/*****************************************************************************
* Function: alltoall_ring_mpi_barrier
* Return: int
* Inputs:
send_buff: send input buffer
send_count: number of elements to send
send_type: data type of elements being sent
recv_buff: receive output buffer
recv_count: number of elements to received
recv_type: data type of elements being received
comm: communicator
* Descrp: Function works in P - 1 steps. In step i, node j - i -> j -> j + i.
MPI barriers are added between each two phases.
* Auther: Ahmad Faraj
****************************************************************************/
int
smpi_coll_tuned_alltoall_ring_mpi_barrier(void *send_buff, int send_count,
MPI_Datatype send_type,
void *recv_buff, int recv_count,
MPI_Datatype recv_type, MPI_Comm comm)
{
MPI_Status s;
MPI_Aint send_chunk, recv_chunk;
int i, src, dst, rank, num_procs;
int tag = COLL_TAG_ALLTOALL;
char *send_ptr = (char *) send_buff;
char *recv_ptr = (char *) recv_buff;
rank = smpi_comm_rank(comm);
num_procs = smpi_comm_size(comm);
send_chunk = smpi_datatype_get_extent(send_type);
recv_chunk = smpi_datatype_get_extent(recv_type);
send_chunk *= send_count;
recv_chunk *= recv_count;
for (i = 0; i < num_procs; i++) {
src = (rank - i + num_procs) % num_procs;
dst = (rank + i) % num_procs;
smpi_mpi_barrier(comm);
smpi_mpi_sendrecv(send_ptr + dst * send_chunk, send_count, send_type, dst,
tag, recv_ptr + src * recv_chunk, recv_count, recv_type,
src, tag, comm, &s);
}
return MPI_SUCCESS;
}
/*****************************************************************************
* Function: alltoall_pair_mpi_barrier
* Return: int
* Inputs:
send_buff: send input buffer
send_count: number of elements to send
send_type: data type of elements being sent
recv_buff: receive output buffer
recv_count: number of elements to received
recv_type: data type of elements being received
comm: communicator
* Descrp: Function works when P is power of two. In each phase of P - 1
phases, nodes in pair communicate their data. MPI barriers are
inserted between each two phases.
* Auther: Ahmad Faraj
****************************************************************************/
int
smpi_coll_tuned_alltoallv_pair_mpi_barrier(void *send_buff, int *send_counts, int *send_disps,
MPI_Datatype send_type,
void *recv_buff, int *recv_counts, int *recv_disps,
MPI_Datatype recv_type, MPI_Comm comm)
{
MPI_Status s;
MPI_Aint send_chunk, recv_chunk;
int i, src, dst, rank, num_procs;
int tag = 101;
char *send_ptr = (char *) send_buff;
char *recv_ptr = (char *) recv_buff;
rank = smpi_comm_rank(comm);
num_procs = smpi_comm_size(comm);
send_chunk = smpi_datatype_get_extent(send_type);
recv_chunk = smpi_datatype_get_extent(recv_type);
for (i = 0; i < num_procs; i++) {
src = dst = rank ^ i;
smpi_mpi_barrier(comm);
smpi_mpi_sendrecv(send_ptr + send_disps[dst] * send_chunk, send_counts[dst], send_type, dst,
tag, recv_ptr + recv_disps[src] * recv_chunk, recv_counts[src], recv_type,
src, tag, comm, &s);
}
return MPI_SUCCESS;
}
static void action_barrier(const char *const *action){
double clock = smpi_process_simulated_elapsed();
#ifdef HAVE_TRACING
int rank = smpi_comm_rank(MPI_COMM_WORLD);
TRACE_smpi_computing_out(rank);
TRACE_smpi_collective_in(rank, -1, __FUNCTION__);
#endif
smpi_mpi_barrier(MPI_COMM_WORLD);
#ifdef HAVE_TRACING
TRACE_smpi_collective_out(rank, -1, __FUNCTION__);
TRACE_smpi_computing_in(rank);
#endif
if (XBT_LOG_ISENABLED(smpi_replay, xbt_log_priority_verbose)){
char *name = xbt_str_join_array(action, " ");
XBT_VERB("%s %f", name, smpi_process_simulated_elapsed()-clock);
free(name);
}
}