X-Git-Url: http://info.iut-bm.univ-fcomte.fr/pub/gitweb/simgrid.git/blobdiff_plain/b7ed19dfcc221d7b3eca182abb5c4a3946671172..5f1dc33c9f76ee99973ba93f034f031451398ebe:/src/smpi/colls/smpi_mvapich2_selector.cpp diff --git a/src/smpi/colls/smpi_mvapich2_selector.cpp b/src/smpi/colls/smpi_mvapich2_selector.cpp index bda8db4287..1d6978c3a1 100644 --- a/src/smpi/colls/smpi_mvapich2_selector.cpp +++ b/src/smpi/colls/smpi_mvapich2_selector.cpp @@ -1,6 +1,6 @@ /* selector for collective algorithms based on mvapich decision logic */ -/* Copyright (c) 2009-2010, 2013-2014. The SimGrid Team. +/* Copyright (c) 2009-2010, 2013-2017. The SimGrid Team. * All rights reserved. */ /* This program is free software; you can redistribute it and/or modify it @@ -10,12 +10,11 @@ #include "smpi_mvapich2_selector_stampede.h" - namespace simgrid{ namespace smpi{ -int Coll_alltoall_mvapich2::alltoall( void *sendbuf, int sendcount, +int Coll_alltoall_mvapich2::alltoall( void *sendbuf, int sendcount, MPI_Datatype sendtype, void* recvbuf, int recvcount, MPI_Datatype recvtype, @@ -65,7 +64,7 @@ int Coll_alltoall_mvapich2::alltoall( void *sendbuf, int sendcount, ||nbytes > mv2_alltoall_thresholds_table[conf_index][range].in_place_algo_table[range_threshold].max ) { tmp_buf = (char *)smpi_get_tmp_sendbuffer( comm_size * recvcount * recvtype_size ); - mpi_errno = Datatype::copy((char *)recvbuf, + Datatype::copy((char *)recvbuf, comm_size*recvcount, recvtype, (char *)tmp_buf, comm_size*recvcount, recvtype); @@ -107,7 +106,7 @@ int Coll_allgather_mvapich2::allgather(void *sendbuf, int sendcount, MPI_Datatyp if(mv2_allgather_table_ppn_conf==NULL) init_mv2_allgather_tables_stampede(); - + if(comm->get_leaders_comm()==MPI_COMM_NULL){ comm->init_smp(); } @@ -135,7 +134,7 @@ int Coll_allgather_mvapich2::allgather(void *sendbuf, int sendcount, MPI_Datatyp if (partial_sub_ok != 1) { conf_index = 0; } - + /* Search for the corresponding system size inside the tuning table */ while ((range < (mv2_size_allgather_tuning_table[conf_index] - 1)) && (comm_size > @@ -240,17 +239,17 @@ int Coll_gather_mvapich2::gather(void *sendbuf, -1)) { range_intra_threshold++; } - + if (comm->is_blocked() ) { - // Set intra-node function pt for gather_two_level - MV2_Gather_intra_node_function = + // Set intra-node function pt for gather_two_level + MV2_Gather_intra_node_function = mv2_gather_thresholds_table[range].intra_node[range_intra_threshold]. MV2_pt_Gather_function; - //Set inter-leader pt + //Set inter-leader pt MV2_Gather_inter_leader_function = mv2_gather_thresholds_table[range].inter_leader[range_threshold]. MV2_pt_Gather_function; - // We call Gather function + // We call Gather function mpi_errno = MV2_Gather_inter_leader_function(sendbuf, sendcnt, sendtype, recvbuf, recvcnt, recvtype, root, comm); @@ -305,13 +304,9 @@ int Coll_allgatherv_mvapich2::allgatherv(void *sendbuf, int sendcount, MPI_Datat if (MV2_Allgatherv_function == &MPIR_Allgatherv_Rec_Doubling_MV2) { - if(!