X-Git-Url: http://info.iut-bm.univ-fcomte.fr/pub/gitweb/simgrid.git/blobdiff_plain/61427a88a76a2c2ef25d0d5b8132995c6f550e5c..1bf13b42c09009ec33fd1928b8bffd2ded6bb931:/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 837584f694..2128f77c74 100644 --- a/src/smpi/colls/smpi_mvapich2_selector.cpp +++ b/src/smpi/colls/smpi_mvapich2_selector.cpp @@ -1,31 +1,30 @@ /* selector for collective algorithms based on mvapich decision logic */ -/* Copyright (c) 2009-2010, 2013-2017. The SimGrid Team. +/* Copyright (c) 2009-2019. The SimGrid Team. * All rights reserved. */ /* This program is free software; you can redistribute it and/or modify it * under the terms of the license (GNU LGPL) which comes with this package. */ -#include "colls_private.h" +#include "colls_private.hpp" -#include "smpi_mvapich2_selector_stampede.h" +#include "smpi_mvapich2_selector_stampede.hpp" -namespace simgrid{ -namespace smpi{ +namespace simgrid { +namespace smpi { -int Coll_alltoall_mvapich2::alltoall( void *sendbuf, int sendcount, - MPI_Datatype sendtype, - void* recvbuf, int recvcount, - MPI_Datatype recvtype, - MPI_Comm comm) +int alltoall__mvapich2( const void *sendbuf, int sendcount, + MPI_Datatype sendtype, + void* recvbuf, int recvcount, + MPI_Datatype recvtype, + MPI_Comm comm) { if(mv2_alltoall_table_ppn_conf==NULL) init_mv2_alltoall_tables_stampede(); int sendtype_size, recvtype_size, comm_size; - char * tmp_buf = NULL; int mpi_errno=MPI_SUCCESS; int range = 0; int range_threshold = 0; @@ -63,16 +62,11 @@ int Coll_alltoall_mvapich2::alltoall( void *sendbuf, int sendcount, mv2_alltoall_thresholds_table[conf_index][range].in_place_algo_table[range_threshold].min ||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, - comm_size*recvcount, recvtype, - (char *)tmp_buf, - comm_size*recvcount, recvtype); + unsigned char* tmp_buf = smpi_get_tmp_sendbuffer(comm_size * recvcount * recvtype_size); + Datatype::copy(recvbuf, comm_size * recvcount, recvtype, tmp_buf, comm_size * recvcount, recvtype); - mpi_errno = MV2_Alltoall_function(tmp_buf, recvcount, recvtype, - recvbuf, recvcount, recvtype, - comm ); - smpi_free_tmp_buffer(tmp_buf); + mpi_errno = MV2_Alltoall_function(tmp_buf, recvcount, recvtype, recvbuf, recvcount, recvtype, comm); + smpi_free_tmp_buffer(tmp_buf); } else { mpi_errno = MPIR_Alltoall_inplace_MV2(sendbuf, sendcount, sendtype, recvbuf, recvcount, recvtype, @@ -84,7 +78,7 @@ int Coll_alltoall_mvapich2::alltoall( void *sendbuf, int sendcount, return (mpi_errno); } -int Coll_allgather_mvapich2::allgather(void *sendbuf, int sendcount, MPI_Datatype sendtype, +int allgather__mvapich2(const void *sendbuf, int sendcount, MPI_Datatype sendtype, void *recvbuf, int recvcount, MPI_Datatype recvtype, MPI_Comm comm) { @@ -96,7 +90,6 @@ int Coll_allgather_mvapich2::allgather(void *sendbuf, int sendcount, MPI_Datatyp int conf_index = 0; int range_threshold = 0; int is_two_level = 0; - int local_size = -1; MPI_Comm shmem_comm; //MPI_Comm *shmem_commptr=NULL; /* Get the size of the communicator */ @@ -111,11 +104,10 @@ int Coll_allgather_mvapich2::allgather(void *sendbuf, int sendcount, MPI_Datatyp comm->init_smp(); } - int i; if (comm->is_uniform()){ shmem_comm = comm->get_intra_comm(); - local_size = shmem_comm->size(); - i = 0; + int local_size = shmem_comm->size(); + int i = 0; if (mv2_allgather_table_ppn_conf[0] == -1) { // Indicating user defined tuning conf_index = 0; @@ -165,7 +157,7 @@ int Coll_allgather_mvapich2::allgather(void *sendbuf, int sendcount, MPI_Datatyp