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_subtype*/ \
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_subtype*/ \
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));
181 if (datatype->sizeof_subtype){
182 (*new_t)->substruct=xbt_malloc(datatype->sizeof_subtype);
183 memcpy((*new_t)->substruct, datatype->substruct, datatype->sizeof_subtype);
186 (*new_t)->name = xbt_strdup(datatype->name);
187 if(datatype->attributes !=NULL){
188 (*new_t)->attributes=xbt_dict_new();
189 xbt_dict_cursor_t cursor = NULL;
194 xbt_dict_foreach(datatype->attributes, cursor, key, value_in){
195 smpi_type_key_elem elem =
196 static_cast<smpi_type_key_elem>(xbt_dict_get_or_null_ext(smpi_type_keyvals, (const char*)key, sizeof(int)));
197 if(elem && elem->copy_fn!=MPI_NULL_COPY_FN){
198 ret = elem->copy_fn(datatype, *key, NULL, value_in, &value_out, &flag );
199 if(ret!=MPI_SUCCESS){
200 *new_t=MPI_DATATYPE_NULL;
204 xbt_dict_set_ext((*new_t)->attributes, (const char*)key, sizeof(int),value_out, NULL);
211 int smpi_datatype_extent(MPI_Datatype datatype, MPI_Aint * lb, MPI_Aint * extent)
213 if(datatype == MPI_DATATYPE_NULL){
219 *extent = datatype->ub - datatype->lb;
223 MPI_Aint smpi_datatype_get_extent(MPI_Datatype datatype){
224 if(datatype == MPI_DATATYPE_NULL){
227 return datatype->ub - datatype->lb;
230 void smpi_datatype_get_name(MPI_Datatype datatype, char* name, int* length){
231 *length = strlen(datatype->name);
232 strcpy(name, datatype->name);
235 void smpi_datatype_set_name(MPI_Datatype datatype, char* name){
236 datatype->name = xbt_strdup(name);;
239 int smpi_datatype_copy(void *sendbuf, int sendcount, MPI_Datatype sendtype,
240 void *recvbuf, int recvcount, MPI_Datatype recvtype)
243 if(smpi_privatize_global_variables){
244 smpi_switch_data_segment(smpi_process_index());
246 /* First check if we really have something to do */
247 if (recvcount > 0 && recvbuf != sendbuf) {
248 /* FIXME: treat packed cases */
249 sendcount *= smpi_datatype_size(sendtype);
250 recvcount *= smpi_datatype_size(recvtype);
251 count = sendcount < recvcount ? sendcount : recvcount;
253 if(sendtype->sizeof_subtype == 0 && recvtype->sizeof_subtype == 0) {
254 if(!smpi_process_get_replaying()) memcpy(recvbuf, sendbuf, count);
256 else if (sendtype->sizeof_subtype == 0)
258 s_smpi_subtype_t *subtype = static_cast<s_smpi_subtype_t*>(recvtype->substruct);
259 subtype->unserialize( sendbuf, recvbuf, recvcount/smpi_datatype_size(recvtype), subtype, MPI_REPLACE);
261 else if (recvtype->sizeof_subtype == 0)
263 s_smpi_subtype_t *subtype = static_cast<s_smpi_subtype_t*>(sendtype->substruct);
264 subtype->serialize(sendbuf, recvbuf, sendcount/smpi_datatype_size(sendtype), subtype);
266 s_smpi_subtype_t *subtype = static_cast<s_smpi_subtype_t*>(sendtype->substruct);
268 void * buf_tmp = xbt_malloc(count);
270 subtype->serialize( sendbuf, buf_tmp,count/smpi_datatype_size(sendtype), subtype);
271 subtype = static_cast<s_smpi_subtype_t*>(recvtype->substruct);
272 subtype->unserialize( buf_tmp, recvbuf,count/smpi_datatype_size(recvtype), subtype, MPI_REPLACE);
278 return sendcount > recvcount ? MPI_ERR_TRUNCATE : MPI_SUCCESS;
282 * Copies noncontiguous data into contiguous memory.
283 * @param contiguous_vector - output vector
284 * @param noncontiguous_vector - input vector
285 * @param type - pointer contening :
286 * - stride - stride of between noncontiguous data
287 * - block_length - the width or height of blocked matrix
288 * - count - the number of rows of matrix
290 void serialize_vector( const void *noncontiguous_vector, void *contiguous_vector, int count, void *type)
292 s_smpi_mpi_vector_t* type_c = (s_smpi_mpi_vector_t*)type;
294 char* contiguous_vector_char = (char*)contiguous_vector;
295 char* noncontiguous_vector_char = (char*)noncontiguous_vector;
297 for (i = 0; i < type_c->block_count * count; i++) {
298 if (type_c->old_type->sizeof_subtype == 0)
299 memcpy(contiguous_vector_char, noncontiguous_vector_char, type_c->block_length * type_c->size_oldtype);
301 ((s_smpi_subtype_t*)type_c->old_type->substruct)->serialize( noncontiguous_vector_char,
302 contiguous_vector_char,
303 type_c->block_length, type_c->old_type->substruct);
305 contiguous_vector_char += type_c->block_length*type_c->size_oldtype;
306 if((i+1)%type_c->block_count ==0)
307 noncontiguous_vector_char += type_c->block_length*smpi_datatype_get_extent(type_c->old_type);
309 noncontiguous_vector_char += type_c->block_stride*smpi_datatype_get_extent(type_c->old_type);
314 * Copies contiguous data into noncontiguous memory.
315 * @param noncontiguous_vector - output vector
316 * @param contiguous_vector - input vector
317 * @param type - pointer contening :
318 * - stride - stride of between noncontiguous data
319 * - block_length - the width or height of blocked matrix
320 * - count - the number of rows of matrix
322 void unserialize_vector( const void *contiguous_vector, void *noncontiguous_vector, int count, void *type, MPI_Op op)
324 s_smpi_mpi_vector_t* type_c = (s_smpi_mpi_vector_t*)type;
327 char* contiguous_vector_char = (char*)contiguous_vector;
328 char* noncontiguous_vector_char = (char*)noncontiguous_vector;
330 for (i = 0; i < type_c->block_count * count; i++) {
331 if (type_c->old_type->sizeof_subtype == 0)
332 smpi_op_apply(op, contiguous_vector_char, noncontiguous_vector_char, &type_c->block_length,
334 /* memcpy(noncontiguous_vector_char,
335 contiguous_vector_char, type_c->block_length * type_c->size_oldtype);*/
337 ((s_smpi_subtype_t*)type_c->old_type->substruct)->unserialize(contiguous_vector_char, noncontiguous_vector_char,
338 type_c->block_length,type_c->old_type->substruct,
340 contiguous_vector_char += type_c->block_length*type_c->size_oldtype;
341 if((i+1)%type_c->block_count ==0)
342 noncontiguous_vector_char += type_c->block_length*smpi_datatype_get_extent(type_c->old_type);
344 noncontiguous_vector_char += type_c->block_stride*smpi_datatype_get_extent(type_c->old_type);
348 /* Create a Sub type vector to be able to serialize and unserialize it the structure s_smpi_mpi_vector_t is derived
349 * from s_smpi_subtype which required the functions unserialize and serialize */
350 s_smpi_mpi_vector_t* smpi_datatype_vector_create( int block_stride, int block_length, int block_count,
351 MPI_Datatype old_type, int size_oldtype){
352 s_smpi_mpi_vector_t *new_t= xbt_new(s_smpi_mpi_vector_t,1);
353 new_t->base.serialize = &serialize_vector;
354 new_t->base.unserialize = &unserialize_vector;
355 new_t->base.subtype_free = &free_vector;
356 new_t->block_stride = block_stride;
357 new_t->block_length = block_length;
358 new_t->block_count = block_count;
359 smpi_datatype_use(old_type);
360 new_t->old_type = old_type;
361 new_t->size_oldtype = size_oldtype;
365 void smpi_datatype_create(MPI_Datatype* new_type, int size,int lb, int ub, int sizeof_subtype, void *struct_type,
367 MPI_Datatype new_t= xbt_new(s_smpi_mpi_datatype_t,1);
370 new_t->sizeof_subtype = size>0? sizeof_subtype:0;
373 new_t->flags = flags;
374 new_t->substruct = struct_type;
376 new_t->attributes=NULL;
381 MC_ignore(&(new_t->in_use), sizeof(new_t->in_use));
385 void smpi_datatype_free(MPI_Datatype* type){
386 xbt_assert((*type)->in_use >= 0);
387 if((*type)->attributes !=NULL){
388 xbt_dict_cursor_t cursor = NULL;
392 xbt_dict_foreach((*type)->attributes, cursor, key, value){
393 smpi_type_key_elem elem =
394 static_cast<smpi_type_key_elem>(xbt_dict_get_or_null_ext(smpi_type_keyvals, (const char*)key, sizeof(int)));
395 if(elem && elem->delete_fn)
396 elem->delete_fn(*type,*key, value, &flag);
400 if((*type)->flags & DT_FLAG_PREDEFINED)return;
402 //if still used, mark for deletion
403 if((*type)->in_use!=0){
404 (*type)->flags |=DT_FLAG_DESTROYED;
408 if ((*type)->sizeof_subtype != 0){
409 ((s_smpi_subtype_t *)(*type)->substruct)->subtype_free(type);
410 xbt_free((*type)->substruct);
412 if ((*type)->name != NULL){
413 xbt_free((*type)->name);
416 *type = MPI_DATATYPE_NULL;
419 void smpi_datatype_use(MPI_Datatype type){
420 if(type)type->in_use++;
424 MC_ignore(&(type->in_use), sizeof(type->in_use));
428 void smpi_datatype_unuse(MPI_Datatype type){
429 if (type->in_use > 0)
432 if(type && type->in_use == 0 && (type->flags & DT_FLAG_DESTROYED))
433 smpi_datatype_free(&type);
437 MC_ignore(&(type->in_use), sizeof(type->in_use));
441 /*Contiguous Implementation*/
443 /* Copies noncontiguous data into contiguous memory.
