static int barrier_smp_intra(MPID_Comm *comm_ptr, mpir_errflag_t *errflag)
{
int mpi_errno=MPI_SUCCESS;
int mpi_errno_ret = MPI_SUCCESS;
MPIU_Assert(MPIR_CVAR_ENABLE_SMP_COLLECTIVES && MPIR_CVAR_ENABLE_SMP_BARRIER &&
MPIR_Comm_is_node_aware(comm_ptr));
/* do the intranode barrier on all nodes */
if (comm_ptr->node_comm != NULL)
{
mpi_errno = MPIR_Barrier_impl(comm_ptr->node_comm, 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);
}
}
/* do the barrier across roots of all nodes */
if (comm_ptr->node_roots_comm != NULL) {
mpi_errno = MPIR_Barrier_impl(comm_ptr->node_roots_comm, 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);
}
}
/* release the local processes on each node with a 1-byte
broadcast (0-byte broadcast just returns without doing
anything) */
if (comm_ptr->node_comm != NULL)
{
int i=0;
mpi_errno = MPIR_Bcast_impl(&i, 1, MPI_BYTE, 0, comm_ptr->node_comm, 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);
}
}
fn_exit:
if (mpi_errno_ret)
mpi_errno = mpi_errno_ret;
else if (*errflag != MPIR_ERR_NONE)
MPIU_ERR_SET(mpi_errno, *errflag, "**coll_fail");
return mpi_errno;
fn_fail:
goto fn_exit;
}
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);
MPIU_ERR_SET(mpi_errno, *errflag, "**fail");
MPIU_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);
MPIU_ERR_SET(mpi_errno, *errflag, "**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 != MPIR_ERR_NONE)
MPIU_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;
}
static int barrier_smp_intra(MPID_Comm *comm_ptr, MPIR_Errflag_t *errflag)
{
int mpi_errno=MPI_SUCCESS;
int mpi_errno_ret = MPI_SUCCESS;
MPIU_Assert(MPIR_CVAR_ENABLE_SMP_COLLECTIVES && MPIR_CVAR_ENABLE_SMP_BARRIER &&
MPIR_Comm_is_node_aware(comm_ptr));
#if defined(FINEGRAIN_MPI)
int colocated_size = -1;
int colocated_sense = -1;
/* do barrier on osproc_colocated_comm */
if (comm_ptr->osproc_colocated_comm != NULL)
{
colocated_size = comm_ptr->osproc_colocated_comm->local_size;
MPIU_Assert( (comm_ptr->osproc_colocated_comm->co_shared_vars != NULL) && (comm_ptr->osproc_colocated_comm->co_shared_vars->co_barrier_vars != NULL) );
MPIU_Assert(colocated_size > 1 );
colocated_sense = comm_ptr->osproc_colocated_comm->co_shared_vars->co_barrier_vars->coproclet_signal;
if( comm_ptr->osproc_colocated_comm->rank != 0 ) { /* non-leader */
(comm_ptr->osproc_colocated_comm->co_shared_vars->co_barrier_vars->coproclet_counter)++;
if (comm_ptr->osproc_colocated_comm->co_shared_vars->co_barrier_vars->coproclet_counter == (colocated_size-1)){ /* excluding the leader */
comm_ptr->osproc_colocated_comm->co_shared_vars->co_barrier_vars->leader_signal = 1;
}
while(comm_ptr->osproc_colocated_comm->co_shared_vars->co_barrier_vars->coproclet_signal == colocated_sense) {
FG_Yield();
}
}
else { /* leader */
while(comm_ptr->osproc_colocated_comm->co_shared_vars->co_barrier_vars->leader_signal == 0) {
FG_Yield();
}
}
#if 0 /* Non-optimized version */
mpi_errno = MPIR_Barrier_impl(comm_ptr->osproc_colocated_comm, 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);
}
#endif
}
#endif
/* do the intranode barrier on all nodes */
if (comm_ptr->node_comm != NULL)
{
mpi_errno = MPIR_Barrier_impl(comm_ptr->node_comm, 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);
}
}
/* do the barrier across roots of all nodes */
if (comm_ptr->node_roots_comm != NULL) {
mpi_errno = MPIR_Barrier_impl(comm_ptr->node_roots_comm, 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);
}
}
/* release the local processes on each node with a 1-byte
broadcast (0-byte broadcast just returns without doing
anything) */
if (comm_ptr->node_comm != NULL)
{
int i=0;
mpi_errno = MPIR_Bcast_impl(&i, 1, MPI_BYTE, 0, comm_ptr->node_comm, 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);
