X-Git-Url: http://info.iut-bm.univ-fcomte.fr/pub/gitweb/simgrid.git/blobdiff_plain/585afb9fd33ca7e70cef86a2a3744598f11a0888..8eb1dd81f53db215cf73319dc485f120addefe2d:/src/smpi/smpi_mpi_dt.c diff --git a/src/smpi/smpi_mpi_dt.c b/src/smpi/smpi_mpi_dt.c index 35707fa93f..22be36c59c 100644 --- a/src/smpi/smpi_mpi_dt.c +++ b/src/smpi/smpi_mpi_dt.c @@ -141,34 +141,22 @@ MPI_Aint smpi_datatype_ub(MPI_Datatype datatype) int smpi_datatype_extent(MPI_Datatype datatype, MPI_Aint * lb, MPI_Aint * extent) { - int retval; - -// if ((datatype->flags & DT_FLAG_COMMITED) != DT_FLAG_COMMITED) { -// retval = MPI_ERR_TYPE; -// } else { - *lb = datatype->lb; - *extent = datatype->ub - datatype->lb; - retval = MPI_SUCCESS; -// } - return retval; + *lb = datatype->lb; + *extent = datatype->ub - datatype->lb; + return MPI_SUCCESS; } MPI_Aint smpi_datatype_get_extent(MPI_Datatype datatype){ - MPI_Aint lb; - MPI_Aint extent; - smpi_datatype_extent(datatype, &lb, &extent); - return extent; + return datatype->ub - datatype->lb; } int smpi_datatype_copy(void *sendbuf, int sendcount, MPI_Datatype sendtype, void *recvbuf, int recvcount, MPI_Datatype recvtype) { - int retval, count; + int count; /* First check if we really have something to do */ - if (recvcount == 0) { - retval = sendcount == 0 ? MPI_SUCCESS : MPI_ERR_TRUNCATE; - } else { + if (recvcount > 0 && recvbuf != sendbuf) { /* FIXME: treat packed cases */ sendcount *= smpi_datatype_size(sendtype); recvcount *= smpi_datatype_size(recvtype); @@ -189,20 +177,18 @@ int smpi_datatype_copy(void *sendbuf, int sendcount, MPI_Datatype sendtype, }else{ s_smpi_subtype_t *subtype = sendtype->substruct; - s_smpi_mpi_vector_t* type_c = (s_smpi_mpi_vector_t*)sendtype; - void * buf_tmp = malloc(count * type_c->size_oldtype); + void * buf_tmp = xbt_malloc(count); subtype->serialize( sendbuf, buf_tmp,1, subtype); subtype = recvtype->substruct; - subtype->unserialize(recvbuf, buf_tmp,1, subtype); + subtype->unserialize( buf_tmp, recvbuf,1, subtype); free(buf_tmp); } - retval = sendcount > recvcount ? MPI_ERR_TRUNCATE : MPI_SUCCESS; } - return retval; + return sendcount > recvcount ? MPI_ERR_TRUNCATE : MPI_SUCCESS; } /* @@ -296,31 +282,141 @@ 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 extent, int has_subtype, +void smpi_datatype_create(MPI_Datatype* new_type, int size,int lb, int ub, 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 = extent; + new_t->has_subtype = size>0? has_subtype:0; + new_t->lb = lb; + new_t->ub = ub; new_t->flags = flags; new_t->substruct = struct_type; + new_t->in_use=0; *new_type = new_t; } void smpi_datatype_free(MPI_Datatype* type){ + + if((*type)->flags & DT_FLAG_PREDEFINED)return; + + //if still used, mark for deletion + if((*type)->in_use!=0){ + (*type)->flags |=DT_FLAG_DESTROYED; + return; + } + if ((*type)->has_subtype == 1){ ((s_smpi_subtype_t *)(*type)->substruct)->subtype_free(type); + xbt_free((*type)->substruct); } xbt_free(*type); + } -int smpi_datatype_contiguous(int count, MPI_Datatype old_type, MPI_Datatype* new_type) +void smpi_datatype_use(MPI_Datatype type){ + if(type)type->in_use++; +} + + +void smpi_datatype_unuse(MPI_Datatype type){ + if(type && type->in_use-- == 0 && (type->flags & DT_FLAG_DESTROYED)) + smpi_datatype_free(&type); +} + + + + +/* +Contiguous Implementation +*/ + + +/* + * Copies noncontiguous data into contiguous memory. + * @param