1 /* smpi_mpi_dt.c -- MPI primitives to handle datatypes */
2 /* FIXME: a very incomplete implementation */
4 /* Copyright (c) 2009-2015. 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, "Logging specific to SMPI (datatype)");
22 xbt_dict_t smpi_type_keyvals = NULL;
23 int type_keyval_id=0;//avoid collisions
25 #define CREATE_MPI_DATATYPE(name, type) \
26 static s_smpi_mpi_datatype_t mpi_##name = { \
28 sizeof(type), /* size */ \
29 0, /*was 1 sizeof_substruct*/ \
31 sizeof(type), /* ub = lb + size */ \
32 DT_FLAG_BASIC, /* flags */ \
33 NULL, /* attributes */ \
34 NULL, /* pointer on extended struct*/ \
35 0 /* in_use counter */ \
37 const MPI_Datatype name = &mpi_##name;
39 #define CREATE_MPI_DATATYPE_NULL(name) \
40 static s_smpi_mpi_datatype_t mpi_##name = { \
43 0, /* was 1 sizeof_substruct*/ \
45 0, /* ub = lb + size */ \
46 DT_FLAG_BASIC, /* flags */ \
47 NULL, /* attributes */ \
48 NULL, /* pointer on extended struct*/ \
49 0 /* in_use counter */ \
51 const MPI_Datatype name = &mpi_##name;
53 //The following are datatypes for the MPI functions MPI_MAXLOC and MPI_MINLOC.
94 // Predefined data types
95 CREATE_MPI_DATATYPE(MPI_CHAR, char);
96 CREATE_MPI_DATATYPE(MPI_SHORT, short);
97 CREATE_MPI_DATATYPE(MPI_INT, int);
98 CREATE_MPI_DATATYPE(MPI_LONG, long);
99 CREATE_MPI_DATATYPE(MPI_LONG_LONG, long long);
100 CREATE_MPI_DATATYPE(MPI_SIGNED_CHAR, signed char);
101 CREATE_MPI_DATATYPE(MPI_UNSIGNED_CHAR, unsigned char);
102 CREATE_MPI_DATATYPE(MPI_UNSIGNED_SHORT, unsigned short);
103 CREATE_MPI_DATATYPE(MPI_UNSIGNED, unsigned int);
104 CREATE_MPI_DATATYPE(MPI_UNSIGNED_LONG, unsigned long);
105 CREATE_MPI_DATATYPE(MPI_UNSIGNED_LONG_LONG, unsigned long long);
106 CREATE_MPI_DATATYPE(MPI_FLOAT, float);
107 CREATE_MPI_DATATYPE(MPI_DOUBLE, double);
108 CREATE_MPI_DATATYPE(MPI_LONG_DOUBLE, long double);
109 CREATE_MPI_DATATYPE(MPI_WCHAR, wchar_t);
110 CREATE_MPI_DATATYPE(MPI_C_BOOL, bool);
111 CREATE_MPI_DATATYPE(MPI_BYTE, int8_t);
112 CREATE_MPI_DATATYPE(MPI_INT8_T, int8_t);
113 CREATE_MPI_DATATYPE(MPI_INT16_T, int16_t);
114 CREATE_MPI_DATATYPE(MPI_INT32_T, int32_t);
115 CREATE_MPI_DATATYPE(MPI_INT64_T, int64_t);
116 CREATE_MPI_DATATYPE(MPI_UINT8_T, uint8_t);
117 CREATE_MPI_DATATYPE(MPI_UINT16_T, uint16_t);
118 CREATE_MPI_DATATYPE(MPI_UINT32_T, uint32_t);
119 CREATE_MPI_DATATYPE(MPI_UINT64_T, uint64_t);
120 CREATE_MPI_DATATYPE(MPI_C_FLOAT_COMPLEX, float _Complex);
121 CREATE_MPI_DATATYPE(MPI_C_DOUBLE_COMPLEX, double _Complex);
122 CREATE_MPI_DATATYPE(MPI_C_LONG_DOUBLE_COMPLEX, long double _Complex);
123 CREATE_MPI_DATATYPE(MPI_AINT, MPI_Aint);
124 CREATE_MPI_DATATYPE(MPI_OFFSET, MPI_Offset);
126 CREATE_MPI_DATATYPE(MPI_FLOAT_INT, float_int);
127 CREATE_MPI_DATATYPE(MPI_LONG_INT, long_int);
128 CREATE_MPI_DATATYPE(MPI_DOUBLE_INT, double_int);
129 CREATE_MPI_DATATYPE(MPI_SHORT_INT, short_int);
130 CREATE_MPI_DATATYPE(MPI_2INT, int_int);
131 CREATE_MPI_DATATYPE(MPI_2FLOAT, float_float);
132 CREATE_MPI_DATATYPE(MPI_2DOUBLE, double_double);
133 CREATE_MPI_DATATYPE(MPI_2LONG, long_long);
135 CREATE_MPI_DATATYPE(MPI_REAL, float);
136 CREATE_MPI_DATATYPE(MPI_REAL4, float);
137 CREATE_MPI_DATATYPE(MPI_REAL8, float);
138 CREATE_MPI_DATATYPE(MPI_REAL16, double);
139 CREATE_MPI_DATATYPE_NULL(MPI_COMPLEX8);
140 CREATE_MPI_DATATYPE_NULL(MPI_COMPLEX16);
141 CREATE_MPI_DATATYPE_NULL(MPI_COMPLEX32);
142 CREATE_MPI_DATATYPE(MPI_INTEGER1, int);
143 CREATE_MPI_DATATYPE(MPI_INTEGER2, int16_t);
144 CREATE_MPI_DATATYPE(MPI_INTEGER4, int32_t);
145 CREATE_MPI_DATATYPE(MPI_INTEGER8, int64_t);
146 CREATE_MPI_DATATYPE(MPI_INTEGER16, integer128_t);
148 CREATE_MPI_DATATYPE(MPI_LONG_DOUBLE_INT, long_double_int);
150 CREATE_MPI_DATATYPE_NULL(MPI_UB);
151 CREATE_MPI_DATATYPE_NULL(MPI_LB);
152 CREATE_MPI_DATATYPE(MPI_PACKED, char);
154 CREATE_MPI_DATATYPE(MPI_PTR, void*);
156 /** Check if the datatype is usable for communications */
157 int is_datatype_valid(MPI_Datatype datatype) {
158 return datatype != MPI_DATATYPE_NULL && (datatype->flags & DT_FLAG_COMMITED);
161 size_t smpi_datatype_size(MPI_Datatype datatype)
163 return datatype->size;
166 MPI_Aint smpi_datatype_lb(MPI_Datatype datatype)
171 MPI_Aint smpi_datatype_ub(MPI_Datatype datatype)
176 int smpi_datatype_dup(MPI_Datatype datatype, MPI_Datatype* new_t)
179 *new_t= xbt_new(s_smpi_mpi_datatype_t,1);
180 memcpy(*new_t, datatype, sizeof(s_smpi_mpi_datatype_t));
182 (*new_t)->flags &= ~DT_FLAG_PREDEFINED;
183 if (datatype->sizeof_substruct){
184 (*new_t)->substruct=xbt_malloc(datatype->sizeof_substruct);
185 memcpy((*new_t)->substruct, datatype->substruct, datatype->sizeof_substruct);
188 (*new_t)->name = xbt_strdup(datatype->name);
189 if(datatype->attributes !=NULL){
190 (*new_t)->attributes=xbt_dict_new();
191 xbt_dict_cursor_t cursor = NULL;
196 xbt_dict_foreach(datatype->attributes, cursor, key, value_in){
197 smpi_type_key_elem elem =
198 static_cast<smpi_type_key_elem>(xbt_dict_get_or_null_ext(smpi_type_keyvals, (const char*)key, sizeof(int)));
199 if(elem && elem->copy_fn!=MPI_NULL_COPY_FN){
200 ret = elem->copy_fn(datatype, *key, NULL, value_in, &value_out, &flag );
201 if(ret!=MPI_SUCCESS){
202 smpi_datatype_unuse(*new_t);
203 *new_t=MPI_DATATYPE_NULL;
204 xbt_dict_cursor_free(&cursor);
208 xbt_dict_set_ext((*new_t)->attributes, (const char*)key, sizeof(int),value_out, NULL);
215 int smpi_datatype_extent(MPI_Datatype datatype, MPI_Aint * lb, MPI_Aint * extent)
217 if(datatype == MPI_DATATYPE_NULL){
223 *extent = datatype->ub - datatype->lb;
227 MPI_Aint smpi_datatype_get_extent(MPI_Datatype datatype){
228 if(datatype == MPI_DATATYPE_NULL){
231 return datatype->ub - datatype->lb;
234 void smpi_datatype_get_name(MPI_Datatype datatype, char* name, int* length){
235 *length = strlen(datatype->name);
236 strcpy(name, datatype->name);
239 void smpi_datatype_set_name(MPI_Datatype datatype, char* name){
240 if(datatype->name!=NULL && !(datatype->flags & DT_FLAG_PREDEFINED))
241 xbt_free(datatype->name);
242 datatype->name = xbt_strdup(name);;
245 int smpi_datatype_copy(void *sendbuf, int sendcount, MPI_Datatype sendtype,
246 void *recvbuf, int recvcount, MPI_Datatype recvtype)
249 if(smpi_privatize_global_variables){
250 smpi_switch_data_segment(smpi_process_index());
252 /* First check if we really have something to do */
253 if (recvcount > 0 && recvbuf != sendbuf) {
254 /* FIXME: treat packed cases */
255 sendcount *= smpi_datatype_size(sendtype);
256 recvcount *= smpi_datatype_size(recvtype);
257 count = sendcount < recvcount ? sendcount : recvcount;
259 if(sendtype->sizeof_substruct == 0 && recvtype->sizeof_substruct == 0) {
260 if(!smpi_process_get_replaying()) memcpy(recvbuf, sendbuf, count);
262 else if (sendtype->sizeof_substruct == 0)
264 s_smpi_subtype_t *subtype = static_cast<s_smpi_subtype_t*>(recvtype->substruct);
265 subtype->unserialize( sendbuf, recvbuf, recvcount/smpi_datatype_size(recvtype), subtype, MPI_REPLACE);
267 else if (recvtype->sizeof_substruct == 0)
269 s_smpi_subtype_t *subtype = static_cast<s_smpi_subtype_t*>(sendtype->substruct);
270 subtype->serialize(sendbuf, recvbuf, sendcount/smpi_datatype_size(sendtype), subtype);
272 s_smpi_subtype_t *subtype = static_cast<s_smpi_subtype_t*>(sendtype->substruct);
274 void * buf_tmp = xbt_malloc(count);
276 subtype->serialize( sendbuf, buf_tmp,count/smpi_datatype_size(sendtype), subtype);
277 subtype = static_cast<s_smpi_subtype_t*>(recvtype->substruct);
278 subtype->unserialize( buf_tmp, recvbuf,count/smpi_datatype_size(recvtype), subtype, MPI_REPLACE);
284 return sendcount > recvcount ? MPI_ERR_TRUNCATE : MPI_SUCCESS;
288 * Copies noncontiguous data into contiguous memory.
