Merge branch 'master' of git+ssh://scm.gforge.inria.fr//gitroot/simgrid/simgrid

[simgrid.git] / src / smpi / smpi_mpi_dt.cpp

diff --git a/src/smpi/smpi_mpi_dt.cpp b/src/smpi/smpi_mpi_dt.cpp
index 81cfd4e..630ca73 100644 (file)
--- a/src/smpi/smpi_mpi_dt.cpp
+++ b/src/smpi/smpi_mpi_dt.cpp
@@ -1,6 +1,5 @@
 /* smpi_mpi_dt.c -- MPI primitives to handle datatypes                      */
-/* Copyright (c) 2009-2015. The SimGrid Team.
- * All rights reserved.                                                     */
+/* Copyright (c) 2009-2017. The SimGrid Team. All rights reserved.          */
 
 /* This program is free software; you can redistribute it and/or modify it
  * under the terms of the license (GNU LGPL) which comes with this package. */
@@ -293,12 +292,12 @@ int smpi_datatype_copy(void *sendbuf, int sendcount, MPI_Datatype sendtype,
  *      - block_length - the width or height of blocked matrix
  *      - count - the number of rows of matrix
  */
-void serialize_vector( const void *noncontiguous_vector, void *contiguous_vector, int count, void *type)
+void serialize_vector( void* noncontiguous_vector, void *contiguous_vector, int count, void *type)
 {
   s_smpi_mpi_vector_t* type_c = reinterpret_cast<s_smpi_mpi_vector_t*>(type);
   int i;
   char* contiguous_vector_char = static_cast<char*>(contiguous_vector);
-  const char* noncontiguous_vector_char = static_cast<const char*>(noncontiguous_vector);
+  char* noncontiguous_vector_char = static_cast<char*>(noncontiguous_vector);
 
   for (i = 0; i < type_c->block_count * count; i++) {
       if (type_c->old_type->sizeof_substruct == 0)
@@ -325,12 +324,12 @@ void serialize_vector( const void *noncontiguous_vector, void *contiguous_vector
  *      - block_length - the width or height of blocked matrix
  *      - count - the number of rows of matrix
  */
-void unserialize_vector( const void *contiguous_vector, void *noncontiguous_vector, int count, void *type, MPI_Op op)
+void unserialize_vector( void* contiguous_vector, void *noncontiguous_vector, int count, void *type, MPI_Op op)
 {
   s_smpi_mpi_vector_t* type_c = reinterpret_cast<s_smpi_mpi_vector_t*>(type);
   int i;
 
-  const char* contiguous_vector_char = static_cast<const char*>(contiguous_vector);
+  char* contiguous_vector_char = static_cast<char*>(contiguous_vector);
   char* noncontiguous_vector_char = static_cast<char*>(noncontiguous_vector);
 
   for (i = 0; i < type_c->block_count * count; i++) {
@@ -424,8 +423,9 @@ void smpi_datatype_free(MPI_Datatype* type){
 
 void smpi_datatype_use(MPI_Datatype type){
 
-  if(type != MPI_DATATYPE_NULL)
-    type->in_use++;
+  if(type == MPI_DATATYPE_NULL)
+    return;
+  type->in_use++;
 
