1 /* smpi_mpi_dt.c -- MPI primitives to handle datatypes */
2 /* Copyright (c) 2009-2015. The SimGrid Team.
3 * All rights reserved. */
5 /* This program is free software; you can redistribute it and/or modify it
6 * under the terms of the license (GNU LGPL) which comes with this package. */
13 #include "smpi_mpi_dt_private.h"
15 #include "xbt/replay.h"
16 #include "simgrid/modelchecker.h"
18 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_mpi_dt, smpi, "Logging specific to SMPI (datatype)");
20 xbt_dict_t smpi_type_keyvals = nullptr;
21 int type_keyval_id=0;//avoid collisions
23 #define CREATE_MPI_DATATYPE(name, type) \
24 static s_smpi_mpi_datatype_t mpi_##name = { \
26 sizeof(type), /* size */ \
27 0, /*was 1 sizeof_substruct*/ \
29 sizeof(type), /* ub = lb + size */ \
30 DT_FLAG_BASIC, /* flags */ \
31 nullptr, /* attributes */ \
32 nullptr, /* pointer on extended struct*/ \
33 0 /* in_use counter */ \
35 const MPI_Datatype name = &mpi_##name;
37 #define CREATE_MPI_DATATYPE_NULL(name) \
38 static s_smpi_mpi_datatype_t mpi_##name = { \
41 0, /* was 1 sizeof_substruct*/ \
43 0, /* ub = lb + size */ \
44 DT_FLAG_BASIC, /* flags */ \
45 nullptr, /* attributes */ \
46 nullptr, /* pointer on extended struct*/ \
47 0 /* in_use counter */ \
49 const MPI_Datatype name = &mpi_##name;
51 //The following are datatypes for the MPI functions MPI_MAXLOC and MPI_MINLOC.
92 // Predefined data types
93 CREATE_MPI_DATATYPE(MPI_CHAR, char);
94 CREATE_MPI_DATATYPE(MPI_SHORT, short);
95 CREATE_MPI_DATATYPE(MPI_INT, int);
96 CREATE_MPI_DATATYPE(MPI_LONG, long);
97 CREATE_MPI_DATATYPE(MPI_LONG_LONG, long long);
98 CREATE_MPI_DATATYPE(MPI_SIGNED_CHAR, signed char);
99 CREATE_MPI_DATATYPE(MPI_UNSIGNED_CHAR, unsigned char);
100 CREATE_MPI_DATATYPE(MPI_UNSIGNED_SHORT, unsigned short);
101 CREATE_MPI_DATATYPE(MPI_UNSIGNED, unsigned int);
102 CREATE_MPI_DATATYPE(MPI_UNSIGNED_LONG, unsigned long);
103 CREATE_MPI_DATATYPE(MPI_UNSIGNED_LONG_LONG, unsigned long long);
104 CREATE_MPI_DATATYPE(MPI_FLOAT, float);
105 CREATE_MPI_DATATYPE(MPI_DOUBLE, double);
106 CREATE_MPI_DATATYPE(MPI_LONG_DOUBLE, long double);
107 CREATE_MPI_DATATYPE(MPI_WCHAR, wchar_t);
108 CREATE_MPI_DATATYPE(MPI_C_BOOL, bool);
109 CREATE_MPI_DATATYPE(MPI_BYTE, int8_t);
110 CREATE_MPI_DATATYPE(MPI_INT8_T, int8_t);
111 CREATE_MPI_DATATYPE(MPI_INT16_T, int16_t);
112 CREATE_MPI_DATATYPE(MPI_INT32_T, int32_t);
113 CREATE_MPI_DATATYPE(MPI_INT64_T, int64_t);
114 CREATE_MPI_DATATYPE(MPI_UINT8_T, uint8_t);
115 CREATE_MPI_DATATYPE(MPI_UINT16_T, uint16_t);
116 CREATE_MPI_DATATYPE(MPI_UINT32_T, uint32_t);
117 CREATE_MPI_DATATYPE(MPI_UINT64_T, uint64_t);
118 CREATE_MPI_DATATYPE(MPI_C_FLOAT_COMPLEX, float _Complex);
119 CREATE_MPI_DATATYPE(MPI_C_DOUBLE_COMPLEX, double _Complex);
120 CREATE_MPI_DATATYPE(MPI_C_LONG_DOUBLE_COMPLEX, long double _Complex);
121 CREATE_MPI_DATATYPE(MPI_AINT, MPI_Aint);
122 CREATE_MPI_DATATYPE(MPI_OFFSET, MPI_Offset);
124 CREATE_MPI_DATATYPE(MPI_FLOAT_INT, float_int);
125 CREATE_MPI_DATATYPE(MPI_LONG_INT, long_int);
126 CREATE_MPI_DATATYPE(MPI_DOUBLE_INT, double_int);
127 CREATE_MPI_DATATYPE(MPI_SHORT_INT, short_int);
128 CREATE_MPI_DATATYPE(MPI_2INT, int_int);
129 CREATE_MPI_DATATYPE(MPI_2FLOAT, float_float);
130 CREATE_MPI_DATATYPE(MPI_2DOUBLE, double_double);
131 CREATE_MPI_DATATYPE(MPI_2LONG, long_long);
133 CREATE_MPI_DATATYPE(MPI_REAL, float);
134 CREATE_MPI_DATATYPE(MPI_REAL4, float);
135 CREATE_MPI_DATATYPE(MPI_REAL8, float);
136 CREATE_MPI_DATATYPE(MPI_REAL16, double);
137 CREATE_MPI_DATATYPE_NULL(MPI_COMPLEX8);
138 CREATE_MPI_DATATYPE_NULL(MPI_COMPLEX16);
139 CREATE_MPI_DATATYPE_NULL(MPI_COMPLEX32);
140 CREATE_MPI_DATATYPE(MPI_INTEGER1, int);
141 CREATE_MPI_DATATYPE(MPI_INTEGER2, int16_t);
142 CREATE_MPI_DATATYPE(MPI_INTEGER4, int32_t);
143 CREATE_MPI_DATATYPE(MPI_INTEGER8, int64_t);
144 CREATE_MPI_DATATYPE(MPI_INTEGER16, integer128_t);
146 CREATE_MPI_DATATYPE(MPI_LONG_DOUBLE_INT, long_double_int);
148 CREATE_MPI_DATATYPE_NULL(MPI_UB);
149 CREATE_MPI_DATATYPE_NULL(MPI_LB);
150 CREATE_MPI_DATATYPE(MPI_PACKED, char);
152 CREATE_MPI_DATATYPE(MPI_PTR, void*);
154 /** Check if the datatype is usable for communications */
155 bool is_datatype_valid(MPI_Datatype datatype) {
156 return datatype != MPI_DATATYPE_NULL && ((datatype->flags & DT_FLAG_COMMITED) != 0);
159 size_t smpi_datatype_size(MPI_Datatype datatype)
161 return datatype->size;
164 MPI_Aint smpi_datatype_lb(MPI_Datatype datatype)
169 MPI_Aint smpi_datatype_ub(MPI_Datatype datatype)
174 int smpi_datatype_dup(MPI_Datatype datatype, MPI_Datatype* new_t)
177 *new_t= xbt_new(s_smpi_mpi_datatype_t,1);
178 memcpy(*new_t, datatype, sizeof(s_smpi_mpi_datatype_t));
180 (*new_t)->flags &= ~DT_FLAG_PREDEFINED;
181 if (datatype->sizeof_substruct){
182 (*new_t)->substruct=xbt_malloc(datatype->sizeof_substruct);
183 memcpy((*new_t)->substruct, datatype->substruct, datatype->sizeof_substruct);
186 (*new_t)->name = xbt_strdup(datatype->name);
187 if(datatype->attributes !=nullptr){
188 (*new_t)->attributes=xbt_dict_new();
189 xbt_dict_cursor_t cursor = nullptr;
194 xbt_dict_foreach(datatype->attributes, cursor, key, value_in){
195 smpi_type_key_elem elem =
196 static_cast<smpi_type_key_elem>(xbt_dict_get_or_null_ext(smpi_type_keyvals, reinterpret_cast<const char*>(key), sizeof(int)));
197 if(elem != nullptr && elem->copy_fn!=MPI_NULL_COPY_FN){
198 ret = elem->copy_fn(datatype, *key, nullptr, value_in, &value_out, &flag );
199 if(ret!=MPI_SUCCESS){
200 smpi_datatype_unuse(*new_t);
201 *new_t=MPI_DATATYPE_NULL;
202 xbt_dict_cursor_free(&cursor);
206 xbt_dict_set_ext((*new_t)->attributes, reinterpret_cast<const char*>(key), sizeof(int),value_out, nullptr);
213 int smpi_datatype_extent(MPI_Datatype datatype, MPI_Aint * lb, MPI_Aint * extent)
215 if(datatype == MPI_DATATYPE_NULL){
221 *extent = datatype->ub - datatype->lb;
225 MPI_Aint smpi_datatype_get_extent(MPI_Datatype datatype){
226 if(datatype == MPI_DATATYPE_NULL){
229 return datatype->ub - datatype->lb;
232 void smpi_datatype_get_name(MPI_Datatype datatype, char* name, int* length){
233 *length = strlen(datatype->name);
234 strncpy(name, datatype->name, *length+1);
237 void smpi_datatype_set_name(MPI_Datatype datatype, char* name){
238 if(datatype->name!=nullptr && (datatype->flags & DT_FLAG_PREDEFINED) == 0)
239 xbt_free(datatype->name);
240 datatype->name = xbt_strdup(name);
243 int smpi_datatype_copy(void *sendbuf, int sendcount, MPI_Datatype sendtype,
244 void *recvbuf, int recvcount, MPI_Datatype recvtype)
247 if(smpi_privatize_global_variables){
248 smpi_switch_data_segment(smpi_process_index());
250 /* First check if we really have something to do */
251 if (recvcount > 0 && recvbuf != sendbuf) {
252 /* FIXME: treat packed cases */
253 sendcount *= smpi_datatype_size(sendtype);
254 recvcount *= smpi_datatype_size(recvtype);
255 count = sendcount < recvcount ? sendcount : recvcount;
257 if(sendtype->sizeof_substruct == 0 && recvtype->sizeof_substruct == 0) {
258 if(!smpi_process_get_replaying())
259 memcpy(recvbuf, sendbuf, count);
261 else if (sendtype->sizeof_substruct == 0)
263 s_smpi_subtype_t *subtype = static_cast<s_smpi_subtype_t*>(recvtype->substruct);
264 subtype->unserialize( sendbuf, recvbuf, recvcount/smpi_datatype_size(recvtype), subtype, MPI_REPLACE);
266 else if (recvtype->sizeof_substruct == 0)
268 s_smpi_subtype_t *subtype = static_cast<s_smpi_subtype_t*>(sendtype->substruct);
269 subtype->serialize(sendbuf, recvbuf, sendcount/smpi_datatype_size(sendtype), subtype);
271 s_smpi_subtype_t *subtype = static_cast<s_smpi_subtype_t*>(sendtype->substruct);
273 void * buf_tmp = xbt_malloc(count);
275 subtype->serialize( sendbuf, buf_tmp,count/smpi_datatype_size(sendtype), subtype);
276 subtype = static_cast<s_smpi_subtype_t*>(recvtype->substruct);
277 subtype->unserialize( buf_tmp, recvbuf,count/smpi_datatype_size(recvtype), subtype, MPI_REPLACE);
283 return sendcount > recvcount ? MPI_ERR_TRUNCATE : MPI_SUCCESS;
287 * Copies noncontiguous data into contiguous memory.
