static int sched_cb_gcn_copy_mask(MPID_Comm * comm, int tag, void *state)
{
int mpi_errno = MPI_SUCCESS;
struct gcn_state *st = state;
if (st->first_iter) {
memset(st->local_mask, 0, (MPIR_MAX_CONTEXT_MASK + 1) * sizeof(int));
st->own_eager_mask = 0;
/* Attempt to reserve the eager mask segment */
if (!eager_in_use && eager_nelem > 0) {
int i;
for (i = 0; i < eager_nelem; i++)
st->local_mask[i] = context_mask[i];
eager_in_use = 1;
st->own_eager_mask = 1;
}
}
else {
/* Same rules as for the blocking case */
if (mask_in_use || st != next_gcn) {
memset(st->local_mask, 0, MPIR_MAX_CONTEXT_MASK * sizeof(int));
st->own_mask = 0;
st->local_mask[ALL_OWN_MASK_FLAG] = 0;
}
else {
/* Copy safe mask segment to local_mask */
int i;
for (i = 0; i < eager_nelem; i++)
st->local_mask[i] = 0;
for (i = eager_nelem; i < MPIR_MAX_CONTEXT_MASK; i++)
st->local_mask[i] = context_mask[i];
mask_in_use = 1;
st->own_mask = 1;
st->local_mask[ALL_OWN_MASK_FLAG] = 1;
}
}
mpi_errno =
st->comm_ptr->coll_fns->Iallreduce_sched(MPI_IN_PLACE, st->local_mask,
MPIR_MAX_CONTEXT_MASK + 1, MPI_UINT32_T, MPI_BAND,
st->comm_ptr, st->s);
if (mpi_errno)
MPIR_ERR_POP(mpi_errno);
MPID_SCHED_BARRIER(st->s);
mpi_errno = MPID_Sched_cb(&sched_cb_gcn_allocate_cid, st, st->s);
if (mpi_errno)
MPIR_ERR_POP(mpi_errno);
MPID_SCHED_BARRIER(st->s);
fn_fail:
return mpi_errno;
}
int MPIR_Iallreduce_naive(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;
rank = comm_ptr->rank;
if ((sendbuf == MPI_IN_PLACE) && (rank != 0)) {
mpi_errno = MPIR_Ireduce_intra(recvbuf, NULL, count, datatype, op, 0, comm_ptr, s);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
}
else {
mpi_errno = MPIR_Ireduce_intra(sendbuf, recvbuf, count, datatype, op, 0, comm_ptr, s);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
}
MPID_SCHED_BARRIER(s);
mpi_errno = MPIR_Ibcast_intra(recvbuf, count, datatype, 0, comm_ptr, s);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
fn_exit:
return mpi_errno;
fn_fail:
goto fn_exit;
}
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);
}
}
static int sched_get_cid_nonblock(MPID_Comm * comm_ptr, MPID_Comm * newcomm,
MPIU_Context_id_t * ctx0, MPIU_Context_id_t * ctx1,
MPID_Sched_t s, MPID_Comm_kind_t gcn_cid_kind)
{
int mpi_errno = MPI_SUCCESS;
struct gcn_state *st = NULL;
MPIU_CHKPMEM_DECL(1);
if (initialize_context_mask) {
context_id_init();
}
MPIU_CHKPMEM_MALLOC(st, struct gcn_state *, sizeof(struct gcn_state), mpi_errno, "gcn_state");
st->ctx0 = ctx0;
st->ctx1 = ctx1;
if (gcn_cid_kind == MPID_INTRACOMM) {
st->comm_ptr = comm_ptr;
st->comm_ptr_inter = NULL;
}
else {
st->comm_ptr = comm_ptr->local_comm;
st->comm_ptr_inter = comm_ptr;
}
st->s = s;
st->gcn_cid_kind = gcn_cid_kind;
*(st->ctx0) = 0;
st->own_eager_mask = 0;
st->first_iter = 1;
st->new_comm = newcomm;
st->own_mask = 0;
if (eager_nelem < 0) {
/* Ensure that at least one word of deadlock-free context IDs is
* always set aside for the base protocol */
MPIU_Assert(MPIR_CVAR_CTXID_EAGER_SIZE >= 0 &&
MPIR_CVAR_CTXID_EAGER_SIZE < MPIR_MAX_CONTEXT_MASK - 1);
eager_nelem = MPIR_CVAR_CTXID_EAGER_SIZE;
}
mpi_errno = MPID_Sched_cb(&sched_cb_gcn_copy_mask, st, s);
if (mpi_errno)
MPIR_ERR_POP(mpi_errno);
MPID_SCHED_BARRIER(s);
MPIU_CHKPMEM_COMMIT();
fn_exit:
return mpi_errno;
/* --BEGIN ERROR HANDLING-- */
fn_fail:
MPIU_CHKPMEM_REAP();
goto fn_exit;
/* --END ERROR HANDLING-- */
}
static int sched_cb_gcn_bcast(MPID_Comm * comm, int tag, void *state)
