int MPIR_Ineighbor_alltoallw_default(const void *sendbuf, const int sendcounts[], const MPI_Aint sdispls[], const MPI_Datatype sendtypes[], void *recvbuf, const int recvcounts[], const MPI_Aint rdispls[], const MPI_Datatype recvtypes[], MPID_Comm *comm_ptr, MPID_Sched_t s)
{
int mpi_errno = MPI_SUCCESS;
int indegree, outdegree, weighted;
int k,l;
int *srcs, *dsts;
MPIU_CHKLMEM_DECL(2);
mpi_errno = MPIR_Topo_canon_nhb_count(comm_ptr, &indegree, &outdegree, &weighted);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
MPIU_CHKLMEM_MALLOC(srcs, int *, indegree*sizeof(int), mpi_errno, "srcs");
MPIU_CHKLMEM_MALLOC(dsts, int *, outdegree*sizeof(int), mpi_errno, "dsts");
mpi_errno = MPIR_Topo_canon_nhb(comm_ptr,
indegree, srcs, MPI_UNWEIGHTED,
outdegree, dsts, MPI_UNWEIGHTED);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
for (k = 0; k < outdegree; ++k) {
char *sb;
MPID_Ensure_Aint_fits_in_pointer(MPI_VOID_PTR_CAST_TO_MPI_AINT sendbuf + sdispls[k]);
sb = ((char *)sendbuf) + sdispls[k];
mpi_errno = MPID_Sched_send(sb, sendcounts[k], sendtypes[k], dsts[k], comm_ptr, s);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
}
for (l = 0; l < indegree; ++l) {
char *rb;
MPID_Ensure_Aint_fits_in_pointer(MPI_VOID_PTR_CAST_TO_MPI_AINT recvbuf + rdispls[l]);
rb = ((char *)recvbuf) + rdispls[l];
mpi_errno = MPID_Sched_recv(rb, recvcounts[l], recvtypes[l], srcs[l], comm_ptr, s);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
}
MPID_SCHED_BARRIER(s);
fn_exit:
MPIU_CHKLMEM_FREEALL();
return mpi_errno;
fn_fail:
goto fn_exit;
}
int MPIR_Scan(
const void *sendbuf,
void *recvbuf,
int count,
MPI_Datatype datatype,
MPI_Op op,
MPID_Comm *comm_ptr,
int *errflag )
{
int mpi_errno = MPI_SUCCESS;
int mpi_errno_ret = MPI_SUCCESS;
MPIU_CHKLMEM_DECL(3);
MPIU_THREADPRIV_DECL;
int rank = comm_ptr->rank;
MPI_Status status;
void *tempbuf = NULL, *localfulldata = NULL, *prefulldata = NULL;
MPI_Aint true_lb, true_extent, extent;
int noneed = 1; /* noneed=1 means no need to bcast tempbuf and
reduce tempbuf & recvbuf */
/* 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 (!MPIR_Comm_is_node_consecutive(comm_ptr)) {
/* We can't use the SMP-aware algorithm, use the generic one */
return MPIR_Scan_generic(sendbuf, recvbuf, count, datatype, op, comm_ptr, errflag);
}
MPIU_THREADPRIV_GET;
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(tempbuf, void *, count*(MPIR_MAX(extent, true_extent)),
mpi_errno, "temporary buffer");
tempbuf = (void *)((char*)tempbuf - true_lb);
/* Create prefulldata and localfulldata on local roots of all nodes */
if (comm_ptr->node_roots_comm != NULL) {
MPIU_CHKLMEM_MALLOC(prefulldata, void *, count*(MPIR_MAX(extent, true_extent)),
mpi_errno, "prefulldata for scan");
prefulldata = (void *)((char*)prefulldata - true_lb);
if (comm_ptr->node_comm != NULL) {
MPIU_CHKLMEM_MALLOC(localfulldata, void *, count*(MPIR_MAX(extent, true_extent)),
mpi_errno, "localfulldata for scan");
localfulldata = (void *)((char*)localfulldata - true_lb);
}
static int MPIR_Reduce_binomial (
const void *sendbuf,
void *recvbuf,
int count,
MPI_Datatype datatype,
MPI_Op op,
int root,
MPID_Comm *comm_ptr,
int *errflag )
{
int mpi_errno = MPI_SUCCESS;
int mpi_errno_ret = MPI_SUCCESS;
MPI_Status status;
int comm_size, rank, is_commutative, type_size ATTRIBUTE((unused));
int mask, relrank, source, lroot;
MPI_Aint true_lb, true_extent, extent;
void *tmp_buf;
MPI_Comm comm;
MPIU_CHKLMEM_DECL(2);
if (count == 0) return MPI_SUCCESS;
comm = comm_ptr->handle;
comm_size = comm_ptr->local_size;
rank = comm_ptr->rank;
/* Create a temporary buffer */
MPIR_Type_get_true_extent_impl(datatype, &true_lb, &true_extent);
MPID_Datatype_get_extent_macro(datatype, extent);
is_commutative = MPIR_Op_is_commutative(op);
/* I think this is the worse case, so we can avoid an assert()
* inside the for loop */
/* should be buf+{this}? */
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);
/* If I'm not the root, then my recvbuf may not be valid, therefore
I have to allocate a temporary one */
if (rank != root) {
MPIU_CHKLMEM_MALLOC(recvbuf, void *,
count*(MPIR_MAX(extent,true_extent)),
mpi_errno, "receive buffer");
recvbuf = (void *)((char*)recvbuf - true_lb);
}
/* ****************************************************************** */
int MPIDO_Allgather_bcast(const void *sendbuf,
int sendcount,
MPI_Datatype sendtype,
void *recvbuf,
int recvcount,
MPI_Datatype recvtype,
MPI_Aint send_true_lb,
MPI_Aint recv_true_lb,
size_t send_size,
size_t recv_size,
MPID_Comm * comm_ptr,
int *mpierrno)
{
int i, np, rc = 0;
MPI_Aint extent;
np = comm_ptr ->local_size;
MPID_Datatype_get_extent_macro(recvtype, extent);
MPID_Ensure_Aint_fits_in_pointer ((MPI_VOID_PTR_CAST_TO_MPI_AINT recvbuf +
np * recvcount * extent));
if (sendbuf != MPI_IN_PLACE)
{
void *destbuf = recvbuf + comm_ptr->rank * recvcount * extent;
MPIR_Localcopy(sendbuf,
sendcount,
sendtype,
destbuf,
recvcount,
recvtype);
}
/* this code should either abort on first error or somehow aggregate
* error codes, esp since it calls internal routines */
for (i = 0; i < np; i++)
{
void *destbuf = recvbuf + i * recvcount * extent;
/* TODO: Change to PAMI */
rc = MPIDO_Bcast(destbuf,
recvcount,
recvtype,
i,
comm_ptr,
mpierrno);
}
return rc;
}
int MPIR_Ineighbor_allgatherv_default(const void *sendbuf, int sendcount, MPI_Datatype sendtype, void *recvbuf, const int recvcounts[], const int displs[], MPI_Datatype recvtype, MPID_Comm *comm_ptr, MPID_Sched_t s)
{
int mpi_errno = MPI_SUCCESS;
int indegree, outdegree, weighted;
int i, k,l;