}
}
#if defined(FINEGRAIN_MPI)
if (comm_ptr->osproc_colocated_comm != NULL)
{
if (comm_ptr->osproc_colocated_comm->rank == 0) { /* leader */
comm_ptr->osproc_colocated_comm->co_shared_vars->co_barrier_vars->leader_signal = 0;
comm_ptr->osproc_colocated_comm->co_shared_vars->co_barrier_vars->coproclet_counter = 0;
comm_ptr->osproc_colocated_comm->co_shared_vars->co_barrier_vars->coproclet_signal = 1 - comm_ptr->osproc_colocated_comm->co_shared_vars->co_barrier_vars->coproclet_signal;
}
#if 0 /* Non-optimized version */
/* release the colocated processes in each OS-process with a 1-byte
broadcast (0-byte broadcast just returns without doing
anything) */
int i=0;
mpi_errno = MPIR_Bcast_impl(&i, 1, MPI_BYTE, 0, comm_ptr->osproc_colocated_comm, 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);
}
#endif
}
#endif
fn_exit:
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;
}
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_Barrier_inter( MPID_Comm *comm_ptr, mpir_errflag_t *errflag )
{
int rank, mpi_errno = MPI_SUCCESS, root;
int mpi_errno_ret = MPI_SUCCESS;
int i = 0;
MPID_Comm *newcomm_ptr = NULL;
rank = comm_ptr->rank;
/* Get the local intracommunicator */
if (!comm_ptr->local_comm) {
mpi_errno = MPIR_Setup_intercomm_localcomm( comm_ptr );
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
}
newcomm_ptr = comm_ptr->local_comm;
/* do a barrier on the local intracommunicator */
mpi_errno = MPIR_Barrier_intra(newcomm_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);
}
/* rank 0 on each group does an intercommunicator broadcast to the
remote group to indicate that all processes in the local group
have reached the barrier. We do a 1-byte bcast because a 0-byte
bcast will just return without doing anything. */
/* first broadcast from left to right group, then from right to
left group */
if (comm_ptr->is_low_group) {
/* bcast to right*/
root = (rank == 0) ? MPI_ROOT : MPI_PROC_NULL;
mpi_errno = MPIR_Bcast_inter(&i, 1, MPI_BYTE, root, 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);
}
/* receive bcast from right */
root = 0;
mpi_errno = MPIR_Bcast_inter(&i, 1, MPI_BYTE, root, 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);
}
}
else {
/* receive bcast from left */
root = 0;
mpi_errno = MPIR_Bcast_inter(&i, 1, MPI_BYTE, root, 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);
}
/* bcast to left */
root = (rank == 0) ? MPI_ROOT : MPI_PROC_NULL;
mpi_errno = MPIR_Bcast_inter(&i, 1, MPI_BYTE, root, 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);
}
}
fn_exit:
if (mpi_errno_ret)
mpi_errno = mpi_errno_ret;
else if (*errflag != MPIR_ERR_NONE)
MPIU_ERR_SET(mpi_errno, *errflag, "**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_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;
}
/* 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);
}
int MPIR_Barrier_impl(MPID_Comm *comm_ptr, int *errflag)
{
int mpi_errno = MPI_SUCCESS;
int mpi_errno_ret = MPI_SUCCESS;
if (comm_ptr->coll_fns != NULL && comm_ptr->coll_fns->Barrier != NULL)
{
mpi_errno = comm_ptr->coll_fns->Barrier(comm_ptr, errflag);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
}
else
{
if (comm_ptr->comm_kind == MPID_INTRACOMM) {
#if defined(USE_SMP_COLLECTIVES)
if (MPIR_Comm_is_node_aware(comm_ptr)) {
/* do the intranode barrier on all nodes */
if (comm_ptr->node_comm != NULL)
{
mpi_errno = MPIR_Barrier_or_coll_fn(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);
}
}
/* do the barrier across roots of all nodes */
if (comm_ptr->node_roots_comm != NULL) {
mpi_errno = MPIR_Barrier_or_coll_fn(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);
}
}
/* release the local processes on each node with a 1-byte broadcast
