/* Algorithm: Short Linear Gather
*
* This linear gather algorithm is tuned for short messages. The remote group
* does a local intracommunicator gather to rank 0. Rank 0 then sends data to
* root.
*
* Cost: (lgp+1).alpha + n.((p-1)/p).beta + n.beta
*/
int MPIR_Igather_sched_inter_short(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
void *recvbuf, int recvcount, MPI_Datatype recvtype, int root,
MPIR_Comm * comm_ptr, MPIR_Sched_t s)
{
int mpi_errno = MPI_SUCCESS;
int rank;
MPI_Aint local_size, remote_size;
MPIR_Comm *newcomm_ptr = NULL;
MPIR_SCHED_CHKPMEM_DECL(1);
remote_size = comm_ptr->remote_size;
local_size = comm_ptr->local_size;
if (root == MPI_ROOT) {
/* root receives data from rank 0 on remote group */
mpi_errno = MPIR_Sched_recv(recvbuf, recvcount * remote_size, recvtype, 0, comm_ptr, s);
if (mpi_errno)
MPIR_ERR_POP(mpi_errno);
} else {
/* remote group. Rank 0 allocates temporary buffer, does
* local intracommunicator gather, and then sends the data
* to root. */
MPI_Aint sendtype_sz;
void *tmp_buf = NULL;
rank = comm_ptr->rank;
if (rank == 0) {
MPIR_Datatype_get_size_macro(sendtype, sendtype_sz);
MPIR_SCHED_CHKPMEM_MALLOC(tmp_buf, void *,
sendcount * local_size * sendtype_sz,
mpi_errno, "tmp_buf", MPL_MEM_BUFFER);
} else {
int MPIR_Iallgather_sched_inter_local_gather_remote_bcast(const void *sendbuf, int sendcount,
MPI_Datatype sendtype, void *recvbuf,
int recvcount, MPI_Datatype recvtype,
MPIR_Comm * comm_ptr, MPIR_Sched_t s)
{
int mpi_errno = MPI_SUCCESS;
int rank, local_size, remote_size, root;
MPI_Aint true_extent, true_lb = 0, extent, send_extent;
void *tmp_buf = NULL;
MPIR_Comm *newcomm_ptr = NULL;
MPIR_SCHED_CHKPMEM_DECL(1);
local_size = comm_ptr->local_size;
remote_size = comm_ptr->remote_size;
rank = comm_ptr->rank;
if ((rank == 0) && (sendcount != 0)) {
/* In each group, rank 0 allocates temp. buffer for local
* gather */
MPIR_Type_get_true_extent_impl(sendtype, &true_lb, &true_extent);
MPIR_Datatype_get_extent_macro(sendtype, send_extent);
extent = MPL_MAX(send_extent, true_extent);
MPIR_Ensure_Aint_fits_in_pointer(extent * sendcount * local_size);
MPIR_SCHED_CHKPMEM_MALLOC(tmp_buf, void *, extent * sendcount * local_size, mpi_errno,
"tmp_buf", MPL_MEM_BUFFER);
/* adjust for potential negative lower bound in datatype */
tmp_buf = (void *) ((char *) tmp_buf - true_lb);
}
int MPIR_Ireduce_sched_intra_binomial(const void *sendbuf, void *recvbuf, int count,
MPI_Datatype datatype, MPI_Op op, int root,
MPIR_Comm * comm_ptr, MPIR_Sched_t s)
{
int mpi_errno = MPI_SUCCESS;
int comm_size, rank, is_commutative;
int mask, relrank, source, lroot;
MPI_Aint true_lb, true_extent, extent;
void *tmp_buf;
MPIR_SCHED_CHKPMEM_DECL(2);
MPIR_Assert(comm_ptr->comm_kind == MPIR_COMM_KIND__INTRACOMM);
if (count == 0)
return MPI_SUCCESS;
comm_size = comm_ptr->local_size;
rank = comm_ptr->rank;
/* set op_errno to 0. stored in perthread structure */
{
MPIR_Per_thread_t *per_thread = NULL;
int err = 0;
MPID_THREADPRIV_KEY_GET_ADDR(MPIR_ThreadInfo.isThreaded, MPIR_Per_thread_key,
MPIR_Per_thread, per_thread, &err);
MPIR_Assert(err == 0);
per_thread->op_errno = 0;
}
/* Create a temporary buffer */
MPIR_Type_get_true_extent_impl(datatype, &true_lb, &true_extent);
MPIR_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}? */
MPIR_Ensure_Aint_fits_in_pointer(count * MPL_MAX(extent, true_extent));
MPIR_SCHED_CHKPMEM_MALLOC(tmp_buf, void *, count * (MPL_MAX(extent, true_extent)),
mpi_errno, "temporary buffer", MPL_MEM_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) {
MPIR_SCHED_CHKPMEM_MALLOC(recvbuf, void *,
count * (MPL_MAX(extent, true_extent)),
mpi_errno, "receive buffer", MPL_MEM_BUFFER);
recvbuf = (void *) ((char *) recvbuf - true_lb);
}
int MPIR_Iscatter_sched_intra_binomial(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
void *recvbuf, int recvcount, MPI_Datatype recvtype,
int root, MPIR_Comm * comm_ptr, MPIR_Sched_t s)
{
int mpi_errno = MPI_SUCCESS;
MPI_Aint extent = 0;
int rank, comm_size, sendtype_size;
int relative_rank;
int mask, recvtype_size = 0, src, dst;
int tmp_buf_size = 0;
void *tmp_buf = NULL;
struct shared_state *ss = NULL;
MPIR_SCHED_CHKPMEM_DECL(4);
comm_size = comm_ptr->local_size;
rank = comm_ptr->rank;
if (((rank == root) && (sendcount == 0)) || ((rank != root) && (recvcount == 0)))
goto fn_exit;
MPIR_SCHED_CHKPMEM_MALLOC(ss, struct shared_state *, sizeof(struct shared_state), mpi_errno,
"shared_state", MPL_MEM_BUFFER);
ss->sendcount = sendcount;
if (rank == root)
MPIR_Datatype_get_extent_macro(sendtype, extent);
relative_rank = (rank >= root) ? rank - root : rank - root + comm_size;
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 */
MPIR_Datatype_get_size_macro(sendtype, sendtype_size);
MPIR_Ensure_Aint_fits_in_pointer(MPIR_VOID_PTR_CAST_TO_MPI_AINT sendbuf +
extent * sendcount * comm_size);
ss->nbytes = sendtype_size * sendcount;
} else {
MPIR_Datatype_get_size_macro(recvtype, recvtype_size);
MPIR_Ensure_Aint_fits_in_pointer(extent * recvcount * comm_size);
ss->nbytes = recvtype_size * recvcount;
}
ss->curr_count = 0;
/* all even nodes other than root need a temporary buffer to
* receive data of max size (ss->nbytes*comm_size)/2 */
if (relative_rank && !(relative_rank % 2)) {
tmp_buf_size = (ss->nbytes * comm_size) / 2;
MPIR_SCHED_CHKPMEM_MALLOC(tmp_buf, void *, tmp_buf_size, mpi_errno, "tmp_buf",
MPL_MEM_BUFFER);
}
int MPIR_Ibcast_binomial(void *buffer, int count, MPI_Datatype datatype, int root, MPID_Comm *comm_ptr, MPID_Sched_t s)
{
int mpi_errno = MPI_SUCCESS;
int mask;
int comm_size, rank;
int is_contig, is_homogeneous;
MPI_Aint nbytes, type_size;
int relative_rank;
int src, dst;
void *tmp_buf = NULL;
MPIR_SCHED_CHKPMEM_DECL(1);
comm_size = comm_ptr->local_size;
rank = comm_ptr->rank;
if (comm_size == 1) {
/* nothing to add, this is a useless broadcast */
goto fn_exit;
}
MPID_Datatype_is_contig(datatype, &is_contig);
is_homogeneous = 1;
#ifdef MPID_HAS_HETERO
if (comm_ptr->is_hetero)
is_homogeneous = 0;
#endif
/* MPI_Type_size() might not give the accurate size of the packed
* datatype for heterogeneous systems (because of padding, encoding,
* etc). On the other hand, MPI_Pack_size() can become very
* expensive, depending on the implementation, especially for
* heterogeneous systems. We want to use MPI_Type_size() wherever
* possible, and MPI_Pack_size() in other places.
