X-Git-Url: http://info.iut-bm.univ-fcomte.fr/pub/gitweb/simgrid.git/blobdiff_plain/150d5c889f5f071ebe2ddcc8c4fcb1335e10707c..4ee149975b19d9143b910c54a0916002451598db:/src/smpi/smpi_mpi_dt.c diff --git a/src/smpi/smpi_mpi_dt.c b/src/smpi/smpi_mpi_dt.c index a266ec9135..d5a1421b7b 100644 --- a/src/smpi/smpi_mpi_dt.c +++ b/src/smpi/smpi_mpi_dt.c @@ -28,6 +28,16 @@ XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_mpi_dt, smpi, }; \ MPI_Datatype name = &mpi_##name; +#define CREATE_MPI_DATATYPE_NULL(name) \ + static s_smpi_mpi_datatype_t mpi_##name = { \ + 0, /* size */ \ + 0, /*was 1 has_subtype*/ \ + 0, /* lb */ \ + 0, /* ub = lb + size */ \ + DT_FLAG_BASIC, /* flags */ \ + NULL /* pointer on extended struct*/ \ + }; \ +MPI_Datatype name = &mpi_##name; //The following are datatypes for the MPI functions MPI_MAXLOC and MPI_MINLOC. typedef struct { @@ -93,6 +103,8 @@ CREATE_MPI_DATATYPE(MPI_SHORT_INT, short_int); CREATE_MPI_DATATYPE(MPI_2INT, int_int); CREATE_MPI_DATATYPE(MPI_LONG_DOUBLE_INT, long_double_int); +CREATE_MPI_DATATYPE_NULL(MPI_UB); +CREATE_MPI_DATATYPE_NULL(MPI_LB); // Internal use only CREATE_MPI_DATATYPE(MPI_PTR, void*); @@ -201,6 +213,7 @@ void serialize_vector( const void *noncontiguous_vector, noncontiguous_vector_char += type_c->block_stride*type_c->size_oldtype; } } + /* * Copies contiguous data into noncontiguous memory. * @param noncontiguous_vector - output vector @@ -232,7 +245,7 @@ void unserialize_vector( const void *contiguous_vector, /* * Create a Sub type vector to be able to serialize and unserialize it - * the structre s_smpi_mpi_vector_t is derived from s_smpi_subtype which + * the structure s_smpi_mpi_vector_t is derived from s_smpi_subtype which * required the functions unserialize and serialize * */ @@ -255,12 +268,12 @@ s_smpi_mpi_vector_t* smpi_datatype_vector_create( int block_stride, void smpi_datatype_create(MPI_Datatype* new_type, int size, int has_subtype, void *struct_type, int flags){ MPI_Datatype new_t= xbt_new(s_smpi_mpi_datatype_t,1); - new_t->size=size; - new_t->has_subtype=has_subtype; - new_t->lb=0; - new_t->ub=size; - new_t->flags=flags; - new_t->substruct=struct_type; + new_t->size = size; + new_t->has_subtype = has_subtype; + new_t->lb = 0; + new_t->ub = size; + new_t->flags = flags; + new_t->substruct = struct_type; *new_type = new_t; } @@ -275,7 +288,7 @@ int smpi_datatype_contiguous(int count, MPI_Datatype old_type, MPI_Datatype* new retval = MPI_ERR_TYPE; } else { smpi_datatype_create(new_type, count * - smpi_datatype_size(old_type),1,NULL, DT_FLAG_CONTIGUOUS); + smpi_datatype_size(old_type),0,NULL, DT_FLAG_CONTIGUOUS); retval=MPI_SUCCESS; } return retval; @@ -289,6 +302,8 @@ int smpi_datatype_vector(int count, int blocklen, int stride, MPI_Datatype old_t retval = MPI_ERR_TYPE; } else { if(stride != blocklen){ +if (old_type->has_subtype == 1) + XBT_WARN("vector contains a complex type - not yet handled"); s_smpi_mpi_vector_t* subtype = smpi_datatype_vector_create( stride, blocklen, count, @@ -308,7 +323,7 @@ int smpi_datatype_vector(int count, int blocklen, int stride, MPI_Datatype old_t smpi_datatype_size(old_type), 0, NULL, - DT_FLAG_VECTOR); + DT_FLAG_VECTOR|DT_FLAG_CONTIGUOUS); retval=MPI_SUCCESS; } } @@ -380,7 +395,7 @@ void unserialize_hvector( const void *contiguous_vector, /* * Create a Sub type vector to be able to serialize and unserialize it - * the structre s_smpi_mpi_vector_t is