int MPIR_Barrier_intra( MPID_Comm *comm_ptr, MPIR_Errflag_t *errflag )
{
int size, rank, src, dst, mask, mpi_errno=MPI_SUCCESS;
int mpi_errno_ret = MPI_SUCCESS;
/* Only one collective operation per communicator can be active at any
time */
MPIDU_ERR_CHECK_MULTIPLE_THREADS_ENTER( comm_ptr );
size = comm_ptr->local_size;
/* Trivial barriers return immediately */
if (size == 1) goto fn_exit;
if (MPIR_CVAR_ENABLE_SMP_COLLECTIVES && MPIR_CVAR_ENABLE_SMP_BARRIER &&
MPIR_Comm_is_node_aware(comm_ptr)) {
mpi_errno = barrier_smp_intra(comm_ptr, errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag = MPIR_ERR_GET_CLASS(mpi_errno);
MPIR_ERR_SET(mpi_errno, *errflag, "**fail");
MPIR_ERR_ADD(mpi_errno_ret, mpi_errno);
}
goto fn_exit;
}
rank = comm_ptr->rank;
mask = 0x1;
while (mask < size) {
dst = (rank + mask) % size;
src = (rank - mask + size) % size;
mpi_errno = MPIC_Sendrecv(NULL, 0, MPI_BYTE, dst,
MPIR_BARRIER_TAG, NULL, 0, MPI_BYTE,
src, MPIR_BARRIER_TAG, comm_ptr,
MPI_STATUS_IGNORE, errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag = MPIR_ERR_GET_CLASS(mpi_errno);
MPIR_ERR_SET(mpi_errno, *errflag, "**fail");
MPIR_ERR_ADD(mpi_errno_ret, mpi_errno);
}
mask <<= 1;
}
fn_exit:
MPIDU_ERR_CHECK_MULTIPLE_THREADS_EXIT( comm_ptr );
if (mpi_errno_ret)
mpi_errno = mpi_errno_ret;
else if (*errflag != MPIR_ERR_NONE)
MPIR_ERR_SET(mpi_errno, *errflag, "**coll_fail");
return mpi_errno;
fn_fail:
goto fn_exit;
}
int MPIR_Barrier_intra( MPID_Comm *comm_ptr, int *errflag )
{
int size, rank, src, dst, mask, mpi_errno=MPI_SUCCESS;
int mpi_errno_ret = MPI_SUCCESS;
MPI_Comm comm;
/* Only one collective operation per communicator can be active at any
time */
MPIDU_ERR_CHECK_MULTIPLE_THREADS_ENTER( comm_ptr );
size = comm_ptr->local_size;
/* Trivial barriers return immediately */
if (size == 1) goto fn_exit;
rank = comm_ptr->rank;
comm = comm_ptr->handle;
mask = 0x1;
while (mask < size) {
dst = (rank + mask) % size;
src = (rank - mask + size) % size;
mpi_errno = MPIC_Sendrecv_ft(NULL, 0, MPI_BYTE, dst,
MPIR_BARRIER_TAG, NULL, 0, MPI_BYTE,
src, MPIR_BARRIER_TAG, comm,
MPI_STATUS_IGNORE, errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag = TRUE;
MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
}
mask <<= 1;
}
fn_exit:
MPIDU_ERR_CHECK_MULTIPLE_THREADS_EXIT( comm_ptr );
if (mpi_errno_ret)
mpi_errno = mpi_errno_ret;
else if (*errflag)
MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**coll_fail");
return mpi_errno;
fn_fail:
goto fn_exit;
}
int MPIR_Alltoallv_intra(const void *sendbuf, const int *sendcounts, const int *sdispls,
MPI_Datatype sendtype, void *recvbuf, const int *recvcounts,
const int *rdispls, MPI_Datatype recvtype, MPID_Comm *comm_ptr,
MPIR_Errflag_t *errflag)
{
int comm_size, i, j;
MPI_Aint send_extent, recv_extent;
int mpi_errno = MPI_SUCCESS;
int mpi_errno_ret = MPI_SUCCESS;
MPI_Status *starray;
MPI_Status status;
MPID_Request **reqarray;
int dst, rank, req_cnt;
int ii, ss, bblock;
int type_size;
MPIU_CHKLMEM_DECL(2);
comm_size = comm_ptr->local_size;
rank = comm_ptr->rank;
/* Get extent of recv type, but send type is only valid if (sendbuf!=MPI_IN_PLACE) */
MPID_Datatype_get_extent_macro(recvtype, recv_extent);
/* check if multiple threads are calling this collective function */
MPIDU_ERR_CHECK_MULTIPLE_THREADS_ENTER( comm_ptr );
if (sendbuf == MPI_IN_PLACE) {
/* We use pair-wise sendrecv_replace in order to conserve memory usage,
* which is keeping with the spirit of the MPI-2.2 Standard. But
* because of this approach all processes must agree on the global
* schedule of sendrecv_replace operations to avoid deadlock.
*
* Note that this is not an especially efficient algorithm in terms of
* time and there will be multiple repeated malloc/free's rather than
* maintaining a single buffer across the whole loop. Something like
* MADRE is probably the best solution for the MPI_IN_PLACE scenario. */
for (i = 0; i < comm_size; ++i) {
/* start inner loop at i to avoid re-exchanging data */
for (j = i; j < comm_size; ++j) {
if (rank == i) {
/* also covers the (rank == i && rank == j) case */
mpi_errno = MPIC_Sendrecv_replace(((char *)recvbuf + rdispls[j]*recv_extent),
recvcounts[j], recvtype,
j, MPIR_ALLTOALLV_TAG,
j, MPIR_ALLTOALLV_TAG,
comm_ptr, &status, errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag = MPIR_ERR_GET_CLASS(mpi_errno);
MPIR_ERR_SET(mpi_errno, *errflag, "**fail");
MPIR_ERR_ADD(mpi_errno_ret, mpi_errno);
}
}
else if (rank == j) {
/* same as above with i/j args reversed */
mpi_errno = MPIC_Sendrecv_replace(((char *)recvbuf + rdispls[i]*recv_extent),
recvcounts[i], recvtype,
i, MPIR_ALLTOALLV_TAG,
i, MPIR_ALLTOALLV_TAG,
comm_ptr, &status, errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag = MPIR_ERR_GET_CLASS(mpi_errno);
MPIR_ERR_SET(mpi_errno, *errflag, "**fail");
MPIR_ERR_ADD(mpi_errno_ret, mpi_errno);
}
}
}
}
}
else {
bblock = MPIR_CVAR_ALLTOALL_THROTTLE;
if (bblock == 0) bblock = comm_size;
MPID_Datatype_get_extent_macro(sendtype, send_extent);
MPIU_CHKLMEM_MALLOC(starray, MPI_Status*, 2*bblock*sizeof(MPI_Status), mpi_errno, "starray");
MPIU_CHKLMEM_MALLOC(reqarray, MPID_Request**, 2*bblock*sizeof(MPID_Request *), mpi_errno, "reqarray");
