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;
207 xbt_dict_set_ext((*new_t)->attributes, (const char*)key, sizeof(int),value_out, NULL);
214 int smpi_datatype_extent(MPI_Datatype datatype, MPI_Aint * lb, MPI_Aint * extent)
216 if(datatype == MPI_DATATYPE_NULL){
222 *extent = datatype->ub - datatype->lb;
226 MPI_Aint smpi_datatype_get_extent(MPI_Datatype datatype){
227 if(datatype == MPI_DATATYPE_NULL){
230 return datatype->ub - datatype->lb;
233 void smpi_datatype_get_name(MPI_Datatype datatype, char* name, int* length){
234 *length = strlen(datatype->name);
235 strcpy(name, datatype->name);
238 void smpi_datatype_set_name(MPI_Datatype datatype, char* name){
239 datatype->name = xbt_strdup(name);;
242 int smpi_datatype_copy(void *sendbuf, int sendcount, MPI_Datatype sendtype,
243 void *recvbuf, int recvcount, MPI_Datatype recvtype)
246 if(smpi_privatize_global_variables){
247 smpi_switch_data_segment(smpi_process_index());
249 /* First check if we really have something to do */
250 if (recvcount > 0 && recvbuf != sendbuf) {
251 /* FIXME: treat packed cases */
252 sendcount *= smpi_datatype_size(sendtype);
253 recvcount *= smpi_datatype_size(recvtype);
254 count = sendcount < recvcount ? sendcount : recvcount;
256 if(sendtype->sizeof_substruct == 0 && recvtype->sizeof_substruct == 0) {
257 if(!smpi_process_get_replaying()) memcpy(recvbuf, sendbuf, count);
259 else if (sendtype->sizeof_substruct == 0)
261 s_smpi_subtype_t *subtype = static_cast<s_smpi_subtype_t*>(recvtype->substruct);
262 subtype->unserialize( sendbuf, recvbuf, recvcount/smpi_datatype_size(recvtype), subtype, MPI_REPLACE);
264 else if (recvtype->sizeof_substruct == 0)
266 s_smpi_subtype_t *subtype = static_cast<s_smpi_subtype_t*>(sendtype->substruct);
267 subtype->serialize(sendbuf, recvbuf, sendcount/smpi_datatype_size(sendtype), subtype);
269 s_smpi_subtype_t *subtype = static_cast<s_smpi_subtype_t*>(sendtype->substruct);
271 void * buf_tmp = xbt_malloc(count);
273 subtype->serialize( sendbuf, buf_tmp,count/smpi_datatype_size(sendtype), subtype);
274 subtype = static_cast<s_smpi_subtype_t*>(recvtype->substruct);
275 subtype->unserialize( buf_tmp, recvbuf,count/smpi_datatype_size(recvtype), subtype, MPI_REPLACE);
281 return sendcount > recvcount ? MPI_ERR_TRUNCATE : MPI_SUCCESS;
285 * Copies noncontiguous data into contiguous memory.
286 * @param contiguous_vector - output vector
287 * @param noncontiguous_vector - input vector
288 * @param type - pointer contening :
289 * - stride - stride of between noncontiguous data
290 * - block_length - the width or height of blocked matrix
291 * - count - the number of rows of matrix
293 void serialize_vector( const void *noncontiguous_vector, void *contiguous_vector, int count, void *type)
295 s_smpi_mpi_vector_t* type_c = (s_smpi_mpi_vector_t*)type;
297 char* contiguous_vector_char = (char*)contiguous_vector;
298 char* noncontiguous_vector_char = (char*)noncontiguous_vector;
300 for (i = 0; i < type_c->block_count * count; i++) {
301 if (type_c->old_type->sizeof_substruct == 0)
302 memcpy(contiguous_vector_char, noncontiguous_vector_char, type_c->block_length * type_c->size_oldtype);
304 ((s_smpi_subtype_t*)type_c->old_type->substruct)->serialize( noncontiguous_vector_char,
305 contiguous_vector_char,
306 type_c->block_length, type_c->old_type->substruct);
308 contiguous_vector_char += type_c->block_length*type_c->size_oldtype;
309 if((i+1)%type_c->block_count ==0)
310 noncontiguous_vector_char += type_c->block_length*smpi_datatype_get_extent(type_c->old_type);
312 noncontiguous_vector_char += type_c->block_stride*smpi_datatype_get_extent(type_c->old_type);
317 * Copies contiguous data into noncontiguous memory.
318 * @param noncontiguous_vector - output vector
319 * @param contiguous_vector - input vector
320 * @param type - pointer contening :
321 * - stride - stride of between noncontiguous data
322 * - block_length - the width or height of blocked matrix
323 * - count - the number of rows of matrix
325 void unserialize_vector( const void *contiguous_vector, void *noncontiguous_vector, int count, void *type, MPI_Op op)
327 s_smpi_mpi_vector_t* type_c = (s_smpi_mpi_vector_t*)type;
330 char* contiguous_vector_char = (char*)contiguous_vector;
331 char* noncontiguous_vector_char = (char*)noncontiguous_vector;
333 for (i = 0; i < type_c->block_count * count; i++) {
334 if (type_c->old_type->sizeof_substruct == 0)
335 smpi_op_apply(op, contiguous_vector_char, noncontiguous_vector_char, &type_c->block_length,
337 /* memcpy(noncontiguous_vector_char,
338 contiguous_vector_char, type_c->block_length * type_c->size_oldtype);*/
340 ((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);
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=NULL;
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);
391 if((*type)->attributes !=NULL){
392 xbt_dict_cursor_t cursor = NULL;
396 xbt_dict_foreach((*type)->attributes, cursor, key, value){
397 smpi_type_key_elem elem =
398 static_cast<smpi_type_key_elem>(xbt_dict_get_or_null_ext(smpi_type_keyvals, (const char*)key, sizeof(int)));
399 if(elem && elem->delete_fn)
400 elem->delete_fn(*type,*key, value, &flag);
404 if((*type)->flags & DT_FLAG_PREDEFINED)return;
406 //if still used, mark for deletion
407 if((*type)->in_use!=0){
408 (*type)->flags |=DT_FLAG_DESTROYED;
412 if ((*type)->sizeof_substruct != 0){
413 //((s_smpi_subtype_t *)(*type)->substruct)->subtype_free(type);
414 xbt_free((*type)->substruct);
416 if ((*type)->name != NULL){
417 xbt_free((*type)->name);
419 *type = MPI_DATATYPE_NULL;
422 void smpi_datatype_use(MPI_Datatype type){
424 if(type)type->in_use++;
426 if(type->sizeof_substruct!=0){
427 ((s_smpi_subtype_t *)(type)->substruct)->subtype_use(&type);
431 MC_ignore(&(type->in_use), sizeof(type->in_use));
435 void smpi_datatype_unuse(MPI_Datatype type){
436 if (type->in_use > 0)
439 if(type->sizeof_substruct!=0){
440 ((s_smpi_subtype_t *)(type)->substruct)->subtype_free(&type);
443 if(type && type->in_use == 0){
444 MPI_Datatype t = type;
445 if (!(type->flags & DT_FLAG_DESTROYED))
446 smpi_datatype_free(&type);
447 if(t->flags & DT_FLAG_PREDEFINED) return;
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 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)
536 if (blocklen<0) return MPI_ERR_ARG;
540 lb=smpi_datatype_lb(old_type);
541 ub=((count-1)*stride+blocklen-1)*smpi_datatype_get_extent(old_type)+smpi_datatype_ub(old_type);
543 if(old_type->sizeof_substruct || stride != blocklen){
545 s_smpi_mpi_vector_t* subtype = smpi_datatype_vector_create(stride, blocklen, count, old_type,
546 smpi_datatype_size(old_type));
547 smpi_datatype_create(new_type, count * (blocklen) * smpi_datatype_size(old_type), lb, ub, sizeof(s_smpi_mpi_vector_t), subtype,
551 /* in this situation the data are contignous thus it's not required to serialize and unserialize it*/
552 smpi_datatype_create(new_type, count * blocklen * smpi_datatype_size(old_type), 0, ((count -1) * stride + blocklen)*
553 smpi_datatype_size(old_type), 0, NULL, DT_FLAG_VECTOR|DT_FLAG_CONTIGUOUS);
559 void free_vector(MPI_Datatype* d){
560 smpi_datatype_unuse(((s_smpi_mpi_indexed_t *)(*d)->substruct)->old_type);
563 void use_vector(MPI_Datatype* d){
564 smpi_datatype_use(((s_smpi_mpi_indexed_t *)(*d)->substruct)->old_type);
567 /* Hvector Implementation - Vector with stride in bytes */
569 /* Copies noncontiguous data into contiguous memory.
