X-Git-Url: http://info.iut-bm.univ-fcomte.fr/pub/gitweb/simgrid.git/blobdiff_plain/cf108868b4eeed4d0d9d343bc68557d7814e18c0..e05ceda6c38a3d48b6a6bb9820101803b002e16e:/src/smpi/smpi_mpi_dt.cpp diff --git a/src/smpi/smpi_mpi_dt.cpp b/src/smpi/smpi_mpi_dt.cpp index bee1d73bc4..36eba9bd54 100644 --- a/src/smpi/smpi_mpi_dt.cpp +++ b/src/smpi/smpi_mpi_dt.cpp @@ -17,8 +17,7 @@ #include "xbt/replay.h" #include "simgrid/modelchecker.h" -XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_mpi_dt, smpi, - "Logging specific to SMPI (datatype)"); +XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_mpi_dt, smpi, "Logging specific to SMPI (datatype)"); xbt_dict_t smpi_type_keyvals = NULL; int type_keyval_id=0;//avoid collisions @@ -27,7 +26,7 @@ int type_keyval_id=0;//avoid collisions static s_smpi_mpi_datatype_t mpi_##name = { \ (char*) # name, \ sizeof(type), /* size */ \ - 0, /*was 1 has_subtype*/ \ + 0, /*was 1 sizeof_substruct*/ \ 0, /* lb */ \ sizeof(type), /* ub = lb + size */ \ DT_FLAG_BASIC, /* flags */ \ @@ -41,7 +40,7 @@ const MPI_Datatype name = &mpi_##name; static s_smpi_mpi_datatype_t mpi_##name = { \ (char*) # name, \ 0, /* size */ \ - 0, /* was 1 has_subtype*/ \ + 0, /* was 1 sizeof_substruct*/ \ 0, /* lb */ \ 0, /* ub = lb + size */ \ DT_FLAG_BASIC, /* flags */ \ @@ -154,11 +153,9 @@ CREATE_MPI_DATATYPE(MPI_PACKED, char); // Internal use only CREATE_MPI_DATATYPE(MPI_PTR, void*); -/** Check if the datatype is usable for communications - */ +/** Check if the datatype is usable for communications */ int is_datatype_valid(MPI_Datatype datatype) { - return datatype != MPI_DATATYPE_NULL - && (datatype->flags & DT_FLAG_COMMITED); + return datatype != MPI_DATATYPE_NULL && (datatype->flags & DT_FLAG_COMMITED); } size_t smpi_datatype_size(MPI_Datatype datatype) @@ -181,10 +178,11 @@ int smpi_datatype_dup(MPI_Datatype datatype, MPI_Datatype* new_t) int ret=MPI_SUCCESS; *new_t= xbt_new(s_smpi_mpi_datatype_t,1); memcpy(*new_t, datatype, sizeof(s_smpi_mpi_datatype_t)); - if (datatype->has_subtype){ - //FIXME: may copy too much information. - (*new_t)->substruct=xbt_new(s_smpi_mpi_struct_t,1); - memcpy((*new_t)->substruct, datatype->substruct, sizeof(s_smpi_mpi_struct_t)); + (*new_t)->in_use=1; + (*new_t)->flags &= ~DT_FLAG_PREDEFINED; + if (datatype->sizeof_substruct){ + (*new_t)->substruct=xbt_malloc(datatype->sizeof_substruct); + memcpy((*new_t)->substruct, datatype->substruct, datatype->sizeof_substruct); } if(datatype->name) (*new_t)->name = xbt_strdup(datatype->name); @@ -196,10 +194,12 @@ int smpi_datatype_dup(MPI_Datatype datatype, MPI_Datatype* new_t) void* value_in; void* value_out; xbt_dict_foreach(datatype->attributes, cursor, key, value_in){ - smpi_type_key_elem elem = static_cast(xbt_dict_get_or_null_ext(smpi_type_keyvals, (const char*)key, sizeof(int))); + smpi_type_key_elem elem = + static_cast(xbt_dict_get_or_null_ext(smpi_type_keyvals, (const char*)key, sizeof(int))); if(elem && elem->copy_fn!=MPI_NULL_COPY_FN){ ret = elem->copy_fn(datatype, *key, NULL, value_in, &value_out, &flag ); if(ret!=MPI_SUCCESS){ + smpi_datatype_unuse(*new_t); *new_t=MPI_DATATYPE_NULL; return ret; } @@ -211,8 +211,7 @@ int smpi_datatype_dup(MPI_Datatype datatype, MPI_Datatype* new_t) return ret; } -int smpi_datatype_extent(MPI_Datatype datatype, MPI_Aint * lb, - MPI_Aint * extent) +int smpi_datatype_extent(MPI_Datatype datatype, MPI_Aint * lb, MPI_Aint * extent) { if(datatype == MPI_DATATYPE_NULL){ *lb=0; @@ -237,6 +236,8 @@ void smpi_datatype_get_name(MPI_Datatype datatype, char* name, int* length){ } void smpi_datatype_set_name(MPI_Datatype datatype, char* name){ + if(datatype->name!=NULL && !(datatype->flags & DT_FLAG_PREDEFINED)) + xbt_free(datatype->name); datatype->name = xbt_strdup(name);; } @@ -254,22 +255,21 @@ int smpi_datatype_copy(void *sendbuf, int sendcount, MPI_Datatype sendtype, recvcount *= smpi_datatype_size(recvtype); count = sendcount < recvcount ? sendcount : recvcount; - if(sendtype->has_subtype == 0 && recvtype->has_subtype == 0) { + if(sendtype->sizeof_substruct == 0 && recvtype->sizeof_substruct == 0) { if(!smpi_process_get_replaying()) memcpy(recvbuf, sendbuf, count); } - else if (sendtype->has_subtype == 0) + else if (sendtype->sizeof_substruct == 0) { s_smpi_subtype_t *subtype = static_cast(recvtype->substruct); subtype->unserialize( sendbuf, recvbuf, recvcount/smpi_datatype_size(recvtype), subtype, MPI_REPLACE); } - else if (recvtype->has_subtype == 0) + else if (recvtype->sizeof_substruct == 0) { s_smpi_subtype_t *subtype = static_cast(sendtype->substruct); subtype->serialize(sendbuf, recvbuf, sendcount/smpi_datatype_size(sendtype), subtype); }else{ s_smpi_subtype_t *subtype = static_cast(sendtype->substruct); - void * buf_tmp = xbt_malloc(count); subtype->serialize( sendbuf, buf_tmp,count/smpi_datatype_size(sendtype), subtype); @@ -292,10 +292,7 @@ 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( const void *noncontiguous_vector, void *contiguous_vector, int count, void *type) { s_smpi_mpi_vector_t* type_c = (s_smpi_mpi_vector_t*)type; int i; @@ -303,20 +300,18 @@ void serialize_vector( const void *noncontiguous_vector, char* noncontiguous_vector_char = (char*)noncontiguous_vector; for (i = 0; i < type_c->block_count * count; i++) { - if (type_c->old_type->has_subtype == 0) - memcpy(contiguous_vector_char, - noncontiguous_vector_char, type_c->block_length * type_c->size_oldtype); + if (type_c->old_type->sizeof_substruct == 0) + memcpy(contiguous_vector_char, noncontiguous_vector_char, type_c->block_length * type_c->size_oldtype); else ((s_smpi_subtype_t*)type_c->old_type->substruct)->serialize( noncontiguous_vector_char, contiguous_vector_char, - type_c->block_length, - type_c->old_type->substruct); + type_c->block_length, type_c->old_type->substruct); contiguous_vector_char += type_c->block_length*type_c->size_oldtype; if((i+1)%type_c->block_count ==0) - noncontiguous_vector_char += type_c->block_length*smpi_datatype_get_extent(type_c->old_type); + noncontiguous_vector_char += type_c->block_length*smpi_datatype_get_extent(type_c->old_type); else - noncontiguous_vector_char += type_c->block_stride*smpi_datatype_get_extent(type_c->old_type); + noncontiguous_vector_char += type_c->block_stride*smpi_datatype_get_extent(type_c->old_type); } } @@ -329,11 +324,7 @@ void serialize_vector( const void *noncontiguous_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( const void *contiguous_vector, void *noncontiguous_vector, int count, void *type, MPI_Op op) { s_smpi_mpi_vector_t* type_c = (s_smpi_mpi_vector_t*)type; int i; @@ -342,40 +333,32 @@ void