/* post only bblock isends/irecvs at a time as suggested by Tony Ladd */
for (ii=0; ii<comm_size; ii+=bblock) {
req_cnt = 0;
ss = comm_size-ii < bblock ? comm_size-ii : bblock;
/* do the communication -- post ss sends and receives: */
for ( i=0; i<ss; i++ ) {
dst = (rank+i+ii) % comm_size;
if (recvcounts[dst]) {
MPID_Datatype_get_size_macro(recvtype, type_size);
if (type_size) {
MPIU_Ensure_Aint_fits_in_pointer(MPIU_VOID_PTR_CAST_TO_MPI_AINT recvbuf +
rdispls[dst]*recv_extent);
mpi_errno = MPIC_Irecv((char *)recvbuf+rdispls[dst]*recv_extent,
recvcounts[dst], recvtype, dst,
MPIR_ALLTOALLV_TAG, comm_ptr,
&reqarray[req_cnt]);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag = MPIR_ERR_GET_CLASS(mpi_errno);
MPIR_ERR_SET(mpi_errno, *errflag, "**fail");
MPIR_ERR_ADD(mpi_errno_ret, mpi_errno);
}
req_cnt++;
}
}
}
for ( i=0; i<ss; i++ ) {
dst = (rank-i-ii+comm_size) % comm_size;
if (sendcounts[dst]) {
MPID_Datatype_get_size_macro(sendtype, type_size);
if (type_size) {
MPIU_Ensure_Aint_fits_in_pointer(MPIU_VOID_PTR_CAST_TO_MPI_AINT sendbuf +
sdispls[dst]*send_extent);
mpi_errno = MPIC_Isend((char *)sendbuf+sdispls[dst]*send_extent,
sendcounts[dst], sendtype, dst,
MPIR_ALLTOALLV_TAG, comm_ptr,
&reqarray[req_cnt], errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag = MPIR_ERR_GET_CLASS(mpi_errno);
MPIR_ERR_SET(mpi_errno, *errflag, "**fail");
MPIR_ERR_ADD(mpi_errno_ret, mpi_errno);
}
req_cnt++;
}
}
}
mpi_errno = MPIC_Waitall(req_cnt, reqarray, starray, errflag);
if (mpi_errno && mpi_errno != MPI_ERR_IN_STATUS) MPIR_ERR_POP(mpi_errno);
/* --BEGIN ERROR HANDLING-- */
if (mpi_errno == MPI_ERR_IN_STATUS) {
for (i=0; i<req_cnt; i++) {
if (starray[i].MPI_ERROR != MPI_SUCCESS) {
mpi_errno = starray[i].MPI_ERROR;
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag = MPIR_ERR_GET_CLASS(mpi_errno);
MPIR_ERR_SET(mpi_errno, *errflag, "**fail");
MPIR_ERR_ADD(mpi_errno_ret, mpi_errno);
}
}
}
}
/* --END ERROR HANDLING-- */
}
}
fn_exit:
/* check if multiple threads are calling this collective function */
MPIDU_ERR_CHECK_MULTIPLE_THREADS_EXIT( comm_ptr );
MPIU_CHKLMEM_FREEALL();
if (mpi_errno_ret)
mpi_errno = mpi_errno_ret;
else if (*errflag != MPIR_ERR_NONE)
MPIR_ERR_SET(mpi_errno, *errflag, "**coll_fail");
return mpi_errno;
fn_fail:
goto fn_exit;
}
int MPIR_Alltoallv_inter(const void *sendbuf, const int *sendcounts, const int *sdispls,
MPI_Datatype sendtype, void *recvbuf, const int *recvcounts,
const int *rdispls, MPI_Datatype recvtype, MPID_Comm *comm_ptr,
MPIR_Errflag_t *errflag)
{
/* Intercommunicator alltoallv. We use a pairwise exchange algorithm
similar to the one used in intracommunicator alltoallv. Since the
local and remote groups can be of different
sizes, we first compute the max of local_group_size,
remote_group_size. At step i, 0 <= i < max_size, each process
receives from src = (rank - i + max_size) % max_size if src <
remote_size, and sends to dst = (rank + i) % max_size if dst <
remote_size.
FIXME: change algorithm to match intracommunicator alltoallv
*/
int local_size, remote_size, max_size, i;
MPI_Aint send_extent, recv_extent;
int mpi_errno = MPI_SUCCESS;
int mpi_errno_ret = MPI_SUCCESS;
MPI_Status status;
int src, dst, rank, sendcount, recvcount;
char *sendaddr, *recvaddr;
local_size = comm_ptr->local_size;
remote_size = comm_ptr->remote_size;
rank = comm_ptr->rank;
/* Get extent of send and recv types */
MPID_Datatype_get_extent_macro(sendtype, send_extent);
MPID_Datatype_get_extent_macro(recvtype, recv_extent);
/* check if multiple threads are calling this collective function */
MPIDU_ERR_CHECK_MULTIPLE_THREADS_ENTER( comm_ptr );
/* Use pairwise exchange algorithm. */
max_size = MPIR_MAX(local_size, remote_size);
for (i=0; i<max_size; i++) {
src = (rank - i + max_size) % max_size;
dst = (rank + i) % max_size;
if (src >= remote_size) {
src = MPI_PROC_NULL;
recvaddr = NULL;
recvcount = 0;
}
else {
MPIU_Ensure_Aint_fits_in_pointer(MPIU_VOID_PTR_CAST_TO_MPI_AINT recvbuf +
rdispls[src]*recv_extent);
recvaddr = (char *)recvbuf + rdispls[src]*recv_extent;
recvcount = recvcounts[src];
}
if (dst >= remote_size) {
dst = MPI_PROC_NULL;
sendaddr = NULL;
sendcount = 0;
}
else {
MPIU_Ensure_Aint_fits_in_pointer(MPIU_VOID_PTR_CAST_TO_MPI_AINT sendbuf +
sdispls[dst]*send_extent);
sendaddr = (char *)sendbuf + sdispls[dst]*send_extent;
sendcount = sendcounts[dst];
}
mpi_errno = MPIC_Sendrecv(sendaddr, sendcount, sendtype, dst,
MPIR_ALLTOALLV_TAG, recvaddr, recvcount,
recvtype, src, MPIR_ALLTOALLV_TAG,
comm_ptr, &status, errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag = MPIR_ERR_GET_CLASS(mpi_errno);
MPIR_ERR_SET(mpi_errno, *errflag, "**fail");
MPIR_ERR_ADD(mpi_errno_ret, mpi_errno);
}
}
fn_exit:
/* check if multiple threads are calling this collective function */
MPIDU_ERR_CHECK_MULTIPLE_THREADS_EXIT( comm_ptr );
if (mpi_errno_ret)
mpi_errno = mpi_errno_ret;
else if (*errflag != MPIR_ERR_NONE)
MPIR_ERR_SET(mpi_errno, *errflag, "**coll_fail");
return mpi_errno;
fn_fail:
goto fn_exit;
}
int MPIR_Bcast_inter_MV2 (
void *buffer,
int count,
MPI_Datatype datatype,
int root,
MPID_Comm *comm_ptr,
int *errflag)
{
/* Intercommunicator broadcast.
Root sends to rank 0 in remote group. Remote group does local
intracommunicator broadcast.
