1 /* smpi_mpi_dt.c -- MPI primitives to handle datatypes */
2 /* Copyright (c) 2009-2017. The SimGrid Team. All rights reserved. */
4 /* This program is free software; you can redistribute it and/or modify it
5 * under the terms of the license (GNU LGPL) which comes with this package. */
12 #include "smpi_mpi_dt_private.h"
14 #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_homogeneous(nullptr);
189 xbt_dict_cursor_t cursor = nullptr;
194 xbt_dict_foreach (datatype->attributes, cursor, key, value_in) {
195 smpi_type_key_elem elem = static_cast<smpi_type_key_elem>(
196 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( 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 char* noncontiguous_vector_char = static_cast<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( 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 char* contiguous_vector_char = static_cast<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)
430 if(type->sizeof_substruct!=0){
431 static_cast<s_smpi_subtype_t *>((type)->substruct)->subtype_use(&type);
435 MC_ignore(&(type->in_use), sizeof(type->in_use));
439 void smpi_datatype_unuse(MPI_Datatype type)
441 if (type == MPI_DATATYPE_NULL)
444 if (type->in_use > 0)
447 if(type->sizeof_substruct!=0){
448 static_cast<s_smpi_subtype_t *>((type)->substruct)->subtype_free(&type);
451 if (type->in_use == 0)
452 smpi_datatype_free(&type);
456 MC_ignore(&(type->in_use), sizeof(type->in_use));
460 /*Contiguous Implementation*/
462 /* Copies noncontiguous data into contiguous memory.
463 * @param contiguous_hvector - output hvector
464 * @param noncontiguous_hvector - input hvector
465 * @param type - pointer contening :
466 * - stride - stride of between noncontiguous data, in bytes
467 * - block_length - the width or height of blocked matrix
468 * - count - the number of rows of matrix
470 void serialize_contiguous( void* noncontiguous_hvector, void *contiguous_hvector, int count, void *type)
472 s_smpi_mpi_contiguous_t* type_c = reinterpret_cast<s_smpi_mpi_contiguous_t*>(type);
473 char* contiguous_vector_char = static_cast<char*>(contiguous_hvector);
474 char* noncontiguous_vector_char = static_cast<char*>(noncontiguous_hvector)+type_c->lb;
475 memcpy(contiguous_vector_char, noncontiguous_vector_char, count* type_c->block_count * type_c->size_oldtype);
477 /* Copies contiguous data into noncontiguous memory.
478 * @param noncontiguous_vector - output hvector
479 * @param contiguous_vector - input hvector
480 * @param type - pointer contening :
481 * - stride - stride of between noncontiguous data, in bytes
482 * - block_length - the width or height of blocked matrix
483 * - count - the number of rows of matrix
485 void unserialize_contiguous(void* contiguous_vector, void *noncontiguous_vector, int count, void *type, MPI_Op op)
487 s_smpi_mpi_contiguous_t* type_c = reinterpret_cast<s_smpi_mpi_contiguous_t*>(type);
488 char* contiguous_vector_char = static_cast<char*>(contiguous_vector);
489 char* noncontiguous_vector_char = static_cast<char*>(noncontiguous_vector)+type_c->lb;
490 int n= count* type_c->block_count;
491 smpi_op_apply(op, contiguous_vector_char, noncontiguous_vector_char, &n, &type_c->old_type);
494 void free_contiguous(MPI_Datatype* d){
495 smpi_datatype_unuse(reinterpret_cast<s_smpi_mpi_contiguous_t*>((*d)->substruct)->old_type);
498 void use_contiguous(MPI_Datatype* d){
499 smpi_datatype_use(reinterpret_cast<s_smpi_mpi_contiguous_t*>((*d)->substruct)->old_type);
502 /* Create a Sub type contiguous to be able to serialize and unserialize it the structure s_smpi_mpi_contiguous_t is
503 * derived from s_smpi_subtype which required the functions unserialize and serialize */
504 s_smpi_mpi_contiguous_t* smpi_datatype_contiguous_create( MPI_Aint lb, int block_count, MPI_Datatype old_type,
508 s_smpi_mpi_contiguous_t *new_t= xbt_new(s_smpi_mpi_contiguous_t,1);
509 new_t->base.serialize = &serialize_contiguous;
510 new_t->base.unserialize = &unserialize_contiguous;
511 new_t->base.subtype_free = &free_contiguous;
512 new_t->base.subtype_use = &use_contiguous;
514 new_t->block_count = block_count;
515 new_t->old_type = old_type;
516 smpi_datatype_use(old_type);
517 new_t->size_oldtype = size_oldtype;
518 smpi_datatype_use(old_type);
522 int smpi_datatype_contiguous(int count, MPI_Datatype old_type, MPI_Datatype* new_type, MPI_Aint lb)
525 if(old_type->sizeof_substruct){
526 //handle this case as a hvector with stride equals to the extent of the datatype
527 return smpi_datatype_hvector(count, 1, smpi_datatype_get_extent(old_type), old_type, new_type);
530 s_smpi_mpi_contiguous_t* subtype = smpi_datatype_contiguous_create( lb, count, old_type,smpi_datatype_size(old_type));
532 smpi_datatype_create(new_type, count * smpi_datatype_size(old_type),lb,lb + count * smpi_datatype_size(old_type),
533 sizeof(s_smpi_mpi_contiguous_t),subtype, DT_FLAG_CONTIGUOUS);
538 int smpi_datatype_vector(int count, int blocklen, int stride, MPI_Datatype old_type, MPI_Datatype* new_type)
546 lb=smpi_datatype_lb(old_type);
547 ub=((count-1)*stride+blocklen-1)*smpi_datatype_get_extent(old_type)+smpi_datatype_ub(old_type);
549 if(old_type->sizeof_substruct != 0 || stride != blocklen){
551 s_smpi_mpi_vector_t* subtype = smpi_datatype_vector_create(stride, blocklen, count, old_type,
552 smpi_datatype_size(old_type));
553 smpi_datatype_create(new_type, count * (blocklen) * smpi_datatype_size(old_type), lb, ub, sizeof(s_smpi_mpi_vector_t), subtype,
557 /* in this situation the data are contiguous thus it's not required to serialize and unserialize it*/
558 smpi_datatype_create(new_type, count * blocklen * smpi_datatype_size(old_type), 0, ((count -1) * stride + blocklen)*
559 smpi_datatype_size(old_type), 0, nullptr, DT_FLAG_VECTOR|DT_FLAG_CONTIGUOUS);
565 void free_vector(MPI_Datatype* d){
566 smpi_datatype_unuse(reinterpret_cast<s_smpi_mpi_indexed_t*>((*d)->substruct)->old_type);
569 void use_vector(MPI_Datatype* d){
570 smpi_datatype_use(reinterpret_cast<s_smpi_mpi_indexed_t*>((*d)->substruct)->old_type);
573 /* Hvector Implementation - Vector with stride in bytes */
575 /* Copies noncontiguous data into contiguous memory.