recvbuf, recvcount, recvtype, comm); }else{ - mpi_errno = Coll_allgather_mpich::allgather(sendbuf, sendcount, sendtype, + mpi_errno = allgather__mpich(sendbuf, sendcount, sendtype, recvbuf, recvcount, recvtype, comm); } @@ -187,7 +179,7 @@ int Coll_allgather_mvapich2::allgather(void *sendbuf, int sendcount, MPI_Datatyp return mpi_errno; } -int Coll_gather_mvapich2::gather(void *sendbuf, +int gather__mvapich2(const void *sendbuf, int sendcnt, MPI_Datatype sendtype, void *recvbuf, @@ -203,18 +195,15 @@ int Coll_gather_mvapich2::gather(void *sendbuf, int range_threshold = 0; int range_intra_threshold = 0; long nbytes = 0; - int comm_size = 0; - int recvtype_size, sendtype_size; - int rank = -1; - comm_size = comm->size(); - rank = comm->rank(); + int comm_size = comm->size(); + int rank = comm->rank(); if (rank == root) { - recvtype_size=recvtype->size(); - nbytes = recvcnt * recvtype_size; + int recvtype_size = recvtype->size(); + nbytes = recvcnt * recvtype_size; } else { - sendtype_size=sendtype->size(); - nbytes = sendcnt * sendtype_size; + int sendtype_size = sendtype->size(); + nbytes = sendcnt * sendtype_size; } /* Search for the corresponding system size inside the tuning table */ @@ -256,7 +245,7 @@ int Coll_gather_mvapich2::gather(void *sendbuf, } else { // Indeed, direct (non SMP-aware)gather is MPICH one - mpi_errno = Coll_gather_mpich::gather(sendbuf, sendcnt, sendtype, + mpi_errno = gather__mpich(sendbuf, sendcnt, sendtype, recvbuf, recvcnt, recvtype, root, comm); } @@ -264,8 +253,8 @@ int Coll_gather_mvapich2::gather(void *sendbuf, return mpi_errno; } -int Coll_allgatherv_mvapich2::allgatherv(void *sendbuf, int sendcount, MPI_Datatype sendtype, - void *recvbuf, int *recvcounts, int *displs, +int allgatherv__mvapich2(const void *sendbuf, int sendcount, MPI_Datatype sendtype, + void *recvbuf, const int *recvcounts, const int *displs, MPI_Datatype recvtype, MPI_Comm comm ) { int mpi_errno = MPI_SUCCESS; @@ -326,7 +315,7 @@ int Coll_allgatherv_mvapich2::allgatherv(void *sendbuf, int sendcount, MPI_Datat -int Coll_allreduce_mvapich2::allreduce(void *sendbuf, +int allreduce__mvapich2(const void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype, @@ -351,8 +340,6 @@ int Coll_allreduce_mvapich2::allreduce(void *sendbuf, MPI_Aint sendtype_size = 0; long nbytes = 0; - int range = 0, range_threshold = 0, range_threshold_intra = 0; - int is_two_level = 0; int is_commutative = 0; MPI_Aint true_lb, true_extent; @@ -360,17 +347,19 @@ int Coll_allreduce_mvapich2::allreduce(void *sendbuf, nbytes = count * sendtype_size; datatype->extent(&true_lb, &true_extent); - //MPI_Op *op_ptr; - //is_commutative = op->is_commutative(); + is_commutative = op->is_commutative(); { + int range = 0, range_threshold = 0, range_threshold_intra = 0; + int is_two_level = 0; + /* Search for the corresponding system size inside the tuning table */ while ((range < (mv2_size_allreduce_tuning_table - 1)) && (comm_size > mv2_allreduce_thresholds_table[range].numproc)) { range++; } /* Search for corresponding inter-leader function */ - /* skip mcast poiters if mcast is not available */ + /* skip mcast pointers if mcast is not available */ if(mv2_allreduce_thresholds_table[range].mcast_enabled != 1){ while ((range_threshold < (mv2_allreduce_thresholds_table[range].size_inter_table - 1)) && ((mv2_allreduce_thresholds_table[range]. @@ -445,37 +434,37 @@ int Coll_allreduce_mvapich2::allreduce(void *sendbuf, } -int Coll_alltoallv_mvapich2::alltoallv(void *sbuf, int *scounts, int *sdisps, +int alltoallv__mvapich2(const void *sbuf, const int *scounts, const int *sdisps, MPI_Datatype sdtype, - void *rbuf, int *rcounts, int *rdisps, + void *rbuf, const int *rcounts, const int *rdisps, MPI_Datatype rdtype, MPI_Comm comm ) { if (sbuf == MPI_IN_PLACE) { - return Coll_alltoallv_ompi_basic_linear::alltoallv(sbuf, scounts, sdisps, sdtype, - rbuf, rcounts, rdisps,rdtype, - comm); + return alltoallv__ompi_basic_linear(sbuf, scounts, sdisps, sdtype, + rbuf, rcounts, rdisps, rdtype, + comm); } else /* For starters, just keep the original algorithm. */ - return Coll_alltoallv_ring::alltoallv(sbuf, scounts, sdisps, sdtype, - rbuf, rcounts, rdisps,rdtype, - comm); + return alltoallv__ring(sbuf, scounts, sdisps, sdtype, + rbuf, rcounts, rdisps, rdtype, + comm); } -int Coll_barrier_mvapich2::barrier(MPI_Comm comm) +int barrier__mvapich2(MPI_Comm comm) { - return Coll_barrier_mvapich2_pair::barrier(comm); + return barrier__mvapich2_pair(comm); } -int Coll_bcast_mvapich2::bcast(void *buffer, - int count, - MPI_Datatype datatype, - int root, MPI_Comm comm) +int bcast__mvapich2(void *buffer, + int count, + MPI_Datatype datatype, + int root, MPI_Comm comm) { int mpi_errno = MPI_SUCCESS; int comm_size/*, rank*/; @@ -484,10 +473,10 @@ int Coll_bcast_mvapich2::bcast(void *buffer, int range = 0; int range_threshold = 0; int range_threshold_intra = 0; - int is_homogeneous, is_contig; + // int is_homogeneous, is_contig; MPI_Aint type_size; //, position; - void *tmp_buf = NULL; + // unsigned char *tmp_buf = NULL; MPI_Comm shmem_comm; //MPID_Datatype *dtp; @@ -501,7 +490,7 @@ int Coll_bcast_mvapich2::bcast(void *buffer, comm_size = comm->size(); //rank = comm->rank(); - is_contig=1; + //is_contig=1; /* if (HANDLE_GET_KIND(datatype) == HANDLE_KIND_BUILTIN)*/ /* is_contig = 1;*/ /* else {*/ @@ -509,7 +498,7 @@ int Coll_bcast_mvapich2::bcast(void *buffer, /* is_contig = dtp->is_contig;*/ /* }*/ - is_homogeneous = 1; + // is_homogeneous = 1; /* MPI_Type_size() might not give the accurate size of the packed * datatype for heterogeneous systems (because of padding, encoding, @@ -597,50 +586,48 @@ int Coll_bcast_mvapich2::bcast(void *buffer, mv2_bcast_thresholds_table[range].is_two_level_bcast[range_threshold]; #endif if (two_level_bcast == 1) { - 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);*/ - /* }*/ - } + // if (not is_contig || not is_homogeneous) { +// tmp_buf = 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 (not is_contig || not is_homogeneous) { - mpi_errno = MPIR_Pipelined_Bcast_Zcpy_MV2(tmp_buf, nbytes, MPI_BYTE, root, comm); - } else { + // 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 #endif /* defined(CHANNEL_MRAIL_GEN2) */ { shmem_comm = comm->get_intra_comm(); - if (not is_contig || not is_homogeneous) { - mpi_errno = 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); - } + // if (not is_contig || not is_homogeneous) { + // MPIR_Bcast_tune_inter_node_helper_MV2(tmp_buf, nbytes, MPI_BYTE, root, comm); + // } else { + 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 (not is_contig || not is_homogeneous) { - mpi_errno = MV2_Bcast_intra_node_function(tmp_buf, nbytes, MPI_BYTE, root, shmem_comm); - } else { + // 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 (not is_contig || not is_homogeneous) {*/ /* if (rank != root) {*/ @@ -669,7 +656,7 @@ int Coll_bcast_mvapich2::bcast(void *buffer, -int Coll_reduce_mvapich2::reduce( void *sendbuf, +int reduce__mvapich2(const