288 * @param contiguous_vector - output vector
289 * @param noncontiguous_vector - input vector
290 * @param type - pointer containing :
291 * - stride - stride of between noncontiguous data
292 * - block_length - the width or height of blocked matrix
293 * - count - the number of rows of matrix
295 void serialize_vector( const void *noncontiguous_vector, void *contiguous_vector, int count, void *type)
297 s_smpi_mpi_vector_t* type_c = reinterpret_cast<s_smpi_mpi_vector_t*>(type);
299 char* contiguous_vector_char = static_cast<char*>(contiguous_vector);
300 const char* noncontiguous_vector_char = static_cast<const char*>(noncontiguous_vector);
302 for (i = 0; i < type_c->block_count * count; i++) {
303 if (type_c->old_type->sizeof_substruct == 0)
304 memcpy(contiguous_vector_char, noncontiguous_vector_char, type_c->block_length * type_c->size_oldtype);
306 static_cast<s_smpi_subtype_t*>(type_c->old_type->substruct)->serialize( noncontiguous_vector_char,
307 contiguous_vector_char,
308 type_c->block_length, type_c->old_type->substruct);
310 contiguous_vector_char += type_c->block_length*type_c->size_oldtype;
311 if((i+1)%type_c->block_count ==0)
312 noncontiguous_vector_char += type_c->block_length*smpi_datatype_get_extent(type_c->old_type);
314 noncontiguous_vector_char += type_c->block_stride*smpi_datatype_get_extent(type_c->old_type);
319 * Copies contiguous data into noncontiguous memory.
320 * @param noncontiguous_vector - output vector
321 * @param contiguous_vector - input vector
322 * @param type - pointer contening :
323 * - stride - stride of between noncontiguous data
324 * - block_length - the width or height of blocked matrix
325 * - count - the number of rows of matrix
327 void unserialize_vector( const void *contiguous_vector, void *noncontiguous_vector, int count, void *type, MPI_Op op)
329 s_smpi_mpi_vector_t* type_c = reinterpret_cast<s_smpi_mpi_vector_t*>(type);
332 const char* contiguous_vector_char = static_cast<const char*>(contiguous_vector);
333 char* noncontiguous_vector_char = static_cast<char*>(noncontiguous_vector);
335 for (i = 0; i < type_c->block_count * count; i++) {
336 if (type_c->old_type->sizeof_substruct == 0)
337 smpi_op_apply(op, contiguous_vector_char, noncontiguous_vector_char, &type_c->block_length,
340 static_cast<s_smpi_subtype_t*>(type_c->old_type->substruct)->unserialize(contiguous_vector_char, noncontiguous_vector_char,
341 type_c->block_length,type_c->old_type->substruct,
343 contiguous_vector_char += type_c->block_length*type_c->size_oldtype;
344 if((i+1)%type_c->block_count ==0)
345 noncontiguous_vector_char += type_c->block_length*smpi_datatype_get_extent(type_c->old_type);
347 noncontiguous_vector_char += type_c->block_stride*smpi_datatype_get_extent(type_c->old_type);
351 /* Create a Sub type vector to be able to serialize and unserialize it the structure s_smpi_mpi_vector_t is derived
352 * from s_smpi_subtype which required the functions unserialize and serialize */
353 s_smpi_mpi_vector_t* smpi_datatype_vector_create( int block_stride, int block_length, int block_count,
354 MPI_Datatype old_type, int size_oldtype){
355 s_smpi_mpi_vector_t *new_t= xbt_new(s_smpi_mpi_vector_t,1);
356 new_t->base.serialize = &serialize_vector;
357 new_t->base.unserialize = &unserialize_vector;
358 new_t->base.subtype_free = &free_vector;
359 new_t->base.subtype_use = &use_vector;
360 new_t->block_stride = block_stride;
361 new_t->block_length = block_length;
362 new_t->block_count = block_count;
363 smpi_datatype_use(old_type);
364 new_t->old_type = old_type;
365 new_t->size_oldtype = size_oldtype;
369 void smpi_datatype_create(MPI_Datatype* new_type, int size,int lb, int ub, int sizeof_substruct, void *struct_type,
371 MPI_Datatype new_t= xbt_new(s_smpi_mpi_datatype_t,1);
372 new_t->name = nullptr;
374 new_t->sizeof_substruct = size>0? sizeof_substruct:0;
377 new_t->flags = flags;
378 new_t->substruct = struct_type;
380 new_t->attributes=nullptr;
385 MC_ignore(&(new_t->in_use), sizeof(new_t->in_use));
389 void smpi_datatype_free(MPI_Datatype* type){
390 xbt_assert((*type)->in_use >= 0);
392 if((*type)->flags & DT_FLAG_PREDEFINED)
395 //if still used, mark for deletion
396 if((*type)->in_use!=0){
397 (*type)->flags |=DT_FLAG_DESTROYED;
401 if((*type)->attributes !=nullptr){
402 xbt_dict_cursor_t cursor = nullptr;
406 xbt_dict_foreach((*type)->attributes, cursor, key, value){
407 smpi_type_key_elem elem =
408 static_cast<smpi_type_key_elem>(xbt_dict_get_or_null_ext(smpi_type_keyvals, reinterpret_cast<const char*>(key), sizeof(int)));
409 if(elem!=nullptr && elem->delete_fn!=nullptr)
410 elem->delete_fn(*type,*key, value, &flag);
412 xbt_dict_free(&(*type)->attributes);
415 if ((*type)->sizeof_substruct != 0){
416 //((s_smpi_subtype_t *)(*type)->substruct)->subtype_free(type);
417 xbt_free((*type)->substruct);
419 xbt_free((*type)->name);
421 *type = MPI_DATATYPE_NULL;
424 void smpi_datatype_use(MPI_Datatype type){
426 if(type != MPI_DATATYPE_NULL)
429 if(type->sizeof_substruct!=0){
430 static_cast<s_smpi_subtype_t *>((type)->substruct)->subtype_use(&type);
434 MC_ignore(&(type->in_use), sizeof(type->in_use));
438 void smpi_datatype_unuse(MPI_Datatype type){
439 if (type->in_use > 0)
442 if(type->sizeof_substruct!=0){
443 static_cast<s_smpi_subtype_t *>((type)->substruct)->subtype_free(&type);
446 if(type != MPI_DATATYPE_NULL && type->in_use == 0){
447 smpi_datatype_free(&type);
451 MC_ignore(&(type->in_use), sizeof(type->in_use));
455 /*Contiguous Implementation*/
457 /* Copies noncontiguous data into contiguous memory.
458 * @param contiguous_hvector - output hvector
459 * @param noncontiguous_hvector - input hvector
460 * @param type - pointer contening :
461 * - stride - stride of between noncontiguous data, in bytes
462 * - block_length - the width or height of blocked matrix
463 * - count - the number of rows of matrix
465 void serialize_contiguous( const void *noncontiguous_hvector, void *contiguous_hvector, int count, void *type)
467 s_smpi_mpi_contiguous_t* type_c = reinterpret_cast<s_smpi_mpi_contiguous_t*>(type);
468 char* contiguous_vector_char = static_cast<char*>(contiguous_hvector);
469 const char* noncontiguous_vector_char = static_cast<const char*>(noncontiguous_hvector)+type_c->lb;
470 memcpy(contiguous_vector_char, noncontiguous_vector_char, count* type_c->block_count * type_c->size_oldtype);
472 /* Copies contiguous data into noncontiguous memory.
