int MPIR_Barrier_intra_smp(MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
{
int mpi_errno = MPI_SUCCESS;
int mpi_errno_ret = MPI_SUCCESS;
MPIR_Assert(MPIR_CVAR_ENABLE_SMP_COLLECTIVES && MPIR_CVAR_ENABLE_SMP_BARRIER &&
MPIR_Comm_is_node_aware(comm_ptr));
/* do the intranode barrier on all nodes */
if (comm_ptr->node_comm != NULL) {
mpi_errno = MPIR_Barrier(comm_ptr->node_comm, errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag =
MPIX_ERR_PROC_FAILED ==
MPIR_ERR_GET_CLASS(mpi_errno) ? MPIR_ERR_PROC_FAILED : MPIR_ERR_OTHER;
MPIR_ERR_SET(mpi_errno, *errflag, "**fail");
MPIR_ERR_ADD(mpi_errno_ret, mpi_errno);
}
}
/* do the barrier across roots of all nodes */
if (comm_ptr->node_roots_comm != NULL) {
mpi_errno = MPIR_Barrier(comm_ptr->node_roots_comm, errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag =
MPIX_ERR_PROC_FAILED ==
MPIR_ERR_GET_CLASS(mpi_errno) ? MPIR_ERR_PROC_FAILED : MPIR_ERR_OTHER;
MPIR_ERR_SET(mpi_errno, *errflag, "**fail");
MPIR_ERR_ADD(mpi_errno_ret, mpi_errno);
}
}
/* release the local processes on each node with a 1-byte
* broadcast (0-byte broadcast just returns without doing
* anything) */
if (comm_ptr->node_comm != NULL) {
int i = 0;
mpi_errno = MPIR_Bcast(&i, 1, MPI_BYTE, 0, comm_ptr->node_comm, errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag =
MPIX_ERR_PROC_FAILED ==
MPIR_ERR_GET_CLASS(mpi_errno) ? MPIR_ERR_PROC_FAILED : MPIR_ERR_OTHER;
MPIR_ERR_SET(mpi_errno, *errflag, "**fail");
MPIR_ERR_ADD(mpi_errno_ret, mpi_errno);
}
}
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;
}
int MPIR_Barrier_intra( MPID_Comm *comm_ptr, MPIR_Errflag_t *errflag )
{
int size, rank, src, dst, mask, mpi_errno=MPI_SUCCESS;
int mpi_errno_ret = MPI_SUCCESS;
/* Only one collective operation per communicator can be active at any
time */
MPIDU_ERR_CHECK_MULTIPLE_THREADS_ENTER( comm_ptr );
size = comm_ptr->local_size;
/* Trivial barriers return immediately */
if (size == 1) goto fn_exit;
if (MPIR_CVAR_ENABLE_SMP_COLLECTIVES && MPIR_CVAR_ENABLE_SMP_BARRIER &&
MPIR_Comm_is_node_aware(comm_ptr)) {
mpi_errno = barrier_smp_intra(comm_ptr, errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag = MPIR_ERR_GET_CLASS(mpi_errno);
MPIR_ERR_SET(mpi_errno, *errflag, "**fail");
MPIR_ERR_ADD(mpi_errno_ret, mpi_errno);
}
goto fn_exit;
}
rank = comm_ptr->rank;
mask = 0x1;
while (mask < size) {
dst = (rank + mask) % size;
src = (rank - mask + size) % size;
mpi_errno = MPIC_Sendrecv(NULL, 0, MPI_BYTE, dst,
MPIR_BARRIER_TAG, NULL, 0, MPI_BYTE,
src, MPIR_BARRIER_TAG, comm_ptr,
MPI_STATUS_IGNORE, errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag = MPIR_ERR_GET_CLASS(mpi_errno);
MPIR_ERR_SET(mpi_errno, *errflag, "**fail");
MPIR_ERR_ADD(mpi_errno_ret, mpi_errno);
}
mask <<= 1;
}
fn_exit:
MPIDU_ERR_CHECK_MULTIPLE_THREADS_EXIT( comm_ptr );
if (mpi_errno_ret)
mpi_errno = mpi_errno_ret;
else if (*errflag != MPIR_ERR_NONE)
MPIR_ERR_SET(mpi_errno, *errflag, "**coll_fail");
return mpi_errno;
fn_fail:
goto fn_exit;
}
int MPIR_Gather_inter_linear(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
void *recvbuf, int recvcount, MPI_Datatype recvtype, int root,
MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
{
int remote_size, mpi_errno = MPI_SUCCESS;
int mpi_errno_ret = MPI_SUCCESS;
int i;
MPI_Status status;
MPI_Aint extent;
if (root == MPI_PROC_NULL) {
/* local processes other than root do nothing */
return MPI_SUCCESS;
}
remote_size = comm_ptr->remote_size;
if (root == MPI_ROOT) {
MPIR_Datatype_get_extent_macro(recvtype, extent);
MPIR_Ensure_Aint_fits_in_pointer(MPIR_VOID_PTR_CAST_TO_MPI_AINT recvbuf +
(recvcount * remote_size * extent));
for (i = 0; i < remote_size; i++) {
mpi_errno =
MPIC_Recv(((char *) recvbuf + recvcount * i * extent), recvcount, recvtype, i,
MPIR_GATHER_TAG, comm_ptr, &status, errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag =
MPIX_ERR_PROC_FAILED ==
MPIR_ERR_GET_CLASS(mpi_errno) ? MPIR_ERR_PROC_FAILED : MPIR_ERR_OTHER;
MPIR_ERR_SET(mpi_errno, *errflag, "**fail");
MPIR_ERR_ADD(mpi_errno_ret, mpi_errno);
}
}
} else {
mpi_errno =
MPIC_Send(sendbuf, sendcount, sendtype, root, MPIR_GATHER_TAG, comm_ptr, errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag =
MPIX_ERR_PROC_FAILED ==
MPIR_ERR_GET_CLASS(mpi_errno) ? MPIR_ERR_PROC_FAILED : MPIR_ERR_OTHER;
MPIR_ERR_SET(mpi_errno, *errflag, "**fail");
MPIR_ERR_ADD(mpi_errno_ret, mpi_errno);
}
}
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;
}
/* This function allocates and generates a template_tree based on k_val and right_skewed for
* max_ranks.