*/
if (is_homogeneous)
MPID_Datatype_get_size_macro(datatype, type_size);
else
MPIR_Pack_size_impl(1, datatype, &type_size);
nbytes = type_size * count;
if (!is_contig || !is_homogeneous)
{
MPIR_SCHED_CHKPMEM_MALLOC(tmp_buf, void *, nbytes, mpi_errno, "tmp_buf");
/* TODO: Pipeline the packing and communication */
if (rank == root) {
mpi_errno = MPID_Sched_copy(buffer, count, datatype, tmp_buf, nbytes, MPI_PACKED, s);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
MPID_SCHED_BARRIER(s);
}
}
int MPIR_Iscan_SMP(const void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype, MPI_Op op, MPIR_Comm *comm_ptr, MPIR_Sched_t s)
{
int mpi_errno = MPI_SUCCESS;
int rank = comm_ptr->rank;
MPIR_Comm *node_comm;
MPIR_Comm *roots_comm;
MPI_Aint true_extent, true_lb, extent;
void *tempbuf = NULL;
void *prefulldata = NULL;
void *localfulldata = NULL;
MPIR_SCHED_CHKPMEM_DECL(3);
/* 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 (!MPII_Comm_is_node_consecutive(comm_ptr)) {
/* We can't use the SMP-aware algorithm, use the generic one */
return MPIR_Iscan_rec_dbl(sendbuf, recvbuf, count, datatype, op, comm_ptr, s);
}
node_comm = comm_ptr->node_comm;
roots_comm = comm_ptr->node_roots_comm;
if (node_comm) {
MPIR_Assert(node_comm->coll_fns && node_comm->coll_fns->Iscan_sched && node_comm->coll_fns->Ibcast_sched);
}
if (roots_comm) {
MPIR_Assert(roots_comm->coll_fns && roots_comm->coll_fns->Iscan_sched);
}
MPIR_Type_get_true_extent_impl(datatype, &true_lb, &true_extent);
MPID_Datatype_get_extent_macro(datatype, extent);
MPIR_Ensure_Aint_fits_in_pointer(count * MPL_MAX(extent, true_extent));
MPIR_SCHED_CHKPMEM_MALLOC(tempbuf, void *, count*(MPL_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) {
MPIR_SCHED_CHKPMEM_MALLOC(prefulldata, void *, count*(MPL_MAX(extent, true_extent)),
mpi_errno, "prefulldata for scan");
prefulldata = (void *)((char*)prefulldata - true_lb);
if (node_comm != NULL) {
MPIR_SCHED_CHKPMEM_MALLOC(localfulldata, void *, count*(MPL_MAX(extent, true_extent)),
mpi_errno, "localfulldata for scan");
localfulldata = (void *)((char*)localfulldata - true_lb);
}
int MPIR_Iallgather_rec_dbl(const void *sendbuf, int sendcount, MPI_Datatype sendtype, void *recvbuf, int recvcount, MPI_Datatype recvtype, MPIR_Comm *comm_ptr, MPIR_Sched_t s)
{
int mpi_errno = MPI_SUCCESS;
struct shared_state *ss = NULL;
int comm_size, rank;
int i, j, k;
int mask, tmp_mask, dst;
int dst_tree_root, my_tree_root, tree_root;
int offset, send_offset, recv_offset;
MPI_Aint recvtype_extent;
MPIR_Datatype *recv_dtp;
MPIR_SCHED_CHKPMEM_DECL(1);
comm_size = comm_ptr->local_size;
rank = comm_ptr->rank;
recv_dtp = NULL;
if (HANDLE_GET_KIND(recvtype) != HANDLE_KIND_BUILTIN) {
MPID_Datatype_get_ptr(recvtype, recv_dtp);
}
MPID_Datatype_get_extent_macro( recvtype, recvtype_extent );
/* This is the largest offset we add to recvbuf */
MPIR_Ensure_Aint_fits_in_pointer(MPIR_VOID_PTR_CAST_TO_MPI_AINT recvbuf +
(comm_size * recvcount * recvtype_extent));
/* copy local data into recvbuf */
if (sendbuf != MPI_IN_PLACE) {
mpi_errno = MPIR_Sched_copy(sendbuf, sendcount, sendtype,
((char *)recvbuf + rank*recvcount*recvtype_extent),
recvcount, recvtype, s);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
MPIR_SCHED_BARRIER(s);
}
MPIR_SCHED_CHKPMEM_MALLOC(ss, struct shared_state *, sizeof(struct shared_state), mpi_errno, "ss");
ss->curr_count = recvcount;
ss->recvtype = recvtype;
/* ensure that recvtype doesn't disappear immediately after last _recv but before _cb */
if (recv_dtp)
MPID_Datatype_add_ref(recv_dtp);
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;
/* saving an MPI_Aint into an int, overflow checked above */
send_offset = my_tree_root * recvcount * recvtype_extent;
recv_offset = dst_tree_root * recvcount * recvtype_extent;
if (dst < comm_size) {
mpi_errno = MPIR_Sched_send_defer(((char *)recvbuf + send_offset),
&ss->curr_count, recvtype, dst, comm_ptr, s);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
/* send-recv, no sched barrier here */
mpi_errno = MPIR_Sched_recv_status(((char *)recvbuf + recv_offset),
((comm_size-dst_tree_root)*recvcount),
recvtype, dst, comm_ptr, &ss->status, s);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
MPIR_SCHED_BARRIER(s);
mpi_errno = MPIR_Sched_cb(&get_count, ss, s);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
MPIR_SCHED_BARRIER(s);
}
/* 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 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--;
/* FIXME: saving an MPI_Aint into an int */
offset = recvcount * (my_tree_root + mask) * recvtype_extent;
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))
{
/* last_recv_count was set in the previous
receive. that's the amount of data to be
sent now. */
mpi_errno = MPIR_Sched_send_defer(((char *)recvbuf + offset),
&ss->last_recv_count,
recvtype, dst, comm_ptr, s);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
MPIR_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))
{
/* nprocs_completed is also equal to the
no. of processes whose data we don't have */
mpi_errno = MPIR_Sched_recv_status(((char *)recvbuf + offset),
((comm_size - (my_tree_root + mask))*recvcount),
recvtype, dst, comm_ptr, &ss->status, s);
MPIR_SCHED_BARRIER(s);
mpi_errno = MPIR_Sched_cb(&get_count, ss, s);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
MPIR_SCHED_BARRIER(s);
}
tmp_mask >>= 1;
k--;
}
}
/* --END EXPERIMENTAL-- */
mask <<= 1;
i++;
}
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_Ibarrier_inter(MPID_Comm *comm_ptr, MPID_Sched_t s)
{
int mpi_errno = MPI_SUCCESS;
int rank, root;
MPIR_SCHED_CHKPMEM_DECL(1);
char *buf = NULL;
MPIU_Assert(comm_ptr->comm_kind == MPID_INTERCOMM);