444 * @param contiguous_hvector - output hvector
445 * @param noncontiguous_hvector - input hvector
446 * @param type - pointer contening :
447 * - stride - stride of between noncontiguous data, in bytes
448 * - block_length - the width or height of blocked matrix
449 * - count - the number of rows of matrix
451 void serialize_contiguous( const void *noncontiguous_hvector, void *contiguous_hvector, int count, void *type)
453 s_smpi_mpi_contiguous_t* type_c = (s_smpi_mpi_contiguous_t*)type;
454 char* contiguous_vector_char = (char*)contiguous_hvector;
455 char* noncontiguous_vector_char = (char*)noncontiguous_hvector+type_c->lb;
456 memcpy(contiguous_vector_char, noncontiguous_vector_char, count* type_c->block_count * type_c->size_oldtype);
458 /* Copies contiguous data into noncontiguous memory.
459 * @param noncontiguous_vector - output hvector
460 * @param contiguous_vector - 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 unserialize_contiguous(const void *contiguous_vector, void *noncontiguous_vector, int count, void *type, MPI_Op op)
468 s_smpi_mpi_contiguous_t* type_c = (s_smpi_mpi_contiguous_t*)type;
469 char* contiguous_vector_char = (char*)contiguous_vector;
470 char* noncontiguous_vector_char = (char*)noncontiguous_vector+type_c->lb;
471 int n= count* type_c->block_count;
472 smpi_op_apply(op, contiguous_vector_char, noncontiguous_vector_char, &n, &type_c->old_type);
473 /*memcpy(noncontiguous_vector_char, contiguous_vector_char, count* type_c->block_count * type_c->size_oldtype);*/
476 void free_contiguous(MPI_Datatype* d){
477 smpi_datatype_unuse(((s_smpi_mpi_indexed_t *)(*d)->substruct)->old_type);
480 /* Create a Sub type contiguous to be able to serialize and unserialize it the structure s_smpi_mpi_contiguous_t is
481 * erived from s_smpi_subtype which required the functions unserialize and serialize */
482 s_smpi_mpi_contiguous_t* smpi_datatype_contiguous_create( MPI_Aint lb, int block_count, MPI_Datatype old_type,
484 s_smpi_mpi_contiguous_t *new_t= xbt_new(s_smpi_mpi_contiguous_t,1);
485 new_t->base.serialize = &serialize_contiguous;
486 new_t->base.unserialize = &unserialize_contiguous;
487 new_t->base.subtype_free = &free_contiguous;
489 new_t->block_count = block_count;
490 new_t->old_type = old_type;
491 new_t->size_oldtype = size_oldtype;
492 smpi_datatype_use(old_type);
496 int smpi_datatype_contiguous(int count, MPI_Datatype old_type, MPI_Datatype* new_type, MPI_Aint lb)
499 if(old_type->sizeof_subtype){
500 //handle this case as a hvector with stride equals to the extent of the datatype
501 return smpi_datatype_hvector(count, 1, smpi_datatype_get_extent(old_type), old_type, new_type);
504 s_smpi_mpi_contiguous_t* subtype = smpi_datatype_contiguous_create( lb, count, old_type,smpi_datatype_size(old_type));
506 smpi_datatype_create(new_type, count * smpi_datatype_size(old_type),lb,lb + count * smpi_datatype_size(old_type),
507 sizeof(s_smpi_mpi_contiguous_t),subtype, DT_FLAG_CONTIGUOUS);
512 int smpi_datatype_vector(int count, int blocklen, int stride, MPI_Datatype old_type, MPI_Datatype* new_type)
515 if (blocklen<0) return MPI_ERR_ARG;
519 lb=smpi_datatype_lb(old_type);
520 ub=((count-1)*stride+blocklen-1)*smpi_datatype_get_extent(old_type)+smpi_datatype_ub(old_type);
522 if(old_type->sizeof_subtype || stride != blocklen){
524 s_smpi_mpi_vector_t* subtype = smpi_datatype_vector_create(stride, blocklen, count, old_type,
525 smpi_datatype_size(old_type));
526 smpi_datatype_create(new_type, count * (blocklen) * smpi_datatype_size(old_type), lb, ub, sizeof(s_smpi_mpi_vector_t), subtype,
530 /* in this situation the data are contignous thus it's not required to serialize and unserialize it*/
531 smpi_datatype_create(new_type, count * blocklen * smpi_datatype_size(old_type), 0, ((count -1) * stride + blocklen)*
532 smpi_datatype_size(old_type), 0, NULL, DT_FLAG_VECTOR|DT_FLAG_CONTIGUOUS);
538 void free_vector(MPI_Datatype* d){
539 smpi_datatype_unuse(((s_smpi_mpi_indexed_t *)(*d)->substruct)->old_type);
542 /* Hvector Implementation - Vector with stride in bytes */
544 /* Copies noncontiguous data into contiguous memory.
545 * @param contiguous_hvector - output hvector
546 * @param noncontiguous_hvector - input hvector
547 * @param type - pointer contening :
548 * - stride - stride of between noncontiguous data, in bytes
549 * - block_length - the width or height of blocked matrix
550 * - count - the number of rows of matrix
552 void serialize_hvector( const void *noncontiguous_hvector, void *contiguous_hvector, int count, void *type)
554 s_smpi_mpi_hvector_t* type_c = (s_smpi_mpi_hvector_t*)type;
556 char* contiguous_vector_char = (char*)contiguous_hvector;
557 char* noncontiguous_vector_char = (char*)noncontiguous_hvector;
559 for (i = 0; i < type_c->block_count * count; i++) {
560 if (type_c->old_type->sizeof_subtype == 0)
561 memcpy(contiguous_vector_char, noncontiguous_vector_char, type_c->block_length * type_c->size_oldtype);
563 ((s_smpi_subtype_t*)type_c->old_type->substruct)->serialize( noncontiguous_vector_char,
564 contiguous_vector_char,
565 type_c->block_length, type_c->old_type->substruct);
567 contiguous_vector_char += type_c->block_length*type_c->size_oldtype;
568 if((i+1)%type_c->block_count ==0)
569 noncontiguous_vector_char += type_c->block_length*type_c->size_oldtype;
571 noncontiguous_vector_char += type_c->block_stride;
574 /* Copies contiguous data into noncontiguous memory.