   if(type->sizeof_substruct!=0){
     static_cast<s_smpi_subtype_t *>((type)->substruct)->subtype_use(&type);  
@@ -467,11 +467,11 @@ void smpi_datatype_unuse(MPI_Datatype type)
  *      - block_length - the width or height of blocked matrix
  *      - count - the number of rows of matrix
  */
-void serialize_contiguous( const void *noncontiguous_hvector, void *contiguous_hvector, int count, void *type)
+void serialize_contiguous( void* noncontiguous_hvector, void *contiguous_hvector, int count, void *type)
 {
   s_smpi_mpi_contiguous_t* type_c = reinterpret_cast<s_smpi_mpi_contiguous_t*>(type);
   char* contiguous_vector_char = static_cast<char*>(contiguous_hvector);
-  const char* noncontiguous_vector_char = static_cast<const char*>(noncontiguous_hvector)+type_c->lb;
+  char* noncontiguous_vector_char = static_cast<char*>(noncontiguous_hvector)+type_c->lb;
   memcpy(contiguous_vector_char, noncontiguous_vector_char, count* type_c->block_count * type_c->size_oldtype);
 }
 /* Copies contiguous data into noncontiguous memory.
@@ -482,10 +482,10 @@ void serialize_contiguous( const void *noncontiguous_hvector, void *contiguous_h
  *      - block_length - the width or height of blocked matrix
  *      - count - the number of rows of matrix
  */
-void unserialize_contiguous(const void *contiguous_vector, void *noncontiguous_vector, int count, void *type, MPI_Op op)
+void unserialize_contiguous(void* contiguous_vector, void *noncontiguous_vector, int count, void *type, MPI_Op op)
 {
   s_smpi_mpi_contiguous_t* type_c = reinterpret_cast<s_smpi_mpi_contiguous_t*>(type);
-  const char* contiguous_vector_char = static_cast<const char*>(contiguous_vector);
+  char* contiguous_vector_char = static_cast<char*>(contiguous_vector);
   char* noncontiguous_vector_char = static_cast<char*>(noncontiguous_vector)+type_c->lb;
   int n= count* type_c->block_count;
   smpi_op_apply(op, contiguous_vector_char, noncontiguous_vector_char, &n, &type_c->old_type);
@@ -580,12 +580,12 @@ void use_vector(MPI_Datatype* d){
  *      - block_length - the width or height of blocked matrix
  *      - count - the number of rows of matrix
  */
-void serialize_hvector( const void *noncontiguous_hvector, void *contiguous_hvector, int count, void *type)
+void serialize_hvector( void* noncontiguous_hvector, void *contiguous_hvector, int count, void *type)
 {
   s_smpi_mpi_hvector_t* type_c = reinterpret_cast<s_smpi_mpi_hvector_t*>(type);
   int i;
   char* contiguous_vector_char = static_cast<char*>(contiguous_hvector);
-  const char* noncontiguous_vector_char = static_cast<const char*>(noncontiguous_hvector);
+  char* noncontiguous_vector_char = static_cast<char*>(noncontiguous_hvector);
 
   for (i = 0; i < type_c->block_count * count; i++) {
     if (type_c->old_type->sizeof_substruct == 0)
@@ -610,12 +610,12 @@ void serialize_hvector( const void *noncontiguous_hvector, void *contiguous_hvec
  *      - block_length - the width or height of blocked matrix
  *      - count - the number of rows of matrix
  */
-void unserialize_hvector( const void *contiguous_vector, void *noncontiguous_vector, int count, void *type, MPI_Op op)
+void unserialize_hvector( void* contiguous_vector, void *noncontiguous_vector, int count, void *type, MPI_Op op)
 {
   s_smpi_mpi_hvector_t* type_c = reinterpret_cast<s_smpi_mpi_hvector_t*>(type);
   int i;
 
-  const char* contiguous_vector_char = static_cast<const char*>(contiguous_vector);
+  char* contiguous_vector_char = static_cast<char*>(contiguous_vector);
   char* noncontiguous_vector_char = static_cast<char*>(noncontiguous_vector);
 