576 * @param contiguous_hvector - output hvector
577 * @param noncontiguous_hvector - input hvector
578 * @param type - pointer contening :
579 * - stride - stride of between noncontiguous data, in bytes
580 * - block_length - the width or height of blocked matrix
581 * - count - the number of rows of matrix
583 void serialize_hvector( void* noncontiguous_hvector, void *contiguous_hvector, int count, void *type)
585 s_smpi_mpi_hvector_t* type_c = reinterpret_cast<s_smpi_mpi_hvector_t*>(type);
587 char* contiguous_vector_char = static_cast<char*>(contiguous_hvector);
588 char* noncontiguous_vector_char = static_cast<char*>(noncontiguous_hvector);
590 for (i = 0; i < type_c->block_count * count; i++) {
591 if (type_c->old_type->sizeof_substruct == 0)
592 memcpy(contiguous_vector_char, noncontiguous_vector_char, type_c->block_length * type_c->size_oldtype);
594 static_cast<s_smpi_subtype_t*>(type_c->old_type->substruct)->serialize( noncontiguous_vector_char,
595 contiguous_vector_char,
596 type_c->block_length, type_c->old_type->substruct);
598 contiguous_vector_char += type_c->block_length*type_c->size_oldtype;
599 if((i+1)%type_c->block_count ==0)
600 noncontiguous_vector_char += type_c->block_length*type_c->size_oldtype;
602 noncontiguous_vector_char += type_c->block_stride;
605 /* Copies contiguous data into noncontiguous memory.
606 * @param noncontiguous_vector - output hvector
607 * @param contiguous_vector - input hvector
608 * @param type - pointer contening :
609 * - stride - stride of between noncontiguous data, in bytes
610 * - block_length - the width or height of blocked matrix
611 * - count - the number of rows of matrix
613 void unserialize_hvector( void* contiguous_vector, void *noncontiguous_vector, int count, void *type, MPI_Op op)
615 s_smpi_mpi_hvector_t* type_c = reinterpret_cast<s_smpi_mpi_hvector_t*>(type);
618 char* contiguous_vector_char = static_cast<char*>(contiguous_vector);
619 char* noncontiguous_vector_char = static_cast<char*>(noncontiguous_vector);
621 for (i = 0; i < type_c->block_count * count; i++) {
622 if (type_c->old_type->sizeof_substruct == 0)
623 smpi_op_apply(op, contiguous_vector_char, noncontiguous_vector_char, &type_c->block_length, &type_c->old_type);
625 static_cast<s_smpi_subtype_t*>(type_c->old_type->substruct)->unserialize( contiguous_vector_char, noncontiguous_vector_char,
626 type_c->block_length, type_c->old_type->substruct,
628 contiguous_vector_char += type_c->block_length*type_c->size_oldtype;
629 if((i+1)%type_c->block_count ==0)
630 noncontiguous_vector_char += type_c->block_length*type_c->size_oldtype;
632 noncontiguous_vector_char += type_c->block_stride;
636 /* Create a Sub type vector to be able to serialize and unserialize it the structure s_smpi_mpi_vector_t is derived
637 * from s_smpi_subtype which required the functions unserialize and serialize
640 s_smpi_mpi_hvector_t* smpi_datatype_hvector_create( MPI_Aint block_stride, int block_length, int block_count,
641 MPI_Datatype old_type, int size_oldtype){
642 s_smpi_mpi_hvector_t *new_t= xbt_new(s_smpi_mpi_hvector_t,1);
643 new_t->base.serialize = &serialize_hvector;
644 new_t->base.unserialize = &unserialize_hvector;
645 new_t->base.subtype_free = &free_hvector;
646 new_t->base.subtype_use = &use_hvector;
647 new_t->block_stride = block_stride;
648 new_t->block_length = block_length;
649 new_t->block_count = block_count;
650 new_t->old_type = old_type;
651 new_t->size_oldtype = size_oldtype;
652 smpi_datatype_use(old_type);
656 //do nothing for vector types
657 void free_hvector(MPI_Datatype* d){
658 smpi_datatype_unuse(reinterpret_cast<s_smpi_mpi_hvector_t*>((*d)->substruct)->old_type);
661 void use_hvector(MPI_Datatype* d){
662 smpi_datatype_use(reinterpret_cast<s_smpi_mpi_hvector_t*>((*d)->substruct)->old_type);
665 int smpi_datatype_hvector(int count, int blocklen, MPI_Aint stride, MPI_Datatype old_type, MPI_Datatype* new_type)
673 lb=smpi_datatype_lb(old_type);
674 ub=((count-1)*stride)+(blocklen-1)*smpi_datatype_get_extent(old_type)+smpi_datatype_ub(old_type);
676 if(old_type->sizeof_substruct != 0 || stride != blocklen*smpi_datatype_get_extent(old_type)){
677 s_smpi_mpi_hvector_t* subtype = smpi_datatype_hvector_create( stride, blocklen, count, old_type,
678 smpi_datatype_size(old_type));
680 smpi_datatype_create(new_type, count * blocklen * smpi_datatype_size(old_type), lb,ub, sizeof(s_smpi_mpi_hvector_t), subtype, DT_FLAG_VECTOR);
683 smpi_datatype_create(new_type, count * blocklen * smpi_datatype_size(old_type),0,count * blocklen *
684 smpi_datatype_size(old_type), 0, nullptr, DT_FLAG_VECTOR|DT_FLAG_CONTIGUOUS);
690 /* Indexed Implementation */
692 /* Copies noncontiguous data into contiguous memory.