{
int mpi_errno = MPI_SUCCESS;
struct gcn_state *st = state;
if (st->gcn_cid_kind == MPID_INTERCOMM) {
if (st->comm_ptr_inter->rank == 0) {
mpi_errno =
MPID_Sched_recv(st->ctx1, 1, MPIU_CONTEXT_ID_T_DATATYPE, 0, st->comm_ptr_inter,
st->s);
if (mpi_errno)
MPIR_ERR_POP(mpi_errno);
mpi_errno =
MPID_Sched_send(st->ctx0, 1, MPIU_CONTEXT_ID_T_DATATYPE, 0, st->comm_ptr_inter,
st->s);
if (mpi_errno)
MPIR_ERR_POP(mpi_errno);
MPID_SCHED_BARRIER(st->s);
}
mpi_errno = st->comm_ptr->coll_fns->Ibcast_sched(st->ctx1, 1,
MPIU_CONTEXT_ID_T_DATATYPE, 0,
st->comm_ptr, st->s);
if (mpi_errno)
MPIR_ERR_POP(mpi_errno);
MPID_SCHED_BARRIER(st->s);
}
mpi_errno = MPID_Sched_cb(&sched_cb_commit_comm, st, st->s);
if (mpi_errno)
MPIR_ERR_POP(mpi_errno);
mpi_errno = MPID_Sched_cb(&MPIR_Sched_cb_free_buf, st, st->s);
if (mpi_errno)
MPIR_ERR_POP(mpi_errno);
fn_fail:
return mpi_errno;
}
int MPIR_Ialltoall_pairwise(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 src, dst, is_pof2;
int rank, comm_size;
MPI_Aint sendtype_extent, recvtype_extent;
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_extent_macro(recvtype, recvtype_extent);
/* Make local copy first */
mpi_errno = MPID_Sched_copy(((char *)sendbuf + rank*sendcount*sendtype_extent),
sendcount, sendtype,
((char *)recvbuf + rank*recvcount*recvtype_extent),
recvcount, recvtype, s);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
is_pof2 = MPIU_is_pof2(comm_size, NULL);
/* Do the pairwise exchanges */
for (i = 1; i < comm_size; i++) {
if (is_pof2 == 1) {
/* use exclusive-or algorithm */
src = dst = rank ^ i;
}
else {
src = (rank - i + comm_size) % comm_size;
dst = (rank + i) % comm_size;
}
mpi_errno = MPID_Sched_send(((char *)sendbuf + dst*sendcount*sendtype_extent),
sendcount, sendtype, dst, comm_ptr, s);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
mpi_errno = MPID_Sched_recv(((char *)recvbuf + src*recvcount*recvtype_extent),
recvcount, recvtype, src, comm_ptr, s);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
MPID_SCHED_BARRIER(s);
}
fn_exit:
return mpi_errno;
fn_fail:
goto fn_exit;
}
int MPIR_Ineighbor_allgatherv_default(const void *sendbuf, int sendcount, MPI_Datatype sendtype, void *recvbuf, const int recvcounts[], const int displs[], MPI_Datatype recvtype, MPID_Comm *comm_ptr, MPID_Sched_t s)
{
int mpi_errno = MPI_SUCCESS;
int indegree, outdegree, weighted;
int i, k,l;
int *srcs, *dsts;
int comm_size;
MPI_Aint recvtype_extent;
MPIU_CHKLMEM_DECL(2);
comm_size = comm_ptr->local_size;
MPID_Datatype_get_extent_macro(recvtype, recvtype_extent);
for (i = 0; i < comm_size; ++i) {
MPID_Ensure_Aint_fits_in_pointer(MPI_VOID_PTR_CAST_TO_MPI_AINT recvbuf +
(displs[i] * recvtype_extent));
}
mpi_errno = MPIR_Topo_canon_nhb_count(comm_ptr, &indegree, &outdegree, &weighted);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
MPIU_CHKLMEM_MALLOC(srcs, int *, indegree*sizeof(int), mpi_errno, "srcs");
MPIU_CHKLMEM_MALLOC(dsts, int *, outdegree*sizeof(int), mpi_errno, "dsts");
mpi_errno = MPIR_Topo_canon_nhb(comm_ptr,
indegree, srcs, MPI_UNWEIGHTED,
outdegree, dsts, MPI_UNWEIGHTED);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
for (k = 0; k < outdegree; ++k) {
mpi_errno = MPID_Sched_send(sendbuf, sendcount, sendtype, dsts[k], comm_ptr, s);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
}
for (l = 0; l < indegree; ++l) {
char *rb = ((char *)recvbuf) + displs[l] * recvtype_extent;