   for (i = 0; i < type_c->block_count * count; i++) {
@@ -697,12 +697,12 @@ int smpi_datatype_hvector(int count, int blocklen, MPI_Aint stride, MPI_Datatype
  *      - block_indices - indices of each data, in element
  *      - count - the number of rows of matrix
  */
-void serialize_indexed( const void *noncontiguous_indexed, void *contiguous_indexed, int count, void *type)
+void serialize_indexed( void* noncontiguous_indexed, void *contiguous_indexed, int count, void *type)
 {
   s_smpi_mpi_indexed_t* type_c = reinterpret_cast<s_smpi_mpi_indexed_t*>(type);
   int i,j;
   char* contiguous_indexed_char = static_cast<char*>(contiguous_indexed);
-  const char* noncontiguous_indexed_char = static_cast<const char*>(noncontiguous_indexed)+type_c->block_indices[0] * type_c->size_oldtype;
+  char* noncontiguous_indexed_char = static_cast<char*>(noncontiguous_indexed)+type_c->block_indices[0] * type_c->size_oldtype;
   for(j=0; j<count;j++){
     for (i = 0; i < type_c->block_count; i++) {
       if (type_c->old_type->sizeof_substruct == 0)
@@ -716,11 +716,11 @@ void serialize_indexed( const void *noncontiguous_indexed, void *contiguous_inde
       contiguous_indexed_char += type_c->block_lengths[i]*type_c->size_oldtype;
       if (i<type_c->block_count-1)
         noncontiguous_indexed_char =
-          static_cast<const char*>(noncontiguous_indexed) + type_c->block_indices[i+1]*smpi_datatype_get_extent(type_c->old_type);
+          static_cast<char*>(noncontiguous_indexed) + type_c->block_indices[i+1]*smpi_datatype_get_extent(type_c->old_type);
       else
         noncontiguous_indexed_char += type_c->block_lengths[i]*smpi_datatype_get_extent(type_c->old_type);
     }
-    noncontiguous_indexed=static_cast<const void*>(noncontiguous_indexed_char);
+    noncontiguous_indexed=static_cast< void*>(noncontiguous_indexed_char);
   }
 }
 /* Copies contiguous data into noncontiguous memory.
@@ -731,11 +731,11 @@ void serialize_indexed( const void *noncontiguous_indexed, void *contiguous_inde
  *      - block_indices - indices of each data, in element
  *      - count - the number of rows of matrix
  */
-void unserialize_indexed( const void *contiguous_indexed, void *noncontiguous_indexed, int count, void *type, MPI_Op op)
+void unserialize_indexed( void* contiguous_indexed, void *noncontiguous_indexed, int count, void *type, MPI_Op op)
 {
   s_smpi_mpi_indexed_t* type_c = reinterpret_cast<s_smpi_mpi_indexed_t*>(type);
   int i,j;
-  const char* contiguous_indexed_char = static_cast<const char*>(contiguous_indexed);
+  char* contiguous_indexed_char = static_cast<char*>(contiguous_indexed);
   char* noncontiguous_indexed_char =
     static_cast<char*>(noncontiguous_indexed)+type_c->block_indices[0]*smpi_datatype_get_extent(type_c->old_type);
   for(j=0; j<count;j++){
@@ -848,12 +848,12 @@ int smpi_datatype_indexed(int count, int* blocklens, int* indices, MPI_Datatype
  *      - block_indices - indices of each data, in bytes
  *      - count - the number of rows of matrix
  */
-void serialize_hindexed( const void *noncontiguous_hindexed, void *contiguous_hindexed, int count, void *type)
+void serialize_hindexed( void* noncontiguous_hindexed, void *contiguous_hindexed, int count, void *type)
 {
   s_smpi_mpi_hindexed_t* type_c = reinterpret_cast<s_smpi_mpi_hindexed_t*>(type);
   int i,j;
   char* contiguous_hindexed_char = static_cast<char*>(contiguous_hindexed);
-  const char* noncontiguous_hindexed_char = static_cast<const char*>(noncontiguous_hindexed)+ type_c->block_indices[0];
+  char* noncontiguous_hindexed_char = static_cast<char*>(noncontiguous_hindexed)+ type_c->block_indices[0];
   for(j=0; j<count;j++){
     for (i = 0; i < type_c->block_count; i++) {
       if (type_c->old_type->sizeof_substruct == 0)
@@ -866,11 +866,11 @@ void serialize_hindexed( const void *noncontiguous_hindexed, void *contiguous_hi
 