693 * @param contiguous_indexed - output indexed
694 * @param noncontiguous_indexed - input indexed
695 * @param type - pointer contening :
696 * - block_lengths - the width or height of blocked matrix
697 * - block_indices - indices of each data, in element
698 * - count - the number of rows of matrix
700 void serialize_indexed( void* noncontiguous_indexed, void *contiguous_indexed, int count, void *type)
702 s_smpi_mpi_indexed_t* type_c = reinterpret_cast<s_smpi_mpi_indexed_t*>(type);
704 char* contiguous_indexed_char = static_cast<char*>(contiguous_indexed);
705 char* noncontiguous_indexed_char = static_cast<char*>(noncontiguous_indexed)+type_c->block_indices[0] * type_c->size_oldtype;
706 for(j=0; j<count;j++){
707 for (i = 0; i < type_c->block_count; i++) {
708 if (type_c->old_type->sizeof_substruct == 0)
709 memcpy(contiguous_indexed_char, noncontiguous_indexed_char, type_c->block_lengths[i] * type_c->size_oldtype);
711 static_cast<s_smpi_subtype_t*>(type_c->old_type->substruct)->serialize( noncontiguous_indexed_char,
712 contiguous_indexed_char,
713 type_c->block_lengths[i],
714 type_c->old_type->substruct);
716 contiguous_indexed_char += type_c->block_lengths[i]*type_c->size_oldtype;
717 if (i<type_c->block_count-1)
718 noncontiguous_indexed_char =
719 static_cast<char*>(noncontiguous_indexed) + type_c->block_indices[i+1]*smpi_datatype_get_extent(type_c->old_type);
721 noncontiguous_indexed_char += type_c->block_lengths[i]*smpi_datatype_get_extent(type_c->old_type);
723 noncontiguous_indexed=static_cast< void*>(noncontiguous_indexed_char);
726 /* Copies contiguous data into noncontiguous memory.
727 * @param noncontiguous_indexed - output indexed
728 * @param contiguous_indexed - input indexed
729 * @param type - pointer contening :
730 * - block_lengths - the width or height of blocked matrix
731 * - block_indices - indices of each data, in element
732 * - count - the number of rows of matrix
734 void unserialize_indexed( void* contiguous_indexed, void *noncontiguous_indexed, int count, void *type, MPI_Op op)
736 s_smpi_mpi_indexed_t* type_c = reinterpret_cast<s_smpi_mpi_indexed_t*>(type);
738 char* contiguous_indexed_char = static_cast<char*>(contiguous_indexed);
739 char* noncontiguous_indexed_char =
740 static_cast<char*>(noncontiguous_indexed)+type_c->block_indices[0]*smpi_datatype_get_extent(type_c->old_type);
741 for(j=0; j<count;j++){
742 for (i = 0; i < type_c->block_count; i++) {
743 if (type_c->old_type->sizeof_substruct == 0)
744 smpi_op_apply(op, contiguous_indexed_char, noncontiguous_indexed_char, &type_c->block_lengths[i],
747 static_cast<s_smpi_subtype_t*>(type_c->old_type->substruct)->unserialize( contiguous_indexed_char,
748 noncontiguous_indexed_char,
749 type_c->block_lengths[i],
750 type_c->old_type->substruct, op);
752 contiguous_indexed_char += type_c->block_lengths[i]*type_c->size_oldtype;
753 if (i<type_c->block_count-1)
754 noncontiguous_indexed_char =
755 static_cast<char*>(noncontiguous_indexed) + type_c->block_indices[i+1]*smpi_datatype_get_extent(type_c->old_type);
757 noncontiguous_indexed_char += type_c->block_lengths[i]*smpi_datatype_get_extent(type_c->old_type);
759 noncontiguous_indexed=static_cast<void*>(noncontiguous_indexed_char);
763 void free_indexed(MPI_Datatype* type){
764 if((*type)->in_use==0){
765 xbt_free(reinterpret_cast<s_smpi_mpi_indexed_t*>((*type)->substruct)->block_lengths);
766 xbt_free(reinterpret_cast<s_smpi_mpi_indexed_t*>((*type)->substruct)->block_indices);
768 smpi_datatype_unuse(reinterpret_cast<s_smpi_mpi_indexed_t*>((*type)->substruct)->old_type);
771 void use_indexed(MPI_Datatype* type){
772 smpi_datatype_use(reinterpret_cast<s_smpi_mpi_indexed_t*>((*type)->substruct)->old_type);
776 /* Create a Sub type indexed to be able to serialize and unserialize it the structure s_smpi_mpi_indexed_t is derived
777 * from s_smpi_subtype which required the functions unserialize and serialize */
778 s_smpi_mpi_indexed_t* smpi_datatype_indexed_create( int* block_lengths, int* block_indices, int block_count,
779 MPI_Datatype old_type, int size_oldtype){
780 s_smpi_mpi_indexed_t *new_t= xbt_new(s_smpi_mpi_indexed_t,1);
781 new_t->base.serialize = &serialize_indexed;
782 new_t->base.unserialize = &unserialize_indexed;
783 new_t->base.subtype_free = &free_indexed;
784 new_t->base.subtype_use = &use_indexed;
785 new_t->block_lengths= xbt_new(int, block_count);
786 new_t->block_indices= xbt_new(int, block_count);
788 for(i=0;i<block_count;i++){
789 new_t->block_lengths[i]=block_lengths[i];
790 new_t->block_indices[i]=block_indices[i];
792 new_t->block_count = block_count;
793 smpi_datatype_use(old_type);
794 new_t->old_type = old_type;
795 new_t->size_oldtype = size_oldtype;
799 int smpi_datatype_indexed(int count, int* blocklens, int* indices, MPI_Datatype old_type, MPI_Datatype* new_type)
804 bool contiguous=true;
808 lb=indices[0]*smpi_datatype_get_extent(old_type);
809 ub=indices[0]*smpi_datatype_get_extent(old_type) + blocklens[0]*smpi_datatype_ub(old_type);
812 for(i=0; i< count; i++){
815 size += blocklens[i];
817 if(indices[i]*smpi_datatype_get_extent(old_type)+smpi_datatype_lb(old_type)<lb)
818 lb = indices[i]*smpi_datatype_get_extent(old_type)+smpi_datatype_lb(old_type);
819 if(indices[i]*smpi_datatype_get_extent(old_type)+blocklens[i]*smpi_datatype_ub(old_type)>ub)
820 ub = indices[i]*smpi_datatype_get_extent(old_type)+blocklens[i]*smpi_datatype_ub(old_type);
822 if ( (i< count -1) && (indices[i]+blocklens[i] != indices[i+1]) )
825 if (old_type->sizeof_substruct != 0)
829 s_smpi_mpi_indexed_t* subtype = smpi_datatype_indexed_create( blocklens, indices, count, 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_indexed_t), subtype, DT_FLAG_DATA);
833 s_smpi_mpi_contiguous_t* subtype = smpi_datatype_contiguous_create( lb, size, old_type,
834 smpi_datatype_size(old_type));
835 smpi_datatype_create(new_type, size * smpi_datatype_size(old_type), lb, ub, sizeof(s_smpi_mpi_contiguous_t), subtype,
836 DT_FLAG_DATA|DT_FLAG_CONTIGUOUS);
841 /* Hindexed Implementation - Indexed with indices in bytes */
843 /* Copies noncontiguous data into contiguous memory.