rank = comm_ptr->rank;
/* Get the local intracommunicator */
if (!comm_ptr->local_comm) {
mpi_errno = MPIR_Setup_intercomm_localcomm(comm_ptr);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
}
/* do a barrier on the local intracommunicator */
MPIU_Assert(comm_ptr->local_comm->coll_fns && comm_ptr->local_comm->coll_fns->Ibarrier_sched);
if(comm_ptr->local_size != 1) {
mpi_errno = comm_ptr->local_comm->coll_fns->Ibarrier_sched(comm_ptr->local_comm, s);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
MPID_SCHED_BARRIER(s);
}
/* 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. */
MPIR_SCHED_CHKPMEM_MALLOC(buf, char *, 1, mpi_errno, "bcast buf");
buf[0] = 'D'; /* avoid valgrind warnings */
/* first broadcast from left to right group, then from right to
left group */
MPIU_Assert(comm_ptr->coll_fns && comm_ptr->coll_fns->Ibcast_sched);
if (comm_ptr->is_low_group) {
root = (rank == 0) ? MPI_ROOT : MPI_PROC_NULL;
mpi_errno = comm_ptr->coll_fns->Ibcast_sched(buf, 1, MPI_BYTE, root, comm_ptr, s);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
MPID_SCHED_BARRIER(s);
/* receive bcast from right */
root = 0;
mpi_errno = comm_ptr->coll_fns->Ibcast_sched(buf, 1, MPI_BYTE, root, comm_ptr, s);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
}
else {
/* receive bcast from left */
root = 0;
mpi_errno = comm_ptr->coll_fns->Ibcast_sched(buf, 1, MPI_BYTE, root, comm_ptr, s);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
MPID_SCHED_BARRIER(s);
/* bcast to left */
root = (rank == 0) ? MPI_ROOT : MPI_PROC_NULL;
mpi_errno = comm_ptr->coll_fns->Ibcast_sched(buf, 1, MPI_BYTE, root, comm_ptr, s);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
}
MPIR_SCHED_CHKPMEM_COMMIT(s);
fn_exit:
return mpi_errno;
fn_fail:
MPIR_SCHED_CHKPMEM_REAP(s);
goto fn_exit;
}
int MPIR_Iscatter_intra(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
void *recvbuf, int recvcount, MPI_Datatype recvtype,
int root, MPID_Comm *comm_ptr, MPID_Sched_t s)
{
int mpi_errno = MPI_SUCCESS;
MPI_Aint extent = 0;
int rank, comm_size, is_homogeneous, sendtype_size;
int relative_rank;
int mask, recvtype_size=0, src, dst;
int tmp_buf_size = 0;
void *tmp_buf = NULL;
struct shared_state *ss = NULL;
MPIR_SCHED_CHKPMEM_DECL(4);
comm_size = comm_ptr->local_size;
rank = comm_ptr->rank;
if (((rank == root) && (sendcount == 0)) || ((rank != root) && (recvcount == 0)))
goto fn_exit;
is_homogeneous = 1;
#ifdef MPID_HAS_HETERO
if (comm_ptr->is_hetero)
is_homogeneous = 0;
#endif
/* Use binomial tree algorithm */
MPIR_SCHED_CHKPMEM_MALLOC(ss, struct shared_state *, sizeof(struct shared_state), mpi_errno, "shared_state");
ss->sendcount = sendcount;
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);
MPIU_Ensure_Aint_fits_in_pointer(MPIU_VOID_PTR_CAST_TO_MPI_AINT sendbuf +
extent*sendcount*comm_size);
ss->nbytes = sendtype_size * sendcount;
}
else {
MPID_Datatype_get_size_macro(recvtype, recvtype_size);
MPIU_Ensure_Aint_fits_in_pointer(extent*recvcount*comm_size);
ss->nbytes = recvtype_size * recvcount;
}
ss->curr_count = 0;
/* all even nodes other than root need a temporary buffer to
receive data of max size (ss->nbytes*comm_size)/2 */
if (relative_rank && !(relative_rank % 2)) {
tmp_buf_size = (ss->nbytes*comm_size)/2;
MPIR_SCHED_CHKPMEM_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 = ss->nbytes*comm_size;
MPIR_SCHED_CHKPMEM_MALLOC(tmp_buf, void *, tmp_buf_size, mpi_errno, "tmp_buf");
if (recvbuf != MPI_IN_PLACE)
mpi_errno = MPID_Sched_copy(((char *) sendbuf + extent*sendcount*rank),
sendcount*(comm_size-rank), sendtype,
tmp_buf, ss->nbytes*(comm_size-rank), MPI_BYTE, s);
else
mpi_errno = MPID_Sched_copy(((char *) sendbuf + extent*sendcount*(rank+1)),
sendcount*(comm_size-rank-1), sendtype,
((char *)tmp_buf + ss->nbytes),
ss->nbytes*(comm_size-rank-1), MPI_BYTE, s);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
mpi_errno = MPID_Sched_copy(sendbuf, sendcount*rank, sendtype,
((char *) tmp_buf + ss->nbytes*(comm_size-rank)),
ss->nbytes*rank, MPI_BYTE, s);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
MPID_SCHED_BARRIER(s);
ss->curr_count = ss->nbytes*comm_size;
}
else
int MPIR_Ialltoall_inplace(const void *sendbuf, int sendcount, MPI_Datatype sendtype, void *recvbuf, int recvcount, MPI_Datatype recvtype, MPID_Comm *comm_ptr, MPID_Sched_t s)
{
int mpi_errno = MPI_SUCCESS;
void *tmp_buf = NULL;
int i, j;
int rank, comm_size;
int nbytes, recvtype_size;
MPI_Aint recvtype_extent;
int peer;
MPIR_SCHED_CHKPMEM_DECL(1);
MPIU_Assert(sendbuf == MPI_IN_PLACE);
if (recvcount == 0)
goto fn_exit;
comm_size = comm_ptr->local_size;
rank = comm_ptr->rank;
MPID_Datatype_get_size_macro(recvtype, recvtype_size);
MPID_Datatype_get_extent_macro(recvtype, recvtype_extent);
nbytes = recvtype_size * recvcount;
MPIR_SCHED_CHKPMEM_MALLOC(tmp_buf, void *, nbytes, mpi_errno, "tmp_buf");
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 && rank == j) {
/* no need to "sendrecv_replace" for ourselves */
}
else if (rank == i || rank == j) {
if (rank == i)
peer = j;
else
peer = i;
/* pack to tmp_buf */
mpi_errno = MPID_Sched_copy(((char *)recvbuf + peer*recvcount*recvtype_extent),
recvcount, recvtype,
tmp_buf, nbytes, MPI_BYTE, s);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
MPID_SCHED_BARRIER(s);
/* now simultaneously send from tmp_buf and recv to recvbuf */
mpi_errno = MPID_Sched_send(tmp_buf, nbytes, MPI_BYTE, peer, comm_ptr, s);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
mpi_errno = MPID_Sched_recv(((char *)recvbuf + peer*recvcount*recvtype_extent),
recvcount, recvtype, peer, comm_ptr, s);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
MPID_SCHED_BARRIER(s);
}
}
}
MPIR_SCHED_CHKPMEM_COMMIT(s);