       contiguous_hindexed_char += type_c->block_lengths[i]*type_c->size_oldtype;
       if (i<type_c->block_count-1)
-        noncontiguous_hindexed_char = static_cast<const char*>(noncontiguous_hindexed) + type_c->block_indices[i+1];
+        noncontiguous_hindexed_char = static_cast<char*>(noncontiguous_hindexed) + type_c->block_indices[i+1];
       else
         noncontiguous_hindexed_char += type_c->block_lengths[i]*smpi_datatype_get_extent(type_c->old_type);
     }
-    noncontiguous_hindexed=reinterpret_cast<const void*>(noncontiguous_hindexed_char);
+    noncontiguous_hindexed=static_cast<void*>(noncontiguous_hindexed_char);
   }
 }
 /* Copies contiguous data into noncontiguous memory.
@@ -881,13 +881,13 @@ void serialize_hindexed( const void *noncontiguous_hindexed, void *contiguous_hi
  *      - block_indices - indices of each data, in bytes
  *      - count - the number of rows of matrix
  */
-void unserialize_hindexed( const void *contiguous_hindexed, void *noncontiguous_hindexed, int count, void *type,
+void unserialize_hindexed( void* contiguous_hindexed, void *noncontiguous_hindexed, int count, void *type,
                          MPI_Op op)
 {
   s_smpi_mpi_hindexed_t* type_c = reinterpret_cast<s_smpi_mpi_hindexed_t*>(type);
   int i,j;
 
-  const char* contiguous_hindexed_char = static_cast<const char*>(contiguous_hindexed);
+  char* contiguous_hindexed_char = static_cast<char*>(contiguous_hindexed);
   char* noncontiguous_hindexed_char = static_cast<char*>(noncontiguous_hindexed)+ type_c->block_indices[0];
   for(j=0; j<count;j++){
     for (i = 0; i < type_c->block_count; i++) {
@@ -906,7 +906,7 @@ void unserialize_hindexed( const void *contiguous_hindexed, void *noncontiguous_
       else
         noncontiguous_hindexed_char += type_c->block_lengths[i]*smpi_datatype_get_extent(type_c->old_type);
     }
-    noncontiguous_hindexed=reinterpret_cast<void*>(noncontiguous_hindexed_char);
+    noncontiguous_hindexed=static_cast<void*>(noncontiguous_hindexed_char);
   }
 }
 
@@ -996,12 +996,12 @@ int smpi_datatype_hindexed(int count, int* blocklens, MPI_Aint* indices, MPI_Dat
  *      - block_length - the width or height of blocked matrix
  *      - count - the number of rows of matrix
  */
-void serialize_struct( const void *noncontiguous_struct, void *contiguous_struct, int count, void *type)
+void serialize_struct( void* noncontiguous_struct, void *contiguous_struct, int count, void *type)
 {
   s_smpi_mpi_struct_t* type_c = reinterpret_cast<s_smpi_mpi_struct_t*>(type);
   int i,j;
   char* contiguous_struct_char = static_cast<char*>(contiguous_struct);
-  const char* noncontiguous_struct_char = static_cast<const char*>(noncontiguous_struct)+ type_c->block_indices[0];
+  char* noncontiguous_struct_char = static_cast<char*>(noncontiguous_struct)+ type_c->block_indices[0];
   for(j=0; j<count;j++){
     for (i = 0; i < type_c->block_count; i++) {
       if (type_c->old_types[i]->sizeof_substruct == 0)
@@ -1016,11 +1016,11 @@ void serialize_struct( const void *noncontiguous_struct, void *contiguous_struct
 
