#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;
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 );
- 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,
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)
{
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);
}
return mpi_errno;
}
-int Coll_gather_mvapich2::gather(void *sendbuf,
+int gather__mvapich2(const void *sendbuf,
int sendcnt,
MPI_Datatype sendtype,
void *recvbuf,
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 */
} 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);
}
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;
-int Coll_allreduce_mvapich2::allreduce(void *sendbuf,
+int allreduce__mvapich2(const void *sendbuf,
void *recvbuf,
int count,
MPI_Datatype datatype,
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;
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].
}
-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*/;
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;
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 {*/
/* 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,
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) {
- MPIR_Bcast_tune_inter_node_helper_MV2(tmp_buf, nbytes, MPI_BYTE, root, comm);
- } else {
+ // 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) {*/
-int Coll_reduce_mvapich2::reduce( void *sendbuf,
+int reduce__mvapich2(const void *sendbuf,
void *recvbuf,
int count,
MPI_Datatype datatype,
}
-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)
{
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);
}
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,
if (comm->is_uniform()) {
shmem_comm = comm->get_intra_comm();
- int local_size = shmem_comm->size();
- int 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;