1 /* smpi_mpi_dt.c -- MPI primitives to handle datatypes */
2 /* FIXME: a very incomplete implementation */
4 /* Copyright (c) 2009-2014. 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,
21 "Logging specific to SMPI (datatype)");
23 #define CREATE_MPI_DATATYPE(name, type) \
24 static s_smpi_mpi_datatype_t mpi_##name = { \
25 sizeof(type), /* size */ \
26 0, /*was 1 has_subtype*/ \
28 sizeof(type), /* ub = lb + size */ \
29 DT_FLAG_BASIC, /* flags */ \
30 NULL /* pointer on extended struct*/ \
32 MPI_Datatype name = &mpi_##name;
34 #define CREATE_MPI_DATATYPE_NULL(name) \
35 static s_smpi_mpi_datatype_t mpi_##name = { \
37 0, /*was 1 has_subtype*/ \
39 0, /* ub = lb + size */ \
40 DT_FLAG_BASIC, /* flags */ \
41 NULL /* pointer on extended struct*/ \
43 MPI_Datatype name = &mpi_##name;
45 //The following are datatypes for the MPI functions MPI_MAXLOC and MPI_MINLOC.
79 // Predefined data types
80 CREATE_MPI_DATATYPE(MPI_CHAR, char);
81 CREATE_MPI_DATATYPE(MPI_SHORT, short);
82 CREATE_MPI_DATATYPE(MPI_INT, int);
83 CREATE_MPI_DATATYPE(MPI_LONG, long);
84 CREATE_MPI_DATATYPE(MPI_LONG_LONG, long long);
85 CREATE_MPI_DATATYPE(MPI_SIGNED_CHAR, signed char);
86 CREATE_MPI_DATATYPE(MPI_UNSIGNED_CHAR, unsigned char);
87 CREATE_MPI_DATATYPE(MPI_UNSIGNED_SHORT, unsigned short);
88 CREATE_MPI_DATATYPE(MPI_UNSIGNED, unsigned int);
89 CREATE_MPI_DATATYPE(MPI_UNSIGNED_LONG, unsigned long);
90 CREATE_MPI_DATATYPE(MPI_UNSIGNED_LONG_LONG, unsigned long long);
91 CREATE_MPI_DATATYPE(MPI_FLOAT, float);
92 CREATE_MPI_DATATYPE(MPI_DOUBLE, double);
93 CREATE_MPI_DATATYPE(MPI_LONG_DOUBLE, long double);
94 CREATE_MPI_DATATYPE(MPI_WCHAR, wchar_t);
95 CREATE_MPI_DATATYPE(MPI_C_BOOL, _Bool);
96 CREATE_MPI_DATATYPE(MPI_INT8_T, int8_t);
97 CREATE_MPI_DATATYPE(MPI_INT16_T, int16_t);
98 CREATE_MPI_DATATYPE(MPI_INT32_T, int32_t);
99 CREATE_MPI_DATATYPE(MPI_INT64_T, int64_t);
100 CREATE_MPI_DATATYPE(MPI_UINT8_T, uint8_t);
101 CREATE_MPI_DATATYPE(MPI_UINT16_T, uint16_t);
102 CREATE_MPI_DATATYPE(MPI_UINT32_T, uint32_t);
103 CREATE_MPI_DATATYPE(MPI_UINT64_T, uint64_t);
104 CREATE_MPI_DATATYPE(MPI_C_FLOAT_COMPLEX, float _Complex);
105 CREATE_MPI_DATATYPE(MPI_C_DOUBLE_COMPLEX, double _Complex);
106 CREATE_MPI_DATATYPE(MPI_C_LONG_DOUBLE_COMPLEX, long double _Complex);
107 CREATE_MPI_DATATYPE(MPI_AINT, MPI_Aint);
108 CREATE_MPI_DATATYPE(MPI_OFFSET, MPI_Offset);
110 CREATE_MPI_DATATYPE(MPI_FLOAT_INT, float_int);
111 CREATE_MPI_DATATYPE(MPI_LONG_INT, long_int);
112 CREATE_MPI_DATATYPE(MPI_DOUBLE_INT, double_int);
113 CREATE_MPI_DATATYPE(MPI_SHORT_INT, short_int);
114 CREATE_MPI_DATATYPE(MPI_2INT, int_int);
115 CREATE_MPI_DATATYPE(MPI_2FLOAT, float_float);
116 CREATE_MPI_DATATYPE(MPI_2DOUBLE, double_double);
118 CREATE_MPI_DATATYPE(MPI_LONG_DOUBLE_INT, long_double_int);
120 CREATE_MPI_DATATYPE_NULL(MPI_UB);
121 CREATE_MPI_DATATYPE_NULL(MPI_LB);
122 CREATE_MPI_DATATYPE_NULL(MPI_PACKED);
124 CREATE_MPI_DATATYPE(MPI_PTR, void*);
126 /** Check if the datatype is usable for communications
128 int is_datatype_valid(MPI_Datatype datatype) {
129 return datatype != MPI_DATATYPE_NULL
130 && (datatype->flags & DT_FLAG_COMMITED);
133 size_t smpi_datatype_size(MPI_Datatype datatype)
135 return datatype->size;
138 MPI_Aint smpi_datatype_lb(MPI_Datatype datatype)
143 MPI_Aint smpi_datatype_ub(MPI_Datatype datatype)
148 int smpi_datatype_extent(MPI_Datatype datatype, MPI_Aint * lb,
152 *extent = datatype->ub - datatype->lb;
156 MPI_Aint smpi_datatype_get_extent(MPI_Datatype datatype){
157 return datatype->ub - datatype->lb;
160 int smpi_datatype_copy(void *sendbuf, int sendcount, MPI_Datatype sendtype,
161 void *recvbuf, int recvcount, MPI_Datatype recvtype)
164 if(smpi_privatize_global_variables){
165 switch_data_segment(smpi_process_index());
167 /* First check if we really have something to do */
168 if (recvcount > 0 && recvbuf != sendbuf) {
169 /* FIXME: treat packed cases */
170 sendcount *= smpi_datatype_size(sendtype);
171 recvcount *= smpi_datatype_size(recvtype);
172 count = sendcount < recvcount ? sendcount : recvcount;
174 if(sendtype->has_subtype == 0 && recvtype->has_subtype == 0) {
175 if(!_xbt_replay_is_active()) memcpy(recvbuf, sendbuf, count);
177 else if (sendtype->has_subtype == 0)
179 s_smpi_subtype_t *subtype = recvtype->substruct;
180 subtype->unserialize( sendbuf, recvbuf,1, subtype);
182 else if (recvtype->has_subtype == 0)
184 s_smpi_subtype_t *subtype = sendtype->substruct;
185 subtype->serialize(sendbuf, recvbuf,1, subtype);
187 s_smpi_subtype_t *subtype = sendtype->substruct;
190 void * buf_tmp = xbt_malloc(count);
192 subtype->serialize( sendbuf, buf_tmp,count/smpi_datatype_size(sendtype), subtype);
193 subtype = recvtype->substruct;
194 subtype->unserialize( buf_tmp, recvbuf,count/smpi_datatype_size(recvtype), subtype);
200 return sendcount > recvcount ? MPI_ERR_TRUNCATE : MPI_SUCCESS;
204 * Copies noncontiguous data into contiguous memory.