(0-byte broadcast just returns without doing anything) */
if (comm_ptr->node_comm != NULL)
{
int i=0;
mpi_errno = MPIR_Bcast_impl(&i, 1, MPI_BYTE, 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_Barrier_intra( comm_ptr, errflag );
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
}
#else
mpi_errno = MPIR_Barrier_intra( comm_ptr, errflag );
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
#endif
}
else {
/* intercommunicator */
mpi_errno = MPIR_Barrier_inter( comm_ptr, errflag );
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
}
}
fn_exit:
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_Comm_agree(MPID_Comm *comm_ptr, int *flag)
{
int mpi_errno = MPI_SUCCESS, mpi_errno_tmp = MPI_SUCCESS;
MPID_Group *comm_grp, *failed_grp, *new_group_ptr, *global_failed;
int result, success = 1;
int errflag = 0;
int values[2];
MPID_MPI_STATE_DECL(MPID_STATE_MPIR_COMM_AGREE);
MPID_MPI_FUNC_ENTER(MPID_STATE_MPIR_COMM_AGREE);
MPIR_Comm_group_impl(comm_ptr, &comm_grp);
/* Get the locally known (not acknowledged) group of failed procs */
mpi_errno = MPID_Comm_failure_get_acked(comm_ptr, &failed_grp);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
/* First decide on the group of failed procs. */
mpi_errno = MPID_Comm_get_all_failed_procs(comm_ptr, &global_failed, MPIR_AGREE_TAG);
if (mpi_errno) errflag = 1;
mpi_errno = MPIR_Group_compare_impl(failed_grp, global_failed, &result);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
/* Create a subgroup without the failed procs */
mpi_errno = MPIR_Group_difference_impl(comm_grp, global_failed, &new_group_ptr);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
/* If that group isn't the same as what we think is failed locally, then
* mark it as such. */
if (result == MPI_UNEQUAL || errflag)
success = 0;
/* Do an allreduce to decide whether or not anyone thinks the group
* has changed */
mpi_errno_tmp = MPIR_Allreduce_group(MPI_IN_PLACE, &success, 1, MPI_INT, MPI_MIN, comm_ptr,
new_group_ptr, MPIR_AGREE_TAG, &errflag);
if (!success || errflag || mpi_errno_tmp)
success = 0;
values[0] = success;
values[1] = *flag;
/* Determine both the result of this function (mpi_errno) and the result
* of flag that will be returned to the user. */
MPIR_Allreduce_group(MPI_IN_PLACE, values, 2, MPI_INT, MPI_BAND, comm_ptr,
new_group_ptr, MPIR_AGREE_TAG, &errflag);
/* Ignore the result of the operation this time. Everyone will either
* return a failure because of !success earlier or they will return
* something useful for flag because of this operation. If there was a new
* failure in between the first allreduce and the second one, it's ignored
* here. */
if (failed_grp != MPID_Group_empty)
MPIR_Group_release(failed_grp);
MPIR_Group_release(new_group_ptr);
MPIR_Group_release(comm_grp);
if (global_failed != MPID_Group_empty)
MPIR_Group_release(global_failed);
success = values[0];
*flag = values[1];
if (!success) {
MPIU_ERR_SET(mpi_errno_tmp, MPIX_ERR_PROC_FAILED, "**mpix_comm_agree");
MPIU_ERR_ADD(mpi_errno, mpi_errno_tmp);
}
fn_exit:
MPID_MPI_FUNC_EXIT(MPID_STATE_MPIR_COMM_AGREE);
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,
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_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_Reduce_scatter_block_noncomm (
const void *sendbuf,
void *recvbuf,
int recvcount,
MPI_Datatype datatype,
MPI_Op op,
MPID_Comm *comm_ptr,
int *errflag )
{
int mpi_errno = MPI_SUCCESS;
int mpi_errno_ret = MPI_SUCCESS;
int comm_size = comm_ptr->local_size;
int rank = comm_ptr->rank;
int pof2;
int log2_comm_size;
int i, k;
int recv_offset, send_offset;
int block_size, total_count, size;
MPI_Aint true_extent, true_lb;
int buf0_was_inout;
void *tmp_buf0;
void *tmp_buf1;