570 * @param contiguous_hvector - output hvector
571 * @param noncontiguous_hvector - input hvector
572 * @param type - pointer contening :
573 * - stride - stride of between noncontiguous data, in bytes
574 * - block_length - the width or height of blocked matrix
575 * - count - the number of rows of matrix
577 void serialize_hvector( const void *noncontiguous_hvector, void *contiguous_hvector, int count, void *type)
579 s_smpi_mpi_hvector_t* type_c = (s_smpi_mpi_hvector_t*)type;
581 char* contiguous_vector_char = (char*)contiguous_hvector;
582 char* noncontiguous_vector_char = (char*)noncontiguous_hvector;
584 for (i = 0; i < type_c->block_count * count; i++) {
585 if (type_c->old_type->sizeof_substruct == 0)
586 memcpy(contiguous_vector_char, noncontiguous_vector_char, type_c->block_length * type_c->size_oldtype);
588 ((s_smpi_subtype_t*)type_c->old_type->substruct)->serialize( noncontiguous_vector_char,
589 contiguous_vector_char,
590 type_c->block_length, type_c->old_type->substruct);
592 contiguous_vector_char += type_c->block_length*type_c->size_oldtype;
593 if((i+1)%type_c->block_count ==0)
594 noncontiguous_vector_char += type_c->block_length*type_c->size_oldtype;
596 noncontiguous_vector_char += type_c->block_stride;
599 /* Copies contiguous data into noncontiguous memory.
600 * @param noncontiguous_vector - output hvector
601 * @param contiguous_vector - input hvector
602 * @param type - pointer contening :
603 * - stride - stride of between noncontiguous data, in bytes
604 * - block_length - the width or height of blocked matrix
605 * - count - the number of rows of matrix
607 void unserialize_hvector( const void *contiguous_vector, void *noncontiguous_vector, int count, void *type, MPI_Op op)
609 s_smpi_mpi_hvector_t* type_c = (s_smpi_mpi_hvector_t*)type;
612 char* contiguous_vector_char = (char*)contiguous_vector;
613 char* noncontiguous_vector_char = (char*)noncontiguous_vector;
615 for (i = 0; i < type_c->block_count * count; i++) {
616 if (type_c->old_type->sizeof_substruct == 0)
617 smpi_op_apply(op, contiguous_vector_char, noncontiguous_vector_char, &type_c->block_length, &type_c->old_type);
618 /*memcpy(noncontiguous_vector_char,
619 contiguous_vector_char, type_c->block_length * type_c->size_oldtype);*/
621 ((s_smpi_subtype_t*)type_c->old_type->substruct)->unserialize( contiguous_vector_char, noncontiguous_vector_char,
622 type_c->block_length, type_c->old_type->substruct,
624 contiguous_vector_char += type_c->block_length*type_c->size_oldtype;
625 if((i+1)%type_c->block_count ==0)
626 noncontiguous_vector_char += type_c->block_length*type_c->size_oldtype;
628 noncontiguous_vector_char += type_c->block_stride;
632 /* Create a Sub type vector to be able to serialize and unserialize it the structure s_smpi_mpi_vector_t is derived
633 * from s_smpi_subtype which required the functions unserialize and serialize
636 s_smpi_mpi_hvector_t* smpi_datatype_hvector_create( MPI_Aint block_stride, int block_length, int block_count,
637 MPI_Datatype old_type, int size_oldtype){
638 s_smpi_mpi_hvector_t *new_t= xbt_new(s_smpi_mpi_hvector_t,1);
639 new_t->base.serialize = &serialize_hvector;
640 new_t->base.unserialize = &unserialize_hvector;
641 new_t->base.subtype_free = &free_hvector;
642 new_t->base.subtype_use = &use_hvector;
643 new_t->block_stride = block_stride;
644 new_t->block_length = block_length;
645 new_t->block_count = block_count;
646 new_t->old_type = old_type;
647 new_t->size_oldtype = size_oldtype;
648 smpi_datatype_use(old_type);
652 //do nothing for vector types
653 void free_hvector(MPI_Datatype* d){
654 smpi_datatype_unuse(((s_smpi_mpi_hvector_t *)(*d)->substruct)->old_type);
657 void use_hvector(MPI_Datatype* d){
658 smpi_datatype_use(((s_smpi_mpi_hvector_t *)(*d)->substruct)->old_type);
661 int smpi_datatype_hvector(int count, int blocklen, MPI_Aint stride, MPI_Datatype old_type, MPI_Datatype* new_type)
664 if (blocklen<0) return MPI_ERR_ARG;
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 || 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, NULL, 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 = (s_smpi_mpi_indexed_t*)type;
699 char* contiguous_indexed_char = (char*)contiguous_indexed;
700 char* noncontiguous_indexed_char = (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 ((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 (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=(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 = (s_smpi_mpi_indexed_t*)type;
733 char* contiguous_indexed_char = (char*)contiguous_indexed;
734 char* noncontiguous_indexed_char =
735 (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],
741 /*memcpy(noncontiguous_indexed_char ,
742 contiguous_indexed_char, type_c->block_lengths[i] * type_c->size_oldtype);*/
744 ((s_smpi_subtype_t*)type_c->old_type->substruct)->unserialize( contiguous_indexed_char,
745 noncontiguous_indexed_char,