*/
int rank, mpi_errno;
int mpi_errno_ret = MPI_SUCCESS;
MPI_Status status;
MPID_Comm *newcomm_ptr = NULL;
MPI_Comm comm;
MPID_MPI_STATE_DECL(MPID_STATE_MPIR_BCAST_INTER);
MPID_MPI_FUNC_ENTER(MPID_STATE_MPIR_BCAST_INTER);
comm = comm_ptr->handle;
if (root == MPI_PROC_NULL)
{
/* local processes other than root do nothing */
mpi_errno = MPI_SUCCESS;
}
else if (root == MPI_ROOT)
{
/* root sends to rank 0 on remote group and returns */
MPIDU_ERR_CHECK_MULTIPLE_THREADS_ENTER( comm_ptr );
mpi_errno = MPIC_Send_ft(buffer, count, datatype, 0,
MPIR_BCAST_TAG, comm, errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag = TRUE;
MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
}
MPIDU_ERR_CHECK_MULTIPLE_THREADS_EXIT( comm_ptr );
}
else
{
/* remote group. rank 0 on remote group receives from root */
rank = comm_ptr->rank;
if (rank == 0)
{
mpi_errno = MPIC_Recv_ft(buffer, count, datatype, root,
MPIR_BCAST_TAG, comm, &status, errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag = TRUE;
MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
}
}
/* Get the local intracommunicator */
if (!comm_ptr->local_comm) {
MPIR_Setup_intercomm_localcomm( comm_ptr );
}
newcomm_ptr = comm_ptr->local_comm;
/* now do the usual broadcast on this intracommunicator
with rank 0 as root. */
mpi_errno = MPIR_Bcast_intra_MV2(buffer, count, datatype, 0, newcomm_ptr, errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag = TRUE;
MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
}
}
MPID_MPI_FUNC_EXIT(MPID_STATE_MPIR_BCAST_INTER);
if (mpi_errno_ret)
mpi_errno = mpi_errno_ret;
else if (*errflag)
MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**coll_fail");
return mpi_errno;
}
int MPIR_Allreduce_intra (
void *sendbuf,
void *recvbuf,
int count,
MPI_Datatype datatype,
MPI_Op op,
MPID_Comm *comm_ptr,
int *errflag )
{
int is_homogeneous;
#ifdef MPID_HAS_HETERO
int rc;
#endif
int comm_size, rank, type_size;
int mpi_errno = MPI_SUCCESS;
int mpi_errno_ret = MPI_SUCCESS;
int mask, dst, is_commutative, pof2, newrank, rem, newdst, i,
send_idx, recv_idx, last_idx, send_cnt, recv_cnt, *cnts, *disps;
MPI_Aint true_extent, true_lb, extent;
void *tmp_buf;
MPI_Comm comm;
MPIU_CHKLMEM_DECL(3);
/* check if multiple threads are calling this collective function */
MPIDU_ERR_CHECK_MULTIPLE_THREADS_ENTER( comm_ptr );
if (count == 0) goto fn_exit;
comm = comm_ptr->handle;
is_commutative = MPIR_Op_is_commutative(op);
#if defined(USE_SMP_COLLECTIVES)
/* is the op commutative? We do SMP optimizations only if it is. */
if (MPIR_Comm_is_node_aware(comm_ptr) && is_commutative) {
/* on each node, do a reduce to the local root */
if (comm_ptr->node_comm != NULL) {
/* take care of the MPI_IN_PLACE case. For reduce,
MPI_IN_PLACE is specified only on the root;
for allreduce it is specified on all processes. */
if ((sendbuf == MPI_IN_PLACE) && (comm_ptr->node_comm->rank != 0)) {
/* IN_PLACE and not root of reduce. Data supplied to this
allreduce is in recvbuf. Pass that as the sendbuf to reduce. */
mpi_errno = MPIR_Reduce_impl(recvbuf, NULL, count, datatype, op, 0, comm_ptr->node_comm, errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag = TRUE;
MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
}
} else {
mpi_errno = MPIR_Reduce_impl(sendbuf, recvbuf, count, datatype, op, 0, comm_ptr->node_comm, errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag = TRUE;
MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
}
}
} else {
/* only one process on the node. copy sendbuf to recvbuf */
if (sendbuf != MPI_IN_PLACE) {
mpi_errno = MPIR_Localcopy(sendbuf, count, datatype, recvbuf, count, datatype);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
}
}
/* now do an IN_PLACE allreduce among the local roots of all nodes */
if (comm_ptr->node_roots_comm != NULL) {
mpi_errno = allreduce_intra_or_coll_fn(MPI_IN_PLACE, recvbuf, count, datatype, op, comm_ptr->node_roots_comm,
errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag = TRUE;
MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
}
}
/* now broadcast the result among local processes */
if (comm_ptr->node_comm != NULL) {
mpi_errno = MPIR_Bcast_impl(recvbuf, count, datatype, 0, comm_ptr->node_comm, errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag = TRUE;
MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
}
}
goto fn_exit;
}
#endif
is_homogeneous = 1;
#ifdef MPID_HAS_HETERO
if (comm_ptr->is_hetero)
is_homogeneous = 0;
#endif
#ifdef MPID_HAS_HETERO
if (!is_homogeneous) {
/* heterogeneous. To get the same result on all processes, we
do a reduce to 0 and then broadcast. */
mpi_errno = MPIR_Reduce_impl ( sendbuf, recvbuf, count, datatype,
op, 0, comm_ptr, errflag );
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag = TRUE;
MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
}
mpi_errno = MPIR_Bcast_impl( recvbuf, count, datatype, 0, comm_ptr, errflag );
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag = TRUE;
MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
}
}
else
#endif /* MPID_HAS_HETERO */
{
/* homogeneous */
comm_size = comm_ptr->local_size;
rank = comm_ptr->rank;
is_commutative = MPIR_Op_is_commutative(op);
/* need to allocate temporary buffer to store incoming data*/
MPIR_Type_get_true_extent_impl(datatype, &true_lb, &true_extent);
MPID_Datatype_get_extent_macro(datatype, extent);
MPID_Ensure_Aint_fits_in_pointer(count * MPIR_MAX(extent, true_extent));
MPIU_CHKLMEM_MALLOC(tmp_buf, void *, count*(MPIR_MAX(extent,true_extent)), mpi_errno, "temporary buffer");
/* adjust for potential negative lower bound in datatype */
tmp_buf = (void *)((char*)tmp_buf - true_lb);
/* copy local data into recvbuf */
if (sendbuf != MPI_IN_PLACE) {
mpi_errno = MPIR_Localcopy(sendbuf, count, datatype, recvbuf,
count, datatype);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
}
MPID_Datatype_get_size_macro(datatype, type_size);
/* find nearest power-of-two less than or equal to comm_size */
pof2 = 1;
while (pof2 <= comm_size) pof2 <<= 1;
pof2 >>=1;
rem = comm_size - pof2;
/* In the non-power-of-two case, all even-numbered
processes of rank < 2*rem send their data to
(rank+1). These even-numbered processes no longer
participate in the algorithm until the very end. The
remaining processes form a nice power-of-two. */
if (rank < 2*rem) {
if (rank % 2 == 0) { /* even */
mpi_errno = MPIC_Send_ft(recvbuf, count,
datatype, rank+1,
MPIR_ALLREDUCE_TAG, comm, errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag = TRUE;
MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
}
/* temporarily set the rank to -1 so that this
process does not pariticipate in recursive
doubling */
newrank = -1;
}
else { /* odd */
mpi_errno = MPIC_Recv_ft(tmp_buf, count,
datatype, rank-1,
MPIR_ALLREDUCE_TAG, comm,
MPI_STATUS_IGNORE, errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag = TRUE;
MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
}
/* do the reduction on received data. since the
ordering is right, it doesn't matter whether
the operation is commutative or not. */
mpi_errno = MPIR_Reduce_local_impl(tmp_buf, recvbuf, count, datatype, op);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
/* change the rank */
newrank = rank / 2;
}
}
else /* rank >= 2*rem */
newrank = rank - rem;
/* If op is user-defined or count is less than pof2, use
recursive doubling algorithm. Otherwise do a reduce-scatter
followed by allgather. (If op is user-defined,
derived datatypes are allowed and the user could pass basic
datatypes on one process and derived on another as long as
the type maps are the same. Breaking up derived
datatypes to do the reduce-scatter is tricky, therefore
using recursive doubling in that case.) */
if (newrank != -1) {
if ((count*type_size <= MPIR_PARAM_ALLREDUCE_SHORT_MSG_SIZE) ||
(HANDLE_GET_KIND(op) != HANDLE_KIND_BUILTIN) ||
(count < pof2)) { /* use recursive doubling */
mask = 0x1;
while (mask < pof2) {
newdst = newrank ^ mask;
/* find real rank of dest */
dst = (newdst < rem) ? newdst*2 + 1 : newdst + rem;
/* Send the most current data, which is in recvbuf. Recv
into tmp_buf */
mpi_errno = MPIC_Sendrecv_ft(recvbuf, count, datatype,
dst, MPIR_ALLREDUCE_TAG, tmp_buf,
count, datatype, dst,
MPIR_ALLREDUCE_TAG, comm,
MPI_STATUS_IGNORE, errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag = TRUE;
MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
}
/* tmp_buf contains data received in this step.