mpi_errno = MPID_Sched_recv(rb, recvcounts[l], recvtype, srcs[l], comm_ptr, s);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
}
MPID_SCHED_BARRIER(s);
fn_exit:
MPIU_CHKLMEM_FREEALL();
return mpi_errno;
fn_fail:
goto fn_exit;
}
int MPIR_Ialltoall_perm_sr(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 rank, comm_size;
int ii, ss, bblock, dst;
MPI_Aint sendtype_extent, recvtype_extent;
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_extent_macro(recvtype, recvtype_extent);
bblock = MPIR_CVAR_ALLTOALL_THROTTLE;
if (bblock == 0) bblock = comm_size;
for (ii = 0; ii < comm_size; ii += bblock) {
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;
mpi_errno = MPID_Sched_recv(((char *)recvbuf + dst*recvcount*recvtype_extent),
recvcount, recvtype, dst, comm_ptr, s);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
}
for (i = 0; i < ss; i++) {
dst = (rank-i-ii+comm_size) % comm_size;
mpi_errno = MPID_Sched_send(((char *)sendbuf + dst*sendcount*sendtype_extent),
sendcount, sendtype, dst, comm_ptr, s);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
}
/* force the (2*ss) sends/recvs above to complete before posting additional ops */
MPID_SCHED_BARRIER(s);
}
fn_exit:
return mpi_errno;
fn_fail:
goto fn_exit;
}
int MPIR_Ineighbor_alltoallw_default(const void *sendbuf, const int sendcounts[], const MPI_Aint sdispls[], const MPI_Datatype sendtypes[], void *recvbuf, const int recvcounts[], const MPI_Aint rdispls[], const MPI_Datatype recvtypes[], MPID_Comm *comm_ptr, MPID_Sched_t s)
{
int mpi_errno = MPI_SUCCESS;
int indegree, outdegree, weighted;
int k,l;
int *srcs, *dsts;
MPIU_CHKLMEM_DECL(2);
mpi_errno = MPIR_Topo_canon_nhb_count(comm_ptr, &indegree, &outdegree, &weighted);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
MPIU_CHKLMEM_MALLOC(srcs, int *, indegree*sizeof(int), mpi_errno, "srcs");
MPIU_CHKLMEM_MALLOC(dsts, int *, outdegree*sizeof(int), mpi_errno, "dsts");
mpi_errno = MPIR_Topo_canon_nhb(comm_ptr,
indegree, srcs, MPI_UNWEIGHTED,
outdegree, dsts, MPI_UNWEIGHTED);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
for (k = 0; k < outdegree; ++k) {
char *sb;
MPID_Ensure_Aint_fits_in_pointer(MPI_VOID_PTR_CAST_TO_MPI_AINT sendbuf + sdispls[k]);
sb = ((char *)sendbuf) + sdispls[k];
mpi_errno = MPID_Sched_send(sb, sendcounts[k], sendtypes[k], dsts[k], comm_ptr, s);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
}
for (l = 0; l < indegree; ++l) {
char *rb;
MPID_Ensure_Aint_fits_in_pointer(MPI_VOID_PTR_CAST_TO_MPI_AINT recvbuf + rdispls[l]);
rb = ((char *)recvbuf) + rdispls[l];
mpi_errno = MPID_Sched_recv(rb, recvcounts[l], recvtypes[l], srcs[l], comm_ptr, s);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
}
MPID_SCHED_BARRIER(s);
fn_exit:
MPIU_CHKLMEM_FREEALL();
return mpi_errno;
fn_fail:
goto fn_exit;
}
int MPIR_Ialltoall_inter(const void *sendbuf, int sendcount, MPI_Datatype sendtype, void *recvbuf, int recvcount, MPI_Datatype recvtype, MPID_Comm *comm_ptr, MPID_Sched_t s)
{
/* Intercommunicator alltoall. We use a pairwise exchange algorithm
similar to the one used in intracommunicator alltoall for long
messages. Since the local and remote groups can be of different
sizes, we first compute the max of local_group_size,
remote_group_size. At step i, 0 <= i < max_size, each process
receives from src = (rank - i + max_size) % max_size if src <
remote_size, and sends to dst = (rank + i) % max_size if dst <
remote_size.