205 * @param contiguous_vector - output vector
206 * @param noncontiguous_vector - input vector
207 * @param type - pointer contening :
208 * - stride - stride of between noncontiguous data
209 * - block_length - the width or height of blocked matrix
210 * - count - the number of rows of matrix
212 void serialize_vector( const void *noncontiguous_vector,
213 void *contiguous_vector,
217 s_smpi_mpi_vector_t* type_c = (s_smpi_mpi_vector_t*)type;
219 char* contiguous_vector_char = (char*)contiguous_vector;
220 char* noncontiguous_vector_char = (char*)noncontiguous_vector;
222 for (i = 0; i < type_c->block_count * count; i++) {
223 if (type_c->old_type->has_subtype == 0)
224 memcpy(contiguous_vector_char,
225 noncontiguous_vector_char, type_c->block_length * type_c->size_oldtype);
227 ((s_smpi_subtype_t*)type_c->old_type->substruct)->serialize( noncontiguous_vector_char,
228 contiguous_vector_char,
229 type_c->block_length,
230 type_c->old_type->substruct);
232 contiguous_vector_char += type_c->block_length*type_c->size_oldtype;
233 if((i+1)%type_c->block_count ==0)
234 noncontiguous_vector_char += type_c->block_length*smpi_datatype_get_extent(type_c->old_type);
236 noncontiguous_vector_char += type_c->block_stride*smpi_datatype_get_extent(type_c->old_type);
241 * Copies contiguous data into noncontiguous memory.
242 * @param noncontiguous_vector - output vector
243 * @param contiguous_vector - input vector
244 * @param type - pointer contening :
245 * - stride - stride of between noncontiguous data
246 * - block_length - the width or height of blocked matrix
247 * - count - the number of rows of matrix
249 void unserialize_vector( const void *contiguous_vector,
250 void *noncontiguous_vector,
254 s_smpi_mpi_vector_t* type_c = (s_smpi_mpi_vector_t*)type;
257 char* contiguous_vector_char = (char*)contiguous_vector;
258 char* noncontiguous_vector_char = (char*)noncontiguous_vector;
260 for (i = 0; i < type_c->block_count * count; i++) {
261 if (type_c->old_type->has_subtype == 0)
262 memcpy(noncontiguous_vector_char,
263 contiguous_vector_char, type_c->block_length * type_c->size_oldtype);
265 ((s_smpi_subtype_t*)type_c->old_type->substruct)->unserialize( contiguous_vector_char,
266 noncontiguous_vector_char,
267 type_c->block_length,
268 type_c->old_type->substruct);
269 contiguous_vector_char += type_c->block_length*type_c->size_oldtype;
270 if((i+1)%type_c->block_count ==0)
271 noncontiguous_vector_char += type_c->block_length*smpi_datatype_get_extent(type_c->old_type);
273 noncontiguous_vector_char += type_c->block_stride*smpi_datatype_get_extent(type_c->old_type);
278 * Create a Sub type vector to be able to serialize and unserialize it
279 * the structure s_smpi_mpi_vector_t is derived from s_smpi_subtype which
280 * required the functions unserialize and serialize
283 s_smpi_mpi_vector_t* smpi_datatype_vector_create( int block_stride,
286 MPI_Datatype old_type,
288 s_smpi_mpi_vector_t *new_t= xbt_new(s_smpi_mpi_vector_t,1);
289 new_t->base.serialize = &serialize_vector;
290 new_t->base.unserialize = &unserialize_vector;
291 new_t->base.subtype_free = &free_vector;
292 new_t->block_stride = block_stride;
293 new_t->block_length = block_length;
294 new_t->block_count = block_count;
295 new_t->old_type = old_type;
296 new_t->size_oldtype = size_oldtype;
300 void smpi_datatype_create(MPI_Datatype* new_type, int size,int lb, int ub, int has_subtype,
301 void *struct_type, int flags){
302 MPI_Datatype new_t= xbt_new(s_smpi_mpi_datatype_t,1);
304 new_t->has_subtype = size>0? has_subtype:0;
307 new_t->flags = flags;
308 new_t->substruct = struct_type;
314 MC_ignore(&(new_t->in_use), sizeof(new_t->in_use));
318 void smpi_datatype_free(MPI_Datatype* type){
320 if((*type)->flags & DT_FLAG_PREDEFINED)return;
322 //if still used, mark for deletion
323 if((*type)->in_use!=0){
324 (*type)->flags |=DT_FLAG_DESTROYED;
328 if ((*type)->has_subtype == 1){
329 ((s_smpi_subtype_t *)(*type)->substruct)->subtype_free(type);
330 xbt_free((*type)->substruct);
333 *type = MPI_DATATYPE_NULL;
336 void smpi_datatype_use(MPI_Datatype type){
337 if(type)type->in_use++;
341 MC_ignore(&(type->in_use), sizeof(type->in_use));
346 void smpi_datatype_unuse(MPI_Datatype type){
347 if(type && type->in_use-- == 0 && (type->flags & DT_FLAG_DESTROYED))
348 smpi_datatype_free(&type);
352 MC_ignore(&(type->in_use), sizeof(type->in_use));
360 Contiguous Implementation
365 * Copies noncontiguous data into contiguous memory.
366 * @param contiguous_hvector - output hvector
367 * @param noncontiguous_hvector - input hvector
368 * @param type - pointer contening :
369 * - stride - stride of between noncontiguous data, in bytes
370 * - block_length - the width or height of blocked matrix
371 * - count - the number of rows of matrix
373 void serialize_contiguous( const void *noncontiguous_hvector,
374 void *contiguous_hvector,
378 s_smpi_mpi_contiguous_t* type_c = (s_smpi_mpi_contiguous_t*)type;
379 char* contiguous_vector_char = (char*)contiguous_hvector;
380 char* noncontiguous_vector_char = (char*)noncontiguous_hvector+type_c->lb;
381 memcpy(contiguous_vector_char,
382 noncontiguous_vector_char, count* type_c->block_count * type_c->size_oldtype);
385 * Copies contiguous data into noncontiguous memory.
386 * @param noncontiguous_vector - output hvector
387 * @param contiguous_vector - input hvector
388 * @param type - pointer contening :
389 * - stride - stride of between noncontiguous data, in bytes
390 * - block_length - the width or height of blocked matrix
391 * - count - the number of rows of matrix
393 void unserialize_contiguous( const void *contiguous_vector,
394 void *noncontiguous_vector,
398 s_smpi_mpi_contiguous_t* type_c = (s_smpi_mpi_contiguous_t*)type;
399 char* contiguous_vector_char = (char*)contiguous_vector;
400 char* noncontiguous_vector_char = (char*)noncontiguous_vector+type_c->lb;
402 memcpy(noncontiguous_vector_char,
403 contiguous_vector_char, count* type_c->block_count * type_c->size_oldtype);
406 void free_contiguous(MPI_Datatype* d){
410 * Create a Sub type contiguous to be able to serialize and unserialize it
411 * the structure s_smpi_mpi_contiguous_t is derived from s_smpi_subtype which
412 * required the functions unserialize and serialize
415 s_smpi_mpi_contiguous_t* smpi_datatype_contiguous_create( MPI_Aint lb,
417 MPI_Datatype old_type,
419 s_smpi_mpi_contiguous_t *new_t= xbt_new(s_smpi_mpi_contiguous_t,1);
420 new_t->base.serialize = &serialize_contiguous;
421 new_t->base.unserialize = &unserialize_contiguous;
422 new_t->base.subtype_free = &free_contiguous;
424 new_t->block_count = block_count;
425 new_t->old_type = old_type;
426 new_t->size_oldtype = size_oldtype;
433 int smpi_datatype_contiguous(int count, MPI_Datatype old_type, MPI_Datatype* new_type, MPI_Aint lb)
436 if(old_type->has_subtype){
437 //handle this case as a hvector with stride equals to the extent of the datatype
438 return smpi_datatype_hvector(count, 1, smpi_datatype_get_extent(old_type), old_type, new_type);
441 s_smpi_mpi_contiguous_t* subtype = smpi_datatype_contiguous_create( lb,
444 smpi_datatype_size(old_type));
446 smpi_datatype_create(new_type,
447 count * smpi_datatype_size(old_type),
448 lb,lb + count * smpi_datatype_size(old_type),
449 1,subtype, DT_FLAG_CONTIGUOUS);
454 int smpi_datatype_vector(int count, int blocklen, int stride, MPI_Datatype old_type, MPI_Datatype* new_type)
457 if (blocklen<0) return MPI_ERR_ARG;
461 lb=smpi_datatype_lb(old_type);
462 ub=((count-1)*stride+blocklen-1)*smpi_datatype_get_extent(old_type)+smpi_datatype_ub(old_type);
464 if(old_type->has_subtype || stride != blocklen){
467 s_smpi_mpi_vector_t* subtype = smpi_datatype_vector_create( stride,
471 smpi_datatype_size(old_type));
472 smpi_datatype_create(new_type,
473 count * (blocklen) * smpi_datatype_size(old_type), lb,
480 /* in this situation the data are contignous thus it's not
481 * required to serialize and unserialize it*/
482 smpi_datatype_create(new_type, count * blocklen *
483 smpi_datatype_size(old_type), 0, ((count -1) * stride + blocklen)*
484 smpi_datatype_size(old_type),
487 DT_FLAG_VECTOR|DT_FLAG_CONTIGUOUS);
493 void free_vector(MPI_Datatype* d){
497 Hvector Implementation - Vector with stride in bytes
502 * Copies noncontiguous data into contiguous memory.