844 * @param contiguous_hindexed - output hindexed
845 * @param noncontiguous_hindexed - input hindexed
846 * @param type - pointer contening :
847 * - block_lengths - the width or height of blocked matrix
848 * - block_indices - indices of each data, in bytes
849 * - count - the number of rows of matrix
851 void serialize_hindexed( void* noncontiguous_hindexed, void *contiguous_hindexed, int count, void *type)
853 s_smpi_mpi_hindexed_t* type_c = reinterpret_cast<s_smpi_mpi_hindexed_t*>(type);
855 char* contiguous_hindexed_char = static_cast<char*>(contiguous_hindexed);
856 char* noncontiguous_hindexed_char = static_cast<char*>(noncontiguous_hindexed)+ type_c->block_indices[0];
857 for(j=0; j<count;j++){
858 for (i = 0; i < type_c->block_count; i++) {
859 if (type_c->old_type->sizeof_substruct == 0)
860 memcpy(contiguous_hindexed_char, noncontiguous_hindexed_char, type_c->block_lengths[i] * type_c->size_oldtype);
862 static_cast<s_smpi_subtype_t*>(type_c->old_type->substruct)->serialize( noncontiguous_hindexed_char,
863 contiguous_hindexed_char,
864 type_c->block_lengths[i],
865 type_c->old_type->substruct);
867 contiguous_hindexed_char += type_c->block_lengths[i]*type_c->size_oldtype;
868 if (i<type_c->block_count-1)
869 noncontiguous_hindexed_char = static_cast<char*>(noncontiguous_hindexed) + type_c->block_indices[i+1];
871 noncontiguous_hindexed_char += type_c->block_lengths[i]*smpi_datatype_get_extent(type_c->old_type);
873 noncontiguous_hindexed=static_cast<void*>(noncontiguous_hindexed_char);
876 /* Copies contiguous data into noncontiguous memory.
877 * @param noncontiguous_hindexed - output hindexed
878 * @param contiguous_hindexed - input hindexed
879 * @param type - pointer contening :
880 * - block_lengths - the width or height of blocked matrix
881 * - block_indices - indices of each data, in bytes
882 * - count - the number of rows of matrix
884 void unserialize_hindexed( void* contiguous_hindexed, void *noncontiguous_hindexed, int count, void *type,
887 s_smpi_mpi_hindexed_t* type_c = reinterpret_cast<s_smpi_mpi_hindexed_t*>(type);
890 char* contiguous_hindexed_char = static_cast<char*>(contiguous_hindexed);
891 char* noncontiguous_hindexed_char = static_cast<char*>(noncontiguous_hindexed)+ type_c->block_indices[0];
892 for(j=0; j<count;j++){
893 for (i = 0; i < type_c->block_count; i++) {
894 if (type_c->old_type->sizeof_substruct == 0)
895 smpi_op_apply(op, contiguous_hindexed_char, noncontiguous_hindexed_char, &type_c->block_lengths[i],
898 static_cast<s_smpi_subtype_t*>(type_c->old_type->substruct)->unserialize( contiguous_hindexed_char,
899 noncontiguous_hindexed_char,
900 type_c->block_lengths[i],
901 type_c->old_type->substruct, op);
903 contiguous_hindexed_char += type_c->block_lengths[i]*type_c->size_oldtype;
904 if (i<type_c->block_count-1)
905 noncontiguous_hindexed_char = static_cast<char*>(noncontiguous_hindexed) + type_c->block_indices[i+1];
907 noncontiguous_hindexed_char += type_c->block_lengths[i]*smpi_datatype_get_extent(type_c->old_type);
909 noncontiguous_hindexed=static_cast<void*>(noncontiguous_hindexed_char);
913 void free_hindexed(MPI_Datatype* type){
914 if((*type)->in_use==0){
915 xbt_free(reinterpret_cast<s_smpi_mpi_hindexed_t*>((*type)->substruct)->block_lengths);
916 xbt_free(reinterpret_cast<s_smpi_mpi_hindexed_t*>((*type)->substruct)->block_indices);
918 smpi_datatype_unuse(reinterpret_cast<s_smpi_mpi_hindexed_t*>((*type)->substruct)->old_type);
921 void use_hindexed(MPI_Datatype* type){
922 smpi_datatype_use(reinterpret_cast<s_smpi_mpi_hindexed_t*>((*type)->substruct)->old_type);
925 /* Create a Sub type hindexed to be able to serialize and unserialize it the structure s_smpi_mpi_hindexed_t is derived
926 * from s_smpi_subtype which required the functions unserialize and serialize
928 s_smpi_mpi_hindexed_t* smpi_datatype_hindexed_create( int* block_lengths, MPI_Aint* block_indices, int block_count,
929 MPI_Datatype old_type, int size_oldtype){
930 s_smpi_mpi_hindexed_t *new_t= xbt_new(s_smpi_mpi_hindexed_t,1);
931 new_t->base.serialize = &serialize_hindexed;
932 new_t->base.unserialize = &unserialize_hindexed;
933 new_t->base.subtype_free = &free_hindexed;
934 new_t->base.subtype_use = &use_hindexed;
935 new_t->block_lengths= xbt_new(int, block_count);
936 new_t->block_indices= xbt_new(MPI_Aint, block_count);
937 for(int i=0;i<block_count;i++){
938 new_t->block_lengths[i]=block_lengths[i];
939 new_t->block_indices[i]=block_indices[i];
941 new_t->block_count = block_count;
942 new_t->old_type = old_type;
943 smpi_datatype_use(old_type);
944 new_t->size_oldtype = size_oldtype;
948 int smpi_datatype_hindexed(int count, int* blocklens, MPI_Aint* indices, MPI_Datatype old_type, MPI_Datatype* new_type)
953 bool contiguous=true;
957 lb=indices[0] + smpi_datatype_lb(old_type);
958 ub=indices[0] + blocklens[0]*smpi_datatype_ub(old_type);
960 for(i=0; i< count; i++){
963 size += blocklens[i];
965 if(indices[i]+smpi_datatype_lb(old_type)<lb)
966 lb = indices[i]+smpi_datatype_lb(old_type);
967 if(indices[i]+blocklens[i]*smpi_datatype_ub(old_type)>ub)
968 ub = indices[i]+blocklens[i]*smpi_datatype_ub(old_type);
970 if ( (i< count -1) && (indices[i]+blocklens[i]*(static_cast<int>(smpi_datatype_size(old_type))) != indices[i+1]) )
973 if (old_type->sizeof_substruct != 0 || lb!=0)
977 s_smpi_mpi_hindexed_t* subtype = smpi_datatype_hindexed_create( blocklens, indices, count, old_type,
978 smpi_datatype_size(old_type));
979 smpi_datatype_create(new_type, size * smpi_datatype_size(old_type), lb, ub ,sizeof(s_smpi_mpi_hindexed_t), subtype, DT_FLAG_DATA);
981 s_smpi_mpi_contiguous_t* subtype = smpi_datatype_contiguous_create(lb,size, old_type, smpi_datatype_size(old_type));
982 smpi_datatype_create(new_type, size * smpi_datatype_size(old_type), 0,size * smpi_datatype_size(old_type),
983 1, subtype, DT_FLAG_DATA|DT_FLAG_CONTIGUOUS);
989 /* struct Implementation - Indexed with indices in bytes */
991 /* Copies noncontiguous data into contiguous memory.