*/
int mpi_errno = MPI_SUCCESS;
int local_size, remote_size, max_size, i;
MPI_Aint sendtype_extent, recvtype_extent;
int src, dst, rank;
char *sendaddr, *recvaddr;
local_size = comm_ptr->local_size;
remote_size = comm_ptr->remote_size;
rank = comm_ptr->rank;
/* Get extent of send and recv types */
MPID_Datatype_get_extent_macro(sendtype, sendtype_extent);
MPID_Datatype_get_extent_macro(recvtype, recvtype_extent);
/* Do the pairwise exchanges */
max_size = MPIR_MAX(local_size, remote_size);
MPID_Ensure_Aint_fits_in_pointer(MPI_VOID_PTR_CAST_TO_MPI_AINT recvbuf +
max_size*recvcount*recvtype_extent);
MPID_Ensure_Aint_fits_in_pointer(MPI_VOID_PTR_CAST_TO_MPI_AINT sendbuf +
max_size*sendcount*sendtype_extent);
for (i = 0; i < max_size; i++) {
src = (rank - i + max_size) % max_size;
dst = (rank + i) % max_size;
if (src >= remote_size) {
src = MPI_PROC_NULL;
recvaddr = NULL;
}
else {
recvaddr = (char *)recvbuf + src*recvcount*recvtype_extent;
}
if (dst >= remote_size) {
dst = MPI_PROC_NULL;
sendaddr = NULL;
}
else {
sendaddr = (char *)sendbuf + dst*sendcount*sendtype_extent;
}
mpi_errno = MPID_Sched_send(sendaddr, sendcount, sendtype, dst, comm_ptr, s);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
mpi_errno = MPID_Sched_recv(recvaddr, recvcount, recvtype, src, comm_ptr, s);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
MPID_SCHED_BARRIER(s);
}
fn_exit:
return mpi_errno;
fn_fail:
goto fn_exit;
}
int MPIR_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_Ireduce_scatter_rec_hlv(const void *sendbuf, void *recvbuf, const int recvcounts[],
MPI_Datatype datatype, MPI_Op op,
MPID_Comm *comm_ptr, MPID_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;
MPID_Datatype_get_extent_macro(datatype, extent);
MPIR_Type_get_true_extent_impl(datatype, &true_lb, &true_extent);
MPIU_Assert(MPIR_Op_is_commutative(op));
MPIR_SCHED_CHKPMEM_MALLOC(disps, int *, comm_size * sizeof(int), mpi_errno, "disps");
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;
}
MPID_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");
/* 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");
/* 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 = MPID_Sched_copy(sendbuf, total_count, datatype,
tmp_results, total_count, datatype, s);
else
mpi_errno = MPID_Sched_copy(recvbuf, total_count, datatype,
tmp_results, total_count, datatype, s);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
MPID_SCHED_BARRIER(s);
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(tmp_results, total_count, datatype, rank+1, comm_ptr, s);
if (mpi_errno) MPIR_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_recvbuf, total_count, datatype, rank-1, comm_ptr, s);
if (mpi_errno) MPIR_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_recvbuf, tmp_results, total_count, datatype, op, s);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
MPID_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");
MPIR_SCHED_CHKPMEM_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] = 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 = MPID_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 = MPID_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);
MPID_SCHED_BARRIER(s);
}
/* tmp_recvbuf contains data received in this step.