289 * @param contiguous_vector - output vector
290 * @param noncontiguous_vector - input vector
291 * @param type - pointer contening :
292 * - stride - stride of between noncontiguous data
293 * - block_length - the width or height of blocked matrix
294 * - count - the number of rows of matrix
296 void serialize_vector( const void *noncontiguous_vector, void *contiguous_vector, int count, void *type)
298 s_smpi_mpi_vector_t* type_c = (s_smpi_mpi_vector_t*)type;
300 char* contiguous_vector_char = (char*)contiguous_vector;
301 char* noncontiguous_vector_char = (char*)noncontiguous_vector;
303 for (i = 0; i < type_c->block_count * count; i++) {
304 if (type_c->old_type->sizeof_substruct == 0)
305 memcpy(contiguous_vector_char, noncontiguous_vector_char, type_c->block_length * type_c->size_oldtype);
307 ((s_smpi_subtype_t*)type_c->old_type->substruct)->serialize( noncontiguous_vector_char,
308 contiguous_vector_char,
309 type_c->block_length, type_c->old_type->substruct);
311 contiguous_vector_char += type_c->block_length*type_c->size_oldtype;
312 if((i+1)%type_c->block_count ==0)
313 noncontiguous_vector_char += type_c->block_length*smpi_datatype_get_extent(type_c->old_type);
315 noncontiguous_vector_char += type_c->block_stride*smpi_datatype_get_extent(type_c->old_type);
320 * Copies contiguous data into noncontiguous memory.
321 * @param noncontiguous_vector - output vector
322 * @param contiguous_vector - input vector
323 * @param type - pointer contening :
324 * - stride - stride of between noncontiguous data
325 * - block_length - the width or height of blocked matrix
326 * - count - the number of rows of matrix
328 void unserialize_vector( const void *contiguous_vector, void *noncontiguous_vector, int count, void *type, MPI_Op op)
330 s_smpi_mpi_vector_t* type_c = (s_smpi_mpi_vector_t*)type;
333 char* contiguous_vector_char = (char*)contiguous_vector;
334 char* noncontiguous_vector_char = (char*)noncontiguous_vector;
336 for (i = 0; i < type_c->block_count * count; i++) {
337 if (type_c->old_type->sizeof_substruct == 0)
338 smpi_op_apply(op, contiguous_vector_char, noncontiguous_vector_char, &type_c->block_length,
340 /* memcpy(noncontiguous_vector_char,
341 contiguous_vector_char, type_c->block_length * type_c->size_oldtype);*/
343 ((s_smpi_subtype_t*)type_c->old_type->substruct)->unserialize(contiguous_vector_char, noncontiguous_vector_char,
344 type_c->block_length,type_c->old_type->substruct,
346 contiguous_vector_char += type_c->block_length*type_c->size_oldtype;
347 if((i+1)%type_c->block_count ==0)
348 noncontiguous_vector_char += type_c->block_length*smpi_datatype_get_extent(type_c->old_type);
350 noncontiguous_vector_char += type_c->block_stride*smpi_datatype_get_extent(type_c->old_type);
354 /* Create a Sub type vector to be able to serialize and unserialize it the structure s_smpi_mpi_vector_t is derived
355 * from s_smpi_subtype which required the functions unserialize and serialize */
356 s_smpi_mpi_vector_t* smpi_datatype_vector_create( int block_stride, int block_length, int block_count,
357 MPI_Datatype old_type, int size_oldtype){
358 s_smpi_mpi_vector_t *new_t= xbt_new(s_smpi_mpi_vector_t,1);
359 new_t->base.serialize = &serialize_vector;
360 new_t->base.unserialize = &unserialize_vector;
361 new_t->base.subtype_free = &free_vector;
362 new_t->base.subtype_use = &use_vector;
363 new_t->block_stride = block_stride;
364 new_t->block_length = block_length;
365 new_t->block_count = block_count;
366 smpi_datatype_use(old_type);
367 new_t->old_type = old_type;
368 new_t->size_oldtype = size_oldtype;
372 void smpi_datatype_create(MPI_Datatype* new_type, int size,int lb, int ub, int sizeof_substruct, void *struct_type,
374 MPI_Datatype new_t= xbt_new(s_smpi_mpi_datatype_t,1);
377 new_t->sizeof_substruct = size>0? sizeof_substruct:0;
380 new_t->flags = flags;
381 new_t->substruct = struct_type;
383 new_t->attributes=NULL;
388 MC_ignore(&(new_t->in_use), sizeof(new_t->in_use));
392 void smpi_datatype_free(MPI_Datatype* type){
393 xbt_assert((*type)->in_use >= 0);
395 if((*type)->flags & DT_FLAG_PREDEFINED)return;
397 //if still used, mark for deletion
398 if((*type)->in_use!=0){
399 (*type)->flags |=DT_FLAG_DESTROYED;
403 if((*type)->attributes !=NULL){
404 xbt_dict_cursor_t cursor = NULL;
408 xbt_dict_foreach((*type)->attributes, cursor, key, value){
409 smpi_type_key_elem elem =
410 static_cast<smpi_type_key_elem>(xbt_dict_get_or_null_ext(smpi_type_keyvals, (const char*)key, sizeof(int)));
411 if(elem && elem->delete_fn)
412 elem->delete_fn(*type,*key, value, &flag);
414 xbt_dict_free(&(*type)->attributes);
417 if ((*type)->sizeof_substruct != 0){
418 //((s_smpi_subtype_t *)(*type)->substruct)->subtype_free(type);
419 xbt_free((*type)->substruct);
421 xbt_free((*type)->name);
423 *type = MPI_DATATYPE_NULL;
426 void smpi_datatype_use(MPI_Datatype type){
428 if(type)type->in_use++;
430 if(type->sizeof_substruct!=0){
431 ((s_smpi_subtype_t *)(type)->substruct)->subtype_use(&type);
435 MC_ignore(&(type->in_use), sizeof(type->in_use));
439 void smpi_datatype_unuse(MPI_Datatype type){
440 if (type->in_use > 0)
443 if(type->sizeof_substruct!=0){
444 ((s_smpi_subtype_t *)(type)->substruct)->subtype_free(&type);
447 if(type && type->in_use == 0){
448 smpi_datatype_free(&type);
452 MC_ignore(&(type->in_use), sizeof(type->in_use));
456 /*Contiguous Implementation*/
458 /* Copies noncontiguous data into contiguous memory.
459 * @param contiguous_hvector - output hvector
460 * @param noncontiguous_hvector - input hvector
461 * @param type - pointer contening :
462 * - stride - stride of between noncontiguous data, in bytes
463 * - block_length - the width or height of blocked matrix
464 * - count - the number of rows of matrix
466 void serialize_contiguous( const void *noncontiguous_hvector, void *contiguous_hvector, int count, void *type)
468 s_smpi_mpi_contiguous_t* type_c = (s_smpi_mpi_contiguous_t*)type;
469 char* contiguous_vector_char = (char*)contiguous_hvector;
470 char* noncontiguous_vector_char = (char*)noncontiguous_hvector+type_c->lb;
471 memcpy(contiguous_vector_char, noncontiguous_vector_char, count* type_c->block_count * type_c->size_oldtype);
473 /* Copies contiguous data into noncontiguous memory.