void *result_ptr;
MPI_Comm comm = comm_ptr->handle;
MPIU_CHKLMEM_DECL(3);
MPIR_Type_get_true_extent_impl(datatype, &true_lb, &true_extent);
pof2 = 1;
log2_comm_size = 0;
while (pof2 < comm_size) {
pof2 <<= 1;
++log2_comm_size;
}
/* begin error checking */
MPIU_Assert(pof2 == comm_size); /* FIXME this version only works for power of 2 procs */
/* end error checking */
/* size of a block (count of datatype per block, NOT bytes per block) */
block_size = recvcount;
total_count = block_size * comm_size;
MPIU_CHKLMEM_MALLOC(tmp_buf0, void *, true_extent * total_count, mpi_errno, "tmp_buf0");
MPIU_CHKLMEM_MALLOC(tmp_buf1, void *, true_extent * total_count, mpi_errno, "tmp_buf1");
/* adjust for potential negative lower bound in datatype */
tmp_buf0 = (void *)((char*)tmp_buf0 - true_lb);
tmp_buf1 = (void *)((char*)tmp_buf1 - true_lb);
/* Copy our send data to tmp_buf0. We do this one block at a time and
permute the blocks as we go according to the mirror permutation. */
for (i = 0; i < comm_size; ++i) {
mpi_errno = MPIR_Localcopy((char *)(sendbuf == MPI_IN_PLACE ? recvbuf : sendbuf) + (i * true_extent * block_size), block_size, datatype,
(char *)tmp_buf0 + (MPIU_Mirror_permutation(i, log2_comm_size) * true_extent * block_size), block_size, datatype);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
}
buf0_was_inout = 1;
send_offset = 0;
recv_offset = 0;
size = total_count;
for (k = 0; k < log2_comm_size; ++k) {
/* use a double-buffering scheme to avoid local copies */
char *incoming_data = (buf0_was_inout ? tmp_buf1 : tmp_buf0);
char *outgoing_data = (buf0_was_inout ? tmp_buf0 : tmp_buf1);
int peer = rank ^ (0x1 << k);
size /= 2;
if (rank > peer) {
/* we have the higher rank: send top half, recv bottom half */
recv_offset += size;
}
else {
/* we have the lower rank: recv top half, send bottom half */
send_offset += size;
}
mpi_errno = MPIC_Sendrecv_ft(outgoing_data + send_offset*true_extent,
size, datatype, peer, MPIR_REDUCE_SCATTER_BLOCK_TAG,
incoming_data + recv_offset*true_extent,
size, datatype, peer, 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);
}
/* always perform the reduction at recv_offset, the data at send_offset
is now our peer's responsibility */
if (rank > peer) {
/* higher ranked value so need to call op(received_data, my_data) */
mpi_errno = MPIR_Reduce_local_impl(
incoming_data + recv_offset*true_extent,
outgoing_data + recv_offset*true_extent,
size, datatype, op);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
buf0_was_inout = buf0_was_inout;
}
else {
/* lower ranked value so need to call op(my_data, received_data) */
mpi_errno = MPIR_Reduce_local_impl(
outgoing_data + recv_offset*true_extent,
incoming_data + recv_offset*true_extent,
size, datatype, op);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
buf0_was_inout = !buf0_was_inout;
}
/* the next round of send/recv needs to happen within the block (of size
"size") that we just received and reduced */
send_offset = recv_offset;
}
MPIU_Assert(size == recvcount);
/* copy the reduced data to the recvbuf */
result_ptr = (char *)(buf0_was_inout ? tmp_buf0 : tmp_buf1) + recv_offset * true_extent;
mpi_errno = MPIR_Localcopy(result_ptr, size, datatype,
recvbuf, size, datatype);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
fn_exit:
MPIU_CHKLMEM_FREEALL();
/* --BEGIN ERROR HANDLING-- */
if (mpi_errno_ret)
mpi_errno = mpi_errno_ret;
else if (*errflag)
MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**coll_fail");
/* --END ERROR HANDLING-- */
return mpi_errno;
fn_fail:
goto fn_exit;
}
int MPIR_Allreduce_inter (
void *sendbuf,
void *recvbuf,
int count,
MPI_Datatype datatype,
MPI_Op op,
MPID_Comm *comm_ptr,
int *errflag )
{
/* Intercommunicator Allreduce.