int *srcs, *dsts;
int comm_size;
MPI_Aint recvtype_extent;
MPIU_CHKLMEM_DECL(2);
comm_size = comm_ptr->local_size;
MPID_Datatype_get_extent_macro(recvtype, recvtype_extent);
for (i = 0; i < comm_size; ++i) {
MPID_Ensure_Aint_fits_in_pointer(MPI_VOID_PTR_CAST_TO_MPI_AINT recvbuf +
(displs[i] * recvtype_extent));
}
mpi_errno = MPIR_Topo_canon_nhb_count(comm_ptr, &indegree, &outdegree, &weighted);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
MPIU_CHKLMEM_MALLOC(srcs, int *, indegree*sizeof(int), mpi_errno, "srcs");
MPIU_CHKLMEM_MALLOC(dsts, int *, outdegree*sizeof(int), mpi_errno, "dsts");
mpi_errno = MPIR_Topo_canon_nhb(comm_ptr,
indegree, srcs, MPI_UNWEIGHTED,
outdegree, dsts, MPI_UNWEIGHTED);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
for (k = 0; k < outdegree; ++k) {
mpi_errno = MPID_Sched_send(sendbuf, sendcount, sendtype, dsts[k], comm_ptr, s);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
}
for (l = 0; l < indegree; ++l) {
char *rb = ((char *)recvbuf) + displs[l] * recvtype_extent;
mpi_errno = MPID_Sched_recv(rb, recvcounts[l], recvtype, srcs[l], comm_ptr, s);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
}
MPID_SCHED_BARRIER(s);
fn_exit:
MPIU_CHKLMEM_FREEALL();
return mpi_errno;
fn_fail:
goto fn_exit;
}
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;
}
int MPIR_Pack_impl(const void *inbuf,
int incount,
MPI_Datatype datatype,
void *outbuf,
MPI_Aint outsize,
MPI_Aint *position)
{
int mpi_errno = MPI_SUCCESS;
MPI_Aint first, last;
MPID_Segment *segp;
int contig;
MPI_Aint dt_true_lb;
MPI_Aint data_sz;
if (incount == 0) {
goto fn_exit;
}
/* Handle contig case quickly */
if (HANDLE_GET_KIND(datatype) == HANDLE_KIND_BUILTIN) {
contig = TRUE;
dt_true_lb = 0;
data_sz = incount * MPID_Datatype_get_basic_size(datatype);
} else {
MPID_Datatype *dt_ptr;
MPID_Datatype_get_ptr(datatype, dt_ptr);
contig = dt_ptr->is_contig;
dt_true_lb = dt_ptr->true_lb;
data_sz = incount * dt_ptr->size;
}
if (contig) {
MPIU_Memcpy((char *) outbuf + *position, (char *)inbuf + dt_true_lb, data_sz);
*position = (int)((MPI_Aint)*position + data_sz);
goto fn_exit;
}
/* non-contig case */
/* TODO: CHECK RETURN VALUES?? */
/* TODO: SHOULD THIS ALL BE IN A MPID_PACK??? */
segp = MPID_Segment_alloc();
MPIU_ERR_CHKANDJUMP1(segp == NULL, mpi_errno, MPI_ERR_OTHER, "**nomem", "**nomem %s", "MPID_Segment");
mpi_errno = MPID_Segment_init(inbuf, incount, datatype, segp, 0);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
/* NOTE: the use of buffer values and positions in MPI_Pack and in
* MPID_Segment_pack are quite different. See code or docs or something.
*/
first = 0;
last = SEGMENT_IGNORE_LAST;
/* Ensure that pointer increment fits in a pointer */
MPID_Ensure_Aint_fits_in_pointer((MPI_VOID_PTR_CAST_TO_MPI_AINT outbuf) +
(MPI_Aint) *position);
MPID_Segment_pack(segp,
first,
&last,
(void *) ((char *) outbuf + *position));
/* Ensure that calculation fits into an int datatype. */
MPID_Ensure_Aint_fits_in_int((MPI_Aint)*position + last);
*position = (int)((MPI_Aint)*position + last);
MPID_Segment_free(segp);
fn_exit:
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_Ialltoall_inter(const void *sendbuf, int sendcount, MPI_Datatype sendtype, void *recvbuf, int recvcount, MPI_Datatype recvtype, MPID_Comm *comm_ptr, MPID_Sched_t s)
{
/* Intercommunicator alltoall. We use a pairwise exchange algorithm
similar to the one used in intracommunicator alltoall for long
messages. 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.
*/
int mpi_errno = MPI_SUCCESS;
int local_size, remote_size, max_size, i;
MPI_Aint sendtype_extent, recvtype_extent;
int src, dst, rank;
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, sendtype_extent);
MPID_Datatype_get_extent_macro(recvtype, recvtype_extent);
/* Do the pairwise exchanges */
max_size = MPIR_MAX(local_size, remote_size);
MPID_Ensure_Aint_fits_in_pointer(MPI_VOID_PTR_CAST_TO_MPI_AINT recvbuf +
max_size*recvcount*recvtype_extent);
MPID_Ensure_Aint_fits_in_pointer(MPI_VOID_PTR_CAST_TO_MPI_AINT sendbuf +
max_size*sendcount*sendtype_extent);
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;
}
else {
recvaddr = (char *)recvbuf + src*recvcount*recvtype_extent;
}
if (dst >= remote_size) {
dst = MPI_PROC_NULL;
sendaddr = NULL;
}
else {
sendaddr = (char *)sendbuf + dst*sendcount*sendtype_extent;
}
mpi_errno = MPID_Sched_send(sendaddr, sendcount, sendtype, dst, comm_ptr, s);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
mpi_errno = MPID_Sched_recv(recvaddr, recvcount, recvtype, src, comm_ptr, s);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
MPID_SCHED_BARRIER(s);
}
fn_exit:
return mpi_errno;
fn_fail:
goto fn_exit;
}
int MPIR_Scatter_intra(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
void *recvbuf, int recvcount, MPI_Datatype recvtype, int root,
MPID_Comm *comm_ptr, int *errflag)
{
MPI_Status status;
MPI_Aint extent=0;
int rank, comm_size, is_homogeneous, sendtype_size;
int curr_cnt, relative_rank, nbytes, send_subtree_cnt;
int mask, recvtype_size=0, src, dst;
int tmp_buf_size = 0;
void *tmp_buf=NULL;
int mpi_errno = MPI_SUCCESS;
int mpi_errno_ret = MPI_SUCCESS;
MPI_Comm comm;
MPIU_CHKLMEM_DECL(4);
comm = comm_ptr->handle;
comm_size = comm_ptr->local_size;
rank = comm_ptr->rank;
if ( ((rank == root) && (sendcount == 0)) ||
((rank != root) && (recvcount == 0)) )
return MPI_SUCCESS;
/* check if multiple threads are calling this collective function */
MPIDU_ERR_CHECK_MULTIPLE_THREADS_ENTER( comm_ptr );
is_homogeneous = 1;
#ifdef MPID_HAS_HETERO