746 type_c->block_lengths[i],
747 type_c->old_type->substruct, op);
749 contiguous_indexed_char += type_c->block_lengths[i]*type_c->size_oldtype;
750 if (i<type_c->block_count-1)
751 noncontiguous_indexed_char =
752 (char*)noncontiguous_indexed + type_c->block_indices[i+1]*smpi_datatype_get_extent(type_c->old_type);
754 noncontiguous_indexed_char += type_c->block_lengths[i]*smpi_datatype_get_extent(type_c->old_type);
756 noncontiguous_indexed=(void*)noncontiguous_indexed_char;
760 void free_indexed(MPI_Datatype* type){
761 if((*type)->in_use==0){
762 xbt_free(((s_smpi_mpi_indexed_t *)(*type)->substruct)->block_lengths);
763 xbt_free(((s_smpi_mpi_indexed_t *)(*type)->substruct)->block_indices);
765 smpi_datatype_unuse(((s_smpi_mpi_indexed_t *)(*type)->substruct)->old_type);
768 void use_indexed(MPI_Datatype* type){
769 smpi_datatype_use(((s_smpi_mpi_indexed_t *)(*type)->substruct)->old_type);
773 /* Create a Sub type indexed to be able to serialize and unserialize it the structure s_smpi_mpi_indexed_t is derived
774 * from s_smpi_subtype which required the functions unserialize and serialize */
775 s_smpi_mpi_indexed_t* smpi_datatype_indexed_create( int* block_lengths, int* block_indices, int block_count,
776 MPI_Datatype old_type, int size_oldtype){
777 s_smpi_mpi_indexed_t *new_t= xbt_new(s_smpi_mpi_indexed_t,1);
778 new_t->base.serialize = &serialize_indexed;
779 new_t->base.unserialize = &unserialize_indexed;
780 new_t->base.subtype_free = &free_indexed;
781 new_t->base.subtype_use = &use_indexed;
782 new_t->block_lengths= xbt_new(int, block_count);
783 new_t->block_indices= xbt_new(int, block_count);
785 for(i=0;i<block_count;i++){
786 new_t->block_lengths[i]=block_lengths[i];
787 new_t->block_indices[i]=block_indices[i];
789 new_t->block_count = block_count;
790 smpi_datatype_use(old_type);
791 new_t->old_type = old_type;
792 new_t->size_oldtype = size_oldtype;
796 int smpi_datatype_indexed(int count, int* blocklens, int* indices, MPI_Datatype old_type, MPI_Datatype* new_type)
805 lb=indices[0]*smpi_datatype_get_extent(old_type);
806 ub=indices[0]*smpi_datatype_get_extent(old_type) + blocklens[0]*smpi_datatype_ub(old_type);
809 for(i=0; i< count; i++){
812 size += blocklens[i];
814 if(indices[i]*smpi_datatype_get_extent(old_type)+smpi_datatype_lb(old_type)<lb)
815 lb = indices[i]*smpi_datatype_get_extent(old_type)+smpi_datatype_lb(old_type);
816 if(indices[i]*smpi_datatype_get_extent(old_type)+blocklens[i]*smpi_datatype_ub(old_type)>ub)
817 ub = indices[i]*smpi_datatype_get_extent(old_type)+blocklens[i]*smpi_datatype_ub(old_type);
819 if ( (i< count -1) && (indices[i]+blocklens[i] != indices[i+1]) )contiguous=0;
821 if (old_type->sizeof_substruct != 0)
825 s_smpi_mpi_indexed_t* subtype = smpi_datatype_indexed_create( blocklens, indices, count, old_type,
826 smpi_datatype_size(old_type));
827 smpi_datatype_create(new_type, size * smpi_datatype_size(old_type),lb,ub,sizeof(s_smpi_mpi_indexed_t), subtype, DT_FLAG_DATA);
829 s_smpi_mpi_contiguous_t* subtype = smpi_datatype_contiguous_create( lb, size, old_type,
830 smpi_datatype_size(old_type));
831 smpi_datatype_create(new_type, size * smpi_datatype_size(old_type), lb, ub, sizeof(s_smpi_mpi_contiguous_t), subtype,
832 DT_FLAG_DATA|DT_FLAG_CONTIGUOUS);
837 /* Hindexed Implementation - Indexed with indices in bytes */
839 /* Copies noncontiguous data into contiguous memory.
840 * @param contiguous_hindexed - output hindexed
841 * @param noncontiguous_hindexed - input hindexed
842 * @param type - pointer contening :
843 * - block_lengths - the width or height of blocked matrix
844 * - block_indices - indices of each data, in bytes
845 * - count - the number of rows of matrix
847 void serialize_hindexed( const void *noncontiguous_hindexed, void *contiguous_hindexed, int count, void *type)
849 s_smpi_mpi_hindexed_t* type_c = (s_smpi_mpi_hindexed_t*)type;
851 char* contiguous_hindexed_char = (char*)contiguous_hindexed;
852 char* noncontiguous_hindexed_char = (char*)noncontiguous_hindexed+ type_c->block_indices[0];
853 for(j=0; j<count;j++){
854 for (i = 0; i < type_c->block_count; i++) {
855 if (type_c->old_type->sizeof_substruct == 0)
856 memcpy(contiguous_hindexed_char, noncontiguous_hindexed_char, type_c->block_lengths[i] * type_c->size_oldtype);
858 ((s_smpi_subtype_t*)type_c->old_type->substruct)->serialize( noncontiguous_hindexed_char,
859 contiguous_hindexed_char,
860 type_c->block_lengths[i],
861 type_c->old_type->substruct);
863 contiguous_hindexed_char += type_c->block_lengths[i]*type_c->size_oldtype;
864 if (i<type_c->block_count-1)
865 noncontiguous_hindexed_char = (char*)noncontiguous_hindexed + type_c->block_indices[i+1];
867 noncontiguous_hindexed_char += type_c->block_lengths[i]*smpi_datatype_get_extent(type_c->old_type);
869 noncontiguous_hindexed=(void*)noncontiguous_hindexed_char;
872 /* Copies contiguous data into noncontiguous memory.