We first do an intercommunicator reduce to rank 0 on left group,
then an intercommunicator reduce to rank 0 on right group, followed
by local intracommunicator broadcasts in each group.
We don't do local reduces first and then intercommunicator
broadcasts because it would require allocation of a temporary buffer.
*/
int rank, mpi_errno, root;
int mpi_errno_ret = MPI_SUCCESS;
MPID_Comm *newcomm_ptr = NULL;
rank = comm_ptr->rank;
/* first do a reduce from right group to rank 0 in left group,
then from left group to rank 0 in right group*/
if (comm_ptr->is_low_group) {
/* reduce from right group to rank 0*/
root = (rank == 0) ? MPI_ROOT : MPI_PROC_NULL;
mpi_errno = MPIR_Reduce_inter(sendbuf, recvbuf, count, datatype, op,
root, 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);
}
/* reduce to rank 0 of right group */
root = 0;
mpi_errno = MPIR_Reduce_inter(sendbuf, recvbuf, count, datatype, op,
root, 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 {
/* reduce to rank 0 of left group */
root = 0;
mpi_errno = MPIR_Reduce_inter(sendbuf, recvbuf, count, datatype, op,
root, 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);
}
/* reduce from right group to rank 0 */
root = (rank == 0) ? MPI_ROOT : MPI_PROC_NULL;
mpi_errno = MPIR_Reduce_inter(sendbuf, recvbuf, count, datatype, op,
root, 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);
}
}
/* Get the local intracommunicator */
if (!comm_ptr->local_comm)
MPIR_Setup_intercomm_localcomm( comm_ptr );
newcomm_ptr = comm_ptr->local_comm;
mpi_errno = MPIR_Bcast_impl(recvbuf, 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);
}
fn_exit:
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 MPIR_Alltoall_intra(
const void *sendbuf,
int sendcount,
MPI_Datatype sendtype,
void *recvbuf,
int recvcount,
MPI_Datatype recvtype,
MPID_Comm *comm_ptr,
int *errflag )
{
int comm_size, i, j, pof2;
MPI_Aint sendtype_extent, recvtype_extent;
MPI_Aint recvtype_true_extent, recvbuf_extent, recvtype_true_lb;
int mpi_errno=MPI_SUCCESS, src, dst, rank, nbytes;
int mpi_errno_ret = MPI_SUCCESS;
MPI_Status status;
int sendtype_size, pack_size, block, position, *displs, count;
MPI_Datatype newtype = MPI_DATATYPE_NULL;
void *tmp_buf;
MPI_Comm comm;
MPI_Request *reqarray;
MPI_Status *starray;
MPIU_CHKLMEM_DECL(6);
#ifdef MPIR_OLD_SHORT_ALLTOALL_ALG
MPI_Aint sendtype_true_extent, sendbuf_extent, sendtype_true_lb;
int k, p, curr_cnt, dst_tree_root, my_tree_root;
int last_recv_cnt, mask, tmp_mask, tree_root, nprocs_completed;
#endif
if (recvcount == 0) return MPI_SUCCESS;
comm = comm_ptr->handle;
comm_size = comm_ptr->local_size;
rank = comm_ptr->rank;
/* Get extent of send and recv types */
MPID_Datatype_get_extent_macro(recvtype, recvtype_extent);
MPID_Datatype_get_extent_macro(sendtype, sendtype_extent);
MPID_Datatype_get_size_macro(sendtype, sendtype_size);
nbytes = sendtype_size * sendcount;
/* 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 + j*recvcount*recvtype_extent),
recvcount, recvtype,
j, MPIR_ALLTOALL_TAG,
j, MPIR_ALLTOALL_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 + i*recvcount*recvtype_extent),
recvcount, recvtype,
i, MPIR_ALLTOALL_TAG,
i, MPIR_ALLTOALL_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 ((nbytes <= MPIR_PARAM_ALLTOALL_SHORT_MSG_SIZE) && (comm_size >= 8)) {
