This is an automated email from the git hooks/post-receive script. It was generated because a ref change was pushed to the repository containing the project "MPICH primary repository". The branch, master has been updated via 051449e7b5421b006556ba4e8c6f3300b6efed06 (commit) from 818b6a61a2f4653dc12da9c14c46ab866ba2a02b (commit) Those revisions listed above that are new to this repository have not appeared on any other notification email; so we list those revisions in full, below. - Log ----------------------------------------------------------------- http://git.mpich.org/mpich.git/commitdiff/051449e7b5421b006556ba4e8c6f3300b6... commit 051449e7b5421b006556ba4e8c6f3300b6efed06 Author: Igor Ivanov <[email protected]> Date: Fri Sep 5 18:18:49 2014 +0300 mpi/comm: Fix MPI_Intercomm_merge Previous code is based on a trick to get true context id using invalid communicator. There is a call of MPIR_Get_contextid() for new communicator that is not completelly built (no comm_commit and comm_hook calls). netmod/mxm uses comm_hook and can not work with this trick. This changes allow to avoid call of invalid communicator using temporary (intermediate) communicator. Signed-off-by: Igor Ivanov <[email protected]> Signed-off-by: Pavan Balaji <[email protected]> diff --git a/src/mpi/comm/intercomm_merge.c b/src/mpi/comm/intercomm_merge.c index 90b7c15..e12d991 100644 --- a/src/mpi/comm/intercomm_merge.c +++ b/src/mpi/comm/intercomm_merge.c @@ -20,6 +20,62 @@ int MPI_Intercomm_merge(MPI_Comm intercomm, int high, MPI_Comm *newintracomm) __ #endif /* -- End Profiling Symbol Block */ +/* These functions help implement the merge procedure */ +static int MPIR_Intercomm_merge_create_and_map_vcrt(MPID_Comm *comm_ptr, int local_high, MPID_Comm *new_intracomm_ptr); + + +/* This function creates VCRT for new communicator + * basing on VCRT of existing communicator. + */ + +#undef FUNCNAME +#define FUNCNAME MPIR_Intercomm_merge_create_and_map_vcrt +#undef FCNAME +#define FCNAME MPIDI_QUOTE(FUNCNAME) +static int MPIR_Intercomm_merge_create_and_map_vcrt(MPID_Comm *comm_ptr, int local_high, MPID_Comm *new_intracomm_ptr) +{ + int mpi_errno = MPI_SUCCESS; + int i, j; + + /* Now we know which group comes first. Build the new vcr + from the existing vcrs */ + mpi_errno = MPID_VCRT_Create( comm_ptr->local_size + comm_ptr->remote_size, &new_intracomm_ptr->vcrt ); + if (mpi_errno) MPIU_ERR_POP(mpi_errno); + + mpi_errno = MPID_VCRT_Get_ptr( new_intracomm_ptr->vcrt, &new_intracomm_ptr->vcr ); + if (mpi_errno) MPIU_ERR_POP(mpi_errno); + + if (local_high) { + /* remote group first */ + j = 0; + for (i = 0; i < comm_ptr->remote_size; i++) { + mpi_errno = MPID_VCR_Dup( comm_ptr->vcr[i], &new_intracomm_ptr->vcr[j++] ); + if (mpi_errno) MPIU_ERR_POP(mpi_errno); + } + for (i = 0; i < comm_ptr->local_size; i++) { + if (i == comm_ptr->rank) new_intracomm_ptr->rank = j; + mpi_errno = MPID_VCR_Dup( comm_ptr->local_vcr[i], &new_intracomm_ptr->vcr[j++] ); + if (mpi_errno) MPIU_ERR_POP(mpi_errno); + } + } + else { + /* local group first */ + j = 0; + for (i = 0; i < comm_ptr->local_size; i++) { + if (i == comm_ptr->rank) new_intracomm_ptr->rank = j; + mpi_errno = MPID_VCR_Dup( comm_ptr->local_vcr[i], &new_intracomm_ptr->vcr[j++] ); + if (mpi_errno) MPIU_ERR_POP(mpi_errno); + } + for (i = 0; i < comm_ptr->remote_size; i++) { + mpi_errno = MPID_VCR_Dup( comm_ptr->vcr[i], &new_intracomm_ptr->vcr[j++] ); + if (mpi_errno) MPIU_ERR_POP(mpi_errno); + } + } + +fn_fail: + return mpi_errno; +} + /* Define MPICH_MPI_FROM_PMPI if weak symbols are not supported to build the MPI routines */ #ifndef MPICH_MPI_FROM_PMPI @@ -33,7 +89,7 @@ int MPI_Intercomm_merge(MPI_Comm