474 * @param noncontiguous_vector - output hvector
475 * @param contiguous_vector - input hvector
476 * @param type - pointer contening :
477 * - stride - stride of between noncontiguous data, in bytes
478 * - block_length - the width or height of blocked matrix
479 * - count - the number of rows of matrix
481 void unserialize_contiguous(const void *contiguous_vector, void *noncontiguous_vector, int count, void *type, MPI_Op op)
483 s_smpi_mpi_contiguous_t* type_c = (s_smpi_mpi_contiguous_t*)type;
484 char* contiguous_vector_char = (char*)contiguous_vector;
485 char* noncontiguous_vector_char = (char*)noncontiguous_vector+type_c->lb;
486 int n= count* type_c->block_count;
487 smpi_op_apply(op, contiguous_vector_char, noncontiguous_vector_char, &n, &type_c->old_type);
488 /*memcpy(noncontiguous_vector_char, contiguous_vector_char, count* type_c->block_count * type_c->size_oldtype);*/
491 void free_contiguous(MPI_Datatype* d){
492 smpi_datatype_unuse(((s_smpi_mpi_contiguous_t *)(*d)->substruct)->old_type);
495 void use_contiguous(MPI_Datatype* d){
496 smpi_datatype_use(((s_smpi_mpi_contiguous_t *)(*d)->substruct)->old_type);
499 /* Create a Sub type contiguous to be able to serialize and unserialize it the structure s_smpi_mpi_contiguous_t is
500 * erived from s_smpi_subtype which required the functions unserialize and serialize */
501 s_smpi_mpi_contiguous_t* smpi_datatype_contiguous_create( MPI_Aint lb, int block_count, MPI_Datatype old_type,
503 if(block_count==0)return NULL;
504 s_smpi_mpi_contiguous_t *new_t= xbt_new(s_smpi_mpi_contiguous_t,1);
505 new_t->base.serialize = &serialize_contiguous;
506 new_t->base.unserialize = &unserialize_contiguous;
507 new_t->base.subtype_free = &free_contiguous;
508 new_t->base.subtype_use = &use_contiguous;
510 new_t->block_count = block_count;
511 new_t->old_type = old_type;
512 smpi_datatype_use(old_type);
513 new_t->size_oldtype = size_oldtype;
514 smpi_datatype_use(old_type);
518 int smpi_datatype_contiguous(int count, MPI_Datatype old_type, MPI_Datatype* new_type, MPI_Aint lb)
521 if(old_type->sizeof_substruct){
522 //handle this case as a hvector with stride equals to the extent of the datatype
523 return smpi_datatype_hvector(count, 1, smpi_datatype_get_extent(old_type), old_type, new_type);
526 s_smpi_mpi_contiguous_t* subtype = smpi_datatype_contiguous_create( lb, count, old_type,smpi_datatype_size(old_type));
528 smpi_datatype_create(new_type, count * smpi_datatype_size(old_type),lb,lb + count * smpi_datatype_size(old_type),
529 sizeof(s_smpi_mpi_contiguous_t),subtype, DT_FLAG_CONTIGUOUS);
534 int smpi_datatype_vector(int count, int blocklen, int stride, MPI_Datatype old_type, MPI_Datatype* new_type)
537 if (blocklen<0) return MPI_ERR_ARG;
541 lb=smpi_datatype_lb(old_type);
542 ub=((count-1)*stride+blocklen-1)*smpi_datatype_get_extent(old_type)+smpi_datatype_ub(old_type);
544 if(old_type->sizeof_substruct || stride != blocklen){
546 s_smpi_mpi_vector_t* subtype = smpi_datatype_vector_create(stride, blocklen, count, old_type,
547 smpi_datatype_size(old_type));
548 smpi_datatype_create(new_type, count * (blocklen) * smpi_datatype_size(old_type), lb, ub, sizeof(s_smpi_mpi_vector_t), subtype,
552 /* in this situation the data are contignous thus it's not required to serialize and unserialize it*/
553 smpi_datatype_create(new_type, count * blocklen * smpi_datatype_size(old_type), 0, ((count -1) * stride + blocklen)*
554 smpi_datatype_size(old_type), 0, NULL, DT_FLAG_VECTOR|DT_FLAG_CONTIGUOUS);
560 void free_vector(MPI_Datatype* d){
561 smpi_datatype_unuse(((s_smpi_mpi_indexed_t *)(*d)->substruct)->old_type);
564 void use_vector(MPI_Datatype* d){
565 smpi_datatype_use(((s_smpi_mpi_indexed_t *)(*d)->substruct)->old_type);
568 /* Hvector Implementation - Vector with stride in bytes */
570 /* Copies noncontiguous data into contiguous memory.
571 * @param contiguous_hvector - output hvector
572 * @param noncontiguous_hvector - input hvector
573 * @param type - pointer contening :
574 * - stride - stride of between noncontiguous data, in bytes
575 * - block_length - the width or height of blocked matrix
576 * - count - the number of rows of matrix
578 void serialize_hvector( const void *noncontiguous_hvector, void *contiguous_hvector, int count, void *type)
580 s_smpi_mpi_hvector_t* type_c = (s_smpi_mpi_hvector_t*)type;
582 char* contiguous_vector_char = (char*)contiguous_hvector;
583 char* noncontiguous_vector_char = (char*)noncontiguous_hvector;
585 for (i = 0; i < type_c->block_count * count; i++) {
586 if (type_c->old_type->sizeof_substruct == 0)
587 memcpy(contiguous_vector_char, noncontiguous_vector_char, type_c->block_length * type_c->size_oldtype);
589 ((s_smpi_subtype_t*)type_c->old_type->substruct)->serialize( noncontiguous_vector_char,
590 contiguous_vector_char,
591 type_c->block_length, type_c->old_type->substruct);
593 contiguous_vector_char += type_c->block_length*type_c->size_oldtype;
594 if((i+1)%type_c->block_count ==0)
595 noncontiguous_vector_char += type_c->block_length*type_c->size_oldtype;
597 noncontiguous_vector_char += type_c->block_stride;
600 /* Copies contiguous data into noncontiguous memory.
601 * @param noncontiguous_vector - output hvector
602 * @param contiguous_vector - input hvector
603 * @param type - pointer contening :
604 * - stride - stride of between noncontiguous data, in bytes
605 * - block_length - the width or height of blocked matrix
606 * - count - the number of rows of matrix
608 void unserialize_hvector( const void *contiguous_vector, void *noncontiguous_vector, int count, void *type, MPI_Op op)
610 s_smpi_mpi_hvector_t* type_c = (s_smpi_mpi_hvector_t*)type;
613 char* contiguous_vector_char = (char*)contiguous_vector;
614 char* noncontiguous_vector_char = (char*)noncontiguous_vector;
616 for (i = 0; i < type_c->block_count * count; i++) {
617 if (type_c->old_type->sizeof_substruct == 0)
618 smpi_op_apply(op, contiguous_vector_char, noncontiguous_vector_char, &type_c->block_length, &type_c->old_type);
619 /*memcpy(noncontiguous_vector_char,
620 contiguous_vector_char, type_c->block_length * type_c->size_oldtype);*/
622 ((s_smpi_subtype_t*)type_c->old_type->substruct)->unserialize( contiguous_vector_char, noncontiguous_vector_char,
623 type_c->block_length, type_c->old_type->substruct,
625 contiguous_vector_char += type_c->block_length*type_c->size_oldtype;
626 if((i+1)%type_c->block_count ==0)
627 noncontiguous_vector_char += type_c->block_length*type_c->size_oldtype;
629 noncontiguous_vector_char += type_c->block_stride;
633 /* Create a Sub type vector to be able to serialize and unserialize it the structure s_smpi_mpi_vector_t is derived
634 * from s_smpi_subtype which required the functions unserialize and serialize
637 s_smpi_mpi_hvector_t* smpi_datatype_hvector_create( MPI_Aint block_stride, int block_length, int block_count,
638 MPI_Datatype old_type, int size_oldtype){
639 s_smpi_mpi_hvector_t *new_t= xbt_new(s_smpi_mpi_hvector_t,1);
640 new_t->base.serialize = &serialize_hvector;
641 new_t->base.unserialize = &unserialize_hvector;
642 new_t->base.subtype_free = &free_hvector;
643 new_t->base.subtype_use = &use_hvector;
644 new_t->block_stride = block_stride;
645 new_t->block_length = block_length;
646 new_t->block_count = block_count;
647 new_t->old_type = old_type;
648 new_t->size_oldtype = size_oldtype;
649 smpi_datatype_use(old_type);
653 //do nothing for vector types
654 void free_hvector(MPI_Datatype* d){
655 smpi_datatype_unuse(((s_smpi_mpi_hvector_t *)(*d)->substruct)->old_type);
658 void use_hvector(MPI_Datatype* d){
659 smpi_datatype_use(((s_smpi_mpi_hvector_t *)(*d)->substruct)->old_type);
662 int smpi_datatype_hvector(int count, int blocklen, MPI_Aint stride, MPI_Datatype old_type, MPI_Datatype* new_type)
665 if (blocklen<0) return MPI_ERR_ARG;
669 lb=smpi_datatype_lb(old_type);
670 ub=((count-1)*stride)+(blocklen-1)*smpi_datatype_get_extent(old_type)+smpi_datatype_ub(old_type);
672 if(old_type->sizeof_substruct || stride != blocklen*smpi_datatype_get_extent(old_type)){
673 s_smpi_mpi_hvector_t* subtype = smpi_datatype_hvector_create( stride, blocklen, count, old_type,
674 smpi_datatype_size(old_type));
676 smpi_datatype_create(new_type, count * blocklen * smpi_datatype_size(old_type), lb,ub, sizeof(s_smpi_mpi_hvector_t), subtype, DT_FLAG_VECTOR);
679 smpi_datatype_create(new_type, count * blocklen * smpi_datatype_size(old_type),0,count * blocklen *
680 smpi_datatype_size(old_type), 0, NULL, DT_FLAG_VECTOR|DT_FLAG_CONTIGUOUS);
686 /* Indexed Implementation */
688 /* Copies noncontiguous data into contiguous memory.