992 * @param contiguous_struct - output struct
993 * @param noncontiguous_struct - input struct
994 * @param type - pointer contening :
995 * - stride - stride of between noncontiguous data
996 * - block_length - the width or height of blocked matrix
997 * - count - the number of rows of matrix
999 void serialize_struct( void* noncontiguous_struct, void *contiguous_struct, int count, void *type)
1001 s_smpi_mpi_struct_t* type_c = reinterpret_cast<s_smpi_mpi_struct_t*>(type);
1003 char* contiguous_struct_char = static_cast<char*>(contiguous_struct);
1004 char* noncontiguous_struct_char = static_cast<char*>(noncontiguous_struct)+ type_c->block_indices[0];
1005 for(j=0; j<count;j++){
1006 for (i = 0; i < type_c->block_count; i++) {
1007 if (type_c->old_types[i]->sizeof_substruct == 0)
1008 memcpy(contiguous_struct_char, noncontiguous_struct_char,
1009 type_c->block_lengths[i] * smpi_datatype_size(type_c->old_types[i]));
1011 static_cast<s_smpi_subtype_t*>(type_c->old_types[i]->substruct)->serialize( noncontiguous_struct_char,
1012 contiguous_struct_char,
1013 type_c->block_lengths[i],
1014 type_c->old_types[i]->substruct);
1017 contiguous_struct_char += type_c->block_lengths[i]*smpi_datatype_size(type_c->old_types[i]);
1018 if (i<type_c->block_count-1)
1019 noncontiguous_struct_char = static_cast<char*>(noncontiguous_struct) + type_c->block_indices[i+1];
1020 else //let's hope this is MPI_UB ?
1021 noncontiguous_struct_char += type_c->block_lengths[i]*smpi_datatype_get_extent(type_c->old_types[i]);
1023 noncontiguous_struct=static_cast<void*>(noncontiguous_struct_char);
1027 /* Copies contiguous data into noncontiguous memory.
1028 * @param noncontiguous_struct - output struct
1029 * @param contiguous_struct - input struct
1030 * @param type - pointer contening :
1031 * - stride - stride of between noncontiguous data
1032 * - block_length - the width or height of blocked matrix
1033 * - count - the number of rows of matrix
1035 void unserialize_struct( void* contiguous_struct, void *noncontiguous_struct, int count, void *type, MPI_Op op)
1037 s_smpi_mpi_struct_t* type_c = reinterpret_cast<s_smpi_mpi_struct_t*>(type);
1040 char* contiguous_struct_char = static_cast<char*>(contiguous_struct);
1041 char* noncontiguous_struct_char = static_cast<char*>(noncontiguous_struct)+ type_c->block_indices[0];
1042 for(j=0; j<count;j++){
1043 for (i = 0; i < type_c->block_count; i++) {
1044 if (type_c->old_types[i]->sizeof_substruct == 0)
1045 smpi_op_apply(op, contiguous_struct_char, noncontiguous_struct_char, &type_c->block_lengths[i],
1046 & type_c->old_types[i]);
1048 static_cast<s_smpi_subtype_t*>(type_c->old_types[i]->substruct)->unserialize( contiguous_struct_char,
1049 noncontiguous_struct_char,
1050 type_c->block_lengths[i],
1051 type_c->old_types[i]->substruct, op);
1053 contiguous_struct_char += type_c->block_lengths[i]*smpi_datatype_size(type_c->old_types[i]);
1054 if (i<type_c->block_count-1)
1055 noncontiguous_struct_char = static_cast<char*>(noncontiguous_struct) + type_c->block_indices[i+1];
1057 noncontiguous_struct_char += type_c->block_lengths[i]*smpi_datatype_get_extent(type_c->old_types[i]);
1059 noncontiguous_struct=reinterpret_cast<void*>(noncontiguous_struct_char);
1063 void free_struct(MPI_Datatype* type){
1065 for (i = 0; i < reinterpret_cast<s_smpi_mpi_struct_t*>((*type)->substruct)->block_count; i++)
1066 smpi_datatype_unuse(reinterpret_cast<s_smpi_mpi_struct_t*>((*type)->substruct)->old_types[i]);
1067 if((*type)->in_use==0){
1068 xbt_free(reinterpret_cast<s_smpi_mpi_struct_t*>((*type)->substruct)->block_lengths);
1069 xbt_free(reinterpret_cast<s_smpi_mpi_struct_t*>((*type)->substruct)->block_indices);
1070 xbt_free(reinterpret_cast<s_smpi_mpi_struct_t*>((*type)->substruct)->old_types);
1074 void use_struct(MPI_Datatype* type){
1076 for (i = 0; i < reinterpret_cast<s_smpi_mpi_struct_t*>((*type)->substruct)->block_count; i++)
1077 smpi_datatype_use(reinterpret_cast<s_smpi_mpi_struct_t*>((*type)->substruct)->old_types[i]);
1080 /* Create a Sub type struct to be able to serialize and unserialize it the structure s_smpi_mpi_struct_t is derived
1081 * from s_smpi_subtype which required the functions unserialize and serialize
1083 s_smpi_mpi_struct_t* smpi_datatype_struct_create( int* block_lengths, MPI_Aint* block_indices, int block_count,
1084 MPI_Datatype* old_types){
1085 s_smpi_mpi_struct_t *new_t= xbt_new(s_smpi_mpi_struct_t,1);