473 * @param noncontiguous_vector - output hvector
474 * @param contiguous_vector - input hvector
475 * @param type - pointer contening :
476 * - stride - stride of between noncontiguous data, in bytes
477 * - block_length - the width or height of blocked matrix
478 * - count - the number of rows of matrix
480 void unserialize_contiguous(const void *contiguous_vector, void *noncontiguous_vector, int count, void *type, MPI_Op op)
482 s_smpi_mpi_contiguous_t* type_c = reinterpret_cast<s_smpi_mpi_contiguous_t*>(type);
483 const char* contiguous_vector_char = static_cast<const char*>(contiguous_vector);
484 char* noncontiguous_vector_char = static_cast<char*>(noncontiguous_vector)+type_c->lb;
485 int n= count* type_c->block_count;
486 smpi_op_apply(op, contiguous_vector_char, noncontiguous_vector_char, &n, &type_c->old_type);
489 void free_contiguous(MPI_Datatype* d){
490 smpi_datatype_unuse(reinterpret_cast<s_smpi_mpi_contiguous_t*>((*d)->substruct)->old_type);
493 void use_contiguous(MPI_Datatype* d){
494 smpi_datatype_use(reinterpret_cast<s_smpi_mpi_contiguous_t*>((*d)->substruct)->old_type);
497 /* Create a Sub type contiguous to be able to serialize and unserialize it the structure s_smpi_mpi_contiguous_t is
498 * derived from s_smpi_subtype which required the functions unserialize and serialize */
499 s_smpi_mpi_contiguous_t* smpi_datatype_contiguous_create( MPI_Aint lb, int block_count, MPI_Datatype old_type,
503 s_smpi_mpi_contiguous_t *new_t= xbt_new(s_smpi_mpi_contiguous_t,1);
504 new_t->base.serialize = &serialize_contiguous;
505 new_t->base.unserialize = &unserialize_contiguous;
506 new_t->base.subtype_free = &free_contiguous;
507 new_t->base.subtype_use = &use_contiguous;
509 new_t->block_count = block_count;
510 new_t->old_type = old_type;
511 smpi_datatype_use(old_type);
512 new_t->size_oldtype = size_oldtype;
513 smpi_datatype_use(old_type);
517 int smpi_datatype_contiguous(int count, MPI_Datatype old_type, MPI_Datatype* new_type, MPI_Aint lb)
520 if(old_type->sizeof_substruct){
521 //handle this case as a hvector with stride equals to the extent of the datatype
522 return smpi_datatype_hvector(count, 1, smpi_datatype_get_extent(old_type), old_type, new_type);
525 s_smpi_mpi_contiguous_t* subtype = smpi_datatype_contiguous_create( lb, count, old_type,smpi_datatype_size(old_type));
527 smpi_datatype_create(new_type, count * smpi_datatype_size(old_type),lb,lb + count * smpi_datatype_size(old_type),
528 sizeof(s_smpi_mpi_contiguous_t),subtype, DT_FLAG_CONTIGUOUS);
533 int smpi_datatype_vector(int count, int blocklen, int stride, MPI_Datatype old_type, MPI_Datatype* new_type)
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 != 0 || 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 contiguous 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, nullptr, DT_FLAG_VECTOR|DT_FLAG_CONTIGUOUS);
560 void free_vector(MPI_Datatype* d){
561 smpi_datatype_unuse(reinterpret_cast<s_smpi_mpi_indexed_t*>((*d)->substruct)->old_type);
564 void use_vector(MPI_Datatype* d){
565 smpi_datatype_use(reinterpret_cast<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 = reinterpret_cast<s_smpi_mpi_hvector_t*>(type);
582 char* contiguous_vector_char = static_cast<char*>(contiguous_hvector);
583 const char* noncontiguous_vector_char = static_cast<const 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 static_cast<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 = reinterpret_cast<s_smpi_mpi_hvector_t*>(type);
613 const char* contiguous_vector_char = static_cast<const char*>(contiguous_vector);
614 char* noncontiguous_vector_char = static_cast<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);
620 static_cast<s_smpi_subtype_t*>(type_c->old_type->substruct)->unserialize( contiguous_vector_char, noncontiguous_vector_char,
621 type_c->block_length, type_c->old_type->substruct,
623 contiguous_vector_char += type_c->block_length*type_c->size_oldtype;
624 if((i+1)%type_c->block_count ==0)
625 noncontiguous_vector_char += type_c->block_length*type_c->size_oldtype;
627 noncontiguous_vector_char += type_c->block_stride;
631 /* Create a Sub type vector to be able to serialize and unserialize it the structure s_smpi_mpi_vector_t is derived
632 * from s_smpi_subtype which required the functions unserialize and serialize
635 s_smpi_mpi_hvector_t* smpi_datatype_hvector_create( MPI_Aint block_stride, int block_length, int block_count,
636 MPI_Datatype old_type, int size_oldtype){
637 s_smpi_mpi_hvector_t *new_t= xbt_new(s_smpi_mpi_hvector_t,1);
638 new_t->base.serialize = &serialize_hvector;
639 new_t->base.unserialize = &unserialize_hvector;
640 new_t->base.subtype_free = &free_hvector;
641 new_t->base.subtype_use = &use_hvector;
642 new_t->block_stride = block_stride;
643 new_t->block_length = block_length;
644 new_t->block_count = block_count;
645 new_t->old_type = old_type;
646 new_t->size_oldtype = size_oldtype;
647 smpi_datatype_use(old_type);
651 //do nothing for vector types
652 void free_hvector(MPI_Datatype* d){
653 smpi_datatype_unuse(reinterpret_cast<s_smpi_mpi_hvector_t*>((*d)->substruct)->old_type);
656 void use_hvector(MPI_Datatype* d){
657 smpi_datatype_use(reinterpret_cast<s_smpi_mpi_hvector_t*>((*d)->substruct)->old_type);
660 int smpi_datatype_hvector(int count, int blocklen, MPI_Aint stride, MPI_Datatype old_type, MPI_Datatype* new_type)
668 lb=smpi_datatype_lb(old_type);
669 ub=((count-1)*stride)+(blocklen-1)*smpi_datatype_get_extent(old_type)+smpi_datatype_ub(old_type);
671 if(old_type->sizeof_substruct != 0 || stride != blocklen*smpi_datatype_get_extent(old_type)){
672 s_smpi_mpi_hvector_t* subtype = smpi_datatype_hvector_create( stride, blocklen, count, old_type,
673 smpi_datatype_size(old_type));
675 smpi_datatype_create(new_type, count * blocklen * smpi_datatype_size(old_type), lb,ub, sizeof(s_smpi_mpi_hvector_t), subtype, DT_FLAG_VECTOR);
678 smpi_datatype_create(new_type, count * blocklen * smpi_datatype_size(old_type),0,count * blocklen *
679 smpi_datatype_size(old_type), 0, nullptr, DT_FLAG_VECTOR|DT_FLAG_CONTIGUOUS);
685 /* Indexed Implementation */
687 /* Copies noncontiguous data into contiguous memory.
688 * @param contiguous_indexed - output indexed
689 * @param noncontiguous_indexed - input indexed
690 * @param type - pointer contening :
691 * - block_lengths - the width or height of blocked matrix
692 * - block_indices - indices of each data, in element
693 * - count - the number of rows of matrix
695 void serialize_indexed( const void *noncontiguous_indexed, void *contiguous_indexed, int count, void *type)
697 s_smpi_mpi_indexed_t* type_c = reinterpret_cast<s_smpi_mpi_indexed_t*>(type);
699 char* contiguous_indexed_char = static_cast<char*>(contiguous_indexed);
700 const char* noncontiguous_indexed_char = static_cast<const char*>(noncontiguous_indexed)+type_c->block_indices[0] * type_c->size_oldtype;
701 for(j=0; j<count;j++){
702 for (i = 0; i < type_c->block_count; i++) {
703 if (type_c->old_type->sizeof_substruct == 0)
704 memcpy(contiguous_indexed_char, noncontiguous_indexed_char, type_c->block_lengths[i] * type_c->size_oldtype);
706 static_cast<s_smpi_subtype_t*>(type_c->old_type->substruct)->serialize( noncontiguous_indexed_char,
707 contiguous_indexed_char,
708 type_c->block_lengths[i],
709 type_c->old_type->substruct);
711 contiguous_indexed_char += type_c->block_lengths[i]*type_c->size_oldtype;
712 if (i<type_c->block_count-1)
713 noncontiguous_indexed_char =
714 static_cast<const char*>(noncontiguous_indexed) + type_c->block_indices[i+1]*smpi_datatype_get_extent(type_c->old_type);
716 noncontiguous_indexed_char += type_c->block_lengths[i]*smpi_datatype_get_extent(type_c->old_type);
718 noncontiguous_indexed=static_cast<const void*>(noncontiguous_indexed_char);
721 /* Copies contiguous data into noncontiguous memory.