503 * @param contiguous_hvector - output hvector
504 * @param noncontiguous_hvector - input hvector
505 * @param type - pointer contening :
506 * - stride - stride of between noncontiguous data, in bytes
507 * - block_length - the width or height of blocked matrix
508 * - count - the number of rows of matrix
510 void serialize_hvector( const void *noncontiguous_hvector,
511 void *contiguous_hvector,
515 s_smpi_mpi_hvector_t* type_c = (s_smpi_mpi_hvector_t*)type;
517 char* contiguous_vector_char = (char*)contiguous_hvector;
518 char* noncontiguous_vector_char = (char*)noncontiguous_hvector;
520 for (i = 0; i < type_c->block_count * count; i++) {
521 if (type_c->old_type->has_subtype == 0)
522 memcpy(contiguous_vector_char,
523 noncontiguous_vector_char, type_c->block_length * type_c->size_oldtype);
525 ((s_smpi_subtype_t*)type_c->old_type->substruct)->serialize( noncontiguous_vector_char,
526 contiguous_vector_char,
527 type_c->block_length,
528 type_c->old_type->substruct);
530 contiguous_vector_char += type_c->block_length*type_c->size_oldtype;
531 if((i+1)%type_c->block_count ==0)
532 noncontiguous_vector_char += type_c->block_length*type_c->size_oldtype;
534 noncontiguous_vector_char += type_c->block_stride;
538 * Copies contiguous data into noncontiguous memory.
539 * @param noncontiguous_vector - output hvector
540 * @param contiguous_vector - input hvector
541 * @param type - pointer contening :
542 * - stride - stride of between noncontiguous data, in bytes
543 * - block_length - the width or height of blocked matrix
544 * - count - the number of rows of matrix
546 void unserialize_hvector( const void *contiguous_vector,
547 void *noncontiguous_vector,
551 s_smpi_mpi_hvector_t* type_c = (s_smpi_mpi_hvector_t*)type;
554 char* contiguous_vector_char = (char*)contiguous_vector;
555 char* noncontiguous_vector_char = (char*)noncontiguous_vector;
557 for (i = 0; i < type_c->block_count * count; i++) {
558 if (type_c->old_type->has_subtype == 0)
559 memcpy(noncontiguous_vector_char,
560 contiguous_vector_char, type_c->block_length * type_c->size_oldtype);
562 ((s_smpi_subtype_t*)type_c->old_type->substruct)->unserialize( contiguous_vector_char,
563 noncontiguous_vector_char,
564 type_c->block_length,
565 type_c->old_type->substruct);
566 contiguous_vector_char += type_c->block_length*type_c->size_oldtype;
567 if((i+1)%type_c->block_count ==0)
568 noncontiguous_vector_char += type_c->block_length*type_c->size_oldtype;
570 noncontiguous_vector_char += type_c->block_stride;
575 * Create a Sub type vector to be able to serialize and unserialize it
576 * the structure s_smpi_mpi_vector_t is derived from s_smpi_subtype which
577 * required the functions unserialize and serialize
580 s_smpi_mpi_hvector_t* smpi_datatype_hvector_create( MPI_Aint block_stride,
583 MPI_Datatype old_type,
585 s_smpi_mpi_hvector_t *new_t= xbt_new(s_smpi_mpi_hvector_t,1);
586 new_t->base.serialize = &serialize_hvector;
587 new_t->base.unserialize = &unserialize_hvector;
588 new_t->base.subtype_free = &free_hvector;
589 new_t->block_stride = block_stride;
590 new_t->block_length = block_length;
591 new_t->block_count = block_count;
592 new_t->old_type = old_type;
593 new_t->size_oldtype = size_oldtype;
597 //do nothing for vector types
598 void free_hvector(MPI_Datatype* d){
601 int smpi_datatype_hvector(int count, int blocklen, MPI_Aint stride, MPI_Datatype old_type, MPI_Datatype* new_type)
604 if (blocklen<0) return MPI_ERR_ARG;
608 lb=smpi_datatype_lb(old_type);
609 ub=((count-1)*stride)+(blocklen-1)*smpi_datatype_get_extent(old_type)+smpi_datatype_ub(old_type);
611 if(old_type->has_subtype || stride != blocklen*smpi_datatype_get_extent(old_type)){
612 s_smpi_mpi_hvector_t* subtype = smpi_datatype_hvector_create( stride,
616 smpi_datatype_size(old_type));
618 smpi_datatype_create(new_type, count * blocklen * smpi_datatype_size(old_type),
625 smpi_datatype_create(new_type, count * blocklen *
626 smpi_datatype_size(old_type),0,count * blocklen *
627 smpi_datatype_size(old_type),
630 DT_FLAG_VECTOR|DT_FLAG_CONTIGUOUS);
638 Indexed Implementation
642 * Copies noncontiguous data into contiguous memory.
643 * @param contiguous_indexed - output indexed
644 * @param noncontiguous_indexed - input indexed
645 * @param type - pointer contening :
646 * - block_lengths - the width or height of blocked matrix
647 * - block_indices - indices of each data, in element
648 * - count - the number of rows of matrix
650 void serialize_indexed( const void *noncontiguous_indexed,
651 void *contiguous_indexed,
655 s_smpi_mpi_indexed_t* type_c = (s_smpi_mpi_indexed_t*)type;
657 char* contiguous_indexed_char = (char*)contiguous_indexed;
658 char* noncontiguous_indexed_char = (char*)noncontiguous_indexed+type_c->block_indices[0] * type_c->size_oldtype;
659 for(j=0; j<count;j++){
660 for (i = 0; i < type_c->block_count; i++) {
661 if (type_c->old_type->has_subtype == 0)
662 memcpy(contiguous_indexed_char,
663 noncontiguous_indexed_char, type_c->block_lengths[i] * type_c->size_oldtype);
665 ((s_smpi_subtype_t*)type_c->old_type->substruct)->serialize( noncontiguous_indexed_char,
666 contiguous_indexed_char,
667 type_c->block_lengths[i],
668 type_c->old_type->substruct);
671 contiguous_indexed_char += type_c->block_lengths[i]*type_c->size_oldtype;
672 if (i<type_c->block_count-1)noncontiguous_indexed_char = (char*)noncontiguous_indexed + type_c->block_indices[i+1]*smpi_datatype_get_extent(type_c->old_type);
673 else noncontiguous_indexed_char += type_c->block_lengths[i]*smpi_datatype_get_extent(type_c->old_type);
675 noncontiguous_indexed=(void*)noncontiguous_indexed_char;
679 * Copies contiguous data into noncontiguous memory.