689 * @param contiguous_indexed - output indexed
690 * @param noncontiguous_indexed - input indexed
691 * @param type - pointer contening :
692 * - block_lengths - the width or height of blocked matrix
693 * - block_indices - indices of each data, in element
694 * - count - the number of rows of matrix
696 void serialize_indexed( const void *noncontiguous_indexed, void *contiguous_indexed, int count, void *type)
698 s_smpi_mpi_indexed_t* type_c = (s_smpi_mpi_indexed_t*)type;
700 char* contiguous_indexed_char = (char*)contiguous_indexed;
701 char* noncontiguous_indexed_char = (char*)noncontiguous_indexed+type_c->block_indices[0] * type_c->size_oldtype;
702 for(j=0; j<count;j++){
703 for (i = 0; i < type_c->block_count; i++) {
704 if (type_c->old_type->sizeof_substruct == 0)
705 memcpy(contiguous_indexed_char, noncontiguous_indexed_char, type_c->block_lengths[i] * type_c->size_oldtype);
707 ((s_smpi_subtype_t*)type_c->old_type->substruct)->serialize( noncontiguous_indexed_char,
708 contiguous_indexed_char,
709 type_c->block_lengths[i],
710 type_c->old_type->substruct);
712 contiguous_indexed_char += type_c->block_lengths[i]*type_c->size_oldtype;
713 if (i<type_c->block_count-1)
714 noncontiguous_indexed_char =
715 (char*)noncontiguous_indexed + type_c->block_indices[i+1]*smpi_datatype_get_extent(type_c->old_type);
717 noncontiguous_indexed_char += type_c->block_lengths[i]*smpi_datatype_get_extent(type_c->old_type);
719 noncontiguous_indexed=(void*)noncontiguous_indexed_char;
722 /* Copies contiguous data into noncontiguous memory.
723 * @param noncontiguous_indexed - output indexed
724 * @param contiguous_indexed - input indexed
725 * @param type - pointer contening :
726 * - block_lengths - the width or height of blocked matrix
727 * - block_indices - indices of each data, in element
728 * - count - the number of rows of matrix
730 void unserialize_indexed( const void *contiguous_indexed, void *noncontiguous_indexed, int count, void *type, MPI_Op op)
732 s_smpi_mpi_indexed_t* type_c = (s_smpi_mpi_indexed_t*)type;
734 char* contiguous_indexed_char = (char*)contiguous_indexed;
735 char* noncontiguous_indexed_char =
736 (char*)noncontiguous_indexed+type_c->block_indices[0]*smpi_datatype_get_extent(type_c->old_type);
737 for(j=0; j<count;j++){
738 for (i = 0; i < type_c->block_count; i++) {
739 if (type_c->old_type->sizeof_substruct == 0)
740 smpi_op_apply(op, contiguous_indexed_char, noncontiguous_indexed_char, &type_c->block_lengths[i],
742 /*memcpy(noncontiguous_indexed_char ,
743 contiguous_indexed_char, type_c->block_lengths[i] * type_c->size_oldtype);*/
745 ((s_smpi_subtype_t*)type_c->old_type->substruct)->unserialize( contiguous_indexed_char,
746 noncontiguous_indexed_char,
747 type_c->block_lengths[i],
748 type_c->old_type->substruct, op);
750 contiguous_indexed_char += type_c->block_lengths[i]*type_c->size_oldtype;
751 if (i<type_c->block_count-1)
752 noncontiguous_indexed_char =
753 (char*)noncontiguous_indexed + type_c->block_indices[i+1]*smpi_datatype_get_extent(type_c->old_type);
755 noncontiguous_indexed_char += type_c->block_lengths[i]*smpi_datatype_get_extent(type_c->old_type);
757 noncontiguous_indexed=(void*)noncontiguous_indexed_char;
761 void free_indexed(MPI_Datatype* type){
762 if((*type)->in_use==0){
763 xbt_free(((s_smpi_mpi_indexed_t *)(*type)->substruct)->block_lengths);
764 xbt_free(((s_smpi_mpi_indexed_t *)(*type)->substruct)->block_indices);
766 smpi_datatype_unuse(((s_smpi_mpi_indexed_t *)(*type)->substruct)->old_type);
769 void use_indexed(MPI_Datatype* type){
770 smpi_datatype_use(((s_smpi_mpi_indexed_t *)(*type)->substruct)->old_type);
774 /* Create a Sub type indexed to be able to serialize and unserialize it the structure s_smpi_mpi_indexed_t is derived
775 * from s_smpi_subtype which required the functions unserialize and serialize */
776 s_smpi_mpi_indexed_t* smpi_datatype_indexed_create( int* block_lengths, int* block_indices, int block_count,
777 MPI_Datatype old_type, int size_oldtype){
778 s_smpi_mpi_indexed_t *new_t= xbt_new(s_smpi_mpi_indexed_t,1);
779 new_t->base.serialize = &serialize_indexed;
780 new_t->base.unserialize = &unserialize_indexed;
781 new_t->base.subtype_free = &free_indexed;
782 new_t->base.subtype_use = &use_indexed;
783 new_t->block_lengths= xbt_new(int, block_count);
784 new_t->block_indices= xbt_new(int, block_count);
786 for(i=0;i<block_count;i++){
787 new_t->block_lengths[i]=block_lengths[i];
788 new_t->block_indices[i]=block_indices[i];
790 new_t->block_count = block_count;
791 smpi_datatype_use(old_type);
792 new_t->old_type = old_type;
793 new_t->size_oldtype = size_oldtype;
797 int smpi_datatype_indexed(int count, int* blocklens, int* indices, MPI_Datatype old_type, MPI_Datatype* new_type)
806 lb=indices[0]*smpi_datatype_get_extent(old_type);
807 ub=indices[0]*smpi_datatype_get_extent(old_type) + blocklens[0]*smpi_datatype_ub(old_type);
810 for(i=0; i< count; i++){
813 size += blocklens[i];
815 if(indices[i]*smpi_datatype_get_extent(old_type)+smpi_datatype_lb(old_type)<lb)
816 lb = indices[i]*smpi_datatype_get_extent(old_type)+smpi_datatype_lb(old_type);
817 if(indices[i]*smpi_datatype_get_extent(old_type)+blocklens[i]*smpi_datatype_ub(old_type)>ub)
818 ub = indices[i]*smpi_datatype_get_extent(old_type)+blocklens[i]*smpi_datatype_ub(old_type);
820 if ( (i< count -1) && (indices[i]+blocklens[i] != indices[i+1]) )contiguous=0;
822 if (old_type->sizeof_substruct != 0)
826 s_smpi_mpi_indexed_t* subtype = smpi_datatype_indexed_create( blocklens, indices, count, old_type,
827 smpi_datatype_size(old_type));
828 smpi_datatype_create(new_type, size * smpi_datatype_size(old_type),lb,ub,sizeof(s_smpi_mpi_indexed_t), subtype, DT_FLAG_DATA);
830 s_smpi_mpi_contiguous_t* subtype = smpi_datatype_contiguous_create( lb, size, old_type,
831 smpi_datatype_size(old_type));
832 smpi_datatype_create(new_type, size * smpi_datatype_size(old_type), lb, ub, sizeof(s_smpi_mpi_contiguous_t), subtype,
833 DT_FLAG_DATA|DT_FLAG_CONTIGUOUS);
838 /* Hindexed Implementation - Indexed with indices in bytes */
840 /* Copies noncontiguous data into contiguous memory.
841 * @param contiguous_hindexed - output hindexed
842 * @param noncontiguous_hindexed - input hindexed
843 * @param type - pointer contening :
844 * - block_lengths - the width or height of blocked matrix
845 * - block_indices - indices of each data, in bytes
846 * - count - the number of rows of matrix
848 void serialize_hindexed( const void *noncontiguous_hindexed, void *contiguous_hindexed, int count, void *type)
850 s_smpi_mpi_hindexed_t* type_c = (s_smpi_mpi_hindexed_t*)type;
852 char* contiguous_hindexed_char = (char*)contiguous_hindexed;
853 char* noncontiguous_hindexed_char = (char*)noncontiguous_hindexed+ type_c->block_indices[0];
854 for(j=0; j<count;j++){
855 for (i = 0; i < type_c->block_count; i++) {
856 if (type_c->old_type->sizeof_substruct == 0)
857 memcpy(contiguous_hindexed_char, noncontiguous_hindexed_char, type_c->block_lengths[i] * type_c->size_oldtype);
859 ((s_smpi_subtype_t*)type_c->old_type->substruct)->serialize( noncontiguous_hindexed_char,
860 contiguous_hindexed_char,
861 type_c->block_lengths[i],
862 type_c->old_type->substruct);
864 contiguous_hindexed_char += type_c->block_lengths[i]*type_c->size_oldtype;
865 if (i<type_c->block_count-1)
866 noncontiguous_hindexed_char = (char*)noncontiguous_hindexed + type_c->block_indices[i+1];
868 noncontiguous_hindexed_char += type_c->block_lengths[i]*smpi_datatype_get_extent(type_c->old_type);
870 noncontiguous_hindexed=(void*)noncontiguous_hindexed_char;
873 /* Copies contiguous data into noncontiguous memory.