722 * @param noncontiguous_indexed - output indexed
723 * @param contiguous_indexed - input indexed
724 * @param type - pointer contening :
725 * - block_lengths - the width or height of blocked matrix
726 * - block_indices - indices of each data, in element
727 * - count - the number of rows of matrix
729 void unserialize_indexed( const void *contiguous_indexed, void *noncontiguous_indexed, int count, void *type, MPI_Op op)
731 s_smpi_mpi_indexed_t* type_c = reinterpret_cast<s_smpi_mpi_indexed_t*>(type);
733 const char* contiguous_indexed_char = static_cast<const char*>(contiguous_indexed);
734 char* noncontiguous_indexed_char =
735 static_cast<char*>(noncontiguous_indexed)+type_c->block_indices[0]*smpi_datatype_get_extent(type_c->old_type);
736 for(j=0; j<count;j++){
737 for (i = 0; i < type_c->block_count; i++) {
738 if (type_c->old_type->sizeof_substruct == 0)
739 smpi_op_apply(op, contiguous_indexed_char, noncontiguous_indexed_char, &type_c->block_lengths[i],
742 static_cast<s_smpi_subtype_t*>(type_c->old_type->substruct)->unserialize( contiguous_indexed_char,
743 noncontiguous_indexed_char,
744 type_c->block_lengths[i],
745 type_c->old_type->substruct, op);
747 contiguous_indexed_char += type_c->block_lengths[i]*type_c->size_oldtype;
748 if (i<type_c->block_count-1)
749 noncontiguous_indexed_char =
750 static_cast<char*>(noncontiguous_indexed) + type_c->block_indices[i+1]*smpi_datatype_get_extent(type_c->old_type);
752 noncontiguous_indexed_char += type_c->block_lengths[i]*smpi_datatype_get_extent(type_c->old_type);
754 noncontiguous_indexed=static_cast<void*>(noncontiguous_indexed_char);
758 void free_indexed(MPI_Datatype* type){
759 if((*type)->in_use==0){
760 xbt_free(reinterpret_cast<s_smpi_mpi_indexed_t*>((*type)->substruct)->block_lengths);
761 xbt_free(reinterpret_cast<s_smpi_mpi_indexed_t*>((*type)->substruct)->block_indices);
763 smpi_datatype_unuse(reinterpret_cast<s_smpi_mpi_indexed_t*>((*type)->substruct)->old_type);
766 void use_indexed(MPI_Datatype* type){
767 smpi_datatype_use(reinterpret_cast<s_smpi_mpi_indexed_t*>((*type)->substruct)->old_type);
771 /* Create a Sub type indexed to be able to serialize and unserialize it the structure s_smpi_mpi_indexed_t is derived
772 * from s_smpi_subtype which required the functions unserialize and serialize */
773 s_smpi_mpi_indexed_t* smpi_datatype_indexed_create( int* block_lengths, int* block_indices, int block_count,
774 MPI_Datatype old_type, int size_oldtype){
775 s_smpi_mpi_indexed_t *new_t= xbt_new(s_smpi_mpi_indexed_t,1);
776 new_t->base.serialize = &serialize_indexed;
777 new_t->base.unserialize = &unserialize_indexed;
778 new_t->base.subtype_free = &free_indexed;
779 new_t->base.subtype_use = &use_indexed;
780 new_t->block_lengths= xbt_new(int, block_count);
781 new_t->block_indices= xbt_new(int, block_count);
783 for(i=0;i<block_count;i++){
784 new_t->block_lengths[i]=block_lengths[i];
785 new_t->block_indices[i]=block_indices[i];
787 new_t->block_count = block_count;
788 smpi_datatype_use(old_type);
789 new_t->old_type = old_type;
790 new_t->size_oldtype = size_oldtype;
794 int smpi_datatype_indexed(int count, int* blocklens, int* indices, MPI_Datatype old_type, MPI_Datatype* new_type)
799 bool contiguous=true;
803 lb=indices[0]*smpi_datatype_get_extent(old_type);
804 ub=indices[0]*smpi_datatype_get_extent(old_type) + blocklens[0]*smpi_datatype_ub(old_type);
807 for(i=0; i< count; i++){
810 size += blocklens[i];
812 if(indices[i]*smpi_datatype_get_extent(old_type)+smpi_datatype_lb(old_type)<lb)
813 lb = indices[i]*smpi_datatype_get_extent(old_type)+smpi_datatype_lb(old_type);
814 if(indices[i]*smpi_datatype_get_extent(old_type)+blocklens[i]*smpi_datatype_ub(old_type)>ub)
815 ub = indices[i]*smpi_datatype_get_extent(old_type)+blocklens[i]*smpi_datatype_ub(old_type);
817 if ( (i< count -1) && (indices[i]+blocklens[i] != indices[i+1]) )
820 if (old_type->sizeof_substruct != 0)
824 s_smpi_mpi_indexed_t* subtype = smpi_datatype_indexed_create( blocklens, indices, count, old_type,
825 smpi_datatype_size(old_type));
826 smpi_datatype_create(new_type, size * smpi_datatype_size(old_type),lb,ub,sizeof(s_smpi_mpi_indexed_t), subtype, DT_FLAG_DATA);
828 s_smpi_mpi_contiguous_t* subtype = smpi_datatype_contiguous_create( lb, size, old_type,
829 smpi_datatype_size(old_type));
830 smpi_datatype_create(new_type, size * smpi_datatype_size(old_type), lb, ub, sizeof(s_smpi_mpi_contiguous_t), subtype,
831 DT_FLAG_DATA|DT_FLAG_CONTIGUOUS);
836 /* Hindexed Implementation - Indexed with indices in bytes */
838 /* Copies noncontiguous data into contiguous memory.
839 * @param contiguous_hindexed - output hindexed
840 * @param noncontiguous_hindexed - input hindexed
841 * @param type - pointer contening :
842 * - block_lengths - the width or height of blocked matrix
843 * - block_indices - indices of each data, in bytes
844 * - count - the number of rows of matrix
846 void serialize_hindexed( const void *noncontiguous_hindexed, void *contiguous_hindexed, int count, void *type)
848 s_smpi_mpi_hindexed_t* type_c = reinterpret_cast<s_smpi_mpi_hindexed_t*>(type);
850 char* contiguous_hindexed_char = static_cast<char*>(contiguous_hindexed);
851 const char* noncontiguous_hindexed_char = static_cast<const char*>(noncontiguous_hindexed)+ type_c->block_indices[0];
852 for(j=0; j<count;j++){
853 for (i = 0; i < type_c->block_count; i++) {
854 if (type_c->old_type->sizeof_substruct == 0)
855 memcpy(contiguous_hindexed_char, noncontiguous_hindexed_char, type_c->block_lengths[i] * type_c->size_oldtype);
857 static_cast<s_smpi_subtype_t*>(type_c->old_type->substruct)->serialize( noncontiguous_hindexed_char,
858 contiguous_hindexed_char,
859 type_c->block_lengths[i],
860 type_c->old_type->substruct);
862 contiguous_hindexed_char += type_c->block_lengths[i]*type_c->size_oldtype;
863 if (i<type_c->block_count-1)
864 noncontiguous_hindexed_char = static_cast<const char*>(noncontiguous_hindexed) + type_c->block_indices[i+1];
866 noncontiguous_hindexed_char += type_c->block_lengths[i]*smpi_datatype_get_extent(type_c->old_type);
868 noncontiguous_hindexed=reinterpret_cast<const void*>(noncontiguous_hindexed_char);
871 /* Copies contiguous data into noncontiguous memory.