       contiguous_struct_char += type_c->block_lengths[i]*smpi_datatype_size(type_c->old_types[i]);
       if (i<type_c->block_count-1)
-        noncontiguous_struct_char = static_cast<const char*>(noncontiguous_struct) + type_c->block_indices[i+1];
+        noncontiguous_struct_char = static_cast<char*>(noncontiguous_struct) + type_c->block_indices[i+1];
       else //let's hope this is MPI_UB ?
         noncontiguous_struct_char += type_c->block_lengths[i]*smpi_datatype_get_extent(type_c->old_types[i]);
     }
-    noncontiguous_struct=reinterpret_cast<const void*>(noncontiguous_struct_char);
+    noncontiguous_struct=static_cast<void*>(noncontiguous_struct_char);
   }
 }
 
@@ -1032,12 +1032,12 @@ void serialize_struct( const void *noncontiguous_struct, void *contiguous_struct
  *      - block_length - the width or height of blocked matrix
  *      - count - the number of rows of matrix
  */
-void unserialize_struct( const void *contiguous_struct, void *noncontiguous_struct, int count, void *type, MPI_Op op)
+void unserialize_struct( void* contiguous_struct, void *noncontiguous_struct, int count, void *type, MPI_Op op)
 {
   s_smpi_mpi_struct_t* type_c = reinterpret_cast<s_smpi_mpi_struct_t*>(type);
   int i,j;
 
-  const char* contiguous_struct_char = static_cast<const char*>(contiguous_struct);
+  char* contiguous_struct_char = static_cast<char*>(contiguous_struct);
   char* noncontiguous_struct_char = static_cast<char*>(noncontiguous_struct)+ type_c->block_indices[0];
   for(j=0; j<count;j++){
     for (i = 0; i < type_c->block_count; i++) {
@@ -1159,6 +1159,7 @@ void smpi_datatype_commit(MPI_Datatype *datatype)
 typedef struct s_smpi_mpi_op {
   MPI_User_function *func;
   bool is_commute;
+  bool is_fortran_op;
 } s_smpi_mpi_op_t;
 
 #define MAX_OP(a, b)  (b) = (a) < (b) ? (b) : (a)
@@ -1203,7 +1204,6 @@ APPLY_OP_LOOP(MPI_UNSIGNED, unsigned int,op)\
 APPLY_OP_LOOP(MPI_UNSIGNED_LONG, unsigned long,op)\
 APPLY_OP_LOOP(MPI_UNSIGNED_LONG_LONG, unsigned long long,op)\
 APPLY_OP_LOOP(MPI_WCHAR, wchar_t,op)\
-APPLY_OP_LOOP(MPI_C_BOOL, bool,op)\
 APPLY_OP_LOOP(MPI_BYTE, int8_t,op)\
 APPLY_OP_LOOP(MPI_INT8_T, int8_t,op)\
 APPLY_OP_LOOP(MPI_INT16_T, int16_t,op)\
@@ -1220,6 +1220,9 @@ APPLY_OP_LOOP(MPI_INTEGER2, int16_t,op)\
 APPLY_OP_LOOP(MPI_INTEGER4, int32_t,op)\
 APPLY_OP_LOOP(MPI_INTEGER8, int64_t,op)
 
+#define APPLY_BOOL_OP_LOOP(op)\
+APPLY_OP_LOOP(MPI_C_BOOL, bool,op)
+
 #define APPLY_FLOAT_OP_LOOP(op)\
 APPLY_OP_LOOP(MPI_FLOAT, float,op)\
 APPLY_OP_LOOP(MPI_DOUBLE, double,op)\
@@ -1284,36 +1287,42 @@ static void prod_func(void *a, void *b, int *length, MPI_Datatype * datatype)
 static void land_func(void *a, void *b, int *length, MPI_Datatype * datatype)
 {
   APPLY_BASIC_OP_LOOP(LAND_OP)
+  APPLY_BOOL_OP_LOOP(LAND_OP)
   APPLY_END_OP_LOOP(LAND_OP)
 }
 
