X-Git-Url: http://info.iut-bm.univ-fcomte.fr/pub/gitweb/simgrid.git/blobdiff_plain/70ed180f3ce3495678d048e9e396ef5eb65a6a99..17af7a6b57c5056f8544c59e9f6c078364ff542b:/src/smpi/smpi_coll.c diff --git a/src/smpi/smpi_coll.c b/src/smpi/smpi_coll.c index c5a935a950..a5d1232719 100644 --- a/src/smpi/smpi_coll.c +++ b/src/smpi/smpi_coll.c @@ -102,7 +102,7 @@ static void tree_bcast(void* buf, int count, MPI_Datatype datatype, int root, MP smpi_mpi_recv(buf, count, datatype, tree->parent, system_tag + rank, comm, MPI_STATUS_IGNORE); } requests = xbt_new(MPI_Request, tree->numChildren); - DEBUG2("<%d> creates %d requests (1 per child)\n", rank, tree->numChildren); + DEBUG2("<%d> creates %d requests (1 per child)", rank, tree->numChildren); /* iniates sends to ranks lower in the tree */ for(i = 0; i < tree->numChildren; i++) { if(tree->child[i] == -1) { @@ -133,7 +133,7 @@ static void tree_antibcast(void* buf, int count, MPI_Datatype datatype, int root } //every one receives as many messages as it has children requests = xbt_new(MPI_Request, tree->numChildren); - DEBUG2("<%d> creates %d requests (1 per child)\n", rank, tree->numChildren); + DEBUG2("<%d> creates %d requests (1 per child)", rank, tree->numChildren); for(i = 0; i < tree->numChildren; i++) { if(tree->child[i] == -1) { requests[i] = MPI_REQUEST_NULL; @@ -182,7 +182,52 @@ void nary_tree_barrier(MPI_Comm comm, int arity) { * Openmpi calls this routine when the message size sent to each rank < 2000 bytes and size < 12 **/ int smpi_coll_tuned_alltoall_bruck(void* sendbuf, int sendcount, MPI_Datatype sendtype, void* recvbuf, int recvcount, MPI_Datatype recvtype, MPI_Comm comm) { - DEBUG0("coll:tuned:alltoall_intra_bruck ** NOT IMPLEMENTED YET**"); + int system_tag = 777; + int i, rank, size, err, count; + MPI_Aint lb, sendextent, recvextent; + MPI_Request* requests; + + // FIXME: check implementation + rank = smpi_comm_rank(comm); + size = smpi_comm_size(comm); + DEBUG1("<%d> algorithm alltoall_bruck() called.", rank); + err = smpi_datatype_extent(sendtype, &lb, &sendextent); + err = smpi_datatype_extent(recvtype, &lb, &recvextent); + /* Local copy from self */ + err = smpi_datatype_copy(&((char*)sendbuf)[rank * sendextent], sendcount, sendtype, &((char*)recvbuf)[rank * recvextent], recvcount, recvtype); + if(err == MPI_SUCCESS && size > 1) { + /* Initiate all send/recv to/from others. */ + requests = xbt_new(MPI_Request, 2 * (size - 1)); + count = 0; + /* Create all receives that will be posted first */ + for(i = 0; i < size; ++i) { + if(i == rank) { + DEBUG3("<%d> skip request creation [src = %d, recvcount = %d]", rank, i, recvcount); + continue; + } + requests[count] = smpi_mpi_irecv(&((char*)recvbuf)[i * recvextent], recvcount, recvtype, i, system_tag, comm); + count++; + } + /* Now create all sends */ + for(i = 0; i < size; ++i) { + if(i == rank) { + DEBUG3("<%d> skip request creation [dst = %d, sendcount = %d]", rank, i, sendcount); + continue; + } + requests[count] = smpi_mpi_isend(&((char*)sendbuf)[i * sendextent], sendcount, sendtype, i, system_tag, comm); + count++; + } + /* Wait for them all. If there's an error, note that we don't + * care what the error was -- just that there *was* an error. The + * PML will finish all requests, even if one or more of them fail. + * i.e., by the end of this call, all the requests are free-able. + * So free them anyway -- even if there was an error, and return + * the error after we free everything. + */ + DEBUG2("<%d> wait for %d requests", rank, count); + smpi_mpi_waitall(count, requests, MPI_STATUS_IGNORE); + xbt_free(requests); + } return MPI_SUCCESS; } @@ -269,7 +314,7 @@ int smpi_coll_tuned_alltoall_pairwise(void* sendbuf, int sendcount, MPI_Datatype int smpi_coll_basic_alltoallv(void* sendbuf, int* sendcounts, int* senddisps, MPI_Datatype sendtype, void* recvbuf, int *recvcounts, int* recvdisps, MPI_Datatype recvtype, MPI_Comm comm) { int system_tag = 889; - int i, rank, size, err, rcount, scount; + int i, rank, size, err, count; MPI_Aint lb, sendextent, recvextent; MPI_Request* requests; @@ -284,28 +329,25 @@ int smpi_coll_basic_alltoallv(void* sendbuf, int* sendcounts, int* senddisps, MP if(err == MPI_SUCCESS && size > 1) { /* Initiate all send/recv to/from others. */ requests = xbt_new(MPI_Request, 2 * (size - 1)); - rcount = 0; + count = 0; /* Create all receives that will be posted first */ for(i = 0; i < size; ++i) { if(i == rank || recvcounts[i] == 0) { DEBUG3("<%d> skip request creation [src = %d, recvcounts[src] = %d]", rank, i, recvcounts[i]); continue; } - requests[rcount] = smpi_mpi_irecv(&((char*)recvbuf)[recvdisps[i] * recvextent], recvcounts[i], recvtype, i, system_tag, comm); - rcount++; + requests[count] = smpi_mpi_irecv(&((char*)recvbuf)[recvdisps[i] * recvextent], recvcounts[i], recvtype, i, system_tag, comm); + count++; } - DEBUG2("<%d> %d irecv requests created", rank, rcount); - scount = rcount; /* Now create all sends */ for(i = 0; i < size; ++i) { if(i == rank || sendcounts[i] == 0) { DEBUG3("<%d> skip request creation [dst = %d, sendcounts[dst] = %d]", rank, i, sendcounts[i]); continue; } - requests[scount] = smpi_mpi_isend(&((char*)sendbuf)[senddisps[i] * sendextent], sendcounts[i], sendtype, i, system_tag, comm); - scount++; + requests[count] = smpi_mpi_isend(&((char*)sendbuf)[senddisps[i] * sendextent], sendcounts[i], sendtype, i, system_tag, comm); + count++; } - DEBUG2("<%d> %d isend requests created", rank, scount); /* Wait for them all. If there's an error, note that we don't * care what the error was -- just that there *was* an error. The * PML will finish all requests, even if one or more of them fail. @@ -313,8 +355,8 @@ int smpi_coll_basic_alltoallv(void* sendbuf, int* sendcounts, int* senddisps, MP * So free them anyway -- even if there was an error, and return * the error after we free everything. */ - DEBUG2("<%d> wait for %d requests", rank, rcount + scount); - smpi_mpi_waitall(rcount + scount, requests, MPI_STATUS_IGNORE); + DEBUG2("<%d> wait for %d requests", rank, count); + smpi_mpi_waitall(count, requests, MPI_STATUS_IGNORE); xbt_free(requests); } return err;