unserialize_vector( const void *contiguous_vector, char* noncontiguous_vector_char = (char*)noncontiguous_vector; for (i = 0; i < type_c->block_count * count; i++) { - if (type_c->old_type->has_subtype == 0) + if (type_c->old_type->sizeof_substruct == 0) smpi_op_apply(op, contiguous_vector_char, noncontiguous_vector_char, &type_c->block_length, &type_c->old_type); /* memcpy(noncontiguous_vector_char, contiguous_vector_char, type_c->block_length * type_c->size_oldtype);*/ else - ((s_smpi_subtype_t*)type_c->old_type->substruct)->unserialize( contiguous_vector_char, - noncontiguous_vector_char, - type_c->block_length, - type_c->old_type->substruct, - op); + ((s_smpi_subtype_t*)type_c->old_type->substruct)->unserialize(contiguous_vector_char, noncontiguous_vector_char, + type_c->block_length,type_c->old_type->substruct, + op); contiguous_vector_char += type_c->block_length*type_c->size_oldtype; if((i+1)%type_c->block_count ==0) - noncontiguous_vector_char += type_c->block_length*smpi_datatype_get_extent(type_c->old_type); + noncontiguous_vector_char += type_c->block_length*smpi_datatype_get_extent(type_c->old_type); else - noncontiguous_vector_char += type_c->block_stride*smpi_datatype_get_extent(type_c->old_type); + noncontiguous_vector_char += type_c->block_stride*smpi_datatype_get_extent(type_c->old_type); } } -/* - * Create a Sub type vector to be able to serialize and unserialize it - * the structure s_smpi_mpi_vector_t is derived from s_smpi_subtype which - * required the functions unserialize and serialize - * - */ -s_smpi_mpi_vector_t* smpi_datatype_vector_create( int block_stride, - int block_length, - int block_count, - MPI_Datatype old_type, - int size_oldtype){ +/* Create a Sub type vector to be able to serialize and unserialize it the structure s_smpi_mpi_vector_t is derived + * from s_smpi_subtype which required the functions unserialize and serialize */ +s_smpi_mpi_vector_t* smpi_datatype_vector_create( int block_stride, int block_length, int block_count, + MPI_Datatype old_type, int size_oldtype){ s_smpi_mpi_vector_t *new_t= xbt_new(s_smpi_mpi_vector_t,1); new_t->base.serialize = &serialize_vector; new_t->base.unserialize = &unserialize_vector; new_t->base.subtype_free = &free_vector; + new_t->base.subtype_use = &use_vector; new_t->block_stride = block_stride; new_t->block_length = block_length; new_t->block_count = block_count; @@ -385,21 +368,21 @@ s_smpi_mpi_vector_t* smpi_datatype_vector_create( int block_stride, return new_t; } -void smpi_datatype_create(MPI_Datatype* new_type, int size,int lb, int ub, int has_subtype, - void *struct_type, int flags){ +void smpi_datatype_create(MPI_Datatype* new_type, int size,int lb, int ub, int sizeof_substruct, void *struct_type, + int flags){ MPI_Datatype new_t= xbt_new(s_smpi_mpi_datatype_t,1); new_t->name = NULL; new_t->size = size; - new_t->has_subtype = size>0? has_subtype:0; + new_t->sizeof_substruct = size>0? sizeof_substruct:0; new_t->lb = lb; new_t->ub = ub; new_t->flags = flags; new_t->substruct = struct_type; - new_t->in_use=0; + new_t->in_use=1; new_t->attributes=NULL; *new_type = new_t; -#ifdef HAVE_MC +#if HAVE_MC if(MC_is_active()) MC_ignore(&(new_t->in_use), sizeof(new_t->in_use)); #endif @@ -407,17 +390,6 @@ void smpi_datatype_create(MPI_Datatype* new_type, int size,int lb, int ub, int h void smpi_datatype_free(MPI_Datatype* type){ xbt_assert((*type)->in_use >= 0); - if((*type)->attributes !=NULL){ - xbt_dict_cursor_t cursor = NULL; - int* key; - void * value; - int flag; - xbt_dict_foreach((*type)->attributes, cursor, key, value){ - smpi_type_key_elem elem = static_cast(xbt_dict_get_or_null_ext(smpi_type_keyvals, (const char*)key, sizeof(int))); - if(elem && elem->delete_fn) - elem->delete_fn(*type,*key, value, &flag); - } - } if((*type)->flags & DT_FLAG_PREDEFINED)return; @@ -427,50 +399,61 @@ void smpi_datatype_free(MPI_Datatype* type){ return; } - if ((*type)->has_subtype == 1){ - ((s_smpi_subtype_t *)(*type)->substruct)->subtype_free(type); - xbt_free((*type)->substruct); + if((*type)->attributes !=NULL){ + xbt_dict_cursor_t cursor = NULL; + int* key; + void * value; + int flag; + xbt_dict_foreach((*type)->attributes, cursor, key, value){ + smpi_type_key_elem elem = + static_cast(xbt_dict_get_or_null_ext(smpi_type_keyvals, (const char*)key, sizeof(int))); + if(elem && elem->delete_fn) + elem->delete_fn(*type,*key, value, &flag); + } } - if ((*type)->name != NULL){ - xbt_free((*type)->name); + + if ((*type)->sizeof_substruct != 0){ + //((s_smpi_subtype_t *)(*type)->substruct)->subtype_free(type); + xbt_free((*type)->substruct); } + xbt_free((*type)->name); xbt_free(*type); *type = MPI_DATATYPE_NULL; } void smpi_datatype_use(MPI_Datatype type){ + if(type)type->in_use++; -#ifdef HAVE_MC + if(type->sizeof_substruct!=0){ + ((s_smpi_subtype_t *)(type)->substruct)->subtype_use(&type); + } +#if HAVE_MC if(MC_is_active()) MC_ignore(&(type->in_use), sizeof(type->in_use)); #endif } - void smpi_datatype_unuse(MPI_Datatype type){ if (type->in_use > 0) type->in_use--; - if(type && type->in_use == 0 && (type->flags & DT_FLAG_DESTROYED)) - smpi_datatype_free(&type); + if(type->sizeof_substruct!=0){ + ((s_smpi_subtype_t *)(type)->substruct)->subtype_free(&type); + } -#ifdef HAVE_MC + if(type && type->in_use == 0){ + smpi_datatype_free(&type); + } +#if HAVE_MC if(MC_is_active()) MC_ignore(&(type->in_use), sizeof(type->in_use)); #endif } +/*Contiguous Implementation*/ - - -/* -Contiguous Implementation -*/ - - -/* - * Copies noncontiguous data into contiguous memory. +/* Copies noncontiguous data into contiguous memory. * @param contiguous_hvector - output hvector * @param noncontiguous_hvector - input hvector * @param type - pointer contening : @@ -478,19 +461,14 @@ Contiguous Implementation * - 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( const void *noncontiguous_hvector, void *contiguous_hvector, int count, void *type) { s_smpi_mpi_contiguous_t* type_c = (s_smpi_mpi_contiguous_t*)type; char* contiguous_vector_char = (char*)contiguous_hvector; char* noncontiguous_vector_char = (char*)noncontiguous_hvector+type_c->lb; - memcpy(contiguous_vector_char, - noncontiguous_vector_char, count* type_c->block_count * type_c->size_oldtype); + memcpy(contiguous_vector_char, noncontiguous_vector_char, count* type_c->block_count * type_c->size_oldtype); } -/* - * Copies contiguous data into noncontiguous memory. +/* Copies contiguous data into noncontiguous memory. * @param noncontiguous_vector - output hvector * @param contiguous_vector - input hvector * @param type - pointer contening : @@ -498,68 +476,54 @@ void