680 * @param noncontiguous_indexed - output indexed
681 * @param contiguous_indexed - input indexed
682 * @param type - pointer contening :
683 * - block_lengths - the width or height of blocked matrix
684 * - block_indices - indices of each data, in element
685 * - count - the number of rows of matrix
687 void unserialize_indexed( const void *contiguous_indexed,
688 void *noncontiguous_indexed,
693 s_smpi_mpi_indexed_t* type_c = (s_smpi_mpi_indexed_t*)type;
695 char* contiguous_indexed_char = (char*)contiguous_indexed;
696 char* noncontiguous_indexed_char = (char*)noncontiguous_indexed+type_c->block_indices[0]*smpi_datatype_get_extent(type_c->old_type);
697 for(j=0; j<count;j++){
698 for (i = 0; i < type_c->block_count; i++) {
699 if (type_c->old_type->has_subtype == 0)
700 memcpy(noncontiguous_indexed_char ,
701 contiguous_indexed_char, type_c->block_lengths[i] * type_c->size_oldtype);
703 ((s_smpi_subtype_t*)type_c->old_type->substruct)->unserialize( contiguous_indexed_char,
704 noncontiguous_indexed_char,
705 type_c->block_lengths[i],
706 type_c->old_type->substruct);
708 contiguous_indexed_char += type_c->block_lengths[i]*type_c->size_oldtype;
709 if (i<type_c->block_count-1)
710 noncontiguous_indexed_char = (char*)noncontiguous_indexed + type_c->block_indices[i+1]*smpi_datatype_get_extent(type_c->old_type);
711 else noncontiguous_indexed_char += type_c->block_lengths[i]*smpi_datatype_get_extent(type_c->old_type);
713 noncontiguous_indexed=(void*)noncontiguous_indexed_char;
717 void free_indexed(MPI_Datatype* type){
718 xbt_free(((s_smpi_mpi_indexed_t *)(*type)->substruct)->block_lengths);
719 xbt_free(((s_smpi_mpi_indexed_t *)(*type)->substruct)->block_indices);
723 * Create a Sub type indexed to be able to serialize and unserialize it
724 * the structure s_smpi_mpi_indexed_t is derived from s_smpi_subtype which
725 * required the functions unserialize and serialize
727 s_smpi_mpi_indexed_t* smpi_datatype_indexed_create( int* block_lengths,
730 MPI_Datatype old_type,
732 s_smpi_mpi_indexed_t *new_t= xbt_new(s_smpi_mpi_indexed_t,1);
733 new_t->base.serialize = &serialize_indexed;
734 new_t->base.unserialize = &unserialize_indexed;
735 new_t->base.subtype_free = &free_indexed;
736 //TODO : add a custom function for each time to clean these
737 new_t->block_lengths= xbt_new(int, block_count);
738 new_t->block_indices= xbt_new(int, block_count);
740 for(i=0;i<block_count;i++){
741 new_t->block_lengths[i]=block_lengths[i];
742 new_t->block_indices[i]=block_indices[i];
744 new_t->block_count = block_count;
745 new_t->old_type = old_type;
746 new_t->size_oldtype = size_oldtype;
751 int smpi_datatype_indexed(int count, int* blocklens, int* indices, MPI_Datatype old_type, MPI_Datatype* new_type)
760 lb=indices[0]*smpi_datatype_get_extent(old_type);
761 ub=indices[0]*smpi_datatype_get_extent(old_type) + blocklens[0]*smpi_datatype_ub(old_type);
764 for(i=0; i< count; i++){
767 size += blocklens[i];
769 if(indices[i]*smpi_datatype_get_extent(old_type)+smpi_datatype_lb(old_type)<lb)
770 lb = indices[i]*smpi_datatype_get_extent(old_type)+smpi_datatype_lb(old_type);
771 if(indices[i]*smpi_datatype_get_extent(old_type)+blocklens[i]*smpi_datatype_ub(old_type)>ub)
772 ub = indices[i]*smpi_datatype_get_extent(old_type)+blocklens[i]*smpi_datatype_ub(old_type);
774 if ( (i< count -1) && (indices[i]+blocklens[i] != indices[i+1]) )contiguous=0;
776 if (old_type->has_subtype == 1)
780 s_smpi_mpi_indexed_t* subtype = smpi_datatype_indexed_create( blocklens,
784 smpi_datatype_size(old_type));
785 smpi_datatype_create(new_type, size *
786 smpi_datatype_size(old_type),lb,ub,1, subtype, DT_FLAG_DATA);
788 s_smpi_mpi_contiguous_t* subtype = smpi_datatype_contiguous_create( lb,
791 smpi_datatype_size(old_type));
792 smpi_datatype_create(new_type, size *
793 smpi_datatype_size(old_type),lb,ub,1, subtype, DT_FLAG_DATA|DT_FLAG_CONTIGUOUS);
801 Hindexed Implementation - Indexed with indices in bytes
805 * Copies noncontiguous data into contiguous memory.
806 * @param contiguous_hindexed - output hindexed
807 * @param noncontiguous_hindexed - input hindexed
808 * @param type - pointer contening :
809 * - block_lengths - the width or height of blocked matrix
810 * - block_indices - indices of each data, in bytes
811 * - count - the number of rows of matrix
813 void serialize_hindexed( const void *noncontiguous_hindexed,
814 void *contiguous_hindexed,
818 s_smpi_mpi_hindexed_t* type_c = (s_smpi_mpi_hindexed_t*)type;
820 char* contiguous_hindexed_char = (char*)contiguous_hindexed;
821 char* noncontiguous_hindexed_char = (char*)noncontiguous_hindexed+ type_c->block_indices[0];
822 for(j=0; j<count;j++){
823 for (i = 0; i < type_c->block_count; i++) {
824 if (type_c->old_type->has_subtype == 0)
825 memcpy(contiguous_hindexed_char,
826 noncontiguous_hindexed_char, type_c->block_lengths[i] * type_c->size_oldtype);
828 ((s_smpi_subtype_t*)type_c->old_type->substruct)->serialize( noncontiguous_hindexed_char,
829 contiguous_hindexed_char,
830 type_c->block_lengths[i],
831 type_c->old_type->substruct);
833 contiguous_hindexed_char += type_c->block_lengths[i]*type_c->size_oldtype;
834 if (i<type_c->block_count-1)noncontiguous_hindexed_char = (char*)noncontiguous_hindexed + type_c->block_indices[i+1];
835 else noncontiguous_hindexed_char += type_c->block_lengths[i]*smpi_datatype_get_extent(type_c->old_type);
837 noncontiguous_hindexed=(void*)noncontiguous_hindexed_char;
841 * Copies contiguous data into noncontiguous memory.
