/* * Copyright (C) by Argonne National Laboratory * See COPYRIGHT in top-level directory */ #include "mpiimpl.h" int MPIR_Iscan_intra_sched_smp(const void *sendbuf, void *recvbuf, MPI_Aint count, MPI_Datatype datatype, MPI_Op op, MPIR_Comm * comm_ptr, MPIR_Sched_t s) { int mpi_errno = MPI_SUCCESS; int rank = comm_ptr->rank; MPIR_Comm *node_comm; MPIR_Comm *roots_comm; MPI_Aint true_extent, true_lb, extent; void *tempbuf = NULL; void *prefulldata = NULL; void *localfulldata = NULL; /* In order to use the SMP-aware algorithm, the "op" can be * either commutative or non-commutative, but we require a * communicator in which all the nodes contain processes with * consecutive ranks. */ if (!MPII_Comm_is_node_consecutive(comm_ptr)) { /* We can't use the SMP-aware algorithm, use the non-SMP-aware * one */ return MPIR_Iscan_intra_sched_recursive_doubling(sendbuf, recvbuf, count, datatype, op, comm_ptr, s); } node_comm = comm_ptr->node_comm; roots_comm = comm_ptr->node_roots_comm; MPIR_Type_get_true_extent_impl(datatype, &true_lb, &true_extent); MPIR_Datatype_get_extent_macro(datatype, extent); tempbuf = MPIR_Sched_alloc_state(s, count * (MPL_MAX(extent, true_extent))); MPIR_ERR_CHKANDJUMP(!tempbuf, mpi_errno, MPI_ERR_OTHER, "**nomem"); tempbuf = (void *) ((char *) tempbuf - true_lb); /* Create prefulldata and localfulldata on local roots of all nodes */ if (comm_ptr->node_roots_comm != NULL) { prefulldata = MPIR_Sched_alloc_state(s, count * (MPL_MAX(extent, true_extent))); MPIR_ERR_CHKANDJUMP(!prefulldata, mpi_errno, MPI_ERR_OTHER, "**nomem"); prefulldata = (void *) ((char *) prefulldata - true_lb); if (node_comm != NULL) { localfulldata = MPIR_Sched_alloc_state(s, count * (MPL_MAX(extent, true_extent))); MPIR_ERR_CHKANDJUMP(!localfulldata, mpi_errno, MPI_ERR_OTHER, "**nomem"); localfulldata = (void *) ((char *) localfulldata - true_lb); } } /* perform intranode scan to get temporary result in recvbuf. if there is only * one process, just copy the raw data. */ if (node_comm != NULL) { mpi_errno = MPIR_Iscan_intra_sched_auto(sendbuf, recvbuf, count, datatype, op, node_comm, s); MPIR_ERR_CHECK(mpi_errno); MPIR_SCHED_BARRIER(s); } else if (sendbuf != MPI_IN_PLACE) { mpi_errno = MPIR_Sched_copy(sendbuf, count, datatype, recvbuf, count, datatype, s); MPIR_ERR_CHECK(mpi_errno); MPIR_SCHED_BARRIER(s); } /* get result from local node's last processor which * contains the reduce result of the whole node. Name it as * localfulldata. For example, localfulldata from node 1 contains * reduced data of rank 1,2,3. */ if (roots_comm != NULL && node_comm != NULL) { mpi_errno = MPIR_Sched_recv(localfulldata, count, datatype, (node_comm->local_size - 1), node_comm, s); MPIR_ERR_CHECK(mpi_errno); MPIR_SCHED_BARRIER(s); } else if (roots_comm == NULL && node_comm != NULL && node_comm->rank == node_comm->local_size - 1) { mpi_errno = MPIR_Sched_send(recvbuf, count, datatype, 0, node_comm, s); MPIR_ERR_CHECK(mpi_errno); MPIR_SCHED_BARRIER(s); } else if (roots_comm != NULL) { localfulldata = recvbuf; } /* do scan on localfulldata to prefulldata. for example, * prefulldata on rank 4 contains reduce result of ranks * 1,2,3,4,5,6. it will be sent to rank 7 which is the * main process of node 3. */ if (roots_comm != NULL) { /* FIXME just use roots_comm->rank instead */ int roots_rank = MPIR_Get_internode_rank(comm_ptr, rank); MPIR_Assert(roots_rank == roots_comm->rank); mpi_errno = MPIR_Iscan_intra_sched_auto(localfulldata, prefulldata, count, datatype, op, roots_comm, s); MPIR_ERR_CHECK(mpi_errno); MPIR_SCHED_BARRIER(s); if (roots_rank != roots_comm->local_size - 1) { mpi_errno = MPIR_Sched_send(prefulldata, count, datatype, (roots_rank + 1), roots_comm, s); MPIR_ERR_CHECK(mpi_errno); MPIR_SCHED_BARRIER(s); } if (roots_rank != 0) { mpi_errno = MPIR_Sched_recv(tempbuf, count, datatype, (roots_rank - 1), roots_comm, s); MPIR_ERR_CHECK(mpi_errno); MPIR_SCHED_BARRIER(s); } } /* now tempbuf contains all the data needed to get the correct * scan result. for example, to node 3, it will have reduce result * of rank 1,2,3,4,5,6 in tempbuf. * then we should broadcast this result in the local node, and * reduce it with recvbuf to get final result if necessary. */ if (MPIR_Get_internode_rank(comm_ptr, rank) != 0) { /* we aren't on "node 0", so our node leader (possibly us) received * "prefulldata" from another leader into "tempbuf" */ if (node_comm != NULL) { mpi_errno = MPIR_Ibcast_intra_sched_auto(tempbuf, count, datatype, 0, node_comm, s); MPIR_ERR_CHECK(mpi_errno); MPIR_SCHED_BARRIER(s); } /* do reduce on tempbuf and recvbuf, finish scan. */ mpi_errno = MPIR_Sched_reduce(tempbuf, recvbuf, count, datatype, op, s); MPIR_ERR_CHECK(mpi_errno); } fn_exit: return mpi_errno; fn_fail: goto fn_exit; }