872 * @param noncontiguous_hindexed - output hindexed
873 * @param contiguous_hindexed - input hindexed
874 * @param type - pointer contening :
875 * - block_lengths - the width or height of blocked matrix
876 * - block_indices - indices of each data, in bytes
877 * - count - the number of rows of matrix
879 void unserialize_hindexed( const void *contiguous_hindexed, void *noncontiguous_hindexed, int count, void *type,
882 s_smpi_mpi_hindexed_t* type_c = reinterpret_cast<s_smpi_mpi_hindexed_t*>(type);
885 const char* contiguous_hindexed_char = static_cast<const char*>(contiguous_hindexed);
886 char* noncontiguous_hindexed_char = static_cast<char*>(noncontiguous_hindexed)+ type_c->block_indices[0];
887 for(j=0; j<count;j++){
888 for (i = 0; i < type_c->block_count; i++) {
889 if (type_c->old_type->sizeof_substruct == 0)
890 smpi_op_apply(op, contiguous_hindexed_char, noncontiguous_hindexed_char, &type_c->block_lengths[i],
893 static_cast<s_smpi_subtype_t*>(type_c->old_type->substruct)->unserialize( contiguous_hindexed_char,
894 noncontiguous_hindexed_char,
895 type_c->block_lengths[i],
896 type_c->old_type->substruct, op);
898 contiguous_hindexed_char += type_c->block_lengths[i]*type_c->size_oldtype;
899 if (i<type_c->block_count-1)
900 noncontiguous_hindexed_char = static_cast<char*>(noncontiguous_hindexed) + type_c->block_indices[i+1];
902 noncontiguous_hindexed_char += type_c->block_lengths[i]*smpi_datatype_get_extent(type_c->old_type);
904 noncontiguous_hindexed=reinterpret_cast<void*>(noncontiguous_hindexed_char);
908 void free_hindexed(MPI_Datatype* type){
909 if((*type)->in_use==0){
910 xbt_free(reinterpret_cast<s_smpi_mpi_hindexed_t*>((*type)->substruct)->block_lengths);
911 xbt_free(reinterpret_cast<s_smpi_mpi_hindexed_t*>((*type)->substruct)->block_indices);
913 smpi_datatype_unuse(reinterpret_cast<s_smpi_mpi_hindexed_t*>((*type)->substruct)->old_type);
916 void use_hindexed(MPI_Datatype* type){
917 smpi_datatype_use(reinterpret_cast<s_smpi_mpi_hindexed_t*>((*type)->substruct)->old_type);
920 /* Create a Sub type hindexed to be able to serialize and unserialize it the structure s_smpi_mpi_hindexed_t is derived
921 * from s_smpi_subtype which required the functions unserialize and serialize
923 s_smpi_mpi_hindexed_t* smpi_datatype_hindexed_create( int* block_lengths, MPI_Aint* block_indices, int block_count,
924 MPI_Datatype old_type, int size_oldtype){
925 s_smpi_mpi_hindexed_t *new_t= xbt_new(s_smpi_mpi_hindexed_t,1);
926 new_t->base.serialize = &serialize_hindexed;
927 new_t->base.unserialize = &unserialize_hindexed;
928 new_t->base.subtype_free = &free_hindexed;
929 new_t->base.subtype_use = &use_hindexed;
930 new_t->block_lengths= xbt_new(int, block_count);
931 new_t->block_indices= xbt_new(MPI_Aint, block_count);
933 for(i=0;i<block_count;i++){
934 new_t->block_lengths[i]=block_lengths[i];
935 new_t->block_indices[i]=block_indices[i];
937 new_t->block_count = block_count;
938 new_t->old_type = old_type;
939 smpi_datatype_use(old_type);
940 new_t->size_oldtype = size_oldtype;
944 int smpi_datatype_hindexed(int count, int* blocklens, MPI_Aint* indices, MPI_Datatype old_type, MPI_Datatype* new_type)
949 bool contiguous=true;
953 lb=indices[0] + smpi_datatype_lb(old_type);
954 ub=indices[0] + blocklens[0]*smpi_datatype_ub(old_type);
956 for(i=0; i< count; i++){
959 size += blocklens[i];
961 if(indices[i]+smpi_datatype_lb(old_type)<lb)
962 lb = indices[i]+smpi_datatype_lb(old_type);
963 if(indices[i]+blocklens[i]*smpi_datatype_ub(old_type)>ub)
964 ub = indices[i]+blocklens[i]*smpi_datatype_ub(old_type);
966 if ( (i< count -1) && (indices[i]+blocklens[i]*(static_cast<int>(smpi_datatype_size(old_type))) != indices[i+1]) )
969 if (old_type->sizeof_substruct != 0 || lb!=0)
973 s_smpi_mpi_hindexed_t* subtype = smpi_datatype_hindexed_create( blocklens, indices, count, old_type,
974 smpi_datatype_size(old_type));
975 smpi_datatype_create(new_type, size * smpi_datatype_size(old_type), lb, ub ,sizeof(s_smpi_mpi_hindexed_t), subtype, DT_FLAG_DATA);
977 s_smpi_mpi_contiguous_t* subtype = smpi_datatype_contiguous_create(lb,size, old_type, smpi_datatype_size(old_type));
978 smpi_datatype_create(new_type, size * smpi_datatype_size(old_type), 0,size * smpi_datatype_size(old_type),
979 1, subtype, DT_FLAG_DATA|DT_FLAG_CONTIGUOUS);
985 /* struct Implementation - Indexed with indices in bytes */
987 /* Copies noncontiguous data into contiguous memory.
988 * @param contiguous_struct - output struct
989 * @param noncontiguous_struct - input struct
990 * @param type - pointer contening :
991 * - stride - stride of between noncontiguous data
992 * - block_length - the width or height of blocked matrix
993 * - count - the number of rows of matrix
995 void serialize_struct( const void *noncontiguous_struct, void *contiguous_struct, int count, void *type)
997 s_smpi_mpi_struct_t* type_c = reinterpret_cast<s_smpi_mpi_struct_t*>(type);
999 char* contiguous_struct_char = static_cast<char*>(contiguous_struct);
1000 const char* noncontiguous_struct_char = static_cast<const char*>(noncontiguous_struct)+ type_c->block_indices[0];
1001 for(j=0; j<count;j++){
1002 for (i = 0; i < type_c->block_count; i++) {
1003 if (type_c->old_types[i]->sizeof_substruct == 0)
1004 memcpy(contiguous_struct_char, noncontiguous_struct_char,
1005 type_c->block_lengths[i] * smpi_datatype_size(type_c->old_types[i]));
1007 static_cast<s_smpi_subtype_t*>(type_c->old_types[i]->substruct)->serialize( noncontiguous_struct_char,
1008 contiguous_struct_char,
1009 type_c->block_lengths[i],
1010 type_c->old_types[i]->substruct);
1013 contiguous_struct_char += type_c->block_lengths[i]*smpi_datatype_size(type_c->old_types[i]);
1014 if (i<type_c->block_count-1)
1015 noncontiguous_struct_char = static_cast<const char*>(noncontiguous_struct) + type_c->block_indices[i+1];
1016 else //let's hope this is MPI_UB ?
1017 noncontiguous_struct_char += type_c->block_lengths[i]*smpi_datatype_get_extent(type_c->old_types[i]);
1019 noncontiguous_struct=reinterpret_cast<const void*>(noncontiguous_struct_char);
1023 /* Copies contiguous data into noncontiguous memory.
1024 * @param noncontiguous_struct - output struct
1025 * @param contiguous_struct - input struct
1026 * @param type - pointer contening :
1027 * - stride - stride of between noncontiguous data
1028 * - block_length - the width or height of blocked matrix
1029 * - count - the number of rows of matrix
1031 void unserialize_struct( const void *contiguous_struct, void *noncontiguous_struct, int count, void *type, MPI_Op op)
1033 s_smpi_mpi_struct_t* type_c = reinterpret_cast<s_smpi_mpi_struct_t*>(type);
1036 const char* contiguous_struct_char = static_cast<const char*>(contiguous_struct);
1037 char* noncontiguous_struct_char = static_cast<char*>(noncontiguous_struct)+ type_c->block_indices[0];
1038 for(j=0; j<count;j++){
1039 for (i = 0; i < type_c->block_count; i++) {
1040 if (type_c->old_types[i]->sizeof_substruct == 0)
1041 smpi_op_apply(op, contiguous_struct_char, noncontiguous_struct_char, &type_c->block_lengths[i],
1042 & type_c->old_types[i]);
1044 static_cast<s_smpi_subtype_t*>(type_c->old_types[i]->substruct)->unserialize( contiguous_struct_char,
1045 noncontiguous_struct_char,
1046 type_c->block_lengths[i],