if (comm_ptr->is_hetero)
is_homogeneous = 0;
#endif
/* Use binomial tree algorithm */
if (rank == root)
MPID_Datatype_get_extent_macro(sendtype, extent);
relative_rank = (rank >= root) ? rank - root : rank - root + comm_size;
if (is_homogeneous) {
/* communicator is homogeneous */
if (rank == root) {
/* We separate the two cases (root and non-root) because
in the event of recvbuf=MPI_IN_PLACE on the root,
recvcount and recvtype are not valid */
MPID_Datatype_get_size_macro(sendtype, sendtype_size);
MPID_Ensure_Aint_fits_in_pointer(MPI_VOID_PTR_CAST_TO_MPI_AINT sendbuf +
extent*sendcount*comm_size);
nbytes = sendtype_size * sendcount;
}
else {
MPID_Datatype_get_size_macro(recvtype, recvtype_size);
MPID_Ensure_Aint_fits_in_pointer(extent*recvcount*comm_size);
nbytes = recvtype_size * recvcount;
}
curr_cnt = 0;
/* all even nodes other than root need a temporary buffer to
receive data of max size (nbytes*comm_size)/2 */
if (relative_rank && !(relative_rank % 2)) {
tmp_buf_size = (nbytes*comm_size)/2;
MPIU_CHKLMEM_MALLOC(tmp_buf, void *, tmp_buf_size, mpi_errno, "tmp_buf");
}
/* if the root is not rank 0, we reorder the sendbuf in order of
relative ranks and copy it into a temporary buffer, so that
all the sends from the root are contiguous and in the right
order. */
if (rank == root) {
if (root != 0) {
tmp_buf_size = nbytes*comm_size;
MPIU_CHKLMEM_MALLOC(tmp_buf, void *, tmp_buf_size, mpi_errno, "tmp_buf");
if (recvbuf != MPI_IN_PLACE)
mpi_errno = MPIR_Localcopy(((char *) sendbuf + extent*sendcount*rank),
sendcount*(comm_size-rank), sendtype, tmp_buf,
nbytes*(comm_size-rank), MPI_BYTE);
else
mpi_errno = MPIR_Localcopy(((char *) sendbuf + extent*sendcount*(rank+1)),
sendcount*(comm_size-rank-1),
sendtype, (char *)tmp_buf + nbytes,
nbytes*(comm_size-rank-1), MPI_BYTE);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
mpi_errno = MPIR_Localcopy(sendbuf, sendcount*rank, sendtype,
((char *) tmp_buf + nbytes*(comm_size-rank)),
nbytes*rank, MPI_BYTE);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
curr_cnt = nbytes*comm_size;
}
else
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,
int *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;
MPI_Request *reqarray;
int dst, rank, req_cnt;
MPI_Comm comm;
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;
/* 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_ft(((char *)recvbuf + rdispls[j]*recv_extent),
recvcounts[j], recvtype,
j, MPIR_ALLTOALLV_TAG,
j, MPIR_ALLTOALLV_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_ft(((char *)recvbuf + rdispls[i]*recv_extent),
recvcounts[i], recvtype,
i, MPIR_ALLTOALLV_TAG,
i, MPIR_ALLTOALLV_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_PARAM_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, 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) {
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) {
MPID_Ensure_Aint_fits_in_pointer(MPI_VOID_PTR_CAST_TO_MPI_AINT recvbuf +
rdispls[dst]*recv_extent);
mpi_errno = MPIC_Irecv_ft((char *)recvbuf+rdispls[dst]*recv_extent,
recvcounts[dst], recvtype, dst,
MPIR_ALLTOALLV_TAG, comm,
&reqarray[req_cnt]);
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);
}
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) {
MPID_Ensure_Aint_fits_in_pointer(MPI_VOID_PTR_CAST_TO_MPI_AINT sendbuf +
sdispls[dst]*send_extent);
mpi_errno = MPIC_Isend_ft((char *)sendbuf+sdispls[dst]*send_extent,
sendcounts[dst], sendtype, dst,
MPIR_ALLTOALLV_TAG, comm,
&reqarray[req_cnt], 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);
}
req_cnt++;
}
}
}
mpi_errno = MPIC_Waitall_ft(req_cnt, 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<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 = TRUE;
MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
MPIU_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)
MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**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,
int *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;
MPI_Comm comm;
local_size = comm_ptr->local_size;
remote_size = comm_ptr->remote_size;
comm = comm_ptr->handle;
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 {
MPID_Ensure_Aint_fits_in_pointer(MPI_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 {
MPID_Ensure_Aint_fits_in_pointer(MPI_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_ft(sendaddr, sendcount, sendtype, dst,
MPIR_ALLTOALLV_TAG, recvaddr, recvcount,
recvtype, src, MPIR_ALLTOALLV_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;
}
int MPIR_Allgatherv_intra (
const void *sendbuf,
int sendcount,
MPI_Datatype sendtype,
void *recvbuf,
const int *recvcounts,
const int *displs,
MPI_Datatype recvtype,
MPID_Comm *comm_ptr,
mpir_errflag_t *errflag )
{
int comm_size, rank, j, i, left, right;
int mpi_errno = MPI_SUCCESS;
int mpi_errno_ret = MPI_SUCCESS;
MPI_Status status;
MPI_Aint recvbuf_extent, recvtype_extent, recvtype_true_extent,
recvtype_true_lb;
int curr_cnt, send_cnt, dst, total_count, recvtype_size, pof2, src, rem;
int recv_cnt;
void *tmp_buf;
int mask, dst_tree_root, my_tree_root, is_homogeneous, position,
send_offset, recv_offset, last_recv_cnt, nprocs_completed, k,
offset, tmp_mask, tree_root;
#ifdef MPID_HAS_HETERO
int tmp_buf_size, nbytes;
#endif
MPIU_CHKLMEM_DECL(1);
/* check if multiple threads are calling this collective function */
MPIDU_ERR_CHECK_MULTIPLE_THREADS_ENTER( comm_ptr );
comm_size = comm_ptr->local_size;
rank = comm_ptr->rank;
total_count = 0;
for (i=0; i<comm_size; i++)
total_count += recvcounts[i];
