int MPID_nem_ptl_init_id(MPIDI_VC_t *vc)
{
int mpi_errno = MPI_SUCCESS;
MPID_nem_ptl_vc_area *const vc_ptl = VC_PTL(vc);
char *bc;
int pmi_errno;
int val_max_sz;
MPIU_CHKLMEM_DECL(1);
MPIDI_STATE_DECL(MPID_STATE_MPID_NEM_PTL_INIT_ID);
MPIDI_FUNC_ENTER(MPID_STATE_MPID_NEM_PTL_INIT_ID);
pmi_errno = PMI_KVS_Get_value_length_max(&val_max_sz);
MPIU_ERR_CHKANDJUMP1(pmi_errno, mpi_errno, MPI_ERR_OTHER, "**fail", "**fail %d", pmi_errno);
MPIU_CHKLMEM_MALLOC(bc, char *, val_max_sz, mpi_errno, "bc");
mpi_errno = vc->pg->getConnInfo(vc->pg_rank, bc, val_max_sz, vc->pg);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
mpi_errno = MPID_nem_ptl_get_id_from_bc(bc, &vc_ptl->id, &vc_ptl->pt, &vc_ptl->ptg, &vc_ptl->ptc, &vc_ptl->ptr, &vc_ptl->ptrg, &vc_ptl->ptrc);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
vc_ptl->id_initialized = TRUE;
MPIDI_CHANGE_VC_STATE(vc, ACTIVE);
fn_exit:
MPIU_CHKLMEM_FREEALL();
MPIDI_FUNC_EXIT(MPID_STATE_MPID_NEM_PTL_INIT_ID);
return mpi_errno;
fn_fail:
goto fn_exit;
}
static int get_addr(MPIDI_VC_t * vc, struct scif_portID *addr)
{
int mpi_errno = MPI_SUCCESS;
char *bc;
int pmi_errno;
int val_max_sz;
MPIU_CHKLMEM_DECL(1);
/* Allocate space for the business card */
pmi_errno = PMI_KVS_Get_value_length_max(&val_max_sz);
MPIU_ERR_CHKANDJUMP1(pmi_errno, mpi_errno, MPI_ERR_OTHER, "**fail",
"**fail %d", pmi_errno);
MPIU_CHKLMEM_MALLOC(bc, char *, val_max_sz, mpi_errno, "bc");
mpi_errno = vc->pg->getConnInfo(vc->pg_rank, bc, val_max_sz, vc->pg);
if (mpi_errno)
MPIU_ERR_POP(mpi_errno);
mpi_errno = scif_addr_from_bc(bc, &addr->node, &addr->port);
if (mpi_errno)
MPIU_ERR_POP(mpi_errno);
fn_exit:
MPIU_CHKLMEM_FREEALL();
return mpi_errno;
fn_fail:
goto fn_exit;
}
int MPIR_Type_indexed_impl(int count, const int *array_of_blocklengths,
const int *array_of_displacements,
MPI_Datatype oldtype, MPI_Datatype *newtype)
{
int mpi_errno = MPI_SUCCESS;
MPI_Datatype new_handle;
MPID_Datatype *new_dtp;
int i, *ints;
MPIU_CHKLMEM_DECL(1);
mpi_errno = MPID_Type_indexed(count,
array_of_blocklengths,
array_of_displacements,
0, /* displacements not in bytes */
oldtype,
&new_handle);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
/* copy all integer values into a temporary buffer; this
* includes the count, the blocklengths, and the displacements.
*/
MPIU_CHKLMEM_MALLOC(ints, int *, (2 * count + 1) * sizeof(int), mpi_errno, "contents integer array");
ints[0] = count;
for (i=0; i < count; i++) {
ints[i+1] = array_of_blocklengths[i];
}
for (i=0; i < count; i++) {
ints[i + count + 1] = array_of_displacements[i];
}
MPID_Datatype_get_ptr(new_handle, new_dtp);
mpi_errno = MPID_Datatype_set_contents(new_dtp,
MPI_COMBINER_INDEXED,
2*count + 1, /* ints */
0, /* aints */
1, /* types */
ints,
NULL,
&oldtype);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
MPIU_OBJ_PUBLISH_HANDLE(*newtype, new_handle);
fn_exit:
MPIU_CHKLMEM_FREEALL();
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_Type_create_struct_impl(int count,
int array_of_blocklengths[],
MPI_Aint array_of_displacements[],
MPI_Datatype array_of_types[],
MPI_Datatype *newtype)
{
int mpi_errno = MPI_SUCCESS;
int i, *ints;
MPI_Datatype new_handle;
MPID_Datatype *new_dtp;
MPIU_CHKLMEM_DECL(1);
mpi_errno = MPID_Type_struct(count,
array_of_blocklengths,
array_of_displacements,
array_of_types,
&new_handle);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
MPIU_CHKLMEM_MALLOC_ORJUMP(ints, int *, (count + 1) * sizeof(int), mpi_errno, "content description");
ints[0] = count;
for (i=0; i < count; i++)
ints[i+1] = array_of_blocklengths[i];
MPID_Datatype_get_ptr(new_handle, new_dtp);
mpi_errno = MPID_Datatype_set_contents(new_dtp,
MPI_COMBINER_STRUCT,
count+1, /* ints (cnt,blklen) */
count, /* aints (disps) */
count, /* types */
ints,
array_of_displacements,
array_of_types);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
MPIU_OBJ_PUBLISH_HANDLE(*newtype, new_handle);
fn_exit:
MPIU_CHKLMEM_FREEALL();
return mpi_errno;
fn_fail:
goto fn_exit;
}
/* this function is not used in pmi2 */
static int publish_node_id(MPIDI_PG_t *pg, int our_pg_rank)
{
int mpi_errno = MPI_SUCCESS;
int pmi_errno;
int ret;
char *key;
int key_max_sz;
char *kvs_name;
MPIU_CHKLMEM_DECL(1);
/* set MPIU_hostname */
ret = gethostname(MPIU_hostname, MAX_HOSTNAME_LEN);
MPIR_ERR_CHKANDJUMP2(ret == -1, mpi_errno, MPI_ERR_OTHER, "**sock_gethost", "**sock_gethost %s %d", MPIU_Strerror(errno), errno);
MPIU_hostname[MAX_HOSTNAME_LEN-1] = '\0';
/* Allocate space for pmi key */
pmi_errno = PMI_KVS_Get_key_length_max(&key_max_sz);
MPIR_ERR_CHKANDJUMP1(pmi_errno, mpi_errno, MPI_ERR_OTHER, "**fail", "**fail %d", pmi_errno);
MPIU_CHKLMEM_MALLOC(key, char *, key_max_sz, mpi_errno, "key");
mpi_errno = MPIDI_PG_GetConnKVSname(&kvs_name);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
/* Put my hostname id */
if (pg->size > 1)
{
memset(key, 0, key_max_sz);
MPL_snprintf(key, key_max_sz, "hostname[%d]", our_pg_rank);
pmi_errno = PMI_KVS_Put(kvs_name, key, MPIU_hostname);
MPIR_ERR_CHKANDJUMP1(pmi_errno != PMI_SUCCESS, mpi_errno, MPI_ERR_OTHER, "**pmi_kvs_put", "**pmi_kvs_put %d", pmi_errno);
pmi_errno = PMI_KVS_Commit(kvs_name);
MPIR_ERR_CHKANDJUMP1(pmi_errno != PMI_SUCCESS, mpi_errno, MPI_ERR_OTHER, "**pmi_kvs_commit", "**pmi_kvs_commit %d", pmi_errno);
pmi_errno = PMI_Barrier();
MPIR_ERR_CHKANDJUMP1(pmi_errno != PMI_SUCCESS, mpi_errno, MPI_ERR_OTHER, "**pmi_barrier", "**pmi_barrier %d", pmi_errno);
}
fn_exit:
MPIU_CHKLMEM_FREEALL();
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_Allreduce_intra (
void *sendbuf,
void *recvbuf,
int count,
MPI_Datatype datatype,
MPI_Op op,
MPID_Comm *comm_ptr,
int *errflag )
{
int is_homogeneous;
#ifdef MPID_HAS_HETERO
int rc;
#endif
int comm_size, rank, type_size;
int mpi_errno = MPI_SUCCESS;
int mpi_errno_ret = MPI_SUCCESS;
int mask, dst, is_commutative, pof2, newrank, rem, newdst, i,
send_idx, recv_idx, last_idx, send_cnt, recv_cnt, *cnts, *disps;
MPI_Aint true_extent, true_lb, extent;
void *tmp_buf;
MPI_Comm comm;
MPIU_CHKLMEM_DECL(3);
/* check if multiple threads are calling this collective function */
MPIDU_ERR_CHECK_MULTIPLE_THREADS_ENTER( comm_ptr );
if (count == 0) goto fn_exit;
comm = comm_ptr->handle;
is_commutative = MPIR_Op_is_commutative(op);
#if defined(USE_SMP_COLLECTIVES)
/* is the op commutative? We do SMP optimizations only if it is. */
if (MPIR_Comm_is_node_aware(comm_ptr) && is_commutative) {
/* on each node, do a reduce to the local root */
if (comm_ptr->node_comm != NULL) {
/* take care of the MPI_IN_PLACE case. For reduce,
MPI_IN_PLACE is specified only on the root;
for allreduce it is specified on all processes. */
if ((sendbuf == MPI_IN_PLACE) && (comm_ptr->node_comm->rank != 0)) {
/* IN_PLACE and not root of reduce. Data supplied to this
allreduce is in recvbuf. Pass that as the sendbuf to reduce. */
mpi_errno = MPIR_Reduce_impl(recvbuf, NULL, count, datatype, op, 0, comm_ptr->node_comm, errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag = TRUE;
MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
}
} else {
mpi_errno = MPIR_Reduce_impl(sendbuf, recvbuf, count, datatype, op, 0, comm_ptr->node_comm, errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag = TRUE;
MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
}
}
} else {
/* only one process on the node. copy sendbuf to recvbuf */
if (sendbuf != MPI_IN_PLACE) {
mpi_errno = MPIR_Localcopy(sendbuf, count, datatype, recvbuf, count, datatype);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
}
}
/* now do an IN_PLACE allreduce among the local roots of all nodes */
if (comm_ptr->node_roots_comm != NULL) {
mpi_errno = allreduce_intra_or_coll_fn(MPI_IN_PLACE, recvbuf, count, datatype, op, comm_ptr->node_roots_comm,
errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag = TRUE;
MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
}
}
/* now broadcast the result among local processes */
if (comm_ptr->node_comm != NULL) {
mpi_errno = MPIR_Bcast_impl(recvbuf, count, datatype, 0, comm_ptr->node_comm, errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag = TRUE;
MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
}
}
goto fn_exit;
}
#endif
is_homogeneous = 1;
#ifdef MPID_HAS_HETERO
if (comm_ptr->is_hetero)
is_homogeneous = 0;
#endif
#ifdef MPID_HAS_HETERO
if (!is_homogeneous) {
/* heterogeneous. To get the same result on all processes, we
do a reduce to 0 and then broadcast. */
mpi_errno = MPIR_Reduce_impl ( sendbuf, recvbuf, count, datatype,
op, 0, comm_ptr, errflag );
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag = TRUE;
MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
}
mpi_errno = MPIR_Bcast_impl( recvbuf, count, datatype, 0, comm_ptr, errflag );
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag = TRUE;
MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
}
}
else
#endif /* MPID_HAS_HETERO */
{
/* homogeneous */
comm_size = comm_ptr->local_size;
rank = comm_ptr->rank;
is_commutative = MPIR_Op_is_commutative(op);
/* need to allocate temporary buffer to store incoming data*/
MPIR_Type_get_true_extent_impl(datatype, &true_lb, &true_extent);
MPID_Datatype_get_extent_macro(datatype, extent);
MPID_Ensure_Aint_fits_in_pointer(count * MPIR_MAX(extent, true_extent));
MPIU_CHKLMEM_MALLOC(tmp_buf, void *, count*(MPIR_MAX(extent,true_extent)), mpi_errno, "temporary buffer");
/* adjust for potential negative lower bound in datatype */
tmp_buf = (void *)((char*)tmp_buf - true_lb);
/* copy local data into recvbuf */
if (sendbuf != MPI_IN_PLACE) {
mpi_errno = MPIR_Localcopy(sendbuf, count, datatype, recvbuf,
count, datatype);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
}
MPID_Datatype_get_size_macro(datatype, type_size);
/* find nearest power-of-two less than or equal to comm_size */
pof2 = 1;
while (pof2 <= comm_size) pof2 <<= 1;
pof2 >>=1;
rem = comm_size - pof2;
/* In the non-power-of-two case, all even-numbered
processes of rank < 2*rem send their data to
(rank+1). These even-numbered processes no longer
participate in the algorithm until the very end. The
remaining processes form a nice power-of-two. */
if (rank < 2*rem) {
if (rank % 2 == 0) { /* even */
mpi_errno = MPIC_Send_ft(recvbuf, count,
datatype, rank+1,
MPIR_ALLREDUCE_TAG, comm, errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag = TRUE;
MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
}
/* temporarily set the rank to -1 so that this
process does not pariticipate in recursive
doubling */
newrank = -1;
}
else { /* odd */
mpi_errno = MPIC_Recv_ft(tmp_buf, count,
datatype, rank-1,
MPIR_ALLREDUCE_TAG, comm,
MPI_STATUS_IGNORE, errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag = TRUE;
MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
}
/* do the reduction on received data. since the
ordering is right, it doesn't matter whether
the operation is commutative or not. */
mpi_errno = MPIR_Reduce_local_impl(tmp_buf, recvbuf, count, datatype, op);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
/* change the rank */
newrank = rank / 2;
}
}
else /* rank >= 2*rem */
newrank = rank - rem;
/* If op is user-defined or count is less than pof2, use
recursive doubling algorithm. Otherwise do a reduce-scatter
followed by allgather. (If op is user-defined,
derived datatypes are allowed and the user could pass basic
datatypes on one process and derived on another as long as
the type maps are the same. Breaking up derived
datatypes to do the reduce-scatter is tricky, therefore
using recursive doubling in that case.) */
if (newrank != -1) {
if ((count*type_size <= MPIR_PARAM_ALLREDUCE_SHORT_MSG_SIZE) ||
(HANDLE_GET_KIND(op) != HANDLE_KIND_BUILTIN) ||
(count < pof2)) { /* use recursive doubling */
mask = 0x1;
while (mask < pof2) {
newdst = newrank ^ mask;
/* find real rank of dest */
dst = (newdst < rem) ? newdst*2 + 1 : newdst + rem;
/* Send the most current data, which is in recvbuf. Recv
into tmp_buf */
mpi_errno = MPIC_Sendrecv_ft(recvbuf, count, datatype,
dst, MPIR_ALLREDUCE_TAG, tmp_buf,
count, datatype, dst,
MPIR_ALLREDUCE_TAG, comm,
MPI_STATUS_IGNORE, errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag = TRUE;
MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
}
/* tmp_buf contains data received in this step.