874 * @param noncontiguous_hindexed - output hindexed
875 * @param contiguous_hindexed - input hindexed
876 * @param type - pointer contening :
877 * - block_lengths - the width or height of blocked matrix
878 * - block_indices - indices of each data, in bytes
879 * - count - the number of rows of matrix
881 void unserialize_hindexed( const void *contiguous_hindexed, void *noncontiguous_hindexed, int count, void *type,
884 s_smpi_mpi_hindexed_t* type_c = (s_smpi_mpi_hindexed_t*)type;
887 char* contiguous_hindexed_char = (char*)contiguous_hindexed;
888 char* noncontiguous_hindexed_char = (char*)noncontiguous_hindexed+ type_c->block_indices[0];
889 for(j=0; j<count;j++){
890 for (i = 0; i < type_c->block_count; i++) {
891 if (type_c->old_type->sizeof_substruct == 0)
892 smpi_op_apply(op, contiguous_hindexed_char, noncontiguous_hindexed_char, &type_c->block_lengths[i],
894 /*memcpy(noncontiguous_hindexed_char,contiguous_hindexed_char,type_c->block_lengths[i]*type_c->size_oldtype);*/
896 ((s_smpi_subtype_t*)type_c->old_type->substruct)->unserialize( contiguous_hindexed_char,
897 noncontiguous_hindexed_char,
898 type_c->block_lengths[i],
899 type_c->old_type->substruct, op);
901 contiguous_hindexed_char += type_c->block_lengths[i]*type_c->size_oldtype;
902 if (i<type_c->block_count-1)
903 noncontiguous_hindexed_char = (char*)noncontiguous_hindexed + type_c->block_indices[i+1];
905 noncontiguous_hindexed_char += type_c->block_lengths[i]*smpi_datatype_get_extent(type_c->old_type);
907 noncontiguous_hindexed=(void*)noncontiguous_hindexed_char;
911 void free_hindexed(MPI_Datatype* type){
912 if((*type)->in_use==0){
913 xbt_free(((s_smpi_mpi_hindexed_t *)(*type)->substruct)->block_lengths);
914 xbt_free(((s_smpi_mpi_hindexed_t *)(*type)->substruct)->block_indices);
916 smpi_datatype_unuse(((s_smpi_mpi_indexed_t *)(*type)->substruct)->old_type);
919 void use_hindexed(MPI_Datatype* type){
920 smpi_datatype_use(((s_smpi_mpi_hindexed_t *)(*type)->substruct)->old_type);
923 /* Create a Sub type hindexed to be able to serialize and unserialize it the structure s_smpi_mpi_hindexed_t is derived
924 * from s_smpi_subtype which required the functions unserialize and serialize
926 s_smpi_mpi_hindexed_t* smpi_datatype_hindexed_create( int* block_lengths, MPI_Aint* block_indices, int block_count,
927 MPI_Datatype old_type, int size_oldtype){
928 s_smpi_mpi_hindexed_t *new_t= xbt_new(s_smpi_mpi_hindexed_t,1);
929 new_t->base.serialize = &serialize_hindexed;
930 new_t->base.unserialize = &unserialize_hindexed;
931 new_t->base.subtype_free = &free_hindexed;
932 new_t->base.subtype_use = &use_hindexed;
933 new_t->block_lengths= xbt_new(int, block_count);
934 new_t->block_indices= xbt_new(MPI_Aint, block_count);
936 for(i=0;i<block_count;i++){
937 new_t->block_lengths[i]=block_lengths[i];
938 new_t->block_indices[i]=block_indices[i];
940 new_t->block_count = block_count;
941 new_t->old_type = old_type;
942 smpi_datatype_use(old_type);
943 new_t->size_oldtype = size_oldtype;
947 int smpi_datatype_hindexed(int count, int* blocklens, MPI_Aint* indices, MPI_Datatype old_type, MPI_Datatype* new_type)
956 lb=indices[0] + smpi_datatype_lb(old_type);
957 ub=indices[0] + blocklens[0]*smpi_datatype_ub(old_type);
959 for(i=0; i< count; i++){
962 size += blocklens[i];
964 if(indices[i]+smpi_datatype_lb(old_type)<lb) lb = indices[i]+smpi_datatype_lb(old_type);
965 if(indices[i]+blocklens[i]*smpi_datatype_ub(old_type)>ub) ub = indices[i]+blocklens[i]*smpi_datatype_ub(old_type);
967 if ( (i< count -1) && (indices[i]+blocklens[i]*static_cast<int>(smpi_datatype_size(old_type)) != indices[i+1]) )
970 if (old_type->sizeof_substruct != 0 || lb!=0)
974 s_smpi_mpi_hindexed_t* subtype = smpi_datatype_hindexed_create( blocklens, indices, count, old_type,
975 smpi_datatype_size(old_type));
976 smpi_datatype_create(new_type, size * smpi_datatype_size(old_type), lb, ub ,sizeof(s_smpi_mpi_hindexed_t), subtype, DT_FLAG_DATA);
978 s_smpi_mpi_contiguous_t* subtype = smpi_datatype_contiguous_create(lb,size, old_type, smpi_datatype_size(old_type));
979 smpi_datatype_create(new_type, size * smpi_datatype_size(old_type), 0,size * smpi_datatype_size(old_type),
980 1, subtype, DT_FLAG_DATA|DT_FLAG_CONTIGUOUS);
986 /* struct Implementation - Indexed with indices in bytes */
988 /* Copies noncontiguous data into contiguous memory.
989 * @param contiguous_struct - output struct
990 * @param noncontiguous_struct - input struct
991 * @param type - pointer contening :
992 * - stride - stride of between noncontiguous data
993 * - block_length - the width or height of blocked matrix
994 * - count - the number of rows of matrix
996 void serialize_struct( const void *noncontiguous_struct, void *contiguous_struct, int count, void *type)
998 s_smpi_mpi_struct_t* type_c = (s_smpi_mpi_struct_t*)type;
1000 char* contiguous_struct_char = (char*)contiguous_struct;
1001 char* noncontiguous_struct_char = (char*)noncontiguous_struct+ type_c->block_indices[0];
1002 for(j=0; j<count;j++){
1003 for (i = 0; i < type_c->block_count; i++) {
1004 if (type_c->old_types[i]->sizeof_substruct == 0)
1005 memcpy(contiguous_struct_char, noncontiguous_struct_char,
1006 type_c->block_lengths[i] * smpi_datatype_size(type_c->old_types[i]));
1008 ((s_smpi_subtype_t*)type_c->old_types[i]->substruct)->serialize( noncontiguous_struct_char,
1009 contiguous_struct_char,
1010 type_c->block_lengths[i],
1011 type_c->old_types[i]->substruct);
1014 contiguous_struct_char += type_c->block_lengths[i]*smpi_datatype_size(type_c->old_types[i]);
1015 if (i<type_c->block_count-1)
1016 noncontiguous_struct_char = (char*)noncontiguous_struct + type_c->block_indices[i+1];
1017 else //let's hope this is MPI_UB ?
1018 noncontiguous_struct_char += type_c->block_lengths[i]*smpi_datatype_get_extent(type_c->old_types[i]);
1020 noncontiguous_struct=(void*)noncontiguous_struct_char;
1024 /* Copies contiguous data into noncontiguous memory.