(comm_size & (comm_size - 1))) - { - mpi_errno = - MPIR_Allgatherv_Rec_Doubling_MV2(sendbuf, sendcount, - sendtype, recvbuf, - recvcounts, displs, - recvtype, comm); + if (not(comm_size & (comm_size - 1))) { + mpi_errno = + MPIR_Allgatherv_Rec_Doubling_MV2(sendbuf, sendcount, sendtype, recvbuf, recvcounts, displs, recvtype, comm); } else { mpi_errno = MPIR_Allgatherv_Bruck_MV2(sendbuf, sendcount, @@ -470,7 +465,7 @@ int Coll_alltoallv_mvapich2::alltoallv(void *sbuf, int *scounts, int *sdisps, int Coll_barrier_mvapich2::barrier(MPI_Comm comm) -{ +{ return Coll_barrier_mvapich2_pair::barrier(comm); } @@ -485,7 +480,7 @@ int Coll_bcast_mvapich2::bcast(void *buffer, int mpi_errno = MPI_SUCCESS; int comm_size/*, rank*/; int two_level_bcast = 1; - long nbytes = 0; + long nbytes = 0; int range = 0; int range_threshold = 0; int range_threshold_intra = 0; @@ -501,7 +496,7 @@ int Coll_bcast_mvapich2::bcast(void *buffer, if(comm->get_leaders_comm()==MPI_COMM_NULL){ comm->init_smp(); } - if(!mv2_bcast_thresholds_table) + if (not mv2_bcast_thresholds_table) init_mv2_bcast_tables_stampede(); comm_size = comm->size(); //rank = comm->rank(); @@ -569,7 +564,7 @@ int Coll_bcast_mvapich2::bcast(void *buffer, if (mv2_bcast_thresholds_table[range].inter_leader[range_threshold]. zcpy_pipelined_knomial_factor != -1) { - zcpy_knomial_factor = + zcpy_knomial_factor = mv2_bcast_thresholds_table[range].inter_leader[range_threshold]. zcpy_pipelined_knomial_factor; } @@ -586,7 +581,7 @@ int Coll_bcast_mvapich2::bcast(void *buffer, /* Set value of pipeline segment size */ bcast_segment_size = mv2_bcast_thresholds_table[range].bcast_segment_size; - + /* Set value of inter node knomial factor */ mv2_inter_node_knomial_factor = mv2_bcast_thresholds_table[range].inter_node_knomial_factor; @@ -596,69 +591,62 @@ int Coll_bcast_mvapich2::bcast(void *buffer, /* Check if we will use a two level algorithm or not */ two_level_bcast = #if defined(_MCST_SUPPORT_) - mv2_bcast_thresholds_table[range].is_two_level_bcast[range_threshold] + mv2_bcast_thresholds_table[range].is_two_level_bcast[range_threshold] || comm->ch.is_mcast_ok; #else mv2_bcast_thresholds_table[range].is_two_level_bcast[range_threshold]; #endif if (two_level_bcast == 1) { - if (!is_contig || !is_homogeneous) { - tmp_buf=(void *)smpi_get_tmp_sendbuffer(nbytes); - -/* position = 0;*/ -/* if (rank == root) {*/ -/* mpi_errno =*/ -/* MPIR_Pack_impl(buffer, count, datatype, tmp_buf, nbytes, &position);*/ -/* if (mpi_errno)*/ -/* MPIU_ERR_POP(mpi_errno);*/ -/* }*/ + if (not is_contig || not is_homogeneous) { + tmp_buf = (void*)smpi_get_tmp_sendbuffer(nbytes); + + /* position = 0;*/ + /* if (rank == root) {*/ + /* mpi_errno =*/ + /* MPIR_Pack_impl(buffer, count, datatype, tmp_buf, nbytes, &position);*/ + /* if (mpi_errno)*/ + /* MPIU_ERR_POP(mpi_errno);*/ + /* }*/ } #ifdef CHANNEL_MRAIL_GEN2 if ((mv2_enable_zcpy_bcast == 1) && - (&MPIR_Pipelined_Bcast_Zcpy_MV2 == MV2_Bcast_function)) { - if (!is_contig || !is_homogeneous) { - mpi_errno = MPIR_Pipelined_Bcast_Zcpy_MV2(tmp_buf, nbytes, MPI_BYTE, - root, comm); - } else { + (&MPIR_Pipelined_Bcast_Zcpy_MV2 == MV2_Bcast_function)) { + if (not is_contig || not is_homogeneous) { + mpi_errno = MPIR_Pipelined_Bcast_Zcpy_MV2(tmp_buf, nbytes, MPI_BYTE, root, comm); + } else { mpi_errno = MPIR_Pipelined_Bcast_Zcpy_MV2(buffer, count, datatype, root, comm); - } - } else + } + } else #endif /* defined(CHANNEL_MRAIL_GEN2) */ - { + { shmem_comm = comm->get_intra_comm(); - if (!is_contig || !is_homogeneous) { - mpi_errno = - MPIR_Bcast_tune_inter_node_helper_MV2(tmp_buf, nbytes, MPI_BYTE, - root, comm); + if (not is_contig || not is_homogeneous) { + MPIR_Bcast_tune_inter_node_helper_MV2(tmp_buf, nbytes, MPI_BYTE, root, comm); } else { - mpi_errno = - MPIR_Bcast_tune_inter_node_helper_MV2(buffer, count, datatype, root, - comm); + MPIR_Bcast_tune_inter_node_helper_MV2(buffer, count, datatype, root, comm); } /* We are now done with the inter-node phase */ root = INTRA_NODE_ROOT; - - if (!is_contig || !is_homogeneous) { - mpi_errno = MV2_Bcast_intra_node_function(tmp_buf, nbytes, - MPI_BYTE, root, shmem_comm); + if (not is_contig || not is_homogeneous) { + mpi_errno = MV2_Bcast_intra_node_function(tmp_buf, nbytes, MPI_BYTE, root, shmem_comm); } else { mpi_errno = MV2_Bcast_intra_node_function(buffer, count, datatype, root, shmem_comm); } - } -/* if (!is_contig || !is_homogeneous) {*/ -/* if (rank != root) {*/ -/* position = 0;*/ -/* mpi_errno = MPIR_Unpack_impl(tmp_buf, nbytes, &position, buffer,*/ -/* count, datatype);*/ -/* }*/ -/* }*/ + } + /* if (not is_contig || not is_homogeneous) {*/ + /* if (rank != root) {*/ + /* position = 0;*/ + /* mpi_errno = MPIR_Unpack_impl(tmp_buf, nbytes, &position, buffer,*/ + /* count, datatype);*/ + /* }*/ + /* }*/ } else { /* We use Knomial for intra node */ MV2_Bcast_intra_node_function = &MPIR_Knomial_Bcast_intra_node_MV2; @@ -762,7 +750,7 @@ int Coll_reduce_mvapich2::reduce( void *sendbuf, if(comm->get_leaders_comm()==MPI_COMM_NULL){ comm->init_smp(); } - mpi_errno = MPIR_Reduce_two_level_helper_MV2(sendbuf, recvbuf, count, + mpi_errno = MPIR_Reduce_two_level_helper_MV2(sendbuf, recvbuf, count, datatype, op, root, comm); } else { mpi_errno = MPIR_Reduce_binomial_MV2(sendbuf, recvbuf, count, @@ -771,7 +759,7 @@ int Coll_reduce_mvapich2::reduce( void *sendbuf, } else if(MV2_Reduce_function == &MPIR_Reduce_inter_knomial_wrapper_MV2 ){ if(is_commutative ==1) { - mpi_errno = MV2_Reduce_function(sendbuf, recvbuf, count, + mpi_errno = MV2_Reduce_function(sendbuf, recvbuf, count, datatype, op, root, comm); } else { mpi_errno = MPIR_Reduce_binomial_MV2(sendbuf, recvbuf, count, @@ -780,14 +768,14 @@ int Coll_reduce_mvapich2::reduce( void *sendbuf, } else if(MV2_Reduce_function == &MPIR_Reduce_redscat_gather_MV2){ if (/*(HANDLE_GET_KIND(op) == HANDLE_KIND_BUILTIN) &&*/ (count >= pof2)) { - mpi_errno = MV2_Reduce_function(sendbuf, recvbuf, count, + mpi_errno = MV2_Reduce_function(sendbuf, recvbuf, count, datatype, op, root, comm); } else { mpi_errno = MPIR_Reduce_binomial_MV2(sendbuf, recvbuf, count, datatype, op, root, comm); } } else { - mpi_errno = MV2_Reduce_function(sendbuf, recvbuf, count, + mpi_errno = MV2_Reduce_function(sendbuf, recvbuf, count, datatype, op, root, comm); } @@ -857,7 +845,7 @@ int Coll_reduce_scatter_mvapich2::reduce_scatter(void *sendbuf, void *recvbuf, i while (pof2 < comm_size) pof2 <<= 1; if (pof2 == comm_size && is_block_regular) { /* noncommutative, pof2 size, and block regular */ - mpi_errno = MPIR_Reduce_scatter_non_comm_MV2(sendbuf, recvbuf, + MPIR_Reduce_scatter_non_comm_MV2(sendbuf, recvbuf, recvcnts, datatype, op, comm); } @@ -897,7 +885,7 @@ int Coll_scatter_mvapich2::scatter(void *sendbuf, if(comm->get_leaders_comm()==MPI_COMM_NULL){ comm->init_smp(); } - + comm_size = comm->size(); rank = comm->rank(); @@ -909,15 +897,15 @@ int Coll_scatter_mvapich2::scatter(void *sendbuf, recvtype_size=recvtype->size(); nbytes = recvcnt * recvtype_size; } - - // check if safe to use partial subscription mode + + // check if safe to use partial subscription mode if (comm->is_uniform()) { shmem_comm = comm->get_intra_comm(); local_size = shmem_comm->size(); i = 0; if (mv2_scatter_table_ppn_conf[0] == -1) { - // Indicating user defined tuning + // Indicating user defined tuning conf_index = 0; }else{ do { @@ -930,7 +918,7 @@ int Coll_scatter_mvapich2::scatter(void *sendbuf, } while(i < mv2_scatter_num_ppn_conf); } } - + if (partial_sub_ok != 1) { conf_index = 0; } @@ -978,7 +966,7 @@ int Coll_scatter_mvapich2::scatter(void *sendbuf, /* Fallback! */ MV2_Scatter_function = &MPIR_Scatter_MV2_Binomial; } - } + } } if( (MV2_Scatter_function == &MPIR_Scatter_MV2_two_level_Direct) || @@ -1007,29 +995,31 @@ int Coll_scatter_mvapich2::scatter(void *sendbuf, } } -void smpi_coll_cleanup_mvapich2(void){ -int i=0; -if(mv2_alltoall_thresholds_table) - xbt_free(mv2_alltoall_thresholds_table[i]); -xbt_free(mv2_alltoall_thresholds_table); -xbt_free(mv2_size_alltoall_tuning_table); -xbt_free(mv2_alltoall_table_ppn_conf); - -xbt_free(mv2_gather_thresholds_table); -if(mv2_allgather_thresholds_table) - xbt_free(mv2_allgather_thresholds_table[0]); -xbt_free(mv2_size_allgather_tuning_table); -xbt_free(mv2_allgather_table_ppn_conf); -xbt_free(mv2_allgather_thresholds_table); - -xbt_free(mv2_allgatherv_thresholds_table); -xbt_free(mv2_reduce_thresholds_table); -xbt_free(mv2_red_scat_thresholds_table); -xbt_free(mv2_allreduce_thresholds_table); -xbt_free(mv2_bcast_thresholds_table); -if(mv2_scatter_thresholds_table) - xbt_free(mv2_scatter_thresholds_table[0]); -xbt_free(mv2_scatter_thresholds_table); -xbt_free(mv2_size_scatter_tuning_table); -xbt_free(mv2_scatter_table_ppn_conf); + +void smpi_coll_cleanup_mvapich2() +{ + int i = 0; + if (mv2_alltoall_thresholds_table) + xbt_free(mv2_alltoall_thresholds_table[i]); + xbt_free(mv2_alltoall_thresholds_table); + xbt_free(mv2_size_alltoall_tuning_table); + xbt_free(mv2_alltoall_table_ppn_conf); + + xbt_free(mv2_gather_thresholds_table); + if (mv2_allgather_thresholds_table) + xbt_free(mv2_allgather_thresholds_table[0]); + xbt_free(mv2_size_allgather_tuning_table); + xbt_free(mv2_allgather_table_ppn_conf); + xbt_free(mv2_allgather_thresholds_table); + + xbt_free(mv2_allgatherv_thresholds_table); + xbt_free(mv2_reduce_thresholds_table); + xbt_free(mv2_red_scat_thresholds_table); + xbt_free(mv2_allreduce_thresholds_table); + xbt_free(mv2_bcast_thresholds_table); + if (mv2_scatter_thresholds_table) + xbt_free(mv2_scatter_thresholds_table[0]); + xbt_free(mv2_scatter_thresholds_table); + xbt_free(mv2_size_scatter_tuning_table); + xbt_free(mv2_scatter_table_ppn_conf); }