fn_exit:
return mpi_errno;
fn_fail:
MPIR_SCHED_CHKPMEM_REAP(s);
goto fn_exit;
}
int MPIR_Ialltoall_bruck(const void *sendbuf, int sendcount, MPI_Datatype sendtype, void *recvbuf, int recvcount, MPI_Datatype recvtype, MPID_Comm *comm_ptr, MPID_Sched_t s)
{
int mpi_errno = MPI_SUCCESS;
int i;
int nbytes, recvtype_size, recvbuf_extent, newtype_size;
int rank, comm_size;
void *tmp_buf = NULL;
MPI_Aint sendtype_extent, recvtype_extent, recvtype_true_lb, recvtype_true_extent;
int pof2, dst, src;
int count, block;
MPI_Datatype newtype;
int *displs;
MPIU_CHKLMEM_DECL(1); /* displs */
MPIR_SCHED_CHKPMEM_DECL(2); /* tmp_buf (2x) */
MPIU_Assert(sendbuf != MPI_IN_PLACE); /* we do not handle in-place */
comm_size = comm_ptr->local_size;
rank = comm_ptr->rank;
MPID_Datatype_get_extent_macro(sendtype, sendtype_extent);
MPID_Datatype_get_size_macro(recvtype, recvtype_size);
MPID_Datatype_get_extent_macro(recvtype, recvtype_extent);
/* allocate temporary buffer */
/* must be same size as entire recvbuf for Phase 3 */
nbytes = recvtype_size * recvcount * comm_size;
MPIR_SCHED_CHKPMEM_MALLOC(tmp_buf, void *, nbytes, 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 = MPID_Sched_copy(((char *) sendbuf + rank*sendcount*sendtype_extent),
(comm_size - rank)*sendcount, sendtype,
recvbuf, (comm_size - rank)*recvcount, recvtype, s);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
mpi_errno = MPID_Sched_copy(sendbuf, rank*sendcount, sendtype,
((char *) recvbuf + (comm_size-rank)*recvcount*recvtype_extent),
rank*recvcount, recvtype, s);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
MPID_SCHED_BARRIER(s);
/* 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);
MPID_Datatype_get_size_macro(newtype, newtype_size);
/* we will usually copy much less than nbytes */
mpi_errno = MPID_Sched_copy(recvbuf, 1, newtype, tmp_buf, newtype_size, MPI_BYTE, s);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
MPID_SCHED_BARRIER(s);
/* now send and recv in parallel */
mpi_errno = MPID_Sched_send(tmp_buf, newtype_size, MPI_BYTE, dst, comm_ptr, s);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
mpi_errno = MPID_Sched_recv(recvbuf, 1, newtype, src, comm_ptr, s);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
MPID_SCHED_BARRIER(s);
MPIR_Type_free_impl(&newtype);
pof2 *= 2;
}
/* Phase 3: 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));
/* not a leak, old tmp_buf value is still tracked by CHKPMEM macros */
MPIR_SCHED_CHKPMEM_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 = MPID_Sched_copy(((char *) recvbuf + (rank+1)*recvcount*recvtype_extent),
(comm_size - rank - 1)*recvcount, recvtype,
tmp_buf, (comm_size - rank - 1)*recvcount, recvtype, s);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
mpi_errno = MPID_Sched_copy(recvbuf, (rank+1)*recvcount, recvtype,
((char *) tmp_buf + (comm_size-rank-1)*recvcount*recvtype_extent),
(rank+1)*recvcount, recvtype, s);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
MPID_SCHED_BARRIER(s);
/* 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 = MPID_Sched_copy(((char *) tmp_buf + i*recvcount*recvtype_extent),
recvcount, recvtype,
((char *) recvbuf + (comm_size-i-1)*recvcount*recvtype_extent),
recvcount, recvtype, s);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
}
MPIR_SCHED_CHKPMEM_COMMIT(s);
fn_exit:
MPIU_CHKLMEM_FREEALL();
return mpi_errno;
fn_fail:
MPIR_SCHED_CHKPMEM_REAP(s);
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++;
}
int MPIR_Ireduce_scatter_sched_intra_recursive_doubling(const void *sendbuf, void *recvbuf, const int recvcounts[],
MPI_Datatype datatype, MPI_Op op, MPIR_Comm *comm_ptr,
MPIR_Sched_t s)
{
int mpi_errno = MPI_SUCCESS;
int rank, comm_size, i;
MPI_Aint extent, true_extent, true_lb;
int *disps;
void *tmp_recvbuf, *tmp_results;
int type_size ATTRIBUTE((unused)), dis[2], blklens[2], total_count, dst;
int mask, dst_tree_root, my_tree_root, j, k;
int received;
MPI_Datatype sendtype, recvtype;
int nprocs_completed, tmp_mask, tree_root, is_commutative;
MPIR_SCHED_CHKPMEM_DECL(5);
comm_size = comm_ptr->local_size;
rank = comm_ptr->rank;
MPIR_Datatype_get_extent_macro(datatype, extent);
MPIR_Type_get_true_extent_impl(datatype, &true_lb, &true_extent);
is_commutative = MPIR_Op_is_commutative(op);
MPIR_SCHED_CHKPMEM_MALLOC(disps, int *, comm_size * sizeof(int), mpi_errno, "disps", MPL_MEM_BUFFER);
total_count = 0;
for (i=0; i<comm_size; i++) {
disps[i] = total_count;
total_count += recvcounts[i];
}
if (total_count == 0) {
goto fn_exit;
}
MPIR_Datatype_get_size_macro(datatype, type_size);
/* total_count*extent eventually gets malloced. it isn't added to
* a user-passed in buffer */
MPIR_Ensure_Aint_fits_in_pointer(total_count * MPL_MAX(true_extent, extent));
/* need to allocate temporary buffer to receive incoming data*/
MPIR_SCHED_CHKPMEM_MALLOC(tmp_recvbuf, void *, total_count*(MPL_MAX(true_extent,extent)), mpi_errno, "tmp_recvbuf", MPL_MEM_BUFFER);
/* 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 */
MPIR_SCHED_CHKPMEM_MALLOC(tmp_results, void *, total_count*(MPL_MAX(true_extent,extent)), mpi_errno, "tmp_results", MPL_MEM_BUFFER);
/* 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_Sched_copy(sendbuf, total_count, datatype,
tmp_results, total_count, datatype, s);
else
mpi_errno = MPIR_Sched_copy(recvbuf, total_count, datatype,
tmp_results, total_count, datatype, s);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
MPIR_SCHED_BARRIER(s);
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;
/* At step 1, processes exchange (n-n/p) amount of
data; at step 2, (n-2n/p) amount of data; at step 3, (n-4n/p)
amount of data, and so forth. We use derived datatypes for this.