tmp_results contains data accumulated so far */
if (recv_cnt) {
mpi_errno = MPID_Sched_reduce(((char *)tmp_recvbuf + newdisps[recv_idx]*extent),
((char *)tmp_results + newdisps[recv_idx]*extent),
recv_cnt, datatype, op, s);
MPID_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 = MPID_Sched_copy(((char *)tmp_results + disps[rank]*extent),
recvcounts[rank], datatype,
recvbuf, recvcounts[rank], datatype, s);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
MPID_SCHED_BARRIER(s);
}
}
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);
}
static int sched_cb_gcn_allocate_cid(MPID_Comm * comm, int tag, void *state)
{
int mpi_errno = MPI_SUCCESS;
struct gcn_state *st = state, *tmp;
MPIU_Context_id_t newctxid;
MPIR_Errflag_t errflag = MPIR_ERR_NONE;
if (st->own_eager_mask) {
newctxid = find_and_allocate_context_id(st->local_mask);
if (st->ctx0)
*st->ctx0 = newctxid;
if (st->ctx1)
*st->ctx1 = newctxid;
st->own_eager_mask = 0;
eager_in_use = 0;
}
else if (st->own_mask) {
newctxid = find_and_allocate_context_id(st->local_mask);
if (st->ctx0)
*st->ctx0 = newctxid;
if (st->ctx1)
*st->ctx1 = newctxid;
/* reset flag for the next try */
mask_in_use = 0;
/* If we found a ctx, remove element form list */
if (newctxid > 0) {
if (next_gcn == st) {
next_gcn = st->next;
}
else {
for (tmp = next_gcn; tmp->next != st; tmp = tmp->next);
tmp->next = st->next;
}
}
}
if (*st->ctx0 == 0) {
if (st->local_mask[ALL_OWN_MASK_FLAG] == 1) {
/* --BEGIN ERROR HANDLING-- */
int nfree = 0;
int ntotal = 0;
int minfree;
context_mask_stats(&nfree, &ntotal);
minfree = nfree;
MPIR_Allreduce_impl(MPI_IN_PLACE, &minfree, 1, MPI_INT,
MPI_MIN, st->comm_ptr, &errflag);
if (minfree > 0) {
MPIR_ERR_SETANDJUMP3(mpi_errno, MPI_ERR_OTHER,
"**toomanycommfrag", "**toomanycommfrag %d %d %d",
nfree, ntotal, minfree);
}
else {
MPIR_ERR_SETANDJUMP3(mpi_errno, MPI_ERR_OTHER,
"**toomanycomm", "**toomanycomm %d %d %d",
nfree, ntotal, minfree);
}
/* --END ERROR HANDLING-- */
}
else {
/* do not own mask, try again */
if (st->first_iter == 1) {
st->first_iter = 0;
/* Set the Tag for the idup-operations. We have two problems here:
* 1.) The tag should not be used by another (blocking) context_id allocation.
* Therefore, we set tag_up as lower bound for the operation. tag_ub is used by
* most of the other blocking operations, but tag is always >0, so this
* should be fine.
* 2.) We need odering between multiple idup operations on the same communicator.
* The problem here is that the iallreduce operations of the first iteration
* are not necessarily completed in the same order as they are issued, also on the
* same communicator. To avoid deadlocks, we cannot add the elements to the
* list bevfore the first iallreduce is completed. The "tag" is created for the
* scheduling - by calling MPID_Sched_next_tag(comm_ptr, &tag) - and the same
* for a idup operation on all processes. So we use it here. */
/* FIXME I'm not sure if there can be an overflows for this tag */
st->tag = (uint64_t) tag + MPIR_Process.attrs.tag_ub;
add_gcn_to_list(st);
}
mpi_errno = MPID_Sched_cb(&sched_cb_gcn_copy_mask, st, st->s);
if (mpi_errno)
MPIR_ERR_POP(mpi_errno);
MPID_SCHED_BARRIER(st->s);
}
}
else {
/* Successfully allocated a context id */
mpi_errno = MPID_Sched_cb(&sched_cb_gcn_bcast, st, st->s);
if (mpi_errno)
MPIR_ERR_POP(mpi_errno);
MPID_SCHED_BARRIER(st->s);
}
fn_exit:
return mpi_errno;
fn_fail:
/* make sure that the pending allocations are scheduled */
if (!st->first_iter) {
if (next_gcn == st) {
next_gcn = st->next;
}
else {
for (tmp = next_gcn; tmp && tmp->next != st; tmp = tmp->next);
tmp->next = st->next;
}
}
/* In the case of failure, the new communicator was half created.
* So we need to clean the memory allocated for it. */
MPIR_Comm_map_free(st->new_comm);
MPIU_Handle_obj_free(&MPID_Comm_mem, st->new_comm);
MPIU_Free(st);
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_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_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;
}
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_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++;
}