recvbuf contains data accumulated so far */
if (is_commutative || (dst < rank)) {
/* op is commutative OR the order is already right */
mpi_errno = MPIR_Reduce_local_impl(tmp_buf, recvbuf, count, datatype, op);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
}
else {
/* op is noncommutative and the order is not right */
mpi_errno = MPIR_Reduce_local_impl(recvbuf, tmp_buf, count, datatype, op);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
/* copy result back into recvbuf */
mpi_errno = MPIR_Localcopy(tmp_buf, count, datatype,
recvbuf, count, datatype);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
}
mask <<= 1;
}
}
else {
/* do a reduce-scatter followed by allgather */
/* for the reduce-scatter, calculate the count that
each process receives and the displacement within
the buffer */
MPIU_CHKLMEM_MALLOC(cnts, int *, pof2*sizeof(int), mpi_errno, "counts");
MPIU_CHKLMEM_MALLOC(disps, int *, pof2*sizeof(int), mpi_errno, "displacements");
for (i=0; i<(pof2-1); i++)
cnts[i] = count/pof2;
cnts[pof2-1] = count - (count/pof2)*(pof2-1);
disps[0] = 0;
for (i=1; i<pof2; i++)
disps[i] = disps[i-1] + cnts[i-1];
mask = 0x1;
send_idx = recv_idx = 0;
last_idx = pof2;
while (mask < pof2) {
newdst = newrank ^ mask;
/* find real rank of dest */
dst = (newdst < rem) ? newdst*2 + 1 : newdst + rem;
send_cnt = recv_cnt = 0;
if (newrank < newdst) {
send_idx = recv_idx + pof2/(mask*2);
for (i=send_idx; i<last_idx; i++)
send_cnt += cnts[i];
for (i=recv_idx; i<send_idx; i++)
recv_cnt += cnts[i];
}
else {
recv_idx = send_idx + pof2/(mask*2);
for (i=send_idx; i<recv_idx; i++)
send_cnt += cnts[i];
for (i=recv_idx; i<last_idx; i++)
recv_cnt += cnts[i];
}
/* printf("Rank %d, send_idx %d, recv_idx %d, send_cnt %d, recv_cnt %d, last_idx %d\n", newrank, send_idx, recv_idx,
send_cnt, recv_cnt, last_idx);
*/
/* Send data from recvbuf. Recv into tmp_buf */
mpi_errno = MPIC_Sendrecv_ft((char *) recvbuf +
disps[send_idx]*extent,
send_cnt, datatype,
dst, MPIR_ALLREDUCE_TAG,
(char *) tmp_buf +
disps[recv_idx]*extent,
recv_cnt, datatype, dst,
MPIR_ALLREDUCE_TAG, comm,
MPI_STATUS_IGNORE, errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag = TRUE;
MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
}
/* tmp_buf contains data received in this step.
recvbuf contains data accumulated so far */
/* This algorithm is used only for predefined ops
and predefined ops are always commutative. */
mpi_errno = MPIR_Reduce_local_impl(((char *) tmp_buf + disps[recv_idx]*extent),
((char *) recvbuf + disps[recv_idx]*extent),
recv_cnt, datatype, op);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
/* update send_idx for next iteration */
send_idx = recv_idx;
mask <<= 1;
/* update last_idx, but not in last iteration
because the value is needed in the allgather
step below. */
if (mask < pof2)
last_idx = recv_idx + pof2/mask;
}
/* now do the allgather */
mask >>= 1;
while (mask > 0) {
newdst = newrank ^ mask;
/* find real rank of dest */
dst = (newdst < rem) ? newdst*2 + 1 : newdst + rem;
send_cnt = recv_cnt = 0;
if (newrank < newdst) {
/* update last_idx except on first iteration */
if (mask != pof2/2)
last_idx = last_idx + pof2/(mask*2);
recv_idx = send_idx + pof2/(mask*2);
for (i=send_idx; i<recv_idx; i++)
send_cnt += cnts[i];
for (i=recv_idx; i<last_idx; i++)
recv_cnt += cnts[i];
}
else {
recv_idx = send_idx - pof2/(mask*2);
for (i=send_idx; i<last_idx; i++)
send_cnt += cnts[i];
for (i=recv_idx; i<send_idx; i++)
recv_cnt += cnts[i];
}
mpi_errno = MPIC_Sendrecv_ft((char *) recvbuf +
disps[send_idx]*extent,
send_cnt, datatype,
dst, MPIR_ALLREDUCE_TAG,
(char *) recvbuf +
disps[recv_idx]*extent,
recv_cnt, datatype, dst,
MPIR_ALLREDUCE_TAG, comm,
MPI_STATUS_IGNORE, errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag = TRUE;
MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
}
if (newrank > newdst) send_idx = recv_idx;
mask >>= 1;
}
}
}
/* In the non-power-of-two case, all odd-numbered
processes of rank < 2*rem send the result to
(rank-1), the ranks who didn't participate above. */
if (rank < 2*rem) {
if (rank % 2) /* odd */
mpi_errno = MPIC_Send_ft(recvbuf, count,
datatype, rank-1,
MPIR_ALLREDUCE_TAG, comm, errflag);
else /* even */
mpi_errno = MPIC_Recv_ft(recvbuf, count,
datatype, rank+1,
MPIR_ALLREDUCE_TAG, comm,
MPI_STATUS_IGNORE, errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag = TRUE;
MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
}
}
}
fn_exit:
/* check if multiple threads are calling this collective function */
MPIDU_ERR_CHECK_MULTIPLE_THREADS_EXIT( comm_ptr );
MPIU_CHKLMEM_FREEALL();
if (mpi_errno_ret)
mpi_errno = mpi_errno_ret;
return (mpi_errno);
fn_fail:
goto fn_exit;
}
int MPIR_Alltoallv_intra(const void *sendbuf, const int *sendcounts, const int *sdispls,
MPI_Datatype sendtype, void *recvbuf, const int *recvcounts,
const int *rdispls, MPI_Datatype recvtype, MPID_Comm *comm_ptr,
MPIR_Errflag_t *errflag)
{
int comm_size, i, j;
MPI_Aint send_extent, recv_extent;
int mpi_errno = MPI_SUCCESS;
int mpi_errno_ret = MPI_SUCCESS;
MPI_Status *starray;
MPI_Status status;
MPID_Request **reqarray;
int dst, rank, req_cnt;
int ii, ss, bblock;
int type_size;
MPIU_CHKLMEM_DECL(2);
comm_size = comm_ptr->local_size;
rank = comm_ptr->rank;
/* Get extent of recv type, but send type is only valid if (sendbuf!=MPI_IN_PLACE) */
MPID_Datatype_get_extent_macro(recvtype, recv_extent);
/* check if multiple threads are calling this collective function */
MPIDU_ERR_CHECK_MULTIPLE_THREADS_ENTER( comm_ptr );
if (sendbuf == MPI_IN_PLACE) {
/* We use pair-wise sendrecv_replace in order to conserve memory usage,
* which is keeping with the spirit of the MPI-2.2 Standard. But
* because of this approach all processes must agree on the global
* schedule of sendrecv_replace operations to avoid deadlock.
*
* Note that this is not an especially efficient algorithm in terms of
* time and there will be multiple repeated malloc/free's rather than
* maintaining a single buffer across the whole loop. Something like
* MADRE is probably the best solution for the MPI_IN_PLACE scenario. */
for (i = 0; i < comm_size; ++i) {
/* start inner loop at i to avoid re-exchanging data */
for (j = i; j < comm_size; ++j) {
if (rank == i) {
/* also covers the (rank == i && rank == j) case */
mpi_errno = MPIC_Sendrecv_replace(((char *)recvbuf + rdispls[j]*recv_extent),
recvcounts[j], recvtype,
j, MPIR_ALLTOALLV_TAG,
j, MPIR_ALLTOALLV_TAG,
comm_ptr, &status, errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag = MPIR_ERR_GET_CLASS(mpi_errno);
MPIR_ERR_SET(mpi_errno, *errflag, "**fail");
MPIR_ERR_ADD(mpi_errno_ret, mpi_errno);
}
}
else if (rank == j) {
/* same as above with i/j args reversed */
mpi_errno = MPIC_Sendrecv_replace(((char *)recvbuf + rdispls[i]*recv_extent),
recvcounts[i], recvtype,
i, MPIR_ALLTOALLV_TAG,
i, MPIR_ALLTOALLV_TAG,
comm_ptr, &status, errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag = MPIR_ERR_GET_CLASS(mpi_errno);
MPIR_ERR_SET(mpi_errno, *errflag, "**fail");
MPIR_ERR_ADD(mpi_errno_ret, mpi_errno);
}
}
}
}
}
else {
bblock = MPIR_CVAR_ALLTOALL_THROTTLE;
if (bblock == 0) bblock = comm_size;
MPID_Datatype_get_extent_macro(sendtype, send_extent);
MPIU_CHKLMEM_MALLOC(starray, MPI_Status*, 2*bblock*sizeof(MPI_Status), mpi_errno, "starray");
MPIU_CHKLMEM_MALLOC(reqarray, MPID_Request**, 2*bblock*sizeof(MPID_Request *), mpi_errno, "reqarray");
/* post only bblock isends/irecvs at a time as suggested by Tony Ladd */
for (ii=0; ii<comm_size; ii+=bblock) {
req_cnt = 0;
ss = comm_size-ii < bblock ? comm_size-ii : bblock;
/* do the communication -- post ss sends and receives: */
for ( i=0; i<ss; i++ ) {
dst = (rank+i+ii) % comm_size;
if (recvcounts[dst]) {
MPID_Datatype_get_size_macro(recvtype, type_size);
if (type_size) {
MPIU_Ensure_Aint_fits_in_pointer(MPIU_VOID_PTR_CAST_TO_MPI_AINT recvbuf +
rdispls[dst]*recv_extent);
mpi_errno = MPIC_Irecv((char *)recvbuf+rdispls[dst]*recv_extent,
recvcounts[dst], recvtype, dst,
MPIR_ALLTOALLV_TAG, comm_ptr,
&reqarray[req_cnt]);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag = MPIR_ERR_GET_CLASS(mpi_errno);
MPIR_ERR_SET(mpi_errno, *errflag, "**fail");
MPIR_ERR_ADD(mpi_errno_ret, mpi_errno);
}
req_cnt++;
}
}
}
for ( i=0; i<ss; i++ ) {
dst = (rank-i-ii+comm_size) % comm_size;
if (sendcounts[dst]) {
MPID_Datatype_get_size_macro(sendtype, type_size);
if (type_size) {
MPIU_Ensure_Aint_fits_in_pointer(MPIU_VOID_PTR_CAST_TO_MPI_AINT sendbuf +
sdispls[dst]*send_extent);
mpi_errno = MPIC_Isend((char *)sendbuf+sdispls[dst]*send_extent,
sendcounts[dst], sendtype, dst,
MPIR_ALLTOALLV_TAG, comm_ptr,
&reqarray[req_cnt], errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag = MPIR_ERR_GET_CLASS(mpi_errno);
MPIR_ERR_SET(mpi_errno, *errflag, "**fail");
MPIR_ERR_ADD(mpi_errno_ret, mpi_errno);
}
req_cnt++;
}
}
}
mpi_errno = MPIC_Waitall(req_cnt, reqarray, starray, errflag);
if (mpi_errno && mpi_errno != MPI_ERR_IN_STATUS) MPIR_ERR_POP(mpi_errno);
/* --BEGIN ERROR HANDLING-- */
if (mpi_errno == MPI_ERR_IN_STATUS) {
for (i=0; i<req_cnt; i++) {
if (starray[i].MPI_ERROR != MPI_SUCCESS) {
mpi_errno = starray[i].MPI_ERROR;
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag = MPIR_ERR_GET_CLASS(mpi_errno);
MPIR_ERR_SET(mpi_errno, *errflag, "**fail");
MPIR_ERR_ADD(mpi_errno_ret, mpi_errno);
}
}
}
}
/* --END ERROR HANDLING-- */
}
}
fn_exit:
/* check if multiple threads are calling this collective function */
MPIDU_ERR_CHECK_MULTIPLE_THREADS_EXIT( comm_ptr );
MPIU_CHKLMEM_FREEALL();
if (mpi_errno_ret)
mpi_errno = mpi_errno_ret;
else if (*errflag != MPIR_ERR_NONE)
MPIR_ERR_SET(mpi_errno, *errflag, "**coll_fail");
return mpi_errno;
fn_fail:
goto fn_exit;
}
/*@
MPI_Dist_graph_create - MPI_DIST_GRAPH_CREATE returns a handle to a new
communicator to which the distributed graph topology information is
attached.
Input Parameters:
+ comm_old - input communicator (handle)
. n - number of source nodes for which this process specifies edges
(non-negative integer)
. sources - array containing the n source nodes for which this process
specifies edges (array of non-negative integers)
. degrees - array specifying the number of destinations for each source node
in the source node array (array of non-negative integers)
. destinations - destination nodes for the source nodes in the source node
array (array of non-negative integers)
. weights - weights for source to destination edges (array of non-negative
integers or MPI_UNWEIGHTED)
. info - hints on optimization and interpretation of weights (handle)
- reorder - the process may be reordered (true) or not (false) (logical)
Output Parameters:
. comm_dist_graph - communicator with distributed graph topology added (handle)
.N ThreadSafe
.N Fortran
.N Errors
.N MPI_SUCCESS
.N MPI_ERR_ARG
.N MPI_ERR_OTHER
@*/
int MPI_Dist_graph_create(MPI_Comm comm_old, int n, const int sources[],
const int degrees[], const int destinations[],
const int weights[],
MPI_Info info, int reorder, MPI_Comm *comm_dist_graph)
{
int mpi_errno = MPI_SUCCESS;
MPID_Comm *comm_ptr = NULL;
MPID_Comm *comm_dist_graph_ptr = NULL;
MPI_Request *reqs = NULL;
MPIR_Topology *topo_ptr = NULL;
MPIR_Dist_graph_topology *dist_graph_ptr = NULL;
int i;
int j;
int idx;
int comm_size = 0;
int in_capacity;
int out_capacity;
int **rout = NULL;
int **rin = NULL;
int *rin_sizes;
int *rout_sizes;
int *rin_idx;
int *rout_idx;
int *rs;
int in_out_peers[2] = {-1, -1};
int errflag = FALSE;
MPIU_CHKLMEM_DECL(9);
MPIU_CHKPMEM_DECL(1);
MPID_MPI_STATE_DECL(MPID_STATE_MPI_DIST_GRAPH_CREATE);
MPIR_ERRTEST_INITIALIZED_ORDIE();
MPIU_THREAD_CS_ENTER(ALLFUNC,);
MPID_MPI_FUNC_ENTER(MPID_STATE_MPI_DIST_GRAPH_CREATE);
/* Validate parameters, especially handles needing to be converted */
# ifdef HAVE_ERROR_CHECKING
{
MPID_BEGIN_ERROR_CHECKS;
{
MPIR_ERRTEST_COMM(comm_old, mpi_errno);
MPIR_ERRTEST_INFO_OR_NULL(info, mpi_errno);
if (mpi_errno != MPI_SUCCESS) goto fn_fail;
}
MPID_END_ERROR_CHECKS;
}
# endif
/* Convert MPI object handles to object pointers */
MPID_Comm_get_ptr(comm_old, comm_ptr);
/* Validate parameters and objects (post conversion) */
# ifdef HAVE_ERROR_CHECKING
{
MPID_BEGIN_ERROR_CHECKS;
{
/* Validate comm_ptr */
MPID_Comm_valid_ptr(comm_ptr, mpi_errno);
/* If comm_ptr is not valid, it will be reset to null */
if (comm_ptr) {
MPIR_ERRTEST_COMM_INTRA(comm_ptr, mpi_errno);
}
MPIR_ERRTEST_ARGNEG(n, "n", mpi_errno);
if (n > 0) {
int have_degrees = 0;
MPIR_ERRTEST_ARGNULL(sources, "sources", mpi_errno);
MPIR_ERRTEST_ARGNULL(degrees, "degrees", mpi_errno);
for (i = 0; i < n; ++i) {
if (degrees[i]) {
have_degrees = 1;
break;
}
}
if (have_degrees) {
MPIR_ERRTEST_ARGNULL(destinations, "destinations", mpi_errno);
if (weights != MPI_UNWEIGHTED)
MPIR_ERRTEST_ARGNULL(weights, "weights", mpi_errno);
}
}
if (mpi_errno != MPI_SUCCESS) goto fn_fail;
}
MPID_END_ERROR_CHECKS;
}
# endif /* HAVE_ERROR_CHECKING */
/* ... body of routine ... */
/* Implementation based on Torsten Hoefler's reference implementation
* attached to MPI-2.2 ticket #33. */
*comm_dist_graph = MPI_COMM_NULL;
comm_size = comm_ptr->local_size;
/* following the spirit of the old topo interface, attributes do not
* propagate to the new communicator (see MPI-2.1 pp. 243 line 11) */
mpi_errno = MPIR_Comm_copy(comm_ptr, comm_size, &comm_dist_graph_ptr);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
MPIU_Assert(comm_dist_graph_ptr != NULL);
/* rin is an array of size comm_size containing pointers to arrays of
* rin_sizes[x]. rin[x] is locally known number of edges into this process
* from rank x.