if (total_count == 0) goto fn_exit;
MPID_Datatype_get_extent_macro( recvtype, recvtype_extent );
MPID_Datatype_get_size_macro(recvtype, recvtype_size);
if ((total_count*recvtype_size < MPIR_CVAR_ALLGATHER_LONG_MSG_SIZE) &&
!(comm_size & (comm_size - 1))) {
/* Short or medium size message and power-of-two no. of processes. Use
* recursive doubling algorithm */
is_homogeneous = 1;
#ifdef MPID_HAS_HETERO
if (comm_ptr->is_hetero)
is_homogeneous = 0;
#endif
if (is_homogeneous) {
/* need to receive contiguously into tmp_buf because
displs could make the recvbuf noncontiguous */
MPIR_Type_get_true_extent_impl(recvtype, &recvtype_true_lb, &recvtype_true_extent);
MPID_Ensure_Aint_fits_in_pointer(total_count *
(MPIR_MAX(recvtype_true_extent, recvtype_extent)));
MPIU_CHKLMEM_MALLOC(tmp_buf, void *, total_count*(MPIR_MAX(recvtype_true_extent,recvtype_extent)), mpi_errno, "tmp_buf");
/* adjust for potential negative lower bound in datatype */
tmp_buf = (void *)((char*)tmp_buf - recvtype_true_lb);
/* copy local data into right location in tmp_buf */
position = 0;
for (i=0; i<rank; i++) position += recvcounts[i];
if (sendbuf != MPI_IN_PLACE)
{
mpi_errno = MPIR_Localcopy(sendbuf, sendcount, sendtype,
((char *)tmp_buf + position*
recvtype_extent),
recvcounts[rank], recvtype);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
}
else
{
/* if in_place specified, local data is found in recvbuf */
mpi_errno = MPIR_Localcopy(((char *)recvbuf +
displs[rank]*recvtype_extent),
recvcounts[rank], recvtype,
((char *)tmp_buf + position*
recvtype_extent),
recvcounts[rank], recvtype);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
}
curr_cnt = recvcounts[rank];
mask = 0x1;
i = 0;
while (mask < comm_size) {
dst = rank ^ mask;
/* find offset into send and recv buffers. zero out
the least significant "i" bits of rank and dst to
find root of src and dst subtrees. Use ranks of
roots as index to send from and recv into buffer */
dst_tree_root = dst >> i;
dst_tree_root <<= i;
my_tree_root = rank >> i;
my_tree_root <<= i;
if (dst < comm_size) {
send_offset = 0;
for (j=0; j<my_tree_root; j++)
send_offset += recvcounts[j];
recv_offset = 0;
for (j=0; j<dst_tree_root; j++)
recv_offset += recvcounts[j];
mpi_errno = MPIC_Sendrecv(((char *)tmp_buf + send_offset * recvtype_extent),
curr_cnt, recvtype, dst,
MPIR_ALLGATHERV_TAG,
((char *)tmp_buf + recv_offset * recvtype_extent),
total_count - recv_offset, recvtype, dst,
MPIR_ALLGATHERV_TAG,
comm_ptr, &status, errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag = MPIR_ERR_GET_CLASS(mpi_errno);
MPIU_ERR_SET(mpi_errno, *errflag, "**fail");
MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
last_recv_cnt = 0;
} else
/* for convenience, recv is posted for a bigger amount
than will be sent */
MPIR_Get_count_impl(&status, recvtype, &last_recv_cnt);
curr_cnt += last_recv_cnt;
}
/* if some processes in this process's subtree in this step
did not have any destination process to communicate with
because of non-power-of-two, we need to send them the
data that they would normally have received from those
processes. That is, the haves in this subtree must send to
the havenots. We use a logarithmic
recursive-halfing algorithm for this. */
/* This part of the code will not currently be
executed because we are not using recursive
doubling for non power of two. Mark it as experimental
so that it doesn't show up as red in the coverage
tests. */
/* --BEGIN EXPERIMENTAL-- */
if (dst_tree_root + mask > comm_size) {
nprocs_completed = comm_size - my_tree_root - mask;
/* nprocs_completed is the number of processes in this
subtree that have all the data. Send data to others
in a tree fashion. First find root of current tree
that is being divided into two. k is the number of
least-significant bits in this process's rank that
must be zeroed out to find the rank of the root */
j = mask;
k = 0;
while (j) {
j >>= 1;
k++;
}
k--;
tmp_mask = mask >> 1;
while (tmp_mask) {
dst = rank ^ tmp_mask;
tree_root = rank >> k;
tree_root <<= k;
/* send only if this proc has data and destination
doesn't have data. at any step, multiple processes
can send if they have the data */
if ((dst > rank) &&
(rank < tree_root + nprocs_completed)
&& (dst >= tree_root + nprocs_completed)) {
offset = 0;
for (j=0; j<(my_tree_root+mask); j++)
offset += recvcounts[j];
offset *= recvtype_extent;
mpi_errno = MPIC_Send(((char *)tmp_buf + offset),
last_recv_cnt,
recvtype, dst,
MPIR_ALLGATHERV_TAG, comm_ptr, errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag = MPIR_ERR_GET_CLASS(mpi_errno);
MPIU_ERR_SET(mpi_errno, *errflag, "**fail");
MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
}
/* last_recv_cnt was set in the previous
receive. that's the amount of data to be
sent now. */
}
/* recv only if this proc. doesn't have data and sender
has data */
else if ((dst < rank) &&
(dst < tree_root + nprocs_completed) &&
(rank >= tree_root + nprocs_completed)) {
offset = 0;
for (j=0; j<(my_tree_root+mask); j++)
offset += recvcounts[j];
mpi_errno = MPIC_Recv(((char *)tmp_buf + offset * recvtype_extent),
total_count - offset, recvtype,
dst, MPIR_ALLGATHERV_TAG,
comm_ptr, &status, errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag = MPIR_ERR_GET_CLASS(mpi_errno);
MPIU_ERR_SET(mpi_errno, *errflag, "**fail");
MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
last_recv_cnt = 0;
} else
/* for convenience, recv is posted for a
bigger amount than will be sent */
MPIR_Get_count_impl(&status, recvtype, &last_recv_cnt);
curr_cnt += last_recv_cnt;