575 * @param noncontiguous_vector - output hvector
576 * @param contiguous_vector - input hvector
577 * @param type - pointer contening :
578 * - stride - stride of between noncontiguous data, in bytes
579 * - block_length - the width or height of blocked matrix
580 * - count - the number of rows of matrix
582 void unserialize_hvector( const void *contiguous_vector, void *noncontiguous_vector, int count, void *type, MPI_Op op)
584 s_smpi_mpi_hvector_t* type_c = (s_smpi_mpi_hvector_t*)type;
587 char* contiguous_vector_char = (char*)contiguous_vector;
588 char* noncontiguous_vector_char = (char*)noncontiguous_vector;
590 for (i = 0; i < type_c->block_count * count; i++) {
591 if (type_c->old_type->sizeof_subtype == 0)
592 smpi_op_apply(op, contiguous_vector_char, noncontiguous_vector_char, &type_c->block_length, &type_c->old_type);
593 /*memcpy(noncontiguous_vector_char,
594 contiguous_vector_char, type_c->block_length * type_c->size_oldtype);*/
596 ((s_smpi_subtype_t*)type_c->old_type->substruct)->unserialize( contiguous_vector_char, noncontiguous_vector_char,
597 type_c->block_length, type_c->old_type->substruct,
599 contiguous_vector_char += type_c->block_length*type_c->size_oldtype;
600 if((i+1)%type_c->block_count ==0)
601 noncontiguous_vector_char += type_c->block_length*type_c->size_oldtype;
603 noncontiguous_vector_char += type_c->block_stride;
607 /* Create a Sub type vector to be able to serialize and unserialize it the structure s_smpi_mpi_vector_t is derived
608 * from s_smpi_subtype which required the functions unserialize and serialize
611 s_smpi_mpi_hvector_t* smpi_datatype_hvector_create( MPI_Aint block_stride, int block_length, int block_count,
612 MPI_Datatype old_type, int size_oldtype){
613 s_smpi_mpi_hvector_t *new_t= xbt_new(s_smpi_mpi_hvector_t,1);
614 new_t->base.serialize = &serialize_hvector;
615 new_t->base.unserialize = &unserialize_hvector;
616 new_t->base.subtype_free = &free_hvector;
617 new_t->block_stride = block_stride;
618 new_t->block_length = block_length;
619 new_t->block_count = block_count;
620 new_t->old_type = old_type;
621 new_t->size_oldtype = size_oldtype;
622 smpi_datatype_use(old_type);
626 //do nothing for vector types
627 void free_hvector(MPI_Datatype* d){
628 smpi_datatype_unuse(((s_smpi_mpi_indexed_t *)(*d)->substruct)->old_type);
631 int smpi_datatype_hvector(int count, int blocklen, MPI_Aint stride, MPI_Datatype old_type, MPI_Datatype* new_type)
634 if (blocklen<0) return MPI_ERR_ARG;
638 lb=smpi_datatype_lb(old_type);
639 ub=((count-1)*stride)+(blocklen-1)*smpi_datatype_get_extent(old_type)+smpi_datatype_ub(old_type);
641 if(old_type->sizeof_subtype || stride != blocklen*smpi_datatype_get_extent(old_type)){
642 s_smpi_mpi_hvector_t* subtype = smpi_datatype_hvector_create( stride, blocklen, count, old_type,
643 smpi_datatype_size(old_type));
645 smpi_datatype_create(new_type, count * blocklen * smpi_datatype_size(old_type), lb,ub, sizeof(s_smpi_mpi_hvector_t), subtype, DT_FLAG_VECTOR);
648 smpi_datatype_create(new_type, count * blocklen * smpi_datatype_size(old_type),0,count * blocklen *
649 smpi_datatype_size(old_type), 0, NULL, DT_FLAG_VECTOR|DT_FLAG_CONTIGUOUS);
655 /* Indexed Implementation */
657 /* Copies noncontiguous data into contiguous memory.
658 * @param contiguous_indexed - output indexed
659 * @param noncontiguous_indexed - input indexed
660 * @param type - pointer contening :
661 * - block_lengths - the width or height of blocked matrix
662 * - block_indices - indices of each data, in element
663 * - count - the number of rows of matrix
665 void serialize_indexed( const void *noncontiguous_indexed, void *contiguous_indexed, int count, void *type)
667 s_smpi_mpi_indexed_t* type_c = (s_smpi_mpi_indexed_t*)type;
669 char* contiguous_indexed_char = (char*)contiguous_indexed;
670 char* noncontiguous_indexed_char = (char*)noncontiguous_indexed+type_c->block_indices[0] * type_c->size_oldtype;
671 for(j=0; j<count;j++){
672 for (i = 0; i < type_c->block_count; i++) {
673 if (type_c->old_type->sizeof_subtype == 0)
674 memcpy(contiguous_indexed_char, noncontiguous_indexed_char, type_c->block_lengths[i] * type_c->size_oldtype);
676 ((s_smpi_subtype_t*)type_c->old_type->substruct)->serialize( noncontiguous_indexed_char,
677 contiguous_indexed_char,
678 type_c->block_lengths[i],
679 type_c->old_type->substruct);
681 contiguous_indexed_char += type_c->block_lengths[i]*type_c->size_oldtype;
682 if (i<type_c->block_count-1)
683 noncontiguous_indexed_char =
684 (char*)noncontiguous_indexed + type_c->block_indices[i+1]*smpi_datatype_get_extent(type_c->old_type);
686 noncontiguous_indexed_char += type_c->block_lengths[i]*smpi_datatype_get_extent(type_c->old_type);
688 noncontiguous_indexed=(void*)noncontiguous_indexed_char;
691 /* Copies contiguous data into noncontiguous memory.
692 * @param noncontiguous_indexed - output indexed
693 * @param contiguous_indexed - input indexed
694 * @param type - pointer contening :
695 * - block_lengths - the width or height of blocked matrix
696 * - block_indices - indices of each data, in element
697 * - count - the number of rows of matrix
699 void unserialize_indexed( const void *contiguous_indexed, void *noncontiguous_indexed, int count, void *type, MPI_Op op)
701 s_smpi_mpi_indexed_t* type_c = (s_smpi_mpi_indexed_t*)type;
703 char* contiguous_indexed_char = (char*)contiguous_indexed;
704 char* noncontiguous_indexed_char =
705 (char*)noncontiguous_indexed+type_c->block_indices[0]*smpi_datatype_get_extent(type_c->old_type);
706 for(j=0; j<count;j++){
707 for (i = 0; i < type_c->block_count; i++) {
708 if (type_c->old_type->sizeof_subtype == 0)
709 smpi_op_apply(op, contiguous_indexed_char, noncontiguous_indexed_char, &type_c->block_lengths[i],
711 /*memcpy(noncontiguous_indexed_char ,
712 contiguous_indexed_char, type_c->block_lengths[i] * type_c->size_oldtype);*/
714 ((s_smpi_subtype_t*)type_c->old_type->substruct)->unserialize( contiguous_indexed_char,
715 noncontiguous_indexed_char,
716 type_c->block_lengths[i],
717 type_c->old_type->substruct, op);
719 contiguous_indexed_char += type_c->block_lengths[i]*type_c->size_oldtype;
720 if (i<type_c->block_count-1)
721 noncontiguous_indexed_char =
722 (char*)noncontiguous_indexed + type_c->block_indices[i+1]*smpi_datatype_get_extent(type_c->old_type);
724 noncontiguous_indexed_char += type_c->block_lengths[i]*smpi_datatype_get_extent(type_c->old_type);
726 noncontiguous_indexed=(void*)noncontiguous_indexed_char;
730 void free_indexed(MPI_Datatype* type){
731 xbt_free(((s_smpi_mpi_indexed_t *)(*type)->substruct)->block_lengths);
732 xbt_free(((s_smpi_mpi_indexed_t *)(*type)->substruct)->block_indices);
733 smpi_datatype_unuse(((s_smpi_mpi_indexed_t *)(*type)->substruct)->old_type);
736 /* Create a Sub type indexed to be able to serialize and unserialize it the structure s_smpi_mpi_indexed_t is derived
737 * from s_smpi_subtype which required the functions unserialize and serialize */
738 s_smpi_mpi_indexed_t* smpi_datatype_indexed_create( int* block_lengths, int* block_indices, int block_count,
739 MPI_Datatype old_type, int size_oldtype){
740 s_smpi_mpi_indexed_t *new_t= xbt_new(s_smpi_mpi_indexed_t,1);
741 new_t->base.serialize = &serialize_indexed;
742 new_t->base.unserialize = &unserialize_indexed;
743 new_t->base.subtype_free = &free_indexed;
744 //TODO : add a custom function for each time to clean these
745 new_t->block_lengths= xbt_new(int, block_count);
746 new_t->block_indices= xbt_new(int, block_count);
748 for(i=0;i<block_count;i++){
749 new_t->block_lengths[i]=block_lengths[i];
750 new_t->block_indices[i]=block_indices[i];
752 new_t->block_count = block_count;
753 smpi_datatype_use(old_type);
754 new_t->old_type = old_type;
755 new_t->size_oldtype = size_oldtype;
759 int smpi_datatype_indexed(int count, int* blocklens, int* indices, MPI_Datatype old_type, MPI_Datatype* new_type)
768 lb=indices[0]*smpi_datatype_get_extent(old_type);
769 ub=indices[0]*smpi_datatype_get_extent(old_type) + blocklens[0]*smpi_datatype_ub(old_type);
772 for(i=0; i< count; i++){
775 size += blocklens[i];
777 if(indices[i]*smpi_datatype_get_extent(old_type)+smpi_datatype_lb(old_type)<lb)
778 lb = indices[i]*smpi_datatype_get_extent(old_type)+smpi_datatype_lb(old_type);
779 if(indices[i]*smpi_datatype_get_extent(old_type)+blocklens[i]*smpi_datatype_ub(old_type)>ub)
780 ub = indices[i]*smpi_datatype_get_extent(old_type)+blocklens[i]*smpi_datatype_ub(old_type);
782 if ( (i< count -1) && (indices[i]+blocklens[i] != indices[i+1]) )contiguous=0;
784 if (old_type->sizeof_subtype != 0)
788 s_smpi_mpi_indexed_t* subtype = smpi_datatype_indexed_create( blocklens, indices, count, old_type,
789 smpi_datatype_size(old_type));
790 smpi_datatype_create(new_type, size * smpi_datatype_size(old_type),lb,ub,sizeof(s_smpi_mpi_indexed_t), subtype, DT_FLAG_DATA);
792 s_smpi_mpi_contiguous_t* subtype = smpi_datatype_contiguous_create( lb, size, old_type,
793 smpi_datatype_size(old_type));
794 smpi_datatype_create(new_type, size * smpi_datatype_size(old_type), lb, ub, sizeof(s_smpi_mpi_contiguous_t), subtype,
795 DT_FLAG_DATA|DT_FLAG_CONTIGUOUS);
800 /* Hindexed Implementation - Indexed with indices in bytes */
802 /* Copies noncontiguous data into contiguous memory.