serialize_contiguous( const void *noncontiguous_hvector, * - 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(const void *contiguous_vector, void *noncontiguous_vector, int count, void *type, MPI_Op op) { s_smpi_mpi_contiguous_t* type_c = (s_smpi_mpi_contiguous_t*)type; char* contiguous_vector_char = (char*)contiguous_vector; char* noncontiguous_vector_char = (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); - /*memcpy(noncontiguous_vector_char, - contiguous_vector_char, count* type_c->block_count * type_c->size_oldtype);*/ + smpi_op_apply(op, contiguous_vector_char, noncontiguous_vector_char, &n, &type_c->old_type); + /*memcpy(noncontiguous_vector_char, contiguous_vector_char, count* type_c->block_count * type_c->size_oldtype);*/ } void free_contiguous(MPI_Datatype* d){ - smpi_datatype_unuse(((s_smpi_mpi_indexed_t *)(*d)->substruct)->old_type); + smpi_datatype_unuse(((s_smpi_mpi_contiguous_t *)(*d)->substruct)->old_type); } -/* - * Create a Sub type contiguous to be able to serialize and unserialize it - * the structure s_smpi_mpi_contiguous_t is derived from s_smpi_subtype which - * required the functions unserialize and serialize - * - */ -s_smpi_mpi_contiguous_t* smpi_datatype_contiguous_create( MPI_Aint lb, - int block_count, - MPI_Datatype old_type, +void use_contiguous(MPI_Datatype* d){ + smpi_datatype_use(((s_smpi_mpi_contiguous_t *)(*d)->substruct)->old_type); +} + +/* Create a Sub type contiguous to be able to serialize and unserialize it the structure s_smpi_mpi_contiguous_t is + * erived from s_smpi_subtype which required the functions unserialize and serialize */ +s_smpi_mpi_contiguous_t* smpi_datatype_contiguous_create( MPI_Aint lb, int block_count, MPI_Datatype old_type, int size_oldtype){ s_smpi_mpi_contiguous_t *new_t= xbt_new(s_smpi_mpi_contiguous_t,1); new_t->base.serialize = &serialize_contiguous; new_t->base.unserialize = &unserialize_contiguous; new_t->base.subtype_free = &free_contiguous; + new_t->base.subtype_use = &use_contiguous; new_t->lb = lb; new_t->block_count = block_count; new_t->old_type = old_type; + smpi_datatype_use(old_type); new_t->size_oldtype = size_oldtype; smpi_datatype_use(old_type); return new_t; } - - - int smpi_datatype_contiguous(int count, MPI_Datatype old_type, MPI_Datatype* new_type, MPI_Aint lb) { int retval; - if(old_type->has_subtype){ + if(old_type->sizeof_substruct){ //handle this case as a hvector with stride equals to the extent of the datatype return smpi_datatype_hvector(count, 1, smpi_datatype_get_extent(old_type), old_type, new_type); } - s_smpi_mpi_contiguous_t* subtype = smpi_datatype_contiguous_create( lb, - count, - old_type, - smpi_datatype_size(old_type)); - - smpi_datatype_create(new_type, - count * smpi_datatype_size(old_type), - lb,lb + count * smpi_datatype_size(old_type), - 1,subtype, DT_FLAG_CONTIGUOUS); + s_smpi_mpi_contiguous_t* subtype = smpi_datatype_contiguous_create( lb, count, old_type,smpi_datatype_size(old_type)); + + smpi_datatype_create(new_type, count * smpi_datatype_size(old_type),lb,lb + count * smpi_datatype_size(old_type), + sizeof(s_smpi_mpi_contiguous_t),subtype, DT_FLAG_CONTIGUOUS); retval=MPI_SUCCESS; return retval; } @@ -574,30 +538,17 @@ int smpi_datatype_vector(int count, int blocklen, int stride, MPI_Datatype old_t lb=smpi_datatype_lb(old_type); ub=((count-1)*stride+blocklen-1)*smpi_datatype_get_extent(old_type)+smpi_datatype_ub(old_type); } - if(old_type->has_subtype || stride != blocklen){ - + if(old_type->sizeof_substruct || stride != blocklen){ - s_smpi_mpi_vector_t* subtype = smpi_datatype_vector_create( stride, - blocklen, - count, - old_type, + s_smpi_mpi_vector_t* subtype = smpi_datatype_vector_create(stride, blocklen, count, old_type, smpi_datatype_size(old_type)); - smpi_datatype_create(new_type, - count * (blocklen) * smpi_datatype_size(old_type), lb, - ub, - 1, - subtype, + smpi_datatype_create(new_type, count * (blocklen) * smpi_datatype_size(old_type), lb, ub, sizeof(s_smpi_mpi_vector_t), subtype, DT_FLAG_VECTOR); retval=MPI_SUCCESS; }else{ - /* in this situation the data are contignous thus it's not - * required to serialize and unserialize it*/ - smpi_datatype_create(new_type, count * blocklen * - smpi_datatype_size(old_type), 0, ((count -1) * stride + blocklen)* - smpi_datatype_size(old_type), - 0, - NULL, - DT_FLAG_VECTOR|DT_FLAG_CONTIGUOUS); + /* in this situation the data are contignous thus it's not required to serialize and unserialize it*/ + smpi_datatype_create(new_type, count * blocklen * smpi_datatype_size(old_type), 0, ((count -1) * stride + blocklen)* + smpi_datatype_size(old_type), 0, NULL, DT_FLAG_VECTOR|DT_FLAG_CONTIGUOUS); retval=MPI_SUCCESS; } return retval; @@ -607,13 +558,13 @@ void free_vector(MPI_Datatype* d){ smpi_datatype_unuse(((s_smpi_mpi_indexed_t *)(*d)->substruct)->old_type); } -/* -Hvector Implementation - Vector with stride in bytes -*/ +void use_vector(MPI_Datatype* d){ + smpi_datatype_use(((s_smpi_mpi_indexed_t *)(*d)->substruct)->old_type); +} +/* Hvector Implementation - Vector with stride in bytes */ -/* - * Copies noncontiguous data into contiguous memory. +/* Copies noncontiguous data into contiguous memory. * @param contiguous_hvector - output hvector * @param noncontiguous_hvector - input hvector * @param type - pointer contening : @@ -621,10 +572,7 @@ Hvector Implementation - Vector with stride in bytes * - 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( const void *noncontiguous_hvector, void *contiguous_hvector, int count, void *type) { s_smpi_mpi_hvector_t* type_c = (s_smpi_mpi_hvector_t*)type; int i; @@ -632,24 +580,21 @@ void serialize_hvector( const void *noncontiguous_hvector, char* noncontiguous_vector_char = (char*)noncontiguous_hvector; for (i = 0; i < type_c->block_count * count; i++) { - if (type_c->old_type->has_subtype == 0) - memcpy(contiguous_vector_char, - noncontiguous_vector_char, type_c->block_length * type_c->size_oldtype); + if (type_c->old_type->sizeof_substruct == 0) + memcpy(contiguous_vector_char, noncontiguous_vector_char, type_c->block_length * type_c->size_oldtype); else ((s_smpi_subtype_t*)type_c->old_type->substruct)->serialize( noncontiguous_vector_char, contiguous_vector_char, - type_c->block_length, - type_c->old_type->substruct); + type_c->block_length, type_c->old_type->substruct); contiguous_vector_char += type_c->block_length*type_c->size_oldtype; if((i+1)%type_c->block_count ==0) - noncontiguous_vector_char += type_c->block_length*type_c->size_oldtype; + noncontiguous_vector_char += type_c->block_length*type_c->size_oldtype; else - noncontiguous_vector_char += type_c->block_stride; + noncontiguous_vector_char += type_c->block_stride; } } -/* - * Copies contiguous data into noncontiguous memory. +/* Copies contiguous data into noncontiguous memory. * @param noncontiguous_vector - output hvector * @param contiguous_vector - input hvector * @param type - pointer contening : @@ -657,11 +602,7 @@ void serialize_hvector( const void *noncontiguous_hvector, * - 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( const void *contiguous_vector, void *noncontiguous_vector, int count, void *type, MPI_Op op) { s_smpi_mpi_hvector_t* type_c = (s_smpi_mpi_hvector_t*)type; int i; @@ -670,40 +611,33 @@ void unserialize_hvector( const void *contiguous_vector, char* noncontiguous_vector_char = (char*)noncontiguous_vector; for (i = 0; i < type_c->block_count * count; i++) { - if (type_c->old_type->has_subtype == 0) - smpi_op_apply(op, contiguous_vector_char, noncontiguous_vector_char, &type_c->block_length, - &type_c->old_type); + if (type_c->old_type->sizeof_substruct == 0) + smpi_op_apply(op, contiguous_vector_char, noncontiguous_vector_char, &type_c->block_length, &type_c->old_type); /*memcpy(noncontiguous_vector_char, contiguous_vector_char, type_c->block_length * type_c->size_oldtype);*/ else - ((s_smpi_subtype_t*)type_c->old_type->substruct)->unserialize( contiguous_vector_char, - noncontiguous_vector_char, - type_c->block_length, - type_c->old_type->substruct, + ((s_smpi_subtype_t*)type_c->old_type->substruct)->unserialize( contiguous_vector_char, noncontiguous_vector_char, + type_c->block_length, type_c->old_type->substruct, op); contiguous_vector_char += type_c->block_length*type_c->size_oldtype; if((i+1)%type_c->block_count ==0) - noncontiguous_vector_char += type_c->block_length*type_c->size_oldtype; + noncontiguous_vector_char += type_c->block_length*type_c->size_oldtype; else - noncontiguous_vector_char += type_c->block_stride; + noncontiguous_vector_char += type_c->block_stride; } } -/* - * Create a Sub type vector to be able to serialize and unserialize it - * the structure s_smpi_mpi_vector_t is derived from s_smpi_subtype which - * required the functions unserialize and serialize +/* Create a Sub type vector to be able to serialize and unserialize it the structure s_smpi_mpi_vector_t is derived + * from s_smpi_subtype which required the functions unserialize and serialize * */ -s_smpi_mpi_hvector_t* smpi_datatype_hvector_create( MPI_Aint block_stride, - int block_length, - int block_count, - MPI_Datatype old_type, - int size_oldtype){ +s_smpi_mpi_hvector_t* smpi_datatype_hvector_create( MPI_Aint block_stride, int block_length, int block_count, + MPI_Datatype old_type, int size_oldtype){ s_smpi_mpi_hvector_t *new_t= xbt_new(s_smpi_mpi_hvector_t,1); new_t->base.serialize = &serialize_hvector; new_t->base.unserialize = &unserialize_hvector; new_t->base.subtype_free = &free_hvector; + new_t->base.subtype_use = &use_hvector; new_t->block_stride = block_stride; new_t->block_length = block_length; new_t->block_count = block_count; @@ -715,7 +649,11 @@ s_smpi_mpi_hvector_t* smpi_datatype_hvector_create( MPI_Aint block_stride, //do nothing for vector types void free_hvector(MPI_Datatype* d){ - smpi_datatype_unuse(((s_smpi_mpi_indexed_t *)(*d)->substruct)->old_type); + smpi_datatype_unuse(((s_smpi_mpi_hvector_t *)(*d)->substruct)->old_type); +} + +void use_hvector(MPI_Datatype* d){ + smpi_datatype_use(((s_smpi_mpi_hvector_t *)(*d)->substruct)->old_type); } int smpi_datatype_hvector(int count, int blocklen, MPI_Aint stride, MPI_Datatype old_type, MPI_Datatype* new_type) @@ -728,38 +666,23 @@ int smpi_datatype_hvector(int count, int blocklen, MPI_Aint stride, MPI_Datatype lb=smpi_datatype_lb(old_type); ub=((count-1)*stride)+(blocklen-1)*smpi_datatype_get_extent(old_type)+smpi_datatype_ub(old_type); } - if(old_type->has_subtype || stride != blocklen*smpi_datatype_get_extent(old_type)){ - s_smpi_mpi_hvector_t* subtype = smpi_datatype_hvector_create( stride, - blocklen, - count, - old_type, + if(old_type->sizeof_substruct || stride != blocklen*smpi_datatype_get_extent(old_type)){ + s_smpi_mpi_hvector_t* subtype = smpi_datatype_hvector_create( stride, blocklen, count, old_type, smpi_datatype_size(old_type)); - smpi_datatype_create(new_type, count * blocklen * smpi_datatype_size(old_type), - lb,ub, - 1, - subtype, - DT_FLAG_VECTOR); + smpi_datatype_create(new_type, count * blocklen * smpi_datatype_size(old_type), lb,ub, sizeof(s_smpi_mpi_hvector_t), subtype, DT_FLAG_VECTOR); retval=MPI_SUCCESS; }else{ - smpi_datatype_create(new_type, count * blocklen * - smpi_datatype_size(old_type),0,count * blocklen * - smpi_datatype_size(old_type), - 0, - NULL, - DT_FLAG_VECTOR|DT_FLAG_CONTIGUOUS); + smpi_datatype_create(new_type, count * blocklen * smpi_datatype_size(old_type),0,count * blocklen * + smpi_datatype_size(old_type), 0, NULL, DT_FLAG_VECTOR|DT_FLAG_CONTIGUOUS); retval=MPI_SUCCESS; } return retval; } +/* Indexed Implementation */ -/* -Indexed Implementation -*/ - -/* - * Copies noncontiguous data into contiguous memory. +/* Copies noncontiguous data into contiguous memory. * @param contiguous_indexed - output indexed * @param noncontiguous_indexed - input indexed * @param type - pointer contening : @@ -767,10 +690,7 @@ Indexed Implementation * - 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( const void *noncontiguous_indexed, void *contiguous_indexed, int count, void *type) { s_smpi_mpi_indexed_t* type_c = (s_smpi_mpi_indexed_t*)type; int i,j; @@ -778,25 +698,25 @@ void serialize_indexed( const void *noncontiguous_indexed, char* noncontiguous_indexed_char = (char*)noncontiguous_indexed+type_c->block_indices[0] * type_c->size_oldtype; for(j=0; jblock_count; i++) { - if (type_c->old_type->has_subtype == 0) - memcpy(contiguous_indexed_char, - noncontiguous_indexed_char, type_c->block_lengths[i] * type_c->size_oldtype); + if (type_c->old_type->sizeof_substruct == 0) + memcpy(contiguous_indexed_char, noncontiguous_indexed_char, type_c->block_lengths[i] * type_c->size_oldtype); else ((s_smpi_subtype_t*)type_c->old_type->substruct)->serialize( noncontiguous_indexed_char, contiguous_indexed_char, type_c->block_lengths[i], type_c->old_type->substruct); - 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]*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); + if (iblock_count-1) + noncontiguous_indexed_char = + (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=(void*)noncontiguous_indexed_char; } } -/* - * Copies contiguous data into noncontiguous memory. +/* Copies contiguous data into noncontiguous memory. * @param noncontiguous_indexed - output indexed * @param