}
tmp_mask >>= 1;
k--;
}
}
/* --END EXPERIMENTAL-- */
mask <<= 1;
i++;
}
/* copy data from tmp_buf to recvbuf */
position = 0;
for (j=0; j<comm_size; j++) {
if ((sendbuf != MPI_IN_PLACE) || (j != rank)) {
/* not necessary to copy if in_place and
j==rank. otherwise copy. */
mpi_errno = MPIR_Localcopy(((char *)tmp_buf + position*recvtype_extent),
recvcounts[j], recvtype,
((char *)recvbuf + displs[j]*recvtype_extent),
recvcounts[j], recvtype);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
}
position += recvcounts[j];
}
}
int MPIR_Gatherv (
void *sendbuf,
int sendcnt,
MPI_Datatype sendtype,
void *recvbuf,
int *recvcnts,
int *displs,
MPI_Datatype recvtype,
int root,
MPID_Comm *comm_ptr,
int *errflag )
{
int comm_size, rank;
int mpi_errno = MPI_SUCCESS;
int mpi_errno_ret = MPI_SUCCESS;
MPI_Comm comm;
MPI_Aint extent;
int i, reqs;
int min_procs;
MPI_Request *reqarray;
MPI_Status *starray;
MPIU_CHKLMEM_DECL(2);
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 );
/* If rank == root, then I recv lots, otherwise I send */
if (((comm_ptr->comm_kind == MPID_INTRACOMM) && (root == rank)) ||
((comm_ptr->comm_kind == MPID_INTERCOMM) && (root == MPI_ROOT))) {
if (comm_ptr->comm_kind == MPID_INTRACOMM)
comm_size = comm_ptr->local_size;
else
comm_size = comm_ptr->remote_size;
MPID_Datatype_get_extent_macro(recvtype, extent);
/* each node can make sure it is not going to overflow aint */
MPID_Ensure_Aint_fits_in_pointer(MPI_VOID_PTR_CAST_TO_MPI_AINT recvbuf +
displs[rank] * extent);
MPIU_CHKLMEM_MALLOC(reqarray, MPI_Request *, comm_size * sizeof(MPI_Request), mpi_errno, "reqarray");
MPIU_CHKLMEM_MALLOC(starray, MPI_Status *, comm_size * sizeof(MPI_Status), mpi_errno, "starray");
reqs = 0;
for (i = 0; i < comm_size; i++) {
if (recvcnts[i]) {
if ((comm_ptr->comm_kind == MPID_INTRACOMM) && (i == rank)) {
if (sendbuf != MPI_IN_PLACE) {
mpi_errno = MPIR_Localcopy(sendbuf, sendcnt, sendtype,
((char *)recvbuf+displs[rank]*extent),
recvcnts[rank], recvtype);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
}
}
else {
mpi_errno = MPIC_Irecv_ft(((char *)recvbuf+displs[i]*extent),
recvcnts[i], recvtype, i,
MPIR_GATHERV_TAG, comm,
&reqarray[reqs++]);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
}
}
}
/* ... then wait for *all* of them to finish: */
mpi_errno = MPIC_Waitall_ft(reqs, 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 < reqs; 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-- */
}
else if (root != MPI_PROC_NULL) { /* non-root nodes, and in the intercomm. case, non-root nodes on remote side */
int MPIR_Iallreduce_redscat_allgather(const void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype, MPI_Op op, MPID_Comm *comm_ptr, MPID_Sched_t s)
{
int mpi_errno = MPI_SUCCESS;
int comm_size, rank, newrank, pof2, rem;
int i, send_idx, recv_idx, last_idx, mask, newdst, dst, send_cnt, recv_cnt;
MPI_Aint true_lb, true_extent, extent;
void *tmp_buf = NULL;
int *cnts, *disps;
MPIR_SCHED_CHKPMEM_DECL(1);
MPIU_CHKLMEM_DECL(2);
/* we only support builtin datatypes for now, breaking up user types to do
* the reduce-scatter is tricky */
MPIU_Assert(HANDLE_GET_KIND(op) == HANDLE_KIND_BUILTIN);
comm_size = comm_ptr->local_size;
rank = comm_ptr->rank;
/* 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));
MPIR_SCHED_CHKPMEM_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 = MPID_Sched_copy(sendbuf, count, datatype,
recvbuf, count, datatype, s);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
MPID_SCHED_BARRIER(s);
}
/* 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 = MPID_Sched_send(recvbuf, count, datatype, rank+1, comm_ptr, s);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
MPID_SCHED_BARRIER(s);
/* temporarily set the rank to -1 so that this
process does not pariticipate in recursive
doubling */
newrank = -1;
}
else { /* odd */
mpi_errno = MPID_Sched_recv(tmp_buf, count, datatype, rank-1, comm_ptr, s);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
MPID_SCHED_BARRIER(s);
/* do the reduction on received data. since the
ordering is right, it doesn't matter whether
the operation is commutative or not. */
mpi_errno = MPID_Sched_reduce(tmp_buf, recvbuf, count, datatype, op, s);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
MPID_SCHED_BARRIER(s);
/* change the rank */
newrank = rank / 2;
}
}
else /* rank >= 2*rem */
newrank = rank - rem;
if (newrank != -1) {
/* for the reduce-scatter, calculate the count that
each process receives and the displacement within
the buffer */
/* TODO I (goodell@) believe that these counts and displacements could be
* calculated directly during the loop, rather than requiring a less-scalable
* "2*pof2"-sized memory allocation */
MPIU_CHKLMEM_MALLOC(cnts, int *, pof2*sizeof(int), mpi_errno, "counts");
MPIU_CHKLMEM_MALLOC(disps, int *, pof2*sizeof(int), mpi_errno, "displacements");
MPIU_Assert(count >= pof2); /* the cnts calculations assume this */
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];
}
/* Send data from recvbuf. Recv into tmp_buf */
mpi_errno = MPID_Sched_recv(((char *)tmp_buf + disps[recv_idx]*extent),
recv_cnt, datatype, dst, comm_ptr, s);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
/* sendrecv, no barrier here */
mpi_errno = MPID_Sched_send(((char *)recvbuf + disps[send_idx]*extent),
send_cnt, datatype, dst, comm_ptr, s);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
MPID_SCHED_BARRIER(s);
/* 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 = MPID_Sched_reduce(((char *)tmp_buf + disps[recv_idx]*extent),
