-/* Copyright (c) 2013-2017. The SimGrid Team.
+/* Copyright (c) 2013-2020. The SimGrid Team.
* All rights reserved. */
/* This program is free software; you can redistribute it and/or modify it
* (C) 2001 by Argonne National Laboratory.
* See COPYRIGHT in top-level directory.
*/
- #include "../colls_private.h"
+#include "../colls_private.hpp"
namespace simgrid{
namespace smpi{
-int Coll_allgather_mvapich2_smp::allgather(void *sendbuf,int sendcnt, MPI_Datatype sendtype,
+int allgather__mvapich2_smp(const void *sendbuf,int sendcnt, MPI_Datatype sendtype,
void *recvbuf, int recvcnt,MPI_Datatype recvtype,
MPI_Comm comm)
{
int rank, size;
int local_rank, local_size;
- int leader_comm_size = 0;
+ int leader_comm_size = 0;
int mpi_errno = MPI_SUCCESS;
MPI_Aint recvtype_extent = 0; /* Datatype extent */
MPI_Comm shmem_comm, leader_comm;
if(comm->get_leaders_comm()==MPI_COMM_NULL){
comm->init_smp();
}
-
- if(!comm->is_uniform() || !comm->is_blocked())
- THROWF(arg_error,0, "allgather MVAPICH2 smp algorithm can't be used with irregular deployment. Please insure that processes deployed on the same node are contiguous and that each node has the same number of processes");
-
- if (recvcnt == 0) {
- return MPI_SUCCESS;
+
+ if (not comm->is_uniform() || not comm->is_blocked())
+ throw std::invalid_argument("allgather MVAPICH2 smp algorithm can't be used with irregular deployment. Please "
+ "insure that processes deployed on the same node are contiguous and that each node has "
+ "the same number of processes");
+
+ if (recvcnt == 0) {
+ return MPI_SUCCESS;
}
rank = comm->rank();
/* extract the rank,size information for the intra-node communicator */
recvtype_extent=recvtype->get_extent();
-
+
shmem_comm = comm->get_intra_comm();
local_rank = shmem_comm->rank();
local_size = shmem_comm->size();
/*If there is just one node, after gather itself,
* root has all the data and it can do bcast*/
if(local_rank == 0) {
- mpi_errno = Colls::gather(sendbuf, sendcnt,sendtype,
- (void*)((char*)recvbuf + (rank * recvcnt * recvtype_extent)),
- recvcnt, recvtype,
- 0, shmem_comm);
+ mpi_errno =
+ colls::gather(sendbuf, sendcnt, sendtype, (void*)((char*)recvbuf + (rank * recvcnt * recvtype_extent)),
+ recvcnt, recvtype, 0, shmem_comm);
} else {
- /*Since in allgather all the processes could have
+ /*Since in allgather all the processes could have
* its own data in place*/
if(sendbuf == MPI_IN_PLACE) {
- mpi_errno = Colls::gather((void*)((char*)recvbuf + (rank * recvcnt * recvtype_extent)),
- recvcnt , recvtype,
- recvbuf, recvcnt, recvtype,
- 0, shmem_comm);
+ mpi_errno = colls::gather((void*)((char*)recvbuf + (rank * recvcnt * recvtype_extent)), recvcnt, recvtype,
+ recvbuf, recvcnt, recvtype, 0, shmem_comm);
} else {
- mpi_errno = Colls::gather(sendbuf, sendcnt,sendtype,
- recvbuf, recvcnt, recvtype,
- 0, shmem_comm);
+ mpi_errno = colls::gather(sendbuf, sendcnt, sendtype, recvbuf, recvcnt, recvtype, 0, shmem_comm);
}
}
/* Exchange the data between the node leaders*/
if (local_rank == 0 && (leader_comm_size > 1)) {
/*When data in each socket is different*/
- if (comm->is_uniform() != 1) {
-
- int *displs = NULL;
- int *recvcnts = NULL;
- int *node_sizes = NULL;
- int i = 0;
-
- node_sizes = comm->get_non_uniform_map();
-
- displs = static_cast<int *>(xbt_malloc(sizeof (int) * leader_comm_size));
- recvcnts = static_cast<int *>(xbt_malloc(sizeof (int) * leader_comm_size));
- if (!displs || !recvcnts) {
- return MPI_ERR_OTHER;
- }
- recvcnts[0] = node_sizes[0] * recvcnt;
- displs[0] = 0;
-
- for (i = 1; i < leader_comm_size; i++) {
- displs[i] = displs[i - 1] + node_sizes[i - 1] * recvcnt;
- recvcnts[i] = node_sizes[i] * recvcnt;
- }
-
-
- void* sendbuf=((char*)recvbuf)+recvtype->get_extent()*displs[leader_comm->rank()];
-
- mpi_errno = Colls::allgatherv(sendbuf,
- (recvcnt*local_size),
- recvtype,
- recvbuf, recvcnts,
- displs, recvtype,
- leader_comm);
- xbt_free(displs);
- xbt_free(recvcnts);
+ if (not comm->is_uniform()) {
+
+ int* node_sizes = nullptr;
+ int i = 0;
+
+ node_sizes = comm->get_non_uniform_map();
+
+ int* displs = new int[leader_comm_size];
+ int* recvcnts = new int[leader_comm_size];
+ recvcnts[0] = node_sizes[0] * recvcnt;
+ displs[0] = 0;
+
+ for (i = 1; i < leader_comm_size; i++) {
+ displs[i] = displs[i - 1] + node_sizes[i - 1] * recvcnt;
+ recvcnts[i] = node_sizes[i] * recvcnt;
+ }
+
+ void* sendtmpbuf = ((char*)recvbuf) + recvtype->get_extent() * displs[leader_comm->rank()];
+
+ mpi_errno = colls::allgatherv(sendtmpbuf, (recvcnt * local_size), recvtype, recvbuf, recvcnts, displs,
+ recvtype, leader_comm);
+ delete[] displs;
+ delete[] recvcnts;
} else {
void* sendtmpbuf=((char*)recvbuf)+recvtype->get_extent()*(recvcnt*local_size)*leader_comm->rank();
-
-
- mpi_errno = Coll_allgather_mpich::allgather(sendtmpbuf,
+
+
+ mpi_errno = allgather__mpich(sendtmpbuf,
(recvcnt*local_size),
recvtype,
recvbuf, (recvcnt*local_size), recvtype,
}
/*Bcast the entire data from node leaders to all other cores*/
- mpi_errno = Colls::bcast (recvbuf, recvcnt * size, recvtype, 0, shmem_comm);
+ mpi_errno = colls::bcast(recvbuf, recvcnt * size, recvtype, 0, shmem_comm);
return mpi_errno;
}