803 * @param contiguous_hindexed - output hindexed
804 * @param noncontiguous_hindexed - input hindexed
805 * @param type - pointer contening :
806 * - block_lengths - the width or height of blocked matrix
807 * - block_indices - indices of each data, in bytes
808 * - count - the number of rows of matrix
810 void serialize_hindexed( const void *noncontiguous_hindexed, void *contiguous_hindexed, int count, void *type)
812 s_smpi_mpi_hindexed_t* type_c = (s_smpi_mpi_hindexed_t*)type;
814 char* contiguous_hindexed_char = (char*)contiguous_hindexed;
815 char* noncontiguous_hindexed_char = (char*)noncontiguous_hindexed+ type_c->block_indices[0];
816 for(j=0; j<count;j++){
817 for (i = 0; i < type_c->block_count; i++) {
818 if (type_c->old_type->sizeof_subtype == 0)
819 memcpy(contiguous_hindexed_char, noncontiguous_hindexed_char, type_c->block_lengths[i] * type_c->size_oldtype);
821 ((s_smpi_subtype_t*)type_c->old_type->substruct)->serialize( noncontiguous_hindexed_char,
822 contiguous_hindexed_char,
823 type_c->block_lengths[i],
824 type_c->old_type->substruct);
826 contiguous_hindexed_char += type_c->block_lengths[i]*type_c->size_oldtype;
827 if (i<type_c->block_count-1)
828 noncontiguous_hindexed_char = (char*)noncontiguous_hindexed + type_c->block_indices[i+1];
830 noncontiguous_hindexed_char += type_c->block_lengths[i]*smpi_datatype_get_extent(type_c->old_type);
832 noncontiguous_hindexed=(void*)noncontiguous_hindexed_char;
835 /* Copies contiguous data into noncontiguous memory.
836 * @param noncontiguous_hindexed - output hindexed
837 * @param contiguous_hindexed - input hindexed
838 * @param type - pointer contening :
839 * - block_lengths - the width or height of blocked matrix
840 * - block_indices - indices of each data, in bytes
841 * - count - the number of rows of matrix
843 void unserialize_hindexed( const void *contiguous_hindexed, void *noncontiguous_hindexed, int count, void *type,
846 s_smpi_mpi_hindexed_t* type_c = (s_smpi_mpi_hindexed_t*)type;
849 char* contiguous_hindexed_char = (char*)contiguous_hindexed;
850 char* noncontiguous_hindexed_char = (char*)noncontiguous_hindexed+ type_c->block_indices[0];
851 for(j=0; j<count;j++){
852 for (i = 0; i < type_c->block_count; i++) {
853 if (type_c->old_type->sizeof_subtype == 0)
854 smpi_op_apply(op, contiguous_hindexed_char, noncontiguous_hindexed_char, &type_c->block_lengths[i],
856 /*memcpy(noncontiguous_hindexed_char,contiguous_hindexed_char,type_c->block_lengths[i]*type_c->size_oldtype);*/
858 ((s_smpi_subtype_t*)type_c->old_type->substruct)->unserialize( contiguous_hindexed_char,
859 noncontiguous_hindexed_char,
860 type_c->block_lengths[i],
861 type_c->old_type->substruct, op);
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;
873 void free_hindexed(MPI_Datatype* type){
874 xbt_free(((s_smpi_mpi_hindexed_t *)(*type)->substruct)->block_lengths);
875 xbt_free(((s_smpi_mpi_hindexed_t *)(*type)->substruct)->block_indices);
876 smpi_datatype_unuse(((s_smpi_mpi_indexed_t *)(*type)->substruct)->old_type);
879 /* Create a Sub type hindexed to be able to serialize and unserialize it the structure s_smpi_mpi_hindexed_t is derived
880 * from s_smpi_subtype which required the functions unserialize and serialize
882 s_smpi_mpi_hindexed_t* smpi_datatype_hindexed_create( int* block_lengths, MPI_Aint* block_indices, int block_count,
883 MPI_Datatype old_type, int size_oldtype){
884 s_smpi_mpi_hindexed_t *new_t= xbt_new(s_smpi_mpi_hindexed_t,1);
885 new_t->base.serialize = &serialize_hindexed;
886 new_t->base.unserialize = &unserialize_hindexed;
887 new_t->base.subtype_free = &free_hindexed;
888 //TODO : add a custom function for each time to clean these
889 new_t->block_lengths= xbt_new(int, block_count);
890 new_t->block_indices= xbt_new(MPI_Aint, block_count);
892 for(i=0;i<block_count;i++){
893 new_t->block_lengths[i]=block_lengths[i];
894 new_t->block_indices[i]=block_indices[i];
896 new_t->block_count = block_count;
897 new_t->old_type = old_type;
898 new_t->size_oldtype = size_oldtype;
902 int smpi_datatype_hindexed(int count, int* blocklens, MPI_Aint* indices, MPI_Datatype old_type, MPI_Datatype* new_type)
911 lb=indices[0] + smpi_datatype_lb(old_type);
912 ub=indices[0] + blocklens[0]*smpi_datatype_ub(old_type);
914 for(i=0; i< count; i++){
917 size += blocklens[i];
919 if(indices[i]+smpi_datatype_lb(old_type)<lb) lb = indices[i]+smpi_datatype_lb(old_type);
920 if(indices[i]+blocklens[i]*smpi_datatype_ub(old_type)>ub) ub = indices[i]+blocklens[i]*smpi_datatype_ub(old_type);
922 if ( (i< count -1) && (indices[i]+blocklens[i]*static_cast<int>(smpi_datatype_size(old_type)) != indices[i+1]) )
925 if (old_type->sizeof_subtype != 0 || lb!=0)
929 s_smpi_mpi_hindexed_t* subtype = smpi_datatype_hindexed_create( blocklens, indices, count, old_type,
930 smpi_datatype_size(old_type));
931 smpi_datatype_create(new_type, size * smpi_datatype_size(old_type), lb, ub ,sizeof(s_smpi_mpi_hindexed_t), subtype, DT_FLAG_DATA);
933 s_smpi_mpi_contiguous_t* subtype = smpi_datatype_contiguous_create(lb,size, old_type, smpi_datatype_size(old_type));
934 smpi_datatype_create(new_type, size * smpi_datatype_size(old_type), 0,size * smpi_datatype_size(old_type),
935 1, subtype, DT_FLAG_DATA|DT_FLAG_CONTIGUOUS);
941 /* struct Implementation - Indexed with indices in bytes */
943 /* Copies noncontiguous data into contiguous memory.