((char *)recvbuf + disps[recv_idx]*extent),
recv_cnt, datatype, op, s);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
MPID_SCHED_BARRIER(s);
/* 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 = MPID_Sched_recv(((char *)recvbuf + disps[recv_idx]*extent),
recv_cnt, datatype, dst, comm_ptr, s);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
/* sendrecv, no barrier here */
mpi_errno = MPID_Sched_send(((char *)recvbuf + disps[send_idx]*extent),
send_cnt, datatype, dst, comm_ptr, s);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
MPID_SCHED_BARRIER(s);
if (newrank > newdst) send_idx = recv_idx;
mask >>= 1;
}
}
int MPIR_Igatherv(const void *sendbuf, int sendcount, MPI_Datatype sendtype, void *recvbuf,
const int recvcounts[], const int displs[], MPI_Datatype recvtype, int root,
MPID_Comm *comm_ptr, MPID_Sched_t s)
{
int mpi_errno = MPI_SUCCESS;
int i;
int comm_size, rank;
MPI_Aint extent;
int min_procs;
rank = comm_ptr->rank;
/* If rank == root, then I recv lots, otherwise I send */
if (((comm_ptr->comm_kind == MPID_INTRACOMM) && (root == rank)) ||
((comm_ptr->comm_kind == MPID_INTERCOMM) && (root == MPI_ROOT)))
{
if (comm_ptr->comm_kind == MPID_INTRACOMM)
comm_size = comm_ptr->local_size;
else
comm_size = comm_ptr->remote_size;
MPID_Datatype_get_extent_macro(recvtype, extent);
/* each node can make sure it is not going to overflow aint */
MPID_Ensure_Aint_fits_in_pointer(MPI_VOID_PTR_CAST_TO_MPI_AINT recvbuf + displs[rank] * extent);
for (i = 0; i < comm_size; i++) {
if (recvcounts[i]) {
if ((comm_ptr->comm_kind == MPID_INTRACOMM) && (i == rank)) {
if (sendbuf != MPI_IN_PLACE) {
mpi_errno = MPID_Sched_copy(sendbuf, sendcount, sendtype,
((char *)recvbuf+displs[rank]*extent),
recvcounts[rank], recvtype, s);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
}
}
else {
mpi_errno = MPID_Sched_recv(((char *)recvbuf+displs[i]*extent),
recvcounts[i], recvtype, i, comm_ptr, s);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
}
}
}
}
else if (root != MPI_PROC_NULL)
{
/* non-root nodes, and in the intercomm. case, non-root nodes on remote side */
if (sendcount) {
/* we want local size in both the intracomm and intercomm cases
because the size of the root's group (group A in the standard) is
irrelevant here. */
comm_size = comm_ptr->local_size;
min_procs = MPIR_PARAM_GATHERV_INTER_SSEND_MIN_PROCS;
if (min_procs == -1)
min_procs = comm_size + 1; /* Disable ssend */
else if (min_procs == 0) /* backwards compatibility, use default value */
MPIR_PARAM_GET_DEFAULT_INT(GATHERV_INTER_SSEND_MIN_PROCS,&min_procs);
if (comm_size >= min_procs)
mpi_errno = MPID_Sched_ssend(sendbuf, sendcount, sendtype, root, comm_ptr, s);
else
mpi_errno = MPID_Sched_send(sendbuf, sendcount, sendtype, root, comm_ptr, s);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
}
}
fn_exit:
return mpi_errno;
fn_fail:
goto fn_exit;
}
int MPIR_Iallgatherv_rec_dbl(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
void *recvbuf, const int recvcounts[], const int displs[],
MPI_Datatype recvtype, MPID_Comm *comm_ptr, MPID_Sched_t s)
{
int mpi_errno = MPI_SUCCESS;
int comm_size, rank, i, j, k;
int curr_count, send_offset, incoming_count, recv_offset;
int mask, dst, total_count, position, offset, my_tree_root, dst_tree_root;
MPI_Aint recvtype_extent, recvtype_true_extent, recvtype_true_lb;
void *tmp_buf = NULL;
int is_homogeneous ATTRIBUTE((unused));
MPIR_SCHED_CHKPMEM_DECL(1);
comm_size = comm_ptr->local_size;
rank = comm_ptr->rank;
is_homogeneous = 1;
#ifdef MPID_HAS_HETERO
if (comm_ptr->is_hetero)
is_homogeneous = 0;
#endif
MPIU_Assert(is_homogeneous); /* we only handle the homogeneous for now */
/* need to receive contiguously into tmp_buf because
displs could make the recvbuf noncontiguous */
MPID_Datatype_get_extent_macro(recvtype, recvtype_extent);
MPIR_Type_get_true_extent_impl(recvtype, &recvtype_true_lb, &recvtype_true_extent);
total_count = 0;
for (i=0; i<comm_size; i++)
total_count += recvcounts[i];
if (total_count == 0)
goto fn_exit;
MPID_Ensure_Aint_fits_in_pointer(total_count*(MPIR_MAX(recvtype_true_extent, recvtype_extent)));
MPIR_SCHED_CHKPMEM_MALLOC(tmp_buf, void *, total_count*(MPIR_MAX(recvtype_true_extent,recvtype_extent)), mpi_errno, "tmp_buf");
/* adjust for potential negative lower bound in datatype */
tmp_buf = (void *)((char*)tmp_buf - recvtype_true_lb);
/* copy local data into right location in tmp_buf */
position = 0;
for (i=0; i<rank; i++) position += recvcounts[i];
if (sendbuf != MPI_IN_PLACE)
{
mpi_errno = MPID_Sched_copy(sendbuf, sendcount, sendtype,
((char *)tmp_buf + position*recvtype_extent),
recvcounts[rank], recvtype, s);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
}
else
{
/* if in_place specified, local data is found in recvbuf */
mpi_errno = MPID_Sched_copy(((char *)recvbuf + displs[rank]*recvtype_extent),
recvcounts[rank], recvtype,
((char *)tmp_buf + position*recvtype_extent),
recvcounts[rank], recvtype, s);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
}
curr_count = recvcounts[rank];
/* never used uninitialized w/o this, but compiler can't tell that */
incoming_count = -1;
/* [goodell@] random notes that help slightly when deciphering this code:
* - mask is also equal to the number of blocks that we are going to recv
* (less if comm_size is non-pof2)
* - FOO_tree_root is the leftmost (lowest ranked) process with whom FOO has
* communicated, directly or indirectly, at the beginning of round the
* round. FOO is either "dst" or "my", where "my" means use my rank.