recvbuf contains data accumulated so far */
if (is_commutative || (dst < rank)) {
/* op is commutative OR the order is already right */
mpi_errno = MPIR_Reduce_local_impl(tmp_buf, recvbuf, count, datatype, op);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
}
else {
/* op is noncommutative and the order is not right */
mpi_errno = MPIR_Reduce_local_impl(recvbuf, tmp_buf, count, datatype, op);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
/* copy result back into recvbuf */
mpi_errno = MPIR_Localcopy(tmp_buf, count, datatype,
recvbuf, count, datatype);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
}
mask <<= 1;
}
}
else {
/* do a reduce-scatter followed by allgather */
/* for the reduce-scatter, calculate the count that
each process receives and the displacement within
the buffer */
MPIU_CHKLMEM_MALLOC(cnts, int *, pof2*sizeof(int), mpi_errno, "counts");
MPIU_CHKLMEM_MALLOC(disps, int *, pof2*sizeof(int), mpi_errno, "displacements");
for (i=0; i<(pof2-1); i++)
cnts[i] = count/pof2;
cnts[pof2-1] = count - (count/pof2)*(pof2-1);
disps[0] = 0;
for (i=1; i<pof2; i++)
disps[i] = disps[i-1] + cnts[i-1];
mask = 0x1;
send_idx = recv_idx = 0;
last_idx = pof2;
while (mask < pof2) {
newdst = newrank ^ mask;
/* find real rank of dest */
dst = (newdst < rem) ? newdst*2 + 1 : newdst + rem;
send_cnt = recv_cnt = 0;
if (newrank < newdst) {
send_idx = recv_idx + pof2/(mask*2);
for (i=send_idx; i<last_idx; i++)
send_cnt += cnts[i];
for (i=recv_idx; i<send_idx; i++)
recv_cnt += cnts[i];
}
else {
recv_idx = send_idx + pof2/(mask*2);
for (i=send_idx; i<recv_idx; i++)
send_cnt += cnts[i];
for (i=recv_idx; i<last_idx; i++)
recv_cnt += cnts[i];
}
/* printf("Rank %d, send_idx %d, recv_idx %d, send_cnt %d, recv_cnt %d, last_idx %d\n", newrank, send_idx, recv_idx,
send_cnt, recv_cnt, last_idx);
*/
/* Send data from recvbuf. Recv into tmp_buf */
mpi_errno = MPIC_Sendrecv_ft((char *) recvbuf +
disps[send_idx]*extent,
send_cnt, datatype,
dst, MPIR_ALLREDUCE_TAG,
(char *) tmp_buf +
disps[recv_idx]*extent,
recv_cnt, datatype, dst,
MPIR_ALLREDUCE_TAG, comm,
MPI_STATUS_IGNORE, errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag = TRUE;
MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
}
/* tmp_buf contains data received in this step.
recvbuf contains data accumulated so far */
/* This algorithm is used only for predefined ops
and predefined ops are always commutative. */
mpi_errno = MPIR_Reduce_local_impl(((char *) tmp_buf + disps[recv_idx]*extent),
((char *) recvbuf + disps[recv_idx]*extent),
recv_cnt, datatype, op);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
/* update send_idx for next iteration */
send_idx = recv_idx;
mask <<= 1;
/* update last_idx, but not in last iteration
because the value is needed in the allgather
step below. */
if (mask < pof2)
last_idx = recv_idx + pof2/mask;
}
/* now do the allgather */
mask >>= 1;
while (mask > 0) {
newdst = newrank ^ mask;
/* find real rank of dest */
dst = (newdst < rem) ? newdst*2 + 1 : newdst + rem;
send_cnt = recv_cnt = 0;
if (newrank < newdst) {
/* update last_idx except on first iteration */
if (mask != pof2/2)
last_idx = last_idx + pof2/(mask*2);
recv_idx = send_idx + pof2/(mask*2);
for (i=send_idx; i<recv_idx; i++)
send_cnt += cnts[i];
for (i=recv_idx; i<last_idx; i++)
recv_cnt += cnts[i];
}
else {
recv_idx = send_idx - pof2/(mask*2);
for (i=send_idx; i<last_idx; i++)
send_cnt += cnts[i];
for (i=recv_idx; i<send_idx; i++)
recv_cnt += cnts[i];
}
mpi_errno = MPIC_Sendrecv_ft((char *) recvbuf +
disps[send_idx]*extent,
send_cnt, datatype,
dst, MPIR_ALLREDUCE_TAG,
(char *) recvbuf +
disps[recv_idx]*extent,
recv_cnt, datatype, dst,
MPIR_ALLREDUCE_TAG, comm,
MPI_STATUS_IGNORE, errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag = TRUE;
MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
}
if (newrank > newdst) send_idx = recv_idx;
mask >>= 1;
}
}
}
/* In the non-power-of-two case, all odd-numbered
processes of rank < 2*rem send the result to
(rank-1), the ranks who didn't participate above. */
if (rank < 2*rem) {
if (rank % 2) /* odd */
mpi_errno = MPIC_Send_ft(recvbuf, count,
datatype, rank-1,
MPIR_ALLREDUCE_TAG, comm, errflag);
else /* even */
mpi_errno = MPIC_Recv_ft(recvbuf, count,
datatype, rank+1,
MPIR_ALLREDUCE_TAG, comm,
MPI_STATUS_IGNORE, errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag = TRUE;
MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
}
}
}
fn_exit:
/* check if multiple threads are calling this collective function */
MPIDU_ERR_CHECK_MULTIPLE_THREADS_EXIT( comm_ptr );
MPIU_CHKLMEM_FREEALL();
if (mpi_errno_ret)
mpi_errno = mpi_errno_ret;
return (mpi_errno);
fn_fail:
goto fn_exit;
}
static int MPIR_Scan_generic (
const void *sendbuf,
void *recvbuf,
int count,
MPI_Datatype datatype,
MPI_Op op,
MPID_Comm *comm_ptr,
int *errflag )
{
MPI_Status status;
int rank, comm_size;
int mpi_errno = MPI_SUCCESS;
int mpi_errno_ret = MPI_SUCCESS;
int mask, dst, is_commutative;
MPI_Aint true_extent, true_lb, extent;
void *partial_scan, *tmp_buf;
MPID_Op *op_ptr;
MPI_Comm comm;
MPIU_THREADPRIV_DECL;
MPIU_CHKLMEM_DECL(2);
if (count == 0) return MPI_SUCCESS;
/* check if multiple threads are calling this collective function */
MPIDU_ERR_CHECK_MULTIPLE_THREADS_ENTER( comm_ptr );
comm = comm_ptr->handle;
comm_size = comm_ptr->local_size;
rank = comm_ptr->rank;
MPIU_THREADPRIV_GET;
/* set op_errno to 0. stored in perthread structure */
MPIU_THREADPRIV_FIELD(op_errno) = 0;
if (HANDLE_GET_KIND(op) == HANDLE_KIND_BUILTIN) {
is_commutative = 1;
}
else {
MPID_Op_get_ptr(op, op_ptr);
if (op_ptr->kind == MPID_OP_USER_NONCOMMUTE)
is_commutative = 0;
else
is_commutative = 1;
}
/* need to allocate temporary buffer to store partial scan*/
MPIR_Type_get_true_extent_impl(datatype, &true_lb, &true_extent);
