src/smpi/smpi_mpi_dt.cpp

   1 /* smpi_mpi_dt.c -- MPI primitives to handle datatypes                      */
   2 /* Copyright (c) 2009-2015. The SimGrid Team.
   3  * All rights reserved.                                                     */
   4
   5 /* This program is free software; you can redistribute it and/or modify it
   6  * under the terms of the license (GNU LGPL) which comes with this package. */
   7
   8 #include <stdio.h>
   9 #include <stdlib.h>
  10 #include <string.h>
  11 #include <limits.h>
  12 #include "private.h"
  13 #include "smpi_mpi_dt_private.h"
  14 #include "mc/mc.h"
  15 #include "xbt/replay.h"
  16 #include <xbt/ex.hpp>
  17 #include "simgrid/modelchecker.h"
  18
  19 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_mpi_dt, smpi, "Logging specific to SMPI (datatype)");
  20
  21 xbt_dict_t smpi_type_keyvals = nullptr;
  22 int type_keyval_id=0;//avoid collisions
  23
  24 #define CREATE_MPI_DATATYPE(name, type)               \
  25   static s_smpi_mpi_datatype_t mpi_##name = {         \
  26     (char*) # name,                                   \
  27     sizeof(type),   /* size */                        \
  28     0,              /*was 1 sizeof_substruct*/        \
  29     0,              /* lb */                          \
  30     sizeof(type),   /* ub = lb + size */              \
  31     DT_FLAG_BASIC,  /* flags */                       \
  32     nullptr,        /* attributes */                  \
  33     nullptr,        /* pointer on extended struct*/   \
  34     0               /* in_use counter */              \
  35   };                                                  \
  36 const MPI_Datatype name = &mpi_##name;
  37
  38 #define CREATE_MPI_DATATYPE_NULL(name)                \
  39   static s_smpi_mpi_datatype_t mpi_##name = {         \
  40     (char*) # name,                                   \
  41     0,              /* size */                        \
  42     0,              /* was 1 sizeof_substruct*/       \
  43     0,              /* lb */                          \
  44     0,              /* ub = lb + size */              \
  45     DT_FLAG_BASIC,  /* flags */                       \
  46     nullptr,        /* attributes */                  \
  47     nullptr,        /* pointer on extended struct*/   \
  48     0               /* in_use counter */              \
  49   };                                                  \
  50 const MPI_Datatype name = &mpi_##name;
  51
  52 //The following are datatypes for the MPI functions MPI_MAXLOC and MPI_MINLOC.
  53 typedef struct {
  54   float value;
  55   int index;
  56 } float_int;
  57 typedef struct {
  58   float value;
  59   float index;
  60 } float_float;
  61 typedef struct {
  62   long value;
  63   long index;
  64 } long_long;
  65 typedef struct {
  66   double value;
  67   double index;
  68 } double_double;
  69 typedef struct {
  70   long value;
  71   int index;
  72 } long_int;
  73 typedef struct {
  74   double value;
  75   int index;
  76 } double_int;
  77 typedef struct {
  78   short value;
  79   int index;
  80 } short_int;
  81 typedef struct {
  82   int value;
  83   int index;
  84 } int_int;
  85 typedef struct {
  86   long double value;
  87   int index;
  88 } long_double_int;
  89 typedef struct {
  90   int64_t value;
  91   int64_t index;
  92 } integer128_t;
  93 // Predefined data types
  94 CREATE_MPI_DATATYPE(MPI_CHAR, char);
  95 CREATE_MPI_DATATYPE(MPI_SHORT, short);
  96 CREATE_MPI_DATATYPE(MPI_INT, int);
  97 CREATE_MPI_DATATYPE(MPI_LONG, long);
  98 CREATE_MPI_DATATYPE(MPI_LONG_LONG, long long);
  99 CREATE_MPI_DATATYPE(MPI_SIGNED_CHAR, signed char);
 100 CREATE_MPI_DATATYPE(MPI_UNSIGNED_CHAR, unsigned char);
 101 CREATE_MPI_DATATYPE(MPI_UNSIGNED_SHORT, unsigned short);
 102 CREATE_MPI_DATATYPE(MPI_UNSIGNED, unsigned int);
 103 CREATE_MPI_DATATYPE(MPI_UNSIGNED_LONG, unsigned long);
 104 CREATE_MPI_DATATYPE(MPI_UNSIGNED_LONG_LONG, unsigned long long);
 105 CREATE_MPI_DATATYPE(MPI_FLOAT, float);
 106 CREATE_MPI_DATATYPE(MPI_DOUBLE, double);
 107 CREATE_MPI_DATATYPE(MPI_LONG_DOUBLE, long double);
 108 CREATE_MPI_DATATYPE(MPI_WCHAR, wchar_t);
 109 CREATE_MPI_DATATYPE(MPI_C_BOOL, bool);
 110 CREATE_MPI_DATATYPE(MPI_BYTE, int8_t);
 111 CREATE_MPI_DATATYPE(MPI_INT8_T, int8_t);
 112 CREATE_MPI_DATATYPE(MPI_INT16_T, int16_t);
 113 CREATE_MPI_DATATYPE(MPI_INT32_T, int32_t);
 114 CREATE_MPI_DATATYPE(MPI_INT64_T, int64_t);
 115 CREATE_MPI_DATATYPE(MPI_UINT8_T, uint8_t);
 116 CREATE_MPI_DATATYPE(MPI_UINT16_T, uint16_t);
 117 CREATE_MPI_DATATYPE(MPI_UINT32_T, uint32_t);
 118 CREATE_MPI_DATATYPE(MPI_UINT64_T, uint64_t);
 119 CREATE_MPI_DATATYPE(MPI_C_FLOAT_COMPLEX, float _Complex);
 120 CREATE_MPI_DATATYPE(MPI_C_DOUBLE_COMPLEX, double _Complex);
 121 CREATE_MPI_DATATYPE(MPI_C_LONG_DOUBLE_COMPLEX, long double _Complex);
 122 CREATE_MPI_DATATYPE(MPI_AINT, MPI_Aint);
 123 CREATE_MPI_DATATYPE(MPI_OFFSET, MPI_Offset);
 124
 125 CREATE_MPI_DATATYPE(MPI_FLOAT_INT, float_int);
 126 CREATE_MPI_DATATYPE(MPI_LONG_INT, long_int);
 127 CREATE_MPI_DATATYPE(MPI_DOUBLE_INT, double_int);
 128 CREATE_MPI_DATATYPE(MPI_SHORT_INT, short_int);
 129 CREATE_MPI_DATATYPE(MPI_2INT, int_int);
 130 CREATE_MPI_DATATYPE(MPI_2FLOAT, float_float);
 131 CREATE_MPI_DATATYPE(MPI_2DOUBLE, double_double);
 132 CREATE_MPI_DATATYPE(MPI_2LONG, long_long);
 133
 134 CREATE_MPI_DATATYPE(MPI_REAL, float);
 135 CREATE_MPI_DATATYPE(MPI_REAL4, float);
 136 CREATE_MPI_DATATYPE(MPI_REAL8, float);
 137 CREATE_MPI_DATATYPE(MPI_REAL16, double);
 138 CREATE_MPI_DATATYPE_NULL(MPI_COMPLEX8);
 139 CREATE_MPI_DATATYPE_NULL(MPI_COMPLEX16);
 140 CREATE_MPI_DATATYPE_NULL(MPI_COMPLEX32);
 141 CREATE_MPI_DATATYPE(MPI_INTEGER1, int);
 142 CREATE_MPI_DATATYPE(MPI_INTEGER2, int16_t);
 143 CREATE_MPI_DATATYPE(MPI_INTEGER4, int32_t);
 144 CREATE_MPI_DATATYPE(MPI_INTEGER8, int64_t);
 145 CREATE_MPI_DATATYPE(MPI_INTEGER16, integer128_t);
 146
 147 CREATE_MPI_DATATYPE(MPI_LONG_DOUBLE_INT, long_double_int);
 148
 149 CREATE_MPI_DATATYPE_NULL(MPI_UB);
 150 CREATE_MPI_DATATYPE_NULL(MPI_LB);
 151 CREATE_MPI_DATATYPE(MPI_PACKED, char);
 152 // Internal use only
 153 CREATE_MPI_DATATYPE(MPI_PTR, void*);
 154
 155 /** Check if the datatype is usable for communications */
 156 bool is_datatype_valid(MPI_Datatype datatype) {
 157     return datatype != MPI_DATATYPE_NULL && ((datatype->flags & DT_FLAG_COMMITED) != 0);
 158 }
 159
 160 size_t smpi_datatype_size(MPI_Datatype datatype)
 161 {
 162   return datatype->size;
 163 }
 164
 165 MPI_Aint smpi_datatype_lb(MPI_Datatype datatype)
 166 {
 167   return datatype->lb;
 168 }
 169
 170 MPI_Aint smpi_datatype_ub(MPI_Datatype datatype)
 171 {
 172   return datatype->ub;
 173 }
 174
 175 int smpi_datatype_dup(MPI_Datatype datatype, MPI_Datatype* new_t)
 176 {
 177   int ret=MPI_SUCCESS;
 178   *new_t= xbt_new(s_smpi_mpi_datatype_t,1);
 179   memcpy(*new_t, datatype, sizeof(s_smpi_mpi_datatype_t));
 180   (*new_t)->in_use=1;
 181   (*new_t)->flags &= ~DT_FLAG_PREDEFINED;
 182   if (datatype->sizeof_substruct){
 183     (*new_t)->substruct=xbt_malloc(datatype->sizeof_substruct);
 184     memcpy((*new_t)->substruct, datatype->substruct, datatype->sizeof_substruct);
 185   }
 186   if(datatype->name)
 187     (*new_t)->name = xbt_strdup(datatype->name);
 188   if(datatype->attributes !=nullptr){
 189     (*new_t)->attributes     = xbt_dict_new_homogeneous(nullptr);
 190     xbt_dict_cursor_t cursor = nullptr;
 191     int* key;
 192     int flag;
 193     void* value_in;
 194     void* value_out;
 195     xbt_dict_foreach (datatype->attributes, cursor, key, value_in) {
 196       smpi_type_key_elem elem = static_cast<smpi_type_key_elem>(
 197           xbt_dict_get_or_null_ext(smpi_type_keyvals, reinterpret_cast<const char*>(key), sizeof(int)));
 198       if (elem != nullptr && elem->copy_fn != MPI_NULL_COPY_FN) {
 199         ret = elem->copy_fn(datatype, *key, nullptr, value_in, &value_out, &flag);
 200         if (ret != MPI_SUCCESS) {
 201           smpi_datatype_unuse(*new_t);
 202           *new_t = MPI_DATATYPE_NULL;
 203           xbt_dict_cursor_free(&cursor);
 204           return ret;
 205         }
 206         if (flag)
 207           xbt_dict_set_ext((*new_t)->attributes, reinterpret_cast<const char*>(key), sizeof(int), value_out, nullptr);
 208       }
 209       }
 210     }
 211   return ret;
 212 }
 213
 214 int smpi_datatype_extent(MPI_Datatype datatype, MPI_Aint * lb, MPI_Aint * extent)
 215 {
 216   if(datatype == MPI_DATATYPE_NULL){
 217     *lb=0;
 218     *extent=0;
 219     return MPI_SUCCESS;
 220   }
 221   *lb = datatype->lb;
 222   *extent = datatype->ub - datatype->lb;
 223   return MPI_SUCCESS;
 224 }
 225
 226 MPI_Aint smpi_datatype_get_extent(MPI_Datatype datatype){
 227   if(datatype == MPI_DATATYPE_NULL){
 228     return 0;
 229   }
 230   return datatype->ub - datatype->lb;
 231 }
 232
 233 void smpi_datatype_get_name(MPI_Datatype datatype, char* name, int* length){
 234   *length = strlen(datatype->name);
 235   strncpy(name, datatype->name, *length+1);
 236 }
 237
 238 void smpi_datatype_set_name(MPI_Datatype datatype, char* name){
 239   if(datatype->name!=nullptr &&  (datatype->flags & DT_FLAG_PREDEFINED) == 0)
 240     xbt_free(datatype->name);