At each step, a process does not need to send data
indexed from my_tree_root to
my_tree_root+mask-1. Similarly, a process won't receive
data indexed from dst_tree_root to dst_tree_root+mask-1. */
/* calculate sendtype */
blklens[0] = blklens[1] = 0;
for (j=0; j<my_tree_root; j++)
blklens[0] += recvcounts[j];
for (j=my_tree_root+mask; j<comm_size; j++)
blklens[1] += recvcounts[j];
dis[0] = 0;
dis[1] = blklens[0];
for (j=my_tree_root; (j<my_tree_root+mask) && (j<comm_size); j++)
dis[1] += recvcounts[j];
mpi_errno = MPIR_Type_indexed_impl(2, blklens, dis, datatype, &sendtype);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
mpi_errno = MPIR_Type_commit_impl(&sendtype);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
/* calculate recvtype */
blklens[0] = blklens[1] = 0;
for (j=0; j<dst_tree_root && j<comm_size; j++)
blklens[0] += recvcounts[j];
for (j=dst_tree_root+mask; j<comm_size; j++)
blklens[1] += recvcounts[j];
dis[0] = 0;
dis[1] = blklens[0];
for (j=dst_tree_root; (j<dst_tree_root+mask) && (j<comm_size); j++)
dis[1] += recvcounts[j];
mpi_errno = MPIR_Type_indexed_impl(2, blklens, dis, datatype, &recvtype);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
mpi_errno = MPIR_Type_commit_impl(&recvtype);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
received = 0;
if (dst < comm_size) {
/* tmp_results contains data to be sent in each step. Data is
received in tmp_recvbuf and then accumulated into
tmp_results. accumulation is done later below. */
mpi_errno = MPIR_Sched_send(tmp_results, 1, sendtype, dst, comm_ptr, s);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
mpi_errno = MPIR_Sched_recv(tmp_recvbuf, 1, recvtype, dst, comm_ptr, s);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
MPIR_SCHED_BARRIER(s);
received = 1;
}
/* 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 current result */
mpi_errno = MPIR_Sched_send(tmp_recvbuf, 1, recvtype, dst, comm_ptr, s);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
MPIR_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))
{
mpi_errno = MPIR_Sched_recv(tmp_recvbuf, 1, recvtype, dst, comm_ptr, s);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
MPIR_SCHED_BARRIER(s);
received = 1;
}
tmp_mask >>= 1;
k--;
}
}
/* N.B. The following comment comes from the FT version of
* MPI_Reduce_scatter. It does not currently apply to this code, but
* will in the future when we update the NBC code to be fault-tolerant
* in roughly the same fashion. [goodell@ 2011-03-03] */
/* The following reduction is done here instead of after
the MPIC_Sendrecv or MPIC_Recv above. This is
because to do it above, in the noncommutative
case, we would need an extra temp buffer so as not to
overwrite temp_recvbuf, because temp_recvbuf may have
to be communicated to other processes in the
non-power-of-two case. To avoid that extra allocation,
we do the reduce here. */
if (received) {
if (is_commutative || (dst_tree_root < my_tree_root)) {
mpi_errno = MPIR_Sched_reduce(tmp_recvbuf, tmp_results, blklens[0], datatype, op, s);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
mpi_errno = MPIR_Sched_reduce(((char *)tmp_recvbuf + dis[1]*extent),
((char *)tmp_results + dis[1]*extent),
blklens[1], datatype, op, s);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
MPIR_SCHED_BARRIER(s);
}
else {
mpi_errno = MPIR_Sched_reduce(tmp_results, tmp_recvbuf, blklens[0], datatype, op, s);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
mpi_errno = MPIR_Sched_reduce(((char *)tmp_results + dis[1]*extent),
((char *)tmp_recvbuf + dis[1]*extent),
blklens[1], datatype, op, s);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
MPIR_SCHED_BARRIER(s);
/* copy result back into tmp_results */
mpi_errno = MPIR_Sched_copy(tmp_recvbuf, 1, recvtype,
tmp_results, 1, recvtype, s);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
MPIR_SCHED_BARRIER(s);
}
}
MPIR_Type_free_impl(&sendtype);
MPIR_Type_free_impl(&recvtype);
mask <<= 1;
i++;
}
int MPIR_Ireduce_scatter_sched_intra_recursive_halving(const void *sendbuf, void *recvbuf,
const int recvcounts[],
MPI_Datatype datatype, MPI_Op op,
MPIR_Comm * comm_ptr, MPIR_Sched_t s)
{
int mpi_errno = MPI_SUCCESS;
int rank, comm_size, i;
MPI_Aint extent, true_extent, true_lb;
int *disps;
void *tmp_recvbuf, *tmp_results;
int type_size ATTRIBUTE((unused)), total_count, dst;
int mask;
int *newcnts, *newdisps, rem, newdst, send_idx, recv_idx, last_idx, send_cnt, recv_cnt;
int pof2, old_i, newrank;
MPIR_SCHED_CHKPMEM_DECL(5);
comm_size = comm_ptr->local_size;
rank = comm_ptr->rank;
MPIR_Datatype_get_extent_macro(datatype, extent);
MPIR_Type_get_true_extent_impl(datatype, &true_lb, &true_extent);
#ifdef HAVE_ERROR_CHECKING
MPIR_Assert(MPIR_Op_is_commutative(op));
#endif
MPIR_SCHED_CHKPMEM_MALLOC(disps, int *, comm_size * sizeof(int), mpi_errno, "disps",
MPL_MEM_BUFFER);
total_count = 0;
for (i = 0; i < comm_size; i++) {
disps[i] = total_count;
total_count += recvcounts[i];
}
if (total_count == 0) {
goto fn_exit;
}
MPIR_Datatype_get_size_macro(datatype, type_size);
/* allocate temp. buffer to receive incoming data */
MPIR_SCHED_CHKPMEM_MALLOC(tmp_recvbuf, void *, total_count * (MPL_MAX(true_extent, extent)),
mpi_errno, "tmp_recvbuf", MPL_MEM_BUFFER);
/* 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 */
MPIR_SCHED_CHKPMEM_MALLOC(tmp_results, void *, total_count * (MPL_MAX(true_extent, extent)),
mpi_errno, "tmp_results", MPL_MEM_BUFFER);
/* 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_Sched_copy(sendbuf, total_count, datatype,
tmp_results, total_count, datatype, s);
