED_SHORT) {
1400 APPLY_FUNC(a, b, length, unsigned short, BXOR_OP);
1401 } else if (*datatype == MPI_UNSIGNED) {
1402 APPLY_FUNC(a, b, length, unsigned int, BXOR_OP);
1403 } else if (*datatype == MPI_UNSIGNED_LONG) {
1404 APPLY_FUNC(a, b, length, unsigned long, BXOR_OP);
1405 } else if (*datatype == MPI_BYTE) {
1406 APPLY_FUNC(a, b, length, uint8_t, BXOR_OP);
1410 static void minloc_func(void *a, void *b, int *length,
1411 MPI_Datatype * datatype)
1413 if (*datatype == MPI_FLOAT_INT) {
1414 APPLY_FUNC(a, b, length, float_int, MINLOC_OP);
1415 } else if (*datatype == MPI_LONG_INT) {
1416 APPLY_FUNC(a, b, length, long_int, MINLOC_OP);
1417 } else if (*datatype == MPI_DOUBLE_INT) {
1418 APPLY_FUNC(a, b, length, double_int, MINLOC_OP);
1419 } else if (*datatype == MPI_SHORT_INT) {
1420 APPLY_FUNC(a, b, length, short_int, MINLOC_OP);
1421 } else if (*datatype == MPI_2INT) {
1422 APPLY_FUNC(a, b, length, int_int, MINLOC_OP);
1423 } else if (*datatype == MPI_LONG_DOUBLE_INT) {
1424 APPLY_FUNC(a, b, length, long_double_int, MINLOC_OP);
1425 } else if (*datatype == MPI_2FLOAT) {
1426 APPLY_FUNC(a, b, length, float_float, MINLOC_OP);
1427 } else if (*datatype == MPI_2DOUBLE) {
1428 APPLY_FUNC(a, b, length, double_double, MINLOC_OP);
1432 static void maxloc_func(void *a, void *b, int *length,
1433 MPI_Datatype * datatype)
1435 if (*datatype == MPI_FLOAT_INT) {
1436 APPLY_FUNC(a, b, length, float_int, MAXLOC_OP);
1437 } else if (*datatype == MPI_LONG_INT) {
1438 APPLY_FUNC(a, b, length, long_int, MAXLOC_OP);
1439 } else if (*datatype == MPI_DOUBLE_INT) {
1440 APPLY_FUNC(a, b, length, double_int, MAXLOC_OP);
1441 } else if (*datatype == MPI_SHORT_INT) {
1442 APPLY_FUNC(a, b, length, short_int, MAXLOC_OP);
1443 } else if (*datatype == MPI_2INT) {
1444 APPLY_FUNC(a, b, length, int_int, MAXLOC_OP);
1445 } else if (*datatype == MPI_LONG_DOUBLE_INT) {
1446 APPLY_FUNC(a, b, length, long_double_int, MAXLOC_OP);
1447 } else if (*datatype == MPI_2FLOAT) {
1448 APPLY_FUNC(a, b, length, float_float, MAXLOC_OP);
1449 } else if (*datatype == MPI_2DOUBLE) {
1450 APPLY_FUNC(a, b, length, double_double, MAXLOC_OP);
1455 #define CREATE_MPI_OP(name, func) \
1456 static s_smpi_mpi_op_t mpi_##name = { &(func) /* func */, TRUE }; \
1457 MPI_Op name = &mpi_##name;
1459 CREATE_MPI_OP(MPI_MAX, max_func);
1460 CREATE_MPI_OP(MPI_MIN, min_func);
1461 CREATE_MPI_OP(MPI_SUM, sum_func);
1462 CREATE_MPI_OP(MPI_PROD, prod_func);
1463 CREATE_MPI_OP(MPI_LAND, land_func);
1464 CREATE_MPI_OP(MPI_LOR, lor_func);
1465 CREATE_MPI_OP(MPI_LXOR, lxor_func);
1466 CREATE_MPI_OP(MPI_BAND, band_func);
1467 CREATE_MPI_OP(MPI_BOR, bor_func);
1468 CREATE_MPI_OP(MPI_BXOR, bxor_func);
1469 CREATE_MPI_OP(MPI_MAXLOC, maxloc_func);
1470 CREATE_MPI_OP(MPI_MINLOC, minloc_func);
1472 MPI_Op smpi_op_new(MPI_User_function * function, int commute)
1475 op = xbt_new(s_smpi_mpi_op_t, 1);
1476 op->func = function;
1477 op-> is_commute = commute;
1481 int smpi_op_is_commute(MPI_Op op)
1483 return (op==MPI_OP_NULL) ? 1 : op-> is_commute;
1486 void smpi_op_destroy(MPI_Op op)
1491 void smpi_op_apply(MPI_Op op, void *invec, void *inoutvec, int *len,
1492 MPI_Datatype * datatype)
1494 if(smpi_privatize_global_variables){ //we need to switch here, as the called function may silently touch global variables
1495 XBT_VERB("Applying operation, switch to the right data frame ");
1496 switch_data_segment(smpi_process_index());
1499 if(!_xbt_replay_is_active())
1500 op->func(invec, inoutvec, len, datatype);