MPID_Datatype_get_extent_macro(datatype, extent);
MPIU_CHKLMEM_MALLOC(partial_scan, void *, count*(MPIR_MAX(extent,true_extent)), mpi_errno, "partial_scan");
/* This eventually gets malloc()ed as a temp buffer, not added to
* any user buffers */
MPID_Ensure_Aint_fits_in_pointer(count * MPIR_MAX(extent, true_extent));
/* adjust for potential negative lower bound in datatype */
partial_scan = (void *)((char*)partial_scan - true_lb);
/* need to allocate temporary buffer to store incoming data*/
MPIU_CHKLMEM_MALLOC(tmp_buf, void *, count*(MPIR_MAX(extent,true_extent)), mpi_errno, "tmp_buf");
/* adjust for potential negative lower bound in datatype */
tmp_buf = (void *)((char*)tmp_buf - true_lb);
/* Since this is an inclusive scan, copy local contribution into
recvbuf. */
if (sendbuf != MPI_IN_PLACE) {
mpi_errno = MPIR_Localcopy(sendbuf, count, datatype,
recvbuf, count, datatype);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
}
if (sendbuf != MPI_IN_PLACE)
mpi_errno = MPIR_Localcopy(sendbuf, count, datatype,
partial_scan, count, datatype);
else
mpi_errno = MPIR_Localcopy(recvbuf, count, datatype,
partial_scan, count, datatype);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
mask = 0x1;
while (mask < comm_size) {
dst = rank ^ mask;
if (dst < comm_size) {
/* Send partial_scan to dst. Recv into tmp_buf */
mpi_errno = MPIC_Sendrecv(partial_scan, count, datatype,
dst, MPIR_SCAN_TAG, tmp_buf,
count, datatype, dst,
MPIR_SCAN_TAG, comm,
&status, errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag = TRUE;
MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
}
if (rank > dst) {
mpi_errno = MPIR_Reduce_local_impl(
tmp_buf, partial_scan, count, datatype, op);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
mpi_errno = MPIR_Reduce_local_impl(
tmp_buf, recvbuf, count, datatype, op);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
}
else {
if (is_commutative) {
mpi_errno = MPIR_Reduce_local_impl(
tmp_buf, partial_scan, count, datatype, op);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
}
else {
mpi_errno = MPIR_Reduce_local_impl(
partial_scan, tmp_buf, count, datatype, op);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
mpi_errno = MPIR_Localcopy(tmp_buf, count, datatype,
partial_scan,
count, datatype);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
}
}
}
mask <<= 1;
}
if (MPIU_THREADPRIV_FIELD(op_errno)) {
mpi_errno = MPIU_THREADPRIV_FIELD(op_errno);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
}
fn_exit:
MPIU_CHKLMEM_FREEALL();
/* check if multiple threads are calling this collective function */
MPIDU_ERR_CHECK_MULTIPLE_THREADS_EXIT( comm_ptr );
if (mpi_errno_ret)
mpi_errno = mpi_errno_ret;
else if (*errflag)
MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**coll_fail");
return mpi_errno;
fn_fail:
goto fn_exit;
}
int MPIR_Exscan (
void *sendbuf,
void *recvbuf,
int count,
MPI_Datatype datatype,
MPI_Op op,
MPID_Comm *comm_ptr,
int *errflag )
{
MPI_Status status;
int rank, comm_size;
int mpi_errno = MPI_SUCCESS;
int mpi_errno_ret = MPI_SUCCESS;
int mask, dst, is_commutative, flag;
MPI_Aint true_extent, true_lb, extent;
void *partial_scan, *tmp_buf;
MPI_User_function *uop;
MPID_Op *op_ptr;
MPI_Comm comm;
MPIU_CHKLMEM_DECL(2);
MPIU_THREADPRIV_DECL;
#ifdef HAVE_CXX_BINDING
int is_cxx_uop = 0;
#endif
if (count == 0) return MPI_SUCCESS;
MPIU_THREADPRIV_GET;
comm = comm_ptr->handle;
comm_size = comm_ptr->local_size;
rank = comm_ptr->rank;
/* set op_errno to 0. stored in perthread structure */
MPIU_THREADPRIV_FIELD(op_errno) = 0;
if (HANDLE_GET_KIND(op) == HANDLE_KIND_BUILTIN) {
is_commutative = 1;
/* get the function by indexing into the op table */
uop = MPIR_Op_table[op%16 - 1];
}
else {
MPID_Op_get_ptr(op, op_ptr);
if (op_ptr->kind == MPID_OP_USER_NONCOMMUTE)
is_commutative = 0;
else
is_commutative = 1;
#ifdef HAVE_CXX_BINDING
if (op_ptr->language == MPID_LANG_CXX) {
uop = (MPI_User_function *) op_ptr->function.c_function;
is_cxx_uop = 1;
}
else
#endif
if ((op_ptr->language == MPID_LANG_C))
uop = (MPI_User_function *) op_ptr->function.c_function;
else
uop = (MPI_User_function *) op_ptr->function.f77_function;
}
/* need to allocate temporary buffer to store partial scan*/
MPIR_Type_get_true_extent_impl(datatype, &true_lb, &true_extent);
MPID_Datatype_get_extent_macro( datatype, extent );
MPIU_CHKLMEM_MALLOC(partial_scan, void *, (count*(MPIR_MAX(true_extent,extent))), mpi_errno, "partial_scan");
/* adjust for potential negative lower bound in datatype */
partial_scan = (void *)((char*)partial_scan - true_lb);
/* need to allocate temporary buffer to store incoming data*/
MPIU_CHKLMEM_MALLOC(tmp_buf, void *, (count*(MPIR_MAX(true_extent,extent))), mpi_errno, "tmp_buf");
/* adjust for potential negative lower bound in datatype */
tmp_buf = (void *)((char*)tmp_buf - true_lb);
mpi_errno = MPIR_Localcopy((sendbuf == MPI_IN_PLACE ? recvbuf : sendbuf), count, datatype,
partial_scan, count, datatype);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
/* check if multiple threads are calling this collective function */
MPIDU_ERR_CHECK_MULTIPLE_THREADS_ENTER( comm_ptr );
flag = 0;
mask = 0x1;
while (mask < comm_size) {
dst = rank ^ mask;
if (dst < comm_size) {
/* Send partial_scan to dst. Recv into tmp_buf */
mpi_errno = MPIC_Sendrecv_ft(partial_scan, count, datatype,
dst, MPIR_EXSCAN_TAG, tmp_buf,
count, datatype, dst,
MPIR_EXSCAN_TAG, comm,
&status, errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag = TRUE;
MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
}
if (rank > dst) {
call_uop(tmp_buf, partial_scan, count, datatype);
/* On rank 0, recvbuf is not defined. For sendbuf==MPI_IN_PLACE
recvbuf must not change (per MPI-2.2).
On rank 1, recvbuf is to be set equal to the value
in sendbuf on rank 0.