1047 type_c->old_types[i]->substruct, op);
1049 contiguous_struct_char += type_c->block_lengths[i]*smpi_datatype_size(type_c->old_types[i]);
1050 if (i<type_c->block_count-1)
1051 noncontiguous_struct_char = static_cast<char*>(noncontiguous_struct) + type_c->block_indices[i+1];
1053 noncontiguous_struct_char += type_c->block_lengths[i]*smpi_datatype_get_extent(type_c->old_types[i]);
1055 noncontiguous_struct=reinterpret_cast<void*>(noncontiguous_struct_char);
1059 void free_struct(MPI_Datatype* type){
1061 for (i = 0; i < reinterpret_cast<s_smpi_mpi_struct_t*>((*type)->substruct)->block_count; i++)
1062 smpi_datatype_unuse(reinterpret_cast<s_smpi_mpi_struct_t*>((*type)->substruct)->old_types[i]);
1063 if((*type)->in_use==0){
1064 xbt_free(reinterpret_cast<s_smpi_mpi_struct_t*>((*type)->substruct)->block_lengths);
1065 xbt_free(reinterpret_cast<s_smpi_mpi_struct_t*>((*type)->substruct)->block_indices);
1066 xbt_free(reinterpret_cast<s_smpi_mpi_struct_t*>((*type)->substruct)->old_types);
1070 void use_struct(MPI_Datatype* type){
1072 for (i = 0; i < reinterpret_cast<s_smpi_mpi_struct_t*>((*type)->substruct)->block_count; i++)
1073 smpi_datatype_use(reinterpret_cast<s_smpi_mpi_struct_t*>((*type)->substruct)->old_types[i]);
1076 /* Create a Sub type struct to be able to serialize and unserialize it the structure s_smpi_mpi_struct_t is derived
1077 * from s_smpi_subtype which required the functions unserialize and serialize
1079 s_smpi_mpi_struct_t* smpi_datatype_struct_create( int* block_lengths, MPI_Aint* block_indices, int block_count,
1080 MPI_Datatype* old_types){
1081 s_smpi_mpi_struct_t *new_t= xbt_new(s_smpi_mpi_struct_t,1);
1082 new_t->base.serialize = &serialize_struct;
1083 new_t->base.unserialize = &unserialize_struct;
1084 new_t->base.subtype_free = &free_struct;
1085 new_t->base.subtype_use = &use_struct;
1086 new_t->block_lengths= xbt_new(int, block_count);
1087 new_t->block_indices= xbt_new(MPI_Aint, block_count);
1088 new_t->old_types= xbt_new(MPI_Datatype, block_count);
1090 for(i=0;i<block_count;i++){
1091 new_t->block_lengths[i]=block_lengths[i];
1092 new_t->block_indices[i]=block_indices[i];
1093 new_t->old_types[i]=old_types[i];
1094 smpi_datatype_use(new_t->old_types[i]);
1096 new_t->block_count = block_count;
1100 int smpi_datatype_struct(int count, int* blocklens, MPI_Aint* indices, MPI_Datatype* old_types, MPI_Datatype* new_type)
1104 bool contiguous=true;
1109 lb=indices[0] + smpi_datatype_lb(old_types[0]);
1110 ub=indices[0] + blocklens[0]*smpi_datatype_ub(old_types[0]);
1112 bool forced_lb=false;
1113 bool forced_ub=false;
1114 for(i=0; i< count; i++){
1117 if (old_types[i]->sizeof_substruct != 0)
1120 size += blocklens[i]*smpi_datatype_size(old_types[i]);
1121 if (old_types[i]==MPI_LB){
1125 if (old_types[i]==MPI_UB){
1130 if(!forced_lb && indices[i]+smpi_datatype_lb(old_types[i])<lb)
1132 if(!forced_ub && indices[i]+blocklens[i]*smpi_datatype_ub(old_types[i])>ub)
1133 ub = indices[i]+blocklens[i]*smpi_datatype_ub(old_types[i]);
1135 if ( (i< count -1) && (indices[i]+blocklens[i]*static_cast<int>(smpi_datatype_size(old_types[i])) != indices[i+1]) )
1140 s_smpi_mpi_struct_t* subtype = smpi_datatype_struct_create( blocklens, indices, count, old_types);
1142 smpi_datatype_create(new_type, size, lb, ub,sizeof(s_smpi_mpi_struct_t), subtype, DT_FLAG_DATA);
1144 s_smpi_mpi_contiguous_t* subtype = smpi_datatype_contiguous_create( lb, size, MPI_CHAR, 1);
1145 smpi_datatype_create(new_type, size, lb, ub,1, subtype, DT_FLAG_DATA|DT_FLAG_CONTIGUOUS);
1150 void smpi_datatype_commit(MPI_Datatype *datatype)
1152 (*datatype)->flags= ((*datatype)->flags | DT_FLAG_COMMITED);
1155 typedef struct s_smpi_mpi_op {
1156 MPI_User_function *func;
1160 #define MAX_OP(a, b) (b) = (a) < (b) ? (b) : (a)
1161 #define MIN_OP(a, b) (b) = (a) < (b) ? (a) : (b)
1162 #define SUM_OP(a, b) (b) += (a)
1163 #define PROD_OP(a, b) (b) *= (a)
1164 #define LAND_OP(a, b) (b) = (a) && (b)
1165 #define LOR_OP(a, b) (b) = (a) || (b)
1166 #define LXOR_OP(a, b) (b) = (!(a) && (b)) || ((a) && !(b))
1167 #define BAND_OP(a, b) (b) &= (a)
1168 #define BOR_OP(a, b) (b) |= (a)
1169 #define BXOR_OP(a, b) (b) ^= (a)
1170 #define MAXLOC_OP(a, b) (b) = (a.value) < (b.value) ? (b) : (a)
1171 #define MINLOC_OP(a, b) (b) = (a.value) < (b.value) ? (a) : (b)
1173 #define APPLY_FUNC(a, b, length, type, func) \
1176 type* x = (type*)(a); \
1177 type* y = (type*)(b); \
1178 for(i = 0; i < *(length); i++) { \
1183 static void max_func(void *a, void *b, int *length, MPI_Datatype * datatype)
1185 if (*datatype == MPI_CHAR) {
1186 APPLY_FUNC(a, b, length, char, MAX_OP)
1187 } else if (*datatype == MPI_SHORT) {
1188 APPLY_FUNC(a, b, length, short, MAX_OP)
1189 } else if (*datatype == MPI_INT) {
1190 APPLY_FUNC(a, b, length, int, MAX_OP)
1191 } else if (*datatype == MPI_LONG) {
1192 APPLY_FUNC(a, b, length, long, MAX_OP)
1193 } else if (*datatype == MPI_UNSIGNED_SHORT) {
1194 APPLY_FUNC(a, b, length, unsigned short, MAX_OP)
1195 } else if (*datatype == MPI_UNSIGNED) {
1196 APPLY_FUNC(a, b, length, unsigned int, MAX_OP)
1197 } else if (*datatype == MPI_UNSIGNED_LONG) {
1198 APPLY_FUNC(a, b, length, unsigned long, MAX_OP)
1199 } else if (*datatype == MPI_UNSIGNED_CHAR) {
1200 APPLY_FUNC(a, b, length, unsigned char, MAX_OP)
1201 } else if (*datatype == MPI_FLOAT) {
1202 APPLY_FUNC(a, b, length, float, MAX_OP)
1203 } else if (*datatype == MPI_DOUBLE) {
1204 APPLY_FUNC(a, b, length, double, MAX_OP)
1205 } else if (*datatype == MPI_LONG_DOUBLE) {
1206 APPLY_FUNC(a, b, length, long double, MAX_OP)
1208 xbt_die("Failed to apply MAX_OP to type %s", (*datatype)->name);
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)
1237 xbt_die("Failed to apply MIN_OP to type %s", (*datatype)->name);
1241 static void sum_func(void *a, void *b, int *length, MPI_Datatype * datatype)
1243 if (*datatype == MPI_CHAR) {
1244 APPLY_FUNC(a, b, length, char, SUM_OP)
1245 } else if (*datatype == MPI_SHORT) {
1246 APPLY_FUNC(a, b, length, short, SUM_OP)
1247 } else if (*datatype == MPI_INT) {
1248 APPLY_FUNC(a, b, length, int, SUM_OP)
1249 } else if (*datatype == MPI_LONG) {
1250 APPLY_FUNC(a, b, length, long, SUM_OP)
1251 } else if (*datatype == MPI_UNSIGNED_SHORT) {
1252 APPLY_FUNC(a, b, length, unsigned short, SUM_OP)
1253 } else if (*datatype == MPI_UNSIGNED) {
1254 APPLY_FUNC(a, b, length, unsigned int, SUM_OP)
1255 } else if (*datatype == MPI_UNSIGNED_LONG) {
1256 APPLY_FUNC(a, b, length, unsigned long, SUM_OP)
1257 } else if (*datatype == MPI_UNSIGNED_CHAR) {
1258 APPLY_FUNC(a, b, length, unsigned char, SUM_OP)
1259 } else if (*datatype == MPI_FLOAT) {
1260 APPLY_FUNC(a, b, length, float, SUM_OP)
1261 } else if (*datatype == MPI_DOUBLE) {
1262 APPLY_FUNC(a, b, length, double, SUM_OP)
1263 } else if (*datatype == MPI_LONG_DOUBLE) {
1264 APPLY_FUNC(a, b, length, long double, SUM_OP)
1265 } else if (*datatype == MPI_C_FLOAT_COMPLEX) {
1266 APPLY_FUNC(a, b, length, float _Complex, SUM_OP)
1267 } else if (*datatype == MPI_C_DOUBLE_COMPLEX) {
1268 APPLY_FUNC(a, b, length, double _Complex, SUM_OP)
1269 } else if (*datatype == MPI_C_LONG_DOUBLE_COMPLEX) {
1270 APPLY_FUNC(a, b, length, long double _Complex, SUM_OP)
1272 xbt_die("Failed to apply SUM_OP to type %s", (*datatype)->name);