contiguous_indexed - input indexed * @param type - pointer contening : @@ -804,20 +724,16 @@ void serialize_indexed( const void *noncontiguous_indexed, * - 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( const void *contiguous_indexed, void *noncontiguous_indexed, int count, void *type, MPI_Op op) { - s_smpi_mpi_indexed_t* type_c = (s_smpi_mpi_indexed_t*)type; int i,j; char* contiguous_indexed_char = (char*)contiguous_indexed; - char* noncontiguous_indexed_char = (char*)noncontiguous_indexed+type_c->block_indices[0]*smpi_datatype_get_extent(type_c->old_type); + char* noncontiguous_indexed_char = + (char*)noncontiguous_indexed+type_c->block_indices[0]*smpi_datatype_get_extent(type_c->old_type); for(j=0; jblock_count; i++) { - if (type_c->old_type->has_subtype == 0) + if (type_c->old_type->sizeof_substruct == 0) smpi_op_apply(op, contiguous_indexed_char, noncontiguous_indexed_char, &type_c->block_lengths[i], &type_c->old_type); /*memcpy(noncontiguous_indexed_char , @@ -826,39 +742,41 @@ void unserialize_indexed( const void *contiguous_indexed, ((s_smpi_subtype_t*)type_c->old_type->substruct)->unserialize( contiguous_indexed_char, noncontiguous_indexed_char, type_c->block_lengths[i], - type_c->old_type->substruct, - op); + type_c->old_type->substruct, op); 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]*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_char = + (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=(void*)noncontiguous_indexed_char; } } void free_indexed(MPI_Datatype* type){ - xbt_free(((s_smpi_mpi_indexed_t *)(*type)->substruct)->block_lengths); - xbt_free(((s_smpi_mpi_indexed_t *)(*type)->substruct)->block_indices); + if((*type)->in_use==0){ + xbt_free(((s_smpi_mpi_indexed_t *)(*type)->substruct)->block_lengths); + xbt_free(((s_smpi_mpi_indexed_t *)(*type)->substruct)->block_indices); + } smpi_datatype_unuse(((s_smpi_mpi_indexed_t *)(*type)->substruct)->old_type); } -/* - * Create a Sub type indexed to be able to serialize and unserialize it - * 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, - int* block_indices, - int block_count, - MPI_Datatype old_type, - int size_oldtype){ +void use_indexed(MPI_Datatype* type){ + smpi_datatype_use(((s_smpi_mpi_indexed_t *)(*type)->substruct)->old_type); +} + + +/* Create a Sub type indexed to be able to serialize and unserialize it 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, int* block_indices, int block_count, + MPI_Datatype old_type, int size_oldtype){ 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; new_t->base.subtype_free = &free_indexed; - //TODO : add a custom function for each time to clean these + new_t->base.subtype_use = &use_indexed; new_t->block_lengths= xbt_new(int, block_count); new_t->block_indices= xbt_new(int, block_count); int i; @@ -873,7 +791,6 @@ s_smpi_mpi_indexed_t* smpi_datatype_indexed_create( int* block_lengths, return new_t; } - int smpi_datatype_indexed(int count, int* blocklens, int* indices, MPI_Datatype old_type, MPI_Datatype* new_type) { int i; @@ -899,36 +816,25 @@ int smpi_datatype_indexed(int count, int* blocklens, int* indices, MPI_Datatype if ( (i< count -1) && (indices[i]+blocklens[i] != indices[i+1]) )contiguous=0; } - if (old_type->has_subtype == 1) + if (old_type->sizeof_substruct != 0) contiguous=0; if(!contiguous){ - s_smpi_mpi_indexed_t* subtype = smpi_datatype_indexed_create( blocklens, - indices, - count, - old_type, + 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),lb,ub,1, subtype, DT_FLAG_DATA); + smpi_datatype_create(new_type, size * smpi_datatype_size(old_type),lb,ub,sizeof(s_smpi_mpi_indexed_t), subtype, DT_FLAG_DATA); }else{ - s_smpi_mpi_contiguous_t* subtype = smpi_datatype_contiguous_create( lb, - size, - old_type, + s_smpi_mpi_contiguous_t* subtype = smpi_datatype_contiguous_create( lb, size, old_type, smpi_datatype_size(old_type)); - smpi_datatype_create(new_type, size * - smpi_datatype_size(old_type),lb,ub,1, subtype, DT_FLAG_DATA|DT_FLAG_CONTIGUOUS); + smpi_datatype_create(new_type, size * smpi_datatype_size(old_type), lb, ub, sizeof(s_smpi_mpi_contiguous_t), subtype, + DT_FLAG_DATA|DT_FLAG_CONTIGUOUS); } retval=MPI_SUCCESS; return retval; } +/* Hindexed Implementation - Indexed with indices in bytes */ - -/* -Hindexed Implementation - Indexed with indices in bytes -*/ - -/* - * Copies noncontiguous data into contiguous memory. +/* Copies noncontiguous data into contiguous memory. * @param contiguous_hindexed - output hindexed * @param noncontiguous_hindexed - input hindexed * @param type - pointer contening : @@ -936,10 +842,7 @@ Hindexed Implementation - Indexed with indices in bytes * - 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( const void *noncontiguous_hindexed, void *contiguous_hindexed, int count, void *type) { s_smpi_mpi_hindexed_t* type_c = (s_smpi_mpi_hindexed_t*)type; int i,j; @@ -947,9 +850,8 @@ void serialize_hindexed( const void *noncontiguous_hindexed, char* noncontiguous_hindexed_char = (char*)noncontiguous_hindexed+ type_c->block_indices[0]; for(j=0; jblock_count; i++) { - if (type_c->old_type->has_subtype == 0) - memcpy(contiguous_hindexed_char, - noncontiguous_hindexed_char, type_c->block_lengths[i] * type_c->size_oldtype); + if (type_c->old_type->sizeof_substruct == 0) + memcpy(contiguous_hindexed_char, noncontiguous_hindexed_char, type_c->block_lengths[i] * type_c->size_oldtype); else ((s_smpi_subtype_t*)type_c->old_type->substruct)->serialize( noncontiguous_hindexed_char, contiguous_hindexed_char, @@ -957,14 +859,15 @@ void serialize_hindexed( const void *noncontiguous_hindexed, type_c->old_type->substruct); 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]*smpi_datatype_get_extent(type_c->old_type); + 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]*smpi_datatype_get_extent(type_c->old_type); } noncontiguous_hindexed=(void*)noncontiguous_hindexed_char; } } -/* - * Copies contiguous data into noncontiguous memory. +/* Copies contiguous data into noncontiguous memory. * @param noncontiguous_hindexed - output hindexed * @param contiguous_hindexed - input hindexed * @param type - pointer contening : @@ -972,10 +875,7 @@ void serialize_hindexed( const void *noncontiguous_hindexed, * - 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( const void *contiguous_hindexed, void *noncontiguous_hindexed, int count, void *type, MPI_Op op) { s_smpi_mpi_hindexed_t* type_c = (s_smpi_mpi_hindexed_t*)type; @@ -985,47 +885,48 @@ void unserialize_hindexed( const void *contiguous_hindexed, char* noncontiguous_hindexed_char = (char*)noncontiguous_hindexed+ type_c->block_indices[0]; for(j=0; jblock_count; i++) { - if (type_c->old_type->has_subtype == 0) + if (type_c->old_type->sizeof_substruct == 0) smpi_op_apply(op, contiguous_hindexed_char, noncontiguous_hindexed_char, &type_c->block_lengths[i], &type_c->old_type); - /*memcpy(noncontiguous_hindexed_char, - contiguous_hindexed_char, type_c->block_lengths[i] * type_c->size_oldtype);*/ + /*memcpy(noncontiguous_hindexed_char,contiguous_hindexed_char,type_c->block_lengths[i]*type_c->size_oldtype);*/ else ((s_smpi_subtype_t*)type_c->old_type->substruct)->unserialize( contiguous_hindexed_char, noncontiguous_hindexed_char, type_c->block_lengths[i], - type_c->old_type->substruct, - op); + type_c->old_type->substruct, op); 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]*smpi_datatype_get_extent(type_c->old_type); + 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]*smpi_datatype_get_extent(type_c->old_type); } noncontiguous_hindexed=(void*)noncontiguous_hindexed_char; } } void free_hindexed(MPI_Datatype* type){ - xbt_free(((s_smpi_mpi_hindexed_t *)(*type)->substruct)->block_lengths); - xbt_free(((s_smpi_mpi_hindexed_t *)(*type)->substruct)->block_indices); + if((*type)->in_use==0){ + xbt_free(((s_smpi_mpi_hindexed_t *)(*type)->substruct)->block_lengths); + xbt_free(((s_smpi_mpi_hindexed_t *)(*type)->substruct)->block_indices); + } smpi_datatype_unuse(((s_smpi_mpi_indexed_t *)(*type)->substruct)->old_type); } -/* - * Create a Sub type hindexed to be able to serialize and unserialize it - * the structure s_smpi_mpi_hindexed_t is derived from s_smpi_subtype which - * required the functions unserialize and serialize +void use_hindexed(MPI_Datatype* type){ + smpi_datatype_use(((s_smpi_mpi_hindexed_t *)(*type)->substruct)->old_type); +} + +/* Create a Sub type hindexed to be able to serialize and unserialize it 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, - MPI_Aint* block_indices, - int block_count, - MPI_Datatype old_type, - int size_oldtype){ +s_smpi_mpi_hindexed_t* smpi_datatype_hindexed_create( int* block_lengths, MPI_Aint* block_indices, int block_count, + MPI_Datatype old_type, int size_oldtype){ 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; new_t->base.subtype_free = &free_hindexed; - //TODO : add a custom function for each time to clean these + new_t->base.subtype_use = &use_hindexed; new_t->block_lengths= xbt_new(int, block_count); new_t->block_indices= xbt_new(MPI_Aint, block_count); int i; @@ -1035,11 +936,11 @@ s_smpi_mpi_hindexed_t* smpi_datatype_hindexed_create( int* block_lengths, } new_t->block_count = block_count; new_t->old_type = old_type; + smpi_datatype_use(old_type); new_t->size_oldtype = size_oldtype; return new_t; } - int smpi_datatype_hindexed(int count, int* blocklens, MPI_Aint* indices, MPI_Datatype old_type, MPI_Datatype* new_type) { int i; @@ -1063,39 +964,25 @@ int smpi_datatype_hindexed(int count, int* blocklens, MPI_Aint* indices, MPI_Dat if ( (i< count -1) && (indices[i]+blocklens[i]*static_cast(smpi_datatype_size(old_type)) != indices[i+1]) ) contiguous=0; } - if (old_type->has_subtype == 1 || lb!=0) + if (old_type->sizeof_substruct != 0 || lb!=0) contiguous=0; if(!contiguous){ - s_smpi_mpi_hindexed_t* subtype = smpi_datatype_hindexed_create( blocklens, - indices, - count, - old_type, + 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), - lb, - ub - ,1, subtype, DT_FLAG_DATA); + smpi_datatype_create(new_type, size * smpi_datatype_size(old_type), lb, ub ,sizeof(s_smpi_mpi_hindexed_t), subtype, DT_FLAG_DATA); }else{ - s_smpi_mpi_contiguous_t* subtype = smpi_datatype_contiguous_create( lb, - size, - old_type, - smpi_datatype_size(old_type)); - smpi_datatype_create(new_type, size * smpi_datatype_size(old_type), - 0,size * smpi_datatype_size(old_type), + s_smpi_mpi_contiguous_t* subtype = smpi_datatype_contiguous_create(lb,size, old_type, smpi_datatype_size(old_type)); + smpi_datatype_create(new_type, size * smpi_datatype_size(old_type), 0,size * smpi_datatype_size(old_type), 1, subtype, DT_FLAG_DATA|DT_FLAG_CONTIGUOUS); } retval=MPI_SUCCESS; return retval; } +/* struct Implementation - Indexed with indices in bytes */ -/* -struct Implementation - Indexed with indices in bytes -*/ - -/* - * Copies noncontiguous data into contiguous memory. +/* Copies noncontiguous data into contiguous memory. * @param contiguous_struct - output struct * @param noncontiguous_struct - input struct * @param type - pointer contening : @@ -1103,10 +990,7 @@ struct Implementation - Indexed with indices in bytes * - 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( const void *noncontiguous_struct, void *contiguous_struct, int count, void *type) { s_smpi_mpi_struct_t* type_c = (s_smpi_mpi_struct_t*)type; int i,j; @@ -1114,9 +998,9 @@ void serialize_struct( const void *noncontiguous_struct, char* noncontiguous_struct_char = (char*)noncontiguous_struct+ type_c->block_indices[0]; for(j=0; jblock_count; i++) { - if (type_c->old_types[i]->has_subtype == 0) - memcpy(contiguous_struct_char, - noncontiguous_struct_char, type_c->block_lengths[i] * smpi_datatype_size(type_c->old_types[i])); + if (type_c->old_types[i]->sizeof_substruct == 0) + memcpy(contiguous_struct_char, noncontiguous_struct_char, + type_c->block_lengths[i] * smpi_datatype_size(type_c->old_types[i])); else ((s_smpi_subtype_t*)type_c->old_types[i]->substruct)->serialize( noncontiguous_struct_char, contiguous_struct_char, @@ -1125,14 +1009,16 @@ void serialize_struct( const void *noncontiguous_struct, 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_get_extent(type_c->old_types[i]);//let's hope this is MPI_UB ? + if (iblock_count-1) + noncontiguous_struct_char = (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=(void*)noncontiguous_struct_char; } } -/* - * Copies contiguous data into noncontiguous memory. + +/* Copies contiguous data into noncontiguous memory. * @param noncontiguous_struct - output struct * @param contiguous_struct - input struct * @param type - pointer contening : @@ -1140,11 +1026,7 @@ void serialize_struct( const void *noncontiguous_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( const void *contiguous_struct, void *noncontiguous_struct, int count, void *type, MPI_Op op) { s_smpi_mpi_struct_t* type_c = (s_smpi_mpi_struct_t*)type; int i,j; @@ -1153,50 +1035,54 @@ void unserialize_struct( const void *contiguous_struct, char* noncontiguous_struct_char = (char*)noncontiguous_struct+ type_c->block_indices[0]; for(j=0; jblock_count; i++) { - if (type_c->old_types[i]->has_subtype == 0) + if (type_c->old_types[i]->sizeof_substruct == 