842 * @param noncontiguous_hindexed - output hindexed
843 * @param contiguous_hindexed - input hindexed
844 * @param type - pointer contening :
845 * - block_lengths - the width or height of blocked matrix
846 * - block_indices - indices of each data, in bytes
847 * - count - the number of rows of matrix
849 void unserialize_hindexed( const void *contiguous_hindexed,
850 void *noncontiguous_hindexed,
854 s_smpi_mpi_hindexed_t* type_c = (s_smpi_mpi_hindexed_t*)type;
857 char* contiguous_hindexed_char = (char*)contiguous_hindexed;
858 char* noncontiguous_hindexed_char = (char*)noncontiguous_hindexed+ type_c->block_indices[0];
859 for(j=0; j<count;j++){
860 for (i = 0; i < type_c->block_count; i++) {
861 if (type_c->old_type->has_subtype == 0)
862 memcpy(noncontiguous_hindexed_char,
863 contiguous_hindexed_char, type_c->block_lengths[i] * type_c->size_oldtype);
865 ((s_smpi_subtype_t*)type_c->old_type->substruct)->unserialize( contiguous_hindexed_char,
866 noncontiguous_hindexed_char,
867 type_c->block_lengths[i],
868 type_c->old_type->substruct);
870 contiguous_hindexed_char += type_c->block_lengths[i]*type_c->size_oldtype;
871 if (i<type_c->block_count-1)noncontiguous_hindexed_char = (char*)noncontiguous_hindexed + type_c->block_indices[i+1];
872 else noncontiguous_hindexed_char += type_c->block_lengths[i]*smpi_datatype_get_extent(type_c->old_type);
874 noncontiguous_hindexed=(void*)noncontiguous_hindexed_char;
878 void free_hindexed(MPI_Datatype* type){
879 xbt_free(((s_smpi_mpi_hindexed_t *)(*type)->substruct)->block_lengths);
880 xbt_free(((s_smpi_mpi_hindexed_t *)(*type)->substruct)->block_indices);
884 * Create a Sub type hindexed to be able to serialize and unserialize it
885 * the structure s_smpi_mpi_hindexed_t is derived from s_smpi_subtype which
886 * required the functions unserialize and serialize
888 s_smpi_mpi_hindexed_t* smpi_datatype_hindexed_create( int* block_lengths,
889 MPI_Aint* block_indices,
891 MPI_Datatype old_type,
893 s_smpi_mpi_hindexed_t *new_t= xbt_new(s_smpi_mpi_hindexed_t,1);
894 new_t->base.serialize = &serialize_hindexed;
895 new_t->base.unserialize = &unserialize_hindexed;
896 new_t->base.subtype_free = &free_hindexed;
897 //TODO : add a custom function for each time to clean these
898 new_t->block_lengths= xbt_new(int, block_count);
899 new_t->block_indices= xbt_new(MPI_Aint, block_count);
901 for(i=0;i<block_count;i++){
902 new_t->block_lengths[i]=block_lengths[i];
903 new_t->block_indices[i]=block_indices[i];
905 new_t->block_count = block_count;
906 new_t->old_type = old_type;
907 new_t->size_oldtype = size_oldtype;
912 int smpi_datatype_hindexed(int count, int* blocklens, MPI_Aint* indices, MPI_Datatype old_type, MPI_Datatype* new_type)
921 lb=indices[0] + smpi_datatype_lb(old_type);
922 ub=indices[0] + blocklens[0]*smpi_datatype_ub(old_type);
924 for(i=0; i< count; i++){
927 size += blocklens[i];
929 if(indices[i]+smpi_datatype_lb(old_type)<lb) lb = indices[i]+smpi_datatype_lb(old_type);
930 if(indices[i]+blocklens[i]*smpi_datatype_ub(old_type)>ub) ub = indices[i]+blocklens[i]*smpi_datatype_ub(old_type);
932 if ( (i< count -1) && (indices[i]+blocklens[i]*smpi_datatype_size(old_type) != indices[i+1]) )contiguous=0;
934 if (old_type->has_subtype == 1 || lb!=0)
938 s_smpi_mpi_hindexed_t* subtype = smpi_datatype_hindexed_create( blocklens,
942 smpi_datatype_size(old_type));
943 smpi_datatype_create(new_type, size * smpi_datatype_size(old_type),
946 ,1, subtype, DT_FLAG_DATA);
948 s_smpi_mpi_contiguous_t* subtype = smpi_datatype_contiguous_create( lb,
951 smpi_datatype_size(old_type));
952 smpi_datatype_create(new_type, size * smpi_datatype_size(old_type),
953 0,size * smpi_datatype_size(old_type),
954 1, subtype, DT_FLAG_DATA|DT_FLAG_CONTIGUOUS);
962 struct Implementation - Indexed with indices in bytes
966 * Copies noncontiguous data into contiguous memory.
967 * @param contiguous_struct - output struct
968 * @param noncontiguous_struct - input struct
969 * @param type - pointer contening :
970 * - stride - stride of between noncontiguous data
971 * - block_length - the width or height of blocked matrix
972 * - count - the number of rows of matrix
974 void serialize_struct( const void *noncontiguous_struct,
975 void *contiguous_struct,
979 s_smpi_mpi_struct_t* type_c = (s_smpi_mpi_struct_t*)type;
981 char* contiguous_struct_char = (char*)contiguous_struct;
982 char* noncontiguous_struct_char = (char*)noncontiguous_struct+ type_c->block_indices[0];
983 for(j=0; j<count;j++){
984 for (i = 0; i < type_c->block_count; i++) {
985 if (type_c->old_types[i]->has_subtype == 0)
986 memcpy(contiguous_struct_char,
987 noncontiguous_struct_char, type_c->block_lengths[i] * smpi_datatype_size(type_c->old_types[i]));
989 ((s_smpi_subtype_t*)type_c->old_types[i]->substruct)->serialize( noncontiguous_struct_char,
990 contiguous_struct_char,
991 type_c->block_lengths[i],
992 type_c->old_types[i]->substruct);
995 contiguous_struct_char += type_c->block_lengths[i]*smpi_datatype_size(type_c->old_types[i]);
996 if (i<type_c->block_count-1)noncontiguous_struct_char = (char*)noncontiguous_struct + type_c->block_indices[i+1];
997 else noncontiguous_struct_char += type_c->block_lengths[i]*smpi_datatype_get_extent(type_c->old_types[i]);//let's hope this is MPI_UB ?
999 noncontiguous_struct=(void*)noncontiguous_struct_char;
1003 * Copies contiguous data into noncontiguous memory.