/* use the indexing algorithm by Jehoshua Bruck et al,
* IEEE TPDS, Nov. 97 */
/* allocate temporary buffer */
MPIR_Pack_size_impl(recvcount*comm_size, recvtype, &pack_size);
MPIU_CHKLMEM_MALLOC(tmp_buf, void *, pack_size, mpi_errno, "tmp_buf");
/* Do Phase 1 of the algorithim. Shift the data blocks on process i
* upwards by a distance of i blocks. Store the result in recvbuf. */
mpi_errno = MPIR_Localcopy((char *) sendbuf +
rank*sendcount*sendtype_extent,
(comm_size - rank)*sendcount, sendtype, recvbuf,
(comm_size - rank)*recvcount, recvtype);
if (mpi_errno) { MPIU_ERR_POP(mpi_errno); }
mpi_errno = MPIR_Localcopy(sendbuf, rank*sendcount, sendtype,
(char *) recvbuf +
(comm_size-rank)*recvcount*recvtype_extent,
rank*recvcount, recvtype);
if (mpi_errno) { MPIU_ERR_POP(mpi_errno); }
/* Input data is now stored in recvbuf with datatype recvtype */
/* Now do Phase 2, the communication phase. It takes
ceiling(lg p) steps. In each step i, each process sends to rank+2^i
and receives from rank-2^i, and exchanges all data blocks
whose ith bit is 1. */
/* allocate displacements array for indexed datatype used in
communication */
MPIU_CHKLMEM_MALLOC(displs, int *, comm_size * sizeof(int), mpi_errno, "displs");
pof2 = 1;
while (pof2 < comm_size) {
dst = (rank + pof2) % comm_size;
src = (rank - pof2 + comm_size) % comm_size;
/* Exchange all data blocks whose ith bit is 1 */
/* Create an indexed datatype for the purpose */
count = 0;
for (block=1; block<comm_size; block++) {
if (block & pof2) {
displs[count] = block * recvcount;
count++;
}
}
mpi_errno = MPIR_Type_create_indexed_block_impl(count, recvcount,
displs, recvtype, &newtype);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
mpi_errno = MPIR_Type_commit_impl(&newtype);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
position = 0;
mpi_errno = MPIR_Pack_impl(recvbuf, 1, newtype, tmp_buf, pack_size, &position);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
mpi_errno = MPIC_Sendrecv_ft(tmp_buf, position, MPI_PACKED, dst,
MPIR_ALLTOALL_TAG, recvbuf, 1, newtype,
src, MPIR_ALLTOALL_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);
}
MPIR_Type_free_impl(&newtype);
pof2 *= 2;
}
/* Rotate blocks in recvbuf upwards by (rank + 1) blocks. Need
* a temporary buffer of the same size as recvbuf. */
/* get true extent of recvtype */
MPIR_Type_get_true_extent_impl(recvtype, &recvtype_true_lb, &recvtype_true_extent);
recvbuf_extent = recvcount * comm_size *
(MPIR_MAX(recvtype_true_extent, recvtype_extent));
MPIU_CHKLMEM_MALLOC(tmp_buf, void *, recvbuf_extent, mpi_errno, "tmp_buf");
/* adjust for potential negative lower bound in datatype */
tmp_buf = (void *)((char*)tmp_buf - recvtype_true_lb);
mpi_errno = MPIR_Localcopy((char *) recvbuf + (rank+1)*recvcount*recvtype_extent,
(comm_size - rank - 1)*recvcount, recvtype, tmp_buf,
(comm_size - rank - 1)*recvcount, recvtype);
if (mpi_errno) { MPIU_ERR_POP(mpi_errno); }
mpi_errno = MPIR_Localcopy(recvbuf, (rank+1)*recvcount, recvtype,
(char *) tmp_buf + (comm_size-rank-1)*recvcount*recvtype_extent,
(rank+1)*recvcount, recvtype);
if (mpi_errno) { MPIU_ERR_POP(mpi_errno); }
/* Blocks are in the reverse order now (comm_size-1 to 0).