 241   datatype->name = xbt_strdup(name);
 242 }
 243
 244 int smpi_datatype_copy(void *sendbuf, int sendcount, MPI_Datatype sendtype,
 245                        void *recvbuf, int recvcount, MPI_Datatype recvtype)
 246 {
 247   int count;
 248   if(smpi_privatize_global_variables){
 249     smpi_switch_data_segment(smpi_process_index());
 250   }
 251   /* First check if we really have something to do */
 252   if (recvcount > 0 && recvbuf != sendbuf) {
 253     /* FIXME: treat packed cases */
 254     sendcount *= smpi_datatype_size(sendtype);
 255     recvcount *= smpi_datatype_size(recvtype);
 256     count = sendcount < recvcount ? sendcount : recvcount;
 257
 258     if(sendtype->sizeof_substruct == 0 && recvtype->sizeof_substruct == 0) {
 259       if(!smpi_process_get_replaying())
 260         memcpy(recvbuf, sendbuf, count);
 261     }
 262     else if (sendtype->sizeof_substruct == 0)
 263     {
 264       s_smpi_subtype_t *subtype =  static_cast<s_smpi_subtype_t*>(recvtype->substruct);
 265       subtype->unserialize( sendbuf, recvbuf, recvcount/smpi_datatype_size(recvtype), subtype, MPI_REPLACE);
 266     }
 267     else if (recvtype->sizeof_substruct == 0)
 268     {
 269       s_smpi_subtype_t *subtype =  static_cast<s_smpi_subtype_t*>(sendtype->substruct);
 270       subtype->serialize(sendbuf, recvbuf, sendcount/smpi_datatype_size(sendtype), subtype);
 271     }else{
 272       s_smpi_subtype_t *subtype =  static_cast<s_smpi_subtype_t*>(sendtype->substruct);
 273
 274       void * buf_tmp = xbt_malloc(count);
 275
 276       subtype->serialize( sendbuf, buf_tmp,count/smpi_datatype_size(sendtype), subtype);
 277       subtype =  static_cast<s_smpi_subtype_t*>(recvtype->substruct);
 278       subtype->unserialize( buf_tmp, recvbuf,count/smpi_datatype_size(recvtype), subtype, MPI_REPLACE);
 279
 280       xbt_free(buf_tmp);
 281     }
 282   }
 283
 284   return sendcount > recvcount ? MPI_ERR_TRUNCATE : MPI_SUCCESS;
 285 }
 286
 287 /*
 288  *  Copies noncontiguous data into contiguous memory.
 289  *  @param contiguous_vector - output vector
 290  *  @param noncontiguous_vector - input vector
 291  *  @param type - pointer containing :
 292  *      - stride - stride of between noncontiguous data
 293  *      - block_length - the width or height of blocked matrix
 294  *      - count - the number of rows of matrix
 295  */
 296 void serialize_vector( void* noncontiguous_vector, void *contiguous_vector, int count, void *type)
 297 {
 298   s_smpi_mpi_vector_t* type_c = reinterpret_cast<s_smpi_mpi_vector_t*>(type);
 299   int i;
 300   char* contiguous_vector_char = static_cast<char*>(contiguous_vector);
 301   char* noncontiguous_vector_char = static_cast<char*>(noncontiguous_vector);
 302
 303   for (i = 0; i < type_c->block_count * count; i++) {
 304       if (type_c->old_type->sizeof_substruct == 0)
 305         memcpy(contiguous_vector_char, noncontiguous_vector_char, type_c->block_length * type_c->size_oldtype);
 306       else
 307         static_cast<s_smpi_subtype_t*>(type_c->old_type->substruct)->serialize( noncontiguous_vector_char,
 308                                                                      contiguous_vector_char,
 309                                                                      type_c->block_length, type_c->old_type->substruct);
 310
 311     contiguous_vector_char += type_c->block_length*type_c->size_oldtype;
 312     if((i+1)%type_c->block_count ==0)
 313       noncontiguous_vector_char += type_c->block_length*smpi_datatype_get_extent(type_c->old_type);
 314     else
 315       noncontiguous_vector_char += type_c->block_stride*smpi_datatype_get_extent(type_c->old_type);
 316   }
 317 }
 318
 319 /*
 320  *  Copies contiguous data into noncontiguous memory.
 321  *  @param noncontiguous_vector - output vector
 322  *  @param contiguous_vector - input vector
 323  *  @param type - pointer contening :
 324  *      - stride - stride of between noncontiguous data
 325  *      - block_length - the width or height of blocked matrix
 326  *      - count - the number of rows of matrix
 327  */
 328 void unserialize_vector( void* contiguous_vector, void *noncontiguous_vector, int count, void *type, MPI_Op op)
 329 {
 330   s_smpi_mpi_vector_t* type_c = reinterpret_cast<s_smpi_mpi_vector_t*>(type);
 331   int i;
 332
 333   char* contiguous_vector_char = static_cast<char*>(contiguous_vector);
 334   char* noncontiguous_vector_char = static_cast<char*>(noncontiguous_vector);
 335
 336   for (i = 0; i < type_c->block_count * count; i++) {
 337     if (type_c->old_type->sizeof_substruct == 0)
 338       smpi_op_apply(op, contiguous_vector_char, noncontiguous_vector_char, &type_c->block_length,
 339           &type_c->old_type);
 340     else
 341       static_cast<s_smpi_subtype_t*>(type_c->old_type->substruct)->unserialize(contiguous_vector_char, noncontiguous_vector_char,
 342                                                                     type_c->block_length,type_c->old_type->substruct,
 343                                                                     op);
 344     contiguous_vector_char += type_c->block_length*type_c->size_oldtype;
 345     if((i+1)%type_c->block_count ==0)
 346       noncontiguous_vector_char += type_c->block_length*smpi_datatype_get_extent(type_c->old_type);
 347     else
 348       noncontiguous_vector_char += type_c->block_stride*smpi_datatype_get_extent(type_c->old_type);
 349   }
 350 }
 351
 352 /* Create a Sub type vector to be able to serialize and unserialize it the structure s_smpi_mpi_vector_t is derived
 353  * from s_smpi_subtype which required the functions unserialize and serialize */
 354 s_smpi_mpi_vector_t* smpi_datatype_vector_create( int block_stride, int block_length, int block_count,
 355                                                   MPI_Datatype old_type, int size_oldtype){
 356   s_smpi_mpi_vector_t *new_t= xbt_new(s_smpi_mpi_vector_t,1);
 357   new_t->base.serialize = &serialize_vector;
 358   new_t->base.unserialize = &unserialize_vector;
 359   new_t->base.subtype_free = &free_vector;
 360   new_t->base.subtype_use = &use_vector;
 361   new_t->block_stride = block_stride;
 362   new_t->block_length = block_length;
 363   new_t->block_count = block_count;
 364   smpi_datatype_use(old_type);
 365   new_t->old_type = old_type;
 366   new_t->size_oldtype = size_oldtype;
 367   return new_t;
 368 }
 369
 370 void smpi_datatype_create(MPI_Datatype* new_type, int size,int lb, int ub, int sizeof_substruct, void *struct_type,
 371                           int flags){
 372   MPI_Datatype new_t= xbt_new(s_smpi_mpi_datatype_t,1);
 373   new_t->name = nullptr;
 374   new_t->size = size;
 375   new_t->sizeof_substruct = size>0? sizeof_substruct:0;
 376   new_t->lb = lb;
 377   new_t->ub = ub;
 378   new_t->flags = flags;
 379   new_t->substruct = struct_type;
 380   new_t->in_use=1;
 381   new_t->attributes=nullptr;
 382   *new_type = new_t;
 383
 384 #if HAVE_MC
 385   if(MC_is_active())
 386     MC_ignore(&(new_t->in_use), sizeof(new_t->in_use));
 387 #endif
 388 }
 389
 390 void smpi_datatype_free(MPI_Datatype* type){
 391   xbt_assert((*type)->in_use >= 0);
 392
 393   if((*type)->flags & DT_FLAG_PREDEFINED)
 394     return;
 395
 396   //if still used, mark for deletion
 397   if((*type)->in_use!=0){
 398       (*type)->flags |=DT_FLAG_DESTROYED;
 399       return;
 400   }
 401
 402   if((*type)->attributes !=nullptr){
 403     xbt_dict_cursor_t cursor = nullptr;
 404     int* key;
 405     void * value;
 406     int flag;
 407     xbt_dict_foreach((*type)->attributes, cursor, key, value){
 408       smpi_type_key_elem elem =
 409           static_cast<smpi_type_key_elem>(xbt_dict_get_or_null_ext(smpi_type_keyvals, reinterpret_cast<const char*>(key), sizeof(int)));
 410       if(elem!=nullptr && elem->delete_fn!=nullptr)
 411         elem->delete_fn(*type,*key, value, &flag);
 412     }
 413     xbt_dict_free(&(*type)->attributes);
 414   }
 415
 416   if ((*type)->sizeof_substruct != 0){
 417     //((s_smpi_subtype_t *)(*type)->substruct)->subtype_free(type);
 418     xbt_free((*type)->substruct);
 419   }
 420   xbt_free((*type)->name);
 421   xbt_free(*type);
 422   *type = MPI_DATATYPE_NULL;
 423 }
 424
 425 void smpi_datatype_use(MPI_Datatype type){
 426
 427   if(type == MPI_DATATYPE_NULL)
 428     return;
 429   type->in_use++;
 430
 431   if(type->sizeof_substruct!=0){
 432     static_cast<s_smpi_subtype_t *>((type)->substruct)->subtype_use(&type);
 433   }
 434 #if HAVE_MC
 435   if(MC_is_active())
 436     MC_ignore(&(type->in_use), sizeof(type->in_use));
 437 #endif
 438 }
 439
 440 void smpi_datatype_unuse(MPI_Datatype type)
 441 {
 442   if (type == MPI_DATATYPE_NULL)
 443     return;
 444
 445   if (type->in_use > 0)
 446     type->in_use--;
 447
 448   if(type->sizeof_substruct!=0){
 449     static_cast<s_smpi_subtype_t *>((type)->substruct)->subtype_free(&type);
 450   }
 451
 452   if (type->in_use == 0)
 453     smpi_datatype_free(&type);
 454
 455 #if HAVE_MC
 456   if(MC_is_active())
 457     MC_ignore(&(type->in_use), sizeof(type->in_use));
 458 #endif
 459 }
 460
 461 /*Contiguous Implementation*/
 462
 463 /* Copies noncontiguous data into contiguous memory.