944 * @param contiguous_struct - output struct
945 * @param noncontiguous_struct - input struct
946 * @param type - pointer contening :
947 * - stride - stride of between noncontiguous data
948 * - block_length - the width or height of blocked matrix
949 * - count - the number of rows of matrix
951 void serialize_struct( const void *noncontiguous_struct, void *contiguous_struct, int count, void *type)
953 s_smpi_mpi_struct_t* type_c = (s_smpi_mpi_struct_t*)type;
955 char* contiguous_struct_char = (char*)contiguous_struct;
956 char* noncontiguous_struct_char = (char*)noncontiguous_struct+ type_c->block_indices[0];
957 for(j=0; j<count;j++){
958 for (i = 0; i < type_c->block_count; i++) {
959 if (type_c->old_types[i]->sizeof_subtype == 0)
960 memcpy(contiguous_struct_char, noncontiguous_struct_char,
961 type_c->block_lengths[i] * smpi_datatype_size(type_c->old_types[i]));
963 ((s_smpi_subtype_t*)type_c->old_types[i]->substruct)->serialize( noncontiguous_struct_char,
964 contiguous_struct_char,
965 type_c->block_lengths[i],
966 type_c->old_types[i]->substruct);
969 contiguous_struct_char += type_c->block_lengths[i]*smpi_datatype_size(type_c->old_types[i]);
970 if (i<type_c->block_count-1)
971 noncontiguous_struct_char = (char*)noncontiguous_struct + type_c->block_indices[i+1];
972 else //let's hope this is MPI_UB ?
973 noncontiguous_struct_char += type_c->block_lengths[i]*smpi_datatype_get_extent(type_c->old_types[i]);
975 noncontiguous_struct=(void*)noncontiguous_struct_char;
979 /* Copies contiguous data into noncontiguous memory.
980 * @param noncontiguous_struct - output struct
981 * @param contiguous_struct - input struct
982 * @param type - pointer contening :
983 * - stride - stride of between noncontiguous data
984 * - block_length - the width or height of blocked matrix
985 * - count - the number of rows of matrix
987 void unserialize_struct( const void *contiguous_struct, void *noncontiguous_struct, int count, void *type, MPI_Op op)
989 s_smpi_mpi_struct_t* type_c = (s_smpi_mpi_struct_t*)type;
992 char* contiguous_struct_char = (char*)contiguous_struct;
993 char* noncontiguous_struct_char = (char*)noncontiguous_struct+ type_c->block_indices[0];
994 for(j=0; j<count;j++){
995 for (i = 0; i < type_c->block_count; i++) {
996 if (type_c->old_types[i]->sizeof_subtype == 0)
997 smpi_op_apply(op, contiguous_struct_char, noncontiguous_struct_char, &type_c->block_lengths[i],
998 & type_c->old_types[i]);
999 /*memcpy(noncontiguous_struct_char,
1000 contiguous_struct_char, type_c->block_lengths[i] * smpi_datatype_size(type_c->old_types[i]));*/
1002 ((s_smpi_subtype_t*)type_c->old_types[i]->substruct)->unserialize( contiguous_struct_char,
1003 noncontiguous_struct_char,
1004 type_c->block_lengths[i],
1005 type_c->old_types[i]->substruct, op);
1007 contiguous_struct_char += type_c->block_lengths[i]*smpi_datatype_size(type_c->old_types[i]);
1008 if (i<type_c->block_count-1)
1009 noncontiguous_struct_char = (char*)noncontiguous_struct + type_c->block_indices[i+1];
1011 noncontiguous_struct_char += type_c->block_lengths[i]*smpi_datatype_get_extent(type_c->old_types[i]);
1013 noncontiguous_struct=(void*)noncontiguous_struct_char;
1017 void free_struct(MPI_Datatype* type){
1018 xbt_free(((s_smpi_mpi_struct_t *)(*type)->substruct)->block_lengths);
1019 xbt_free(((s_smpi_mpi_struct_t *)(*type)->substruct)->block_indices);
1021 for (i = 0; i < ((s_smpi_mpi_struct_t *)(*type)->substruct)->block_count; i++)
1022 smpi_datatype_unuse(((s_smpi_mpi_struct_t *)(*type)->substruct)->old_types[i]);
1023 xbt_free(((s_smpi_mpi_struct_t *)(*type)->substruct)->old_types);
1026 /* Create a Sub type struct to be able to serialize and unserialize it the structure s_smpi_mpi_struct_t is derived
1027 * from s_smpi_subtype which required the functions unserialize and serialize
1029 s_smpi_mpi_struct_t* smpi_datatype_struct_create( int* block_lengths, MPI_Aint* block_indices, int block_count,
1030 MPI_Datatype* old_types){
1031 s_smpi_mpi_struct_t *new_t= xbt_new(s_smpi_mpi_struct_t,1);
1032 new_t->base.serialize = &serialize_struct;
1033 new_t->base.unserialize = &unserialize_struct;
1034 new_t->base.subtype_free = &free_struct;
1035 //TODO : add a custom function for each time to clean these
1036 new_t->block_lengths= xbt_new(int, block_count);
1037 new_t->block_indices= xbt_new(MPI_Aint, block_count);
1038 new_t->old_types= xbt_new(MPI_Datatype, block_count);
1040 for(i=0;i<block_count;i++){
1041 new_t->block_lengths[i]=block_lengths[i];
1042 new_t->block_indices[i]=block_indices[i];
1043 new_t->old_types[i]=old_types[i];
1044 smpi_datatype_use(new_t->old_types[i]);
1046 //new_t->block_lengths = block_lengths;
1047 //new_t->block_indices = block_indices;
1048 new_t->block_count = block_count;
1049 //new_t->old_types = old_types;
1053 int smpi_datatype_struct(int count, int* blocklens, MPI_Aint* indices, MPI_Datatype* old_types, MPI_Datatype* new_type)
1062 lb=indices[0] + smpi_datatype_lb(old_types[0]);
1063 ub=indices[0] + blocklens[0]*smpi_datatype_ub(old_types[0]);
1067 for(i=0; i< count; i++){
1070 if (old_types[i]->sizeof_subtype != 0)
1073 size += blocklens[i]*smpi_datatype_size(old_types[i]);
1074 if (old_types[i]==MPI_LB){
1078 if (old_types[i]==MPI_UB){
1083 if(!forced_lb && indices[i]+smpi_datatype_lb(old_types[i])<lb) lb = indices[i];
1084 if(!forced_ub && indices[i]+blocklens[i]*smpi_datatype_ub(old_types[i])>ub)
1085 ub = indices[i]+blocklens[i]*smpi_datatype_ub(old_types[i]);
1087 if ( (i< count -1) && (indices[i]+blocklens[i]*static_cast<int>(smpi_datatype_size(old_types[i])) != indices[i+1]) )
1092 s_smpi_mpi_struct_t* subtype = smpi_datatype_struct_create( blocklens, indices, count, old_types);
1094 smpi_datatype_create(new_type, size, lb, ub,sizeof(s_smpi_mpi_struct_t), subtype, DT_FLAG_DATA);
1096 s_smpi_mpi_contiguous_t* subtype = smpi_datatype_contiguous_create( lb, size, MPI_CHAR, 1);
1097 smpi_datatype_create(new_type, size, lb, ub,1, subtype, DT_FLAG_DATA|DT_FLAG_CONTIGUOUS);
1102 void smpi_datatype_commit(MPI_Datatype *datatype)