873 * @param noncontiguous_hindexed - output hindexed
874 * @param contiguous_hindexed - input hindexed
875 * @param type - pointer contening :
876 * - block_lengths - the width or height of blocked matrix
877 * - block_indices - indices of each data, in bytes
878 * - count - the number of rows of matrix
880 void unserialize_hindexed( const void *contiguous_hindexed, void *noncontiguous_hindexed, int count, void *type,
883 s_smpi_mpi_hindexed_t* type_c = (s_smpi_mpi_hindexed_t*)type;
886 char* contiguous_hindexed_char = (char*)contiguous_hindexed;
887 char* noncontiguous_hindexed_char = (char*)noncontiguous_hindexed+ type_c->block_indices[0];
888 for(j=0; j<count;j++){
889 for (i = 0; i < type_c->block_count; i++) {
890 if (type_c->old_type->sizeof_substruct == 0)
891 smpi_op_apply(op, contiguous_hindexed_char, noncontiguous_hindexed_char, &type_c->block_lengths[i],
893 /*memcpy(noncontiguous_hindexed_char,contiguous_hindexed_char,type_c->block_lengths[i]*type_c->size_oldtype);*/
895 ((s_smpi_subtype_t*)type_c->old_type->substruct)->unserialize( contiguous_hindexed_char,
896 noncontiguous_hindexed_char,
897 type_c->block_lengths[i],
898 type_c->old_type->substruct, op);
900 contiguous_hindexed_char += type_c->block_lengths[i]*type_c->size_oldtype;
901 if (i<type_c->block_count-1)
902 noncontiguous_hindexed_char = (char*)noncontiguous_hindexed + type_c->block_indices[i+1];
904 noncontiguous_hindexed_char += type_c->block_lengths[i]*smpi_datatype_get_extent(type_c->old_type);
906 noncontiguous_hindexed=(void*)noncontiguous_hindexed_char;
910 void free_hindexed(MPI_Datatype* type){
911 if((*type)->in_use==0){
912 xbt_free(((s_smpi_mpi_hindexed_t *)(*type)->substruct)->block_lengths);
913 xbt_free(((s_smpi_mpi_hindexed_t *)(*type)->substruct)->block_indices);
915 smpi_datatype_unuse(((s_smpi_mpi_indexed_t *)(*type)->substruct)->old_type);
918 void use_hindexed(MPI_Datatype* type){
919 smpi_datatype_use(((s_smpi_mpi_hindexed_t *)(*type)->substruct)->old_type);
922 /* Create a Sub type hindexed to be able to serialize and unserialize it the structure s_smpi_mpi_hindexed_t is derived
923 * from s_smpi_subtype which required the functions unserialize and serialize
925 s_smpi_mpi_hindexed_t* smpi_datatype_hindexed_create( int* block_lengths, MPI_Aint* block_indices, int block_count,
926 MPI_Datatype old_type, int size_oldtype){
927 s_smpi_mpi_hindexed_t *new_t= xbt_new(s_smpi_mpi_hindexed_t,1);
928 new_t->base.serialize = &serialize_hindexed;
929 new_t->base.unserialize = &unserialize_hindexed;
930 new_t->base.subtype_free = &free_hindexed;
931 new_t->base.subtype_use = &use_hindexed;
932 new_t->block_lengths= xbt_new(int, block_count);
933 new_t->block_indices= xbt_new(MPI_Aint, block_count);
935 for(i=0;i<block_count;i++){
936 new_t->block_lengths[i]=block_lengths[i];
937 new_t->block_indices[i]=block_indices[i];
939 new_t->block_count = block_count;
940 new_t->old_type = old_type;
941 smpi_datatype_use(old_type);
942 new_t->size_oldtype = size_oldtype;
946 int smpi_datatype_hindexed(int count, int* blocklens, MPI_Aint* indices, MPI_Datatype old_type, MPI_Datatype* new_type)
955 lb=indices[0] + smpi_datatype_lb(old_type);
956 ub=indices[0] + blocklens[0]*smpi_datatype_ub(old_type);
958 for(i=0; i< count; i++){
961 size += blocklens[i];
963 if(indices[i]+smpi_datatype_lb(old_type)<lb) lb = indices[i]+smpi_datatype_lb(old_type);
964 if(indices[i]+blocklens[i]*smpi_datatype_ub(old_type)>ub) 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 = (s_smpi_mpi_struct_t*)type;
999 char* contiguous_struct_char = (char*)contiguous_struct;
1000 char* noncontiguous_struct_char = (char*)noncontiguous_struct+ type_c->block_indices[0];
1001 for(j=0; j<count;j++){
1002 for (i = 0; i < type_c->block_count; i++) {
1003 if (type_c->old_types[i]->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 ((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 = (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=(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 = (s_smpi_mpi_struct_t*)type;
1036 char* contiguous_struct_char = (char*)contiguous_struct;
1037 char* noncontiguous_struct_char = (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]);
1043 /*memcpy(noncontiguous_struct_char,
1044 contiguous_struct_char, type_c->block_lengths[i] * smpi_datatype_size(type_c->old_types[i]));*/
1046 ((s_smpi_subtype_t*)type_c->old_types[i]->substruct)->unserialize( contiguous_struct_char,
1047 noncontiguous_struct_char,
1048 type_c->block_lengths[i],
1049 type_c->old_types[i]->substruct, op);
1051 contiguous_struct_char += type_c->block_lengths[i]*smpi_datatype_size(type_c->old_types[i]);
1052 if (i<type_c->block_count-1)
1053 noncontiguous_struct_char = (char*)noncontiguous_struct + type_c->block_indices[i+1];
1055 noncontiguous_struct_char += type_c->block_lengths[i]*smpi_datatype_get_extent(type_c->old_types[i]);
1057 noncontiguous_struct=(void*)noncontiguous_struct_char;
1061 void free_struct(MPI_Datatype* type){
1063 for (i = 0; i < ((s_smpi_mpi_struct_t *)(*type)->substruct)->block_count; i++)
1064 smpi_datatype_unuse(((s_smpi_mpi_struct_t *)(*type)->substruct)->old_types[i]);
1065 if((*type)->in_use==0){