 464  *  @param contiguous_hvector - output hvector
 465  *  @param noncontiguous_hvector - input hvector
 466  *  @param type - pointer contening :
 467  *      - stride - stride of between noncontiguous data, in bytes
 468  *      - block_length - the width or height of blocked matrix
 469  *      - count - the number of rows of matrix
 470  */
 471 void serialize_contiguous( void* noncontiguous_hvector, void *contiguous_hvector, int count, void *type)
 472 {
 473   s_smpi_mpi_contiguous_t* type_c = reinterpret_cast<s_smpi_mpi_contiguous_t*>(type);
 474   char* contiguous_vector_char = static_cast<char*>(contiguous_hvector);
 475   char* noncontiguous_vector_char = static_cast<char*>(noncontiguous_hvector)+type_c->lb;
 476   memcpy(contiguous_vector_char, noncontiguous_vector_char, count* type_c->block_count * type_c->size_oldtype);
 477 }
 478 /* Copies contiguous data into noncontiguous memory.
 479  *  @param noncontiguous_vector - output hvector
 480  *  @param contiguous_vector - input hvector
 481  *  @param type - pointer contening :
 482  *      - stride - stride of between noncontiguous data, in bytes
 483  *      - block_length - the width or height of blocked matrix
 484  *      - count - the number of rows of matrix
 485  */
 486 void unserialize_contiguous(void* contiguous_vector, void *noncontiguous_vector, int count, void *type, MPI_Op op)
 487 {
 488   s_smpi_mpi_contiguous_t* type_c = reinterpret_cast<s_smpi_mpi_contiguous_t*>(type);
 489   char* contiguous_vector_char = static_cast<char*>(contiguous_vector);
 490   char* noncontiguous_vector_char = static_cast<char*>(noncontiguous_vector)+type_c->lb;
 491   int n= count* type_c->block_count;
 492   smpi_op_apply(op, contiguous_vector_char, noncontiguous_vector_char, &n, &type_c->old_type);
 493 }
 494
 495 void free_contiguous(MPI_Datatype* d){
 496   smpi_datatype_unuse(reinterpret_cast<s_smpi_mpi_contiguous_t*>((*d)->substruct)->old_type);
 497 }
 498
 499 void use_contiguous(MPI_Datatype* d){
 500   smpi_datatype_use(reinterpret_cast<s_smpi_mpi_contiguous_t*>((*d)->substruct)->old_type);
 501 }
 502
 503 /* Create a Sub type contiguous to be able to serialize and unserialize it the structure s_smpi_mpi_contiguous_t is
 504  * derived from s_smpi_subtype which required the functions unserialize and serialize */
 505 s_smpi_mpi_contiguous_t* smpi_datatype_contiguous_create( MPI_Aint lb, int block_count, MPI_Datatype old_type,
 506                                                   int size_oldtype){
 507   if(block_count==0)
 508     return nullptr;
 509   s_smpi_mpi_contiguous_t *new_t= xbt_new(s_smpi_mpi_contiguous_t,1);
 510   new_t->base.serialize = &serialize_contiguous;
 511   new_t->base.unserialize = &unserialize_contiguous;
 512   new_t->base.subtype_free = &free_contiguous;
 513   new_t->base.subtype_use = &use_contiguous;
 514   new_t->lb = lb;
 515   new_t->block_count = block_count;
 516   new_t->old_type = old_type;
 517   smpi_datatype_use(old_type);
 518   new_t->size_oldtype = size_oldtype;
 519   smpi_datatype_use(old_type);
 520   return new_t;
 521 }
 522
 523 int smpi_datatype_contiguous(int count, MPI_Datatype old_type, MPI_Datatype* new_type, MPI_Aint lb)
 524 {
 525   int retval;
 526   if(old_type->sizeof_substruct){
 527     //handle this case as a hvector with stride equals to the extent of the datatype
 528     return smpi_datatype_hvector(count, 1, smpi_datatype_get_extent(old_type), old_type, new_type);
 529   }
 530
 531   s_smpi_mpi_contiguous_t* subtype = smpi_datatype_contiguous_create( lb, count, old_type,smpi_datatype_size(old_type));
 532
 533   smpi_datatype_create(new_type, count * smpi_datatype_size(old_type),lb,lb + count * smpi_datatype_size(old_type),
 534             sizeof(s_smpi_mpi_contiguous_t),subtype, DT_FLAG_CONTIGUOUS);
 535   retval=MPI_SUCCESS;
 536   return retval;
 537 }
 538
 539 int smpi_datatype_vector(int count, int blocklen, int stride, MPI_Datatype old_type, MPI_Datatype* new_type)
 540 {
 541   int retval;
 542   if (blocklen<0)
 543     return MPI_ERR_ARG;
 544   MPI_Aint lb = 0;
 545   MPI_Aint ub = 0;
 546   if(count>0){
 547     lb=smpi_datatype_lb(old_type);
 548     ub=((count-1)*stride+blocklen-1)*smpi_datatype_get_extent(old_type)+smpi_datatype_ub(old_type);
 549   }
 550   if(old_type->sizeof_substruct != 0 || stride != blocklen){
 551
 552     s_smpi_mpi_vector_t* subtype = smpi_datatype_vector_create(stride, blocklen, count, old_type,
 553                                                                 smpi_datatype_size(old_type));
 554     smpi_datatype_create(new_type, count * (blocklen) * smpi_datatype_size(old_type), lb, ub, sizeof(s_smpi_mpi_vector_t), subtype,
 555                          DT_FLAG_VECTOR);
 556     retval=MPI_SUCCESS;
 557   }else{
 558     /* in this situation the data are contiguous thus it's not required to serialize and unserialize it*/
 559     smpi_datatype_create(new_type, count * blocklen * smpi_datatype_size(old_type), 0, ((count -1) * stride + blocklen)*
 560                          smpi_datatype_size(old_type), 0, nullptr, DT_FLAG_VECTOR|DT_FLAG_CONTIGUOUS);
 561     retval=MPI_SUCCESS;
 562   }
 563   return retval;
 564 }
 565
 566 void free_vector(MPI_Datatype* d){
 567   smpi_datatype_unuse(reinterpret_cast<s_smpi_mpi_indexed_t*>((*d)->substruct)->old_type);
 568 }
 569
 570 void use_vector(MPI_Datatype* d){
 571   smpi_datatype_use(reinterpret_cast<s_smpi_mpi_indexed_t*>((*d)->substruct)->old_type);
 572 }
 573
 574 /* Hvector Implementation - Vector with stride in bytes */
 575
 576 /* Copies noncontiguous data into contiguous memory.
 577  *  @param contiguous_hvector - output hvector
 578  *  @param noncontiguous_hvector - input hvector
 579  *  @param type - pointer contening :
 580  *      - stride - stride of between noncontiguous data, in bytes
 581  *      - block_length - the width or height of blocked matrix
 582  *      - count - the number of rows of matrix
 583  */
 584 void serialize_hvector( void* noncontiguous_hvector, void *contiguous_hvector, int count, void *type)
 585 {
 586   s_smpi_mpi_hvector_t* type_c = reinterpret_cast<s_smpi_mpi_hvector_t*>(type);
 587   int i;
 588   char* contiguous_vector_char = static_cast<char*>(contiguous_hvector);
 589   char* noncontiguous_vector_char = static_cast<char*>(noncontiguous_hvector);
 590
 591   for (i = 0; i < type_c->block_count * count; i++) {
 592     if (type_c->old_type->sizeof_substruct == 0)
 593       memcpy(contiguous_vector_char, noncontiguous_vector_char, type_c->block_length * type_c->size_oldtype);
 594     else
 595       static_cast<s_smpi_subtype_t*>(type_c->old_type->substruct)->serialize( noncontiguous_vector_char,
 596                                                                    contiguous_vector_char,
 597                                                                    type_c->block_length, type_c->old_type->substruct);
 598
 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;
 602     else
 603       noncontiguous_vector_char += type_c->block_stride;
 604   }
 605 }
 606 /* Copies contiguous data into noncontiguous memory.