1025 * @param noncontiguous_struct - output struct
1026 * @param contiguous_struct - input struct
1027 * @param type - pointer contening :
1028 * - stride - stride of between noncontiguous data
1029 * - block_length - the width or height of blocked matrix
1030 * - count - the number of rows of matrix
1032 void unserialize_struct( const void *contiguous_struct, void *noncontiguous_struct, int count, void *type, MPI_Op op)
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]->sizeof_substruct == 0)
1042 smpi_op_apply(op, contiguous_struct_char, noncontiguous_struct_char, &type_c->block_lengths[i],
1043 & type_c->old_types[i]);
1044 /*memcpy(noncontiguous_struct_char,
1045 contiguous_struct_char, type_c->block_lengths[i] * smpi_datatype_size(type_c->old_types[i]));*/
1047 ((s_smpi_subtype_t*)type_c->old_types[i]->substruct)->unserialize( contiguous_struct_char,
1048 noncontiguous_struct_char,
1049 type_c->block_lengths[i],
1050 type_c->old_types[i]->substruct, op);
1052 contiguous_struct_char += type_c->block_lengths[i]*smpi_datatype_size(type_c->old_types[i]);
1053 if (i<type_c->block_count-1)
1054 noncontiguous_struct_char = (char*)noncontiguous_struct + type_c->block_indices[i+1];
1056 noncontiguous_struct_char += type_c->block_lengths[i]*smpi_datatype_get_extent(type_c->old_types[i]);
1058 noncontiguous_struct=(void*)noncontiguous_struct_char;
1062 void free_struct(MPI_Datatype* type){
1064 for (i = 0; i < ((s_smpi_mpi_struct_t *)(*type)->substruct)->block_count; i++)
1065 smpi_datatype_unuse(((s_smpi_mpi_struct_t *)(*type)->substruct)->old_types[i]);
1066 if((*type)->in_use==0){
1067 xbt_free(((s_smpi_mpi_struct_t *)(*type)->substruct)->block_lengths);
1068 xbt_free(((s_smpi_mpi_struct_t *)(*type)->substruct)->block_indices);
1069 xbt_free(((s_smpi_mpi_struct_t *)(*type)->substruct)->old_types);
1073 void use_struct(MPI_Datatype* type){
1075 for (i = 0; i < ((s_smpi_mpi_struct_t *)(*type)->substruct)->block_count; i++)
1076 smpi_datatype_use(((s_smpi_mpi_struct_t *)(*type)->substruct)->old_types[i]);
1079 /* Create a Sub type struct to be able to serialize and unserialize it the structure s_smpi_mpi_struct_t is derived
1080 * from s_smpi_subtype which required the functions unserialize and serialize
1082 s_smpi_mpi_struct_t* smpi_datatype_struct_create( int* block_lengths, MPI_Aint* block_indices, int block_count,
1083 MPI_Datatype* old_types){
1084 s_smpi_mpi_struct_t *new_t= xbt_new(s_smpi_mpi_struct_t,1);
1085 new_t->base.serialize = &serialize_struct;
1086 new_t->base.unserialize = &unserialize_struct;
1087 new_t->base.subtype_free = &free_struct;
1088 new_t->base.subtype_use = &use_struct;
1089 new_t->block_lengths= xbt_new(int, block_count);
1090 new_t->block_indices= xbt_new(MPI_Aint, block_count);
1091 new_t->old_types= xbt_new(MPI_Datatype, block_count);
1093 for(i=0;i<block_count;i++){
1094 new_t->block_lengths[i]=block_lengths[i];
1095 new_t->block_indices[i]=block_indices[i];
1096 new_t->old_types[i]=old_types[i];
1097 smpi_datatype_use(new_t->old_types[i]);
1099 new_t->block_count = block_count;
1103 int smpi_datatype_struct(int count, int* blocklens, MPI_Aint* indices, MPI_Datatype* old_types, MPI_Datatype* new_type)
1112 lb=indices[0] + smpi_datatype_lb(old_types[0]);
1113 ub=indices[0] + blocklens[0]*smpi_datatype_ub(old_types[0]);
1117 for(i=0; i< count; i++){
1120 if (old_types[i]->sizeof_substruct != 0)
1123 size += blocklens[i]*smpi_datatype_size(old_types[i]);
1124 if (old_types[i]==MPI_LB){
1128 if (old_types[i]==MPI_UB){
1133 if(!forced_lb && indices[i]+smpi_datatype_lb(old_types[i])<lb) lb = indices[i];
1134 if(!forced_ub && indices[i]+blocklens[i]*smpi_datatype_ub(old_types[i])>ub)
1135 ub = indices[i]+blocklens[i]*smpi_datatype_ub(old_types[i]);
1137 if ( (i< count -1) && (indices[i]+blocklens[i]*static_cast<int>(smpi_datatype_size(old_types[i])) != indices[i+1]) )
1142 s_smpi_mpi_struct_t* subtype = smpi_datatype_struct_create( blocklens, indices, count, old_types);
1144 smpi_datatype_create(new_type, size, lb, ub,sizeof(s_smpi_mpi_struct_t), subtype, DT_FLAG_DATA);
1146 s_smpi_mpi_contiguous_t* subtype = smpi_datatype_contiguous_create( lb, size, MPI_CHAR, 1);
1147 smpi_datatype_create(new_type, size, lb, ub,1, subtype, DT_FLAG_DATA|DT_FLAG_CONTIGUOUS);
1152 void smpi_datatype_commit(MPI_Datatype *datatype)
1154 (*datatype)->flags= ((*datatype)->flags | DT_FLAG_COMMITED);
1157 typedef struct s_smpi_mpi_op {
1158 MPI_User_function *func;
1162 #define MAX_OP(a, b) (b) = (a) < (b) ? (b) : (a)
1163 #define MIN_OP(a, b) (b) = (a) < (b) ? (a) : (b)
1164 #define SUM_OP(a, b) (b) += (a)
1165 #define PROD_OP(a, b) (b) *= (a)
1166 #define LAND_OP(a, b) (b) = (a) && (b)
1167 #define LOR_OP(a, b) (b) = (a) || (b)
1168 #define LXOR_OP(a, b) (b) = (!(a) && (b)) || ((a) && !(b))
1169 #define BAND_OP(a, b) (b) &= (a)
1170 #define BOR_OP(a, b) (b) |= (a)
1171 #define BXOR_OP(a, b) (b) ^= (a)
1172 #define MAXLOC_OP(a, b) (b) = (a.value) < (b.value) ? (b) : (a)
1173 #define MINLOC_OP(a, b) (b) = (a.value) < (b.value) ? (a) : (b)
1175 #define APPLY_FUNC(a, b, length, type, func) \
1178 type* x = (type*)(a); \
1179 type* y = (type*)(b); \
1180 for(i = 0; i < *(length); i++) { \
1185 static void max_func(void *a, void *b, int *length, MPI_Datatype * datatype)
1187 if (*datatype == MPI_CHAR) {
1188 APPLY_FUNC(a, b, length, char, MAX_OP);
1189 } else if (*datatype == MPI_SHORT) {
1190 APPLY_FUNC(a, b, length, short, MAX_OP);
1191 } else if (*datatype == MPI_INT) {
1192 APPLY_FUNC(a, b, length, int, MAX_OP);
1193 } else if (*datatype == MPI_LONG) {
1194 APPLY_FUNC(a, b, length, long, MAX_OP);
1195 } else if (*datatype == MPI_UNSIGNED_SHORT) {
1196 APPLY_FUNC(a, b, length, unsigned short, MAX_OP);
1197 } else if (*datatype == MPI_UNSIGNED) {
1198 APPLY_FUNC(a, b, length, unsigned int, MAX_OP);
1199 } else if (*datatype == MPI_UNSIGNED_LONG) {
1200 APPLY_FUNC(a, b, length, unsigned long, MAX_OP);
1201 } else if (*datatype == MPI_UNSIGNED_CHAR) {
1202 APPLY_FUNC(a, b, length, unsigned char, MAX_OP);
1203 } else if (*datatype == MPI_FLOAT) {
1204 APPLY_FUNC(a, b, length, float, MAX_OP);
1205 } else if (*datatype == MPI_DOUBLE) {
1206 APPLY_FUNC(a, b, length, double, MAX_OP);
1207 } else if (*datatype == MPI_LONG_DOUBLE) {
1208 APPLY_FUNC(a, b, length, long double, MAX_OP);
1212 static void min_func(void *a, void *b, int *length, MPI_Datatype * datatype)
1214 if (*datatype == MPI_CHAR) {
1215 APPLY_FUNC(a, b, length, char, MIN_OP);
1216 } else if (*datatype == MPI_SHORT) {
1217 APPLY_FUNC(a, b, length, short, MIN_OP);
1218 } else if (*datatype == MPI_INT) {
1219 APPLY_FUNC(a, b, length, int, MIN_OP);
1220 } else if (*datatype == MPI_LONG) {
1221 APPLY_FUNC(a, b, length, long, MIN_OP);
1222 } else if (*datatype == MPI_UNSIGNED_SHORT) {
1223 APPLY_FUNC(a, b, length, unsigned short, MIN_OP);
1224 } else if (*datatype == MPI_UNSIGNED) {
1225 APPLY_FUNC(a, b, length, unsigned int, MIN_OP);
1226 } else if (*datatype == MPI_UNSIGNED_LONG) {
1227 APPLY_FUNC(a, b, length, unsigned long, MIN_OP);
1228 } else if (*datatype == MPI_UNSIGNED_CHAR) {
1229 APPLY_FUNC(a, b, length, unsigned char, 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, MPI_Datatype * datatype)