1066 xbt_free(((s_smpi_mpi_struct_t *)(*type)->substruct)->block_lengths);
1067 xbt_free(((s_smpi_mpi_struct_t *)(*type)->substruct)->block_indices);
1068 xbt_free(((s_smpi_mpi_struct_t *)(*type)->substruct)->old_types);
1072 void use_struct(MPI_Datatype* type){
1074 for (i = 0; i < ((s_smpi_mpi_struct_t *)(*type)->substruct)->block_count; i++)
1075 smpi_datatype_use(((s_smpi_mpi_struct_t *)(*type)->substruct)->old_types[i]);
1078 /* Create a Sub type struct to be able to serialize and unserialize it the structure s_smpi_mpi_struct_t is derived
1079 * from s_smpi_subtype which required the functions unserialize and serialize
1081 s_smpi_mpi_struct_t* smpi_datatype_struct_create( int* block_lengths, MPI_Aint* block_indices, int block_count,
1082 MPI_Datatype* old_types){
1083 s_smpi_mpi_struct_t *new_t= xbt_new(s_smpi_mpi_struct_t,1);
1084 new_t->base.serialize = &serialize_struct;
1085 new_t->base.unserialize = &unserialize_struct;
1086 new_t->base.subtype_free = &free_struct;
1087 new_t->base.subtype_use = &use_struct;
1088 new_t->block_lengths= xbt_new(int, block_count);
1089 new_t->block_indices= xbt_new(MPI_Aint, block_count);
1090 new_t->old_types= xbt_new(MPI_Datatype, block_count);
1092 for(i=0;i<block_count;i++){
1093 new_t->block_lengths[i]=block_lengths[i];
1094 new_t->block_indices[i]=block_indices[i];
1095 new_t->old_types[i]=old_types[i];
1096 smpi_datatype_use(new_t->old_types[i]);
1098 new_t->block_count = block_count;
1102 int smpi_datatype_struct(int count, int* blocklens, MPI_Aint* indices, MPI_Datatype* old_types, MPI_Datatype* new_type)
1111 lb=indices[0] + smpi_datatype_lb(old_types[0]);
1112 ub=indices[0] + blocklens[0]*smpi_datatype_ub(old_types[0]);
1116 for(i=0; i< count; i++){
1119 if (old_types[i]->sizeof_substruct != 0)
1122 size += blocklens[i]*smpi_datatype_size(old_types[i]);
1123 if (old_types[i]==MPI_LB){
1127 if (old_types[i]==MPI_UB){
1132 if(!forced_lb && indices[i]+smpi_datatype_lb(old_types[i])<lb) lb = indices[i];
1133 if(!forced_ub && indices[i]+blocklens[i]*smpi_datatype_ub(old_types[i])>ub)
1134 ub = indices[i]+blocklens[i]*smpi_datatype_ub(old_types[i]);
1136 if ( (i< count -1) && (indices[i]+blocklens[i]*static_cast<int>(smpi_datatype_size(old_types[i])) != indices[i+1]) )
1141 s_smpi_mpi_struct_t* subtype = smpi_datatype_struct_create( blocklens, indices, count, old_types);
1143 smpi_datatype_create(new_type, size, lb, ub,sizeof(s_smpi_mpi_struct_t), subtype, DT_FLAG_DATA);
1145 s_smpi_mpi_contiguous_t* subtype = smpi_datatype_contiguous_create( lb, size, MPI_CHAR, 1);
1146 smpi_datatype_create(new_type, size, lb, ub,1, subtype, DT_FLAG_DATA|DT_FLAG_CONTIGUOUS);
1151 void smpi_datatype_commit(MPI_Datatype *datatype)
1153 (*datatype)->flags= ((*datatype)->flags | DT_FLAG_COMMITED);
1156 typedef struct s_smpi_mpi_op {
1157 MPI_User_function *func;
1161 #define MAX_OP(a, b) (b) = (a) < (b) ? (b) : (a)
1162 #define MIN_OP(a, b) (b) = (a) < (b) ? (a) : (b)
1163 #define SUM_OP(a, b) (b) += (a)
1164 #define PROD_OP(a, b) (b) *= (a)
1165 #define LAND_OP(a, b) (b) = (a) && (b)
1166 #define LOR_OP(a, b) (b) = (a) || (b)
1167 #define LXOR_OP(a, b) (b) = (!(a) && (b)) || ((a) && !(b))
1168 #define BAND_OP(a, b) (b) &= (a)
1169 #define BOR_OP(a, b) (b) |= (a)
1170 #define BXOR_OP(a, b) (b) ^= (a)
1171 #define MAXLOC_OP(a, b) (b) = (a.value) < (b.value) ? (b) : (a)
1172 #define MINLOC_OP(a, b) (b) = (a.value) < (b.value) ? (a) : (b)
1174 #define APPLY_FUNC(a, b, length, type, func) \
1177 type* x = (type*)(a); \
1178 type* y = (type*)(b); \
1179 for(i = 0; i < *(length); i++) { \
1184 static void max_func(void *a, void *b, int *length, MPI_Datatype * datatype)
1186 if (*datatype == MPI_CHAR) {
1187 APPLY_FUNC(a, b, length, char, MAX_OP);
1188 } else if (*datatype == MPI_SHORT) {
1189 APPLY_FUNC(a, b, length, short, MAX_OP);
1190 } else if (*datatype == MPI_INT) {
1191 APPLY_FUNC(a, b, length, int, MAX_OP);
1192 } else if (*datatype == MPI_LONG) {
1193 APPLY_FUNC(a, b, length, long, MAX_OP);
1194 } else if (*datatype == MPI_UNSIGNED_SHORT) {
1195 APPLY_FUNC(a, b, length, unsigned short, MAX_OP);
1196 } else if (*datatype == MPI_UNSIGNED) {
1197 APPLY_FUNC(a, b, length, unsigned int, MAX_OP);
1198 } else if (*datatype == MPI_UNSIGNED_LONG) {
1199 APPLY_FUNC(a, b, length, unsigned long, MAX_OP);
1200 } else if (*datatype == MPI_UNSIGNED_CHAR) {
1201 APPLY_FUNC(a, b, length, unsigned char, MAX_OP);
1202 } else if (*datatype == MPI_FLOAT) {
1203 APPLY_FUNC(a, b, length, float, MAX_OP);
1204 } else if (*datatype == MPI_DOUBLE) {
1205 APPLY_FUNC(a, b, length, double, MAX_OP);
1206 } else if (*datatype == MPI_LONG_DOUBLE) {
1207 APPLY_FUNC(a, b, length, long double, MAX_OP);
1211 static void min_func(void *a, void *b, int *length, MPI_Datatype * datatype)
1213 if (*datatype == MPI_CHAR) {
1214 APPLY_FUNC(a, b, length, char, MIN_OP);
1215 } else if (*datatype == MPI_SHORT) {
1216 APPLY_FUNC(a, b, length, short, MIN_OP);
1217 } else if (*datatype == MPI_INT) {
1218 APPLY_FUNC(a, b, length, int, MIN_OP);
1219 } else if (*datatype == MPI_LONG) {
1220 APPLY_FUNC(a, b, length, long, MIN_OP);
1221 } else if (*datatype == MPI_UNSIGNED_SHORT) {
1222 APPLY_FUNC(a, b, length, unsigned short, MIN_OP);