*
* rout is an array of comm_size containing pointers to arrays of
* rout_sizes[x]. rout[x] is the locally known number of edges out of this
* process to rank x. */
MPIU_CHKLMEM_MALLOC(rout, int **, comm_size*sizeof(int*), mpi_errno, "rout");
MPIU_CHKLMEM_MALLOC(rin, int **, comm_size*sizeof(int*), mpi_errno, "rin");
MPIU_CHKLMEM_MALLOC(rin_sizes, int *, comm_size*sizeof(int), mpi_errno, "rin_sizes");
MPIU_CHKLMEM_MALLOC(rout_sizes, int *, comm_size*sizeof(int), mpi_errno, "rout_sizes");
MPIU_CHKLMEM_MALLOC(rin_idx, int *, comm_size*sizeof(int), mpi_errno, "rin_idx");
MPIU_CHKLMEM_MALLOC(rout_idx, int *, comm_size*sizeof(int), mpi_errno, "rout_idx");
memset(rout, 0, comm_size*sizeof(int*));
memset(rin, 0, comm_size*sizeof(int*));
memset(rin_sizes, 0, comm_size*sizeof(int));
memset(rout_sizes, 0, comm_size*sizeof(int));
memset(rin_idx, 0, comm_size*sizeof(int));
memset(rout_idx, 0, comm_size*sizeof(int));
/* compute array sizes */
idx = 0;
for (i = 0; i < n; ++i) {
MPIU_Assert(sources[i] < comm_size);
for (j = 0; j < degrees[i]; ++j) {
MPIU_Assert(destinations[idx] < comm_size);
/* rout_sizes[i] is twice as long as the number of edges to be
* sent to rank i by this process */
rout_sizes[sources[i]] += 2;
rin_sizes[destinations[idx]] += 2;
++idx;
}
}
/* allocate arrays */
for (i = 0; i < comm_size; ++i) {
/* can't use CHKLMEM macros b/c we are in a loop */
if (rin_sizes[i]) {
rin[i] = MPIU_Malloc(rin_sizes[i] * sizeof(int));
}
if (rout_sizes[i]) {
rout[i] = MPIU_Malloc(rout_sizes[i] * sizeof(int));
}
}
/* populate arrays */
idx = 0;
for (i = 0; i < n; ++i) {
/* TODO add this assert as proper error checking above */
int s_rank = sources[i];
MPIU_Assert(s_rank < comm_size);
MPIU_Assert(s_rank >= 0);
for (j = 0; j < degrees[i]; ++j) {
int d_rank = destinations[idx];
int weight = (weights == MPI_UNWEIGHTED ? 0 : weights[idx]);
/* TODO add this assert as proper error checking above */
MPIU_Assert(d_rank < comm_size);
MPIU_Assert(d_rank >= 0);
/* XXX DJG what about self-edges? do we need to drop one of these
* cases when there is a self-edge to avoid double-counting? */
/* rout[s][2*x] is the value of d for the j'th edge between (s,d)
* with weight rout[s][2*x+1], where x is the current end of the
* outgoing edge list for s. x==(rout_idx[s]/2) */
rout[s_rank][rout_idx[s_rank]++] = d_rank;
rout[s_rank][rout_idx[s_rank]++] = weight;
/* rin[d][2*x] is the value of s for the j'th edge between (s,d)
* with weight rout[d][2*x+1], where x is the current end of the
* incoming edge list for d. x==(rin_idx[d]/2) */
rin[d_rank][rin_idx[d_rank]++] = s_rank;
rin[d_rank][rin_idx[d_rank]++] = weight;
++idx;
}
}
for (i = 0; i < comm_size; ++i) {
/* sanity check that all arrays are fully populated*/
MPIU_Assert(rin_idx[i] == rin_sizes[i]);
MPIU_Assert(rout_idx[i] == rout_sizes[i]);
}
MPIU_CHKLMEM_MALLOC(rs, int *, 2*comm_size*sizeof(int), mpi_errno, "red-scat source buffer");
for (i = 0; i < comm_size; ++i) {
rs[2*i] = (rin_sizes[i] ? 1 : 0);
rs[2*i+1] = (rout_sizes[i] ? 1 : 0);
}
/* compute the number of peers I will recv from */
mpi_errno = MPIR_Reduce_scatter_block_impl(rs, in_out_peers, 2, MPI_INT, MPI_SUM, comm_ptr, &errflag);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
MPIU_ERR_CHKANDJUMP(errflag, mpi_errno, MPI_ERR_OTHER, "**coll_fail");
MPIU_Assert(in_out_peers[0] <= comm_size && in_out_peers[0] >= 0);
MPIU_Assert(in_out_peers[1] <= comm_size && in_out_peers[1] >= 0);
idx = 0;
/* must be 2*comm_size requests because we will possibly send inbound and
* outbound edges to everyone in our communicator */
MPIU_CHKLMEM_MALLOC(reqs, MPI_Request *, 2*comm_size*sizeof(MPI_Request), mpi_errno, "temp request array");
for (i = 0; i < comm_size; ++i) {
if (rin_sizes[i]) {
/* send edges where i is a destination to process i */
mpi_errno = MPIC_Isend(&rin[i][0], rin_sizes[i], MPI_INT, i, MPIR_TOPO_A_TAG, comm_old, &reqs[idx++]);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
}
if (rout_sizes[i]) {
/* send edges where i is a source to process i */
mpi_errno = MPIC_Isend(&rout[i][0], rout_sizes[i], MPI_INT, i, MPIR_TOPO_B_TAG, comm_old, &reqs[idx++]);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
}
}
MPIU_Assert(idx <= (2 * comm_size));
/* Create the topology structure */
MPIU_CHKPMEM_MALLOC(topo_ptr, MPIR_Topology *, sizeof(MPIR_Topology), mpi_errno, "topo_ptr");
topo_ptr->kind = MPI_DIST_GRAPH;
dist_graph_ptr = &topo_ptr->topo.dist_graph;
dist_graph_ptr->indegree = 0;
dist_graph_ptr->in = NULL;
dist_graph_ptr->in_weights = NULL;
dist_graph_ptr->outdegree = 0;
dist_graph_ptr->out = NULL;
dist_graph_ptr->out_weights = NULL;
dist_graph_ptr->is_weighted = (weights != MPI_UNWEIGHTED);
/* can't use CHKPMEM macros for this b/c we need to realloc */
in_capacity = 10; /* arbitrary */
dist_graph_ptr->in = MPIU_Malloc(in_capacity*sizeof(int));
if (dist_graph_ptr->is_weighted)
dist_graph_ptr->in_weights = MPIU_Malloc(in_capacity*sizeof(int));
out_capacity = 10; /* arbitrary */
dist_graph_ptr->out = MPIU_Malloc(out_capacity*sizeof(int));
if (dist_graph_ptr->is_weighted)
dist_graph_ptr->out_weights = MPIU_Malloc(out_capacity*sizeof(int));
for (i = 0; i < in_out_peers[0]; ++i) {
MPI_Status status;
int count;
int *buf;
/* receive inbound edges */
mpi_errno = MPIC_Probe(MPI_ANY_SOURCE, MPIR_TOPO_A_TAG, comm_old, &status);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
MPIR_Get_count_impl(&status, MPI_INT, &count);
/* can't use CHKLMEM macros b/c we are in a loop */
buf = MPIU_Malloc(count*sizeof(int));
MPIU_ERR_CHKANDJUMP(!buf, mpi_errno, MPIR_ERR_RECOVERABLE, "**nomem");
mpi_errno = MPIC_Recv(buf, count, MPI_INT, MPI_ANY_SOURCE, MPIR_TOPO_A_TAG, comm_old, MPI_STATUS_IGNORE);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
for (j = 0; j < count/2; ++j) {
int deg = dist_graph_ptr->indegree++;
if (deg >= in_capacity) {
in_capacity *= 2;
MPIU_REALLOC_ORJUMP(dist_graph_ptr->in, in_capacity*sizeof(int), mpi_errno);
if (dist_graph_ptr->is_weighted)
MPIU_REALLOC_ORJUMP(dist_graph_ptr->in_weights, in_capacity*sizeof(int), mpi_errno);
}
dist_graph_ptr->in[deg] = buf[2*j];
if (dist_graph_ptr->is_weighted)
dist_graph_ptr->in_weights[deg] = buf[2*j+1];
}
MPIU_Free(buf);
}
for (i = 0; i < in_out_peers[1]; ++i) {
MPI_Status status;
int count;
int *buf;
/* receive outbound edges */
mpi_errno = MPIC_Probe(MPI_ANY_SOURCE, MPIR_TOPO_B_TAG, comm_old, &status);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
MPIR_Get_count_impl(&status, MPI_INT, &count);
/* can't use CHKLMEM macros b/c we are in a loop */
buf = MPIU_Malloc(count*sizeof(int));
MPIU_ERR_CHKANDJUMP(!buf, mpi_errno, MPIR_ERR_RECOVERABLE, "**nomem");
mpi_errno = MPIC_Recv(buf, count, MPI_INT, MPI_ANY_SOURCE, MPIR_TOPO_B_TAG, comm_old, MPI_STATUS_IGNORE);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
for (j = 0; j < count/2; ++j) {
int deg = dist_graph_ptr->outdegree++;
if (deg >= out_capacity) {
out_capacity *= 2;
MPIU_REALLOC_ORJUMP(dist_graph_ptr->out, out_capacity*sizeof(int), mpi_errno);
if (dist_graph_ptr->is_weighted)
MPIU_REALLOC_ORJUMP(dist_graph_ptr->out_weights, out_capacity*sizeof(int), mpi_errno);
}
dist_graph_ptr->out[deg] = buf[2*j];
if (dist_graph_ptr->is_weighted)
dist_graph_ptr->out_weights[deg] = buf[2*j+1];
}
MPIU_Free(buf);
}
mpi_errno = MPIR_Waitall_impl(idx, reqs, MPI_STATUSES_IGNORE);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
/* remove any excess memory allocation */
MPIU_REALLOC_ORJUMP(dist_graph_ptr->in, dist_graph_ptr->indegree*sizeof(int), mpi_errno);
MPIU_REALLOC_ORJUMP(dist_graph_ptr->out, dist_graph_ptr->outdegree*sizeof(int), mpi_errno);
if (dist_graph_ptr->is_weighted) {
MPIU_REALLOC_ORJUMP(dist_graph_ptr->in_weights, dist_graph_ptr->indegree*sizeof(int), mpi_errno);
MPIU_REALLOC_ORJUMP(dist_graph_ptr->out_weights, dist_graph_ptr->outdegree*sizeof(int), mpi_errno);
}
mpi_errno = MPIR_Topology_put(comm_dist_graph_ptr, topo_ptr);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
MPIU_CHKPMEM_COMMIT();
MPIU_OBJ_PUBLISH_HANDLE(*comm_dist_graph, comm_dist_graph_ptr->handle);
/* ... end of body of routine ... */
fn_exit:
for (i = 0; i < comm_size; ++i) {
if (rin[i])
MPIU_Free(rin[i]);
if (rout[i])
MPIU_Free(rout[i]);
}
MPIU_CHKLMEM_FREEALL();
MPID_MPI_FUNC_EXIT(MPID_STATE_MPI_DIST_GRAPH_CREATE);
MPIU_THREAD_CS_EXIT(ALLFUNC,);
return mpi_errno;
/* --BEGIN ERROR HANDLING-- */
fn_fail:
if (dist_graph_ptr && dist_graph_ptr->in)
MPIU_Free(dist_graph_ptr->in);
if (dist_graph_ptr && dist_graph_ptr->in_weights)
MPIU_Free(dist_graph_ptr->in_weights);
if (dist_graph_ptr && dist_graph_ptr->out)
MPIU_Free(dist_graph_ptr->out);
if (dist_graph_ptr && dist_graph_ptr->out_weights)
MPIU_Free(dist_graph_ptr->out_weights);
MPIU_CHKPMEM_REAP();
#ifdef HAVE_ERROR_CHECKING
mpi_errno = MPIR_Err_create_code(
mpi_errno, MPIR_ERR_RECOVERABLE, FCNAME, __LINE__, MPI_ERR_OTHER,
"**mpi_dist_graph_create", "**mpi_dist_graph_create %C %d %p %p %p %p %I %d %p",
comm_old, n, sources, degrees, destinations, weights, info, reorder, comm_dist_graph);
#endif
mpi_errno = MPIR_Err_return_comm(comm_ptr, FCNAME, mpi_errno);
goto fn_exit;
/* --END ERROR HANDLING-- */
}
static int MPIR_Scan_generic (
const void *sendbuf,
void *recvbuf,
int count,
MPI_Datatype datatype,
MPI_Op op,
MPID_Comm *comm_ptr,
int *errflag )
{
MPI_Status status;
int rank, comm_size;
int mpi_errno = MPI_SUCCESS;
int mpi_errno_ret = MPI_SUCCESS;
int mask, dst, is_commutative;
MPI_Aint true_extent, true_lb, extent;
void *partial_scan, *tmp_buf;
MPID_Op *op_ptr;
MPI_Comm comm;
MPIU_THREADPRIV_DECL;
MPIU_CHKLMEM_DECL(2);
if (count == 0) return MPI_SUCCESS;
/* check if multiple threads are calling this collective function */
MPIDU_ERR_CHECK_MULTIPLE_THREADS_ENTER( comm_ptr );
comm = comm_ptr->handle;
comm_size = comm_ptr->local_size;
rank = comm_ptr->rank;
MPIU_THREADPRIV_GET;
/* set op_errno to 0. stored in perthread structure */
MPIU_THREADPRIV_FIELD(op_errno) = 0;
if (HANDLE_GET_KIND(op) == HANDLE_KIND_BUILTIN) {
is_commutative = 1;
}
else {
MPID_Op_get_ptr(op, op_ptr);
if (op_ptr->kind == MPID_OP_USER_NONCOMMUTE)
is_commutative = 0;
else
is_commutative = 1;
}
/* need to allocate temporary buffer to store partial scan*/
MPIR_Type_get_true_extent_impl(datatype, &true_lb, &true_extent);
MPID_Datatype_get_extent_macro(datatype, extent);
MPIU_CHKLMEM_MALLOC(partial_scan, void *, count*(MPIR_MAX(extent,true_extent)), mpi_errno, "partial_scan");
/* This eventually gets malloc()ed as a temp buffer, not added to
* any user buffers */
MPID_Ensure_Aint_fits_in_pointer(count * MPIR_MAX(extent, true_extent));