intercomm, int high, MPI_Comm *newintracomm) __ int MPIR_Intercomm_merge_impl(MPID_Comm *comm_ptr, int high, MPID_Comm **new_intracomm_ptr) { int mpi_errno = MPI_SUCCESS; - int local_high, remote_high, i, j, new_size; + int local_high, remote_high, new_size; MPIR_Context_id_t new_context_id; int errflag = FALSE; MPID_MPI_STATE_DECL(MPID_STATE_MPIR_INTERCOMM_MERGE_IMPL); @@ -118,44 +174,17 @@ int MPIR_Intercomm_merge_impl(MPID_Comm *comm_ptr, int high, MPID_Comm **new_int /* Now we know which group comes first. Build the new vcr from the existing vcrs */ - mpi_errno = MPID_VCRT_Create( new_size, &(*new_intracomm_ptr)->vcrt ); - if (mpi_errno) MPIU_ERR_POP(mpi_errno); - - mpi_errno = MPID_VCRT_Get_ptr( (*new_intracomm_ptr)->vcrt, &(*new_intracomm_ptr)->vcr ); + mpi_errno = MPIR_Intercomm_merge_create_and_map_vcrt(comm_ptr, local_high, (*new_intracomm_ptr)); if (mpi_errno) MPIU_ERR_POP(mpi_errno); - - if (local_high) { - /* remote group first */ - j = 0; - for (i = 0; i < comm_ptr->remote_size; i++) { - mpi_errno = MPID_VCR_Dup( comm_ptr->vcr[i], &(*new_intracomm_ptr)->vcr[j++] ); - if (mpi_errno) MPIU_ERR_POP(mpi_errno); - } - for (i = 0; i < comm_ptr->local_size; i++) { - if (i == comm_ptr->rank) (*new_intracomm_ptr)->rank = j; - mpi_errno = MPID_VCR_Dup( comm_ptr->local_vcr[i], &(*new_intracomm_ptr)->vcr[j++] ); - if (mpi_errno) MPIU_ERR_POP(mpi_errno); - } - } - else { - /* local group first */ - j = 0; - for (i = 0; i < comm_ptr->local_size; i++) { - if (i == comm_ptr->rank) (*new_intracomm_ptr)->rank = j; - mpi_errno = MPID_VCR_Dup( comm_ptr->local_vcr[i], &(*new_intracomm_ptr)->vcr[j++] ); - if (mpi_errno) MPIU_ERR_POP(mpi_errno); - } - for (i = 0; i < comm_ptr->remote_size; i++) { - mpi_errno = MPID_VCR_Dup( comm_ptr->vcr[i], &(*new_intracomm_ptr)->vcr[j++] ); - if (mpi_errno) MPIU_ERR_POP(mpi_errno); - } - } /* We've setup a temporary context id, based on the context id used by the intercomm. This allows us to perform the allreduce operations within the context id algorithm, since we already have a valid (almost - see comm_create_hook) communicator. */ + mpi_errno = MPIR_Comm_commit((*new_intracomm_ptr)); + if (mpi_errno) MPIU_ERR_POP(mpi_errno); + /* printf( "About to get context id \n" ); fflush( stdout ); */ /* In the multi-threaded case, MPIR_Get_contextid assumes that the calling routine already holds the single criticial section */ @@ -164,9 +193,24 @@ int MPIR_Intercomm_merge_impl(MPID_Comm *comm_ptr, int high, MPID_Comm **new_int if (mpi_errno) MPIU_ERR_POP(mpi_errno); MPIU_Assert(new_context_id != 0); + /* We release this communicator that was involved just to + * get valid context id and create true one + */ + mpi_errno = MPIR_Comm_release( (*new_intracomm_ptr),0); + if (mpi_errno) MPIU_ERR_POP(mpi_errno); + + mpi_errno = MPIR_Comm_create( new_intracomm_ptr ); + if (mpi_errno) MPIU_ERR_POP(mpi_errno); + + (*new_intracomm_ptr)->remote_size = (*new_intracomm_ptr)->local_size = new_size; + (*new_intracomm_ptr)->rank = -1; + (*new_intracomm_ptr)->comm_kind = MPID_INTRACOMM; (*new_intracomm_ptr)->context_id = new_context_id; (*new_intracomm_ptr)->recvcontext_id = new_context_id; + mpi_errno = MPIR_Intercomm_merge_create_and_map_vcrt(comm_ptr, local_high, (*new_intracomm_ptr)); + if (mpi_errno) MPIU_ERR_POP(mpi_errno); + mpi_errno = MPIR_Comm_commit((*new_intracomm_ptr)); if (mpi_errno) MPIU_ERR_POP(mpi_errno); ----------------------------------------------------------------------- Summary of changes: src/mpi/comm/intercomm_merge.c | 108 ++++++++++++++++++++++++++++------------ 1 files changed, 76 insertions(+), 32 deletions(-) hooks/post-receive -- MPICH primary repository