0) smpi_op_apply(op, contiguous_struct_char, noncontiguous_struct_char, &type_c->block_lengths[i], & type_c->old_types[i]); - /*memcpy(noncontiguous_struct_char, - contiguous_struct_char, type_c->block_lengths[i] * smpi_datatype_size(type_c->old_types[i]));*/ + /*memcpy(noncontiguous_struct_char, + contiguous_struct_char, type_c->block_lengths[i] * smpi_datatype_size(type_c->old_types[i]));*/ else ((s_smpi_subtype_t*)type_c->old_types[i]->substruct)->unserialize( contiguous_struct_char, noncontiguous_struct_char, type_c->block_lengths[i], - type_c->old_types[i]->substruct, - op); + type_c->old_types[i]->substruct, op); 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_get_extent(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_get_extent(type_c->old_types[i]); } noncontiguous_struct=(void*)noncontiguous_struct_char; - } } void free_struct(MPI_Datatype* type){ - xbt_free(((s_smpi_mpi_struct_t *)(*type)->substruct)->block_lengths); - xbt_free(((s_smpi_mpi_struct_t *)(*type)->substruct)->block_indices); int i=0; for (i = 0; i < ((s_smpi_mpi_struct_t *)(*type)->substruct)->block_count; i++) smpi_datatype_unuse(((s_smpi_mpi_struct_t *)(*type)->substruct)->old_types[i]); - xbt_free(((s_smpi_mpi_struct_t *)(*type)->substruct)->old_types); + if((*type)->in_use==0){ + xbt_free(((s_smpi_mpi_struct_t *)(*type)->substruct)->block_lengths); + xbt_free(((s_smpi_mpi_struct_t *)(*type)->substruct)->block_indices); + xbt_free(((s_smpi_mpi_struct_t *)(*type)->substruct)->old_types); + } } -/* - * Create a Sub type struct to be able to serialize and unserialize it - * the structure s_smpi_mpi_struct_t is derived from s_smpi_subtype which - * required the functions unserialize and serialize +void use_struct(MPI_Datatype* type){ + int i=0; + for (i = 0; i < ((s_smpi_mpi_struct_t *)(*type)->substruct)->block_count; i++) + smpi_datatype_use(((s_smpi_mpi_struct_t *)(*type)->substruct)->old_types[i]); +} + +/* Create a Sub type struct to be able to serialize and unserialize it 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, - MPI_Aint* block_indices, - int block_count, +s_smpi_mpi_struct_t* smpi_datatype_struct_create( int* block_lengths, MPI_Aint* block_indices, int block_count, MPI_Datatype* old_types){ 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; new_t->base.subtype_free = &free_struct; - //TODO : add a custom function for each time to clean these + new_t->base.subtype_use = &use_struct; 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); @@ -1207,14 +1093,10 @@ s_smpi_mpi_struct_t* smpi_datatype_struct_create( int* block_lengths, new_t->old_types[i]=old_types[i]; smpi_datatype_use(new_t->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; return new_t; } - int smpi_datatype_struct(int count, int* blocklens, MPI_Aint* indices, MPI_Datatype* old_types, MPI_Datatype* new_type) { int i; @@ -1232,7 +1114,7 @@ int smpi_datatype_struct(int count, int* blocklens, MPI_Aint* indices, MPI_Datat for(i=0; i< count; i++){ if (blocklens[i]<0) return MPI_ERR_ARG; - if (old_types[i]->has_subtype == 1) + if (old_types[i]->sizeof_substruct != 0) contiguous=0; size += blocklens[i]*smpi_datatype_size(old_types[i]); @@ -1246,24 +1128,19 @@ int smpi_datatype_struct(int count, int* blocklens, MPI_Aint* indices, MPI_Datat } if(!forced_lb && indices[i]+smpi_datatype_lb(old_types[i])ub) ub = indices[i]+blocklens[i]*smpi_datatype_ub(old_types[i]); + if(!forced_ub && indices[i]+blocklens[i]*smpi_datatype_ub(old_types[i])>ub) + ub = indices[i]+blocklens[i]*smpi_datatype_ub(old_types[i]); if ( (i< count -1) && (indices[i]+blocklens[i]*static_cast(smpi_datatype_size(old_types[i])) != indices[i+1]) ) contiguous=0; } if(!contiguous){ - s_smpi_mpi_struct_t* subtype = smpi_datatype_struct_create( blocklens, - indices, - count, - old_types); + s_smpi_mpi_struct_t* subtype = smpi_datatype_struct_create( blocklens, indices, count, old_types); - smpi_datatype_create(new_type, size, lb, ub,1, subtype, DT_FLAG_DATA); + smpi_datatype_create(new_type, size, lb, ub,sizeof(s_smpi_mpi_struct_t), subtype, DT_FLAG_DATA); }else{ - s_smpi_mpi_contiguous_t* subtype = smpi_datatype_contiguous_create( lb, - size, - MPI_CHAR, - 1); + s_smpi_mpi_contiguous_t* subtype = smpi_datatype_contiguous_create( lb, size, MPI_CHAR, 1); smpi_datatype_create(new_type, size, lb, ub,1, subtype, DT_FLAG_DATA|DT_FLAG_CONTIGUOUS); } return MPI_SUCCESS; @@ -1302,8 +1179,7 @@ typedef struct s_smpi_mpi_op { } \ } -static void max_func(void *a, void *b, int *length, - MPI_Datatype * datatype) +static void max_func(void *a, void *b, int *length, MPI_Datatype * datatype) { if (*datatype == MPI_CHAR) { APPLY_FUNC(a, b, length, char, MAX_OP); @@ -1330,8 +1206,7 @@ static void max_func(void *a, void *b, int *length, } } -static void min_func(void *a, void *b, int *length, - MPI_Datatype * datatype) +static void min_func(void *a, void *b, int *length, MPI_Datatype * datatype) { if (*datatype == MPI_CHAR) { APPLY_FUNC(a, b, length, char, MIN_OP); @@ -1358,8 +1233,7 @@ static void min_func(void *a, void *b, int *length, } } -static void sum_func(void *a, void *b, int *length, - MPI_Datatype * datatype) +static void sum_func(void *a, void *b, int *length, MPI_Datatype * datatype) { if (*datatype == MPI_CHAR) { APPLY_FUNC(a, b, length, char, SUM_OP); @@ -1392,8 +1266,7 @@ static void sum_func(void *a, void *b, int *length, } } -static void prod_func(void *a, void *b, int *length, - MPI_Datatype * datatype) +static void prod_func(void *a, void *b, int *length, MPI_Datatype * datatype) { if (*datatype == MPI_CHAR) { APPLY_FUNC(a, b, length, char, PROD_OP); @@ -1426,8 +1299,7 @@ static void prod_func(void *a, void *b, int *length, } } -static void land_func(void *a, void *b, int *length, - MPI_Datatype * datatype) +static void land_func(void *a, void *b, int *length, MPI_Datatype * datatype) { if (*datatype == MPI_CHAR) { APPLY_FUNC(a, b, length, char, LAND_OP); @@ -1450,8 +1322,7 @@ static void land_func(void *a, void *b, int *length, } } -static void lor_func(void *a, void *b, int *length, - MPI_Datatype * datatype) +static void lor_func(void *a, void *b, int *length, MPI_Datatype * datatype) { if (*datatype == MPI_CHAR) { APPLY_FUNC(a, b, length, char, LOR_OP); @@ -1474,8 +1345,7 @@ static void lor_func(void *a, void *b, int *length, } } -static void lxor_func(void *a, void *b, int *length, - MPI_Datatype * datatype) +static void lxor_func(void *a, void *b, int *length, MPI_Datatype * datatype) { if (*datatype == MPI_CHAR) { APPLY_FUNC(a, b, length, char, LXOR_OP); @@ -1498,8 +1368,7 @@ static void lxor_func(void *a, void *b, int *length, } } -static void band_func(void *a, void *b, int *length, - MPI_Datatype * datatype) +static void band_func(void *a, void *b, int *length, MPI_Datatype * datatype) { if (*datatype == MPI_CHAR) { APPLY_FUNC(a, b, length, char, BAND_OP); @@ -1522,8 +1391,7 @@ static void band_func(void *a, void *b, int *length, } } -static void bor_func(void *a, void *b, int *length, - MPI_Datatype * datatype) +static void bor_func(void *a, void *b, int *length, MPI_Datatype * datatype) { if (*datatype == MPI_CHAR) { APPLY_FUNC(a, b, length, char, BOR_OP); @@ -1546,8 +1414,7 @@ static void bor_func(void *a, void *b, int *length, } } -static void bxor_func(void *a, void *b, int *length, - MPI_Datatype * datatype) +static void bxor_func(void *a, void *b, int *length, MPI_Datatype * datatype) { if (*datatype == MPI_CHAR) { APPLY_FUNC(a, b, length, char, BXOR_OP); @@ -1570,8 +1437,7 @@ static void bxor_func(void *a, void *b, int *length, } } -static void minloc_func(void *a, void *b, int *length, - MPI_Datatype * datatype) +static void minloc_func(void *a, void *b, int *length, MPI_Datatype * datatype) { if (*datatype == MPI_FLOAT_INT) { APPLY_FUNC(a, b, length, float_int, MINLOC_OP); @@ -1594,8 +1460,7 @@ static void minloc_func(void *a, void *b, int *length, } } -static void maxloc_func(void *a, void *b, int *length, - MPI_Datatype * datatype) +static void maxloc_func(void *a, void *b, int *length, MPI_Datatype * datatype) { if (*datatype == MPI_FLOAT_INT) { APPLY_FUNC(a, b, length, float_int, MAXLOC_OP); @@ -1618,14 +1483,13 @@ static void maxloc_func(void *a, void *b, int *length, } } -static void replace_func(void *a, void *b, int *length, - MPI_Datatype * datatype) +static void replace_func(void *a, void *b, int *length, MPI_Datatype * datatype) { memcpy(b, a, *length * smpi_datatype_size(*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 }; \ MPI_Op name = &mpi_##name; CREATE_MPI_OP(MPI_MAX, max_func); @@ -1642,7 +1506,6 @@ CREATE_MPI_OP(MPI_MAXLOC, maxloc_func); CREATE_MPI_OP(MPI_MINLOC, minloc_func); CREATE_MPI_OP(MPI_REPLACE, replace_func); - MPI_Op smpi_op_new(MPI_User_function * function, int commute) { MPI_Op op; @@ -1662,13 +1525,12 @@ void smpi_op_destroy(MPI_Op op) xbt_free(op); } -void smpi_op_apply(MPI_Op op, void *invec, void *inoutvec, int *len, - MPI_Datatype * datatype) +void smpi_op_apply(MPI_Op op, void *invec, void *inoutvec, int *len, MPI_Datatype * datatype) { if(op==MPI_OP_NULL) return; - if(smpi_privatize_global_variables){ //we need to switch here, as the called function may silently touch global variables + if(smpi_privatize_global_variables){//we need to switch as the called function may silently touch global variables XBT_DEBUG("Applying operation, switch to the right data frame "); smpi_switch_data_segment(smpi_process_index()); } @@ -1678,7 +1540,8 @@ void smpi_op_apply(MPI_Op op, void *invec, void *inoutvec, int *len, } int smpi_type_attr_delete(MPI_Datatype type, int keyval){ - smpi_type_key_elem elem = static_cast(xbt_dict_get_or_null_ext(smpi_type_keyvals, (const char*)&keyval, sizeof(int))); + smpi_type_key_elem elem = + static_cast(xbt_dict_get_or_null_ext(smpi_type_keyvals, (const char*)&keyval, sizeof(int))); if(!elem) return MPI_ERR_ARG; if(elem->delete_fn!=MPI_NULL_DELETE_FN){ @@ -1697,7 +1560,8 @@ int smpi_type_attr_delete(MPI_Datatype type, int keyval){ } int smpi_type_attr_get(MPI_Datatype type, int keyval, void* attr_value, int* flag){ - smpi_type_key_elem elem = static_cast(xbt_dict_get_or_null_ext(smpi_type_keyvals, (const char*)&keyval, sizeof(int))); + smpi_type_key_elem elem = + static_cast(xbt_dict_get_or_null_ext(smpi_type_keyvals, (const char*)&keyval, sizeof(int))); if(!elem) return MPI_ERR_ARG; xbt_ex_t ex; @@ -1719,7 +1583,8 @@ int smpi_type_attr_get(MPI_Datatype type, int keyval, void* attr_value, int* fla int smpi_type_attr_put(MPI_Datatype type, int keyval, void* attr_value){ if(!smpi_type_keyvals) smpi_type_keyvals = xbt_dict_new(); - smpi_type_key_elem elem = static_cast(xbt_dict_get_or_null_ext(smpi_type_keyvals, (const char*)&keyval, sizeof(int))); + smpi_type_key_elem elem = + static_cast(xbt_dict_get_or_null_ext(smpi_type_keyvals, (const char*)&keyval, sizeof(int))); if(!elem ) return MPI_ERR_ARG; int flag; @@ -1736,16 +1601,16 @@ int smpi_type_attr_put(MPI_Datatype type, int keyval, void* attr_value){ return MPI_SUCCESS; } -int smpi_type_keyval_create(MPI_Type_copy_attr_function* copy_fn, MPI_Type_delete_attr_function* delete_fn, int* keyval, void* extra_state){ - +int smpi_type_keyval_create(MPI_Type_copy_attr_function* copy_fn, MPI_Type_delete_attr_function* delete_fn, int* keyval, + void* extra_state){ if(!smpi_type_keyvals) smpi_type_keyvals = xbt_dict_new(); - + smpi_type_key_elem value = (smpi_type_key_elem) xbt_new0(s_smpi_mpi_type_key_elem_t,1); - + value->copy_fn=copy_fn; value->delete_fn=delete_fn; - + *keyval = type_keyval_id; xbt_dict_set_ext(smpi_type_keyvals,(const char*)keyval, sizeof(int),(void*)value, NULL); type_keyval_id++; @@ -1753,7 +1618,8 @@ int smpi_type_keyval_create(MPI_Type_copy_attr_function* copy_fn, MPI_Type_delet } int smpi_type_keyval_free(int* keyval){ - smpi_type_key_elem elem = static_cast(xbt_dict_get_or_null_ext(smpi_type_keyvals, (const char*)keyval, sizeof(int))); + smpi_type_key_elem elem = + static_cast(xbt_dict_get_or_null_ext(smpi_type_keyvals, (const char*)keyval, sizeof(int))); if(!elem){ return MPI_ERR_ARG; } @@ -1762,22 +1628,20 @@ int smpi_type_keyval_free(int* keyval){ return MPI_SUCCESS; } -int smpi_mpi_pack(void* inbuf, int incount, MPI_Datatype type, void* outbuf, int outcount, int* position, MPI_Comm comm){ +int smpi_mpi_pack(void* inbuf, int incount, MPI_Datatype type, void* outbuf, int outcount, int* position,MPI_Comm comm){ size_t size = smpi_datatype_size(type); if (outcount - *position < incount*static_cast(size)) return MPI_ERR_BUFFER; - smpi_datatype_copy(inbuf, incount, type, - (char*)outbuf + *position, outcount, MPI_CHAR); + smpi_datatype_copy(inbuf, incount, type, (char*)outbuf + *position, outcount, MPI_CHAR); *position += incount * size; return MPI_SUCCESS; } -int smpi_mpi_unpack(void* inbuf, int insize, int* position, void* outbuf, int outcount, MPI_Datatype type, MPI_Comm comm){ +int smpi_mpi_unpack(void* inbuf, int insize, int* position, void* outbuf, int outcount, MPI_Datatype type,MPI_Comm comm){ int size = static_cast(smpi_datatype_size(type)); if (outcount*size> insize) return MPI_ERR_BUFFER; - smpi_datatype_copy((char*)inbuf + *position, insize, MPI_CHAR, - outbuf, outcount, type); + smpi_datatype_copy((char*)inbuf + *position, insize, MPI_CHAR, outbuf, outcount, type); *position += outcount * size; return MPI_SUCCESS; }