/* adjust for potential negative lower bound in datatype */
partial_scan = (void *)((char*)partial_scan - true_lb);
/* need to allocate temporary buffer to store incoming data*/
MPIU_CHKLMEM_MALLOC(tmp_buf, void *, count*(MPIR_MAX(extent,true_extent)), mpi_errno, "tmp_buf");
/* adjust for potential negative lower bound in datatype */
tmp_buf = (void *)((char*)tmp_buf - true_lb);
/* Since this is an inclusive scan, copy local contribution into
recvbuf. */
if (sendbuf != MPI_IN_PLACE) {
mpi_errno = MPIR_Localcopy(sendbuf, count, datatype,
recvbuf, count, datatype);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
}
if (sendbuf != MPI_IN_PLACE)
mpi_errno = MPIR_Localcopy(sendbuf, count, datatype,
partial_scan, count, datatype);
else
mpi_errno = MPIR_Localcopy(recvbuf, count, datatype,
partial_scan, count, datatype);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
mask = 0x1;
while (mask < comm_size) {
dst = rank ^ mask;
if (dst < comm_size) {
/* Send partial_scan to dst. Recv into tmp_buf */
mpi_errno = MPIC_Sendrecv(partial_scan, count, datatype,
dst, MPIR_SCAN_TAG, tmp_buf,
count, datatype, dst,
MPIR_SCAN_TAG, comm,
&status, errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag = TRUE;
MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
}
if (rank > dst) {
mpi_errno = MPIR_Reduce_local_impl(
tmp_buf, partial_scan, count, datatype, op);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
mpi_errno = MPIR_Reduce_local_impl(
tmp_buf, recvbuf, count, datatype, op);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
}
else {
if (is_commutative) {
mpi_errno = MPIR_Reduce_local_impl(
tmp_buf, partial_scan, count, datatype, op);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
}
else {
mpi_errno = MPIR_Reduce_local_impl(
partial_scan, tmp_buf, count, datatype, op);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
mpi_errno = MPIR_Localcopy(tmp_buf, count, datatype,
partial_scan,
count, datatype);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
}
}
}
mask <<= 1;
}
if (MPIU_THREADPRIV_FIELD(op_errno)) {
mpi_errno = MPIU_THREADPRIV_FIELD(op_errno);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
}
fn_exit:
MPIU_CHKLMEM_FREEALL();
/* check if multiple threads are calling this collective function */
MPIDU_ERR_CHECK_MULTIPLE_THREADS_EXIT( comm_ptr );
if (mpi_errno_ret)
mpi_errno = mpi_errno_ret;
else if (*errflag)
MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**coll_fail");
return mpi_errno;
fn_fail:
goto fn_exit;
}
int MPIR_Ialltoall_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_Alltoallw_intra(const void *sendbuf, const int sendcounts[], const int sdispls[],
const MPI_Datatype sendtypes[], void *recvbuf, const int recvcounts[],
const int rdispls[], const MPI_Datatype recvtypes[], MPID_Comm *comm_ptr,
int *errflag)
{
int comm_size, i, j;
int mpi_errno = MPI_SUCCESS;
int mpi_errno_ret = MPI_SUCCESS;
MPI_Status status;
MPI_Status *starray;
MPI_Request *reqarray;
int dst, rank;
MPI_Comm comm;
int outstanding_requests;
int ii, ss, bblock;
int type_size;
MPIU_CHKLMEM_DECL(2);
comm = comm_ptr->handle;
comm_size = comm_ptr->local_size;
rank = comm_ptr->rank;
/* check if multiple threads are calling this collective function */
MPIDU_ERR_CHECK_MULTIPLE_THREADS_ENTER( comm_ptr );
if (sendbuf == MPI_IN_PLACE) {
/* We use pair-wise sendrecv_replace in order to conserve memory usage,
* which is keeping with the spirit of the MPI-2.2 Standard. But
* because of this approach all processes must agree on the global
* schedule of sendrecv_replace operations to avoid deadlock.
*
* Note that this is not an especially efficient algorithm in terms of
* time and there will be multiple repeated malloc/free's rather than
* maintaining a single buffer across the whole loop. Something like
* MADRE is probably the best solution for the MPI_IN_PLACE scenario. */
for (i = 0; i < comm_size; ++i) {
/* start inner loop at i to avoid re-exchanging data */
for (j = i; j < comm_size; ++j) {
if (rank == i) {
/* also covers the (rank == i && rank == j) case */
mpi_errno = MPIC_Sendrecv_replace(((char *)recvbuf + rdispls[j]),
recvcounts[j], recvtypes[j],
j, MPIR_ALLTOALLW_TAG,
j, MPIR_ALLTOALLW_TAG,
comm, &status, errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag = TRUE;
MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
}
}
else if (rank == j) {
/* same as above with i/j args reversed */
mpi_errno = MPIC_Sendrecv_replace(((char *)recvbuf + rdispls[i]),
recvcounts[i], recvtypes[i],
i, MPIR_ALLTOALLW_TAG,
i, MPIR_ALLTOALLW_TAG,
comm, &status, errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag = TRUE;
MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
}
}
}
}
}
else {
bblock = MPIR_CVAR_ALLTOALL_THROTTLE;
if (bblock == 0) bblock = comm_size;
MPIU_CHKLMEM_MALLOC(starray, MPI_Status*, 2*bblock*sizeof(MPI_Status), mpi_errno, "starray");
MPIU_CHKLMEM_MALLOC(reqarray, MPI_Request*, 2*bblock*sizeof(MPI_Request), mpi_errno, "reqarray");
/* post only bblock isends/irecvs at a time as suggested by Tony Ladd */
for (ii=0; ii<comm_size; ii+=bblock) {
outstanding_requests = 0;
ss = comm_size-ii < bblock ? comm_size-ii : bblock;
/* do the communication -- post ss sends and receives: */
for ( i=0; i<ss; i++ ) {
dst = (rank+i+ii) % comm_size;
if (recvcounts[dst]) {
MPID_Datatype_get_size_macro(recvtypes[dst], type_size);
if (type_size) {
mpi_errno = MPIC_Irecv((char *)recvbuf+rdispls[dst],
recvcounts[dst], recvtypes[dst], dst,
MPIR_ALLTOALLW_TAG, comm,
&reqarray[outstanding_requests]);
if (mpi_errno) { MPIU_ERR_POP(mpi_errno); }
outstanding_requests++;
}
}
}
for ( i=0; i<ss; i++ ) {
dst = (rank-i-ii+comm_size) % comm_size;
if (sendcounts[dst]) {
MPID_Datatype_get_size_macro(sendtypes[dst], type_size);
if (type_size) {
mpi_errno = MPIC_Isend((char *)sendbuf+sdispls[dst],
sendcounts[dst], sendtypes[dst], dst,
MPIR_ALLTOALLW_TAG, comm,
&reqarray[outstanding_requests], errflag);
if (mpi_errno) { MPIU_ERR_POP(mpi_errno); }
outstanding_requests++;
}
}
}
mpi_errno = MPIC_Waitall(outstanding_requests, reqarray, starray, errflag);
if (mpi_errno && mpi_errno != MPI_ERR_IN_STATUS) MPIU_ERR_POP(mpi_errno);
/* --BEGIN ERROR HANDLING-- */
if (mpi_errno == MPI_ERR_IN_STATUS) {
for (i=0; i<outstanding_requests; i++) {
if (starray[i].MPI_ERROR != MPI_SUCCESS) {
mpi_errno = starray[i].MPI_ERROR;
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag = TRUE;
MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
}
}
}
}
/* --END ERROR HANDLING-- */
}
#ifdef FOO
/* Use pairwise exchange algorithm. */
/* Make local copy first */
mpi_errno = MPIR_Localcopy(((char *)sendbuf+sdispls[rank]),
sendcounts[rank], sendtypes[rank],
((char *)recvbuf+rdispls[rank]),
recvcounts[rank], recvtypes[rank]);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
/* Do the pairwise exchange. */
for (i=1; i<comm_size; i++) {
src = (rank - i + comm_size) % comm_size;
dst = (rank + i) % comm_size;
mpi_errno = MPIC_Sendrecv(((char *)sendbuf+sdispls[dst]),
sendcounts[dst], sendtypes[dst], dst,
MPIR_ALLTOALLW_TAG,
((char *)recvbuf+rdispls[src]),
recvcounts[src], recvtypes[dst], src,
MPIR_ALLTOALLW_TAG, comm, &status, errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag = TRUE;
MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
}
}
#endif
}
/* check if multiple threads are calling this collective function */
fn_exit:
MPIDU_ERR_CHECK_MULTIPLE_THREADS_EXIT( comm_ptr );
MPIU_CHKLMEM_FREEALL();
if (mpi_errno_ret)
mpi_errno = mpi_errno_ret;
else if (*errflag)
MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**coll_fail");
return mpi_errno;
fn_fail:
goto fn_exit;
}
int MPIR_Exscan (
void *sendbuf,
void *recvbuf,
int count,
MPI_Datatype datatype,
MPI_Op op,
MPID_Comm *comm_ptr,
int *errflag )
{
MPI_Status status;
int rank, comm_size;
int mpi_errno = MPI_SUCCESS;
int mpi_errno_ret = MPI_SUCCESS;
int mask, dst, is_commutative, flag;
MPI_Aint true_extent, true_lb, extent;
void *partial_scan, *tmp_buf;
MPI_User_function *uop;
MPID_Op *op_ptr;
MPI_Comm comm;
MPIU_CHKLMEM_DECL(2);
MPIU_THREADPRIV_DECL;
#ifdef HAVE_CXX_BINDING
int is_cxx_uop = 0;
#endif
if (count == 0) return MPI_SUCCESS;
MPIU_THREADPRIV_GET;
comm = comm_ptr->handle;
comm_size = comm_ptr->local_size;
rank = comm_ptr->rank;
/* set op_errno to 0. stored in perthread structure */
MPIU_THREADPRIV_FIELD(op_errno) = 0;
if (HANDLE_GET_KIND(op) == HANDLE_KIND_BUILTIN) {
is_commutative = 1;
/* get the function by indexing into the op table */
uop = MPIR_Op_table[op%16 - 1];
}
else {
MPID_Op_get_ptr(op, op_ptr);
if (op_ptr->kind == MPID_OP_USER_NONCOMMUTE)
is_commutative = 0;
else
is_commutative = 1;
#ifdef HAVE_CXX_BINDING
if (op_ptr->language == MPID_LANG_CXX) {
uop = (MPI_User_function *) op_ptr->function.c_function;
is_cxx_uop = 1;
}
else
#endif
if ((op_ptr->language == MPID_LANG_C))
uop = (MPI_User_function *) op_ptr->function.c_function;
else
uop = (MPI_User_function *) op_ptr->function.f77_function;
}
/* need to allocate temporary buffer to store partial scan*/
MPIR_Type_get_true_extent_impl(datatype, &true_lb, &true_extent);
MPID_Datatype_get_extent_macro( datatype, extent );
MPIU_CHKLMEM_MALLOC(partial_scan, void *, (count*(MPIR_MAX(true_extent,extent))), mpi_errno, "partial_scan");
/* adjust for potential negative lower bound in datatype */
partial_scan = (void *)((char*)partial_scan - true_lb);
/* need to allocate temporary buffer to store incoming data*/
MPIU_CHKLMEM_MALLOC(tmp_buf, void *, (count*(MPIR_MAX(true_extent,extent))), mpi_errno, "tmp_buf");
/* adjust for potential negative lower bound in datatype */
tmp_buf = (void *)((char*)tmp_buf - true_lb);
mpi_errno = MPIR_Localcopy((sendbuf == MPI_IN_PLACE ? recvbuf : sendbuf), count, datatype,
partial_scan, count, datatype);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
/* check if multiple threads are calling this collective function */
MPIDU_ERR_CHECK_MULTIPLE_THREADS_ENTER( comm_ptr );
flag = 0;
mask = 0x1;
while (mask < comm_size) {
dst = rank ^ mask;
if (dst < comm_size) {
/* Send partial_scan to dst. Recv into tmp_buf */
mpi_errno = MPIC_Sendrecv_ft(partial_scan, count, datatype,
dst, MPIR_EXSCAN_TAG, tmp_buf,
count, datatype, dst,
MPIR_EXSCAN_TAG, comm,
&status, errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag = TRUE;
MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
}
if (rank > dst) {
call_uop(tmp_buf, partial_scan, count, datatype);
/* On rank 0, recvbuf is not defined. For sendbuf==MPI_IN_PLACE
recvbuf must not change (per MPI-2.2).
On rank 1, recvbuf is to be set equal to the value
in sendbuf on rank 0.