1276 static void prod_func(void *a, void *b, int *length, MPI_Datatype * datatype)
1278 if (*datatype == MPI_CHAR) {
1279 APPLY_FUNC(a, b, length, char, PROD_OP)
1280 } else if (*datatype == MPI_SHORT) {
1281 APPLY_FUNC(a, b, length, short, PROD_OP)
1282 } else if (*datatype == MPI_INT) {
1283 APPLY_FUNC(a, b, length, int, PROD_OP)
1284 } else if (*datatype == MPI_LONG) {
1285 APPLY_FUNC(a, b, length, long, PROD_OP)
1286 } else if (*datatype == MPI_UNSIGNED_SHORT) {
1287 APPLY_FUNC(a, b, length, unsigned short, PROD_OP)
1288 } else if (*datatype == MPI_UNSIGNED) {
1289 APPLY_FUNC(a, b, length, unsigned int, PROD_OP)
1290 } else if (*datatype == MPI_UNSIGNED_LONG) {
1291 APPLY_FUNC(a, b, length, unsigned long, PROD_OP)
1292 } else if (*datatype == MPI_UNSIGNED_CHAR) {
1293 APPLY_FUNC(a, b, length, unsigned char, PROD_OP)
1294 } else if (*datatype == MPI_FLOAT) {
1295 APPLY_FUNC(a, b, length, float, PROD_OP)
1296 } else if (*datatype == MPI_DOUBLE) {
1297 APPLY_FUNC(a, b, length, double, PROD_OP)
1298 } else if (*datatype == MPI_LONG_DOUBLE) {
1299 APPLY_FUNC(a, b, length, long double, PROD_OP)
1300 } else if (*datatype == MPI_C_FLOAT_COMPLEX) {
1301 APPLY_FUNC(a, b, length, float _Complex, PROD_OP)
1302 } else if (*datatype == MPI_C_DOUBLE_COMPLEX) {
1303 APPLY_FUNC(a, b, length, double _Complex, PROD_OP)
1304 } else if (*datatype == MPI_C_LONG_DOUBLE_COMPLEX) {
1305 APPLY_FUNC(a, b, length, long double _Complex, PROD_OP)
1307 xbt_die("Failed to apply PROD_OP to type %s", (*datatype)->name);
1311 static void land_func(void *a, void *b, int *length, MPI_Datatype * datatype)
1313 if (*datatype == MPI_CHAR) {
1314 APPLY_FUNC(a, b, length, char, LAND_OP)
1315 } else if (*datatype == MPI_SHORT) {
1316 APPLY_FUNC(a, b, length, short, LAND_OP)
1317 } else if (*datatype == MPI_INT) {
1318 APPLY_FUNC(a, b, length, int, LAND_OP)
1319 } else if (*datatype == MPI_LONG) {
1320 APPLY_FUNC(a, b, length, long, LAND_OP)
1321 } else if (*datatype == MPI_UNSIGNED_SHORT) {
1322 APPLY_FUNC(a, b, length, unsigned short, LAND_OP)
1323 } else if (*datatype == MPI_UNSIGNED) {
1324 APPLY_FUNC(a, b, length, unsigned int, LAND_OP)
1325 } else if (*datatype == MPI_UNSIGNED_LONG) {
1326 APPLY_FUNC(a, b, length, unsigned long, LAND_OP)
1327 } else if (*datatype == MPI_UNSIGNED_CHAR) {
1328 APPLY_FUNC(a, b, length, unsigned char, LAND_OP)
1329 } else if (*datatype == MPI_C_BOOL) {
1330 APPLY_FUNC(a, b, length, bool, LAND_OP)
1332 xbt_die("Failed to apply LAND_OP to type %s", (*datatype)->name);
1336 static void lor_func(void *a, void *b, int *length, MPI_Datatype * datatype)
1338 if (*datatype == MPI_CHAR) {
1339 APPLY_FUNC(a, b, length, char, LOR_OP)
1340 } else if (*datatype == MPI_SHORT) {
1341 APPLY_FUNC(a, b, length, short, LOR_OP)
1342 } else if (*datatype == MPI_INT) {
1343 APPLY_FUNC(a, b, length, int, LOR_OP)
1344 } else if (*datatype == MPI_LONG) {
1345 APPLY_FUNC(a, b, length, long, LOR_OP)
1346 } else if (*datatype == MPI_UNSIGNED_SHORT) {
1347 APPLY_FUNC(a, b, length, unsigned short, LOR_OP)
1348 } else if (*datatype == MPI_UNSIGNED) {
1349 APPLY_FUNC(a, b, length, unsigned int, LOR_OP)
1350 } else if (*datatype == MPI_UNSIGNED_LONG) {
1351 APPLY_FUNC(a, b, length, unsigned long, LOR_OP)
1352 } else if (*datatype == MPI_UNSIGNED_CHAR) {
1353 APPLY_FUNC(a, b, length, unsigned char, LOR_OP)
1354 } else if (*datatype == MPI_C_BOOL) {
1355 APPLY_FUNC(a, b, length, bool, LOR_OP)
1357 xbt_die("Failed to apply LOR_OP to type %s", (*datatype)->name);
1361 static void lxor_func(void *a, void *b, int *length, MPI_Datatype * datatype)
1363 if (*datatype == MPI_CHAR) {
1364 APPLY_FUNC(a, b, length, char, LXOR_OP)
1365 } else if (*datatype == MPI_SHORT) {
1366 APPLY_FUNC(a, b, length, short, LXOR_OP)
1367 } else if (*datatype == MPI_INT) {
1368 APPLY_FUNC(a, b, length, int, LXOR_OP)
1369 } else if (*datatype == MPI_LONG) {
1370 APPLY_FUNC(a, b, length, long, LXOR_OP)
1371 } else if (*datatype == MPI_UNSIGNED_SHORT) {
1372 APPLY_FUNC(a, b, length, unsigned short, LXOR_OP)
1373 } else if (*datatype == MPI_UNSIGNED) {
1374 APPLY_FUNC(a, b, length, unsigned int, LXOR_OP)
1375 } else if (*datatype == MPI_UNSIGNED_LONG) {
1376 APPLY_FUNC(a, b, length, unsigned long, LXOR_OP)
1377 } else if (*datatype == MPI_UNSIGNED_CHAR) {
1378 APPLY_FUNC(a, b, length, unsigned char, LXOR_OP)
1379 } else if (*datatype == MPI_C_BOOL) {
1380 APPLY_FUNC(a, b, length, bool, LXOR_OP)
1382 xbt_die("Failed to apply LXOR_OP to type %s", (*datatype)->name);
1386 static void band_func(void *a, void *b, int *length, MPI_Datatype * datatype)
1388 if (*datatype == MPI_CHAR) {
1389 APPLY_FUNC(a, b, length, char, BAND_OP)
1390 }else if (*datatype == MPI_SHORT) {
1391 APPLY_FUNC(a, b, length, short, BAND_OP)
1392 } else if (*datatype == MPI_INT) {
1393 APPLY_FUNC(a, b, length, int, BAND_OP)
1394 } else if (*datatype == MPI_LONG) {
1395 APPLY_FUNC(a, b, length, long, BAND_OP)
1396 } else if (*datatype == MPI_UNSIGNED_SHORT) {
1397 APPLY_FUNC(a, b, length, unsigned short, BAND_OP)
1398 } else if (*datatype == MPI_UNSIGNED) {
1399 APPLY_FUNC(a, b, length, unsigned int, BAND_OP)
1400 } else if (*datatype == MPI_UNSIGNED_LONG) {
1401 APPLY_FUNC(a, b, length, unsigned long, BAND_OP)
1402 } else if (*datatype == MPI_UNSIGNED_CHAR) {
1403 APPLY_FUNC(a, b, length, unsigned char, BAND_OP)
1404 } else if (*datatype == MPI_BYTE) {
1405 APPLY_FUNC(a, b, length, uint8_t, BAND_OP)
1407 xbt_die("Failed to apply BAND_OP to type %s", (*datatype)->name);
1411 static void bor_func(void *a, void *b, int *length, MPI_Datatype * datatype)
1413 if (*datatype == MPI_CHAR) {
1414 APPLY_FUNC(a, b, length, char, BOR_OP)
1415 } else if (*datatype == MPI_SHORT) {
1416 APPLY_FUNC(a, b, length, short, BOR_OP)
1417 } else if (*datatype == MPI_INT) {
1418 APPLY_FUNC(a, b, length, int, BOR_OP)
1419 } else if (*datatype == MPI_LONG) {
1420 APPLY_FUNC(a, b, length, long, BOR_OP)
1421 } else if (*datatype == MPI_UNSIGNED_SHORT) {
1422 APPLY_FUNC(a, b, length, unsigned short, BOR_OP)
1423 } else if (*datatype == MPI_UNSIGNED) {
1424 APPLY_FUNC(a, b, length, unsigned int, BOR_OP)
1425 } else if (*datatype == MPI_UNSIGNED_LONG) {
1426 APPLY_FUNC(a, b, length, unsigned long, BOR_OP)
1427 } else if (*datatype == MPI_UNSIGNED_CHAR) {
1428 APPLY_FUNC(a, b, length, unsigned char, BOR_OP)
1429 } else if (*datatype == MPI_BYTE) {
1430 APPLY_FUNC(a, b, length, uint8_t, BOR_OP)
1432 xbt_die("Failed to apply BOR_OP to type %s", (*datatype)->name);
1436 static void bxor_func(void *a, void *b, int *length, MPI_Datatype * datatype)
1438 if (*datatype == MPI_CHAR) {
1439 APPLY_FUNC(a, b, length, char, BXOR_OP)
1440 } else if (*datatype == MPI_SHORT) {
1441 APPLY_FUNC(a, b, length, short, BXOR_OP)
1442 } else if (*datatype == MPI_INT) {
1443 APPLY_FUNC(a, b, length, int, BXOR_OP)
1444 } else if (*datatype == MPI_LONG) {
1445 APPLY_FUNC(a, b, length, long, BXOR_OP)
1446 } else if (*datatype == MPI_UNSIGNED_SHORT) {
1447 APPLY_FUNC(a, b, length, unsigned short, BXOR_OP)
1448 } else if (*datatype == MPI_UNSIGNED) {
1449 APPLY_FUNC(a, b, length, unsigned int, BXOR_OP)
1450 } else if (*datatype == MPI_UNSIGNED_LONG) {
1451 APPLY_FUNC(a, b, length, unsigned long, BXOR_OP)
1452 } else if (*datatype == MPI_UNSIGNED_CHAR) {