1104 (*datatype)->flags= ((*datatype)->flags | DT_FLAG_COMMITED);
1107 typedef struct s_smpi_mpi_op {
1108 MPI_User_function *func;
1112 #define MAX_OP(a, b) (b) = (a) < (b) ? (b) : (a)
1113 #define MIN_OP(a, b) (b) = (a) < (b) ? (a) : (b)
1114 #define SUM_OP(a, b) (b) += (a)
1115 #define PROD_OP(a, b) (b) *= (a)
1116 #define LAND_OP(a, b) (b) = (a) && (b)
1117 #define LOR_OP(a, b) (b) = (a) || (b)
1118 #define LXOR_OP(a, b) (b) = (!(a) && (b)) || ((a) && !(b))
1119 #define BAND_OP(a, b) (b) &= (a)
1120 #define BOR_OP(a, b) (b) |= (a)
1121 #define BXOR_OP(a, b) (b) ^= (a)
1122 #define MAXLOC_OP(a, b) (b) = (a.value) < (b.value) ? (b) : (a)
1123 #define MINLOC_OP(a, b) (b) = (a.value) < (b.value) ? (a) : (b)
1125 #define APPLY_FUNC(a, b, length, type, func) \
1128 type* x = (type*)(a); \
1129 type* y = (type*)(b); \
1130 for(i = 0; i < *(length); i++) { \
1135 static void max_func(void *a, void *b, int *length, MPI_Datatype * datatype)
1137 if (*datatype == MPI_CHAR) {
1138 APPLY_FUNC(a, b, length, char, MAX_OP);
1139 } else if (*datatype == MPI_SHORT) {
1140 APPLY_FUNC(a, b, length, short, MAX_OP);
1141 } else if (*datatype == MPI_INT) {
1142 APPLY_FUNC(a, b, length, int, MAX_OP);
1143 } else if (*datatype == MPI_LONG) {
1144 APPLY_FUNC(a, b, length, long, MAX_OP);
1145 } else if (*datatype == MPI_UNSIGNED_SHORT) {
1146 APPLY_FUNC(a, b, length, unsigned short, MAX_OP);
1147 } else if (*datatype == MPI_UNSIGNED) {
1148 APPLY_FUNC(a, b, length, unsigned int, MAX_OP);
1149 } else if (*datatype == MPI_UNSIGNED_LONG) {
1150 APPLY_FUNC(a, b, length, unsigned long, MAX_OP);
1151 } else if (*datatype == MPI_UNSIGNED_CHAR) {
1152 APPLY_FUNC(a, b, length, unsigned char, MAX_OP);
1153 } else if (*datatype == MPI_FLOAT) {
1154 APPLY_FUNC(a, b, length, float, MAX_OP);
1155 } else if (*datatype == MPI_DOUBLE) {
1156 APPLY_FUNC(a, b, length, double, MAX_OP);
1157 } else if (*datatype == MPI_LONG_DOUBLE) {
1158 APPLY_FUNC(a, b, length, long double, MAX_OP);
1162 static void min_func(void *a, void *b, int *length, MPI_Datatype * datatype)
1164 if (*datatype == MPI_CHAR) {
1165 APPLY_FUNC(a, b, length, char, MIN_OP);
1166 } else if (*datatype == MPI_SHORT) {
1167 APPLY_FUNC(a, b, length, short, MIN_OP);
1168 } else if (*datatype == MPI_INT) {
1169 APPLY_FUNC(a, b, length, int, MIN_OP);
1170 } else if (*datatype == MPI_LONG) {
1171 APPLY_FUNC(a, b, length, long, MIN_OP);
1172 } else if (*datatype == MPI_UNSIGNED_SHORT) {
1173 APPLY_FUNC(a, b, length, unsigned short, MIN_OP);
1174 } else if (*datatype == MPI_UNSIGNED) {
1175 APPLY_FUNC(a, b, length, unsigned int, MIN_OP);
1176 } else if (*datatype == MPI_UNSIGNED_LONG) {
1177 APPLY_FUNC(a, b, length, unsigned long, MIN_OP);
1178 } else if (*datatype == MPI_UNSIGNED_CHAR) {
1179 APPLY_FUNC(a, b, length, unsigned char, MIN_OP);
1180 } else if (*datatype == MPI_FLOAT) {
1181 APPLY_FUNC(a, b, length, float, MIN_OP);
1182 } else if (*datatype == MPI_DOUBLE) {
1183 APPLY_FUNC(a, b, length, double, MIN_OP);
1184 } else if (*datatype == MPI_LONG_DOUBLE) {
1185 APPLY_FUNC(a, b, length, long double, MIN_OP);
1189 static void sum_func(void *a, void *b, int *length, MPI_Datatype * datatype)
1191 if (*datatype == MPI_CHAR) {
1192 APPLY_FUNC(a, b, length, char, SUM_OP);
1193 } else if (*datatype == MPI_SHORT) {
1194 APPLY_FUNC(a, b, length, short, SUM_OP);
1195 } else if (*datatype == MPI_INT) {
1196 APPLY_FUNC(a, b, length, int, SUM_OP);
1197 } else if (*datatype == MPI_LONG) {
1198 APPLY_FUNC(a, b, length, long, SUM_OP);
1199 } else if (*datatype == MPI_UNSIGNED_SHORT) {
1200 APPLY_FUNC(a, b, length, unsigned short, SUM_OP);
1201 } else if (*datatype == MPI_UNSIGNED) {
1202 APPLY_FUNC(a, b, length, unsigned int, SUM_OP);
1203 } else if (*datatype == MPI_UNSIGNED_LONG) {
1204 APPLY_FUNC(a, b, length, unsigned long, SUM_OP);
1205 } else if (*datatype == MPI_UNSIGNED_CHAR) {
1206 APPLY_FUNC(a, b, length, unsigned char, SUM_OP);
1207 } else if (*datatype == MPI_FLOAT) {
1208 APPLY_FUNC(a, b, length, float, SUM_OP);
1209 } else if (*datatype == MPI_DOUBLE) {
1210 APPLY_FUNC(a, b, length, double, SUM_OP);
1211 } else if (*datatype == MPI_LONG_DOUBLE) {
1212 APPLY_FUNC(a, b, length, long double, SUM_OP);
1213 } else if (*datatype == MPI_C_FLOAT_COMPLEX) {
1214 APPLY_FUNC(a, b, length, float _Complex, SUM_OP);
1215 } else if (*datatype == MPI_C_DOUBLE_COMPLEX) {
1216 APPLY_FUNC(a, b, length, double _Complex, SUM_OP);
1217 } else if (*datatype == MPI_C_LONG_DOUBLE_COMPLEX) {
1218 APPLY_FUNC(a, b, length, long double _Complex, SUM_OP);
1222 static void prod_func(void *a, void *b, int *length, MPI_Datatype * datatype)
1224 if (*datatype == MPI_CHAR) {
1225 APPLY_FUNC(a, b, length, char, PROD_OP);
1226 } else if (*datatype == MPI_SHORT) {
1227 APPLY_FUNC(a, b, length, short, PROD_OP);
1228 } else if (*datatype == MPI_INT) {
1229 APPLY_FUNC(a, b, length, int, PROD_OP);
1230 } else if (*datatype == MPI_LONG) {
1231 APPLY_FUNC(a, b, length, long, PROD_OP);
1232 } else if (*datatype == MPI_UNSIGNED_SHORT) {
1233 APPLY_FUNC(a, b, length, unsigned short, PROD_OP);
1234 } else if (*datatype == MPI_UNSIGNED) {
1235 APPLY_FUNC(a, b, length, unsigned int, PROD_OP);
1236 } else if (*datatype == MPI_UNSIGNED_LONG) {
1237 APPLY_FUNC(a, b, length, unsigned long, PROD_OP);
1238 } else if (*datatype == MPI_UNSIGNED_CHAR) {
1239 APPLY_FUNC(a, b, length, unsigned char, PROD_OP);
1240 } else if (*datatype == MPI_FLOAT) {
1241 APPLY_FUNC(a, b, length, float, PROD_OP);
1242 } else if (*datatype == MPI_DOUBLE) {
1243 APPLY_FUNC(a, b, length, double, PROD_OP);
1244 } else if (*datatype == MPI_LONG_DOUBLE) {
1245 APPLY_FUNC(a, b, length, long double, PROD_OP);
1246 } else if (*datatype == MPI_C_FLOAT_COMPLEX) {
1247 APPLY_FUNC(a, b, length, float _Complex, PROD_OP);
1248 } else if (*datatype == MPI_C_DOUBLE_COMPLEX) {
1249 APPLY_FUNC(a, b, length, double _Complex, PROD_OP);
1250 } else if (*datatype == MPI_C_LONG_DOUBLE_COMPLEX) {
1251 APPLY_FUNC(a, b, length, long double _Complex, PROD_OP);
1255 static void land_func(void *a, void *b, int *length, MPI_Datatype * datatype)