On others, recvbuf is the scan of values in the
sendbufs on lower ranks. */
if (rank != 0) {
if (flag == 0) {
/* simply copy data recd from rank 0 into recvbuf */
mpi_errno = MPIR_Localcopy(tmp_buf, count, datatype,
recvbuf, count, datatype);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
flag = 1;
}
else {
call_uop(tmp_buf, recvbuf, count, datatype);
}
}
}
else {
if (is_commutative) {
call_uop(tmp_buf, partial_scan, count, datatype);
}
else {
call_uop(partial_scan, tmp_buf, count, datatype);
mpi_errno = MPIR_Localcopy(tmp_buf, count, datatype,
partial_scan,
count, datatype);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
}
}
}
mask <<= 1;
}
/* check if multiple threads are calling this collective function */
MPIDU_ERR_CHECK_MULTIPLE_THREADS_EXIT( comm_ptr );
if (MPIU_THREADPRIV_FIELD(op_errno))
mpi_errno = MPIU_THREADPRIV_FIELD(op_errno);
fn_exit:
MPIU_CHKLMEM_FREEALL();
if (mpi_errno_ret)
mpi_errno = mpi_errno_ret;
else if (*errflag)
MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**coll_fail");
return mpi_errno;
fn_fail:
goto fn_exit;
}
/*@
MPI_Type_create_darray - Create a datatype representing a distributed array
Input Parameters:
+ size - size of process group (positive integer)
. rank - rank in process group (nonnegative integer)
. ndims - number of array dimensions as well as process grid dimensions (positive integer)
. array_of_gsizes - number of elements of type oldtype in each dimension of global array (array of positive integers)
. array_of_distribs - distribution of array in each dimension (array of state)
. array_of_dargs - distribution argument in each dimension (array of positive integers)
. array_of_psizes - size of process grid in each dimension (array of positive integers)
. order - array storage order flag (state)
- oldtype - old datatype (handle)
Output Parameters:
. newtype - new datatype (handle)
.N ThreadSafe
.N Fortran
.N Errors
.N MPI_SUCCESS
.N MPI_ERR_TYPE
.N MPI_ERR_ARG
@*/
int MPI_Type_create_darray(int size,
int rank,
int ndims,
const int array_of_gsizes[],
const int array_of_distribs[],
const int array_of_dargs[],
const int array_of_psizes[],
int order,
MPI_Datatype oldtype,
MPI_Datatype *newtype)
{
static const char FCNAME[] = "MPI_Type_create_darray";
int mpi_errno = MPI_SUCCESS, i;
MPI_Datatype new_handle;
int procs, tmp_rank, tmp_size, blklens[3], *coords;
MPI_Aint *st_offsets, orig_extent, disps[3];
MPI_Datatype type_old, type_new = MPI_DATATYPE_NULL, types[3];
# ifdef HAVE_ERROR_CHECKING
MPI_Aint size_with_aint;
MPI_Offset size_with_offset;
# endif
int *ints;
MPID_Datatype *datatype_ptr = NULL;
MPIU_CHKLMEM_DECL(3);
MPID_MPI_STATE_DECL(MPID_STATE_MPI_TYPE_CREATE_DARRAY);
MPIR_ERRTEST_INITIALIZED_ORDIE();
MPIU_THREAD_CS_ENTER(ALLFUNC,);
MPID_MPI_FUNC_ENTER(MPID_STATE_MPI_TYPE_CREATE_DARRAY);
/* Validate parameters, especially handles needing to be converted */
# ifdef HAVE_ERROR_CHECKING
{
MPID_BEGIN_ERROR_CHECKS;
{
MPIR_ERRTEST_DATATYPE(oldtype, "datatype", mpi_errno);
}
MPID_END_ERROR_CHECKS;
}
# endif
/* Convert MPI object handles to object pointers */
MPID_Datatype_get_ptr(oldtype, datatype_ptr);
MPID_Datatype_get_extent_macro(oldtype, orig_extent);
/* Validate parameters and objects (post conversion) */
# ifdef HAVE_ERROR_CHECKING
{
MPID_BEGIN_ERROR_CHECKS;
{
/* Check parameters */
MPIR_ERRTEST_ARGNONPOS(size, "size", mpi_errno, MPI_ERR_ARG);
/* use MPI_ERR_RANK class for PE-MPI compatibility */
MPIU_ERR_CHKANDJUMP3((rank < 0 || rank >= size), mpi_errno, MPI_ERR_RANK,
"**argrange", "**argrange %s %d %d", "rank", rank, (size-1));
MPIR_ERRTEST_ARGNONPOS(ndims, "ndims", mpi_errno, MPI_ERR_DIMS);
MPIR_ERRTEST_ARGNULL(array_of_gsizes, "array_of_gsizes", mpi_errno);
MPIR_ERRTEST_ARGNULL(array_of_distribs, "array_of_distribs", mpi_errno);
MPIR_ERRTEST_ARGNULL(array_of_dargs, "array_of_dargs", mpi_errno);
MPIR_ERRTEST_ARGNULL(array_of_psizes, "array_of_psizes", mpi_errno);
if (order != MPI_ORDER_C && order != MPI_ORDER_FORTRAN) {
mpi_errno = MPIR_Err_create_code(MPI_SUCCESS,
MPIR_ERR_RECOVERABLE,
FCNAME,
__LINE__,
MPI_ERR_ARG,
"**arg",
"**arg %s",
"order");
goto fn_fail;
}
tmp_size = 1;
for (i=0; mpi_errno == MPI_SUCCESS && i < ndims; i++) {
MPIR_ERRTEST_ARGNONPOS(array_of_gsizes[i], "gsize", mpi_errno, MPI_ERR_ARG);
MPIR_ERRTEST_ARGNONPOS(array_of_psizes[i], "psize", mpi_errno, MPI_ERR_ARG);
if ((array_of_distribs[i] != MPI_DISTRIBUTE_NONE) &&
(array_of_distribs[i] != MPI_DISTRIBUTE_BLOCK) &&
(array_of_distribs[i] != MPI_DISTRIBUTE_CYCLIC))
{
mpi_errno = MPIR_Err_create_code(MPI_SUCCESS,
MPIR_ERR_RECOVERABLE,
FCNAME,
__LINE__,
MPI_ERR_ARG,
"**darrayunknown",
0);
goto fn_fail;
}
if ((array_of_dargs[i] != MPI_DISTRIBUTE_DFLT_DARG) &&
(array_of_dargs[i] <= 0))
{
mpi_errno = MPIR_Err_create_code(MPI_SUCCESS,
MPIR_ERR_RECOVERABLE,
FCNAME,
__LINE__,
MPI_ERR_ARG,
"**arg",
"**arg %s",
"array_of_dargs");
goto fn_fail;
}
if ((array_of_distribs[i] == MPI_DISTRIBUTE_NONE) &&
(array_of_psizes[i] != 1))
{
mpi_errno = MPIR_Err_create_code(MPI_SUCCESS,
MPIR_ERR_RECOVERABLE,
FCNAME,
__LINE__,
MPI_ERR_ARG,
"**darraydist",
"**darraydist %d %d",
i, array_of_psizes[i]);
goto fn_fail;
}
tmp_size *= array_of_psizes[i];
}
MPIU_ERR_CHKANDJUMP1((tmp_size != size), mpi_errno, MPI_ERR_ARG,
"**arg", "**arg %s", "array_of_psizes");
/* TODO: GET THIS CHECK IN ALSO */
/* check if MPI_Aint is large enough for size of global array.
if not, complain. */
size_with_aint = orig_extent;
for (i=0; i<ndims; i++) size_with_aint *= array_of_gsizes[i];
size_with_offset = orig_extent;
for (i=0; i<ndims; i++) size_with_offset *= array_of_gsizes[i];
if (size_with_aint != size_with_offset) {
mpi_errno = MPIR_Err_create_code(MPI_SUCCESS,
MPIR_ERR_FATAL,
FCNAME,
__LINE__,
MPI_ERR_ARG,
"**darrayoverflow",
"**darrayoverflow %L",
size_with_offset);
goto fn_fail;
}
/* Validate datatype_ptr */
MPID_Datatype_valid_ptr(datatype_ptr, mpi_errno);
/* If datatype_ptr is not valid, it will be reset to null */
/* --BEGIN ERROR HANDLING-- */
if (mpi_errno) goto fn_fail;
/* --END ERROR HANDLING-- */
}
MPID_END_ERROR_CHECKS;
}
# endif /* HAVE_ERROR_CHECKING */
/* ... body of routine ... */
/* calculate position in Cartesian grid as MPI would (row-major
ordering) */
MPIU_CHKLMEM_MALLOC_ORJUMP(coords, int *, ndims * sizeof(int), mpi_errno, "position is Cartesian grid");
procs = size;
tmp_rank = rank;
for (i=0; i<ndims; i++) {
procs = procs/array_of_psizes[i];
coords[i] = tmp_rank/procs;
tmp_rank = tmp_rank % procs;
}
MPIU_CHKLMEM_MALLOC_ORJUMP(st_offsets, MPI_Aint *, ndims * sizeof(MPI_Aint), mpi_errno, "st_offsets");
type_old = oldtype;
if (order == MPI_ORDER_FORTRAN) {
/* dimension 0 changes fastest */
for (i=0; i<ndims; i++) {
switch(array_of_distribs[i]) {
case MPI_DISTRIBUTE_BLOCK:
mpi_errno = MPIR_Type_block(array_of_gsizes,
i,
ndims,
array_of_psizes[i],
coords[i],
array_of_dargs[i],
order,
orig_extent,
type_old,
&type_new,
st_offsets+i);
break;
case MPI_DISTRIBUTE_CYCLIC:
mpi_errno = MPIR_Type_cyclic(array_of_gsizes,
i,
ndims,
array_of_psizes[i],
coords[i],
array_of_dargs[i],
order,
orig_extent,
type_old,
&type_new,
st_offsets+i);
break;
case MPI_DISTRIBUTE_NONE:
/* treat it as a block distribution on 1 process */
mpi_errno = MPIR_Type_block(array_of_gsizes,
i,
ndims,
1,
0,
MPI_DISTRIBUTE_DFLT_DARG,
order,
orig_extent,
type_old,
&type_new,
st_offsets+i);
break;
}
if (i)
{
MPIR_Type_free_impl(&type_old);
}
type_old = type_new;
/* --BEGIN ERROR HANDLING-- */
if (mpi_errno != MPI_SUCCESS) goto fn_fail;
/* --END ERROR HANDLING-- */
}
/* add displacement and UB */
disps[1] = st_offsets[0];
tmp_size = 1;
for (i=1; i<ndims; i++) {
tmp_size *= array_of_gsizes[i-1];
disps[1] += (MPI_Aint) tmp_size * st_offsets[i];
}
/* rest done below for both Fortran and C order */
}
else /* order == MPI_ORDER_C */ {
/* dimension ndims-1 changes fastest */
for (i=ndims-1; i>=0; i--) {
switch(array_of_distribs[i]) {
case MPI_DISTRIBUTE_BLOCK:
mpi_errno = MPIR_Type_block(array_of_gsizes,
i,
ndims,
array_of_psizes[i],
coords[i],
array_of_dargs[i],
order,
orig_extent,
type_old,
&type_new,
st_offsets+i);
break;
case MPI_DISTRIBUTE_CYCLIC:
mpi_errno = MPIR_Type_cyclic(array_of_gsizes,
i,
ndims,
array_of_psizes[i],
coords[i],
array_of_dargs[i],
order,
orig_extent,
type_old,
&type_new,
st_offsets+i);
break;
case MPI_DISTRIBUTE_NONE:
/* treat it as a block distribution on 1 process */
mpi_errno = MPIR_Type_block(array_of_gsizes,
i,
ndims,
array_of_psizes[i],
coords[i],
MPI_DISTRIBUTE_DFLT_DARG,
order,
orig_extent,
type_old,
&type_new,
st_offsets+i);
break;
}
if (i != ndims-1)
{
MPIR_Type_free_impl(&type_old);
}
type_old = type_new;
/* --BEGIN ERROR HANDLING-- */
if (mpi_errno != MPI_SUCCESS) goto fn_fail;
/* --END ERROR HANDLING-- */
}
/* add displacement and UB */
disps[1] = st_offsets[ndims-1];
tmp_size = 1;
for (i=ndims-2; i>=0; i--) {
tmp_size *= array_of_gsizes[i+1];
disps[1] += (MPI_Aint) tmp_size * st_offsets[i];
}
}
disps[1] *= orig_extent;
disps[2] = orig_extent;
for (i=0; i<ndims; i++) disps[2] *= (MPI_Aint)(array_of_gsizes[i]);
disps[0] = 0;
blklens[0] = blklens[1] = blklens[2] = 1;
types[0] = MPI_LB;
types[1] = type_new;
types[2] = MPI_UB;
mpi_errno = MPID_Type_struct(3,
blklens,
disps,
types,
&new_handle);
/* --BEGIN ERROR HANDLING-- */
if (mpi_errno != MPI_SUCCESS) goto fn_fail;
/* --END ERROR HANDLING-- */
MPIR_Type_free_impl(&type_new);
/* at this point we have the new type, and we've cleaned up any
* intermediate types created in the process. we just need to save
* all our contents/envelope information.