 607  *  @param noncontiguous_vector - output hvector
 608  *  @param contiguous_vector - input hvector
 609  *  @param type - pointer contening :
 610  *      - stride - stride of between noncontiguous data, in bytes
 611  *      - block_length - the width or height of blocked matrix
 612  *      - count - the number of rows of matrix
 613  */
 614 void unserialize_hvector( void* contiguous_vector, void *noncontiguous_vector, int count, void *type, MPI_Op op)
 615 {
 616   s_smpi_mpi_hvector_t* type_c = reinterpret_cast<s_smpi_mpi_hvector_t*>(type);
 617   int i;
 618
 619   char* contiguous_vector_char = static_cast<char*>(contiguous_vector);
 620   char* noncontiguous_vector_char = static_cast<char*>(noncontiguous_vector);
 621
 622   for (i = 0; i < type_c->block_count * count; i++) {
 623     if (type_c->old_type->sizeof_substruct == 0)
 624       smpi_op_apply(op, contiguous_vector_char, noncontiguous_vector_char, &type_c->block_length, &type_c->old_type);
 625     else
 626       static_cast<s_smpi_subtype_t*>(type_c->old_type->substruct)->unserialize( contiguous_vector_char, noncontiguous_vector_char,
 627                                                                      type_c->block_length, type_c->old_type->substruct,
 628                                                                      op);
 629     contiguous_vector_char += type_c->block_length*type_c->size_oldtype;
 630     if((i+1)%type_c->block_count ==0)
 631       noncontiguous_vector_char += type_c->block_length*type_c->size_oldtype;
 632     else
 633       noncontiguous_vector_char += type_c->block_stride;
 634   }
 635 }
 636
 637 /* Create a Sub type vector to be able to serialize and unserialize it the structure s_smpi_mpi_vector_t is derived
 638  * from s_smpi_subtype which required the functions unserialize and serialize
 639  *
 640  */
 641 s_smpi_mpi_hvector_t* smpi_datatype_hvector_create( MPI_Aint block_stride, int block_length, int block_count,
 642                                                   MPI_Datatype old_type, int size_oldtype){
 643   s_smpi_mpi_hvector_t *new_t= xbt_new(s_smpi_mpi_hvector_t,1);
 644   new_t->base.serialize = &serialize_hvector;
 645   new_t->base.unserialize = &unserialize_hvector;
 646   new_t->base.subtype_free = &free_hvector;
 647   new_t->base.subtype_use = &use_hvector;
 648   new_t->block_stride = block_stride;
 649   new_t->block_length = block_length;
 650   new_t->block_count = block_count;
 651   new_t->old_type = old_type;
 652   new_t->size_oldtype = size_oldtype;
 653   smpi_datatype_use(old_type);
 654   return new_t;
 655 }
 656
 657 //do nothing for vector types
 658 void free_hvector(MPI_Datatype* d){
 659   smpi_datatype_unuse(reinterpret_cast<s_smpi_mpi_hvector_t*>((*d)->substruct)->old_type);
 660 }
 661
 662 void use_hvector(MPI_Datatype* d){
 663   smpi_datatype_use(reinterpret_cast<s_smpi_mpi_hvector_t*>((*d)->substruct)->old_type);
 664 }
 665
 666 int smpi_datatype_hvector(int count, int blocklen, MPI_Aint stride, MPI_Datatype old_type, MPI_Datatype* new_type)
 667 {
 668   int retval;
 669   if (blocklen<0)
 670     return MPI_ERR_ARG;
 671   MPI_Aint lb = 0;
 672   MPI_Aint ub = 0;
 673   if(count>0){
 674     lb=smpi_datatype_lb(old_type);
 675     ub=((count-1)*stride)+(blocklen-1)*smpi_datatype_get_extent(old_type)+smpi_datatype_ub(old_type);
 676   }
 677   if(old_type->sizeof_substruct != 0 || stride != blocklen*smpi_datatype_get_extent(old_type)){
 678     s_smpi_mpi_hvector_t* subtype = smpi_datatype_hvector_create( stride, blocklen, count, old_type,
 679                                                                   smpi_datatype_size(old_type));
 680
 681     smpi_datatype_create(new_type, count * blocklen * smpi_datatype_size(old_type), lb,ub, sizeof(s_smpi_mpi_hvector_t), subtype, DT_FLAG_VECTOR);
 682     retval=MPI_SUCCESS;
 683   }else{
 684     smpi_datatype_create(new_type, count * blocklen * smpi_datatype_size(old_type),0,count * blocklen *
 685                                              smpi_datatype_size(old_type), 0, nullptr, DT_FLAG_VECTOR|DT_FLAG_CONTIGUOUS);
 686     retval=MPI_SUCCESS;
 687   }
 688   return retval;
 689 }
 690
 691 /* Indexed Implementation */
 692
 693 /* Copies noncontiguous data into contiguous memory.
 694  *  @param contiguous_indexed - output indexed
 695  *  @param noncontiguous_indexed - input indexed
 696  *  @param type - pointer contening :
 697  *      - block_lengths - the width or height of blocked matrix
 698  *      - block_indices - indices of each data, in element
 699  *      - count - the number of rows of matrix
 700  */
 701 void serialize_indexed( void* noncontiguous_indexed, void *contiguous_indexed, int count, void *type)
 702 {
 703   s_smpi_mpi_indexed_t* type_c = reinterpret_cast<s_smpi_mpi_indexed_t*>(type);
 704   int i,j;
 705   char* contiguous_indexed_char = static_cast<char*>(contiguous_indexed);
 706   char* noncontiguous_indexed_char = static_cast<char*>(noncontiguous_indexed)+type_c->block_indices[0] * type_c->size_oldtype;
 707   for(j=0; j<count;j++){
 708     for (i = 0; i < type_c->block_count; i++) {
 709       if (type_c->old_type->sizeof_substruct == 0)
 710         memcpy(contiguous_indexed_char, noncontiguous_indexed_char, type_c->block_lengths[i] * type_c->size_oldtype);
 711       else
 712         static_cast<s_smpi_subtype_t*>(type_c->old_type->substruct)->serialize( noncontiguous_indexed_char,
 713                                                                      contiguous_indexed_char,
 714                                                                      type_c->block_lengths[i],
 715                                                                      type_c->old_type->substruct);
 716
 717       contiguous_indexed_char += type_c->block_lengths[i]*type_c->size_oldtype;
 718       if (i<type_c->block_count-1)
 719         noncontiguous_indexed_char =
 720           static_cast<char*>(noncontiguous_indexed) + type_c->block_indices[i+1]*smpi_datatype_get_extent(type_c->old_type);
 721       else
 722         noncontiguous_indexed_char += type_c->block_lengths[i]*smpi_datatype_get_extent(type_c->old_type);
 723     }
 724     noncontiguous_indexed=static_cast< void*>(noncontiguous_indexed_char);
 725   }
 726 }
 727 /* Copies contiguous data into noncontiguous memory.
 728  *  @param noncontiguous_indexed - output indexed
 729  *  @param contiguous_indexed - input indexed
 730  *  @param type - pointer contening :
 731  *      - block_lengths - the width or height of blocked matrix
 732  *      - block_indices - indices of each data, in element
 733  *      - count - the number of rows of matrix
 734  */
 735 void unserialize_indexed( void* contiguous_indexed, void *noncontiguous_indexed, int count, void *type, MPI_Op op)
 736 {
 737   s_smpi_mpi_indexed_t* type_c = reinterpret_cast<s_smpi_mpi_indexed_t*>(type);
 738   int i,j;
 739   char* contiguous_indexed_char = static_cast<char*>(contiguous_indexed);
 740   char* noncontiguous_indexed_char =
 741     static_cast<char*>(noncontiguous_indexed)+type_c->block_indices[0]*smpi_datatype_get_extent(type_c->old_type);
 742   for(j=0; j<count;j++){
 743     for (i = 0; i < type_c->block_count; i++) {
 744       if (type_c->old_type->sizeof_substruct == 0)
 745         smpi_op_apply(op, contiguous_indexed_char, noncontiguous_indexed_char, &type_c->block_lengths[i],
 746                     &type_c->old_type);
 747       else
 748         static_cast<s_smpi_subtype_t*>(type_c->old_type->substruct)->unserialize( contiguous_indexed_char,
 749                                                                        noncontiguous_indexed_char,
 750                                                                        type_c->block_lengths[i],
 751                                                                        type_c->old_type->substruct, op);
 752
 753       contiguous_indexed_char += type_c->block_lengths[i]*type_c->size_oldtype;
 754       if (i<type_c->block_count-1)
 755         noncontiguous_indexed_char =
 756           static_cast<char*>(noncontiguous_indexed) + type_c->block_indices[i+1]*smpi_datatype_get_extent(type_c->old_type);
 757       else
 758         noncontiguous_indexed_char += type_c->block_lengths[i]*smpi_datatype_get_extent(type_c->old_type);
 759     }
 760     noncontiguous_indexed=static_cast<void*>(noncontiguous_indexed_char);
 761   }
 762 }
 763
 764 void free_indexed(MPI_Datatype* type){
 765   if((*type)->in_use==0){
 766     xbt_free(reinterpret_cast<s_smpi_mpi_indexed_t*>((*type)->substruct)->block_lengths);
 767     xbt_free(reinterpret_cast<s_smpi_mpi_indexed_t*>((*type)->substruct)->block_indices);
 768   }
 769   smpi_datatype_unuse(reinterpret_cast<s_smpi_mpi_indexed_t*>((*type)->substruct)->old_type);
 770 }
 771
 772 void use_indexed(MPI_Datatype* type){
 773   smpi_datatype_use(reinterpret_cast<s_smpi_mpi_indexed_t*>((*type)->substruct)->old_type);
 774 }
 775
 776
 777 /* Create a Sub type indexed to be able to serialize and unserialize it the structure s_smpi_mpi_indexed_t is derived
 778  * from s_smpi_subtype which required the functions unserialize and serialize */
 779 s_smpi_mpi_indexed_t* smpi_datatype_indexed_create( int* block_lengths, int* block_indices, int block_count,
 780                                                   MPI_Datatype old_type, int size_oldtype){
 781   s_smpi_mpi_indexed_t *new_t= xbt_new(s_smpi_mpi_indexed_t,1);
 782   new_t->base.serialize = &serialize_indexed;
 783   new_t->base.unserialize = &unserialize_indexed;
 784   new_t->base.subtype_free = &free_indexed;
 785   new_t->base.subtype_use = &use_indexed;
 786   new_t->block_lengths= xbt_new(int, block_count);
 787   new_t->block_indices= xbt_new(int, block_count);
 788   int i;
 789   for(i=0;i<block_count;i++){
 790     new_t->block_lengths[i]=block_lengths[i];
 791     new_t->block_indices[i]=block_indices[i];
 792   }
 793   new_t->block_count = block_count;