On others, recvbuf is the scan of values in the
sendbufs on lower ranks. */
if (rank != 0) {
if (flag == 0) {
/* simply copy data recd from rank 0 into recvbuf */
mpi_errno = MPIR_Localcopy(tmp_buf, count, datatype,
recvbuf, count, datatype);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
flag = 1;
}
else {
call_uop(tmp_buf, recvbuf, count, datatype);
}
}
}
else {
if (is_commutative) {
call_uop(tmp_buf, partial_scan, count, datatype);
}
else {
call_uop(partial_scan, tmp_buf, count, datatype);
mpi_errno = MPIR_Localcopy(tmp_buf, count, datatype,
partial_scan,
count, datatype);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
}
}
}
mask <<= 1;
}
/* check if multiple threads are calling this collective function */
MPIDU_ERR_CHECK_MULTIPLE_THREADS_EXIT( comm_ptr );
if (MPIU_THREADPRIV_FIELD(op_errno))
mpi_errno = MPIU_THREADPRIV_FIELD(op_errno);
fn_exit:
MPIU_CHKLMEM_FREEALL();
if (mpi_errno_ret)
mpi_errno = mpi_errno_ret;
else if (*errflag)
MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**coll_fail");
return mpi_errno;
fn_fail:
goto fn_exit;
}
int create_2level_comm (MPI_Comm comm, int size, int my_rank)
{
static const char FCNAME[] = "create_2level_comm";
int mpi_errno = MPI_SUCCESS;
MPID_Comm* comm_ptr;
MPID_Comm* comm_world_ptr;
MPI_Group subgroup1, comm_group;
MPID_Group *group_ptr=NULL;
int leader_comm_size, my_local_size, my_local_id, input_flag =0, output_flag=0;
int errflag = FALSE;
int leader_group_size=0;
MPIU_THREADPRIV_DECL;
MPIU_THREADPRIV_GET;
MPID_Comm_get_ptr( comm, comm_ptr );
MPID_Comm_get_ptr( MPI_COMM_WORLD, comm_world_ptr );
int* shmem_group = MPIU_Malloc(sizeof(int) * size);
if (NULL == shmem_group){
printf("Couldn't malloc shmem_group\n");
ibv_error_abort (GEN_EXIT_ERR, "create_2level_com");
}
/* Creating local shmem group */
int i = 0;
int local_rank = 0;
int grp_index = 0;
comm_ptr->ch.leader_comm=MPI_COMM_NULL;
comm_ptr->ch.shmem_comm=MPI_COMM_NULL;
MPIDI_VC_t* vc = NULL;
for (; i < size ; ++i){
MPIDI_Comm_get_vc(comm_ptr, i, &vc);
if (my_rank == i || vc->smp.local_rank >= 0){
shmem_group[grp_index] = i;
if (my_rank == i){
local_rank = grp_index;
}
++grp_index;
}
}
/* Creating leader group */
int leader = 0;
leader = shmem_group[0];
/* Gives the mapping to any process's leader in comm */
comm_ptr->ch.leader_map = MPIU_Malloc(sizeof(int) * size);
if (NULL == comm_ptr->ch.leader_map){
printf("Couldn't malloc group\n");
ibv_error_abort (GEN_EXIT_ERR, "create_2level_com");
}
mpi_errno = MPIR_Allgather_impl (&leader, 1, MPI_INT , comm_ptr->ch.leader_map, 1, MPI_INT, comm_ptr, &errflag);
if(mpi_errno) {
MPIU_ERR_POP(mpi_errno);
}
int* leader_group = MPIU_Malloc(sizeof(int) * size);
if (NULL == leader_group){
printf("Couldn't malloc leader_group\n");
ibv_error_abort (GEN_EXIT_ERR, "create_2level_com");
}
/* Gives the mapping from leader's rank in comm to
* leader's rank in leader_comm */
comm_ptr->ch.leader_rank = MPIU_Malloc(sizeof(int) * size);
if (NULL == comm_ptr->ch.leader_rank){
printf("Couldn't malloc marker\n");
ibv_error_abort (GEN_EXIT_ERR, "create_2level_com");
}
for (i=0; i < size ; ++i){
comm_ptr->ch.leader_rank[i] = -1;
}
int* group = comm_ptr->ch.leader_map;
grp_index = 0;
for (i=0; i < size ; ++i){
if (comm_ptr->ch.leader_rank[(group[i])] == -1){
comm_ptr->ch.leader_rank[(group[i])] = grp_index;
leader_group[grp_index++] = group[i];
}
}
leader_group_size = grp_index;
comm_ptr->ch.leader_group_size = leader_group_size;
mpi_errno = PMPI_Comm_group(comm, &comm_group);
if(mpi_errno) {
MPIU_ERR_POP(mpi_errno);
}
mpi_errno = PMPI_Group_incl(comm_group, leader_group_size, leader_group, &subgroup1);
if(mpi_errno) {
MPIU_ERR_POP(mpi_errno);
}
mpi_errno = PMPI_Comm_create(comm, subgroup1, &(comm_ptr->ch.leader_comm));
if(mpi_errno) {
MPIU_ERR_POP(mpi_errno);
}
MPID_Comm *leader_ptr;
MPID_Comm_get_ptr( comm_ptr->ch.leader_comm, leader_ptr );
MPIU_Free(leader_group);
MPID_Group_get_ptr( subgroup1, group_ptr );
if(group_ptr != NULL) {
mpi_errno = PMPI_Group_free(&subgroup1);
if(mpi_errno) {
MPIU_ERR_POP(mpi_errno);
}
}
mpi_errno = PMPI_Comm_split(comm, leader, local_rank, &(comm_ptr->ch.shmem_comm));
if(mpi_errno) {
MPIU_ERR_POP(mpi_errno);
}
MPID_Comm *shmem_ptr;
MPID_Comm_get_ptr(comm_ptr->ch.shmem_comm, shmem_ptr);
mpi_errno = PMPI_Comm_rank(comm_ptr->ch.shmem_comm, &my_local_id);
if(mpi_errno) {
MPIU_ERR_POP(mpi_errno);
}
mpi_errno = PMPI_Comm_size(comm_ptr->ch.shmem_comm, &my_local_size);
if(mpi_errno) {
MPIU_ERR_POP(mpi_errno);
}
if(my_local_id == 0) {
int array_index=0;
mpi_errno = PMPI_Comm_size(comm_ptr->ch.leader_comm, &leader_comm_size);
if(mpi_errno) {
MPIU_ERR_POP(mpi_errno);
}
comm_ptr->ch.node_sizes = MPIU_Malloc(sizeof(int)*leader_comm_size);
mpi_errno = PMPI_Allgather(&my_local_size, 1, MPI_INT,
comm_ptr->ch.node_sizes, 1, MPI_INT, comm_ptr->ch.leader_comm);
if(mpi_errno) {
MPIU_ERR_POP(mpi_errno);
}
comm_ptr->ch.is_uniform = 1;
for(array_index=0; array_index < leader_comm_size; array_index++) {
if(comm_ptr->ch.node_sizes[0] != comm_ptr->ch.node_sizes[array_index]) {
comm_ptr->ch.is_uniform = 0;
break;
}
}
}
comm_ptr->ch.is_global_block = 0;
/* We need to check to see if the ranks are block or not. Each node leader
* gets the global ranks of all of its children processes. It scans through
* this array to see if the ranks are in block order. The node-leaders then
* do an allreduce to see if all the other nodes are also in block order.