*/
/* Save contents */
MPIU_CHKLMEM_MALLOC_ORJUMP(ints, int *, (4 * ndims + 4) * sizeof(int), mpi_errno, "content description");
ints[0] = size;
ints[1] = rank;
ints[2] = ndims;
for (i=0; i < ndims; i++) {
ints[i + 3] = array_of_gsizes[i];
}
for (i=0; i < ndims; i++) {
ints[i + ndims + 3] = array_of_distribs[i];
}
for (i=0; i < ndims; i++) {
ints[i + 2*ndims + 3] = array_of_dargs[i];
}
for (i=0; i < ndims; i++) {
ints[i + 3*ndims + 3] = array_of_psizes[i];
}
ints[4*ndims + 3] = order;
MPID_Datatype_get_ptr(new_handle, datatype_ptr);
mpi_errno = MPID_Datatype_set_contents(datatype_ptr,
MPI_COMBINER_DARRAY,
4*ndims + 4,
0,
1,
ints,
NULL,
&oldtype);
/* --BEGIN ERROR HANDLING-- */
if (mpi_errno != MPI_SUCCESS) goto fn_fail;
/* --END ERROR HANDLING-- */
MPIU_OBJ_PUBLISH_HANDLE(*newtype, new_handle);
/* ... end of body of routine ... */
fn_exit:
MPIU_CHKLMEM_FREEALL();
MPID_MPI_FUNC_EXIT(MPID_STATE_MPI_TYPE_CREATE_DARRAY);
MPIU_THREAD_CS_EXIT(ALLFUNC,);
return mpi_errno;
fn_fail:
/* --BEGIN ERROR HANDLING-- */
# ifdef HAVE_ERROR_CHECKING
{
mpi_errno = MPIR_Err_create_code(
mpi_errno, MPIR_ERR_RECOVERABLE, FCNAME, __LINE__, MPI_ERR_OTHER, "**mpi_type_create_darray",
"**mpi_type_create_darray %d %d %d %p %p %p %p %d %D %p", size, rank, ndims, array_of_gsizes,
array_of_distribs, array_of_dargs, array_of_psizes, order, oldtype, newtype);
}
# endif
mpi_errno = MPIR_Err_return_comm(NULL, FCNAME, mpi_errno);
goto fn_exit;
/* --END ERROR HANDLING-- */
}
int MPID_Comm_failed_bitarray(MPID_Comm *comm_ptr, uint32_t **bitarray, int acked)
{
int mpi_errno = MPI_SUCCESS;
int size, i;
uint32_t bit;
int *failed_procs, *group_procs;
MPID_Group *failed_group, *comm_group;
MPIU_CHKLMEM_DECL(2);
MPIDI_STATE_DECL(MPID_STATE_COMM_FAILED_BITARRAY);
MPIDI_FUNC_ENTER(MPID_STATE_COMM_FAILED_BITARRAY);
/* TODO - Fix this for intercommunicators */
size = comm_ptr->local_size;
/* We can fit sizeof(uint32_t) * 8 ranks in one uint64_t so divide the
* size by that */
/* This buffer will be handed back to the calling function so we use a
* "real" malloc here and expect the caller to free the buffer later. The
* other buffers in this function are temporary and will be automatically
* cleaned up at the end of the function. */
*bitarray = (uint32_t *) MPIU_Malloc(sizeof(uint32_t) * (size / (sizeof(uint32_t) * 8)+1));
if (!(*bitarray)) {
fprintf(stderr, "Could not allocate space for bitarray\n");
PMPI_Abort(MPI_COMM_WORLD, 1);
}
for (i = 0; i <= size/(sizeof(uint32_t)*8); i++) *bitarray[i] = 0;
mpi_errno = MPIDI_CH3U_Check_for_failed_procs();
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
if (acked)
MPIDI_CH3U_Get_failed_group(comm_ptr->dev.last_ack_rank, &failed_group);
else
MPIDI_CH3U_Get_failed_group(-2, &failed_group);
if (failed_group == MPID_Group_empty) goto fn_exit;
MPIU_CHKLMEM_MALLOC(group_procs, int *, sizeof(int)*failed_group->size, mpi_errno, "group_procs");
for (i = 0; i < failed_group->size; i++) group_procs[i] = i;
MPIU_CHKLMEM_MALLOC(failed_procs, int *, sizeof(int)*failed_group->size, mpi_errno, "failed_procs");
MPIR_Comm_group_impl(comm_ptr, &comm_group);
MPIR_Group_translate_ranks_impl(failed_group, failed_group->size, group_procs, comm_group, failed_procs);
/* The bits will actually be ordered in decending order rather than
* ascending. This is purely for readability since it makes no practical
* difference. So if the bits look like this:
*
* 10001100 01001000 00000000 00000001
*
* Then processes 1, 5, 6, 9, 12, and 32 have failed. */
for (i = 0; i < failed_group->size; i++) {
bit = 0x80000000;
bit >>= failed_procs[i] % (sizeof(uint32_t) * 8);
*bitarray[failed_procs[i] / (sizeof(uint32_t) * 8)] |= bit;
}
MPIR_Group_free_impl(comm_group);
fn_exit:
MPIU_CHKLMEM_FREEALL();
MPIDI_FUNC_EXIT(MPID_STATE_COMM_FAILED_BITARRAY);
return mpi_errno;
fn_fail:
goto fn_exit;
}
static int MPIR_Reduce_scatter_block_noncomm (
const void *sendbuf,
void *recvbuf,
int recvcount,
MPI_Datatype datatype,
MPI_Op op,
MPID_Comm *comm_ptr,
int *errflag )
{
int mpi_errno = MPI_SUCCESS;
int mpi_errno_ret = MPI_SUCCESS;
int comm_size = comm_ptr->local_size;
int rank = comm_ptr->rank;
int pof2;
int log2_comm_size;
int i, k;
int recv_offset, send_offset;
int block_size, total_count, size;
MPI_Aint true_extent, true_lb;
int buf0_was_inout;
void *tmp_buf0;
void *tmp_buf1;
void *result_ptr;
MPI_Comm comm = comm_ptr->handle;
MPIU_CHKLMEM_DECL(3);
MPIR_Type_get_true_extent_impl(datatype, &true_lb, &true_extent);
pof2 = 1;
log2_comm_size = 0;
while (pof2 < comm_size) {
pof2 <<= 1;
++log2_comm_size;
}
/* begin error checking */
MPIU_Assert(pof2 == comm_size); /* FIXME this version only works for power of 2 procs */
/* end error checking */
/* size of a block (count of datatype per block, NOT bytes per block) */
block_size = recvcount;
total_count = block_size * comm_size;
MPIU_CHKLMEM_MALLOC(tmp_buf0, void *, true_extent * total_count, mpi_errno, "tmp_buf0");
MPIU_CHKLMEM_MALLOC(tmp_buf1, void *, true_extent * total_count, mpi_errno, "tmp_buf1");
/* adjust for potential negative lower bound in datatype */
tmp_buf0 = (void *)((char*)tmp_buf0 - true_lb);
tmp_buf1 = (void *)((char*)tmp_buf1 - true_lb);
/* Copy our send data to tmp_buf0. We do this one block at a time and
permute the blocks as we go according to the mirror permutation. */
for (i = 0; i < comm_size; ++i) {
mpi_errno = MPIR_Localcopy((char *)(sendbuf == MPI_IN_PLACE ? recvbuf : sendbuf) + (i * true_extent * block_size), block_size, datatype,
(char *)tmp_buf0 + (MPIU_Mirror_permutation(i, log2_comm_size) * true_extent * block_size), block_size, datatype);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
}
buf0_was_inout = 1;
send_offset = 0;
recv_offset = 0;
size = total_count;
for (k = 0; k < log2_comm_size; ++k) {
/* use a double-buffering scheme to avoid local copies */
char *incoming_data = (buf0_was_inout ? tmp_buf1 : tmp_buf0);
char *outgoing_data = (buf0_was_inout ? tmp_buf0 : tmp_buf1);
int peer = rank ^ (0x1 << k);
size /= 2;
if (rank > peer) {
/* we have the higher rank: send top half, recv bottom half */
recv_offset += size;
}
else {
/* we have the lower rank: recv top half, send bottom half */
send_offset += size;
}
mpi_errno = MPIC_Sendrecv_ft(outgoing_data + send_offset*true_extent,
size, datatype, peer, MPIR_REDUCE_SCATTER_BLOCK_TAG,
incoming_data + recv_offset*true_extent,
size, datatype, peer, MPIR_REDUCE_SCATTER_BLOCK_TAG,
comm, MPI_STATUS_IGNORE, errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag = TRUE;
MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
}
/* always perform the reduction at recv_offset, the data at send_offset
is now our peer's responsibility */
if (rank > peer) {
/* higher ranked value so need to call op(received_data, my_data) */
mpi_errno = MPIR_Reduce_local_impl(
incoming_data + recv_offset*true_extent,
outgoing_data + recv_offset*true_extent,
size, datatype, op);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
buf0_was_inout = buf0_was_inout;
}
else {
/* lower ranked value so need to call op(my_data, received_data) */
mpi_errno = MPIR_Reduce_local_impl(
outgoing_data + recv_offset*true_extent,
incoming_data + recv_offset*true_extent,
size, datatype, op);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
buf0_was_inout = !buf0_was_inout;
}
/* the next round of send/recv needs to happen within the block (of size
"size") that we just received and reduced */
send_offset = recv_offset;
}
MPIU_Assert(size == recvcount);
/* copy the reduced data to the recvbuf */
result_ptr = (char *)(buf0_was_inout ? tmp_buf0 : tmp_buf1) + recv_offset * true_extent;
mpi_errno = MPIR_Localcopy(result_ptr, size, datatype,
recvbuf, size, datatype);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
fn_exit:
MPIU_CHKLMEM_FREEALL();
/* --BEGIN ERROR HANDLING-- */
if (mpi_errno_ret)
mpi_errno = mpi_errno_ret;
else if (*errflag)
MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**coll_fail");
/* --END ERROR HANDLING-- */
return mpi_errno;
fn_fail:
goto fn_exit;
}
/*@
MPI_Comm_join - Create a communicator by joining two processes connected by
a socket.
Input Parameter:
. fd - socket file descriptor
Output Parameter:
. intercomm - new intercommunicator (handle)
Notes:
The socket must be quiescent before 'MPI_COMM_JOIN' is called and after
'MPI_COMM_JOIN' returns. More specifically, on entry to 'MPI_COMM_JOIN', a
read on the socket will not read any data that was written to the socket
before the remote process called 'MPI_COMM_JOIN'.
.N ThreadSafe
.N Fortran
.N Errors
.N MPI_SUCCESS
.N MPI_ERR_ARG
@*/
int MPI_Comm_join(int fd, MPI_Comm *intercomm)
{
static const char FCNAME[] = "MPI_Comm_join";
int mpi_errno = MPI_SUCCESS, err;
MPID_Comm *intercomm_ptr;
char *local_port, *remote_port;
MPIU_CHKLMEM_DECL(2);
MPID_MPI_STATE_DECL(MPID_STATE_MPI_COMM_JOIN);
MPIR_ERRTEST_INITIALIZED_ORDIE();
MPIU_THREAD_CS_ENTER(ALLFUNC,);
MPID_MPI_FUNC_ENTER(MPID_STATE_MPI_COMM_JOIN);
/* ... body of routine ... */
MPIU_CHKLMEM_MALLOC(local_port, char *, MPI_MAX_PORT_NAME, mpi_errno, "local port name");
MPIU_CHKLMEM_MALLOC(remote_port, char *, MPI_MAX_PORT_NAME, mpi_errno, "remote port name");
mpi_errno = MPIR_Open_port_impl(NULL, local_port);
MPIU_ERR_CHKANDJUMP((mpi_errno != MPI_SUCCESS), mpi_errno, MPI_ERR_OTHER, "**openportfailed");
err = MPIR_fd_send(fd, local_port, MPI_MAX_PORT_NAME);
MPIU_ERR_CHKANDJUMP1((err != 0), mpi_errno, MPI_ERR_INTERN, "**join_send", "**join_send %d", err);
err = MPIR_fd_recv(fd, remote_port, MPI_MAX_PORT_NAME);
MPIU_ERR_CHKANDJUMP1((err != 0), mpi_errno, MPI_ERR_INTERN, "**join_recv", "**join_recv %d", err);
MPIU_ERR_CHKANDJUMP2((strcmp(local_port, remote_port) == 0), mpi_errno, MPI_ERR_INTERN, "**join_portname",
"**join_portname %s %s", local_port, remote_port);
if (strcmp(local_port, remote_port) < 0) {
MPID_Comm *comm_self_ptr;
MPID_Comm_get_ptr( MPI_COMM_SELF, comm_self_ptr );
mpi_errno = MPIR_Comm_accept_impl(local_port, NULL, 0, comm_self_ptr, &intercomm_ptr);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
} else {
MPID_Comm *comm_self_ptr;
MPID_Comm_get_ptr( MPI_COMM_SELF, comm_self_ptr );
mpi_errno = MPIR_Comm_connect_impl(remote_port, NULL, 0, comm_self_ptr, &intercomm_ptr);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
}
mpi_errno = MPIR_Close_port_impl(local_port);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
MPIU_OBJ_PUBLISH_HANDLE(*intercomm, intercomm_ptr->handle);
/* ... end of body of routine ... */
fn_exit:
MPIU_CHKLMEM_FREEALL();
MPID_MPI_FUNC_EXIT(MPID_STATE_MPI_COMM_JOIN);
MPIU_THREAD_CS_EXIT(ALLFUNC,);
return mpi_errno;
fn_fail:
/* --BEGIN ERROR HANDLING-- */
# ifdef HAVE_ERROR_CHECKING
{
mpi_errno = MPIR_Err_create_code(
mpi_errno, MPIR_ERR_RECOVERABLE, FCNAME, __LINE__, MPI_ERR_OTHER, "**mpi_comm_join",
"**mpi_comm_join %d %p", fd, intercomm);
}
# endif
mpi_errno = MPIR_Err_return_comm( NULL, FCNAME, mpi_errno );
goto fn_exit;
/* --END ERROR HANDLING-- */
}
int MPIDI_CH3U_Win_gather_info(void *base, MPI_Aint size, int disp_unit,
MPID_Info * info, MPID_Comm * comm_ptr, MPID_Win ** win_ptr)
{
int mpi_errno = MPI_SUCCESS, i, k, comm_size, rank;
MPI_Aint *tmp_buf;
MPIR_Errflag_t errflag = MPIR_ERR_NONE;
MPIU_CHKPMEM_DECL(1);
MPIU_CHKLMEM_DECL(1);
MPIDI_STATE_DECL(MPID_STATE_MPIDI_CH3U_WIN_GATHER_INFO);
MPIDI_RMA_FUNC_ENTER(MPID_STATE_MPIDI_CH3U_WIN_GATHER_INFO);
comm_size = (*win_ptr)->comm_ptr->local_size;
rank = (*win_ptr)->comm_ptr->rank;
MPIR_T_PVAR_TIMER_START(RMA, rma_wincreate_allgather);
/* allocate memory for the base addresses, disp_units, and
* completion counters of all processes */
MPIU_CHKPMEM_MALLOC((*win_ptr)->basic_info_table, MPIDI_Win_basic_info_t *,
comm_size * sizeof(MPIDI_Win_basic_info_t),
mpi_errno, "(*win_ptr)->basic_info_table");
/* get the addresses of the windows, window objects, and completion
* counters of all processes. allocate temp. buffer for communication */
MPIU_CHKLMEM_MALLOC(tmp_buf, MPI_Aint *, 4 * comm_size * sizeof(MPI_Aint),
mpi_errno, "tmp_buf");
/* FIXME: This needs to be fixed for heterogeneous systems */
/* FIXME: If we wanted to validate the transfer as within range at the
* origin, we'd also need the window size. */
tmp_buf[4 * rank] = MPIU_PtrToAint(base);
tmp_buf[4 * rank + 1] = size;
tmp_buf[4 * rank + 2] = (MPI_Aint) disp_unit;
tmp_buf[4 * rank + 3] = (MPI_Aint) (*win_ptr)->handle;
mpi_errno = MPIR_Allgather_impl(MPI_IN_PLACE, 0, MPI_DATATYPE_NULL,
tmp_buf, 4, MPI_AINT, (*win_ptr)->comm_ptr, &errflag);
MPIR_T_PVAR_TIMER_END(RMA, rma_wincreate_allgather);
if (mpi_errno) {
MPIR_ERR_POP(mpi_errno);
}
MPIR_ERR_CHKANDJUMP(errflag, mpi_errno, MPI_ERR_OTHER, "**coll_fail");
k = 0;
for (i = 0; i < comm_size; i++) {
(*win_ptr)->basic_info_table[i].base_addr = MPIU_AintToPtr(tmp_buf[k++]);
(*win_ptr)->basic_info_table[i].size = tmp_buf[k++];
(*win_ptr)->basic_info_table[i].disp_unit = (int) tmp_buf[k++];
(*win_ptr)->basic_info_table[i].win_handle = (MPI_Win) tmp_buf[k++];
}
fn_exit:
MPIU_CHKLMEM_FREEALL();
MPIDI_RMA_FUNC_EXIT(MPID_STATE_MPIDI_CH3U_WIN_GATHER_INFO);
return mpi_errno;
/* --BEGIN ERROR HANDLING-- */
fn_fail:
MPIU_CHKPMEM_REAP();
goto fn_exit;