else
mpi_errno = MPIR_Sched_copy(recvbuf, total_count, datatype,
tmp_results, total_count, datatype, s);
if (mpi_errno)
MPIR_ERR_POP(mpi_errno);
MPIR_SCHED_BARRIER(s);
pof2 = comm_ptr->pof2;
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 = MPIR_Sched_send(tmp_results, total_count, datatype, rank + 1, comm_ptr, s);
if (mpi_errno)
MPIR_ERR_POP(mpi_errno);
MPIR_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 = MPIR_Sched_recv(tmp_recvbuf, total_count, datatype, rank - 1, comm_ptr, s);
if (mpi_errno)
MPIR_ERR_POP(mpi_errno);
MPIR_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 = MPIR_Sched_reduce(tmp_recvbuf, tmp_results, total_count, datatype, op, s);
if (mpi_errno)
MPIR_ERR_POP(mpi_errno);
MPIR_SCHED_BARRIER(s);
/* change the rank */
newrank = rank / 2;
}
} else /* rank >= 2*rem */
newrank = rank - rem;
if (newrank != -1) {
/* recalculate the recvcounts 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). */
MPIR_SCHED_CHKPMEM_MALLOC(newcnts, int *, pof2 * sizeof(int), mpi_errno, "newcnts",
MPL_MEM_BUFFER);
MPIR_SCHED_CHKPMEM_MALLOC(newdisps, int *, pof2 * sizeof(int), mpi_errno, "newdisps",
MPL_MEM_BUFFER);
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] = recvcounts[old_i] + recvcounts[old_i - 1];
} else
newcnts[i] = recvcounts[old_i];
}
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];
}
/* Send data from tmp_results. Recv into tmp_recvbuf */
{
/* avoid sending and receiving pointless 0-byte messages */
int send_dst = (send_cnt ? dst : MPI_PROC_NULL);
int recv_dst = (recv_cnt ? dst : MPI_PROC_NULL);
mpi_errno = MPIR_Sched_send(((char *) tmp_results + newdisps[send_idx] * extent),
send_cnt, datatype, send_dst, comm_ptr, s);
if (mpi_errno)
MPIR_ERR_POP(mpi_errno);
mpi_errno = MPIR_Sched_recv(((char *) tmp_recvbuf + newdisps[recv_idx] * extent),
recv_cnt, datatype, recv_dst, comm_ptr, s);
if (mpi_errno)
MPIR_ERR_POP(mpi_errno);
MPIR_SCHED_BARRIER(s);
}
/* tmp_recvbuf contains data received in this step.
* tmp_results contains data accumulated so far */
if (recv_cnt) {
mpi_errno = MPIR_Sched_reduce(((char *) tmp_recvbuf + newdisps[recv_idx] * extent),
((char *) tmp_results + newdisps[recv_idx] * extent),
recv_cnt, datatype, op, s);
MPIR_SCHED_BARRIER(s);
}
/* 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 */
if (recvcounts[rank]) {
mpi_errno = MPIR_Sched_copy(((char *) tmp_results + disps[rank] * extent),
recvcounts[rank], datatype,
recvbuf, recvcounts[rank], datatype, s);
if (mpi_errno)
MPIR_ERR_POP(mpi_errno);
MPIR_SCHED_BARRIER(s);
}
}
int MPIR_Igather_sched_intra_binomial(const void *sendbuf, int sendcount, MPI_Datatype sendtype, void *recvbuf, int recvcount, MPI_Datatype recvtype, int root, MPIR_Comm *comm_ptr, MPIR_Sched_t s)
{
int mpi_errno = MPI_SUCCESS;
int comm_size, rank;
int relative_rank, is_homogeneous;
int mask, src, dst, relative_src;
MPI_Aint recvtype_size, sendtype_size, curr_cnt=0, nbytes;
int recvblks;
int tmp_buf_size, missing;
void *tmp_buf = NULL;
int blocks[2];
int displs[2];
MPI_Aint struct_displs[2];
MPI_Aint extent=0;
int copy_offset = 0, copy_blks = 0;
MPI_Datatype types[2], tmp_type;
MPIR_SCHED_CHKPMEM_DECL(1);
comm_size = comm_ptr->local_size;
rank = comm_ptr->rank;
if (((rank == root) && (recvcount == 0)) || ((rank != root) && (sendcount == 0)))
goto fn_exit;
is_homogeneous = TRUE;
#ifdef MPID_HAS_HETERO
is_homogeneous = !comm_ptr->is_hetero;
#endif
MPIR_Assert(comm_ptr->comm_kind == MPIR_COMM_KIND__INTRACOMM);
/* Use binomial tree algorithm. */
relative_rank = (rank >= root) ? rank - root : rank - root + comm_size;
if (rank == root)
{
MPIR_Datatype_get_extent_macro(recvtype, extent);
MPIR_Ensure_Aint_fits_in_pointer(MPIR_VOID_PTR_CAST_TO_MPI_AINT recvbuf+
(extent*recvcount*comm_size));
}
if (is_homogeneous)
{
/* communicator is homogeneous. no need to pack buffer. */
if (rank == root)
{
MPIR_Datatype_get_size_macro(recvtype, recvtype_size);
nbytes = recvtype_size * recvcount;
}
else
{
MPIR_Datatype_get_size_macro(sendtype, sendtype_size);
nbytes = sendtype_size * sendcount;
}
/* Find the number of missing nodes in my sub-tree compared to
* a balanced tree */
for (mask = 1; mask < comm_size; mask <<= 1);
--mask;
while (relative_rank & mask) mask >>= 1;
missing = (relative_rank | mask) - comm_size + 1;
if (missing < 0) missing = 0;
tmp_buf_size = (mask - missing);
/* If the message is smaller than the threshold, we will copy
* our message in there too */
if (nbytes < MPIR_CVAR_GATHER_VSMALL_MSG_SIZE) tmp_buf_size++;
tmp_buf_size *= nbytes;
/* For zero-ranked root, we don't need any temporary buffer */
if ((rank == root) && (!root || (nbytes >= MPIR_CVAR_GATHER_VSMALL_MSG_SIZE)))
tmp_buf_size = 0;
if (tmp_buf_size) {
MPIR_SCHED_CHKPMEM_MALLOC(tmp_buf, void *, tmp_buf_size, mpi_errno, "tmp_buf", MPL_MEM_BUFFER);
}
if (rank == root) {
if (sendbuf != MPI_IN_PLACE) {
mpi_errno = MPIR_Localcopy(sendbuf, sendcount, sendtype,
((char *) recvbuf + extent*recvcount*rank), recvcount, recvtype);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
}
}
else if (tmp_buf_size && (nbytes < MPIR_CVAR_GATHER_VSMALL_MSG_SIZE)) {
/* copy from sendbuf into tmp_buf */
mpi_errno = MPIR_Localcopy(sendbuf, sendcount, sendtype,
tmp_buf, nbytes, MPI_BYTE);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
}
curr_cnt = nbytes;
mask = 0x1;
while (mask < comm_size) {
if ((mask & relative_rank) == 0) {
src = relative_rank | mask;
if (src < comm_size) {
src = (src + root) % comm_size;
if (rank == root) {
recvblks = mask;
if ((2 * recvblks) > comm_size)
recvblks = comm_size - recvblks;