1004 * @param noncontiguous_struct - output struct
1005 * @param contiguous_struct - input struct
1006 * @param type - pointer contening :
1007 * - stride - stride of between noncontiguous data
1008 * - block_length - the width or height of blocked matrix
1009 * - count - the number of rows of matrix
1011 void unserialize_struct( const void *contiguous_struct,
1012 void *noncontiguous_struct,
1016 s_smpi_mpi_struct_t* type_c = (s_smpi_mpi_struct_t*)type;
1019 char* contiguous_struct_char = (char*)contiguous_struct;
1020 char* noncontiguous_struct_char = (char*)noncontiguous_struct+ type_c->block_indices[0];
1021 for(j=0; j<count;j++){
1022 for (i = 0; i < type_c->block_count; i++) {
1023 if (type_c->old_types[i]->has_subtype == 0)
1024 memcpy(noncontiguous_struct_char,
1025 contiguous_struct_char, type_c->block_lengths[i] * smpi_datatype_size(type_c->old_types[i]));
1027 ((s_smpi_subtype_t*)type_c->old_types[i]->substruct)->unserialize( contiguous_struct_char,
1028 noncontiguous_struct_char,
1029 type_c->block_lengths[i],
1030 type_c->old_types[i]->substruct);
1032 contiguous_struct_char += type_c->block_lengths[i]*smpi_datatype_size(type_c->old_types[i]);
1033 if (i<type_c->block_count-1)noncontiguous_struct_char = (char*)noncontiguous_struct + type_c->block_indices[i+1];
1034 else noncontiguous_struct_char += type_c->block_lengths[i]*smpi_datatype_get_extent(type_c->old_types[i]);
1036 noncontiguous_struct=(void*)noncontiguous_struct_char;
1041 void free_struct(MPI_Datatype* type){
1042 xbt_free(((s_smpi_mpi_struct_t *)(*type)->substruct)->block_lengths);
1043 xbt_free(((s_smpi_mpi_struct_t *)(*type)->substruct)->block_indices);
1044 xbt_free(((s_smpi_mpi_struct_t *)(*type)->substruct)->old_types);
1048 * Create a Sub type struct to be able to serialize and unserialize it
1049 * the structure s_smpi_mpi_struct_t is derived from s_smpi_subtype which
1050 * required the functions unserialize and serialize
1052 s_smpi_mpi_struct_t* smpi_datatype_struct_create( int* block_lengths,
1053 MPI_Aint* block_indices,
1055 MPI_Datatype* old_types){
1056 s_smpi_mpi_struct_t *new_t= xbt_new(s_smpi_mpi_struct_t,1);
1057 new_t->base.serialize = &serialize_struct;
1058 new_t->base.unserialize = &unserialize_struct;
1059 new_t->base.subtype_free = &free_struct;
1060 //TODO : add a custom function for each time to clean these
1061 new_t->block_lengths= xbt_new(int, block_count);
1062 new_t->block_indices= xbt_new(MPI_Aint, block_count);
1063 new_t->old_types= xbt_new(MPI_Datatype, block_count);
1065 for(i=0;i<block_count;i++){
1066 new_t->block_lengths[i]=block_lengths[i];
1067 new_t->block_indices[i]=block_indices[i];
1068 new_t->old_types[i]=old_types[i];
1070 //new_t->block_lengths = block_lengths;
1071 //new_t->block_indices = block_indices;
1072 new_t->block_count = block_count;
1073 //new_t->old_types = old_types;
1078 int smpi_datatype_struct(int count, int* blocklens, MPI_Aint* indices, MPI_Datatype* old_types, MPI_Datatype* new_type)
1087 lb=indices[0] + smpi_datatype_lb(old_types[0]);
1088 ub=indices[0] + blocklens[0]*smpi_datatype_ub(old_types[0]);
1092 for(i=0; i< count; i++){
1095 if (old_types[i]->has_subtype == 1)
1098 size += blocklens[i]*smpi_datatype_size(old_types[i]);
1099 if (old_types[i]==MPI_LB){
1103 if (old_types[i]==MPI_UB){
1108 if(!forced_lb && indices[i]+smpi_datatype_lb(old_types[i])<lb) lb = indices[i];
1109 if(!forced_ub && indices[i]+blocklens[i]*smpi_datatype_ub(old_types[i])>ub) ub = indices[i]+blocklens[i]*smpi_datatype_ub(old_types[i]);
1111 if ( (i< count -1) && (indices[i]+blocklens[i]*smpi_datatype_size(old_types[i]) != indices[i+1]) )contiguous=0;
1115 s_smpi_mpi_struct_t* subtype = smpi_datatype_struct_create( blocklens,
1120 smpi_datatype_create(new_type, size, lb, ub,1, subtype, DT_FLAG_DATA);
1122 s_smpi_mpi_contiguous_t* subtype = smpi_datatype_contiguous_create( lb,
1126 smpi_datatype_create(new_type, size, lb, ub,1, subtype, DT_FLAG_DATA|DT_FLAG_CONTIGUOUS);
1131 void smpi_datatype_commit(MPI_Datatype *datatype)
1133 (*datatype)->flags= ((*datatype)->flags | DT_FLAG_COMMITED);
1136 typedef struct s_smpi_mpi_op {
1137 MPI_User_function *func;
1141 #define MAX_OP(a, b) (b) = (a) < (b) ? (b) : (a)
1142 #define MIN_OP(a, b) (b) = (a) < (b) ? (a) : (b)
1143 #define SUM_OP(a, b) (b) += (a)
1144 #define PROD_OP(a, b) (b) *= (a)
1145 #define LAND_OP(a, b) (b) = (a) && (b)
1146 #define LOR_OP(a, b) (b) = (a) || (b)
1147 #define LXOR_OP(a, b) (b) = (!(a) && (b)) || ((a) && !(b))
1148 #define BAND_OP(a, b) (b) &= (a)
1149 #define BOR_OP(a, b) (b) |= (a)
1150 #define BXOR_OP(a, b) (b) ^= (a)
1151 #define MAXLOC_OP(a, b) (b) = (a.value) < (b.value) ? (b) : (a)
1152 #define MINLOC_OP(a, b) (b) = (a.value) < (b.value) ? (a) : (b)
1153 //TODO : MINLOC & MAXLOC
1155 #define APPLY_FUNC(a, b, length, type, func) \
1158 type* x = (type*)(a); \
1159 type* y = (type*)(b); \
1160 for(i = 0; i < *(length); i++) { \
1165 static void max_func(void *a, void *b, int *length,
1166 MPI_Datatype * datatype)
1168 if (*datatype == MPI_CHAR) {
1169 APPLY_FUNC(a, b, length, char, MAX_OP);
1170 } else if (*datatype == MPI_SHORT) {
1171 APPLY_FUNC(a, b, length, short, MAX_OP);
1172 } else if (*datatype == MPI_INT) {
1173 APPLY_FUNC(a, b, length, int, MAX_OP);
1174 } else if (*datatype == MPI_LONG) {
1175 APPLY_FUNC(a, b, length, long, MAX_OP);
1176 } else if (*datatype == MPI_UNSIGNED_SHORT) {
1177 APPLY_FUNC(a, b, length, unsigned short, MAX_OP);
1178 } else if (*datatype == MPI_UNSIGNED) {
1179 APPLY_FUNC(a, b, length, unsigned int, MAX_OP);
1180 } else if (*datatype == MPI_UNSIGNED_LONG) {
1181 APPLY_FUNC(a, b, length, unsigned long, MAX_OP);
1182 } else if (*datatype == MPI_FLOAT) {
1183 APPLY_FUNC(a, b, length, float, MAX_OP);
1184 } else if (*datatype == MPI_DOUBLE) {
1185 APPLY_FUNC(a, b, length, double, MAX_OP);
1186 } else if (*datatype == MPI_LONG_DOUBLE) {
1187 APPLY_FUNC(a, b, length, long double, MAX_OP);