1257 if (*datatype == MPI_CHAR) {
1258 APPLY_FUNC(a, b, length, char, LAND_OP);
1259 } else if (*datatype == MPI_SHORT) {
1260 APPLY_FUNC(a, b, length, short, LAND_OP);
1261 } else if (*datatype == MPI_INT) {
1262 APPLY_FUNC(a, b, length, int, LAND_OP);
1263 } else if (*datatype == MPI_LONG) {
1264 APPLY_FUNC(a, b, length, long, LAND_OP);
1265 } else if (*datatype == MPI_UNSIGNED_SHORT) {
1266 APPLY_FUNC(a, b, length, unsigned short, LAND_OP);
1267 } else if (*datatype == MPI_UNSIGNED) {
1268 APPLY_FUNC(a, b, length, unsigned int, LAND_OP);
1269 } else if (*datatype == MPI_UNSIGNED_LONG) {
1270 APPLY_FUNC(a, b, length, unsigned long, LAND_OP);
1271 } else if (*datatype == MPI_UNSIGNED_CHAR) {
1272 APPLY_FUNC(a, b, length, unsigned char, LAND_OP);
1273 } else if (*datatype == MPI_C_BOOL) {
1274 APPLY_FUNC(a, b, length, bool, LAND_OP);
1278 static void lor_func(void *a, void *b, int *length, MPI_Datatype * datatype)
1280 if (*datatype == MPI_CHAR) {
1281 APPLY_FUNC(a, b, length, char, LOR_OP);
1282 } else if (*datatype == MPI_SHORT) {
1283 APPLY_FUNC(a, b, length, short, LOR_OP);
1284 } else if (*datatype == MPI_INT) {
1285 APPLY_FUNC(a, b, length, int, LOR_OP);
1286 } else if (*datatype == MPI_LONG) {
1287 APPLY_FUNC(a, b, length, long, LOR_OP);
1288 } else if (*datatype == MPI_UNSIGNED_SHORT) {
1289 APPLY_FUNC(a, b, length, unsigned short, LOR_OP);
1290 } else if (*datatype == MPI_UNSIGNED) {
1291 APPLY_FUNC(a, b, length, unsigned int, LOR_OP);
1292 } else if (*datatype == MPI_UNSIGNED_LONG) {
1293 APPLY_FUNC(a, b, length, unsigned long, LOR_OP);
1294 } else if (*datatype == MPI_UNSIGNED_CHAR) {
1295 APPLY_FUNC(a, b, length, unsigned char, LOR_OP);
1296 } else if (*datatype == MPI_C_BOOL) {
1297 APPLY_FUNC(a, b, length, bool, LOR_OP);
1301 static void lxor_func(void *a, void *b, int *length, MPI_Datatype * datatype)
1303 if (*datatype == MPI_CHAR) {
1304 APPLY_FUNC(a, b, length, char, LXOR_OP);
1305 } else if (*datatype == MPI_SHORT) {
1306 APPLY_FUNC(a, b, length, short, LXOR_OP);
1307 } else if (*datatype == MPI_INT) {
1308 APPLY_FUNC(a, b, length, int, LXOR_OP);
1309 } else if (*datatype == MPI_LONG) {
1310 APPLY_FUNC(a, b, length, long, LXOR_OP);
1311 } else if (*datatype == MPI_UNSIGNED_SHORT) {
1312 APPLY_FUNC(a, b, length, unsigned short, LXOR_OP);
1313 } else if (*datatype == MPI_UNSIGNED) {
1314 APPLY_FUNC(a, b, length, unsigned int, LXOR_OP);
1315 } else if (*datatype == MPI_UNSIGNED_LONG) {
1316 APPLY_FUNC(a, b, length, unsigned long, LXOR_OP);
1317 } else if (*datatype == MPI_UNSIGNED_CHAR) {
1318 APPLY_FUNC(a, b, length, unsigned char, LXOR_OP);
1319 } else if (*datatype == MPI_C_BOOL) {
1320 APPLY_FUNC(a, b, length, bool, LXOR_OP);
1324 static void band_func(void *a, void *b, int *length, MPI_Datatype * datatype)
1326 if (*datatype == MPI_CHAR) {
1327 APPLY_FUNC(a, b, length, char, BAND_OP);
1328 }else if (*datatype == MPI_SHORT) {
1329 APPLY_FUNC(a, b, length, short, BAND_OP);
1330 } else if (*datatype == MPI_INT) {
1331 APPLY_FUNC(a, b, length, int, BAND_OP);
1332 } else if (*datatype == MPI_LONG) {
1333 APPLY_FUNC(a, b, length, long, BAND_OP);
1334 } else if (*datatype == MPI_UNSIGNED_SHORT) {
1335 APPLY_FUNC(a, b, length, unsigned short, BAND_OP);
1336 } else if (*datatype == MPI_UNSIGNED) {
1337 APPLY_FUNC(a, b, length, unsigned int, BAND_OP);
1338 } else if (*datatype == MPI_UNSIGNED_LONG) {
1339 APPLY_FUNC(a, b, length, unsigned long, BAND_OP);
1340 } else if (*datatype == MPI_UNSIGNED_CHAR) {
1341 APPLY_FUNC(a, b, length, unsigned char, BAND_OP);
1342 } else if (*datatype == MPI_BYTE) {
1343 APPLY_FUNC(a, b, length, uint8_t, BAND_OP);
1347 static void bor_func(void *a, void *b, int *length, MPI_Datatype * datatype)
1349 if (*datatype == MPI_CHAR) {
1350 APPLY_FUNC(a, b, length, char, BOR_OP);
1351 } else if (*datatype == MPI_SHORT) {
1352 APPLY_FUNC(a, b, length, short, BOR_OP);
1353 } else if (*datatype == MPI_INT) {
1354 APPLY_FUNC(a, b, length, int, BOR_OP);
1355 } else if (*datatype == MPI_LONG) {
1356 APPLY_FUNC(a, b, length, long, BOR_OP);
1357 } else if (*datatype == MPI_UNSIGNED_SHORT) {
1358 APPLY_FUNC(a, b, length, unsigned short, BOR_OP);
1359 } else if (*datatype == MPI_UNSIGNED) {
1360 APPLY_FUNC(a, b, length, unsigned int, BOR_OP);
1361 } else if (*datatype == MPI_UNSIGNED_LONG) {
1362 APPLY_FUNC(a, b, length, unsigned long, BOR_OP);
1363 } else if (*datatype == MPI_UNSIGNED_CHAR) {
1364 APPLY_FUNC(a, b, length, unsigned char, BOR_OP);
1365 } else if (*datatype == MPI_BYTE) {
1366 APPLY_FUNC(a, b, length, uint8_t, BOR_OP);
1370 static void bxor_func(void *a, void *b, int *length, MPI_Datatype * datatype)
1372 if (*datatype == MPI_CHAR) {
1373 APPLY_FUNC(a, b, length, char, BXOR_OP);
1374 } else if (*datatype == MPI_SHORT) {
1375 APPLY_FUNC(a, b, length, short, BXOR_OP);
1376 } else if (*datatype == MPI_INT) {
1377 APPLY_FUNC(a, b, length, int, BXOR_OP);
1378 } else if (*datatype == MPI_LONG) {
1379 APPLY_FUNC(a, b, length, long, BXOR_OP);
1380 } else if (*datatype == MPI_UNSIGNED_SHORT) {
1381 APPLY_FUNC(a, b, length, unsigned short, BXOR_OP);
1382 } else if (*datatype == MPI_UNSIGNED) {
1383 APPLY_FUNC(a, b, length, unsigned int, BXOR_OP);
1384 } else if (*datatype == MPI_UNSIGNED_LONG) {
1385 APPLY_FUNC(a, b, length, unsigned long, BXOR_OP);
1386 } else if (*datatype == MPI_UNSIGNED_CHAR) {
1387 APPLY_FUNC(a, b, length, unsigned char, BXOR_OP);
1388 } else if (*datatype == MPI_BYTE) {
1389 APPLY_FUNC(a, b, length, uint8_t, BXOR_OP);
1393 static void minloc_func(void *a, void *b, int *length, MPI_Datatype * datatype)
1395 if (*datatype == MPI_FLOAT_INT) {
1396 APPLY_FUNC(a, b, length, float_int, MINLOC_OP);
1397 } else if (*datatype == MPI_LONG_INT) {
1398 APPLY_FUNC(a, b, length, long_int, MINLOC_OP);
1399 } else if (*datatype == MPI_DOUBLE_INT) {
1400 APPLY_FUNC(a, b, length, double_int, MINLOC_OP);
1401 } else if (*datatype == MPI_SHORT_INT) {
1402 APPLY_FUNC(a, b, length, short_int, MINLOC_OP);
1403 } else if (*datatype == MPI_2LONG) {