 static void lor_func(void *a, void *b, int *length, MPI_Datatype * datatype)
 {
   APPLY_BASIC_OP_LOOP(LOR_OP)
+  APPLY_BOOL_OP_LOOP(LOR_OP)
   APPLY_END_OP_LOOP(LOR_OP)
 }
 
 static void lxor_func(void *a, void *b, int *length, MPI_Datatype * datatype)
 {
   APPLY_BASIC_OP_LOOP(LXOR_OP)
+  APPLY_BOOL_OP_LOOP(LXOR_OP)
   APPLY_END_OP_LOOP(LXOR_OP)
 }
 
 static void band_func(void *a, void *b, int *length, MPI_Datatype * datatype)
 {
   APPLY_BASIC_OP_LOOP(BAND_OP)
+  APPLY_BOOL_OP_LOOP(BAND_OP)
   APPLY_END_OP_LOOP(BAND_OP)
 }
 
 static void bor_func(void *a, void *b, int *length, MPI_Datatype * datatype)
 {
   APPLY_BASIC_OP_LOOP(BOR_OP)
+  APPLY_BOOL_OP_LOOP(BOR_OP)
   APPLY_END_OP_LOOP(BOR_OP)
 }
 
 static void bxor_func(void *a, void *b, int *length, MPI_Datatype * datatype)
 {
   APPLY_BASIC_OP_LOOP(BXOR_OP)
+  APPLY_BOOL_OP_LOOP(BXOR_OP)
   APPLY_END_OP_LOOP(BXOR_OP)
 }
 
@@ -1335,7 +1344,7 @@ static void replace_func(void *a, void *b, int *length, MPI_Datatype * datatype)
 }
 
 #define CREATE_MPI_OP(name, func)                             \
-  static s_smpi_mpi_op_t mpi_##name = { &(func) /* func */, true }; \
+  static s_smpi_mpi_op_t mpi_##name = { &(func) /* func */, true, false }; \
 MPI_Op name = &mpi_##name;
 
 CREATE_MPI_OP(MPI_MAX, max_func);
@@ -1358,6 +1367,7 @@ MPI_Op smpi_op_new(MPI_User_function * function, bool commute)
   op = xbt_new(s_smpi_mpi_op_t, 1);
   op->func = function;
   op-> is_commute = commute;
+  op-> is_fortran_op = false;
   return op;
 }
 
@@ -1371,7 +1381,13 @@ void smpi_op_destroy(MPI_Op op)
   xbt_free(op);
 }
 
-void smpi_op_apply(MPI_Op op, const void *invec, void *inoutvec, int *len, MPI_Datatype * datatype)
+void smpi_op_set_fortran(MPI_Op op)
+{
+  //tell that we were created from fortran, so we need to translate the type to fortran when called
+  op->is_fortran_op = true;
+}
+
+void smpi_op_apply(MPI_Op op, void *invec, void *inoutvec, int *len, MPI_Datatype * datatype)
 {
   if(op==MPI_OP_NULL)
     return;
@@ -1381,8 +1397,16 @@ void smpi_op_apply(MPI_Op op, const void *invec, void *inoutvec, int *len, MPI_D
     smpi_switch_data_segment(smpi_process_index());
   }
 
-  if(!smpi_process_get_replaying())
-    op->func(const_cast<void*>(invec), inoutvec, len, datatype);
+  if(!smpi_process_get_replaying()){
+    if(! op->is_fortran_op)
+      op->func(invec, inoutvec, len, datatype);
+    else{
+      int tmp = smpi_type_c2f(*datatype);
+      /* Unfortunately, the C and Fortran version of the MPI standard do not agree on the type here,
+         thus the reinterpret_cast. */
+      op->func(invec, inoutvec, len, reinterpret_cast<MPI_Datatype*>(&tmp) );
+    }
+  }
 }
 
 int smpi_type_attr_delete(MPI_Datatype type, int keyval){