* - in each round we are going to recv the blocks
* B[dst_tree_root],B[dst_tree_root+1],...,B[min(dst_tree_root+mask,comm_size)]
*/
mask = 0x1;
i = 0;
while (mask < comm_size) {
dst = rank ^ mask;
/* find offset into send and recv buffers. zero out
the least significant "i" bits of rank and dst to
find root of src and dst subtrees. Use ranks of
roots as index to send from and recv into buffer */
dst_tree_root = dst >> i;
dst_tree_root <<= i;
my_tree_root = rank >> i;
my_tree_root <<= i;
if (dst < comm_size) {
send_offset = 0;
for (j = 0; j < my_tree_root; j++)
send_offset += recvcounts[j];
recv_offset = 0;
for (j = 0; j < dst_tree_root; j++)
recv_offset += recvcounts[j];
incoming_count = 0;
for (j = dst_tree_root; j < (dst_tree_root + mask) && j < comm_size; ++j)
incoming_count += recvcounts[j];
mpi_errno = MPID_Sched_send(((char *)tmp_buf + send_offset * recvtype_extent),
curr_count, recvtype, dst, comm_ptr, s);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
/* sendrecv, no barrier here */
mpi_errno = MPID_Sched_recv(((char *)tmp_buf + recv_offset * recvtype_extent),
incoming_count, recvtype, dst, comm_ptr, s);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
MPID_SCHED_BARRIER(s);
curr_count += incoming_count;
}
/* if some processes in this process's subtree in this step
did not have any destination process to communicate with
because of non-power-of-two, we need to send them the
data that they would normally have received from those
processes. That is, the haves in this subtree must send to
the havenots. We use a logarithmic
recursive-halfing algorithm for this. */
/* This part of the code will not currently be
executed because we are not using recursive
doubling for non power of two. Mark it as experimental
so that it doesn't show up as red in the coverage
tests. */
/* --BEGIN EXPERIMENTAL-- */
if (dst_tree_root + mask > comm_size) {
int tmp_mask, tree_root;
int nprocs_completed = comm_size - my_tree_root - mask;
/* nprocs_completed is the number of processes in this
subtree that have all the data. Send data to others
in a tree fashion. First find root of current tree
that is being divided into two. k is the number of
least-significant bits in this process's rank that
must be zeroed out to find the rank of the root */
/* [goodell@] it looks like (k==i) is always true, could possibly
* skip the loop below */
j = mask;
k = 0;
while (j) {
j >>= 1;
k++;
}
k--;
tmp_mask = mask >> 1;
while (tmp_mask) {
dst = rank ^ tmp_mask;
tree_root = rank >> k;
tree_root <<= k;
/* send only if this proc has data and destination
doesn't have data. at any step, multiple processes
can send if they have the data */
if ((dst > rank) &&
(rank < tree_root + nprocs_completed) &&
(dst >= tree_root + nprocs_completed))
{
offset = 0;
for (j = 0; j < (my_tree_root+mask); j++)
offset += recvcounts[j];
offset *= recvtype_extent;
/* incoming_count was set in the previous
receive. that's the amount of data to be
sent now. */
mpi_errno = MPID_Sched_send(((char *)tmp_buf + offset),
incoming_count, recvtype, dst, comm_ptr, s);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
MPID_SCHED_BARRIER(s);
}
/* recv only if this proc. doesn't have data and sender
has data */
else if ((dst < rank) &&
(dst < tree_root + nprocs_completed) &&
(rank >= tree_root + nprocs_completed))
{
offset = 0;
for (j = 0; j < (my_tree_root+mask); j++)
offset += recvcounts[j];
/* recalculate incoming_count, since not all processes will have
* this value */
incoming_count = 0;
for (j = dst_tree_root; j < (dst_tree_root + mask) && j < comm_size; ++j)
incoming_count += recvcounts[j];
mpi_errno = MPID_Sched_recv(((char *)tmp_buf + offset * recvtype_extent),
incoming_count, recvtype,
dst, comm_ptr, s);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
MPID_SCHED_BARRIER(s);
curr_count += incoming_count;
}
tmp_mask >>= 1;
k--;
}
}
/* --END EXPERIMENTAL-- */
mask <<= 1;
i++;
}
/*@
MPI_Pack_external - Packs a datatype into contiguous memory, using the
external32 format
Input Parameters:
+ datarep - data representation (string)
. inbuf - input buffer start (choice)
. incount - number of input data items (integer)
. datatype - datatype of each input data item (handle)
- outsize - output buffer size, in bytes (address integer)
Output Parameters:
. outbuf - output buffer start (choice)
Input/Output Parameters:
. position - current position in buffer, in bytes (address integer)
.N ThreadSafe
.N Fortran
.N Errors
.N MPI_SUCCESS
.N MPI_ERR_TYPE
.N MPI_ERR_ARG
.N MPI_ERR_COUNT
@*/
int MPI_Pack_external(const char datarep[],
const void *inbuf,
int incount,
MPI_Datatype datatype,
void *outbuf,
MPI_Aint outsize,
MPI_Aint *position)
{
static const char FCNAME[] = "MPI_Pack_external";
int mpi_errno = MPI_SUCCESS;
MPI_Aint first, last;
MPID_Segment *segp;
MPID_MPI_STATE_DECL(MPID_STATE_MPI_PACK_EXTERNAL);
MPIR_ERRTEST_INITIALIZED_ORDIE();
MPID_MPI_FUNC_ENTER(MPID_STATE_MPI_PACK_EXTERNAL);
/* Validate parameters and objects (post conversion) */
# ifdef HAVE_ERROR_CHECKING
{
MPID_BEGIN_ERROR_CHECKS;
{
MPIR_ERRTEST_COUNT(incount, mpi_errno);
MPIR_ERRTEST_COUNT(outsize, mpi_errno);
/* NOTE: inbuf could be null (MPI_BOTTOM) */
if (incount > 0) {
MPIR_ERRTEST_ARGNULL(outbuf, "output buffer", mpi_errno);
}
MPIR_ERRTEST_ARGNULL(position, "position", mpi_errno);
MPIR_ERRTEST_DATATYPE(datatype, "datatype", mpi_errno);
if (HANDLE_GET_KIND(datatype) != HANDLE_KIND_BUILTIN) {
MPID_Datatype *datatype_ptr = NULL;
MPID_Datatype_get_ptr(datatype, datatype_ptr);
MPID_Datatype_valid_ptr(datatype_ptr, mpi_errno);
MPID_Datatype_committed_ptr(datatype_ptr, mpi_errno);
if (mpi_errno != MPI_SUCCESS) goto fn_fail;
}
}
MPID_END_ERROR_CHECKS;
}
# endif /* HAVE_ERROR_CHECKING */
/* ... body of routine ... */
if (incount == 0) {
goto fn_exit;
}
segp = MPID_Segment_alloc();
/* --BEGIN ERROR HANDLING-- */
if (segp == NULL)
{
mpi_errno = MPIR_Err_create_code(MPI_SUCCESS,
MPIR_ERR_RECOVERABLE,
FCNAME,
__LINE__,
MPI_ERR_OTHER,
"**nomem",
"**nomem %s",
"MPID_Segment");
goto fn_fail;
}
/* --END ERROR HANDLING-- */
mpi_errno = MPID_Segment_init(inbuf, incount, datatype, segp, 1);
if (mpi_errno != MPI_SUCCESS) goto fn_fail;
/* NOTE: the use of buffer values and positions in MPI_Pack_external and
* in MPID_Segment_pack_external are quite different. See code or docs
* or something.