 794   smpi_datatype_use(old_type);
 795   new_t->old_type = old_type;
 796   new_t->size_oldtype = size_oldtype;
 797   return new_t;
 798 }
 799
 800 int smpi_datatype_indexed(int count, int* blocklens, int* indices, MPI_Datatype old_type, MPI_Datatype* new_type)
 801 {
 802   int i;
 803   int retval;
 804   int size = 0;
 805   bool contiguous=true;
 806   MPI_Aint lb = 0;
 807   MPI_Aint ub = 0;
 808   if(count>0){
 809     lb=indices[0]*smpi_datatype_get_extent(old_type);
 810     ub=indices[0]*smpi_datatype_get_extent(old_type) + blocklens[0]*smpi_datatype_ub(old_type);
 811   }
 812
 813   for(i=0; i< count; i++){
 814     if   (blocklens[i]<0)
 815       return MPI_ERR_ARG;
 816     size += blocklens[i];
 817
 818     if(indices[i]*smpi_datatype_get_extent(old_type)+smpi_datatype_lb(old_type)<lb)
 819       lb = indices[i]*smpi_datatype_get_extent(old_type)+smpi_datatype_lb(old_type);
 820     if(indices[i]*smpi_datatype_get_extent(old_type)+blocklens[i]*smpi_datatype_ub(old_type)>ub)
 821       ub = indices[i]*smpi_datatype_get_extent(old_type)+blocklens[i]*smpi_datatype_ub(old_type);
 822
 823     if ( (i< count -1) && (indices[i]+blocklens[i] != indices[i+1]) )
 824       contiguous=false;
 825   }
 826   if (old_type->sizeof_substruct != 0)
 827     contiguous=false;
 828
 829   if(!contiguous){
 830     s_smpi_mpi_indexed_t* subtype = smpi_datatype_indexed_create( blocklens, indices, count, old_type,
 831                                                                   smpi_datatype_size(old_type));
 832      smpi_datatype_create(new_type,  size * smpi_datatype_size(old_type),lb,ub,sizeof(s_smpi_mpi_indexed_t), subtype, DT_FLAG_DATA);
 833   }else{
 834     s_smpi_mpi_contiguous_t* subtype = smpi_datatype_contiguous_create( lb, size, old_type,
 835                                                                   smpi_datatype_size(old_type));
 836     smpi_datatype_create(new_type, size * smpi_datatype_size(old_type), lb, ub, sizeof(s_smpi_mpi_contiguous_t), subtype,
 837                          DT_FLAG_DATA|DT_FLAG_CONTIGUOUS);
 838   }
 839   retval=MPI_SUCCESS;
 840   return retval;
 841 }
 842 /* Hindexed Implementation - Indexed with indices in bytes */
 843
 844 /* Copies noncontiguous data into contiguous memory.
 845  *  @param contiguous_hindexed - output hindexed
 846  *  @param noncontiguous_hindexed - input hindexed
 847  *  @param type - pointer contening :
 848  *      - block_lengths - the width or height of blocked matrix
 849  *      - block_indices - indices of each data, in bytes
 850  *      - count - the number of rows of matrix
 851  */
 852 void serialize_hindexed( void* noncontiguous_hindexed, void *contiguous_hindexed, int count, void *type)
 853 {
 854   s_smpi_mpi_hindexed_t* type_c = reinterpret_cast<s_smpi_mpi_hindexed_t*>(type);
 855   int i,j;
 856   char* contiguous_hindexed_char = static_cast<char*>(contiguous_hindexed);
 857   char* noncontiguous_hindexed_char = static_cast<char*>(noncontiguous_hindexed)+ type_c->block_indices[0];
 858   for(j=0; j<count;j++){
 859     for (i = 0; i < type_c->block_count; i++) {
 860       if (type_c->old_type->sizeof_substruct == 0)
 861         memcpy(contiguous_hindexed_char, noncontiguous_hindexed_char, type_c->block_lengths[i] * type_c->size_oldtype);
 862       else
 863         static_cast<s_smpi_subtype_t*>(type_c->old_type->substruct)->serialize( noncontiguous_hindexed_char,
 864                                                                      contiguous_hindexed_char,
 865                                                                      type_c->block_lengths[i],
 866                                                                      type_c->old_type->substruct);
 867
 868       contiguous_hindexed_char += type_c->block_lengths[i]*type_c->size_oldtype;
 869       if (i<type_c->block_count-1)
 870         noncontiguous_hindexed_char = static_cast<char*>(noncontiguous_hindexed) + type_c->block_indices[i+1];
 871       else
 872         noncontiguous_hindexed_char += type_c->block_lengths[i]*smpi_datatype_get_extent(type_c->old_type);
 873     }
 874     noncontiguous_hindexed=static_cast<void*>(noncontiguous_hindexed_char);
 875   }
 876 }
 877 /* Copies contiguous data into noncontiguous memory.
 878  *  @param noncontiguous_hindexed - output hindexed
 879  *  @param contiguous_hindexed - input hindexed
 880  *  @param type - pointer contening :
 881  *      - block_lengths - the width or height of blocked matrix
 882  *      - block_indices - indices of each data, in bytes
 883  *      - count - the number of rows of matrix
 884  */
 885 void unserialize_hindexed( void* contiguous_hindexed, void *noncontiguous_hindexed, int count, void *type,
 886                          MPI_Op op)
 887 {
 888   s_smpi_mpi_hindexed_t* type_c = reinterpret_cast<s_smpi_mpi_hindexed_t*>(type);
 889   int i,j;
 890
 891   char* contiguous_hindexed_char = static_cast<char*>(contiguous_hindexed);
 892   char* noncontiguous_hindexed_char = static_cast<char*>(noncontiguous_hindexed)+ type_c->block_indices[0];
 893   for(j=0; j<count;j++){
 894     for (i = 0; i < type_c->block_count; i++) {
 895       if (type_c->old_type->sizeof_substruct == 0)
 896         smpi_op_apply(op, contiguous_hindexed_char, noncontiguous_hindexed_char, &type_c->block_lengths[i],
 897                             &type_c->old_type);
 898       else
 899         static_cast<s_smpi_subtype_t*>(type_c->old_type->substruct)->unserialize( contiguous_hindexed_char,
 900                                                                        noncontiguous_hindexed_char,
 901                                                                        type_c->block_lengths[i],
 902                                                                        type_c->old_type->substruct, op);
 903
 904       contiguous_hindexed_char += type_c->block_lengths[i]*type_c->size_oldtype;
 905       if (i<type_c->block_count-1)
 906         noncontiguous_hindexed_char = static_cast<char*>(noncontiguous_hindexed) + type_c->block_indices[i+1];
 907       else
 908         noncontiguous_hindexed_char += type_c->block_lengths[i]*smpi_datatype_get_extent(type_c->old_type);
 909     }
 910     noncontiguous_hindexed=static_cast<void*>(noncontiguous_hindexed_char);
 911   }
 912 }
 913
 914 void free_hindexed(MPI_Datatype* type){
 915   if((*type)->in_use==0){
 916     xbt_free(reinterpret_cast<s_smpi_mpi_hindexed_t*>((*type)->substruct)->block_lengths);
 917     xbt_free(reinterpret_cast<s_smpi_mpi_hindexed_t*>((*type)->substruct)->block_indices);
 918   }
 919   smpi_datatype_unuse(reinterpret_cast<s_smpi_mpi_hindexed_t*>((*type)->substruct)->old_type);
 920 }
 921
 922 void use_hindexed(MPI_Datatype* type){
 923   smpi_datatype_use(reinterpret_cast<s_smpi_mpi_hindexed_t*>((*type)->substruct)->old_type);
 924 }
 925
 926 /* Create a Sub type hindexed to be able to serialize and unserialize it the structure s_smpi_mpi_hindexed_t is derived
 927  * from s_smpi_subtype which required the functions unserialize and serialize
 928  */
 929 s_smpi_mpi_hindexed_t* smpi_datatype_hindexed_create( int* block_lengths, MPI_Aint* block_indices, int block_count,
 930                                                   MPI_Datatype old_type, int size_oldtype){
 931   s_smpi_mpi_hindexed_t *new_t= xbt_new(s_smpi_mpi_hindexed_t,1);
 932   new_t->base.serialize = &serialize_hindexed;
 933   new_t->base.unserialize = &unserialize_hindexed;
 934   new_t->base.subtype_free = &free_hindexed;
 935   new_t->base.subtype_use = &use_hindexed;
 936   new_t->block_lengths= xbt_new(int, block_count);
 937   new_t->block_indices= xbt_new(MPI_Aint, block_count);
 938   for(int i=0;i<block_count;i++){
 939     new_t->block_lengths[i]=block_lengths[i];
 940     new_t->block_indices[i]=block_indices[i];
 941   }
 942   new_t->block_count = block_count;
 943   new_t->old_type = old_type;
 944   smpi_datatype_use(old_type);
 945   new_t->size_oldtype = size_oldtype;
 946   return new_t;
 947 }
 948
 949 int smpi_datatype_hindexed(int count, int* blocklens, MPI_Aint* indices, MPI_Datatype old_type, MPI_Datatype* new_type)
 950 {
 951   int i;
 952   int retval;
 953   int size = 0;
 954   bool contiguous=true;
 955   MPI_Aint lb = 0;
 956   MPI_Aint ub = 0;
 957   if(count>0){
 958     lb=indices[0] + smpi_datatype_lb(old_type);
 959     ub=indices[0] + blocklens[0]*smpi_datatype_ub(old_type);
 960   }
 961   for(i=0; i< count; i++){
 962     if   (blocklens[i]<0)
 963       return MPI_ERR_ARG;
 964     size += blocklens[i];
 965
 966     if(indices[i]+smpi_datatype_lb(old_type)<lb)
 967       lb = indices[i]+smpi_datatype_lb(old_type);
 968     if(indices[i]+blocklens[i]*smpi_datatype_ub(old_type)>ub)
 969       ub = indices[i]+blocklens[i]*smpi_datatype_ub(old_type);
 970
 971     if ( (i< count -1) && (indices[i]+blocklens[i]*(static_cast<int>(smpi_datatype_size(old_type))) != indices[i+1]) )
 972       contiguous=false;
 973   }
 974   if (old_type->sizeof_substruct != 0 || lb!=0)
 975     contiguous=false;
 976
 977   if(!contiguous){
 978     s_smpi_mpi_hindexed_t* subtype = smpi_datatype_hindexed_create( blocklens, indices, count, old_type,
 979                                                                   smpi_datatype_size(old_type));
 980     smpi_datatype_create(new_type,  size * smpi_datatype_size(old_type), lb, ub ,sizeof(s_smpi_mpi_hindexed_t), subtype, DT_FLAG_DATA);
 981   }else{
 982     s_smpi_mpi_contiguous_t* subtype = smpi_datatype_contiguous_create(lb,size, old_type, smpi_datatype_size(old_type));
 983     smpi_datatype_create(new_type,  size * smpi_datatype_size(old_type), 0,size * smpi_datatype_size(old_type),
 984                1, subtype, DT_FLAG_DATA|DT_FLAG_CONTIGUOUS);
 985   }
 986   retval=MPI_SUCCESS;
 987   return retval;
 988 }
 989
 990 /* struct Implementation - Indexed with indices in bytes */
 991
 992 /* Copies noncontiguous data into contiguous memory.