derived from s_smpi_subtype which + * the structure s_smpi_mpi_vector_t is derived from s_smpi_subtype which * required the functions unserialize and serialize * */ @@ -407,6 +422,8 @@ int smpi_datatype_hvector(int count, int blocklen, MPI_Aint stride, MPI_Datatype if ((old_type->flags & DT_FLAG_COMMITED) != DT_FLAG_COMMITED) { retval = MPI_ERR_TYPE; } else { +if (old_type->has_subtype == 1) + XBT_WARN("hvector contains a complex type - not yet handled"); if(stride != blocklen*smpi_datatype_size(old_type)){ s_smpi_mpi_hvector_t* subtype = smpi_datatype_hvector_create( stride, blocklen, @@ -425,7 +442,7 @@ int smpi_datatype_hvector(int count, int blocklen, MPI_Aint stride, MPI_Datatype smpi_datatype_size(old_type), 0, NULL, - DT_FLAG_VECTOR); + DT_FLAG_VECTOR|DT_FLAG_CONTIGUOUS); retval=MPI_SUCCESS; } } @@ -442,8 +459,8 @@ Indexed Implementation * @param contiguous_indexed - output indexed * @param noncontiguous_indexed - input indexed * @param type - pointer contening : - * - stride - stride of between noncontiguous data - * - block_length - the width or height of blocked matrix + * - block_lengths - the width or height of blocked matrix + * - block_indices - indices of each data, in element * - count - the number of rows of matrix */ void serialize_indexed( const void *noncontiguous_indexed, @@ -452,16 +469,19 @@ void serialize_indexed( const void *noncontiguous_indexed, void *type) { s_smpi_mpi_indexed_t* type_c = (s_smpi_mpi_indexed_t*)type; - int i; + int i,j; char* contiguous_indexed_char = (char*)contiguous_indexed; char* noncontiguous_indexed_char = (char*)noncontiguous_indexed; - - for (i = 0; i < type_c->block_count * count; i++) { - memcpy(contiguous_indexed_char, - noncontiguous_indexed_char, type_c->block_lengths[i] * type_c->size_oldtype); - - contiguous_indexed_char += type_c->block_lengths[i]*type_c->size_oldtype; - noncontiguous_indexed_char = (char*)noncontiguous_indexed + type_c->block_indices[i+1]*type_c->size_oldtype; + for(j=0; jblock_count; i++) { + memcpy(contiguous_indexed_char, + noncontiguous_indexed_char, type_c->block_lengths[i] * type_c->size_oldtype); + + contiguous_indexed_char += type_c->block_lengths[i]*type_c->size_oldtype; + if (iblock_count-1)noncontiguous_indexed_char = (char*)noncontiguous_indexed + type_c->block_indices[i+1]*type_c->size_oldtype; + else noncontiguous_indexed_char += type_c->block_lengths[i]*type_c->size_oldtype; + } + noncontiguous_indexed=(void*)noncontiguous_indexed_char; } } /* @@ -469,8 +489,8 @@ void serialize_indexed( const void *noncontiguous_indexed, * @param noncontiguous_indexed - output indexed * @param contiguous_indexed - input indexed * @param type - pointer contening : - * - stride - stride of between noncontiguous data - * - block_length - the width or height of blocked matrix + * - block_lengths - the width or height of blocked matrix + * - block_indices - indices of each data, in element * - count - the number of rows of matrix */ void unserialize_indexed( const void *contiguous_indexed, @@ -479,23 +499,26 @@ void unserialize_indexed( const void *contiguous_indexed, void *type) { s_smpi_mpi_indexed_t* type_c = (s_smpi_mpi_indexed_t*)type; - int i; + int i,j; char* contiguous_indexed_char = (char*)contiguous_indexed; char* noncontiguous_indexed_char = (char*)noncontiguous_indexed; - - for (i = 0; i < type_c->block_count * count; i++) { - memcpy(noncontiguous_indexed_char, - contiguous_indexed_char, type_c->block_lengths[i] * type_c->size_oldtype); - - contiguous_indexed_char += type_c->block_lengths[i]*type_c->size_oldtype; - noncontiguous_indexed_char = (char*)noncontiguous_indexed + type_c->block_indices[i+1]*type_c->size_oldtype; + for(j=0; jblock_count; i++) { + memcpy(noncontiguous_indexed_char, + contiguous_indexed_char, type_c->block_lengths[i] * type_c->size_oldtype); + + contiguous_indexed_char += type_c->block_lengths[i]*type_c->size_oldtype; + if (iblock_count-1)noncontiguous_indexed_char = (char*)noncontiguous_indexed + type_c->block_indices[i+1]*type_c->size_oldtype; + else noncontiguous_indexed_char += type_c->block_lengths[i]*type_c->size_oldtype; + } + noncontiguous_indexed=(void*)noncontiguous_indexed_char; } } /* * Create a Sub type indexed to be able to serialize and unserialize it - * the structre s_smpi_mpi_indexed_t is derived from s_smpi_subtype which + * the structure s_smpi_mpi_indexed_t is derived from s_smpi_subtype which * required the functions unserialize and serialize */ s_smpi_mpi_indexed_t* smpi_datatype_indexed_create( int* block_lengths, @@ -506,9 +529,14 @@ s_smpi_mpi_indexed_t* smpi_datatype_indexed_create( int* block_lengths, s_smpi_mpi_indexed_t *new_t= xbt_new(s_smpi_mpi_indexed_t,1); new_t->base.serialize = &serialize_indexed; new_t->base.unserialize = &unserialize_indexed; - //FIXME : copy those or assume they won't be freed ? - new_t->block_lengths = block_lengths; - new_t->block_indices = block_indices; + //TODO : add a custom function for each time to clean these + new_t->block_lengths= xbt_new(int, block_count); + new_t->block_indices= xbt_new(int, block_count); + int i; + for(i=0;iblock_lengths[i]=block_lengths[i]; + new_t->block_indices[i]=block_indices[i]; + } new_t->block_count = block_count; new_t->old_type = old_type; new_t->size_oldtype = size_oldtype; @@ -521,22 +549,34 @@ int smpi_datatype_indexed(int count, int* blocklens, int* indices, MPI_Datatype int i; int retval; int size = 0; + int contiguous=1; for(i=0; i< count; i++){ if (blocklens[i]<=0) return MPI_ERR_ARG; size += blocklens[i]; + + if ( (i< count -1) && (indices[i]+blocklens[i] != indices[i+1]) )contiguous=0; } if ((old_type->flags & DT_FLAG_COMMITED) != DT_FLAG_COMMITED) { retval = MPI_ERR_TYPE; } else { - s_smpi_mpi_indexed_t* subtype = smpi_datatype_indexed_create( blocklens, - indices, - count, - old_type, - smpi_datatype_size(old_type)); - smpi_datatype_create(new_type, size * - smpi_datatype_size(old_type),1, subtype, DT_FLAG_DATA); + if (old_type->has_subtype == 1) + XBT_WARN("indexed contains a complex type - not yet handled"); + + if(!contiguous){ + s_smpi_mpi_indexed_t* subtype = smpi_datatype_indexed_create( blocklens, + indices, + count, + old_type, + smpi_datatype_size(old_type)); + + smpi_datatype_create(new_type, size * + smpi_datatype_size(old_type),1, subtype, DT_FLAG_DATA); +}else{ + smpi_datatype_create(new_type, size * + smpi_datatype_size(old_type),0, NULL, DT_FLAG_DATA|DT_FLAG_CONTIGUOUS); +} retval=MPI_SUCCESS; } return retval; @@ -552,8 +592,8 @@ Hindexed Implementation - Indexed with indices in bytes * @param contiguous_hindexed - output hindexed * @param noncontiguous_hindexed - input hindexed * @param type - pointer contening : - * - stride - stride of between noncontiguous data - * - block_length - the width or height of blocked matrix + * - block_lengths - the width or height