1404 APPLY_FUNC(a, b, length, long_long, MINLOC_OP);
1405 } else if (*datatype == MPI_2INT) {
1406 APPLY_FUNC(a, b, length, int_int, MINLOC_OP);
1407 } else if (*datatype == MPI_LONG_DOUBLE_INT) {
1408 APPLY_FUNC(a, b, length, long_double_int, MINLOC_OP);
1409 } else if (*datatype == MPI_2FLOAT) {
1410 APPLY_FUNC(a, b, length, float_float, MINLOC_OP);
1411 } else if (*datatype == MPI_2DOUBLE) {
1412 APPLY_FUNC(a, b, length, double_double, MINLOC_OP);
1416 static void maxloc_func(void *a, void *b, int *length, MPI_Datatype * datatype)
1418 if (*datatype == MPI_FLOAT_INT) {
1419 APPLY_FUNC(a, b, length, float_int, MAXLOC_OP);
1420 } else if (*datatype == MPI_LONG_INT) {
1421 APPLY_FUNC(a, b, length, long_int, MAXLOC_OP);
1422 } else if (*datatype == MPI_DOUBLE_INT) {
1423 APPLY_FUNC(a, b, length, double_int, MAXLOC_OP);
1424 } else if (*datatype == MPI_SHORT_INT) {
1425 APPLY_FUNC(a, b, length, short_int, MAXLOC_OP);
1426 } else if (*datatype == MPI_2LONG) {
1427 APPLY_FUNC(a, b, length, long_long, MAXLOC_OP);
1428 } else if (*datatype == MPI_2INT) {
1429 APPLY_FUNC(a, b, length, int_int, MAXLOC_OP);
1430 } else if (*datatype == MPI_LONG_DOUBLE_INT) {
1431 APPLY_FUNC(a, b, length, long_double_int, MAXLOC_OP);
1432 } else if (*datatype == MPI_2FLOAT) {
1433 APPLY_FUNC(a, b, length, float_float, MAXLOC_OP);
1434 } else if (*datatype == MPI_2DOUBLE) {
1435 APPLY_FUNC(a, b, length, double_double, MAXLOC_OP);
1439 static void replace_func(void *a, void *b, int *length, MPI_Datatype * datatype)
1441 memcpy(b, a, *length * smpi_datatype_size(*datatype));
1444 #define CREATE_MPI_OP(name, func) \
1445 static s_smpi_mpi_op_t mpi_##name = { &(func) /* func */, true }; \
1446 MPI_Op name = &mpi_##name;
1448 CREATE_MPI_OP(MPI_MAX, max_func);
1449 CREATE_MPI_OP(MPI_MIN, min_func);
1450 CREATE_MPI_OP(MPI_SUM, sum_func);
1451 CREATE_MPI_OP(MPI_PROD, prod_func);
1452 CREATE_MPI_OP(MPI_LAND, land_func);
1453 CREATE_MPI_OP(MPI_LOR, lor_func);
1454 CREATE_MPI_OP(MPI_LXOR, lxor_func);
1455 CREATE_MPI_OP(MPI_BAND, band_func);
1456 CREATE_MPI_OP(MPI_BOR, bor_func);
1457 CREATE_MPI_OP(MPI_BXOR, bxor_func);
1458 CREATE_MPI_OP(MPI_MAXLOC, maxloc_func);
1459 CREATE_MPI_OP(MPI_MINLOC, minloc_func);
1460 CREATE_MPI_OP(MPI_REPLACE, replace_func);
1462 MPI_Op smpi_op_new(MPI_User_function * function, int commute)
1465 op = xbt_new(s_smpi_mpi_op_t, 1);
1466 op->func = function;
1467 op-> is_commute = commute;
1471 int smpi_op_is_commute(MPI_Op op)
1473 return (op==MPI_OP_NULL) ? 1 : op-> is_commute;
1476 void smpi_op_destroy(MPI_Op op)
1481 void smpi_op_apply(MPI_Op op, void *invec, void *inoutvec, int *len, MPI_Datatype * datatype)
1486 if(smpi_privatize_global_variables){//we need to switch as the called function may silently touch global variables
1487 XBT_DEBUG("Applying operation, switch to the right data frame ");
1488 smpi_switch_data_segment(smpi_process_index());
1491 if(!smpi_process_get_replaying())
1492 op->func(invec, inoutvec, len, datatype);
1495 int smpi_type_attr_delete(MPI_Datatype type, int keyval){
1496 smpi_type_key_elem elem =
1497 static_cast<smpi_type_key_elem>(xbt_dict_get_or_null_ext(smpi_type_keyvals, (const char*)&keyval, sizeof(int)));
1500 if(elem->delete_fn!=MPI_NULL_DELETE_FN){
1503 if(smpi_type_attr_get(type, keyval, &value, &flag)==MPI_SUCCESS){
1504 int ret = elem->delete_fn(type, keyval, value, &flag);
1505 if(ret!=MPI_SUCCESS) return ret;
1508 if(type->attributes==NULL)
1511 xbt_dict_remove_ext(type->attributes, (const char*)&keyval, sizeof(int));
1515 int smpi_type_attr_get(MPI_Datatype type, int keyval, void* attr_value, int* flag){
1516 smpi_type_key_elem elem =
1517 static_cast<smpi_type_key_elem>(xbt_dict_get_or_null_ext(smpi_type_keyvals, (const char*)&keyval, sizeof(int)));
1521 if(type->attributes==NULL){
1526 *(void**)attr_value = xbt_dict_get_ext(type->attributes, (const char*)&keyval, sizeof(int));
1536 int smpi_type_attr_put(MPI_Datatype type, int keyval, void* attr_value){
1537 if(!smpi_type_keyvals)
1538 smpi_type_keyvals = xbt_dict_new();
1539 smpi_type_key_elem elem =
1540 static_cast<smpi_type_key_elem>(xbt_dict_get_or_null_ext(smpi_type_keyvals, (const char*)&keyval, sizeof(int)));
1545 smpi_type_attr_get(type, keyval, &value, &flag);
1546 if(flag && elem->delete_fn!=MPI_NULL_DELETE_FN){
1547 int ret = elem->delete_fn(type, keyval, value, &flag);
1548 if(ret!=MPI_SUCCESS) return ret;
1550 if(type->attributes==NULL)
1551 type->attributes=xbt_dict_new();
1553 xbt_dict_set_ext(type->attributes, (const char*)&keyval, sizeof(int), attr_value, NULL);
1557 int smpi_type_keyval_create(MPI_Type_copy_attr_function* copy_fn, MPI_Type_delete_attr_function* delete_fn, int* keyval,
1559 if(!smpi_type_keyvals)
1560 smpi_type_keyvals = xbt_dict_new();
1562 smpi_type_key_elem value = (smpi_type_key_elem) xbt_new0(s_smpi_mpi_type_key_elem_t,1);
1564 value->copy_fn=copy_fn;
1565 value->delete_fn=delete_fn;
1567 *keyval = type_keyval_id;
1568 xbt_dict_set_ext(smpi_type_keyvals,(const char*)keyval, sizeof(int),(void*)value, NULL);
1573 int smpi_type_keyval_free(int* keyval){
1574 smpi_type_key_elem elem =
1575 static_cast<smpi_type_key_elem>(xbt_dict_get_or_null_ext(smpi_type_keyvals, (const char*)keyval, sizeof(int)));
1579 xbt_dict_remove_ext(smpi_type_keyvals, (const char*)keyval, sizeof(int));
1584 int smpi_mpi_pack(void* inbuf, int incount, MPI_Datatype type, void* outbuf, int outcount, int* position,MPI_Comm comm){
1585 size_t size = smpi_datatype_size(type);
1586 if (outcount - *position < incount*static_cast<int>(size))
1587 return MPI_ERR_BUFFER;
1588 smpi_datatype_copy(inbuf, incount, type, (char*)outbuf + *position, outcount, MPI_CHAR);
1589 *position += incount * size;
1593 int smpi_mpi_unpack(void* inbuf, int insize, int* position, void* outbuf, int outcount, MPI_Datatype type,MPI_Comm comm){
1594 int size = static_cast<int>(smpi_datatype_size(type));
1595 if (outcount*size> insize)
1596 return MPI_ERR_BUFFER;
1597 smpi_datatype_copy((char*)inbuf + *position, insize, MPI_CHAR, outbuf, outcount, type);
1598 *position += outcount * size;