1223 } else if (*datatype == MPI_UNSIGNED) {
1224 APPLY_FUNC(a, b, length, unsigned int, MIN_OP);
1225 } else if (*datatype == MPI_UNSIGNED_LONG) {
1226 APPLY_FUNC(a, b, length, unsigned long, MIN_OP);
1227 } else if (*datatype == MPI_UNSIGNED_CHAR) {
1228 APPLY_FUNC(a, b, length, unsigned char, MIN_OP);
1229 } else if (*datatype == MPI_FLOAT) {
1230 APPLY_FUNC(a, b, length, float, MIN_OP);
1231 } else if (*datatype == MPI_DOUBLE) {
1232 APPLY_FUNC(a, b, length, double, MIN_OP);
1233 } else if (*datatype == MPI_LONG_DOUBLE) {
1234 APPLY_FUNC(a, b, length, long double, MIN_OP);
1238 static void sum_func(void *a, void *b, int *length, MPI_Datatype * datatype)
1240 if (*datatype == MPI_CHAR) {
1241 APPLY_FUNC(a, b, length, char, SUM_OP);
1242 } else if (*datatype == MPI_SHORT) {
1243 APPLY_FUNC(a, b, length, short, SUM_OP);
1244 } else if (*datatype == MPI_INT) {
1245 APPLY_FUNC(a, b, length, int, SUM_OP);
1246 } else if (*datatype == MPI_LONG) {
1247 APPLY_FUNC(a, b, length, long, SUM_OP);
1248 } else if (*datatype == MPI_UNSIGNED_SHORT) {
1249 APPLY_FUNC(a, b, length, unsigned short, SUM_OP);
1250 } else if (*datatype == MPI_UNSIGNED) {
1251 APPLY_FUNC(a, b, length, unsigned int, SUM_OP);
1252 } else if (*datatype == MPI_UNSIGNED_LONG) {
1253 APPLY_FUNC(a, b, length, unsigned long, SUM_OP);
1254 } else if (*datatype == MPI_UNSIGNED_CHAR) {
1255 APPLY_FUNC(a, b, length, unsigned char, SUM_OP);
1256 } else if (*datatype == MPI_FLOAT) {
1257 APPLY_FUNC(a, b, length, float, SUM_OP);
1258 } else if (*datatype == MPI_DOUBLE) {
1259 APPLY_FUNC(a, b, length, double, SUM_OP);
1260 } else if (*datatype == MPI_LONG_DOUBLE) {
1261 APPLY_FUNC(a, b, length, long double, SUM_OP);
1262 } else if (*datatype == MPI_C_FLOAT_COMPLEX) {
1263 APPLY_FUNC(a, b, length, float _Complex, SUM_OP);
1264 } else if (*datatype == MPI_C_DOUBLE_COMPLEX) {
1265 APPLY_FUNC(a, b, length, double _Complex, SUM_OP);
1266 } else if (*datatype == MPI_C_LONG_DOUBLE_COMPLEX) {
1267 APPLY_FUNC(a, b, length, long double _Complex, SUM_OP);
1271 static void prod_func(void *a, void *b, int *length, MPI_Datatype * datatype)
1273 if (*datatype == MPI_CHAR) {
1274 APPLY_FUNC(a, b, length, char, PROD_OP);
1275 } else if (*datatype == MPI_SHORT) {
1276 APPLY_FUNC(a, b, length, short, PROD_OP);
1277 } else if (*datatype == MPI_INT) {
1278 APPLY_FUNC(a, b, length, int, PROD_OP);
1279 } else if (*datatype == MPI_LONG) {
1280 APPLY_FUNC(a, b, length, long, PROD_OP);
1281 } else if (*datatype == MPI_UNSIGNED_SHORT) {
1282 APPLY_FUNC(a, b, length, unsigned short, PROD_OP);
1283 } else if (*datatype == MPI_UNSIGNED) {
1284 APPLY_FUNC(a, b, length, unsigned int, PROD_OP);
1285 } else if (*datatype == MPI_UNSIGNED_LONG) {
1286 APPLY_FUNC(a, b, length, unsigned long, PROD_OP);
1287 } else if (*datatype == MPI_UNSIGNED_CHAR) {
1288 APPLY_FUNC(a, b, length, unsigned char, PROD_OP);
1289 } else if (*datatype == MPI_FLOAT) {
1290 APPLY_FUNC(a, b, length, float, PROD_OP);
1291 } else if (*datatype == MPI_DOUBLE) {
1292 APPLY_FUNC(a, b, length, double, PROD_OP);
1293 } else if (*datatype == MPI_LONG_DOUBLE) {
1294 APPLY_FUNC(a, b, length, long double, PROD_OP);
1295 } else if (*datatype == MPI_C_FLOAT_COMPLEX) {
1296 APPLY_FUNC(a, b, length, float _Complex, PROD_OP);
1297 } else if (*datatype == MPI_C_DOUBLE_COMPLEX) {
1298 APPLY_FUNC(a, b, length, double _Complex, PROD_OP);
1299 } else if (*datatype == MPI_C_LONG_DOUBLE_COMPLEX) {
1300 APPLY_FUNC(a, b, length, long double _Complex, PROD_OP);
1304 static void land_func(void *a, void *b, int *length, MPI_Datatype * datatype)
1306 if (*datatype == MPI_CHAR) {
1307 APPLY_FUNC(a, b, length, char, LAND_OP);
1308 } else if (*datatype == MPI_SHORT) {
1309 APPLY_FUNC(a, b, length, short, LAND_OP);
1310 } else if (*datatype == MPI_INT) {
1311 APPLY_FUNC(a, b, length, int, LAND_OP);
1312 } else if (*datatype == MPI_LONG) {
1313 APPLY_FUNC(a, b, length, long, LAND_OP);
1314 } else if (*datatype == MPI_UNSIGNED_SHORT) {
1315 APPLY_FUNC(a, b, length, unsigned short, LAND_OP);
1316 } else if (*datatype == MPI_UNSIGNED) {
1317 APPLY_FUNC(a, b, length, unsigned int, LAND_OP);
1318 } else if (*datatype == MPI_UNSIGNED_LONG) {
1319 APPLY_FUNC(a, b, length, unsigned long, LAND_OP);
1320 } else if (*datatype == MPI_UNSIGNED_CHAR) {
1321 APPLY_FUNC(a, b, length, unsigned char, LAND_OP);
1322 } else if (*datatype == MPI_C_BOOL) {
1323 APPLY_FUNC(a, b, length, bool, LAND_OP);
1327 static void lor_func(void *a, void *b, int *length, MPI_Datatype * datatype)
1329 if (*datatype == MPI_CHAR) {
1330 APPLY_FUNC(a, b, length, char, LOR_OP);
1331 } else if (*datatype == MPI_SHORT) {
1332 APPLY_FUNC(a, b, length, short, LOR_OP);
1333 } else if (*datatype == MPI_INT) {
1334 APPLY_FUNC(a, b, length, int, LOR_OP);
1335 } else if (*datatype == MPI_LONG) {
1336 APPLY_FUNC(a, b, length, long, LOR_OP);
1337 } else if (*datatype == MPI_UNSIGNED_SHORT) {
1338 APPLY_FUNC(a, b, length, unsigned short, LOR_OP);
1339 } else if (*datatype == MPI_UNSIGNED) {
1340 APPLY_FUNC(a, b, length, unsigned int, LOR_OP);
1341 } else if (*datatype == MPI_UNSIGNED_LONG) {
1342 APPLY_FUNC(a, b, length, unsigned long, LOR_OP);
1343 } else if (*datatype == MPI_UNSIGNED_CHAR) {