if ((rank + mask + recvblks == comm_size) ||
(((rank + mask) % comm_size) < ((rank + mask + recvblks) % comm_size)))
{
/* If the data contiguously fits into the
* receive buffer, place it directly. This
* should cover the case where the root is
* rank 0. */
char *rp = (char *)recvbuf + (((rank + mask) % comm_size)*recvcount*extent);
mpi_errno = MPIR_Sched_recv(rp, (recvblks * recvcount), recvtype, src, comm_ptr, s);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
mpi_errno = MPIR_Sched_barrier(s);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
}
else if (nbytes < MPIR_CVAR_GATHER_VSMALL_MSG_SIZE) {
mpi_errno = MPIR_Sched_recv(tmp_buf, (recvblks * nbytes), MPI_BYTE, src, comm_ptr, s);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
mpi_errno = MPIR_Sched_barrier(s);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
copy_offset = rank + mask;
copy_blks = recvblks;
}
else {
blocks[0] = recvcount * (comm_size - root - mask);
displs[0] = recvcount * (root + mask);
blocks[1] = (recvcount * recvblks) - blocks[0];
displs[1] = 0;
mpi_errno = MPIR_Type_indexed_impl(2, blocks, displs, recvtype, &tmp_type);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
mpi_errno = MPIR_Type_commit_impl(&tmp_type);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
mpi_errno = MPIR_Sched_recv(recvbuf, 1, tmp_type, src, comm_ptr, s);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
mpi_errno = MPIR_Sched_barrier(s);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
/* this "premature" free is safe b/c the sched holds an actual ref to keep it alive */
MPIR_Type_free_impl(&tmp_type);
}
}
else { /* Intermediate nodes store in temporary buffer */
MPI_Aint offset;
/* Estimate the amount of data that is going to come in */
recvblks = mask;
relative_src = ((src - root) < 0) ? (src - root + comm_size) : (src - root);
if (relative_src + mask > comm_size)
recvblks -= (relative_src + mask - comm_size);
if (nbytes < MPIR_CVAR_GATHER_VSMALL_MSG_SIZE)
offset = mask * nbytes;
else
offset = (mask - 1) * nbytes;
mpi_errno = MPIR_Sched_recv(((char *)tmp_buf + offset), (recvblks * nbytes),
MPI_BYTE, src, comm_ptr, s);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
mpi_errno = MPIR_Sched_barrier(s);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
curr_cnt += (recvblks * nbytes);
}
}
}
else {
dst = relative_rank ^ mask;
dst = (dst + root) % comm_size;
if (!tmp_buf_size) {
/* leaf nodes send directly from sendbuf */
mpi_errno = MPIR_Sched_send(sendbuf, sendcount, sendtype, dst, comm_ptr, s);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
mpi_errno = MPIR_Sched_barrier(s);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
}
else if (nbytes < MPIR_CVAR_GATHER_VSMALL_MSG_SIZE) {
mpi_errno = MPIR_Sched_send(tmp_buf, curr_cnt, MPI_BYTE, dst, comm_ptr, s);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
mpi_errno = MPIR_Sched_barrier(s);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
}
else {
blocks[0] = sendcount;
struct_displs[0] = MPIR_VOID_PTR_CAST_TO_MPI_AINT sendbuf;
types[0] = sendtype;
/* check for overflow. work around int limits if needed*/
if (curr_cnt - nbytes != (int)(curr_cnt-nbytes)) {
blocks[1] = 1;
MPIR_Type_contiguous_x_impl(curr_cnt - nbytes,
MPI_BYTE, &(types[1]));
} else {
MPIR_Assign_trunc(blocks[1], curr_cnt - nbytes, int);
types[1] = MPI_BYTE;
}
struct_displs[1] = MPIR_VOID_PTR_CAST_TO_MPI_AINT tmp_buf;
mpi_errno = MPIR_Type_create_struct_impl(2, blocks, struct_displs, types, &tmp_type);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
mpi_errno = MPIR_Type_commit_impl(&tmp_type);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
mpi_errno = MPIR_Sched_send(MPI_BOTTOM, 1, tmp_type, dst, comm_ptr, s);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
MPIR_SCHED_BARRIER(s);
/* this "premature" free is safe b/c the sched holds an actual ref to keep it alive */
MPIR_Type_free_impl(&tmp_type);
}
break;
}
mask <<= 1;
}
if ((rank == root) && root && (nbytes < MPIR_CVAR_GATHER_VSMALL_MSG_SIZE) && copy_blks) {
/* reorder and copy from tmp_buf into recvbuf */
/* FIXME why are there two copies here? */
mpi_errno = MPIR_Sched_copy(tmp_buf, nbytes * (comm_size - copy_offset), MPI_BYTE,
((char *)recvbuf + extent * recvcount * copy_offset),
recvcount * (comm_size - copy_offset), recvtype, s);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
mpi_errno = MPIR_Sched_copy((char *)tmp_buf + nbytes * (comm_size - copy_offset),
nbytes * (copy_blks - comm_size + copy_offset), MPI_BYTE,
recvbuf, recvcount * (copy_blks - comm_size + copy_offset),
recvtype, s);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
}
}
#ifdef MPID_HAS_HETERO
else {
int MPIR_Ialltoallv_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,
MPID_Sched_t s)
{
int mpi_errno = MPI_SUCCESS;
int comm_size;
int i, j;
int ii, ss, bblock;
MPI_Aint send_extent, recv_extent, sendtype_size, recvtype_size;
int dst, rank;
MPIR_SCHED_CHKPMEM_DECL(1);
MPIU_Assert(comm_ptr->comm_kind == MPID_INTRACOMM);
comm_size = comm_ptr->local_size;
rank = comm_ptr->rank;
/* Get extent and size of recvtype, don't look at sendtype for MPI_IN_PLACE */
MPID_Datatype_get_extent_macro(recvtype, recv_extent);
MPID_Datatype_get_size_macro(recvtype, recvtype_size);
if (sendbuf == MPI_IN_PLACE) {
int max_count;
void *tmp_buf = NULL;
/* The regular MPI_Alltoallv handles MPI_IN_PLACE using pairwise
* sendrecv_replace calls. We don't have a sendrecv_replace, so just
* malloc the maximum of the counts array entries and then perform the
* pairwise exchanges manually with schedule barriers instead.