 993  *  @param contiguous_struct - output struct
 994  *  @param noncontiguous_struct - input struct
 995  *  @param type - pointer contening :
 996  *      - stride - stride of between noncontiguous data
 997  *      - block_length - the width or height of blocked matrix
 998  *      - count - the number of rows of matrix
 999  */
1000 void serialize_struct( void* noncontiguous_struct, void *contiguous_struct, int count, void *type)
1001 {
1002   s_smpi_mpi_struct_t* type_c = reinterpret_cast<s_smpi_mpi_struct_t*>(type);
1003   int i,j;
1004   char* contiguous_struct_char = static_cast<char*>(contiguous_struct);
1005   char* noncontiguous_struct_char = static_cast<char*>(noncontiguous_struct)+ type_c->block_indices[0];
1006   for(j=0; j<count;j++){
1007     for (i = 0; i < type_c->block_count; i++) {
1008       if (type_c->old_types[i]->sizeof_substruct == 0)
1009         memcpy(contiguous_struct_char, noncontiguous_struct_char,
1010                type_c->block_lengths[i] * smpi_datatype_size(type_c->old_types[i]));
1011       else
1012         static_cast<s_smpi_subtype_t*>(type_c->old_types[i]->substruct)->serialize( noncontiguous_struct_char,
1013                                                                          contiguous_struct_char,
1014                                                                          type_c->block_lengths[i],
1015                                                                          type_c->old_types[i]->substruct);
1016
1017
1018       contiguous_struct_char += type_c->block_lengths[i]*smpi_datatype_size(type_c->old_types[i]);
1019       if (i<type_c->block_count-1)
1020         noncontiguous_struct_char = static_cast<char*>(noncontiguous_struct) + type_c->block_indices[i+1];
1021       else //let's hope this is MPI_UB ?
1022         noncontiguous_struct_char += type_c->block_lengths[i]*smpi_datatype_get_extent(type_c->old_types[i]);
1023     }
1024     noncontiguous_struct=static_cast<void*>(noncontiguous_struct_char);
1025   }
1026 }
1027
1028 /* Copies contiguous data into noncontiguous memory.
1029  *  @param noncontiguous_struct - output struct
1030  *  @param contiguous_struct - input struct
1031  *  @param type - pointer contening :
1032  *      - stride - stride of between noncontiguous data
1033  *      - block_length - the width or height of blocked matrix
1034  *      - count - the number of rows of matrix
1035  */
1036 void unserialize_struct( void* contiguous_struct, void *noncontiguous_struct, int count, void *type, MPI_Op op)
1037 {
1038   s_smpi_mpi_struct_t* type_c = reinterpret_cast<s_smpi_mpi_struct_t*>(type);
1039   int i,j;
1040
1041   char* contiguous_struct_char = static_cast<char*>(contiguous_struct);
1042   char* noncontiguous_struct_char = static_cast<char*>(noncontiguous_struct)+ type_c->block_indices[0];
1043   for(j=0; j<count;j++){
1044     for (i = 0; i < type_c->block_count; i++) {
1045       if (type_c->old_types[i]->sizeof_substruct == 0)
1046         smpi_op_apply(op, contiguous_struct_char, noncontiguous_struct_char, &type_c->block_lengths[i],
1047            & type_c->old_types[i]);
1048       else
1049         static_cast<s_smpi_subtype_t*>(type_c->old_types[i]->substruct)->unserialize( contiguous_struct_char,
1050                                                                            noncontiguous_struct_char,
1051                                                                            type_c->block_lengths[i],
1052                                                                            type_c->old_types[i]->substruct, op);
1053
1054       contiguous_struct_char += type_c->block_lengths[i]*smpi_datatype_size(type_c->old_types[i]);
1055       if (i<type_c->block_count-1)
1056         noncontiguous_struct_char =  static_cast<char*>(noncontiguous_struct) + type_c->block_indices[i+1];
1057       else
1058         noncontiguous_struct_char += type_c->block_lengths[i]*smpi_datatype_get_extent(type_c->old_types[i]);
1059     }
1060     noncontiguous_struct=reinterpret_cast<void*>(noncontiguous_struct_char);
1061   }
1062 }
1063
1064 void free_struct(MPI_Datatype* type){
1065   int i=0;
1066   for (i = 0; i < reinterpret_cast<s_smpi_mpi_struct_t*>((*type)->substruct)->block_count; i++)
1067     smpi_datatype_unuse(reinterpret_cast<s_smpi_mpi_struct_t*>((*type)->substruct)->old_types[i]);
1068   if((*type)->in_use==0){
1069     xbt_free(reinterpret_cast<s_smpi_mpi_struct_t*>((*type)->substruct)->block_lengths);
1070     xbt_free(reinterpret_cast<s_smpi_mpi_struct_t*>((*type)->substruct)->block_indices);
1071     xbt_free(reinterpret_cast<s_smpi_mpi_struct_t*>((*type)->substruct)->old_types);
1072   }
1073 }
1074
1075 void use_struct(MPI_Datatype* type){
1076   int i=0;
1077   for (i = 0; i < reinterpret_cast<s_smpi_mpi_struct_t*>((*type)->substruct)->block_count; i++)
1078     smpi_datatype_use(reinterpret_cast<s_smpi_mpi_struct_t*>((*type)->substruct)->old_types[i]);
1079 }
1080
1081 /* Create a Sub type struct to be able to serialize and unserialize it the structure s_smpi_mpi_struct_t is derived
1082  * from s_smpi_subtype which required the functions unserialize and serialize
1083  */
1084 s_smpi_mpi_struct_t* smpi_datatype_struct_create( int* block_lengths, MPI_Aint* block_indices, int block_count,
1085                                                   MPI_Datatype* old_types){
1086   s_smpi_mpi_struct_t *new_t= xbt_new(s_smpi_mpi_struct_t,1);
1087   new_t->base.serialize = &serialize_struct;
1088   new_t->base.unserialize = &unserialize_struct;
1089   new_t->base.subtype_free = &free_struct;
1090   new_t->base.subtype_use = &use_struct;
1091   new_t->block_lengths= xbt_new(int, block_count);
1092   new_t->block_indices= xbt_new(MPI_Aint, block_count);
1093   new_t->old_types=  xbt_new(MPI_Datatype, block_count);
1094   int i;
1095   for(i=0;i<block_count;i++){
1096     new_t->block_lengths[i]=block_lengths[i];
1097     new_t->block_indices[i]=block_indices[i];
1098     new_t->old_types[i]=old_types[i];
1099     smpi_datatype_use(new_t->old_types[i]);
1100   }
1101   new_t->block_count = block_count;
1102   return new_t;
1103 }
1104
1105 int smpi_datatype_struct(int count, int* blocklens, MPI_Aint* indices, MPI_Datatype* old_types, MPI_Datatype* new_type)
1106 {
1107   int i;
1108   size_t size = 0;
1109   bool contiguous=true;
1110   size = 0;
1111   MPI_Aint lb = 0;
1112   MPI_Aint ub = 0;
1113   if(count>0){
1114     lb=indices[0] + smpi_datatype_lb(old_types[0]);
1115     ub=indices[0] + blocklens[0]*smpi_datatype_ub(old_types[0]);
1116   }
1117   bool forced_lb=false;
1118   bool forced_ub=false;
1119   for(i=0; i< count; i++){
1120     if (blocklens[i]<0)
1121       return MPI_ERR_ARG;
1122     if (old_types[i]->sizeof_substruct != 0)
1123       contiguous=false;
1124
1125     size += blocklens[i]*smpi_datatype_size(old_types[i]);
1126     if (old_types[i]==MPI_LB){
1127       lb=indices[i];
1128       forced_lb=true;
1129     }
1130     if (old_types[i]==MPI_UB){
1131       ub=indices[i];
1132       forced_ub=true;
1133     }
1134