1453 APPLY_FUNC(a, b, length, unsigned char, BXOR_OP)
1454 } else if (*datatype == MPI_BYTE) {
1455 APPLY_FUNC(a, b, length, uint8_t, BXOR_OP)
1457 xbt_die("Failed to apply BXOR_OP to type %s", (*datatype)->name);
1461 static void minloc_func(void *a, void *b, int *length, MPI_Datatype * datatype)
1463 if (*datatype == MPI_FLOAT_INT) {
1464 APPLY_FUNC(a, b, length, float_int, MINLOC_OP)
1465 } else if (*datatype == MPI_LONG_INT) {
1466 APPLY_FUNC(a, b, length, long_int, MINLOC_OP)
1467 } else if (*datatype == MPI_DOUBLE_INT) {
1468 APPLY_FUNC(a, b, length, double_int, MINLOC_OP)
1469 } else if (*datatype == MPI_SHORT_INT) {
1470 APPLY_FUNC(a, b, length, short_int, MINLOC_OP)
1471 } else if (*datatype == MPI_2LONG) {
1472 APPLY_FUNC(a, b, length, long_long, MINLOC_OP)
1473 } else if (*datatype == MPI_2INT) {
1474 APPLY_FUNC(a, b, length, int_int, MINLOC_OP)
1475 } else if (*datatype == MPI_LONG_DOUBLE_INT) {
1476 APPLY_FUNC(a, b, length, long_double_int, MINLOC_OP)
1477 } else if (*datatype == MPI_2FLOAT) {
1478 APPLY_FUNC(a, b, length, float_float, MINLOC_OP)
1479 } else if (*datatype == MPI_2DOUBLE) {
1480 APPLY_FUNC(a, b, length, double_double, MINLOC_OP)
1482 xbt_die("Failed to apply MINLOC_OP to type %s", (*datatype)->name);
1486 static void maxloc_func(void *a, void *b, int *length, MPI_Datatype * datatype)
1488 if (*datatype == MPI_FLOAT_INT) {
1489 APPLY_FUNC(a, b, length, float_int, MAXLOC_OP)
1490 } else if (*datatype == MPI_LONG_INT) {
1491 APPLY_FUNC(a, b, length, long_int, MAXLOC_OP)
1492 } else if (*datatype == MPI_DOUBLE_INT) {
1493 APPLY_FUNC(a, b, length, double_int, MAXLOC_OP)
1494 } else if (*datatype == MPI_SHORT_INT) {
1495 APPLY_FUNC(a, b, length, short_int, MAXLOC_OP)
1496 } else if (*datatype == MPI_2LONG) {
1497 APPLY_FUNC(a, b, length, long_long, MAXLOC_OP)
1498 } else if (*datatype == MPI_2INT) {
1499 APPLY_FUNC(a, b, length, int_int, MAXLOC_OP)
1500 } else if (*datatype == MPI_LONG_DOUBLE_INT) {
1501 APPLY_FUNC(a, b, length, long_double_int, MAXLOC_OP)
1502 } else if (*datatype == MPI_2FLOAT) {
1503 APPLY_FUNC(a, b, length, float_float, MAXLOC_OP)
1504 } else if (*datatype == MPI_2DOUBLE) {
1505 APPLY_FUNC(a, b, length, double_double, MAXLOC_OP)
1507 xbt_die("Failed to apply MAXLOC_OP to type %s", (*datatype)->name);
1511 static void replace_func(void *a, void *b, int *length, MPI_Datatype * datatype)
1513 memcpy(b, a, *length * smpi_datatype_size(*datatype));
1516 #define CREATE_MPI_OP(name, func) \
1517 static s_smpi_mpi_op_t mpi_##name = { &(func) /* func */, true }; \
1518 MPI_Op name = &mpi_##name;
1520 CREATE_MPI_OP(MPI_MAX, max_func);
1521 CREATE_MPI_OP(MPI_MIN, min_func);
1522 CREATE_MPI_OP(MPI_SUM, sum_func);
1523 CREATE_MPI_OP(MPI_PROD, prod_func);
1524 CREATE_MPI_OP(MPI_LAND, land_func);
1525 CREATE_MPI_OP(MPI_LOR, lor_func);
1526 CREATE_MPI_OP(MPI_LXOR, lxor_func);
1527 CREATE_MPI_OP(MPI_BAND, band_func);
1528 CREATE_MPI_OP(MPI_BOR, bor_func);
1529 CREATE_MPI_OP(MPI_BXOR, bxor_func);
1530 CREATE_MPI_OP(MPI_MAXLOC, maxloc_func);
1531 CREATE_MPI_OP(MPI_MINLOC, minloc_func);
1532 CREATE_MPI_OP(MPI_REPLACE, replace_func);
1534 MPI_Op smpi_op_new(MPI_User_function * function, bool commute)
1537 op = xbt_new(s_smpi_mpi_op_t, 1);
1538 op->func = function;
1539 op-> is_commute = commute;
1543 bool smpi_op_is_commute(MPI_Op op)
1545 return (op==MPI_OP_NULL) ? true : op-> is_commute;
1548 void smpi_op_destroy(MPI_Op op)
1553 void smpi_op_apply(MPI_Op op, const void *invec, void *inoutvec, int *len, MPI_Datatype * datatype)
1558 if(smpi_privatize_global_variables){//we need to switch as the called function may silently touch global variables
1559 XBT_DEBUG("Applying operation, switch to the right data frame ");
1560 smpi_switch_data_segment(smpi_process_index());
1563 if(!smpi_process_get_replaying())
1564 op->func(const_cast<void*>(invec), inoutvec, len, datatype);
1567 int smpi_type_attr_delete(MPI_Datatype type, int keyval){
1568 smpi_type_key_elem elem =
1569 static_cast<smpi_type_key_elem>(xbt_dict_get_or_null_ext(smpi_type_keyvals, reinterpret_cast<const char*>(&keyval), sizeof(int)));
1572 if(elem->delete_fn!=MPI_NULL_DELETE_FN){
1573 void * value = nullptr;
1575 if(smpi_type_attr_get(type, keyval, &value, &flag)==MPI_SUCCESS){
1576 int ret = elem->delete_fn(type, keyval, value, &flag);
1577 if(ret!=MPI_SUCCESS)
1581 if(type->attributes==nullptr)
1584 xbt_dict_remove_ext(type->attributes, reinterpret_cast<const char*>(&keyval), sizeof(int));
1588 int smpi_type_attr_get(MPI_Datatype type, int keyval, void* attr_value, int* flag){
1589 smpi_type_key_elem elem =
1590 static_cast<smpi_type_key_elem>(xbt_dict_get_or_null_ext(smpi_type_keyvals, reinterpret_cast<const char*>(&keyval), sizeof(int)));
1593 if(type->attributes==nullptr){
1598 *static_cast<void**>(attr_value) = xbt_dict_get_ext(type->attributes, reinterpret_cast<const char*>(&keyval), sizeof(int));
1601 catch (xbt_ex& ex) {
1607 int smpi_type_attr_put(MPI_Datatype type, int keyval, void* attr_value){
1608 if(smpi_type_keyvals==nullptr)
1609 smpi_type_keyvals = xbt_dict_new();
1610 smpi_type_key_elem elem =
1611 static_cast<smpi_type_key_elem>(xbt_dict_get_or_null_ext(smpi_type_keyvals, reinterpret_cast<const char*>(&keyval), sizeof(int)));
1615 void* value = nullptr;
1616 smpi_type_attr_get(type, keyval, &value, &flag);
1617 if(flag!=0 && elem->delete_fn!=MPI_NULL_DELETE_FN){
1618 int ret = elem->delete_fn(type, keyval, value, &flag);
1619 if(ret!=MPI_SUCCESS)
1622 if(type->attributes==nullptr)
1623 type->attributes=xbt_dict_new();
1625 xbt_dict_set_ext(type->attributes, reinterpret_cast<const char*>(&keyval), sizeof(int), attr_value, nullptr);
1629 int smpi_type_keyval_create(MPI_Type_copy_attr_function* copy_fn, MPI_Type_delete_attr_function* delete_fn, int* keyval,
1631 if(smpi_type_keyvals==nullptr)
1632 smpi_type_keyvals = xbt_dict_new();
1634 smpi_type_key_elem value = (smpi_type_key_elem) xbt_new0(s_smpi_mpi_type_key_elem_t,1);
1636 value->copy_fn=copy_fn;
1637 value->delete_fn=delete_fn;
1639 *keyval = type_keyval_id;
1640 xbt_dict_set_ext(smpi_type_keyvals,reinterpret_cast<const char*>(keyval), sizeof(int),reinterpret_cast<void*>(value), nullptr);
1645 int smpi_type_keyval_free(int* keyval){
1646 smpi_type_key_elem elem =
1647 static_cast<smpi_type_key_elem>(xbt_dict_get_or_null_ext(smpi_type_keyvals, reinterpret_cast<const char*>(keyval), sizeof(int)));
1651 xbt_dict_remove_ext(smpi_type_keyvals, reinterpret_cast<const char*>(keyval), sizeof(int));
1656 int smpi_mpi_pack(void* inbuf, int incount, MPI_Datatype type, void* outbuf, int outcount, int* position,MPI_Comm comm){
1657 size_t size = smpi_datatype_size(type);
1658 if (outcount - *position < incount*static_cast<int>(size))
1659 return MPI_ERR_BUFFER;
1660 smpi_datatype_copy(inbuf, incount, type, static_cast<char*>(outbuf) + *position, outcount, MPI_CHAR);
1661 *position += incount * size;
1665 int smpi_mpi_unpack(void* inbuf, int insize, int* position, void* outbuf, int outcount, MPI_Datatype type,MPI_Comm comm){
1666 int size = static_cast<int>(smpi_datatype_size(type));
1667 if (outcount*size> insize)
1668 return MPI_ERR_BUFFER;
1669 smpi_datatype_copy(static_cast<char*>(inbuf) + *position, insize, MPI_CHAR, outbuf, outcount, type);
1670 *position += outcount * size;