1191 static void min_func(void *a, void *b, int *length,
1192 MPI_Datatype * datatype)
1194 if (*datatype == MPI_CHAR) {
1195 APPLY_FUNC(a, b, length, char, MIN_OP);
1196 } else if (*datatype == MPI_SHORT) {
1197 APPLY_FUNC(a, b, length, short, MIN_OP);
1198 } else if (*datatype == MPI_INT) {
1199 APPLY_FUNC(a, b, length, int, MIN_OP);
1200 } else if (*datatype == MPI_LONG) {
1201 APPLY_FUNC(a, b, length, long, MIN_OP);
1202 } else if (*datatype == MPI_UNSIGNED_SHORT) {
1203 APPLY_FUNC(a, b, length, unsigned short, MIN_OP);
1204 } else if (*datatype == MPI_UNSIGNED) {
1205 APPLY_FUNC(a, b, length, unsigned int, MIN_OP);
1206 } else if (*datatype == MPI_UNSIGNED_LONG) {
1207 APPLY_FUNC(a, b, length, unsigned long, MIN_OP);
1208 } else if (*datatype == MPI_FLOAT) {
1209 APPLY_FUNC(a, b, length, float, MIN_OP);
1210 } else if (*datatype == MPI_DOUBLE) {
1211 APPLY_FUNC(a, b, length, double, MIN_OP);
1212 } else if (*datatype == MPI_LONG_DOUBLE) {
1213 APPLY_FUNC(a, b, length, long double, MIN_OP);
1217 static void sum_func(void *a, void *b, int *length,
1218 MPI_Datatype * datatype)
1220 if (*datatype == MPI_CHAR) {
1221 APPLY_FUNC(a, b, length, char, SUM_OP);
1222 } else if (*datatype == MPI_SHORT) {
1223 APPLY_FUNC(a, b, length, short, SUM_OP);
1224 } else if (*datatype == MPI_INT) {
1225 APPLY_FUNC(a, b, length, int, SUM_OP);
1226 } else if (*datatype == MPI_LONG) {
1227 APPLY_FUNC(a, b, length, long, SUM_OP);
1228 } else if (*datatype == MPI_UNSIGNED_SHORT) {
1229 APPLY_FUNC(a, b, length, unsigned short, SUM_OP);
1230 } else if (*datatype == MPI_UNSIGNED) {
1231 APPLY_FUNC(a, b, length, unsigned int, SUM_OP);
1232 } else if (*datatype == MPI_UNSIGNED_LONG) {
1233 APPLY_FUNC(a, b, length, unsigned long, SUM_OP);
1234 } else if (*datatype == MPI_FLOAT) {
1235 APPLY_FUNC(a, b, length, float, SUM_OP);
1236 } else if (*datatype == MPI_DOUBLE) {
1237 APPLY_FUNC(a, b, length, double, SUM_OP);
1238 } else if (*datatype == MPI_LONG_DOUBLE) {
1239 APPLY_FUNC(a, b, length, long double, SUM_OP);
1240 } else if (*datatype == MPI_C_FLOAT_COMPLEX) {
1241 APPLY_FUNC(a, b, length, float _Complex, SUM_OP);
1242 } else if (*datatype == MPI_C_DOUBLE_COMPLEX) {
1243 APPLY_FUNC(a, b, length, double _Complex, SUM_OP);
1244 } else if (*datatype == MPI_C_LONG_DOUBLE_COMPLEX) {
1245 APPLY_FUNC(a, b, length, long double _Complex, SUM_OP);
1249 static void prod_func(void *a, void *b, int *length,
1250 MPI_Datatype * datatype)
1252 if (*datatype == MPI_CHAR) {
1253 APPLY_FUNC(a, b, length, char, PROD_OP);
1254 } else if (*datatype == MPI_SHORT) {
1255 APPLY_FUNC(a, b, length, short, PROD_OP);
1256 } else if (*datatype == MPI_INT) {
1257 APPLY_FUNC(a, b, length, int, PROD_OP);
1258 } else if (*datatype == MPI_LONG) {
1259 APPLY_FUNC(a, b, length, long, PROD_OP);
1260 } else if (*datatype == MPI_UNSIGNED_SHORT) {
1261 APPLY_FUNC(a, b, length, unsigned short, PROD_OP);
1262 } else if (*datatype == MPI_UNSIGNED) {
1263 APPLY_FUNC(a, b, length, unsigned int, PROD_OP);
1264 } else if (*datatype == MPI_UNSIGNED_LONG) {
1265 APPLY_FUNC(a, b, length, unsigned long, PROD_OP);
1266 } else if (*datatype == MPI_FLOAT) {
1267 APPLY_FUNC(a, b, length, float, PROD_OP);
1268 } else if (*datatype == MPI_DOUBLE) {
1269 APPLY_FUNC(a, b, length, double, PROD_OP);
1270 } else if (*datatype == MPI_LONG_DOUBLE) {
1271 APPLY_FUNC(a, b, length, long double, PROD_OP);
1272 } else if (*datatype == MPI_C_FLOAT_COMPLEX) {
1273 APPLY_FUNC(a, b, length, float _Complex, PROD_OP);
1274 } else if (*datatype == MPI_C_DOUBLE_COMPLEX) {
1275 APPLY_FUNC(a, b, length, double _Complex, PROD_OP);
1276 } else if (*datatype == MPI_C_LONG_DOUBLE_COMPLEX) {
1277 APPLY_FUNC(a, b, length, long double _Complex, PROD_OP);
1281 static void land_func(void *a, void *b, int *length,
1282 MPI_Datatype * datatype)
1284 if (*datatype == MPI_CHAR) {
1285 APPLY_FUNC(a, b, length, char, LAND_OP);
1286 } else if (*datatype == MPI_SHORT) {
1287 APPLY_FUNC(a, b, length, short, LAND_OP);
1288 } else if (*datatype == MPI_INT) {
1289 APPLY_FUNC(a, b, length, int, LAND_OP);
1290 } else if (*datatype == MPI_LONG) {
1291 APPLY_FUNC(a, b, length, long, LAND_OP);
1292 } else if (*datatype == MPI_UNSIGNED_SHORT) {
1293 APPLY_FUNC(a, b, length, unsigned short, LAND_OP);
1294 } else if (*datatype == MPI_UNSIGNED) {
1295 APPLY_FUNC(a, b, length, unsigned int, LAND_OP);
1296 } else if (*datatype == MPI_UNSIGNED_LONG) {
1297 APPLY_FUNC(a, b, length, unsigned long, LAND_OP);
1298 } else if (*datatype == MPI_C_BOOL) {
1299 APPLY_FUNC(a, b, length, _Bool, LAND_OP);
1303 static void lor_func(void *a, void *b, int *length,
1304 MPI_Datatype * datatype)
1306 if (*datatype == MPI_CHAR) {
1307 APPLY_FUNC(a, b, length, char, LOR_OP);
1308 } else if (*datatype == MPI_SHORT) {
1309 APPLY_FUNC(a, b, length, short, LOR_OP);
1310 } else if (*datatype == MPI_INT) {
1311 APPLY_FUNC(a, b, length, int, LOR_OP);
1312 } else if (*datatype == MPI_LONG) {
1313 APPLY_FUNC(a, b, length, long, LOR_OP);
1314 } else if (*datatype == MPI_UNSIGNED_SHORT) {
1315 APPLY_FUNC(a, b, length, unsigned short, LOR_OP);
1316 } else if (*datatype == MPI_UNSIGNED) {
1317 APPLY_FUNC(a, b, length, unsigned int, LOR_OP);
1318 } else if (*datatype == MPI_UNSIGNED_LONG) {
1319 APPLY_FUNC(a, b, length, unsigned long, LOR_OP);
1320 } else if (*datatype == MPI_C_BOOL) {
1321 APPLY_FUNC(a, b, length, _Bool, LOR_OP);
1325 static void lxor_func(void *a, void *b, int *length,
1326 MPI_Datatype * datatype)
1328 if (*datatype == MPI_CHAR) {
1329 APPLY_FUNC(a, b, length, char, LXOR_OP);
1330 } else if (*datatype == MPI_SHORT) {
1331 APPLY_FUNC(a, b, length, short, LXOR_OP);
1332 } else if (*datatype == MPI_INT) {
1333 APPLY_FUNC(a, b, length, int, LXOR_OP);
1334 } else if (*datatype == MPI_LONG) {
1335 APPLY_FUNC(a, b, length, long, LXOR_OP);
1336 } else if (*datatype == MPI_UNSIGNED_SHORT) {
1337 APPLY_FUNC(a, b, length, unsigned short, LXOR_OP);