1135     if(!forced_lb && indices[i]+smpi_datatype_lb(old_types[i])<lb)
1136       lb = indices[i];
1137     if(!forced_ub &&  indices[i]+blocklens[i]*smpi_datatype_ub(old_types[i])>ub)
1138       ub = indices[i]+blocklens[i]*smpi_datatype_ub(old_types[i]);
1139
1140     if ( (i< count -1) && (indices[i]+blocklens[i]*static_cast<int>(smpi_datatype_size(old_types[i])) != indices[i+1]) )
1141       contiguous=false;
1142   }
1143
1144   if(!contiguous){
1145     s_smpi_mpi_struct_t* subtype = smpi_datatype_struct_create( blocklens, indices, count, old_types);
1146
1147     smpi_datatype_create(new_type,  size, lb, ub,sizeof(s_smpi_mpi_struct_t), subtype, DT_FLAG_DATA);
1148   }else{
1149     s_smpi_mpi_contiguous_t* subtype = smpi_datatype_contiguous_create( lb, size, MPI_CHAR, 1);
1150     smpi_datatype_create(new_type,  size, lb, ub,1, subtype, DT_FLAG_DATA|DT_FLAG_CONTIGUOUS);
1151   }
1152   return MPI_SUCCESS;
1153 }
1154
1155 void smpi_datatype_commit(MPI_Datatype *datatype)
1156 {
1157   (*datatype)->flags=  ((*datatype)->flags | DT_FLAG_COMMITED);
1158 }
1159
1160 typedef struct s_smpi_mpi_op {
1161   MPI_User_function *func;
1162   bool is_commute;
1163   bool is_fortran_op;
1164 } s_smpi_mpi_op_t;
1165
1166 #define MAX_OP(a, b)  (b) = (a) < (b) ? (b) : (a)
1167 #define MIN_OP(a, b)  (b) = (a) < (b) ? (a) : (b)
1168 #define SUM_OP(a, b)  (b) += (a)
1169 #define PROD_OP(a, b) (b) *= (a)
1170 #define LAND_OP(a, b) (b) = (a) && (b)
1171 #define LOR_OP(a, b)  (b) = (a) || (b)
1172 #define LXOR_OP(a, b) (b) = (!(a) && (b)) || ((a) && !(b))
1173 #define BAND_OP(a, b) (b) &= (a)
1174 #define BOR_OP(a, b)  (b) |= (a)
1175 #define BXOR_OP(a, b) (b) ^= (a)
1176 #define MAXLOC_OP(a, b)  (b) = (a.value) < (b.value) ? (b) : (a)
1177 #define MINLOC_OP(a, b)  (b) = (a.value) < (b.value) ? (a) : (b)
1178
1179 #define APPLY_FUNC(a, b, length, type, func) \
1180 {                                          \
1181   int i;                                   \
1182   type* x = (type*)(a);                    \
1183   type* y = (type*)(b);                    \
1184   for(i = 0; i < *(length); i++) {         \
1185     func(x[i], y[i]);                      \
1186   }                                        \
1187 }
1188
1189 #define APPLY_OP_LOOP(dtype, type, op) \
1190   if (*datatype == dtype) {\
1191     APPLY_FUNC(a, b, length, type, op)\
1192   } else \
1193
1194
1195 #define APPLY_BASIC_OP_LOOP(op)\
1196 APPLY_OP_LOOP(MPI_CHAR, char,op)\
1197 APPLY_OP_LOOP(MPI_SHORT, short,op)\
1198 APPLY_OP_LOOP(MPI_INT, int,op)\
1199 APPLY_OP_LOOP(MPI_LONG, long,op)\
1200 APPLY_OP_LOOP(MPI_LONG_LONG, long long,op)\
1201 APPLY_OP_LOOP(MPI_SIGNED_CHAR, signed char,op)\
1202 APPLY_OP_LOOP(MPI_UNSIGNED_CHAR, unsigned char,op)\
1203 APPLY_OP_LOOP(MPI_UNSIGNED_SHORT, unsigned short,op)\
1204 APPLY_OP_LOOP(MPI_UNSIGNED, unsigned int,op)\
1205 APPLY_OP_LOOP(MPI_UNSIGNED_LONG, unsigned long,op)\
1206 APPLY_OP_LOOP(MPI_UNSIGNED_LONG_LONG, unsigned long long,op)\
1207 APPLY_OP_LOOP(MPI_WCHAR, wchar_t,op)\
1208 APPLY_OP_LOOP(MPI_BYTE, int8_t,op)\
1209 APPLY_OP_LOOP(MPI_INT8_T, int8_t,op)\
1210 APPLY_OP_LOOP(MPI_INT16_T, int16_t,op)\
1211 APPLY_OP_LOOP(MPI_INT32_T, int32_t,op)\
1212 APPLY_OP_LOOP(MPI_INT64_T, int64_t,op)\
1213 APPLY_OP_LOOP(MPI_UINT8_T, uint8_t,op)\
1214 APPLY_OP_LOOP(MPI_UINT16_T, uint16_t,op)\
1215 APPLY_OP_LOOP(MPI_UINT32_T, uint32_t,op)\
1216 APPLY_OP_LOOP(MPI_UINT64_T, uint64_t,op)\
1217 APPLY_OP_LOOP(MPI_AINT, MPI_Aint,op)\
1218 APPLY_OP_LOOP(MPI_OFFSET, MPI_Offset,op)\
1219 APPLY_OP_LOOP(MPI_INTEGER1, int,op)\
1220 APPLY_OP_LOOP(MPI_INTEGER2, int16_t,op)\
1221 APPLY_OP_LOOP(MPI_INTEGER4, int32_t,op)\
1222 APPLY_OP_LOOP(MPI_INTEGER8, int64_t,op)
1223
1224 #define APPLY_BOOL_OP_LOOP(op)\
1225 APPLY_OP_LOOP(MPI_C_BOOL, bool,op)
1226
1227 #define APPLY_FLOAT_OP_LOOP(op)\
1228 APPLY_OP_LOOP(MPI_FLOAT, float,op)\
1229 APPLY_OP_LOOP(MPI_DOUBLE, double,op)\
1230 APPLY_OP_LOOP(MPI_LONG_DOUBLE, long double,op)\
1231 APPLY_OP_LOOP(MPI_REAL, float,op)\
1232 APPLY_OP_LOOP(MPI_REAL4, float,op)\
1233 APPLY_OP_LOOP(MPI_REAL8, float,op)\
1234 APPLY_OP_LOOP(MPI_REAL16, double,op)
1235
1236 #define APPLY_COMPLEX_OP_LOOP(op)\
1237 APPLY_OP_LOOP(MPI_C_FLOAT_COMPLEX, float _Complex,op)\
1238 APPLY_OP_LOOP(MPI_C_DOUBLE_COMPLEX, double _Complex,op)\
1239 APPLY_OP_LOOP(MPI_C_LONG_DOUBLE_COMPLEX, long double _Complex,op)
1240
1241 #define APPLY_PAIR_OP_LOOP(op)\
1242 APPLY_OP_LOOP(MPI_FLOAT_INT, float_int,op)\
1243 APPLY_OP_LOOP(MPI_LONG_INT, long_int,op)\
1244 APPLY_OP_LOOP(MPI_DOUBLE_INT, double_int,op)\
1245 APPLY_OP_LOOP(MPI_SHORT_INT, short_int,op)\
1246 APPLY_OP_LOOP(MPI_2INT, int_int,op)\
1247 APPLY_OP_LOOP(MPI_2FLOAT, float_float,op)\
1248 APPLY_OP_LOOP(MPI_2DOUBLE, double_double,op)\
1249 APPLY_OP_LOOP(MPI_LONG_DOUBLE_INT, long_double_int,op)\
1250 APPLY_OP_LOOP(MPI_2LONG, long_long,op)
1251
1252 #define APPLY_END_OP_LOOP(op)\
1253   {\
1254     xbt_die("Failed to apply " #op " to type %s", (*datatype)->name);\
1255   }
1256
1257
1258 static void max_func(void *a, void *b, int *length, MPI_Datatype * datatype)
1259 {
1260   APPLY_BASIC_OP_LOOP(MAX_OP)
1261   APPLY_FLOAT_OP_LOOP(MAX_OP)
1262   APPLY_END_OP_LOOP(MAX_OP)
1263 }
1264
1265 static void min_func(void *a, void *b, int *length, MPI_Datatype * datatype)
1266 {
1267   APPLY_BASIC_OP_LOOP(MIN_OP)
1268   APPLY_FLOAT_OP_LOOP(MIN_OP)
1269   APPLY_END_OP_LOOP(MIN_OP)
1270 }
1271
1272 static void sum_func(void *a, void *b, int *length, MPI_Datatype * datatype)
1273 {
1274   APPLY_BASIC_OP_LOOP(SUM_OP)
1275   APPLY_FLOAT_OP_LOOP(SUM_OP)
1276   APPLY_COMPLEX_OP_LOOP(SUM_OP)
1277   APPLY_END_OP_LOOP(SUM_OP)
1278 }
1279
1280 static void prod_func(void *a, void *b, int *length, MPI_Datatype * datatype)
1281 {
1282   APPLY_BASIC_OP_LOOP(PROD_OP)
1283   APPLY_FLOAT_OP_LOOP(PROD_OP)
1284   APPLY_COMPLEX_OP_LOOP(PROD_OP)
1285   APPLY_END_OP_LOOP(PROD_OP)
1286 }
1287
1288 static void land_func(void *a, void *b, int *length, MPI_Datatype * datatype)
1289 {
1290   APPLY_BASIC_OP_LOOP(LAND_OP)
1291   APPLY_BOOL_OP_LOOP(LAND_OP)
1292   APPLY_END_OP_LOOP(LAND_OP)
1293 }
1294
1295 static void lor_func(void *a, void *b, int *length, MPI_Datatype * datatype)
1296 {
1297   APPLY_BASIC_OP_LOOP(LOR_OP)
1298   APPLY_BOOL_OP_LOOP(LOR_OP)
1299   APPLY_END_OP_LOOP(LOR_OP)
1300 }
1301
1302 static void lxor_func(void *a, void *b, int *length, MPI_Datatype * datatype)
1303 {
1304   APPLY_BASIC_OP_LOOP(LXOR_OP)
1305   APPLY_BOOL_OP_LOOP(LXOR_OP)
1306   APPLY_END_OP_LOOP(LXOR_OP)
1307 }
1308
1309 static void band_func(void *a, void *b, int *length, MPI_Datatype * datatype)
1310 {
1311   APPLY_BASIC_OP_LOOP(BAND_OP)
1312   APPLY_BOOL_OP_LOOP(BAND_OP)
1313   APPLY_END_OP_LOOP(BAND_OP)
1314 }
1315
1316 static void bor_func(void *a, void *b, int *length, MPI_Datatype * datatype)
1317 {
1318   APPLY_BASIC_OP_LOOP(BOR_OP)
1319   APPLY_BOOL_OP_LOOP(BOR_OP)
1320   APPLY_END_OP_LOOP(BOR_OP)
1321 }
1322
1323 static void bxor_func(void *a, void *b, int *length, MPI_Datatype * datatype)
1324 {
1325   APPLY_BASIC_OP_LOOP(BXOR_OP)
1326   APPLY_BOOL_OP_LOOP(BXOR_OP)
1327   APPLY_END_OP_LOOP(BXOR_OP)
1328 }
1329