*/
first = 0;
last = SEGMENT_IGNORE_LAST;
/* Ensure that pointer increment fits in a pointer */
MPID_Ensure_Aint_fits_in_pointer((MPI_VOID_PTR_CAST_TO_MPI_AINT outbuf) + *position);
MPID_Segment_pack_external32(segp,
first,
&last,
(void *)((char *) outbuf + *position));
*position += last;
MPID_Segment_free(segp);
/* ... end of body of routine ... */
fn_exit:
MPID_MPI_FUNC_EXIT(MPID_STATE_MPI_PACK_EXTERNAL);
return mpi_errno;
fn_fail:
/* --BEGIN ERROR HANDLING-- */
# ifdef HAVE_ERROR_CHECKING
{
mpi_errno = MPIR_Err_create_code(
mpi_errno, MPIR_ERR_RECOVERABLE, FCNAME, __LINE__, MPI_ERR_OTHER, "**mpi_pack_external",
"**mpi_pack_external %s %p %d %D %p %d %p", datarep, inbuf, incount, datatype, outbuf, outsize, position);
}
# endif
mpi_errno = MPIR_Err_return_comm(0, FCNAME, mpi_errno);
goto fn_exit;
/* --END ERROR HANDLING-- */
}
/* not declared static because a machine-specific function may call this one in some cases */
int MPIR_Reduce_scatter_block_intra (
const void *sendbuf,
void *recvbuf,
int recvcount,
MPI_Datatype datatype,
MPI_Op op,
MPID_Comm *comm_ptr,
int *errflag )
{
int rank, comm_size, i;
MPI_Aint extent, true_extent, true_lb;
int *disps;
void *tmp_recvbuf, *tmp_results;
int mpi_errno = MPI_SUCCESS;
int mpi_errno_ret = MPI_SUCCESS;
int type_size, dis[2], blklens[2], total_count, nbytes, src, dst;
int mask, dst_tree_root, my_tree_root, j, k;
int *newcnts, *newdisps, rem, newdst, send_idx, recv_idx,
last_idx, send_cnt, recv_cnt;
int pof2, old_i, newrank, received;
MPI_Datatype sendtype, recvtype;
int nprocs_completed, tmp_mask, tree_root, is_commutative;
MPID_Op *op_ptr;
MPI_Comm comm;
MPIU_THREADPRIV_DECL;
MPIU_CHKLMEM_DECL(5);
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_GET;
MPIU_THREADPRIV_FIELD(op_errno) = 0;
if (recvcount == 0) {
goto fn_exit;
}
MPID_Datatype_get_extent_macro(datatype, extent);
MPIR_Type_get_true_extent_impl(datatype, &true_lb, &true_extent);
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;
}
MPIU_CHKLMEM_MALLOC(disps, int *, comm_size * sizeof(int), mpi_errno, "disps");
total_count = comm_size*recvcount;
for (i=0; i<comm_size; i++) {
disps[i] = i*recvcount;
}
MPID_Datatype_get_size_macro(datatype, type_size);
nbytes = total_count * type_size;
/* check if multiple threads are calling this collective function */
MPIDU_ERR_CHECK_MULTIPLE_THREADS_ENTER( comm_ptr );
/* total_count*extent eventually gets malloced. it isn't added to
* a user-passed in buffer */
MPID_Ensure_Aint_fits_in_pointer(total_count * MPIR_MAX(true_extent, extent));
if ((is_commutative) && (nbytes < MPIR_PARAM_REDSCAT_COMMUTATIVE_LONG_MSG_SIZE)) {
/* commutative and short. use recursive halving algorithm */
/* allocate temp. buffer to receive incoming data */
MPIU_CHKLMEM_MALLOC(tmp_recvbuf, void *, total_count*(MPIR_MAX(true_extent,extent)), mpi_errno, "tmp_recvbuf");
/* adjust for potential negative lower bound in datatype */
tmp_recvbuf = (void *)((char*)tmp_recvbuf - true_lb);
/* need to allocate another temporary buffer to accumulate
results because recvbuf may not be big enough */
MPIU_CHKLMEM_MALLOC(tmp_results, void *, total_count*(MPIR_MAX(true_extent,extent)), mpi_errno, "tmp_results");
/* adjust for potential negative lower bound in datatype */
tmp_results = (void *)((char*)tmp_results - true_lb);
/* copy sendbuf into tmp_results */
if (sendbuf != MPI_IN_PLACE)
mpi_errno = MPIR_Localcopy(sendbuf, total_count, datatype,
tmp_results, total_count, datatype);
else
mpi_errno = MPIR_Localcopy(recvbuf, total_count, datatype,
tmp_results, total_count, datatype);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
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(tmp_results, total_count,
datatype, rank+1,
MPIR_REDUCE_SCATTER_BLOCK_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_recvbuf, total_count,
datatype, rank-1,
MPIR_REDUCE_SCATTER_BLOCK_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_recvbuf, tmp_results,
total_count, datatype, op);
/* change the rank */
newrank = rank / 2;
}
}
else /* rank >= 2*rem */
newrank = rank - rem;
if (newrank != -1) {
/* recalculate the recvcnts and disps arrays because the
even-numbered processes who no longer participate will
have their result calculated by the process to their
right (rank+1). */
MPIU_CHKLMEM_MALLOC(newcnts, int *, pof2*sizeof(int), mpi_errno, "newcnts");
MPIU_CHKLMEM_MALLOC(newdisps, int *, pof2*sizeof(int), mpi_errno, "newdisps");
for (i=0; i<pof2; i++) {
/* what does i map to in the old ranking? */
old_i = (i < rem) ? i*2 + 1 : i + rem;
if (old_i < 2*rem) {
/* This process has to also do its left neighbor's
work */
newcnts[i] = 2 * recvcount;
}
else
newcnts[i] = recvcount;
}
newdisps[0] = 0;
for (i=1; i<pof2; i++)
newdisps[i] = newdisps[i-1] + newcnts[i-1];
mask = pof2 >> 1;
send_idx = recv_idx = 0;
last_idx = pof2;
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) {
send_idx = recv_idx + mask;
for (i=send_idx; i<last_idx; i++)
send_cnt += newcnts[i];
for (i=recv_idx; i<send_idx; i++)
recv_cnt += newcnts[i];
}
else {
recv_idx = send_idx + mask;
for (i=send_idx; i<recv_idx; i++)
send_cnt += newcnts[i];
for (i=recv_idx; i<last_idx; i++)
recv_cnt += newcnts[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 tmp_results. Recv into tmp_recvbuf */
if ((send_cnt != 0) && (recv_cnt != 0))
mpi_errno = MPIC_Sendrecv_ft((char *) tmp_results +
newdisps[send_idx]*extent,
send_cnt, datatype,
dst, MPIR_REDUCE_SCATTER_BLOCK_TAG,
(char *) tmp_recvbuf +
newdisps[recv_idx]*extent,
recv_cnt, datatype, dst,
MPIR_REDUCE_SCATTER_BLOCK_TAG, comm,
MPI_STATUS_IGNORE, errflag);
else if ((send_cnt == 0) && (recv_cnt != 0))
mpi_errno = MPIC_Recv_ft((char *) tmp_recvbuf +
newdisps[recv_idx]*extent,
recv_cnt, datatype, dst,
MPIR_REDUCE_SCATTER_BLOCK_TAG, comm,
MPI_STATUS_IGNORE, errflag);
else if ((recv_cnt == 0) && (send_cnt != 0))
mpi_errno = MPIC_Send_ft((char *) tmp_results +
newdisps[send_idx]*extent,
send_cnt, datatype,
dst, MPIR_REDUCE_SCATTER_BLOCK_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);
}
/* tmp_recvbuf contains data received in this step.
tmp_results contains data accumulated so far */
if (recv_cnt) {
mpi_errno = MPIR_Reduce_local_impl(
(char *) tmp_recvbuf + newdisps[recv_idx]*extent,
(char *) tmp_results + newdisps[recv_idx]*extent,
recv_cnt, datatype, op);
}
/* update send_idx for next iteration */
send_idx = recv_idx;
last_idx = recv_idx + mask;
mask >>= 1;
}
/* copy this process's result from tmp_results to recvbuf */
mpi_errno = MPIR_Localcopy((char *)tmp_results +
disps[rank]*extent,
recvcount, datatype, recvbuf,
recvcount, datatype);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
}