void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype, @@ -787,17 +774,15 @@ int Coll_reduce_mvapich2::reduce( void *sendbuf, } -int Coll_reduce_scatter_mvapich2::reduce_scatter(void *sendbuf, void *recvbuf, int *recvcnts, +int reduce_scatter__mvapich2(const void *sendbuf, void *recvbuf, const int *recvcnts, MPI_Datatype datatype, MPI_Op op, MPI_Comm comm) { int mpi_errno = MPI_SUCCESS; int i = 0, comm_size = comm->size(), total_count = 0, type_size = 0, nbytes = 0; - int range = 0; - int range_threshold = 0; int is_commutative = 0; - int *disps = static_cast(xbt_malloc(comm_size * sizeof (int))); + int* disps = new int[comm_size]; if(mv2_red_scat_thresholds_table==NULL) init_mv2_reduce_scatter_tables_stampede(); @@ -812,6 +797,8 @@ int Coll_reduce_scatter_mvapich2::reduce_scatter(void *sendbuf, void *recvbuf, i nbytes = total_count * type_size; if (is_commutative) { + int range = 0; + int range_threshold = 0; /* Search for the corresponding system size inside the tuning table */ while ((range < (mv2_size_red_scat_tuning_table - 1)) && @@ -847,22 +834,22 @@ 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); } - mpi_errno = Coll_reduce_scatter_mpich_rdb::reduce_scatter(sendbuf, recvbuf, - recvcnts, datatype, - op, comm); + mpi_errno = reduce_scatter__mpich_rdb(sendbuf, recvbuf, + recvcnts, datatype, + op, comm); } - xbt_free(disps); + delete[] disps; return mpi_errno; } -int Coll_scatter_mvapich2::scatter(void *sendbuf, +int scatter__mvapich2(const void *sendbuf, int sendcnt, MPI_Datatype sendtype, void *recvbuf, @@ -874,11 +861,8 @@ int Coll_scatter_mvapich2::scatter(void *sendbuf, int mpi_errno = MPI_SUCCESS; // int mpi_errno_ret = MPI_SUCCESS; int rank, nbytes, comm_size; - int recvtype_size, sendtype_size; int partial_sub_ok = 0; int conf_index = 0; - int local_size = -1; - int i; MPI_Comm shmem_comm; // MPID_Comm *shmem_commptr=NULL; if(mv2_scatter_thresholds_table==NULL) @@ -893,23 +877,23 @@ int Coll_scatter_mvapich2::scatter(void *sendbuf, rank = comm->rank(); if (rank == root) { - sendtype_size=sendtype->size(); - nbytes = sendcnt * sendtype_size; + int sendtype_size = sendtype->size(); + nbytes = sendcnt * sendtype_size; } else { - recvtype_size=recvtype->size(); - nbytes = recvcnt * recvtype_size; + int recvtype_size = recvtype->size(); + nbytes = recvcnt * recvtype_size; } // 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 conf_index = 0; }else{ + int local_size = shmem_comm->size(); + int i = 0; do { if (local_size == mv2_scatter_table_ppn_conf[i]) { conf_index = i; @@ -1000,28 +984,27 @@ int Coll_scatter_mvapich2::scatter(void *sendbuf, 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); + delete[] mv2_alltoall_thresholds_table[0]; + delete[] mv2_alltoall_thresholds_table; + delete[] mv2_size_alltoall_tuning_table; + delete[] mv2_alltoall_table_ppn_conf; - xbt_free(mv2_gather_thresholds_table); + delete[] 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); + delete[] mv2_allgather_thresholds_table[0]; + delete[] mv2_size_allgather_tuning_table; + delete[] mv2_allgather_table_ppn_conf; + delete[] mv2_allgather_thresholds_table; + + delete[] mv2_allgatherv_thresholds_table; + delete[] mv2_reduce_thresholds_table; + delete[] mv2_red_scat_thresholds_table; + delete[] mv2_allreduce_thresholds_table; + delete[] 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); + delete[] mv2_scatter_thresholds_table[0]; + delete[] mv2_scatter_thresholds_table; + delete[] mv2_size_scatter_tuning_table; + delete[] mv2_scatter_table_ppn_conf; }