1330 static void minloc_func(void *a, void *b, int *length, MPI_Datatype * datatype)
1331 {
1332   APPLY_PAIR_OP_LOOP(MINLOC_OP)
1333   APPLY_END_OP_LOOP(MINLOC_OP)
1334 }
1335
1336 static void maxloc_func(void *a, void *b, int *length, MPI_Datatype * datatype)
1337 {
1338   APPLY_PAIR_OP_LOOP(MAXLOC_OP)
1339   APPLY_END_OP_LOOP(MAXLOC_OP)
1340 }
1341
1342 static void replace_func(void *a, void *b, int *length, MPI_Datatype * datatype)
1343 {
1344   memcpy(b, a, *length * smpi_datatype_size(*datatype));
1345 }
1346
1347 #define CREATE_MPI_OP(name, func)                             \
1348   static s_smpi_mpi_op_t mpi_##name = { &(func) /* func */, true, false }; \
1349 MPI_Op name = &mpi_##name;
1350
1351 CREATE_MPI_OP(MPI_MAX, max_func);
1352 CREATE_MPI_OP(MPI_MIN, min_func);
1353 CREATE_MPI_OP(MPI_SUM, sum_func);
1354 CREATE_MPI_OP(MPI_PROD, prod_func);
1355 CREATE_MPI_OP(MPI_LAND, land_func);
1356 CREATE_MPI_OP(MPI_LOR, lor_func);
1357 CREATE_MPI_OP(MPI_LXOR, lxor_func);
1358 CREATE_MPI_OP(MPI_BAND, band_func);
1359 CREATE_MPI_OP(MPI_BOR, bor_func);
1360 CREATE_MPI_OP(MPI_BXOR, bxor_func);
1361 CREATE_MPI_OP(MPI_MAXLOC, maxloc_func);
1362 CREATE_MPI_OP(MPI_MINLOC, minloc_func);
1363 CREATE_MPI_OP(MPI_REPLACE, replace_func);
1364
1365 MPI_Op smpi_op_new(MPI_User_function * function, bool commute)
1366 {
1367   MPI_Op op;
1368   op = xbt_new(s_smpi_mpi_op_t, 1);
1369   op->func = function;
1370   op-> is_commute = commute;
1371   op-> is_fortran_op = false;
1372   return op;
1373 }
1374
1375 bool smpi_op_is_commute(MPI_Op op)
1376 {
1377   return (op==MPI_OP_NULL) ? true : op-> is_commute;
1378 }
1379
1380 void smpi_op_destroy(MPI_Op op)
1381 {
1382   xbt_free(op);
1383 }
1384
1385 void smpi_op_set_fortran(MPI_Op op)
1386 {
1387   //tell that we were created from fortran, so we need to translate the type to fortran when called
1388   op->is_fortran_op = true;
1389 }
1390
1391 void smpi_op_apply(MPI_Op op, void *invec, void *inoutvec, int *len, MPI_Datatype * datatype)
1392 {
1393   if(op==MPI_OP_NULL)
1394     return;
1395
1396   if(smpi_privatize_global_variables){//we need to switch as the called function may silently touch global variables
1397     XBT_DEBUG("Applying operation, switch to the right data frame ");
1398     smpi_switch_data_segment(smpi_process_index());
1399   }
1400
1401   if(!smpi_process_get_replaying()){
1402     if(! op->is_fortran_op)
1403       op->func(invec, inoutvec, len, datatype);
1404     else{
1405       int tmp = smpi_type_c2f(*datatype);
1406       op->func(invec, inoutvec, len, reinterpret_cast<MPI_Datatype*>(&tmp) );
1407     }
1408   }
1409 }
1410
1411 int smpi_type_attr_delete(MPI_Datatype type, int keyval){
1412   smpi_type_key_elem elem =
1413     static_cast<smpi_type_key_elem>(xbt_dict_get_or_null_ext(smpi_type_keyvals, reinterpret_cast<const char*>(&keyval), sizeof(int)));
1414   if(elem==nullptr)
1415     return MPI_ERR_ARG;
1416   if(elem->delete_fn!=MPI_NULL_DELETE_FN){
1417     void * value = nullptr;
1418     int flag;
1419     if(smpi_type_attr_get(type, keyval, &value, &flag)==MPI_SUCCESS){
1420       int ret = elem->delete_fn(type, keyval, value, &flag);
1421       if(ret!=MPI_SUCCESS)
1422         return ret;
1423     }
1424   }
1425   if(type->attributes==nullptr)
1426     return MPI_ERR_ARG;
1427
1428   xbt_dict_remove_ext(type->attributes, reinterpret_cast<const char*>(&keyval), sizeof(int));
1429   return MPI_SUCCESS;
1430 }
1431
1432 int smpi_type_attr_get(MPI_Datatype type, int keyval, void* attr_value, int* flag){
1433   smpi_type_key_elem elem =
1434     static_cast<smpi_type_key_elem>(xbt_dict_get_or_null_ext(smpi_type_keyvals, reinterpret_cast<const char*>(&keyval), sizeof(int)));
1435   if(elem==nullptr)
1436     return MPI_ERR_ARG;
1437   if(type->attributes==nullptr){
1438     *flag=0;
1439     return MPI_SUCCESS;
1440   }
1441   try {
1442     *static_cast<void**>(attr_value) = xbt_dict_get_ext(type->attributes, reinterpret_cast<const char*>(&keyval), sizeof(int));
1443     *flag=1;
1444   }
1445   catch (xbt_ex& ex) {
1446     *flag=0;
1447   }
1448   return MPI_SUCCESS;
1449 }
1450
1451 int smpi_type_attr_put(MPI_Datatype type, int keyval, void* attr_value){
1452   if(smpi_type_keyvals==nullptr)
1453     smpi_type_keyvals = xbt_dict_new_homogeneous(nullptr);
1454   smpi_type_key_elem elem =
1455      static_cast<smpi_type_key_elem>(xbt_dict_get_or_null_ext(smpi_type_keyvals, reinterpret_cast<const char*>(&keyval), sizeof(int)));
1456   if(elem==nullptr)
1457     return MPI_ERR_ARG;
1458   int flag;
1459   void* value = nullptr;
1460   smpi_type_attr_get(type, keyval, &value, &flag);
1461   if(flag!=0 && elem->delete_fn!=MPI_NULL_DELETE_FN){
1462     int ret = elem->delete_fn(type, keyval, value, &flag);
1463     if(ret!=MPI_SUCCESS)
1464       return ret;
1465   }
1466   if(type->attributes==nullptr)
1467     type->attributes = xbt_dict_new_homogeneous(nullptr);
1468
1469   xbt_dict_set_ext(type->attributes, reinterpret_cast<const char*>(&keyval), sizeof(int), attr_value, nullptr);
1470   return MPI_SUCCESS;
1471 }
1472
1473 int smpi_type_keyval_create(MPI_Type_copy_attr_function* copy_fn, MPI_Type_delete_attr_function* delete_fn, int* keyval,
1474                             void* extra_state){
1475   if(smpi_type_keyvals==nullptr)
1476     smpi_type_keyvals = xbt_dict_new_homogeneous(nullptr);
1477
1478   smpi_type_key_elem value = (smpi_type_key_elem) xbt_new0(s_smpi_mpi_type_key_elem_t,1);
1479
1480   value->copy_fn=copy_fn;
1481   value->delete_fn=delete_fn;
1482
1483   *keyval = type_keyval_id;
1484   xbt_dict_set_ext(smpi_type_keyvals,reinterpret_cast<const char*>(keyval), sizeof(int),reinterpret_cast<void*>(value), nullptr);
1485   type_keyval_id++;
1486   return MPI_SUCCESS;
1487 }
1488
1489 int smpi_type_keyval_free(int* keyval){
1490   smpi_type_key_elem elem =
1491     static_cast<smpi_type_key_elem>(xbt_dict_get_or_null_ext(smpi_type_keyvals, reinterpret_cast<const char*>(keyval), sizeof(int)));
1492   if(elem==0){
1493     return MPI_ERR_ARG;
1494   }
1495   xbt_dict_remove_ext(smpi_type_keyvals, reinterpret_cast<const char*>(keyval), sizeof(int));
1496   xbt_free(elem);
1497   return MPI_SUCCESS;
1498 }
1499
1500 int smpi_mpi_pack(void* inbuf, int incount, MPI_Datatype type, void* outbuf, int outcount, int* position,MPI_Comm comm){
1501   size_t size = smpi_datatype_size(type);
1502   if (outcount - *position < incount*static_cast<int>(size))
1503     return MPI_ERR_BUFFER;
1504   smpi_datatype_copy(inbuf, incount, type, static_cast<char*>(outbuf) + *position, outcount, MPI_CHAR);
1505   *position += incount * size;
1506   return MPI_SUCCESS;
1507 }
1508
1509 int smpi_mpi_unpack(void* inbuf, int insize, int* position, void* outbuf, int outcount, MPI_Datatype type,MPI_Comm comm){
1510   int size = static_cast<int>(smpi_datatype_size(type));
1511   if (outcount*size> insize)
1512     return MPI_ERR_BUFFER;
1513   smpi_datatype_copy(static_cast<char*>(inbuf) + *position, insize, MPI_CHAR, outbuf, outcount, type);
1514   *position += outcount * size;
1515   return MPI_SUCCESS;
1516 }