1338 } else if (*datatype == MPI_UNSIGNED) {
1339 APPLY_FUNC(a, b, length, unsigned int, LXOR_OP);
1340 } else if (*datatype == MPI_UNSIGNED_LONG) {
1341 APPLY_FUNC(a, b, length, unsigned long, LXOR_OP);
1342 } else if (*datatype == MPI_C_BOOL) {
1343 APPLY_FUNC(a, b, length, _Bool, LXOR_OP);
1347 static void band_func(void *a, void *b, int *length,
1348 MPI_Datatype * datatype)
1350 if (*datatype == MPI_CHAR) {
1351 APPLY_FUNC(a, b, length, char, BAND_OP);
1353 if (*datatype == MPI_SHORT) {
1354 APPLY_FUNC(a, b, length, short, BAND_OP);
1355 } else if (*datatype == MPI_INT) {
1356 APPLY_FUNC(a, b, length, int, BAND_OP);
1357 } else if (*datatype == MPI_LONG) {
1358 APPLY_FUNC(a, b, length, long, BAND_OP);
1359 } else if (*datatype == MPI_UNSIGNED_SHORT) {
1360 APPLY_FUNC(a, b, length, unsigned short, BAND_OP);
1361 } else if (*datatype == MPI_UNSIGNED) {
1362 APPLY_FUNC(a, b, length, unsigned int, BAND_OP);
1363 } else if (*datatype == MPI_UNSIGNED_LONG) {
1364 APPLY_FUNC(a, b, length, unsigned long, BAND_OP);
1365 } else if (*datatype == MPI_BYTE) {
1366 APPLY_FUNC(a, b, length, uint8_t, BAND_OP);
1370 static void bor_func(void *a, void *b, int *length,
1371 MPI_Datatype * datatype)
1373 if (*datatype == MPI_CHAR) {
1374 APPLY_FUNC(a, b, length, char, BOR_OP);
1375 } else if (*datatype == MPI_SHORT) {
1376 APPLY_FUNC(a, b, length, short, BOR_OP);
1377 } else if (*datatype == MPI_INT) {
1378 APPLY_FUNC(a, b, length, int, BOR_OP);
1379 } else if (*datatype == MPI_LONG) {
1380 APPLY_FUNC(a, b, length, long, BOR_OP);
1381 } else if (*datatype == MPI_UNSIGNED_SHORT) {
1382 APPLY_FUNC(a, b, length, unsigned short, BOR_OP);
1383 } else if (*datatype == MPI_UNSIGNED) {
1384 APPLY_FUNC(a, b, length, unsigned int, BOR_OP);
1385 } else if (*datatype == MPI_UNSIGNED_LONG) {
1386 APPLY_FUNC(a, b, length, unsigned long, BOR_OP);
1387 } else if (*datatype == MPI_BYTE) {
1388 APPLY_FUNC(a, b, length, uint8_t, BOR_OP);
1392 static void bxor_func(void *a, void *b, int *length,
1393 MPI_Datatype * datatype)
1395 if (*datatype == MPI_CHAR) {
1396 APPLY_FUNC(a, b, length, char, BXOR_OP);
1397 } else if (*datatype == MPI_SHORT) {
1398 APPLY_FUNC(a, b, length, short, BXOR_OP);
1399 } else if (*datatype == MPI_INT) {
1400 APPLY_FUNC(a, b, length, int, BXOR_OP);
1401 } else if (*datatype == MPI_LONG) {
1402 APPLY_FUNC(a, b, length, long, BXOR_OP);
1403 } else if (*datatype == MPI_UNSIGNED_SHORT) {
1404 APPLY_FUNC(a, b, length, unsigned short, BXOR_OP);
1405 } else if (*datatype == MPI_UNSIGNED) {
1406 APPLY_FUNC(a, b, length, unsigned int, BXOR_OP);
1407 } else if (*datatype == MPI_UNSIGNED_LONG) {
1408 APPLY_FUNC(a, b, length, unsigned long, BXOR_OP);
1409 } else if (*datatype == MPI_BYTE) {
1410 APPLY_FUNC(a, b, length, uint8_t, BXOR_OP);
1414 static void minloc_func(void *a, void *b, int *length,
1415 MPI_Datatype * datatype)
1417 if (*datatype == MPI_FLOAT_INT) {
1418 APPLY_FUNC(a, b, length, float_int, MINLOC_OP);
1419 } else if (*datatype == MPI_LONG_INT) {
1420 APPLY_FUNC(a, b, length, long_int, MINLOC_OP);
1421 } else if (*datatype == MPI_DOUBLE_INT) {
1422 APPLY_FUNC(a, b, length, double_int, MINLOC_OP);
1423 } else if (*datatype == MPI_SHORT_INT) {
1424 APPLY_FUNC(a, b, length, short_int, MINLOC_OP);
1425 } else if (*datatype == MPI_2INT) {
1426 APPLY_FUNC(a, b, length, int_int, MINLOC_OP);
1427 } else if (*datatype == MPI_LONG_DOUBLE_INT) {
1428 APPLY_FUNC(a, b, length, long_double_int, MINLOC_OP);
1429 } else if (*datatype == MPI_2FLOAT) {
1430 APPLY_FUNC(a, b, length, float_float, MINLOC_OP);
1431 } else if (*datatype == MPI_2DOUBLE) {
1432 APPLY_FUNC(a, b, length, double_double, MINLOC_OP);
1436 static void maxloc_func(void *a, void *b, int *length,
1437 MPI_Datatype * datatype)
1439 if (*datatype == MPI_FLOAT_INT) {
1440 APPLY_FUNC(a, b, length, float_int, MAXLOC_OP);
1441 } else if (*datatype == MPI_LONG_INT) {
1442 APPLY_FUNC(a, b, length, long_int, MAXLOC_OP);
1443 } else if (*datatype == MPI_DOUBLE_INT) {
1444 APPLY_FUNC(a, b, length, double_int, MAXLOC_OP);
1445 } else if (*datatype == MPI_SHORT_INT) {
1446 APPLY_FUNC(a, b, length, short_int, MAXLOC_OP);
1447 } else if (*datatype == MPI_2INT) {
1448 APPLY_FUNC(a, b, length, int_int, MAXLOC_OP);
1449 } else if (*datatype == MPI_LONG_DOUBLE_INT) {
1450 APPLY_FUNC(a, b, length, long_double_int, MAXLOC_OP);
1451 } else if (*datatype == MPI_2FLOAT) {
1452 APPLY_FUNC(a, b, length, float_float, MAXLOC_OP);
1453 } else if (*datatype == MPI_2DOUBLE) {
1454 APPLY_FUNC(a, b, length, double_double, MAXLOC_OP);
1459 #define CREATE_MPI_OP(name, func) \
1460 static s_smpi_mpi_op_t mpi_##name = { &(func) /* func */, TRUE }; \
1461 MPI_Op name = &mpi_##name;
1463 CREATE_MPI_OP(MPI_MAX, max_func);
1464 CREATE_MPI_OP(MPI_MIN, min_func);
1465 CREATE_MPI_OP(MPI_SUM, sum_func);
1466 CREATE_MPI_OP(MPI_PROD, prod_func);
1467 CREATE_MPI_OP(MPI_LAND, land_func);
1468 CREATE_MPI_OP(MPI_LOR, lor_func);
1469 CREATE_MPI_OP(MPI_LXOR, lxor_func);
1470 CREATE_MPI_OP(MPI_BAND, band_func);
1471 CREATE_MPI_OP(MPI_BOR, bor_func);
1472 CREATE_MPI_OP(MPI_BXOR, bxor_func);
1473 CREATE_MPI_OP(MPI_MAXLOC, maxloc_func);
1474 CREATE_MPI_OP(MPI_MINLOC, minloc_func);
1476 MPI_Op smpi_op_new(MPI_User_function * function, int commute)
1479 op = xbt_new(s_smpi_mpi_op_t, 1);
1480 op->func = function;
1481 op-> is_commute = commute;
1485 int smpi_op_is_commute(MPI_Op op)
1487 return (op==MPI_OP_NULL) ? 1 : op-> is_commute;
1490 void smpi_op_destroy(MPI_Op op)
1495 void smpi_op_apply(MPI_Op op, void *invec, void *inoutvec, int *len,
1496 MPI_Datatype * datatype)
1498 if(smpi_privatize_global_variables){ //we need to switch here, as the called function may silently touch global variables
1499 XBT_VERB("Applying operation, switch to the right data frame ");
1500 switch_data_segment(smpi_process_index());
1503 if(!_xbt_replay_is_active())
1504 op->func(invec, inoutvec, len, datatype);