1086 new_t->base.serialize = &serialize_struct;
1087 new_t->base.unserialize = &unserialize_struct;
1088 new_t->base.subtype_free = &free_struct;
1089 new_t->base.subtype_use = &use_struct;
1090 new_t->block_lengths= xbt_new(int, block_count);
1091 new_t->block_indices= xbt_new(MPI_Aint, block_count);
1092 new_t->old_types= xbt_new(MPI_Datatype, block_count);
1094 for(i=0;i<block_count;i++){
1095 new_t->block_lengths[i]=block_lengths[i];
1096 new_t->block_indices[i]=block_indices[i];
1097 new_t->old_types[i]=old_types[i];
1098 smpi_datatype_use(new_t->old_types[i]);
1100 new_t->block_count = block_count;
1104 int smpi_datatype_struct(int count, int* blocklens, MPI_Aint* indices, MPI_Datatype* old_types, MPI_Datatype* new_type)
1108 bool contiguous=true;
1113 lb=indices[0] + smpi_datatype_lb(old_types[0]);
1114 ub=indices[0] + blocklens[0]*smpi_datatype_ub(old_types[0]);
1116 bool forced_lb=false;
1117 bool forced_ub=false;
1118 for(i=0; i< count; i++){
1121 if (old_types[i]->sizeof_substruct != 0)
1124 size += blocklens[i]*smpi_datatype_size(old_types[i]);
1125 if (old_types[i]==MPI_LB){
1129 if (old_types[i]==MPI_UB){
1134 if(!forced_lb && indices[i]+smpi_datatype_lb(old_types[i])<lb)
1136 if(!forced_ub && indices[i]+blocklens[i]*smpi_datatype_ub(old_types[i])>ub)
1137 ub = indices[i]+blocklens[i]*smpi_datatype_ub(old_types[i]);
1139 if ( (i< count -1) && (indices[i]+blocklens[i]*static_cast<int>(smpi_datatype_size(old_types[i])) != indices[i+1]) )
1144 s_smpi_mpi_struct_t* subtype = smpi_datatype_struct_create( blocklens, indices, count, old_types);
1146 smpi_datatype_create(new_type, size, lb, ub,sizeof(s_smpi_mpi_struct_t), subtype, DT_FLAG_DATA);
1148 s_smpi_mpi_contiguous_t* subtype = smpi_datatype_contiguous_create( lb, size, MPI_CHAR, 1);
1149 smpi_datatype_create(new_type, size, lb, ub,1, subtype, DT_FLAG_DATA|DT_FLAG_CONTIGUOUS);
1154 void smpi_datatype_commit(MPI_Datatype *datatype)
1156 (*datatype)->flags= ((*datatype)->flags | DT_FLAG_COMMITED);
1159 typedef struct s_smpi_mpi_op {
1160 MPI_User_function *func;
1165 #define MAX_OP(a, b) (b) = (a) < (b) ? (b) : (a)
1166 #define MIN_OP(a, b) (b) = (a) < (b) ? (a) : (b)
1167 #define SUM_OP(a, b) (b) += (a)
1168 #define PROD_OP(a, b) (b) *= (a)
1169 #define LAND_OP(a, b) (b) = (a) && (b)
1170 #define LOR_OP(a, b) (b) = (a) || (b)
1171 #define LXOR_OP(a, b) (b) = (!(a) && (b)) || ((a) && !(b))
1172 #define BAND_OP(a, b) (b) &= (a)
1173 #define BOR_OP(a, b) (b) |= (a)
1174 #define BXOR_OP(a, b) (b) ^= (a)
1175 #define MAXLOC_OP(a, b) (b) = (a.value) < (b.value) ? (b) : (a)
1176 #define MINLOC_OP(a, b) (b) = (a.value) < (b.value) ? (a) : (b)
1178 #define APPLY_FUNC(a, b, length, type, func) \
1181 type* x = (type*)(a); \
1182 type* y = (type*)(b); \
1183 for(i = 0; i < *(length); i++) { \
1188 #define APPLY_OP_LOOP(dtype, type, op) \
1189 if (*datatype == dtype) {\
1190 APPLY_FUNC(a, b, length, type, op)\
1194 #define APPLY_BASIC_OP_LOOP(op)\
1195 APPLY_OP_LOOP(MPI_CHAR, char,op)\
1196 APPLY_OP_LOOP(MPI_SHORT, short,op)\
1197 APPLY_OP_LOOP(MPI_INT, int,op)\
1198 APPLY_OP_LOOP(MPI_LONG, long,op)\
1199 APPLY_OP_LOOP(MPI_LONG_LONG, long long,op)\
1200 APPLY_OP_LOOP(MPI_SIGNED_CHAR, signed char,op)\
1201 APPLY_OP_LOOP(MPI_UNSIGNED_CHAR, unsigned char,op)\
1202 APPLY_OP_LOOP(MPI_UNSIGNED_SHORT, unsigned short,op)\
1203 APPLY_OP_LOOP(MPI_UNSIGNED, unsigned int,op)\
1204 APPLY_OP_LOOP(MPI_UNSIGNED_LONG, unsigned long,op)\
1205 APPLY_OP_LOOP(MPI_UNSIGNED_LONG_LONG, unsigned long long,op)\
1206 APPLY_OP_LOOP(MPI_WCHAR, wchar_t,op)\
1207 APPLY_OP_LOOP(MPI_BYTE, int8_t,op)\
1208 APPLY_OP_LOOP(MPI_INT8_T, int8_t,op)\
1209 APPLY_OP_LOOP(MPI_INT16_T, int16_t,op)\
1210 APPLY_OP_LOOP(MPI_INT32_T, int32_t,op)\
1211 APPLY_OP_LOOP(MPI_INT64_T, int64_t,op)\
1212 APPLY_OP_LOOP(MPI_UINT8_T, uint8_t,op)\
1213 APPLY_OP_LOOP(MPI_UINT16_T, uint16_t,op)\
1214 APPLY_OP_LOOP(MPI_UINT32_T, uint32_t,op)\
1215 APPLY_OP_LOOP(MPI_UINT64_T, uint64_t,op)\
1216 APPLY_OP_LOOP(MPI_AINT, MPI_Aint,op)\
1217 APPLY_OP_LOOP(MPI_OFFSET, MPI_Offset,op)\
1218 APPLY_OP_LOOP(MPI_INTEGER1, int,op)\
1219 APPLY_OP_LOOP(MPI_INTEGER2, int16_t,op)\
1220 APPLY_OP_LOOP(MPI_INTEGER4, int32_t,op)\
1221 APPLY_OP_LOOP(MPI_INTEGER8, int64_t,op)
1223 #define APPLY_BOOL_OP_LOOP(op)\
1224 APPLY_OP_LOOP(MPI_C_BOOL, bool,op)
1226 #define APPLY_FLOAT_OP_LOOP(op)\
1227 APPLY_OP_LOOP(MPI_FLOAT, float,op)\