of blocked matrix + * - block_indices - indices of each data, in bytes * - count - the number of rows of matrix */ void serialize_hindexed( const void *noncontiguous_hindexed, @@ -562,16 +602,19 @@ void serialize_hindexed( const void *noncontiguous_hindexed, void *type) { s_smpi_mpi_hindexed_t* type_c = (s_smpi_mpi_hindexed_t*)type; - int i; + int i,j; char* contiguous_hindexed_char = (char*)contiguous_hindexed; char* noncontiguous_hindexed_char = (char*)noncontiguous_hindexed; - - for (i = 0; i < type_c->block_count * count; i++) { - memcpy(contiguous_hindexed_char, - noncontiguous_hindexed_char, type_c->block_lengths[i] * type_c->size_oldtype); - - contiguous_hindexed_char += type_c->block_lengths[i]*type_c->size_oldtype; - noncontiguous_hindexed_char = (char*)noncontiguous_hindexed + type_c->block_indices[i+1]; + for(j=0; jblock_count; i++) { + memcpy(contiguous_hindexed_char, + noncontiguous_hindexed_char, type_c->block_lengths[i] * type_c->size_oldtype); + + contiguous_hindexed_char += type_c->block_lengths[i]*type_c->size_oldtype; + if (iblock_count-1)noncontiguous_hindexed_char = (char*)noncontiguous_hindexed + type_c->block_indices[i+1]; + else noncontiguous_hindexed_char += type_c->block_lengths[i]*type_c->size_oldtype; + } + noncontiguous_hindexed=(void*)noncontiguous_hindexed_char; } } /* @@ -579,8 +622,8 @@ void serialize_hindexed( const void *noncontiguous_hindexed, * @param noncontiguous_hindexed - output hindexed * @param contiguous_hindexed - input hindexed * @param type - pointer contening : - * - stride - stride of between noncontiguous data - * - block_length - the width or height of blocked matrix + * - block_lengths - the width or height of blocked matrix + * - block_indices - indices of each data, in bytes * - count - the number of rows of matrix */ void unserialize_hindexed( const void *contiguous_hindexed, @@ -589,23 +632,26 @@ void unserialize_hindexed( const void *contiguous_hindexed, void *type) { s_smpi_mpi_hindexed_t* type_c = (s_smpi_mpi_hindexed_t*)type; - int i; + int i,j; char* contiguous_hindexed_char = (char*)contiguous_hindexed; char* noncontiguous_hindexed_char = (char*)noncontiguous_hindexed; - - for (i = 0; i < type_c->block_count * count; i++) { - memcpy(noncontiguous_hindexed_char, - contiguous_hindexed_char, type_c->block_lengths[i] * type_c->size_oldtype); - - contiguous_hindexed_char += type_c->block_lengths[i]*type_c->size_oldtype; - noncontiguous_hindexed_char = (char*)noncontiguous_hindexed + type_c->block_indices[i+1]; + for(j=0; jblock_count; i++) { + memcpy(noncontiguous_hindexed_char, + contiguous_hindexed_char, type_c->block_lengths[i] * type_c->size_oldtype); + + contiguous_hindexed_char += type_c->block_lengths[i]*type_c->size_oldtype; + if (iblock_count-1)noncontiguous_hindexed_char = (char*)noncontiguous_hindexed + type_c->block_indices[i+1]; + else noncontiguous_hindexed_char += type_c->block_lengths[i]*type_c->size_oldtype; + } + noncontiguous_hindexed=(void*)noncontiguous_hindexed_char; } } /* * Create a Sub type hindexed to be able to serialize and unserialize it - * the structre s_smpi_mpi_hindexed_t is derived from s_smpi_subtype which + * the structure s_smpi_mpi_hindexed_t is derived from s_smpi_subtype which * required the functions unserialize and serialize */ s_smpi_mpi_hindexed_t* smpi_datatype_hindexed_create( int* block_lengths, @@ -616,9 +662,14 @@ s_smpi_mpi_hindexed_t* smpi_datatype_hindexed_create( int* block_lengths, s_smpi_mpi_hindexed_t *new_t= xbt_new(s_smpi_mpi_hindexed_t,1); new_t->base.serialize = &serialize_hindexed; new_t->base.unserialize = &unserialize_hindexed; - //FIXME : copy those or assume they won't be freed ? - new_t->block_lengths = block_lengths; - new_t->block_indices = block_indices; + //TODO : add a custom function for each time to clean these + new_t->block_lengths= xbt_new(int, block_count); + new_t->block_indices= xbt_new(MPI_Aint, block_count); + int i; + for(i=0;iblock_lengths[i]=block_lengths[i]; + new_t->block_indices[i]=block_indices[i]; + } new_t->block_count = block_count; new_t->old_type = old_type; new_t->size_oldtype = size_oldtype; @@ -631,22 +682,34 @@ int smpi_datatype_hindexed(int count, int* blocklens, MPI_Aint* indices, MPI_Dat int i; int retval; int size = 0; + int contiguous=1; for(i=0; i< count; i++){ if (blocklens[i]<=0) return MPI_ERR_ARG; size += blocklens[i]; + + + if ( (i< count -1) && (indices[i]+blocklens[i]*smpi_datatype_size(old_type) != indices[i+1]) )contiguous=0; } if ((old_type->flags & DT_FLAG_COMMITED) != DT_FLAG_COMMITED) { retval = MPI_ERR_TYPE; } else { - s_smpi_mpi_hindexed_t* subtype = smpi_datatype_hindexed_create( blocklens, - indices, - count, - old_type, - smpi_datatype_size(old_type)); + if (old_type->has_subtype == 1) + XBT_WARN("hindexed contains a complex type - not yet handled"); + + if(!contiguous){ + s_smpi_mpi_hindexed_t* subtype = smpi_datatype_hindexed_create( blocklens, + indices, + count, + old_type, + smpi_datatype_size(old_type)); - smpi_datatype_create(new_type, size * - smpi_datatype_size(old_type),1, subtype, DT_FLAG_DATA); + smpi_datatype_create(new_type, size * + smpi_datatype_size(old_type),1, subtype, DT_FLAG_DATA); + }else{ + smpi_datatype_create(new_type, size * + smpi_datatype_size(old_type),0, NULL, DT_FLAG_DATA|DT_FLAG_CONTIGUOUS); + } retval=MPI_SUCCESS; } return retval; @@ -672,15 +735,18 @@ void serialize_struct( const void *noncontiguous_struct, void *type) { s_smpi_mpi_struct_t* type_c = (s_smpi_mpi_struct_t*)type; - int i; + int i,j; char* contiguous_struct_char = (char*)contiguous_struct; char* noncontiguous_struct_char = (char*)noncontiguous_struct; - - for (i = 0; i < type_c->block_count * count; i++) { - memcpy(contiguous_struct_char, - noncontiguous_struct_char, type_c->block_lengths[i] * smpi_datatype_size(type_c->old_types[i])); - contiguous_struct_char += type_c->block_lengths[i]*smpi_datatype_size(type_c->old_types[i]); - noncontiguous_struct_char = (char*)noncontiguous_struct + type_c->block_indices[i+1]; + for(j=0; jblock_count; i++) { + memcpy(contiguous_struct_char, + noncontiguous_struct_char, type_c->block_lengths[i] * smpi_datatype_size(type_c->old_types[i])); + contiguous_struct_char += type_c->block_lengths[i]*smpi_datatype_size(type_c->old_types[i]); + if (iblock_count-1)noncontiguous_struct_char = (char*)noncontiguous_struct + type_c->block_indices[i+1]; + else noncontiguous_struct_char += type_c->block_lengths[i]*smpi_datatype_size(type_c->old_types[i]);//let's hope this is MPI_UB ? + } + noncontiguous_struct=(void*)noncontiguous_struct_char; } } /* @@ -698,22 +764,26 @@ void unserialize_struct( const void *contiguous_struct, void *type) { s_smpi_mpi_struct_t* type_c = (s_smpi_mpi_struct_t*)type; - int i; + int i,j; char* contiguous_struct_char = (char*)contiguous_struct; char* noncontiguous_struct_char = (char*)noncontiguous_struct; - - for (i = 0; i < type_c->block_count * count; i++) { - memcpy(noncontiguous_struct_char, - contiguous_struct_char, type_c->block_lengths[i] * smpi_datatype_size(type_c->old_types[i])); - contiguous_struct_char += type_c->block_lengths[i]*smpi_datatype_size(type_c->old_types[i]); - noncontiguous_struct_char = (char*)noncontiguous_struct + type_c->block_indices[i+1]; + for(j=0; jblock_count; i++) { + memcpy(noncontiguous_struct_char, + contiguous_struct_char, type_c->block_lengths[i] * smpi_datatype_size(type_c->old_types[i])); + contiguous_struct_char += type_c->block_lengths[i]*smpi_datatype_size(type_c->old_types[i]); + if (iblock_count-1)noncontiguous_struct_char = (char*)noncontiguous_struct + type_c->block_indices[i+1]; + else noncontiguous_struct_char += type_c->block_lengths[i]*smpi_datatype_size(type_c->old_types[i]); + } + noncontiguous_struct=(void*)noncontiguous_struct_char; + } } /* * Create a Sub type struct to be able to serialize and unserialize it - * the structre s_smpi_mpi_struct_t is derived from s_smpi_subtype which + * the structure s_smpi_mpi_struct_t is derived from s_smpi_subtype which * required the functions unserialize and serialize */ s_smpi_mpi_struct_t* smpi_datatype_struct_create( int* block_lengths, @@ -723,11 +793,20 @@ s_smpi_mpi_struct_t* smpi_datatype_struct_create( int* block_lengths, s_smpi_mpi_struct_t *new_t= xbt_new(s_smpi_mpi_struct_t,1); new_t->base.serialize = &serialize_struct; new_t->base.unserialize = &unserialize_struct; - //FIXME : copy those or assume they won't be freed ? - new_t->block_lengths = block_lengths; - new_t->block_indices = block_indices; + //TODO : add a custom function for each time to clean these + new_t->block_lengths= xbt_new(int, block_count); + new_t->block_indices= xbt_new(MPI_Aint, block_count); + new_t->old_types= xbt_new(MPI_Datatype, block_count); + int i; + for(i=0;iblock_lengths[i]=block_lengths[i]; + new_t->block_indices[i]=block_indices[i]; + new_t->old_types[i]=old_types[i]; + } + //new_t->block_lengths = block_lengths; + //new_t->block_indices = block_indices; new_t->block_count = block_count; - new_t->old_types = old_types; + //new_t->old_types = old_types; return new_t; } @@ -735,25 +814,31 @@ s_smpi_mpi_struct_t* smpi_datatype_struct_create( int* block_lengths, int smpi_datatype_struct(int count, int* blocklens, MPI_Aint* indices, MPI_Datatype* old_types, MPI_Datatype* new_type) { int i; - size_t size; //Khalid added this - + size_t size = 0; + int contiguous=1; size = 0; for(i=0; i< count; i++){ if (blocklens[i]<=0) return MPI_ERR_ARG; if ((old_types[i]->flags & DT_FLAG_COMMITED) != DT_FLAG_COMMITED) return MPI_ERR_TYPE; + if (old_types[i]->has_subtype == 1) + XBT_WARN("Struct contains a complex type - not yet handled"); size += blocklens[i]*smpi_datatype_size(old_types[i]); - } + if ( (i< count -1) && (indices[i]+blocklens[i]*smpi_datatype_size(old_types[i]) != indices[i+1]) )contiguous=0; + } - s_smpi_mpi_struct_t* subtype = smpi_datatype_struct_create( blocklens, + if(!contiguous){ + s_smpi_mpi_struct_t* subtype = smpi_datatype_struct_create( blocklens, indices, count, old_types); - smpi_datatype_create(new_type, size ,1, subtype, DT_FLAG_DATA); - + smpi_datatype_create(new_type, size ,1, subtype, DT_FLAG_DATA); + }else{ + smpi_datatype_create(new_type, size,0, NULL, DT_FLAG_DATA|DT_FLAG_CONTIGUOUS); + } return MPI_SUCCESS; }