contiguous_hvector - output hvector + * @param noncontiguous_hvector - input hvector + * @param type - pointer contening : + * - stride - stride of between noncontiguous data, in bytes + * - 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, + size_t 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); +} +/* + * Copies contiguous data into noncontiguous memory. + * @param noncontiguous_vector - output hvector + * @param contiguous_vector - input hvector + * @param type - pointer contening : + * - stride - stride of between noncontiguous data, in bytes + * - 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, + size_t count, + void *type) +{ + 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; + + memcpy(noncontiguous_vector_char, + contiguous_vector_char, count* type_c->block_count * type_c->size_oldtype); +} + +void free_contiguous(MPI_Datatype* d){ +} + +/* + * 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, + 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->lb = lb; + new_t->block_count = block_count; + new_t->old_type = old_type; + new_t->size_oldtype = size_oldtype; + return new_t; +} + + + + +int smpi_datatype_contiguous(int count, MPI_Datatype old_type, MPI_Datatype* new_type, MPI_Aint lb) { int retval; - smpi_datatype_create(new_type, count * - smpi_datatype_size(old_type),count * - smpi_datatype_size(old_type),0,NULL, DT_FLAG_CONTIGUOUS); + if(old_type->has_subtype){ + //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); retval=MPI_SUCCESS; return retval; } @@ -328,16 +424,24 @@ int smpi_datatype_contiguous(int count, MPI_Datatype old_type, MPI_Datatype* new int smpi_datatype_vector(int count, int blocklen, int stride, MPI_Datatype old_type, MPI_Datatype* new_type) { int retval; - if (blocklen<=0) return MPI_ERR_ARG; + if (blocklen<0) return MPI_ERR_ARG; + MPI_Aint lb = 0; + MPI_Aint ub = 0; + if(count>0){ + 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){ + + 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), - ((count -1) * stride + blocklen) * smpi_datatype_get_extent(old_type), + count * (blocklen) * smpi_datatype_size(old_type), lb, + ub, 1, subtype, DT_FLAG_VECTOR); @@ -346,7 +450,7 @@ int smpi_datatype_vector(int count, int blocklen, int stride, MPI_Datatype old_t /* 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), ((count -1) * stride + blocklen)* + smpi_datatype_size(old_type), 0, ((count -1) * stride + blocklen)* smpi_datatype_size(old_type), 0, NULL, @@ -461,7 +565,13 @@ void free_hvector(MPI_Datatype* d){ int smpi_datatype_hvector(int count, int blocklen, MPI_Aint stride, MPI_Datatype old_type, MPI_Datatype* new_type) { int retval; - if (blocklen<=0) return MPI_ERR_ARG; + if (blocklen<0) return MPI_ERR_ARG; + MPI_Aint lb = 0; + MPI_Aint ub = 0; + if(count>0){ + 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, @@ -469,16 +579,15 @@ int smpi_datatype_hvector(int count, int blocklen, MPI_Aint stride, MPI_Datatype old_type, smpi_datatype_size(old_type)); - smpi_datatype_create(new_type, count * blocklen * - smpi_datatype_size(old_type), (count-1) * stride + blocklen * - smpi_datatype_size(old_type), + smpi_datatype_create(new_type, count * blocklen * smpi_datatype_size(old_type), + lb,ub, 1, subtype, DT_FLAG_VECTOR); retval=MPI_SUCCESS; }else{ smpi_datatype_create(new_type, count * blocklen * - smpi_datatype_size(old_type),count * blocklen * + smpi_datatype_size(old_type),0,count * blocklen * smpi_datatype_size(old_type), 0, NULL, @@ -510,7 +619,7 @@ void serialize_indexed( const void *noncontiguous_indexed, 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; + 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) @@ -544,15 +653,15 @@ void unserialize_indexed( const void *contiguous_indexed, size_t count, void *type) { + 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; + 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) - memcpy(noncontiguous_indexed_char, + memcpy(noncontiguous_indexed_char , contiguous_indexed_char, type_c->block_lengths[i] * type_c->size_oldtype); else ((s_smpi_subtype_t*)type_c->old_type->substruct)->unserialize( contiguous_indexed_char, @@ -609,11 +718,23 @@ int smpi_datatype_indexed(int count, int* blocklens, int* indices, MPI_Datatype int retval; int size = 0; int contiguous=1; + MPI_Aint lb = 0; + MPI_Aint ub = 0; + if(count>0){ + lb=indices[0]*smpi_datatype_get_extent(old_type); + ub=indices[0]*smpi_datatype_get_extent(old_type) + blocklens[0]*smpi_datatype_ub(old_type); + } + for(i=0; i< count; i++){ - if (blocklens[i]<=0) + if (blocklens[i]<0) return MPI_ERR_ARG; size += blocklens[i]; + if(indices[i]*smpi_datatype_get_extent(old_type)+smpi_datatype_lb(old_type)ub) + ub = indices[i]*smpi_datatype_get_extent(old_type)+blocklens[i]*smpi_datatype_ub(old_type); + if ( (i< count -1) && (indices[i]+blocklens[i] != indices[i+1]) )contiguous=0; } if (old_type->has_subtype == 1) @@ -626,11 +747,14 @@ int smpi_datatype_indexed(int count, int* blocklens, int* indices, MPI_Datatype old_type, smpi_datatype_size(old_type)); smpi_datatype_create(new_type, size * - smpi_datatype_size(old_type),(indices[count-1]+blocklens[count-1])*smpi_datatype_size(old_type),1, subtype, DT_FLAG_DATA); + smpi_datatype_size(old_type),lb,ub,1, 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),size * - smpi_datatype_size(old_type),0, NULL, DT_FLAG_DATA|DT_FLAG_CONTIGUOUS); + smpi_datatype_size(old_type),lb,ub,1, subtype, DT_FLAG_DATA|DT_FLAG_CONTIGUOUS); } retval=MPI_SUCCESS; return retval; @@ -658,7 +782,7 @@ void serialize_hindexed( const void *noncontiguous_hindexed, s_smpi_mpi_hindexed_t* type_c = (s_smpi_mpi_hindexed_t*)type; int i,j; char* contiguous_hindexed_char = (char*)contiguous_hindexed; - char* noncontiguous_hindexed_char = (char*)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) @@ -695,7 +819,7 @@ void unserialize_hindexed( const void *contiguous_hindexed, int i,j; char* contiguous_hindexed_char = (char*)contiguous_hindexed; - char* noncontiguous_hindexed_char = (char*)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) @@ -755,27 +879,43 @@ int smpi_datatype_hindexed(int count, int* blocklens, MPI_Aint* indices, MPI_Dat int retval; int size = 0; int contiguous=1; + MPI_Aint lb = 0; + MPI_Aint ub = 0; + if(count>0){ + lb=indices[0] + smpi_datatype_lb(old_type); + ub=indices[0] + blocklens[0]*smpi_datatype_ub(old_type); + } for(i=0; i< count; i++){ - if (blocklens[i]<=0) + if (blocklens[i]<0) return MPI_ERR_ARG; size += blocklens[i]; + + if(indices[i]+smpi_datatype_lb(old_type)ub) ub = indices[i]+blocklens[i]*smpi_datatype_ub(old_type); + if ( (i< count -1) && (indices[i]+blocklens[i]*smpi_datatype_size(old_type) != indices[i+1]) )contiguous=0; } - if (old_type->has_subtype == 1) + if (old_type->has_subtype == 1 || lb!