* This is followed by an intra-node bcast to let the children processes
* know of the result of this step */
if(my_local_id == 0) {
int is_local_block = 1;
int index = 1;
while( index < my_local_size) {
if( (shmem_group[index] - 1) !=
shmem_group[index - 1]) {
is_local_block = 0;
break;
}
index++;
}
comm_ptr->ch.shmem_coll_ok = 0;/* To prevent Allreduce taking shmem route*/
mpi_errno = MPIR_Allreduce_impl(&(is_local_block),
&(comm_ptr->ch.is_global_block), 1,
MPI_INT, MPI_LAND, leader_ptr, &errflag);
if(mpi_errno) {
MPIU_ERR_POP(mpi_errno);
}
mpi_errno = MPIR_Bcast_impl(&(comm_ptr->ch.is_global_block),1, MPI_INT, 0,
shmem_ptr, &errflag);
if(mpi_errno) {
MPIU_ERR_POP(mpi_errno);
}
} else {
mpi_errno = MPIR_Bcast_impl(&(comm_ptr->ch.is_global_block),1, MPI_INT, 0,
shmem_ptr, &errflag);
if(mpi_errno) {
MPIU_ERR_POP(mpi_errno);
}
}
if (my_local_id == 0){
lock_shmem_region();
increment_shmem_comm_count();
shmem_comm_count = get_shmem_comm_count();
unlock_shmem_region();
}
shmem_ptr->ch.shmem_coll_ok = 0;
/* To prevent Bcast taking the knomial_2level_bcast route */
mpi_errno = MPIR_Bcast_impl (&shmem_comm_count, 1, MPI_INT, 0, shmem_ptr, &errflag);
if(mpi_errno) {
MPIU_ERR_POP(mpi_errno);
}
if (shmem_comm_count <= g_shmem_coll_blocks){
shmem_ptr->ch.shmem_comm_rank = shmem_comm_count-1;
input_flag = 1;
} else{
input_flag = 0;
}
comm_ptr->ch.shmem_coll_ok = 0;/* To prevent Allreduce taking shmem route*/
mpi_errno = MPIR_Allreduce_impl(&input_flag, &output_flag, 1, MPI_INT, MPI_LAND, comm_ptr, &errflag);
if(mpi_errno) {
MPIU_ERR_POP(mpi_errno);
}
comm_ptr->ch.allgather_comm_ok = 0;
if (allgather_ranking){
int is_contig =1, check_leader =1, check_size=1, is_local_ok=0,is_block=0;
int PPN;
int shmem_grp_size = my_local_size;
int leader_rank;
MPI_Group allgather_group;
comm_ptr->ch.allgather_comm=MPI_COMM_NULL;
comm_ptr->ch.allgather_new_ranks=NULL;
if(comm_ptr->ch.leader_comm != MPI_COMM_NULL) {
PMPI_Comm_rank(comm_ptr->ch.leader_comm, &leader_rank);
}
mpi_errno=MPIR_Bcast_impl(&leader_rank, 1, MPI_INT, 0, shmem_ptr, &errflag);
if(mpi_errno) {
MPIU_ERR_POP(mpi_errno);
}
for (i=1; i < shmem_grp_size; i++ ){
if (shmem_group[i] != shmem_group[i-1]+1){
is_contig =0;
break;
}
}
if (leader != (shmem_grp_size*leader_rank)){
check_leader=0;
}
if (shmem_grp_size != (size/leader_group_size)){
check_size=0;
}
is_local_ok = is_contig && check_leader && check_size;
mpi_errno = MPIR_Allreduce_impl(&is_local_ok, &is_block, 1, MPI_INT, MPI_LAND, comm_ptr, &errflag);
if(mpi_errno) {
MPIU_ERR_POP(mpi_errno);
}
if (is_block){
int counter=0,j;
comm_ptr->ch.allgather_new_ranks = MPIU_Malloc(sizeof(int)*size);
if (NULL == comm_ptr->ch.allgather_new_ranks){
mpi_errno = MPIR_Err_create_code( MPI_SUCCESS, MPIR_ERR_RECOVERABLE,
FCNAME, __LINE__, MPI_ERR_OTHER,
"**nomem", 0 );
return mpi_errno;
}
PPN = shmem_grp_size;
for (j=0; j < PPN; j++){
for (i=0; i < leader_group_size; i++){
comm_ptr->ch.allgather_new_ranks[counter] = j + i*PPN;
counter++;
}
}
mpi_errno = PMPI_Group_incl(comm_group, size, comm_ptr->ch.allgather_new_ranks, &allgather_group);
if(mpi_errno) {
MPIU_ERR_POP(mpi_errno);
}
mpi_errno = PMPI_Comm_create(comm_ptr->handle, allgather_group, &(comm_ptr->ch.allgather_comm));
if(mpi_errno) {
MPIU_ERR_POP(mpi_errno);
}
comm_ptr->ch.allgather_comm_ok = 1;
mpi_errno=PMPI_Group_free(&allgather_group);
if(mpi_errno) {
MPIU_ERR_POP(mpi_errno);
}
}
}
mpi_errno=PMPI_Group_free(&comm_group);
if(mpi_errno) {
MPIU_ERR_POP(mpi_errno);
}
if (output_flag == 1){
comm_ptr->ch.shmem_coll_ok = 1;
comm_registry[comm_registered++] = comm_ptr->context_id;
} else{
comm_ptr->ch.shmem_coll_ok = 0;
MPID_Group_get_ptr( subgroup1, group_ptr );
if(group_ptr != NULL) {
mpi_errno = PMPI_Group_free(&subgroup1);
if(mpi_errno) {
MPIU_ERR_POP(mpi_errno);
}
}
MPID_Group_get_ptr( comm_group, group_ptr );
if(group_ptr != NULL) {
mpi_errno = PMPI_Group_free(&comm_group);
if(mpi_errno) {
MPIU_ERR_POP(mpi_errno);
}
}
free_2level_comm(comm_ptr);
comm_ptr->ch.shmem_comm = MPI_COMM_NULL;
comm_ptr->ch.leader_comm = MPI_COMM_NULL;
}
++comm_count;
MPIU_Free(shmem_group);
fn_fail:
MPIDU_ERR_CHECK_MULTIPLE_THREADS_EXIT( comm_ptr );
return (mpi_errno);
}
int MPIR_Alltoallw_intra(const void *sendbuf, const int sendcounts[], const int sdispls[],
const MPI_Datatype sendtypes[], void *recvbuf, const int recvcounts[],
const int rdispls[], const MPI_Datatype recvtypes[], MPID_Comm *comm_ptr,
int *errflag)
{
int comm_size, i, j;
int mpi_errno = MPI_SUCCESS;
int mpi_errno_ret = MPI_SUCCESS;
MPI_Status status;
MPI_Status *starray;
MPI_Request *reqarray;
int dst, rank;
MPI_Comm comm;
int outstanding_requests;
int ii, ss, bblock;
int type_size;
MPIU_CHKLMEM_DECL(2);
comm = comm_ptr->handle;
comm_size = comm_ptr->local_size;
rank = comm_ptr->rank;
/* check if multiple threads are calling this collective function */
MPIDU_ERR_CHECK_MULTIPLE_THREADS_ENTER( comm_ptr );
if (sendbuf == MPI_IN_PLACE) {
/* We use pair-wise sendrecv_replace in order to conserve memory usage,
* which is keeping with the spirit of the MPI-2.2 Standard. But
* because of this approach all processes must agree on the global
* schedule of sendrecv_replace operations to avoid deadlock.