1344 APPLY_FUNC(a, b, length, unsigned char, LOR_OP);
1345 } else if (*datatype == MPI_C_BOOL) {
1346 APPLY_FUNC(a, b, length, bool, LOR_OP);
1350 static void lxor_func(void *a, void *b, int *length, MPI_Datatype * datatype)
1352 if (*datatype == MPI_CHAR) {
1353 APPLY_FUNC(a, b, length, char, LXOR_OP);
1354 } else if (*datatype == MPI_SHORT) {
1355 APPLY_FUNC(a, b, length, short, LXOR_OP);
1356 } else if (*datatype == MPI_INT) {
1357 APPLY_FUNC(a, b, length, int, LXOR_OP);
1358 } else if (*datatype == MPI_LONG) {
1359 APPLY_FUNC(a, b, length, long, LXOR_OP);
1360 } else if (*datatype == MPI_UNSIGNED_SHORT) {
1361 APPLY_FUNC(a, b, length, unsigned short, LXOR_OP);
1362 } else if (*datatype == MPI_UNSIGNED) {
1363 APPLY_FUNC(a, b, length, unsigned int, LXOR_OP);
1364 } else if (*datatype == MPI_UNSIGNED_LONG) {
1365 APPLY_FUNC(a, b, length, unsigned long, LXOR_OP);
1366 } else if (*datatype == MPI_UNSIGNED_CHAR) {
1367 APPLY_FUNC(a, b, length, unsigned char, LXOR_OP);
1368 } else if (*datatype == MPI_C_BOOL) {
1369 APPLY_FUNC(a, b, length, bool, LXOR_OP);
1373 static void band_func(void *a, void *b, int *length, MPI_Datatype * datatype)
1375 if (*datatype == MPI_CHAR) {
1376 APPLY_FUNC(a, b, length, char, BAND_OP);
1377 }else if (*datatype == MPI_SHORT) {
1378 APPLY_FUNC(a, b, length, short, BAND_OP);
1379 } else if (*datatype == MPI_INT) {
1380 APPLY_FUNC(a, b, length, int, BAND_OP);
1381 } else if (*datatype == MPI_LONG) {
1382 APPLY_FUNC(a, b, length, long, BAND_OP);
1383 } else if (*datatype == MPI_UNSIGNED_SHORT) {
1384 APPLY_FUNC(a, b, length, unsigned short, BAND_OP);
1385 } else if (*datatype == MPI_UNSIGNED) {
1386 APPLY_FUNC(a, b, length, unsigned int, BAND_OP);
1387 } else if (*datatype == MPI_UNSIGNED_LONG) {
1388 APPLY_FUNC(a, b, length, unsigned long, BAND_OP);
1389 } else if (*datatype == MPI_UNSIGNED_CHAR) {
1390 APPLY_FUNC(a, b, length, unsigned char, BAND_OP);
1391 } else if (*datatype == MPI_BYTE) {
1392 APPLY_FUNC(a, b, length, uint8_t, BAND_OP);
1396 static void bor_func(void *a, void *b, int *length, MPI_Datatype * datatype)
1398 if (*datatype == MPI_CHAR) {
1399 APPLY_FUNC(a, b, length, char, BOR_OP);
1400 } else if (*datatype == MPI_SHORT) {
1401 APPLY_FUNC(a, b, length, short, BOR_OP);
1402 } else if (*datatype == MPI_INT) {
1403 APPLY_FUNC(a, b, length, int, BOR_OP);
1404 } else if (*datatype == MPI_LONG) {
1405 APPLY_FUNC(a, b, length, long, BOR_OP);
1406 } else if (*datatype == MPI_UNSIGNED_SHORT) {
1407 APPLY_FUNC(a, b, length, unsigned short, BOR_OP);
1408 } else if (*datatype == MPI_UNSIGNED) {
1409 APPLY_FUNC(a, b, length, unsigned int, BOR_OP);
1410 } else if (*datatype == MPI_UNSIGNED_LONG) {
1411 APPLY_FUNC(a, b, length, unsigned long, BOR_OP);
1412 } else if (*datatype == MPI_UNSIGNED_CHAR) {
1413 APPLY_FUNC(a, b, length, unsigned char, BOR_OP);
1414 } else if (*datatype == MPI_BYTE) {
1415 APPLY_FUNC(a, b, length, uint8_t, BOR_OP);
1419 static void bxor_func(void *a, void *b, int *length, MPI_Datatype * datatype)
1421 if (*datatype == MPI_CHAR) {
1422 APPLY_FUNC(a, b, length, char, BXOR_OP);
1423 } else if (*datatype == MPI_SHORT) {
1424 APPLY_FUNC(a, b, length, short, BXOR_OP);
1425 } else if (*datatype == MPI_INT) {
1426 APPLY_FUNC(a, b, length, int, BXOR_OP);
1427 } else if (*datatype == MPI_LONG) {
1428 APPLY_FUNC(a, b, length, long, BXOR_OP);
1429 } else if (*datatype == MPI_UNSIGNED_SHORT) {
1430 APPLY_FUNC(a, b, length, unsigned short, BXOR_OP);
1431 } else if (*datatype == MPI_UNSIGNED) {
1432 APPLY_FUNC(a, b, length, unsigned int, BXOR_OP);
1433 } else if (*datatype == MPI_UNSIGNED_LONG) {
1434 APPLY_FUNC(a, b, length, unsigned long, BXOR_OP);
1435 } else if (*datatype == MPI_UNSIGNED_CHAR) {
1436 APPLY_FUNC(a, b, length, unsigned char, BXOR_OP);
1437 } else if (*datatype == MPI_BYTE) {
1438 APPLY_FUNC(a, b, length, uint8_t, BXOR_OP);
1442 static void minloc_func(void *a, void *b, int *length, MPI_Datatype * datatype)
1444 if (*datatype == MPI_FLOAT_INT) {
1445 APPLY_FUNC(a, b, length, float_int, MINLOC_OP);
1446 } else if (*datatype == MPI_LONG_INT) {
1447 APPLY_FUNC(a, b, length, long_int, MINLOC_OP);
1448 } else if (*datatype == MPI_DOUBLE_INT) {
1449 APPLY_FUNC(a, b, length, double_int, MINLOC_OP);
1450 } else if (*datatype == MPI_SHORT_INT) {
1451 APPLY_FUNC(a, b, length, short_int, MINLOC_OP);
1452 } else if (*datatype == MPI_2LONG) {
1453 APPLY_FUNC(a, b, length, long_long, MINLOC_OP);
1454 } else if (*datatype == MPI_2INT) {
1455 APPLY_FUNC(a, b, length, int_int, MINLOC_OP);
1456 } else if (*datatype == MPI_LONG_DOUBLE_INT) {
1457 APPLY_FUNC(a, b, length, long_double_int, MINLOC_OP);
1458 } else if (*datatype == MPI_2FLOAT) {
1459 APPLY_FUNC(a, b, length, float_float, MINLOC_OP);
1460 } else if (*datatype == MPI_2DOUBLE) {
1461 APPLY_FUNC(a, b, length, double_double, MINLOC_OP);
1465 static void maxloc_func(void *a, void *b, int *length, MPI_Datatype * datatype)
1467 if (*datatype == MPI_FLOAT_INT) {
1468 APPLY_FUNC(a, b, length, float_int, MAXLOC_OP);
1469 } else if (*datatype == MPI_LONG_INT) {
1470 APPLY_FUNC(a, b, length, long_int, MAXLOC_OP);
1471 } else if (*datatype == MPI_DOUBLE_INT) {
1472 APPLY_FUNC(a, b, length, double_int, MAXLOC_OP);
1473 } else if (*datatype == MPI_SHORT_INT) {
