diff --git a/ompi/datatype/ompi_datatype_sndrcv.c b/ompi/datatype/ompi_datatype_sndrcv.c index 8cbbac038c..6cc0b1f7b3 100644 --- a/ompi/datatype/ompi_datatype_sndrcv.c +++ b/ompi/datatype/ompi_datatype_sndrcv.c @@ -11,6 +11,8 @@ * Copyright (c) 2004-2006 The Regents of the University of California. * All rights reserved. * Copyright (c) 2009 Oak Ridge National Labs. All rights reserved. + * Copyright (c) 2014 Research Organization for Information Science + * and Technology (RIST). All rights reserved. * $COPYRIGHT$ * * Additional copyrights may follow @@ -50,8 +52,8 @@ int32_t ompi_datatype_sndrcv( void *sbuf, int32_t scount, const ompi_datatype_t* size_t max_data; /* First check if we really have something to do */ - if (0 == rcount) { - return ((0 == scount) ? MPI_SUCCESS : MPI_ERR_TRUNCATE); + if (0 == rcount || 0 == rdtype->super.size) { + return ((0 == scount || 0 == sdtype->super.size) ? MPI_SUCCESS : MPI_ERR_TRUNCATE); } /* If same datatypes used, just copy. */ diff --git a/ompi/mca/coll/basic/coll_basic_alltoallv.c b/ompi/mca/coll/basic/coll_basic_alltoallv.c index 5e27fc40cb..9d7bd8388f 100644 --- a/ompi/mca/coll/basic/coll_basic_alltoallv.c +++ b/ompi/mca/coll/basic/coll_basic_alltoallv.c @@ -13,6 +13,9 @@ * Copyright (c) 2013 Los Alamos National Security, LLC. All rights * reserved. * Copyright (c) 2013 FUJITSU LIMITED. All rights reserved. + * Copyright (c) 2014 Research Organization for Information Science + * and Technology (RIST). All rights reserved. + * Copyright (c) 2014 Cisco Systems, Inc. All rights reserved. * $COPYRIGHT$ * * Additional copyrights may follow @@ -38,10 +41,10 @@ mca_coll_basic_alltoallv_intra_inplace(void *rbuf, const int *rcounts, const int mca_coll_base_module_t *module) { mca_coll_basic_module_t *basic_module = (mca_coll_basic_module_t*) module; - int i, j, size, rank, err=MPI_SUCCESS; + int i, j, size, rank, err=MPI_SUCCESS, nreqs; MPI_Request *preq; char *tmp_buffer; - size_t max_size; + size_t type_size, max_size; ptrdiff_t ext; /* Initialize. */ @@ -68,13 +71,19 @@ mca_coll_basic_alltoallv_intra_inplace(void *rbuf, const int *rcounts, const int return OMPI_ERR_OUT_OF_RESOURCE; } + ompi_datatype_type_size(rdtype, &type_size); + /* in-place alltoallv slow algorithm (but works) */ for (i = 0 ; i < size ; ++i) { + size_t msg_size_i = type_size * rcounts[i]; + for (j = i+1 ; j < size ; ++j) { + size_t msg_size_j = type_size * rcounts[j]; /* Initiate all send/recv to/from others. */ + nreqs = 0; preq = basic_module->mccb_reqs; - if (i == rank && rcounts[j]) { + if (i == rank && msg_size_j > 0) { /* Copy the data into the temporary buffer */ err = ompi_datatype_copy_content_same_ddt (rdtype, rcounts[j], tmp_buffer, (char *) rbuf + rdisps[j] * ext); @@ -83,13 +92,15 @@ mca_coll_basic_alltoallv_intra_inplace(void *rbuf, const int *rcounts, const int /* Exchange data with the peer */ err = MCA_PML_CALL(irecv ((char *) rbuf + rdisps[j] * ext, rcounts[j], rdtype, j, MCA_COLL_BASE_TAG_ALLTOALLV, comm, preq++)); + ++nreqs; if (MPI_SUCCESS != err) { goto error_hndl; } err = MCA_PML_CALL(isend ((void *) tmp_buffer, rcounts[j], rdtype, j, MCA_COLL_BASE_TAG_ALLTOALLV, MCA_PML_BASE_SEND_STANDARD, comm, preq++)); + ++nreqs; if (MPI_SUCCESS != err) { goto error_hndl; } - } else if (j == rank && rcounts[i]) { + } else if (j == rank && msg_size_i > 0) { /* Copy the data into the temporary buffer */ err = ompi_datatype_copy_content_same_ddt (rdtype, rcounts[i], tmp_buffer, (char *) rbuf + rdisps[i] * ext); @@ -98,22 +109,24 @@ mca_coll_basic_alltoallv_intra_inplace(void *rbuf, const int *rcounts, const int /* Exchange data with the peer */ err = MCA_PML_CALL(irecv ((char *) rbuf + rdisps[i] * ext, rcounts[i], rdtype, i, MCA_COLL_BASE_TAG_ALLTOALLV, comm, preq++)); + ++nreqs; if (MPI_SUCCESS != err) { goto error_hndl; } err = MCA_PML_CALL(isend ((void *) tmp_buffer, rcounts[i], rdtype, i, MCA_COLL_BASE_TAG_ALLTOALLV, MCA_PML_BASE_SEND_STANDARD, comm, preq++)); + ++nreqs; if (MPI_SUCCESS != err) { goto error_hndl; } } else { continue; } /* Wait for the requests to complete */ - err = ompi_request_wait_all (2, basic_module->mccb_reqs, MPI_STATUS_IGNORE); + err = ompi_request_wait_all (nreqs, basic_module->mccb_reqs, MPI_STATUSES_IGNORE); if (MPI_SUCCESS != err) { goto error_hndl; } /* Free the requests. */ - mca_coll_basic_free_reqs(basic_module->mccb_reqs, 2); + mca_coll_basic_free_reqs(basic_module->mccb_reqs, nreqs); } } @@ -147,6 +160,7 @@ mca_coll_basic_alltoallv_intra(void *sbuf, int *scounts, int *sdisps, int err; char *psnd; char *prcv; + size_t type_size; int nreqs; MPI_Aint sndextent; MPI_Aint rcvextent; @@ -171,12 +185,10 @@ mca_coll_basic_alltoallv_intra(void *sbuf, int *scounts, int *sdisps, psnd = ((char *) sbuf) + (sdisps[rank] * sndextent); prcv = ((char *) rbuf) + (rdisps[rank] * rcvextent); - if (0 != scounts[rank]) { - err = ompi_datatype_sndrcv(psnd, scounts[rank], sdtype, - prcv, rcounts[rank], rdtype); - if (MPI_SUCCESS != err) { - return err; - } + err = ompi_datatype_sndrcv(psnd, scounts[rank], sdtype, + prcv, rcounts[rank], rdtype); + if (MPI_SUCCESS != err) { + return err; } /* If only one process, we're done. */ @@ -192,8 +204,10 @@ mca_coll_basic_alltoallv_intra(void *sbuf, int *scounts, int *sdisps, /* Post all receives first -- a simple optimization */ + ompi_datatype_type_size(rdtype, &type_size); for (i = 0; i < size; ++i) { - if (i == rank || 0 == rcounts[i]) { + size_t msg_size = type_size * rcounts[i]; + if (i == rank || 0 == msg_size) { continue; } @@ -210,8 +224,10 @@ mca_coll_basic_alltoallv_intra(void *sbuf, int *scounts, int *sdisps, /* Now post all sends */ + ompi_datatype_type_size(sdtype, &type_size); for (i = 0; i < size; ++i) { - if (i == rank || 0 == scounts[i]) { + size_t msg_size = type_size * scounts[i]; + if (i == rank || 0 == msg_size) { continue; } @@ -271,7 +287,7 @@ mca_coll_basic_alltoallv_inter(void *sbuf, int *scounts, int *sdisps, int err; char *psnd; char *prcv; - size_t nreqs; + size_t type_size, nreqs; MPI_Aint sndextent; MPI_Aint rcvextent; @@ -290,9 +306,11 @@ mca_coll_basic_alltoallv_inter(void *sbuf, int *scounts, int *sdisps, /* Post all receives first */ /* A simple optimization: do not send and recv msgs of length zero */ + ompi_datatype_type_size(rdtype, &type_size); for (i = 0; i < rsize; ++i) { + size_t msg_size = type_size * rcounts[i]; prcv = ((char *) rbuf) + (rdisps[i] * rcvextent); - if (rcounts[i] > 0) { + if (msg_size > 0) { err = MCA_PML_CALL(irecv(prcv, rcounts[i], rdtype, i, MCA_COLL_BASE_TAG_ALLTOALLV, comm, &preq[i])); @@ -305,9 +323,11 @@ mca_coll_basic_alltoallv_inter(void *sbuf, int *scounts, int *sdisps, } /* Now post all sends */ + ompi_datatype_type_size(sdtype, &type_size); for (i = 0; i < rsize; ++i) { + size_t msg_size = type_size * scounts[i]; psnd = ((char *) sbuf) + (sdisps[i] * sndextent); - if (scounts[i] > 0) { + if (msg_size > 0) { err = MCA_PML_CALL(isend(psnd, scounts[i], sdtype, i, MCA_COLL_BASE_TAG_ALLTOALLV, MCA_PML_BASE_SEND_STANDARD, comm, diff --git a/ompi/mca/coll/basic/coll_basic_alltoallw.c b/ompi/mca/coll/basic/coll_basic_alltoallw.c index 832ab4e1e5..cc54fed142 100644 --- a/ompi/mca/coll/basic/coll_basic_alltoallw.c +++ b/ompi/mca/coll/basic/coll_basic_alltoallw.c @@ -14,6 +14,9 @@ * Copyright (c) 2013 Los Alamos National Security, LLC. All rights * reserved. * Copyright (c) 2013 FUJITSU LIMITED. All rights reserved. + * Copyright (c) 2014 Research Organization for Information Science + * and Technology (RIST). All rights reserved. + * Copyright (c) 2014 Cisco Systems, Inc. All rights reserved. * $COPYRIGHT$ * * Additional copyrights may follow @@ -39,7 +42,7 @@ mca_coll_basic_alltoallw_intra_inplace(void *rbuf, int *rcounts, const int *rdis mca_coll_base_module_t *module) { mca_coll_basic_module_t *basic_module = (mca_coll_basic_module_t*) module; - int i, j, size, rank, err=MPI_SUCCESS, max_size; + int i, j, size, rank, err=MPI_SUCCESS, max_size, nreqs; MPI_Request *preq; char *tmp_buffer; ptrdiff_t ext; @@ -70,13 +73,19 @@ mca_coll_basic_alltoallw_intra_inplace(void *rbuf, int *rcounts, const int *rdis /* in-place alltoallw slow algorithm (but works) */ for (i = 0 ; i < size ; ++i) { + size_t msg_size_i; + ompi_datatype_type_size(rdtypes[i], &msg_size_i); + msg_size_i *= rcounts[i]; for (j = i+1 ; j < size ; ++j) { - ompi_datatype_type_extent (rdtypes[j], &ext); + size_t msg_size_j; + ompi_datatype_type_size(rdtypes[j], &msg_size_j); + msg_size_j *= rcounts[j]; /* Initiate all send/recv to/from others. */ + nreqs = 0; preq = basic_module->mccb_reqs; - if (i == rank && rcounts[j] != 0) { + if (i == rank && msg_size_j != 0) { /* Copy the data into the temporary buffer */ err = ompi_datatype_copy_content_same_ddt (rdtypes[j], rcounts[j], tmp_buffer, (char *) rbuf + rdisps[j]); @@ -85,13 +94,15 @@ mca_coll_basic_alltoallw_intra_inplace(void *rbuf, int *rcounts, const int *rdis /* Exchange data with the peer */ err = MCA_PML_CALL(irecv ((char *) rbuf + rdisps[j], rcounts[j], rdtypes[j], j, MCA_COLL_BASE_TAG_ALLTOALLW, comm, preq++)); + ++nreqs; if (MPI_SUCCESS != err) { goto error_hndl; } err = MCA_PML_CALL(isend ((void *) tmp_buffer, rcounts[j], rdtypes[j], j, MCA_COLL_BASE_TAG_ALLTOALLW, MCA_PML_BASE_SEND_STANDARD, comm, preq++)); + ++nreqs; if (MPI_SUCCESS != err) { goto error_hndl; } - } else if (j == rank && rcounts[i] != 0) { + } else if (j == rank && msg_size_i != 0) { /* Copy the data into the temporary buffer */ err = ompi_datatype_copy_content_same_ddt (rdtypes[i], rcounts[i], tmp_buffer, (char *) rbuf + rdisps[i]); @@ -100,22 +111,24 @@ mca_coll_basic_alltoallw_intra_inplace(void *rbuf, int *rcounts, const int *rdis /* Exchange data with the peer */ err = MCA_PML_CALL(irecv ((char *) rbuf + rdisps[i], rcounts[i], rdtypes[i], i, MCA_COLL_BASE_TAG_ALLTOALLW, comm, preq++)); + ++nreqs; if (MPI_SUCCESS != err) { goto error_hndl; } err = MCA_PML_CALL(isend ((void *) tmp_buffer, rcounts[i], rdtypes[i], i, MCA_COLL_BASE_TAG_ALLTOALLW, MCA_PML_BASE_SEND_STANDARD, comm, preq++)); + ++nreqs; if (MPI_SUCCESS != err) { goto error_hndl; } } else { continue; } /* Wait for the requests to complete */ - err = ompi_request_wait_all (2, basic_module->mccb_reqs, MPI_STATUS_IGNORE); + err = ompi_request_wait_all (nreqs, basic_module->mccb_reqs, MPI_STATUSES_IGNORE); if (MPI_SUCCESS != err) { goto error_hndl; } /* Free the requests. */ - mca_coll_basic_free_reqs(basic_module->mccb_reqs, 2); + mca_coll_basic_free_reqs(basic_module->mccb_reqs, nreqs); } } @@ -168,12 +181,10 @@ mca_coll_basic_alltoallw_intra(void *sbuf, int *scounts, int *sdisps, psnd = ((char *) sbuf) + sdisps[rank]; prcv = ((char *) rbuf) + rdisps[rank]; - if (0 != scounts[rank]) { - err = ompi_datatype_sndrcv(psnd, scounts[rank], sdtypes[rank], - prcv, rcounts[rank], rdtypes[rank]); - if (MPI_SUCCESS != err) { - return err; - } + err = ompi_datatype_sndrcv(psnd, scounts[rank], sdtypes[rank], + prcv, rcounts[rank], rdtypes[rank]); + if (MPI_SUCCESS != err) { + return err; } /* If only one process, we're done. */ @@ -190,7 +201,11 @@ mca_coll_basic_alltoallw_intra(void *sbuf, int *scounts, int *sdisps, /* Post all receives first -- a simple optimization */ for (i = 0; i < size; ++i) { - if (i == rank || 0 == rcounts[i]) + size_t msg_size; + ompi_datatype_type_size(rdtypes[i], &msg_size); + msg_size *= rcounts[i]; + + if (i == rank || 0 == msg_size) continue; prcv = ((char *) rbuf) + rdisps[i]; @@ -208,7 +223,11 @@ mca_coll_basic_alltoallw_intra(void *sbuf, int *scounts, int *sdisps, /* Now post all sends */ for (i = 0; i < size; ++i) { - if (i == rank || 0 == scounts[i]) + size_t msg_size; + ompi_datatype_type_size(sdtypes[i], &msg_size); + msg_size *= scounts[i]; + + if (i == rank || 0 == msg_size) continue; psnd = ((char *) sbuf) + sdisps[i];