1241 if (*datatype == MPI_CHAR) {
1242 APPLY_FUNC(a, b, length, char, SUM_OP);
1243 } else if (*datatype == MPI_SHORT) {
1244 APPLY_FUNC(a, b, length, short, SUM_OP);
1245 } else if (*datatype == MPI_INT) {
1246 APPLY_FUNC(a, b, length, int, SUM_OP);
1247 } else if (*datatype == MPI_LONG) {
1248 APPLY_FUNC(a, b, length, long, SUM_OP);
1249 } else if (*datatype == MPI_UNSIGNED_SHORT) {
1250 APPLY_FUNC(a, b, length, unsigned short, SUM_OP);
1251 } else if (*datatype == MPI_UNSIGNED) {
1252 APPLY_FUNC(a, b, length, unsigned int, SUM_OP);
1253 } else if (*datatype == MPI_UNSIGNED_LONG) {
1254 APPLY_FUNC(a, b, length, unsigned long, SUM_OP);
1255 } else if (*datatype == MPI_UNSIGNED_CHAR) {
1256 APPLY_FUNC(a, b, length, unsigned char, SUM_OP);
1257 } else if (*datatype == MPI_FLOAT) {
1258 APPLY_FUNC(a, b, length, float, SUM_OP);
1259 } else if (*datatype == MPI_DOUBLE) {
1260 APPLY_FUNC(a, b, length, double, SUM_OP);
1261 } else if (*datatype == MPI_LONG_DOUBLE) {
1262 APPLY_FUNC(a, b, length, long double, SUM_OP);
1263 } else if (*datatype == MPI_C_FLOAT_COMPLEX) {
1264 APPLY_FUNC(a, b, length, float _Complex, SUM_OP);
1265 } else if (*datatype == MPI_C_DOUBLE_COMPLEX) {
1266 APPLY_FUNC(a, b, length, double _Complex, SUM_OP);
1267 } else if (*datatype == MPI_C_LONG_DOUBLE_COMPLEX) {
1268 APPLY_FUNC(a, b, length, long double _Complex, SUM_OP);
1272 static void prod_func(void *a, void *b, int *length, 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_UNSIGNED_CHAR) {
1289 APPLY_FUNC(a, b, length, unsigned char, PROD_OP);
1290 } else if (*datatype == MPI_FLOAT) {
1291 APPLY_FUNC(a, b, length, float, PROD_OP);
1292 } else if (*datatype == MPI_DOUBLE) {
1293 APPLY_FUNC(a, b, length, double, PROD_OP);
1294 } else if (*datatype == MPI_LONG_DOUBLE) {
1295 APPLY_FUNC(a, b, length, long double, PROD_OP);
1296 } else if (*datatype == MPI_C_FLOAT_COMPLEX) {
1297 APPLY_FUNC(a, b, length, float _Complex, PROD_OP);
1298 } else if (*datatype == MPI_C_DOUBLE_COMPLEX) {
1299 APPLY_FUNC(a, b, length, double _Complex, PROD_OP);
1300 } else if (*datatype == MPI_C_LONG_DOUBLE_COMPLEX) {
1301 APPLY_FUNC(a, b, length, long double _Complex, PROD_OP);
1305 static void land_func(void *a, void *b, int *length, MPI_Datatype * datatype)
1307 if (*datatype == MPI_CHAR) {
1308 APPLY_FUNC(a, b, length, char, LAND_OP);
1309 } else if (*datatype == MPI_SHORT) {
1310 APPLY_FUNC(a, b, length, short, LAND_OP);
1311 } else if (*datatype == MPI_INT) {
1312 APPLY_FUNC(a, b, length, int, LAND_OP);
1313 } else if (*datatype == MPI_LONG) {
1314 APPLY_FUNC(a, b, length, long, LAND_OP);
1315 } else if (*datatype == MPI_UNSIGNED_SHORT) {
1316 APPLY_FUNC(a, b, length, unsigned short, LAND_OP);
1317 } else if (*datatype == MPI_UNSIGNED) {
1318 APPLY_FUNC(a, b, length, unsigned int, LAND_OP);
1319 } else if (*datatype == MPI_UNSIGNED_LONG) {
1320 APPLY_FUNC(a, b, length, unsigned long, LAND_OP);
1321 } else if (*datatype == MPI_UNSIGNED_CHAR) {
1322 APPLY_FUNC(a, b, length, unsigned char, LAND_OP);
1323 } else if (*datatype == MPI_C_BOOL) {
1324 APPLY_FUNC(a, b, length, bool, LAND_OP);
1328 static void lor_func(void *a, void *b, int *length, MPI_Datatype * datatype)
1330 if (*datatype == MPI_CHAR) {
1331 APPLY_FUNC(a, b, length, char, LOR_OP);
1332 } else if (*datatype == MPI_SHORT) {
1333 APPLY_FUNC(a, b, length, short, LOR_OP);
1334 } else if (*datatype == MPI_INT) {
1335 APPLY_FUNC(a, b, length, int, LOR_OP);
1336 } else if (*datatype == MPI_LONG) {
1337 APPLY_FUNC(a, b, length, long, LOR_OP);
1338 } else if (*datatype == MPI_UNSIGNED_SHORT) {
1339 APPLY_FUNC(a, b, length, unsigned short, LOR_OP);
1340 } else if (*datatype == MPI_UNSIGNED) {
1341 APPLY_FUNC(a, b, length, unsigned int, LOR_OP);
1342 } else if (*datatype == MPI_UNSIGNED_LONG) {
1343 APPLY_FUNC(a, b, length, unsigned long, LOR_OP);
1344 } else if (*datatype == MPI_UNSIGNED_CHAR) {
1345 APPLY_FUNC(a, b, length, unsigned char, LOR_OP);
1346 } else if (*datatype == MPI_C_BOOL) {
1347 APPLY_FUNC(a, b, length, bool, LOR_OP);
1351 static void lxor_func(void *a, void *b, int *length, MPI_Datatype * datatype)
1353 if (*datatype == MPI_CHAR) {
1354 APPLY_FUNC(a, b, length, char, LXOR_OP);
1355 } else if (*datatype == MPI_SHORT) {
1356 APPLY_FUNC(a, b, length, short, LXOR_OP);
1357 } else if (*datatype == MPI_INT) {
1358 APPLY_FUNC(a, b, length, int, LXOR_OP);
1359 } else if (*datatype == MPI_LONG) {
1360 APPLY_FUNC(a, b, length, long, LXOR_OP);
1361 } else if (*datatype == MPI_UNSIGNED_SHORT) {
1362 APPLY_FUNC(a, b, length, unsigned short, LXOR_OP);
1363 } else if (*datatype == MPI_UNSIGNED) {
1364 APPLY_FUNC(a, b, length, unsigned int, LXOR_OP);
1365 } else if (*datatype == MPI_UNSIGNED_LONG) {
1366 APPLY_FUNC(a, b, length, unsigned long, LXOR_OP);
1367 } else if (*datatype == MPI_UNSIGNED_CHAR) {
1368 APPLY_FUNC(a, b, length, unsigned char, LXOR_OP);
1369 } else if (*datatype == MPI_C_BOOL) {
1370 APPLY_FUNC(a, b, length, bool, LXOR_OP);
1374 static void band_func(void *a, void *b, int *length, MPI_Datatype * datatype)
1376 if (*datatype == MPI_CHAR) {
1377 APPLY_FUNC(a, b, length, char, BAND_OP);
1378 }else if (*datatype == MPI_SHORT) {
1379 APPLY_FUNC(a, b, length, short, BAND_OP);
1380 } else if (*datatype == MPI_INT) {
1381 APPLY_FUNC(a, b, length, int, BAND_OP);
1382 } else if (*datatype == MPI_LONG) {
1383 APPLY_FUNC(a, b, length, long, BAND_OP);
1384 } else if (*datatype == MPI_UNSIGNED_SHORT) {
1385 APPLY_FUNC(a, b, length, unsigned short, BAND_OP);
1386 } else if (*datatype == MPI_UNSIGNED) {
1387 APPLY_FUNC(a, b, length, unsigned int, BAND_OP);
1388 } else if (*datatype == MPI_UNSIGNED_LONG) {
1389 APPLY_FUNC(a, b, length, unsigned long, BAND_OP);
1390 } else if (*datatype == MPI_UNSIGNED_CHAR) {
1391 APPLY_FUNC(a, b, length, unsigned char, BAND_OP);
1392 } else if (*datatype == MPI_BYTE) {
1393 APPLY_FUNC(a, b, length, uint8_t, BAND_OP);
1397 static void bor_func(void *a, void *b, int *length, MPI_Datatype * datatype)
1399 if (*datatype == MPI_CHAR) {
1400 APPLY_FUNC(a, b, length, char, BOR_OP);
1401 } else if (*datatype == MPI_SHORT) {
1402 APPLY_FUNC(a, b, length, short, BOR_OP);
1403 } else if (*datatype == MPI_INT) {
1404 APPLY_FUNC(a, b, length, int, BOR_OP);
1405 } else if (*datatype == MPI_LONG) {
1406 APPLY_FUNC(a, b, length, long, BOR_OP);
1407 } else if (*datatype == MPI_UNSIGNED_SHORT) {
1408 APPLY_FUNC(a, b, length, unsigned short, BOR_OP);
1409 } else if (*datatype == MPI_UNSIGNED) {
1410 APPLY_FUNC(a, b, length, unsigned int, BOR_OP);
1411 } else if (*datatype == MPI_UNSIGNED_LONG) {
1412 APPLY_FUNC(a, b, length, unsigned long, BOR_OP);
1413 } else if (*datatype == MPI_UNSIGNED_CHAR) {
1414 APPLY_FUNC(a, b, length, unsigned char, BOR_OP);
1415 } else if (*datatype == MPI_BYTE) {
1416 APPLY_FUNC(a, b, length, uint8_t, BOR_OP);
1420 static void bxor_func(void *a, void *b, int *length, MPI_Datatype * datatype)
1422 if (*datatype == MPI_CHAR) {
1423 APPLY_FUNC(a, b, length, char, BXOR_OP);
1424 } else if (*datatype == MPI_SHORT) {
1425 APPLY_FUNC(a, b, length, short, BXOR_OP);
1426 } else if (*datatype == MPI_INT) {
1427 APPLY_FUNC(a, b, length, int, BXOR_OP);