1474 APPLY_FUNC(a, b, length, short_int, MAXLOC_OP);
1475 } else if (*datatype == MPI_2LONG) {
1476 APPLY_FUNC(a, b, length, long_long, MAXLOC_OP);
1477 } else if (*datatype == MPI_2INT) {
1478 APPLY_FUNC(a, b, length, int_int, MAXLOC_OP);
1479 } else if (*datatype == MPI_LONG_DOUBLE_INT) {
1480 APPLY_FUNC(a, b, length, long_double_int, MAXLOC_OP);
1481 } else if (*datatype == MPI_2FLOAT) {
1482 APPLY_FUNC(a, b, length, float_float, MAXLOC_OP);
1483 } else if (*datatype == MPI_2DOUBLE) {
1484 APPLY_FUNC(a, b, length, double_double, MAXLOC_OP);
1488 static void replace_func(void *a, void *b, int *length, MPI_Datatype * datatype)
1490 memcpy(b, a, *length * smpi_datatype_size(*datatype));
1493 #define CREATE_MPI_OP(name, func) \
1494 static s_smpi_mpi_op_t mpi_##name = { &(func) /* func */, true }; \
1495 MPI_Op name = &mpi_##name;
1497 CREATE_MPI_OP(MPI_MAX, max_func);
1498 CREATE_MPI_OP(MPI_MIN, min_func);
1499 CREATE_MPI_OP(MPI_SUM, sum_func);
1500 CREATE_MPI_OP(MPI_PROD, prod_func);
1501 CREATE_MPI_OP(MPI_LAND, land_func);
1502 CREATE_MPI_OP(MPI_LOR, lor_func);
1503 CREATE_MPI_OP(MPI_LXOR, lxor_func);
1504 CREATE_MPI_OP(MPI_BAND, band_func);
1505 CREATE_MPI_OP(MPI_BOR, bor_func);
1506 CREATE_MPI_OP(MPI_BXOR, bxor_func);
1507 CREATE_MPI_OP(MPI_MAXLOC, maxloc_func);
1508 CREATE_MPI_OP(MPI_MINLOC, minloc_func);
1509 CREATE_MPI_OP(MPI_REPLACE, replace_func);
1511 MPI_Op smpi_op_new(MPI_User_function * function, int commute)
1514 op = xbt_new(s_smpi_mpi_op_t, 1);
1515 op->func = function;
1516 op-> is_commute = commute;
1520 int smpi_op_is_commute(MPI_Op op)
1522 return (op==MPI_OP_NULL) ? 1 : op-> is_commute;
1525 void smpi_op_destroy(MPI_Op op)
1530 void smpi_op_apply(MPI_Op op, void *invec, void *inoutvec, int *len, MPI_Datatype * datatype)
1535 if(smpi_privatize_global_variables){//we need to switch as the called function may silently touch global variables
1536 XBT_DEBUG("Applying operation, switch to the right data frame ");
1537 smpi_switch_data_segment(smpi_process_index());
1540 if(!smpi_process_get_replaying())
1541 op->func(invec, inoutvec, len, datatype);
1544 int smpi_type_attr_delete(MPI_Datatype type, int keyval){
1545 smpi_type_key_elem elem =
1546 static_cast<smpi_type_key_elem>(xbt_dict_get_or_null_ext(smpi_type_keyvals, (const char*)&keyval, sizeof(int)));
1549 if(elem->delete_fn!=MPI_NULL_DELETE_FN){
1552 if(smpi_type_attr_get(type, keyval, &value, &flag)==MPI_SUCCESS){
1553 int ret = elem->delete_fn(type, keyval, value, &flag);
1554 if(ret!=MPI_SUCCESS) return ret;
1557 if(type->attributes==NULL)
1560 xbt_dict_remove_ext(type->attributes, (const char*)&keyval, sizeof(int));
1564 int smpi_type_attr_get(MPI_Datatype type, int keyval, void* attr_value, int* flag){
1565 smpi_type_key_elem elem =
1566 static_cast<smpi_type_key_elem>(xbt_dict_get_or_null_ext(smpi_type_keyvals, (const char*)&keyval, sizeof(int)));
1570 if(type->attributes==NULL){
1575 *(void**)attr_value = xbt_dict_get_ext(type->attributes, (const char*)&keyval, sizeof(int));
1585 int smpi_type_attr_put(MPI_Datatype type, int keyval, void* attr_value){
1586 if(!smpi_type_keyvals)
1587 smpi_type_keyvals = xbt_dict_new();
1588 smpi_type_key_elem elem =
1589 static_cast<smpi_type_key_elem>(xbt_dict_get_or_null_ext(smpi_type_keyvals, (const char*)&keyval, sizeof(int)));
1594 smpi_type_attr_get(type, keyval, &value, &flag);
1595 if(flag && elem->delete_fn!=MPI_NULL_DELETE_FN){
1596 int ret = elem->delete_fn(type, keyval, value, &flag);
1597 if(ret!=MPI_SUCCESS) return ret;
1599 if(type->attributes==NULL)
1600 type->attributes=xbt_dict_new();
1602 xbt_dict_set_ext(type->attributes, (const char*)&keyval, sizeof(int), attr_value, NULL);
1606 int smpi_type_keyval_create(MPI_Type_copy_attr_function* copy_fn, MPI_Type_delete_attr_function* delete_fn, int* keyval,
1608 if(!smpi_type_keyvals)
1609 smpi_type_keyvals = xbt_dict_new();
1611 smpi_type_key_elem value = (smpi_type_key_elem) xbt_new0(s_smpi_mpi_type_key_elem_t,1);
1613 value->copy_fn=copy_fn;
1614 value->delete_fn=delete_fn;
1616 *keyval = type_keyval_id;
1617 xbt_dict_set_ext(smpi_type_keyvals,(const char*)keyval, sizeof(int),(void*)value, NULL);
1622 int smpi_type_keyval_free(int* keyval){
1623 smpi_type_key_elem elem =
1624 static_cast<smpi_type_key_elem>(xbt_dict_get_or_null_ext(smpi_type_keyvals, (const char*)keyval, sizeof(int)));
1628 xbt_dict_remove_ext(smpi_type_keyvals, (const char*)keyval, sizeof(int));
1633 int smpi_mpi_pack(void* inbuf, int incount, MPI_Datatype type, void* outbuf, int outcount, int* position,MPI_Comm comm){
1634 size_t size = smpi_datatype_size(type);
1635 if (outcount - *position < incount*static_cast<int>(size))
1636 return MPI_ERR_BUFFER;
1637 smpi_datatype_copy(inbuf, incount, type, (char*)outbuf + *position, outcount, MPI_CHAR);
1638 *position += incount * size;
1642 int smpi_mpi_unpack(void* inbuf, int insize, int* position, void* outbuf, int outcount, MPI_Datatype type,MPI_Comm comm){
1643 int size = static_cast<int>(smpi_datatype_size(type));
1644 if (outcount*size> insize)
1645 return MPI_ERR_BUFFER;
1646 smpi_datatype_copy((char*)inbuf + *position, insize, MPI_CHAR, outbuf, outcount, type);
1647 *position += outcount * size;