*
* Note that this is not an especially efficient algorithm in terms of
* time and there will be multiple repeated malloc/free's rather than
* maintaining a single buffer across the whole loop. Something like
* MADRE is probably the best solution for the MPI_IN_PLACE scenario. */
for (i = 0; i < comm_size; ++i) {
/* start inner loop at i to avoid re-exchanging data */
for (j = i; j < comm_size; ++j) {
if (rank == i) {
/* also covers the (rank == i && rank == j) case */
mpi_errno = MPIC_Sendrecv_replace(((char *)recvbuf + rdispls[j]),
recvcounts[j], recvtypes[j],
j, MPIR_ALLTOALLW_TAG,
j, MPIR_ALLTOALLW_TAG,
comm, &status, errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag = TRUE;
MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
}
}
else if (rank == j) {
/* same as above with i/j args reversed */
mpi_errno = MPIC_Sendrecv_replace(((char *)recvbuf + rdispls[i]),
recvcounts[i], recvtypes[i],
i, MPIR_ALLTOALLW_TAG,
i, MPIR_ALLTOALLW_TAG,
comm, &status, errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag = TRUE;
MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
}
}
}
}
}
else {
bblock = MPIR_CVAR_ALLTOALL_THROTTLE;
if (bblock == 0) bblock = comm_size;
MPIU_CHKLMEM_MALLOC(starray, MPI_Status*, 2*bblock*sizeof(MPI_Status), mpi_errno, "starray");
MPIU_CHKLMEM_MALLOC(reqarray, MPI_Request*, 2*bblock*sizeof(MPI_Request), mpi_errno, "reqarray");
/* post only bblock isends/irecvs at a time as suggested by Tony Ladd */
for (ii=0; ii<comm_size; ii+=bblock) {
outstanding_requests = 0;
ss = comm_size-ii < bblock ? comm_size-ii : bblock;
/* do the communication -- post ss sends and receives: */
for ( i=0; i<ss; i++ ) {
dst = (rank+i+ii) % comm_size;
if (recvcounts[dst]) {
MPID_Datatype_get_size_macro(recvtypes[dst], type_size);
if (type_size) {
mpi_errno = MPIC_Irecv((char *)recvbuf+rdispls[dst],
recvcounts[dst], recvtypes[dst], dst,
MPIR_ALLTOALLW_TAG, comm,
&reqarray[outstanding_requests]);
if (mpi_errno) { MPIU_ERR_POP(mpi_errno); }
outstanding_requests++;
}
}
}
for ( i=0; i<ss; i++ ) {
dst = (rank-i-ii+comm_size) % comm_size;
if (sendcounts[dst]) {
MPID_Datatype_get_size_macro(sendtypes[dst], type_size);
if (type_size) {
mpi_errno = MPIC_Isend((char *)sendbuf+sdispls[dst],
sendcounts[dst], sendtypes[dst], dst,
MPIR_ALLTOALLW_TAG, comm,
&reqarray[outstanding_requests], errflag);
if (mpi_errno) { MPIU_ERR_POP(mpi_errno); }
outstanding_requests++;
}
}
}
mpi_errno = MPIC_Waitall(outstanding_requests, reqarray, starray, errflag);
if (mpi_errno && mpi_errno != MPI_ERR_IN_STATUS) MPIU_ERR_POP(mpi_errno);
/* --BEGIN ERROR HANDLING-- */
if (mpi_errno == MPI_ERR_IN_STATUS) {
for (i=0; i<outstanding_requests; i++) {
if (starray[i].MPI_ERROR != MPI_SUCCESS) {
mpi_errno = starray[i].MPI_ERROR;
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag = TRUE;
MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
}
}
}
}
/* --END ERROR HANDLING-- */
}
#ifdef FOO
/* Use pairwise exchange algorithm. */
/* Make local copy first */
mpi_errno = MPIR_Localcopy(((char *)sendbuf+sdispls[rank]),
sendcounts[rank], sendtypes[rank],
((char *)recvbuf+rdispls[rank]),
recvcounts[rank], recvtypes[rank]);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
/* Do the pairwise exchange. */
for (i=1; i<comm_size; i++) {
src = (rank - i + comm_size) % comm_size;
dst = (rank + i) % comm_size;
mpi_errno = MPIC_Sendrecv(((char *)sendbuf+sdispls[dst]),
sendcounts[dst], sendtypes[dst], dst,
MPIR_ALLTOALLW_TAG,
((char *)recvbuf+rdispls[src]),
recvcounts[src], recvtypes[dst], src,
MPIR_ALLTOALLW_TAG, comm, &status, errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag = TRUE;
MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
}
}
#endif
}
/* check if multiple threads are calling this collective function */
fn_exit:
MPIDU_ERR_CHECK_MULTIPLE_THREADS_EXIT( comm_ptr );
MPIU_CHKLMEM_FREEALL();
if (mpi_errno_ret)
mpi_errno = mpi_errno_ret;
else if (*errflag)
MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**coll_fail");
return mpi_errno;
fn_fail:
goto fn_exit;
}
int MPIR_Alltoallw_inter(const void *sendbuf, const int sendcounts[], const int sdispls[],
const MPI_Datatype sendtypes[], void *recvbuf, const int recvcounts[],
const int rdispls[], const MPI_Datatype recvtypes[], MPID_Comm *comm_ptr,
int *errflag)
{
/* Intercommunicator alltoallw. We use a pairwise exchange algorithm
similar to the one used in intracommunicator alltoallw. Since the
local and remote groups can be of different
sizes, we first compute the max of local_group_size,
remote_group_size. At step i, 0 <= i < max_size, each process
receives from src = (rank - i + max_size) % max_size if src <
remote_size, and sends to dst = (rank + i) % max_size if dst <
remote_size.
FIXME: change algorithm to match intracommunicator alltoallv
*/
int local_size, remote_size, max_size, i;
int mpi_errno = MPI_SUCCESS;
int mpi_errno_ret = MPI_SUCCESS;
MPI_Status status;
int src, dst, rank, sendcount, recvcount;
char *sendaddr, *recvaddr;
MPI_Datatype sendtype, recvtype;
MPI_Comm comm;
local_size = comm_ptr->local_size;
remote_size = comm_ptr->remote_size;
comm = comm_ptr->handle;
rank = comm_ptr->rank;
/* check if multiple threads are calling this collective function */
MPIDU_ERR_CHECK_MULTIPLE_THREADS_ENTER( comm_ptr );
/* Use pairwise exchange algorithm. */
max_size = MPIR_MAX(local_size, remote_size);
for (i=0; i<max_size; i++) {
src = (rank - i + max_size) % max_size;
dst = (rank + i) % max_size;
if (src >= remote_size) {
src = MPI_PROC_NULL;
recvaddr = NULL;
recvcount = 0;
recvtype = MPI_DATATYPE_NULL;
}
else {
recvaddr = (char *)recvbuf + rdispls[src];
recvcount = recvcounts[src];
recvtype = recvtypes[src];
}
if (dst >= remote_size) {
dst = MPI_PROC_NULL;
sendaddr = NULL;
sendcount = 0;
sendtype = MPI_DATATYPE_NULL;
}
else {
sendaddr = (char *)sendbuf+sdispls[dst];
sendcount = sendcounts[dst];
sendtype = sendtypes[dst];
}
mpi_errno = MPIC_Sendrecv(sendaddr, sendcount, sendtype,
dst, MPIR_ALLTOALLW_TAG, recvaddr,
recvcount, recvtype, src,
MPIR_ALLTOALLW_TAG, comm, &status, errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag = TRUE;
MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
}
}
fn_exit:
/* check if multiple threads are calling this collective function */
MPIDU_ERR_CHECK_MULTIPLE_THREADS_EXIT( comm_ptr );
if (mpi_errno_ret)
mpi_errno = mpi_errno_ret;
else if (*errflag)
MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**coll_fail");
return mpi_errno;
fn_fail:
goto fn_exit;
}