1428 } else if (*datatype == MPI_LONG) {
1429 APPLY_FUNC(a, b, length, long, BXOR_OP);
1430 } else if (*datatype == MPI_UNSIGNED_SHORT) {
1431 APPLY_FUNC(a, b, length, unsigned short, BXOR_OP);
1432 } else if (*datatype == MPI_UNSIGNED) {
1433 APPLY_FUNC(a, b, length, unsigned int, BXOR_OP);
1434 } else if (*datatype == MPI_UNSIGNED_LONG) {
1435 APPLY_FUNC(a, b, length, unsigned long, BXOR_OP);
1436 } else if (*datatype == MPI_UNSIGNED_CHAR) {
1437 APPLY_FUNC(a, b, length, unsigned char, BXOR_OP);
1438 } else if (*datatype == MPI_BYTE) {
1439 APPLY_FUNC(a, b, length, uint8_t, BXOR_OP);
1443 static void minloc_func(void *a, void *b, int *length, MPI_Datatype * datatype)
1445 if (*datatype == MPI_FLOAT_INT) {
1446 APPLY_FUNC(a, b, length, float_int, MINLOC_OP);
1447 } else if (*datatype == MPI_LONG_INT) {
1448 APPLY_FUNC(a, b, length, long_int, MINLOC_OP);
1449 } else if (*datatype == MPI_DOUBLE_INT) {
1450 APPLY_FUNC(a, b, length, double_int, MINLOC_OP);
1451 } else if (*datatype == MPI_SHORT_INT) {
1452 APPLY_FUNC(a, b, length, short_int, MINLOC_OP);
1453 } else if (*datatype == MPI_2LONG) {
1454 APPLY_FUNC(a, b, length, long_long, MINLOC_OP);
1455 } else if (*datatype == MPI_2INT) {
1456 APPLY_FUNC(a, b, length, int_int, MINLOC_OP);
1457 } else if (*datatype == MPI_LONG_DOUBLE_INT) {
1458 APPLY_FUNC(a, b, length, long_double_int, MINLOC_OP);
1459 } else if (*datatype == MPI_2FLOAT) {
1460 APPLY_FUNC(a, b, length, float_float, MINLOC_OP);
1461 } else if (*datatype == MPI_2DOUBLE) {
1462 APPLY_FUNC(a, b, length, double_double, MINLOC_OP);
1466 static void maxloc_func(void *a, void *b, int *length, MPI_Datatype * datatype)
1468 if (*datatype == MPI_FLOAT_INT) {
1469 APPLY_FUNC(a, b, length, float_int, MAXLOC_OP);
1470 } else if (*datatype == MPI_LONG_INT) {
1471 APPLY_FUNC(a, b, length, long_int, MAXLOC_OP);
1472 } else if (*datatype == MPI_DOUBLE_INT) {
1473 APPLY_FUNC(a, b, length, double_int, MAXLOC_OP);
1474 } else if (*datatype == MPI_SHORT_INT) {
1475 APPLY_FUNC(a, b, length, short_int, MAXLOC_OP);
1476 } else if (*datatype == MPI_2LONG) {
1477 APPLY_FUNC(a, b, length, long_long, MAXLOC_OP);
1478 } else if (*datatype == MPI_2INT) {
1479 APPLY_FUNC(a, b, length, int_int, MAXLOC_OP);
1480 } else if (*datatype == MPI_LONG_DOUBLE_INT) {
1481 APPLY_FUNC(a, b, length, long_double_int, MAXLOC_OP);
1482 } else if (*datatype == MPI_2FLOAT) {
1483 APPLY_FUNC(a, b, length, float_float, MAXLOC_OP);
1484 } else if (*datatype == MPI_2DOUBLE) {
1485 APPLY_FUNC(a, b, length, double_double, MAXLOC_OP);
1489 static void replace_func(void *a, void *b, int *length, MPI_Datatype * datatype)
1491 memcpy(b, a, *length * smpi_datatype_size(*datatype));
1494 #define CREATE_MPI_OP(name, func) \
1495 static s_smpi_mpi_op_t mpi_##name = { &(func) /* func */, true }; \
1496 MPI_Op name = &mpi_##name;
1498 CREATE_MPI_OP(MPI_MAX, max_func);
1499 CREATE_MPI_OP(MPI_MIN, min_func);
1500 CREATE_MPI_OP(MPI_SUM, sum_func);
1501 CREATE_MPI_OP(MPI_PROD, prod_func);
1502 CREATE_MPI_OP(MPI_LAND, land_func);
1503 CREATE_MPI_OP(MPI_LOR, lor_func);
1504 CREATE_MPI_OP(MPI_LXOR, lxor_func);
1505 CREATE_MPI_OP(MPI_BAND, band_func);
1506 CREATE_MPI_OP(MPI_BOR, bor_func);
1507 CREATE_MPI_OP(MPI_BXOR, bxor_func);
1508 CREATE_MPI_OP(MPI_MAXLOC, maxloc_func);
1509 CREATE_MPI_OP(MPI_MINLOC, minloc_func);
1510 CREATE_MPI_OP(MPI_REPLACE, replace_func);
1512 MPI_Op smpi_op_new(MPI_User_function * function, int commute)
1515 op = xbt_new(s_smpi_mpi_op_t, 1);
1516 op->func = function;
1517 op-> is_commute = commute;
1521 int smpi_op_is_commute(MPI_Op op)
1523 return (op==MPI_OP_NULL) ? 1 : op-> is_commute;
1526 void smpi_op_destroy(MPI_Op op)
1531 void smpi_op_apply(MPI_Op op, void *invec, void *inoutvec, int *len, MPI_Datatype * datatype)
1536 if(smpi_privatize_global_variables){//we need to switch as the called function may silently touch global variables
1537 XBT_DEBUG("Applying operation, switch to the right data frame ");
1538 smpi_switch_data_segment(smpi_process_index());
1541 if(!smpi_process_get_replaying())
1542 op->func(invec, inoutvec, len, datatype);
1545 int smpi_type_attr_delete(MPI_Datatype type, int keyval){
1546 smpi_type_key_elem elem =
1547 static_cast<smpi_type_key_elem>(xbt_dict_get_or_null_ext(smpi_type_keyvals, (const char*)&keyval, sizeof(int)));
1550 if(elem->delete_fn!=MPI_NULL_DELETE_FN){
1553 if(smpi_type_attr_get(type, keyval, &value, &flag)==MPI_SUCCESS){
1554 int ret = elem->delete_fn(type, keyval, value, &flag);
1555 if(ret!=MPI_SUCCESS) return ret;
1558 if(type->attributes==NULL)
1561 xbt_dict_remove_ext(type->attributes, (const char*)&keyval, sizeof(int));
1565 int smpi_type_attr_get(MPI_Datatype type, int keyval, void* attr_value, int* flag){
1566 smpi_type_key_elem elem =
1567 static_cast<smpi_type_key_elem>(xbt_dict_get_or_null_ext(smpi_type_keyvals, (const char*)&keyval, sizeof(int)));
1571 if(type->attributes==NULL){
1576 *(void**)attr_value = xbt_dict_get_ext(type->attributes, (const char*)&keyval, sizeof(int));
1586 int smpi_type_attr_put(MPI_Datatype type, int keyval, void* attr_value){
1587 if(!smpi_type_keyvals)
1588 smpi_type_keyvals = xbt_dict_new();
1589 smpi_type_key_elem elem =
1590 static_cast<smpi_type_key_elem>(xbt_dict_get_or_null_ext(smpi_type_keyvals, (const char*)&keyval, sizeof(int)));
1595 smpi_type_attr_get(type, keyval, &value, &flag);
1596 if(flag && elem->delete_fn!=MPI_NULL_DELETE_FN){
1597 int ret = elem->delete_fn(type, keyval, value, &flag);
1598 if(ret!=MPI_SUCCESS) return ret;
1600 if(type->attributes==NULL)
1601 type->attributes=xbt_dict_new();
1603 xbt_dict_set_ext(type->attributes, (const char*)&keyval, sizeof(int), attr_value, NULL);
1607 int smpi_type_keyval_create(MPI_Type_copy_attr_function* copy_fn, MPI_Type_delete_attr_function* delete_fn, int* keyval,
1609 if(!smpi_type_keyvals)
1610 smpi_type_keyvals = xbt_dict_new();
1612 smpi_type_key_elem value = (smpi_type_key_elem) xbt_new0(s_smpi_mpi_type_key_elem_t,1);
1614 value->copy_fn=copy_fn;
1615 value->delete_fn=delete_fn;
1617 *keyval = type_keyval_id;
1618 xbt_dict_set_ext(smpi_type_keyvals,(const char*)keyval, sizeof(int),(void*)value, NULL);
1623 int smpi_type_keyval_free(int* keyval){
1624 smpi_type_key_elem elem =
1625 static_cast<smpi_type_key_elem>(xbt_dict_get_or_null_ext(smpi_type_keyvals, (const char*)keyval, sizeof(int)));
1629 xbt_dict_remove_ext(smpi_type_keyvals, (const char*)keyval, sizeof(int));
1634 int smpi_mpi_pack(void* inbuf, int incount, MPI_Datatype type, void* outbuf, int outcount, int* position,MPI_Comm comm){
1635 size_t size = smpi_datatype_size(type);
1636 if (outcount - *position < incount*static_cast<int>(size))
1637 return MPI_ERR_BUFFER;
1638 smpi_datatype_copy(inbuf, incount, type, (char*)outbuf + *position, outcount, MPI_CHAR);
1639 *position += incount * size;
1643 int smpi_mpi_unpack(void* inbuf, int insize, int* position, void* outbuf, int outcount, MPI_Datatype type,MPI_Comm comm){
1644 int size = static_cast<int>(smpi_datatype_size(type));
1645 if (outcount*size> insize)
1646 return MPI_ERR_BUFFER;
1647 smpi_datatype_copy((char*)inbuf + *position, insize, MPI_CHAR, outbuf, outcount, type);
1648 *position += outcount * size;