=0) contiguous=0; + 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),indices[count-1]+blocklens[count-1]*smpi_datatype_size(old_type) + smpi_datatype_create(new_type, size * smpi_datatype_size(old_type), + lb, + ub ,1, subtype, DT_FLAG_DATA); }else{ - smpi_datatype_create(new_type, size * - smpi_datatype_size(old_type),size * - smpi_datatype_size(old_type),0, NULL, DT_FLAG_DATA|DT_FLAG_CONTIGUOUS); + 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; @@ -803,7 +943,7 @@ void serialize_struct( const void *noncontiguous_struct, s_smpi_mpi_struct_t* type_c = (s_smpi_mpi_struct_t*)type; int i,j; char* contiguous_struct_char = (char*)contiguous_struct; - char* noncontiguous_struct_char = (char*)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) @@ -841,7 +981,7 @@ void unserialize_struct( const void *contiguous_struct, int i,j; char* contiguous_struct_char = (char*)contiguous_struct; - char* noncontiguous_struct_char = (char*)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) @@ -905,12 +1045,32 @@ int smpi_datatype_struct(int count, int* blocklens, MPI_Aint* indices, MPI_Datat size_t size = 0; int contiguous=1; size = 0; + MPI_Aint lb = 0; + MPI_Aint ub = 0; + if(count>0){ + lb=indices[0] + smpi_datatype_lb(old_types[0]); + ub=indices[0] + blocklens[0]*smpi_datatype_ub(old_types[0]); + } + int forced_lb=0; + int forced_ub=0; for(i=0; i< count; i++){ - if (blocklens[i]<=0) + if (blocklens[i]<0) return MPI_ERR_ARG; if (old_types[i]->has_subtype == 1) contiguous=0; + size += blocklens[i]*smpi_datatype_size(old_types[i]); + if (old_types[i]==MPI_LB){ + lb=indices[i]; + forced_lb=1; + } + if (old_types[i]==MPI_UB){ + ub=indices[i]; + forced_ub=1; + } + + if(!forced_lb && indices[i]+smpi_datatype_lb(old_types[i])ub) ub = indices[i]+blocklens[i]*smpi_datatype_ub(old_types[i]); if ( (i< count -1) && (indices[i]+blocklens[i]*smpi_datatype_size(old_types[i]) != indices[i+1]) )contiguous=0; } @@ -921,9 +1081,13 @@ int smpi_datatype_struct(int count, int* blocklens, MPI_Aint* indices, MPI_Datat count, old_types); - smpi_datatype_create(new_type, size, indices[count-1] + blocklens[count-1]*smpi_datatype_size(old_types[count-1]),1, subtype, DT_FLAG_DATA); + smpi_datatype_create(new_type, size, lb, ub,1, subtype, DT_FLAG_DATA); }else{ - smpi_datatype_create(new_type, size, indices[count-1] + blocklens[count-1]*smpi_datatype_size(old_types[count-1]),0, NULL, DT_FLAG_DATA|DT_FLAG_CONTIGUOUS); + 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; } @@ -935,6 +1099,7 @@ void smpi_datatype_commit(MPI_Datatype *datatype) typedef struct s_smpi_mpi_op { MPI_User_function *func; + int is_commute; } s_smpi_mpi_op_t; #define MAX_OP(a, b) (b) = (a) < (b) ? (b) : (a) @@ -1256,7 +1421,7 @@ static void maxloc_func(void *a, void *b, int *length, #define CREATE_MPI_OP(name, func) \ - static s_smpi_mpi_op_t mpi_##name = { &(func) /* func */ }; \ + static s_smpi_mpi_op_t mpi_##name = { &(func) /* func */, TRUE }; \ MPI_Op name = &mpi_##name; CREATE_MPI_OP(MPI_MAX, max_func); @@ -1275,13 +1440,17 @@ CREATE_MPI_OP(MPI_MINLOC, minloc_func); MPI_Op smpi_op_new(MPI_User_function * function, int commute) { MPI_Op op; - - //FIXME: add commute param op = xbt_new(s_smpi_mpi_op_t, 1); op->func = function; + op-> is_commute = commute; return op; } +int smpi_op_is_commute(MPI_Op op) +{ + return op-> is_commute; +} + void smpi_op_destroy(MPI_Op op) { xbt_free(op);