/* --END ERROR HANDLING-- */
}
int MPIDI_Win_create(void *base, MPI_Aint size, int disp_unit, MPID_Info *info,
MPID_Comm *comm_ptr, MPID_Win **win_ptr )
{
int mpi_errno=MPI_SUCCESS, i, k, comm_size, rank;
MPI_Aint *tmp_buf;
MPID_Comm *win_comm_ptr;
int errflag = FALSE;
MPIU_CHKPMEM_DECL(4);
MPIU_CHKLMEM_DECL(1);
MPIDI_STATE_DECL(MPID_STATE_MPIDI_WIN_CREATE);
MPIDI_RMA_FUNC_ENTER(MPID_STATE_MPIDI_WIN_CREATE);
/* FIXME: There should be no unreferenced args */
MPIU_UNREFERENCED_ARG(info);
if(initRMAoptions) {
int rc;
MPIU_THREADSAFE_INIT_BLOCK_BEGIN(initRMAoptions);
/* Default is to enable the use of the immediate accumulate feature */
if (!MPL_env2bool( "MPICH_RMA_ACC_IMMED", &rc ))
rc = 1;
MPIDI_CH3_RMA_SetAccImmed(rc);
#ifdef USE_MPIU_INSTR
/* Define all instrumentation handle used in the CH3 RMA here*/
MPIU_INSTR_DURATION_INIT(wincreate_allgather,0,"WIN_CREATE:Allgather");
MPIU_INSTR_DURATION_INIT(winfree_rs,0,"WIN_FREE:ReduceScatterBlock");
MPIU_INSTR_DURATION_INIT(winfree_complete,0,"WIN_FREE:Complete");
MPIU_INSTR_DURATION_INIT(rmaqueue_alloc,0,"Allocate RMA Queue element");
MPIDI_CH3_RMA_InitInstr();
#endif
MPIU_THREADSAFE_INIT_CLEAR(initRMAoptions);
MPIU_THREADSAFE_INIT_BLOCK_END(initRMAoptions);
}
comm_size = comm_ptr->local_size;
rank = comm_ptr->rank;
*win_ptr = (MPID_Win *)MPIU_Handle_obj_alloc( &MPID_Win_mem );
MPIU_ERR_CHKANDJUMP1(!(*win_ptr),mpi_errno,MPI_ERR_OTHER,"**nomem",
"**nomem %s","MPID_Win_mem");
MPIU_Object_set_ref(*win_ptr, 1);
(*win_ptr)->fence_cnt = 0;
(*win_ptr)->base = base;
(*win_ptr)->size = size;
(*win_ptr)->disp_unit = disp_unit;
(*win_ptr)->start_group_ptr = NULL;
(*win_ptr)->start_assert = 0;
(*win_ptr)->attributes = NULL;
(*win_ptr)->rma_ops_list_head = NULL;
(*win_ptr)->rma_ops_list_tail = NULL;
(*win_ptr)->lock_granted = 0;
(*win_ptr)->current_lock_type = MPID_LOCK_NONE;
(*win_ptr)->shared_lock_ref_cnt = 0;
(*win_ptr)->lock_queue = NULL;
(*win_ptr)->my_counter = 0;
(*win_ptr)->my_pt_rma_puts_accs = 0;
mpi_errno = MPIR_Comm_dup_impl(comm_ptr, &win_comm_ptr);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
(*win_ptr)->comm_ptr = win_comm_ptr;
(*win_ptr)->myrank = rank;
MPIU_INSTR_DURATION_START(wincreate_allgather);
/* allocate memory for the base addresses, disp_units, and
completion counters of all processes */
MPIU_CHKPMEM_MALLOC((*win_ptr)->base_addrs, void **,
comm_size*sizeof(void *),
mpi_errno, "(*win_ptr)->base_addrs");
MPIU_CHKPMEM_MALLOC((*win_ptr)->disp_units, int *, comm_size*sizeof(int),
mpi_errno, "(*win_ptr)->disp_units");
MPIU_CHKPMEM_MALLOC((*win_ptr)->all_win_handles, MPI_Win *,
comm_size*sizeof(MPI_Win),
mpi_errno, "(*win_ptr)->all_win_handles");
MPIU_CHKPMEM_MALLOC((*win_ptr)->pt_rma_puts_accs, int *,
comm_size*sizeof(int),
mpi_errno, "(*win_ptr)->pt_rma_puts_accs");
for (i=0; i<comm_size; i++) (*win_ptr)->pt_rma_puts_accs[i] = 0;
/* get the addresses of the windows, window objects, and completion
counters of all processes. allocate temp. buffer for communication */
MPIU_CHKLMEM_MALLOC(tmp_buf, MPI_Aint *, 3*comm_size*sizeof(MPI_Aint),
mpi_errno, "tmp_buf");
/* FIXME: This needs to be fixed for heterogeneous systems */
tmp_buf[3*rank] = MPIU_PtrToAint(base);
tmp_buf[3*rank+1] = (MPI_Aint) disp_unit;
tmp_buf[3*rank+2] = (MPI_Aint) (*win_ptr)->handle;
mpi_errno = MPIR_Allgather_impl(MPI_IN_PLACE, 0, MPI_DATATYPE_NULL,
tmp_buf, 3 * sizeof(MPI_Aint), MPI_BYTE,
comm_ptr, &errflag);
MPIU_INSTR_DURATION_END(wincreate_allgather);
if (mpi_errno) { MPIU_ERR_POP(mpi_errno); }
MPIU_ERR_CHKANDJUMP(errflag, mpi_errno, MPI_ERR_OTHER, "**coll_fail");
k = 0;
for (i=0; i<comm_size; i++)
{
(*win_ptr)->base_addrs[i] = MPIU_AintToPtr(tmp_buf[k++]);
(*win_ptr)->disp_units[i] = (int) tmp_buf[k++];
(*win_ptr)->all_win_handles[i] = (MPI_Win) tmp_buf[k++];
}
fn_exit:
MPIU_CHKLMEM_FREEALL();
MPIDI_RMA_FUNC_EXIT(MPID_STATE_MPIDI_WIN_CREATE);
return mpi_errno;
/* --BEGIN ERROR HANDLING-- */
fn_fail:
MPIU_CHKPMEM_REAP();
goto fn_exit;
/* --END ERROR HANDLING-- */
}
int MPIDI_CH3U_Win_create_gather( void *base, MPI_Aint size, int disp_unit,
MPID_Info *info, MPID_Comm *comm_ptr, MPID_Win **win_ptr )
{
int mpi_errno=MPI_SUCCESS, i, k, comm_size, rank;
MPI_Aint *tmp_buf;
int errflag = FALSE;
MPIU_CHKPMEM_DECL(5);
MPIU_CHKLMEM_DECL(1);
MPIDI_STATE_DECL(MPID_STATE_MPIDI_CH3U_WIN_CREATE_GATHER);
MPIDI_RMA_FUNC_ENTER(MPID_STATE_MPIDI_CH3U_WIN_CREATE_GATHER);
comm_size = (*win_ptr)->comm_ptr->local_size;
rank = (*win_ptr)->comm_ptr->rank;
/* RMA handlers should be set before calling this function */
mpi_errno = (*win_ptr)->RMAFns.Win_set_info(*win_ptr, info);
MPIU_INSTR_DURATION_START(wincreate_allgather);
/* allocate memory for the base addresses, disp_units, and
completion counters of all processes */
MPIU_CHKPMEM_MALLOC((*win_ptr)->base_addrs, void **,
comm_size*sizeof(void *),
mpi_errno, "(*win_ptr)->base_addrs");
MPIU_CHKPMEM_MALLOC((*win_ptr)->sizes, MPI_Aint *, comm_size*sizeof(MPI_Aint),
mpi_errno, "(*win_ptr)->sizes");
MPIU_CHKPMEM_MALLOC((*win_ptr)->disp_units, int *, comm_size*sizeof(int),
mpi_errno, "(*win_ptr)->disp_units");
MPIU_CHKPMEM_MALLOC((*win_ptr)->all_win_handles, MPI_Win *,
comm_size*sizeof(MPI_Win),
mpi_errno, "(*win_ptr)->all_win_handles");
MPIU_CHKPMEM_MALLOC((*win_ptr)->pt_rma_puts_accs, int *,
comm_size*sizeof(int),
mpi_errno, "(*win_ptr)->pt_rma_puts_accs");
for (i=0; i<comm_size; i++) (*win_ptr)->pt_rma_puts_accs[i] = 0;
/* get the addresses of the windows, window objects, and completion
counters of all processes. allocate temp. buffer for communication */
MPIU_CHKLMEM_MALLOC(tmp_buf, MPI_Aint *, 4*comm_size*sizeof(MPI_Aint),
mpi_errno, "tmp_buf");
/* FIXME: This needs to be fixed for heterogeneous systems */
/* FIXME: If we wanted to validate the transfer as within range at the
origin, we'd also need the window size. */
tmp_buf[4*rank] = MPIU_PtrToAint(base);
tmp_buf[4*rank+1] = size;
tmp_buf[4*rank+2] = (MPI_Aint) disp_unit;
tmp_buf[4*rank+3] = (MPI_Aint) (*win_ptr)->handle;
mpi_errno = MPIR_Allgather_impl(MPI_IN_PLACE, 0, MPI_DATATYPE_NULL,
tmp_buf, 4, MPI_AINT,
(*win_ptr)->comm_ptr, &errflag);
MPIU_INSTR_DURATION_END(wincreate_allgather);
if (mpi_errno) { MPIU_ERR_POP(mpi_errno); }
MPIU_ERR_CHKANDJUMP(errflag, mpi_errno, MPI_ERR_OTHER, "**coll_fail");
k = 0;
for (i=0; i<comm_size; i++)
{
(*win_ptr)->base_addrs[i] = MPIU_AintToPtr(tmp_buf[k++]);
(*win_ptr)->sizes[i] = tmp_buf[k++];
(*win_ptr)->disp_units[i] = (int) tmp_buf[k++];
(*win_ptr)->all_win_handles[i] = (MPI_Win) tmp_buf[k++];
}
fn_exit:
MPIU_CHKLMEM_FREEALL();
MPIDI_RMA_FUNC_EXIT(MPID_STATE_MPIDI_CH3U_WIN_CREATE_GATHER);
return mpi_errno;
/* --BEGIN ERROR HANDLING-- */
fn_fail:
MPIU_CHKPMEM_REAP();
goto fn_exit;
/* --END ERROR HANDLING-- */
}
/*@
MPI_Type_create_subarray - Create a datatype for a subarray of a regular,
multidimensional array
Input Parameters:
+ ndims - number of array dimensions (positive integer)
. array_of_sizes - number of elements of type oldtype in each dimension of the
full array (array of positive integers)
. array_of_subsizes - number of elements of type oldtype in each dimension of
the subarray (array of positive integers)
. array_of_starts - starting coordinates of the subarray in each dimension
(array of nonnegative integers)
. order - array storage order flag (state)
- oldtype - array element datatype (handle)
Output Parameters:
. newtype - new datatype (handle)
.N ThreadSafe
.N Fortran
.N Errors
.N MPI_SUCCESS
.N MPI_ERR_TYPE
.N MPI_ERR_ARG
@*/
int MPI_Type_create_subarray(int ndims,
const int array_of_sizes[],
const int array_of_subsizes[],
const int array_of_starts[],
int order,
MPI_Datatype oldtype,
MPI_Datatype *newtype)
{
static const char FCNAME[] = "MPI_Type_create_subarray";
int mpi_errno = MPI_SUCCESS, i;
MPI_Datatype new_handle;
/* these variables are from the original version in ROMIO */
MPI_Aint size, extent, disps[3];
MPI_Datatype tmp1, tmp2;
# ifdef HAVE_ERROR_CHECKING
MPI_Aint size_with_aint;
MPI_Offset size_with_offset;
# endif
/* for saving contents */
int *ints;
MPID_Datatype *new_dtp;
MPIU_CHKLMEM_DECL(1);
MPID_MPI_STATE_DECL(MPID_STATE_MPI_TYPE_CREATE_SUBARRAY);
MPIR_ERRTEST_INITIALIZED_ORDIE();
MPID_THREAD_CS_ENTER(GLOBAL, MPIR_THREAD_GLOBAL_ALLFUNC_MUTEX);
MPID_MPI_FUNC_ENTER(MPID_STATE_MPI_TYPE_CREATE_SUBARRAY);
# ifdef HAVE_ERROR_CHECKING
{
MPID_BEGIN_ERROR_CHECKS;
{
MPID_Datatype *datatype_ptr = NULL;
/* Check parameters */
MPIR_ERRTEST_ARGNONPOS(ndims, "ndims", mpi_errno, MPI_ERR_DIMS);
MPIR_ERRTEST_ARGNULL(array_of_sizes, "array_of_sizes", mpi_errno);
MPIR_ERRTEST_ARGNULL(array_of_subsizes, "array_of_subsizes", mpi_errno);
MPIR_ERRTEST_ARGNULL(array_of_starts, "array_of_starts", mpi_errno);
for (i=0; mpi_errno == MPI_SUCCESS && i < ndims; i++) {
MPIR_ERRTEST_ARGNONPOS(array_of_sizes[i], "size", mpi_errno, MPI_ERR_ARG);
MPIR_ERRTEST_ARGNONPOS(array_of_subsizes[i], "subsize", mpi_errno, MPI_ERR_ARG);
MPIR_ERRTEST_ARGNEG(array_of_starts[i], "start", mpi_errno);
if (array_of_subsizes[i] > array_of_sizes[i]) {
mpi_errno = MPIR_Err_create_code(MPI_SUCCESS,
MPIR_ERR_RECOVERABLE,
FCNAME,
__LINE__,
MPI_ERR_ARG,
"**argrange",
"**argrange %s %d %d",
"array_of_subsizes",
array_of_subsizes[i],
array_of_sizes[i]);
goto fn_fail;
}
if (array_of_starts[i] > (array_of_sizes[i] - array_of_subsizes[i]))
{
mpi_errno = MPIR_Err_create_code(MPI_SUCCESS,
MPIR_ERR_RECOVERABLE,
FCNAME,
__LINE__,
MPI_ERR_ARG,
"**argrange",
"**argrange %s %d %d",
"array_of_starts",
array_of_starts[i],
array_of_sizes[i] -
array_of_subsizes[i]);
goto fn_fail;
}
}
if (order != MPI_ORDER_FORTRAN && order != MPI_ORDER_C) {
mpi_errno = MPIR_Err_create_code(MPI_SUCCESS,
MPIR_ERR_RECOVERABLE,
FCNAME,
__LINE__,
MPI_ERR_ARG,
"**arg",
"**arg %s",
"order");
goto fn_fail;
}
MPIR_Type_extent_impl(oldtype, &extent);
/* check if MPI_Aint is large enough for size of global array.
if not, complain. */
size_with_aint = extent;
for (i=0; i<ndims; i++) size_with_aint *= array_of_sizes[i];
size_with_offset = extent;
for (i=0; i<ndims; i++) size_with_offset *= array_of_sizes[i];
if (size_with_aint != size_with_offset) {
mpi_errno = MPIR_Err_create_code(MPI_SUCCESS,
MPIR_ERR_FATAL,
FCNAME,
__LINE__,
MPI_ERR_ARG,
"**subarrayoflow",
"**subarrayoflow %L",
size_with_offset);
goto fn_fail;
}
/* Get handles to MPI objects. */
MPID_Datatype_get_ptr(oldtype, datatype_ptr);
/* Validate datatype_ptr */
MPID_Datatype_valid_ptr(datatype_ptr, mpi_errno);
/* If datatype_ptr is not valid, it will be reset to null */
if (mpi_errno != MPI_SUCCESS) goto fn_fail;
}
MPID_END_ERROR_CHECKS;
}
# endif /* HAVE_ERROR_CHECKING */
/* ... body of routine ... */
/* TODO: CHECK THE ERROR RETURNS FROM ALL THESE!!! */
/* TODO: GRAB EXTENT WITH A MACRO OR SOMETHING FASTER */
MPIR_Type_extent_impl(oldtype, &extent);
if (order == MPI_ORDER_FORTRAN) {
if (ndims == 1)
mpi_errno = MPID_Type_contiguous(array_of_subsizes[0],
oldtype,
&tmp1);
else {
mpi_errno = MPID_Type_vector(array_of_subsizes[1],
array_of_subsizes[0],
(MPI_Aint)(array_of_sizes[0]),
0, /* stride in types */
oldtype,
&tmp1);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
size = ((MPI_Aint)(array_of_sizes[0])) * extent;
for (i=2; i<ndims; i++) {
size *= (MPI_Aint)(array_of_sizes[i-1]);
mpi_errno = MPID_Type_vector(array_of_subsizes[i],
1,
size,
1, /* stride in bytes */
tmp1,
&tmp2);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
MPIR_Type_free_impl(&tmp1);
tmp1 = tmp2;
}
}
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
/* add displacement and UB */
disps[1] = (MPI_Aint)(array_of_starts[0]);
size = 1;
for (i=1; i<ndims; i++) {
size *= (MPI_Aint)(array_of_sizes[i-1]);
disps[1] += size * (MPI_Aint)(array_of_starts[i]);
}
/* rest done below for both Fortran and C order */
}
else /* MPI_ORDER_C */ {
/* dimension ndims-1 changes fastest */
if (ndims == 1) {
mpi_errno = MPID_Type_contiguous(array_of_subsizes[0],
oldtype,
&tmp1);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
}
else {
mpi_errno = MPID_Type_vector(array_of_subsizes[ndims-2],
array_of_subsizes[ndims-1],
(MPI_Aint)(array_of_sizes[ndims-1]),
0, /* stride in types */
oldtype,
&tmp1);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
size = (MPI_Aint)(array_of_sizes[ndims-1]) * extent;
for (i=ndims-3; i>=0; i--) {
size *= (MPI_Aint)(array_of_sizes[i+1]);
mpi_errno = MPID_Type_vector(array_of_subsizes[i],
1, /* blocklen */
size, /* stride */
1, /* stride in bytes */
tmp1, /* old type */
&tmp2);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
MPIR_Type_free_impl(&tmp1);
tmp1 = tmp2;
}
}
/* add displacement and UB */
disps[1] = (MPI_Aint)(array_of_starts[ndims-1]);
size = 1;
for (i=ndims-2; i>=0; i--) {
size *= (MPI_Aint)(array_of_sizes[i+1]);
disps[1] += size * (MPI_Aint)(array_of_starts[i]);
}
}
disps[1] *= extent;
disps[2] = extent;
for (i=0; i<ndims; i++) disps[2] *= (MPI_Aint)(array_of_sizes[i]);
disps[0] = 0;
/* Instead of using MPI_LB/MPI_UB, which have been removed from MPI in MPI-3,
use MPI_Type_create_resized. Use hindexed_block to set the starting displacement
of the datatype (disps[1]) and type_create_resized to set lb to 0 (disps[0])
and extent to disps[2], which makes ub = disps[2].