1228 APPLY_OP_LOOP(MPI_DOUBLE, double,op)\
1229 APPLY_OP_LOOP(MPI_LONG_DOUBLE, long double,op)\
1230 APPLY_OP_LOOP(MPI_REAL, float,op)\
1231 APPLY_OP_LOOP(MPI_REAL4, float,op)\
1232 APPLY_OP_LOOP(MPI_REAL8, float,op)\
1233 APPLY_OP_LOOP(MPI_REAL16, double,op)
1235 #define APPLY_COMPLEX_OP_LOOP(op)\
1236 APPLY_OP_LOOP(MPI_C_FLOAT_COMPLEX, float _Complex,op)\
1237 APPLY_OP_LOOP(MPI_C_DOUBLE_COMPLEX, double _Complex,op)\
1238 APPLY_OP_LOOP(MPI_C_LONG_DOUBLE_COMPLEX, long double _Complex,op)
1240 #define APPLY_PAIR_OP_LOOP(op)\
1241 APPLY_OP_LOOP(MPI_FLOAT_INT, float_int,op)\
1242 APPLY_OP_LOOP(MPI_LONG_INT, long_int,op)\
1243 APPLY_OP_LOOP(MPI_DOUBLE_INT, double_int,op)\
1244 APPLY_OP_LOOP(MPI_SHORT_INT, short_int,op)\
1245 APPLY_OP_LOOP(MPI_2INT, int_int,op)\
1246 APPLY_OP_LOOP(MPI_2FLOAT, float_float,op)\
1247 APPLY_OP_LOOP(MPI_2DOUBLE, double_double,op)\
1248 APPLY_OP_LOOP(MPI_LONG_DOUBLE_INT, long_double_int,op)\
1249 APPLY_OP_LOOP(MPI_2LONG, long_long,op)
1251 #define APPLY_END_OP_LOOP(op)\
1253 xbt_die("Failed to apply " #op " to type %s", (*datatype)->name);\
1257 static void max_func(void *a, void *b, int *length, MPI_Datatype * datatype)
1259 APPLY_BASIC_OP_LOOP(MAX_OP)
1260 APPLY_FLOAT_OP_LOOP(MAX_OP)
1261 APPLY_END_OP_LOOP(MAX_OP)
1264 static void min_func(void *a, void *b, int *length, MPI_Datatype * datatype)
1266 APPLY_BASIC_OP_LOOP(MIN_OP)
1267 APPLY_FLOAT_OP_LOOP(MIN_OP)
1268 APPLY_END_OP_LOOP(MIN_OP)
1271 static void sum_func(void *a, void *b, int *length, MPI_Datatype * datatype)
1273 APPLY_BASIC_OP_LOOP(SUM_OP)
1274 APPLY_FLOAT_OP_LOOP(SUM_OP)
1275 APPLY_COMPLEX_OP_LOOP(SUM_OP)
1276 APPLY_END_OP_LOOP(SUM_OP)
1279 static void prod_func(void *a, void *b, int *length, MPI_Datatype * datatype)
1281 APPLY_BASIC_OP_LOOP(PROD_OP)
1282 APPLY_FLOAT_OP_LOOP(PROD_OP)
1283 APPLY_COMPLEX_OP_LOOP(PROD_OP)
1284 APPLY_END_OP_LOOP(PROD_OP)
1287 static void land_func(void *a, void *b, int *length, MPI_Datatype * datatype)
1289 APPLY_BASIC_OP_LOOP(LAND_OP)
1290 APPLY_BOOL_OP_LOOP(LAND_OP)
1291 APPLY_END_OP_LOOP(LAND_OP)
1294 static void lor_func(void *a, void *b, int *length, MPI_Datatype * datatype)
1296 APPLY_BASIC_OP_LOOP(LOR_OP)
1297 APPLY_BOOL_OP_LOOP(LOR_OP)
1298 APPLY_END_OP_LOOP(LOR_OP)
1301 static void lxor_func(void *a, void *b, int *length, MPI_Datatype * datatype)
1303 APPLY_BASIC_OP_LOOP(LXOR_OP)
1304 APPLY_BOOL_OP_LOOP(LXOR_OP)
1305 APPLY_END_OP_LOOP(LXOR_OP)
1308 static void band_func(void *a, void *b, int *length, MPI_Datatype * datatype)
1310 APPLY_BASIC_OP_LOOP(BAND_OP)
1311 APPLY_BOOL_OP_LOOP(BAND_OP)
1312 APPLY_END_OP_LOOP(BAND_OP)
1315 static void bor_func(void *a, void *b, int *length, MPI_Datatype * datatype)
1317 APPLY_BASIC_OP_LOOP(BOR_OP)
1318 APPLY_BOOL_OP_LOOP(BOR_OP)
1319 APPLY_END_OP_LOOP(BOR_OP)
1322 static void bxor_func(void *a, void *b, int *length, MPI_Datatype * datatype)
1324 APPLY_BASIC_OP_LOOP(BXOR_OP)
1325 APPLY_BOOL_OP_LOOP(BXOR_OP)
1326 APPLY_END_OP_LOOP(BXOR_OP)
1329 static void minloc_func(void *a, void *b, int *length, MPI_Datatype * datatype)
1331 APPLY_PAIR_OP_LOOP(MINLOC_OP)
1332 APPLY_END_OP_LOOP(MINLOC_OP)
1335 static void maxloc_func(void *a, void *b, int *length, MPI_Datatype * datatype)
1337 APPLY_PAIR_OP_LOOP(MAXLOC_OP)
1338 APPLY_END_OP_LOOP(MAXLOC_OP)
1341 static void replace_func(void *a, void *b, int *length, MPI_Datatype * datatype)
1343 memcpy(b, a, *length * smpi_datatype_size(*datatype));
1346 #define CREATE_MPI_OP(name, func) \
1347 static s_smpi_mpi_op_t mpi_##name = { &(func) /* func */, true, false }; \
1348 MPI_Op name = &mpi_##name;
1350 CREATE_MPI_OP(MPI_MAX, max_func);
1351 CREATE_MPI_OP(MPI_MIN, min_func);
1352 CREATE_MPI_OP(MPI_SUM, sum_func);
1353 CREATE_MPI_OP(MPI_PROD, prod_func);
1354 CREATE_MPI_OP(MPI_LAND, land_func);
1355 CREATE_MPI_OP(MPI_LOR, lor_func);
1356 CREATE_MPI_OP(MPI_LXOR, lxor_func);
1357 CREATE_MPI_OP(MPI_BAND, band_func);
1358 CREATE_MPI_OP(MPI_BOR, bor_func);
1359 CREATE_MPI_OP(MPI_BXOR, bxor_func);
1360 CREATE_MPI_OP(MPI_MAXLOC, maxloc_func);
1361 CREATE_MPI_OP(MPI_MINLOC, minloc_func);
1362 CREATE_MPI_OP(MPI_REPLACE, replace_func);
1364 MPI_Op smpi_op_new(MPI_User_function * function, bool commute)
1367 op = xbt_new(s_smpi_mpi_op_t, 1);
1368 op->func = function;
1369 op-> is_commute = commute;
1370 op-> is_fortran_op = false;
1374 bool smpi_op_is_commute(MPI_Op op)
1376 return (op==MPI_OP_NULL) ? true : op-> is_commute;