*
* Because of this approach all processes must agree on the global
* schedule of "sendrecv_replace" operations to avoid deadlock.
*
* This keeps with the spirit of the MPI-2.2 standard, which is to
* conserve memory when using MPI_IN_PLACE for these routines.
* Something like MADRE would probably generate a more optimal
* algorithm. */
max_count = 0;
for (i = 0; i < comm_size; ++i) {
max_count = MPIU_MAX(max_count, recvcounts[i]);
}
MPIR_SCHED_CHKPMEM_MALLOC(tmp_buf, void *, max_count*recv_extent, mpi_errno, "Ialltoallv tmp_buf");
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 && rank == j) {
/* no need to "sendrecv_replace" for ourselves */
}
else if (rank == i || rank == j) {
if (rank == i)
dst = j;
else
dst = i;
mpi_errno = MPID_Sched_send(((char *)recvbuf + rdispls[dst]*recv_extent),
recvcounts[dst], recvtype, dst, comm_ptr, s);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
mpi_errno = MPID_Sched_recv(tmp_buf, recvcounts[dst], recvtype, dst, comm_ptr, s);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
MPID_SCHED_BARRIER(s);
mpi_errno = MPID_Sched_copy(tmp_buf, recvcounts[dst], recvtype,
((char *)recvbuf + rdispls[dst]*recv_extent),
recvcounts[dst], recvtype, s);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
MPID_SCHED_BARRIER(s);
}
}
}
MPID_SCHED_BARRIER(s);
}
int MPIR_Iscan_rec_dbl(const void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype, MPI_Op op, MPIR_Comm *comm_ptr, MPIR_Sched_t s)
{
int mpi_errno = MPI_SUCCESS;
MPI_Aint true_extent, true_lb, extent;
int is_commutative;
int mask, dst, rank, comm_size;
void *partial_scan = NULL;
void *tmp_buf = NULL;
MPIR_SCHED_CHKPMEM_DECL(2);
if (count == 0)
goto fn_exit;
comm_size = comm_ptr->local_size;
rank = comm_ptr->rank;
is_commutative = MPIR_Op_is_commutative(op);
/* 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);
MPIR_SCHED_CHKPMEM_MALLOC(partial_scan, void *, count*(MPL_MAX(extent,true_extent)), mpi_errno, "partial_scan");
/* This eventually gets malloc()ed as a temp buffer, not added to
* any user buffers */
MPIR_Ensure_Aint_fits_in_pointer(count * MPL_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*/
MPIR_SCHED_CHKPMEM_MALLOC(tmp_buf, void *, count*(MPL_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_Sched_copy(sendbuf, count, datatype,
recvbuf, count, datatype, s);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
}
if (sendbuf != MPI_IN_PLACE)
mpi_errno = MPIR_Sched_copy(sendbuf, count, datatype,
partial_scan, count, datatype, s);
else
mpi_errno = MPIR_Sched_copy(recvbuf, count, datatype,
partial_scan, count, datatype, s);
if (mpi_errno) MPIR_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 = MPIR_Sched_send(partial_scan, count, datatype, dst, comm_ptr, s);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
/* sendrecv, no barrier here */
mpi_errno = MPIR_Sched_recv(tmp_buf, count, datatype, dst, comm_ptr, s);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
MPIR_SCHED_BARRIER(s);
if (rank > dst) {
mpi_errno = MPIR_Sched_reduce(tmp_buf, partial_scan, count, datatype, op, s);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
mpi_errno = MPIR_Sched_reduce(tmp_buf, recvbuf, count, datatype, op, s);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
MPIR_SCHED_BARRIER(s);
}
else {
if (is_commutative) {
mpi_errno = MPIR_Sched_reduce(tmp_buf, partial_scan, count, datatype, op, s);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
MPIR_SCHED_BARRIER(s);
}
else {
mpi_errno = MPIR_Sched_reduce(partial_scan, tmp_buf, count, datatype, op, s);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
MPIR_SCHED_BARRIER(s);
mpi_errno = MPIR_Sched_copy(tmp_buf, count, datatype,
partial_scan, count, datatype, s);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
MPIR_SCHED_BARRIER(s);
}
}
}
mask <<= 1;
}
MPIR_SCHED_CHKPMEM_COMMIT(s);
fn_exit:
return mpi_errno;
fn_fail:
MPIR_SCHED_CHKPMEM_REAP(s);
goto fn_exit;
}
int MPIR_Iexscan(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 rank, comm_size;
int mask, dst, is_commutative, flag;
MPI_Aint true_extent, true_lb, extent;
void *partial_scan, *tmp_buf;
MPIR_SCHED_CHKPMEM_DECL(2);
if (count == 0)
goto fn_exit;
comm_size = comm_ptr->local_size;
rank = comm_ptr->rank;
is_commutative = MPIR_Op_is_commutative(op);
/* 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);
MPIR_SCHED_CHKPMEM_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*/
MPIR_SCHED_CHKPMEM_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 = MPID_Sched_copy((sendbuf == MPI_IN_PLACE ? recvbuf : sendbuf), count, datatype,
partial_scan, count, datatype, s);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
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 = MPID_Sched_send(partial_scan, count, datatype, dst, comm_ptr, s);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
/* sendrecv, no barrier here */
mpi_errno = MPID_Sched_recv(tmp_buf, count, datatype, dst, comm_ptr, s);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
MPID_SCHED_BARRIER(s);
if (rank > dst) {
mpi_errno = MPID_Sched_reduce(tmp_buf, partial_scan, count, datatype, op, s);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
MPID_SCHED_BARRIER(s);
/* 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 = MPID_Sched_copy(tmp_buf, count, datatype,
recvbuf, count, datatype, s);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
MPID_SCHED_BARRIER(s);
flag = 1;
}
else {
mpi_errno = MPID_Sched_reduce(tmp_buf, recvbuf, count, datatype, op, s);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
MPID_SCHED_BARRIER(s);
}
}
}
else {
if (is_commutative) {
mpi_errno = MPID_Sched_reduce(tmp_buf, partial_scan, count, datatype, op, s);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
MPID_SCHED_BARRIER(s);
}
else {
mpi_errno = MPID_Sched_reduce(partial_scan, tmp_buf, count, datatype, op, s);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
MPID_SCHED_BARRIER(s);
mpi_errno = MPID_Sched_copy(tmp_buf, count, datatype,
partial_scan, count, datatype, s);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
MPID_SCHED_BARRIER(s);
}
}
}
mask <<= 1;
}
MPIR_SCHED_CHKPMEM_COMMIT(s);
fn_exit:
return mpi_errno;
fn_fail:
MPIR_SCHED_CHKPMEM_REAP(s);
goto fn_exit;
}