*/
mpi_errno = MPID_Type_blockindexed(1, 1, &disps[1],
1, /* 1 means disp is in bytes */
tmp1, &tmp2);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
mpi_errno = MPID_Type_create_resized(tmp2, 0, disps[2], &new_handle);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
MPIR_Type_free_impl(&tmp1);
MPIR_Type_free_impl(&tmp2);
/* at this point we have the new type, and we've cleaned up any
* intermediate types created in the process. we just need to save
* all our contents/envelope information.
*/
/* Save contents */
MPIU_CHKLMEM_MALLOC_ORJUMP(ints, int *, (3 * ndims + 2) * sizeof(int), mpi_errno, "content description");
ints[0] = ndims;
for (i=0; i < ndims; i++) {
ints[i + 1] = array_of_sizes[i];
}
for(i=0; i < ndims; i++) {
ints[i + ndims + 1] = array_of_subsizes[i];
}
for(i=0; i < ndims; i++) {
ints[i + 2*ndims + 1] = array_of_starts[i];
}
ints[3*ndims + 1] = order;
MPID_Datatype_get_ptr(new_handle, new_dtp);
mpi_errno = MPID_Datatype_set_contents(new_dtp,
MPI_COMBINER_SUBARRAY,
3 * ndims + 2, /* ints */
0, /* aints */
1, /* types */
ints,
NULL,
&oldtype);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
MPIR_OBJ_PUBLISH_HANDLE(*newtype, new_handle);
/* ... end of body of routine ... */
fn_exit:
MPIU_CHKLMEM_FREEALL();
MPID_MPI_FUNC_EXIT(MPID_STATE_MPI_TYPE_CREATE_SUBARRAY);
MPID_THREAD_CS_EXIT(GLOBAL, MPIR_THREAD_GLOBAL_ALLFUNC_MUTEX);
return mpi_errno;
fn_fail:
/* --BEGIN ERROR HANDLING-- */
# ifdef HAVE_ERROR_CHECKING
{
mpi_errno = MPIR_Err_create_code(
mpi_errno, MPIR_ERR_RECOVERABLE, FCNAME, __LINE__, MPI_ERR_OTHER, "**mpi_type_create_subarray",
"**mpi_type_create_subarray %d %p %p %p %d %D %p", ndims, array_of_sizes, array_of_subsizes,
array_of_starts, order, oldtype, newtype);
}
# endif
mpi_errno = MPIR_Err_return_comm(NULL, FCNAME, mpi_errno);
goto fn_exit;
/* --END ERROR HANDLING-- */
}
/*@
MPI_Type_hindexed - Creates an indexed datatype with offsets in bytes
Input Parameters:
+ count - number of blocks -- also number of entries in indices and blocklens
. blocklens - number of elements in each block (array of nonnegative integers)
. indices - byte displacement of each block (array of MPI_Aint)
- old_type - old datatype (handle)
Output Parameter:
. newtype - new datatype (handle)
.N Deprecated
This routine is replaced by 'MPI_Type_create_hindexed'.
.N ThreadSafe
.N Fortran
The indices are displacements, and are based on a zero origin. A common error
is to do something like to following
.vb
integer a(100)
integer blens(10), indices(10)
do i=1,10
blens(i) = 1
10 indices(i) = (1 + (i-1)*10) * sizeofint
call MPI_TYPE_HINDEXED(10,blens,indices,MPI_INTEGER,newtype,ierr)
call MPI_TYPE_COMMIT(newtype,ierr)
call MPI_SEND(a,1,newtype,...)
.ve
expecting this to send 'a(1),a(11),...' because the indices have values
'1,11,...'. Because these are `displacements` from the beginning of 'a',
it actually sends 'a(1+1),a(1+11),...'.
If you wish to consider the displacements as indices into a Fortran array,
consider declaring the Fortran array with a zero origin
.vb
integer a(0:99)
.ve
.N Errors
.N MPI_SUCCESS
.N MPI_ERR_TYPE
.N MPI_ERR_COUNT
.N MPI_ERR_EXHAUSTED
.N MPI_ERR_ARG
@*/
int MPI_Type_hindexed(int count,
int blocklens[],
MPI_Aint indices[],
MPI_Datatype old_type,
MPI_Datatype *newtype)
{
static const char FCNAME[] = "MPI_Type_hindexed";
int mpi_errno = MPI_SUCCESS;
MPI_Datatype new_handle;
MPID_Datatype *new_dtp;
int i, *ints;
MPIU_CHKLMEM_DECL(1);
MPID_MPI_STATE_DECL(MPID_STATE_MPI_TYPE_HINDEXED);
MPIR_ERRTEST_INITIALIZED_ORDIE();
MPIU_THREAD_CS_ENTER(ALLFUNC,);
MPID_MPI_FUNC_ENTER(MPID_STATE_MPI_TYPE_HINDEXED);
# ifdef HAVE_ERROR_CHECKING
{
MPID_BEGIN_ERROR_CHECKS;
{
int j;
MPID_Datatype *datatype_ptr = NULL;
MPIR_ERRTEST_COUNT(count, mpi_errno);
MPIR_ERRTEST_DATATYPE(old_type, "datatype", mpi_errno);
if (count > 0) {
MPIR_ERRTEST_ARGNULL(blocklens, "blocklens", mpi_errno);
MPIR_ERRTEST_ARGNULL(indices, "indices", mpi_errno);
}
if (mpi_errno == MPI_SUCCESS) {
if (HANDLE_GET_KIND(old_type) != HANDLE_KIND_BUILTIN) {
MPID_Datatype_get_ptr( old_type, datatype_ptr );
MPID_Datatype_valid_ptr(datatype_ptr, mpi_errno);
}
/* verify that all blocklengths are >= 0 */
for (j=0; j < count; j++) {
MPIR_ERRTEST_ARGNEG(blocklens[j], "blocklen", mpi_errno);
}
}
MPIR_ERRTEST_ARGNULL(newtype, "newtype", mpi_errno);
if (mpi_errno != MPI_SUCCESS) goto fn_fail;
}
MPID_END_ERROR_CHECKS;
}
# endif /* HAVE_ERROR_CHECKING */
/* ... body of routine ... */
mpi_errno = MPID_Type_indexed(count,
blocklens,
indices,
1, /* displacements in bytes */
old_type,
&new_handle);
if (mpi_errno != MPI_SUCCESS) goto fn_fail;
MPIU_CHKLMEM_MALLOC(ints, int *, (count + 1) * sizeof(int), mpi_errno, "contents integer array");
/* copy ints into temporary buffer (count and blocklengths) */
ints[0] = count;
for (i=0; i < count; i++)
{
ints[i+1] = blocklens[i];
}
MPID_Datatype_get_ptr(new_handle, new_dtp);
mpi_errno = MPID_Datatype_set_contents(new_dtp,
MPI_COMBINER_HINDEXED,
count+1, /* ints */
count, /* aints (displs) */
1, /* types */
ints,
indices,
&old_type);
if (mpi_errno != MPI_SUCCESS) goto fn_fail;
MPIU_OBJ_PUBLISH_HANDLE(*newtype, new_handle);
/* ... end of body of routine ... */
fn_exit:
MPIU_CHKLMEM_FREEALL();
MPID_MPI_FUNC_EXIT(MPID_STATE_MPI_TYPE_HINDEXED);
MPIU_THREAD_CS_EXIT(ALLFUNC,);
return mpi_errno;
fn_fail:
/* --BEGIN ERROR HANDLING-- */
# ifdef HAVE_ERROR_CHECKING
{
mpi_errno = MPIR_Err_create_code(
mpi_errno, MPIR_ERR_RECOVERABLE, FCNAME, __LINE__, MPI_ERR_OTHER, "**mpi_type_hindexed",
"**mpi_type_hindexed %d %p %p %D %p", count, blocklens, indices, old_type, newtype);
}
# endif
mpi_errno = MPIR_Err_return_comm( NULL, FCNAME, mpi_errno );
goto fn_exit;
/* --END ERROR HANDLING-- */
}
int MPIR_Localcopy(const void *sendbuf, MPI_Aint sendcount, MPI_Datatype sendtype,
void *recvbuf, MPI_Aint recvcount, MPI_Datatype recvtype)
{
int mpi_errno = MPI_SUCCESS;
int sendtype_iscontig, recvtype_iscontig;
MPI_Aint sendsize, recvsize, sdata_sz, rdata_sz, copy_sz;
MPI_Aint true_extent, sendtype_true_lb, recvtype_true_lb;
MPIU_CHKLMEM_DECL(1);
MPID_MPI_STATE_DECL(MPID_STATE_MPIR_LOCALCOPY);
MPID_MPI_FUNC_ENTER(MPID_STATE_MPIR_LOCALCOPY);
MPID_Datatype_get_size_macro(sendtype, sendsize);
MPID_Datatype_get_size_macro(recvtype, recvsize);
sdata_sz = sendsize * sendcount;
rdata_sz = recvsize * recvcount;
/* if there is no data to copy, bail out */
if (!sdata_sz || !rdata_sz)
goto fn_exit;
#if defined(HAVE_ERROR_CHECKING)
if (sdata_sz > rdata_sz) {
MPIU_ERR_SET2(mpi_errno, MPI_ERR_TRUNCATE, "**truncate", "**truncate %d %d", sdata_sz, rdata_sz);
copy_sz = rdata_sz;
}
else
#endif /* HAVE_ERROR_CHECKING */
copy_sz = sdata_sz;
/* Builtin types is the common case; optimize for it */
if ((HANDLE_GET_KIND(sendtype) == HANDLE_KIND_BUILTIN) &&
HANDLE_GET_KIND(recvtype) == HANDLE_KIND_BUILTIN) {
MPIU_Memcpy(recvbuf, sendbuf, copy_sz);
goto fn_exit;
}
MPIR_Datatype_iscontig(sendtype, &sendtype_iscontig);
MPIR_Datatype_iscontig(recvtype, &recvtype_iscontig);
MPIR_Type_get_true_extent_impl(sendtype, &sendtype_true_lb, &true_extent);
MPIR_Type_get_true_extent_impl(recvtype, &recvtype_true_lb, &true_extent);
if (sendtype_iscontig && recvtype_iscontig)
{
#if defined(HAVE_ERROR_CHECKING)
MPIU_ERR_CHKMEMCPYANDJUMP(mpi_errno,
((char *)recvbuf + recvtype_true_lb),
((char *)sendbuf + sendtype_true_lb),
copy_sz);
#endif
MPIU_Memcpy(((char *) recvbuf + recvtype_true_lb),
((char *) sendbuf + sendtype_true_lb),
copy_sz);
}
else if (sendtype_iscontig)
{
MPID_Segment seg;
MPI_Aint last;
MPID_Segment_init(recvbuf, recvcount, recvtype, &seg, 0);
last = copy_sz;
MPID_Segment_unpack(&seg, 0, &last, (char*)sendbuf + sendtype_true_lb);
MPIU_ERR_CHKANDJUMP(last != copy_sz, mpi_errno, MPI_ERR_TYPE, "**dtypemismatch");
}
else if (recvtype_iscontig)
{
MPID_Segment seg;
MPI_Aint last;
MPID_Segment_init(sendbuf, sendcount, sendtype, &seg, 0);
last = copy_sz;
MPID_Segment_pack(&seg, 0, &last, (char*)recvbuf + recvtype_true_lb);
MPIU_ERR_CHKANDJUMP(last != copy_sz, mpi_errno, MPI_ERR_TYPE, "**dtypemismatch");
}
else
{
char * buf;
MPIDI_msg_sz_t buf_off;
MPID_Segment sseg;
MPIDI_msg_sz_t sfirst;
MPID_Segment rseg;
MPIDI_msg_sz_t rfirst;
MPIU_CHKLMEM_MALLOC(buf, char *, COPY_BUFFER_SZ, mpi_errno, "buf");
MPID_Segment_init(sendbuf, sendcount, sendtype, &sseg, 0);
MPID_Segment_init(recvbuf, recvcount, recvtype, &rseg, 0);
sfirst = 0;
rfirst = 0;
buf_off = 0;
while (1)
{
MPI_Aint last;
char * buf_end;
if (copy_sz - sfirst > COPY_BUFFER_SZ - buf_off)
{
last = sfirst + (COPY_BUFFER_SZ - buf_off);
}
else
{
last = copy_sz;
}
MPID_Segment_pack(&sseg, sfirst, &last, buf + buf_off);
MPIU_Assert(last > sfirst);
buf_end = buf + buf_off + (last - sfirst);
sfirst = last;
MPID_Segment_unpack(&rseg, rfirst, &last, buf);
MPIU_Assert(last > rfirst);
rfirst = last;
if (rfirst == copy_sz)
{
/* successful completion */
break;
}
/* if the send side finished, but the recv side couldn't unpack it, there's a datatype mismatch */
MPIU_ERR_CHKANDJUMP(sfirst == copy_sz, mpi_errno, MPI_ERR_TYPE, "**dtypemismatch");
/* if not all data was unpacked, copy it to the front of the buffer for next time */
buf_off = sfirst - rfirst;
if (buf_off > 0)
{
memmove(buf, buf_end - buf_off, buf_off);
}
}
}
fn_exit:
MPIU_CHKLMEM_FREEALL();
MPID_MPI_FUNC_EXIT(MPID_STATE_MPIR_LOCALCOPY);
return mpi_errno;
fn_fail:
goto fn_exit;
}