1379 void smpi_op_destroy(MPI_Op op)
1384 void smpi_op_set_fortran(MPI_Op op)
1386 //tell that we were created from fortran, so we need to translate the type to fortran when called
1387 op->is_fortran_op = true;
1390 void smpi_op_apply(MPI_Op op, void *invec, void *inoutvec, int *len, MPI_Datatype * datatype)
1395 if(smpi_privatize_global_variables){//we need to switch as the called function may silently touch global variables
1396 XBT_DEBUG("Applying operation, switch to the right data frame ");
1397 smpi_switch_data_segment(smpi_process_index());
1400 if(!smpi_process_get_replaying()){
1401 if(! op->is_fortran_op)
1402 op->func(invec, inoutvec, len, datatype);
1404 int tmp = smpi_type_c2f(*datatype);
1405 /* Unfortunately, the C and Fortran version of the MPI standard do not agree on the type here,
1406 thus the reinterpret_cast. */
1407 op->func(invec, inoutvec, len, reinterpret_cast<MPI_Datatype*>(&tmp) );
1412 int smpi_type_attr_delete(MPI_Datatype type, int keyval){
1413 smpi_type_key_elem elem =
1414 static_cast<smpi_type_key_elem>(xbt_dict_get_or_null_ext(smpi_type_keyvals, reinterpret_cast<const char*>(&keyval), sizeof(int)));
1417 if(elem->delete_fn!=MPI_NULL_DELETE_FN){
1418 void * value = nullptr;
1420 if(smpi_type_attr_get(type, keyval, &value, &flag)==MPI_SUCCESS){
1421 int ret = elem->delete_fn(type, keyval, value, &flag);
1422 if(ret!=MPI_SUCCESS)
1426 if(type->attributes==nullptr)
1429 xbt_dict_remove_ext(type->attributes, reinterpret_cast<const char*>(&keyval), sizeof(int));
1433 int smpi_type_attr_get(MPI_Datatype type, int keyval, void* attr_value, int* flag){
1434 smpi_type_key_elem elem =
1435 static_cast<smpi_type_key_elem>(xbt_dict_get_or_null_ext(smpi_type_keyvals, reinterpret_cast<const char*>(&keyval), sizeof(int)));
1438 if(type->attributes==nullptr){
1443 *static_cast<void**>(attr_value) = xbt_dict_get_ext(type->attributes, reinterpret_cast<const char*>(&keyval), sizeof(int));
1446 catch (xbt_ex& ex) {
1452 int smpi_type_attr_put(MPI_Datatype type, int keyval, void* attr_value){
1453 if(smpi_type_keyvals==nullptr)
1454 smpi_type_keyvals = xbt_dict_new_homogeneous(nullptr);
1455 smpi_type_key_elem elem =
1456 static_cast<smpi_type_key_elem>(xbt_dict_get_or_null_ext(smpi_type_keyvals, reinterpret_cast<const char*>(&keyval), sizeof(int)));
1460 void* value = nullptr;
1461 smpi_type_attr_get(type, keyval, &value, &flag);
1462 if(flag!=0 && elem->delete_fn!=MPI_NULL_DELETE_FN){
1463 int ret = elem->delete_fn(type, keyval, value, &flag);
1464 if(ret!=MPI_SUCCESS)
1467 if(type->attributes==nullptr)
1468 type->attributes = xbt_dict_new_homogeneous(nullptr);
1470 xbt_dict_set_ext(type->attributes, reinterpret_cast<const char*>(&keyval), sizeof(int), attr_value, nullptr);
1474 int smpi_type_keyval_create(MPI_Type_copy_attr_function* copy_fn, MPI_Type_delete_attr_function* delete_fn, int* keyval,
1476 if(smpi_type_keyvals==nullptr)
1477 smpi_type_keyvals = xbt_dict_new_homogeneous(nullptr);
1479 smpi_type_key_elem value = (smpi_type_key_elem) xbt_new0(s_smpi_mpi_type_key_elem_t,1);
1481 value->copy_fn=copy_fn;
1482 value->delete_fn=delete_fn;
1484 *keyval = type_keyval_id;
1485 xbt_dict_set_ext(smpi_type_keyvals,reinterpret_cast<const char*>(keyval), sizeof(int),reinterpret_cast<void*>(value), nullptr);
1490 int smpi_type_keyval_free(int* keyval){
1491 smpi_type_key_elem elem =
1492 static_cast<smpi_type_key_elem>(xbt_dict_get_or_null_ext(smpi_type_keyvals, reinterpret_cast<const char*>(keyval), sizeof(int)));
1496 xbt_dict_remove_ext(smpi_type_keyvals, reinterpret_cast<const char*>(keyval), sizeof(int));
1501 int smpi_mpi_pack(void* inbuf, int incount, MPI_Datatype type, void* outbuf, int outcount, int* position,MPI_Comm comm){
1502 size_t size = smpi_datatype_size(type);
1503 if (outcount - *position < incount*static_cast<int>(size))
1504 return MPI_ERR_BUFFER;
1505 smpi_datatype_copy(inbuf, incount, type, static_cast<char*>(outbuf) + *position, outcount, MPI_CHAR);
1506 *position += incount * size;
1510 int smpi_mpi_unpack(void* inbuf, int insize, int* position, void* outbuf, int outcount, MPI_Datatype type,MPI_Comm comm){
1511 int size = static_cast<int>(smpi_datatype_size(type));
1512 if (outcount*size> insize)
1513 return MPI_ERR_BUFFER;
1514 smpi_datatype_copy(static_cast<char*>(inbuf) + *position, insize, MPI_CHAR, outbuf, outcount, type);
1515 *position += outcount * size;