* Reorder them to (0 to comm_size-1) and store them in recvbuf. */
for (i=0; i<comm_size; i++){
mpi_errno = MPIR_Localcopy((char *) tmp_buf + i*recvcount*recvtype_extent,
recvcount, recvtype,
(char *) recvbuf + (comm_size-i-1)*recvcount*recvtype_extent,
recvcount, recvtype);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
}
#ifdef MPIR_OLD_SHORT_ALLTOALL_ALG
/* Short message. Use recursive doubling. Each process sends all
its data at each step along with all data it received in
previous steps. */
/* need to allocate temporary buffer of size
sendbuf_extent*comm_size */
/* get true extent of sendtype */
MPIR_Type_get_true_extent_impl(sendtype, &sendtype_true_lb, &sendtype_true_extent);
sendbuf_extent = sendcount * comm_size *
(MPIR_MAX(sendtype_true_extent, sendtype_extent));
MPIU_CHKLMEM_MALLOC(tmp_buf, void *, sendbuf_extent*comm_size, mpi_errno, "tmp_buf");
/* adjust for potential negative lower bound in datatype */
tmp_buf = (void *)((char*)tmp_buf - sendtype_true_lb);
/* copy local sendbuf into tmp_buf at location indexed by rank */
curr_cnt = sendcount*comm_size;
mpi_errno = MPIR_Localcopy(sendbuf, curr_cnt, sendtype,
((char *)tmp_buf + rank*sendbuf_extent),
curr_cnt, sendtype);
if (mpi_errno) { MPIU_ERR_POP(mpi_errno);}
mask = 0x1;
i = 0;
while (mask < comm_size) {
dst = rank ^ mask;
dst_tree_root = dst >> i;
dst_tree_root <<= i;
my_tree_root = rank >> i;
my_tree_root <<= i;
if (dst < comm_size) {
mpi_errno = MPIC_Sendrecv_ft(((char *)tmp_buf +
my_tree_root*sendbuf_extent),
curr_cnt, sendtype,
dst, MPIR_ALLTOALL_TAG,
((char *)tmp_buf +
dst_tree_root*sendbuf_extent),
sendbuf_extent*(comm_size-dst_tree_root),
sendtype, dst, MPIR_ALLTOALL_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);
last_recv_cnt = 0;
} else
/* in case of non-power-of-two nodes, less data may be
received than specified */
MPIR_Get_count_impl(&status, sendtype, &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
result. We use a logarithmic recursive-halfing algorithm
for this. */
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)) {
/* send the data received in this step above */
mpi_errno = MPIC_Send_ft(((char *)tmp_buf +
dst_tree_root*sendbuf_extent),
last_recv_cnt, sendtype,
dst, MPIR_ALLTOALL_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);
}
}
/* 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)) {
mpi_errno = MPIC_Recv_ft(((char *)tmp_buf +
dst_tree_root*sendbuf_extent),
sendbuf_extent*(comm_size-dst_tree_root),
sendtype,
dst, MPIR_ALLTOALL_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);
last_recv_cnt = 0;
} else
MPIR_Get_count_impl(&status, sendtype, &last_recv_cnt);
curr_cnt += last_recv_cnt;
}
tmp_mask >>= 1;
k--;
}
}
mask <<= 1;
i++;
}
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 */