* */
int MPIDI_SHM_create_template_tree(MPIDI_SHM_topotree_t * template_tree, int k_val,
bool right_skewed, int max_ranks, MPIR_Errflag_t * errflag)
{
int mpi_errno = MPI_SUCCESS, mpi_errno_ret = MPI_SUCCESS;
int i, j, child_id, child_idx;
mpi_errno = MPIDI_SHM_topotree_allocate(template_tree, max_ranks, k_val);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag =
MPIX_ERR_PROC_FAILED ==
MPIR_ERR_GET_CLASS(mpi_errno) ? MPIR_ERR_PROC_FAILED : MPIR_ERR_OTHER;
MPIR_ERR_SET(mpi_errno, *errflag, "**fail");
MPIR_ERR_ADD(mpi_errno_ret, mpi_errno);
}
for (i = 0; i < max_ranks; ++i) {
MPIDI_SHM_TOPOTREE_PARENT(template_tree, i) = ceilf(i / (float) (k_val)) - 1;
MPIDI_SHM_TOPOTREE_NUM_CHILD(template_tree, i) = 0;
if (!right_skewed) {
for (j = 0; j < k_val; ++j) {
child_id = i * k_val + 1 + j;
if (child_id < max_ranks) {
child_idx = MPIDI_SHM_TOPOTREE_NUM_CHILD(template_tree, i)++;
MPIDI_SHM_TOPOTREE_CHILD(template_tree, i, child_idx) = child_id;
}
}
} else if (right_skewed) {
for (j = k_val - 1; j >= 0; --j) {
child_id = i * k_val + 1 + j;
if (child_id < max_ranks) {
child_idx = MPIDI_SHM_TOPOTREE_NUM_CHILD(template_tree, i)++;
MPIDI_SHM_TOPOTREE_CHILD(template_tree, i, child_idx) = child_id;
}
}
}
}
MPIDI_SHM_TOPOTREE_PARENT(template_tree, 0) = -1;
if (MPIDI_SHM_TOPOTREE_DEBUG) {
fprintf(stderr, "TemplateTree, %d\n", max_ranks);
MPIDI_SHM_print_topotree("TemplateTree", template_tree);
for (i = 0; i < max_ranks; ++i) {
fprintf(stderr, "TemplateR, %d, P=%d, C=%d, [", i,
MPIDI_SHM_TOPOTREE_PARENT(template_tree, i),
MPIDI_SHM_TOPOTREE_NUM_CHILD(template_tree, i));
for (j = 0; j < MPIDI_SHM_TOPOTREE_NUM_CHILD(template_tree, i); ++j) {
fprintf(stderr, "%d, ", MPIDI_SHM_TOPOTREE_CHILD(template_tree, i, j));
}
fprintf(stderr, "]\n");
}
}
/* template tree is ready here */
return mpi_errno;
}
int MPIC_Ssend(const void *buf, MPI_Aint count, MPI_Datatype datatype, int dest, int tag,
MPID_Comm *comm_ptr, mpir_errflag_t *errflag)
{
int mpi_errno = MPI_SUCCESS;
int context_id;
MPID_Request *request_ptr = NULL;
MPIDI_STATE_DECL(MPID_STATE_MPIC_SSEND);
MPIDI_FUNC_ENTER(MPID_STATE_MPIC_SSEND);
MPIU_DBG_MSG_D(PT2PT, TYPICAL, "IN: errflag = %d", *errflag);
MPIU_ERR_CHKANDJUMP1((count < 0), mpi_errno, MPI_ERR_COUNT,
"**countneg", "**countneg %d", count);
context_id = (comm_ptr->comm_kind == MPID_INTRACOMM) ?
MPID_CONTEXT_INTRA_COLL : MPID_CONTEXT_INTER_COLL;
switch(*errflag) {
case MPIR_ERR_NONE:
break;
case MPIR_ERR_PROC_FAILED:
MPIR_TAG_SET_PROC_FAILURE_BIT(tag);
default:
MPIR_TAG_SET_ERROR_BIT(tag);
}
mpi_errno = MPID_Ssend(buf, count, datatype, dest, tag, comm_ptr,
context_id, &request_ptr);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
if (request_ptr) {
mpi_errno = MPIC_Wait(request_ptr, errflag);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
MPID_Request_release(request_ptr);
}
fn_exit:
MPIU_DBG_MSG_D(PT2PT, TYPICAL, "OUT: errflag = %d", *errflag);
MPIDI_FUNC_EXIT(MPID_STATE_MPIC_SSEND);
return mpi_errno;
fn_fail:
/* --BEGIN ERROR HANDLING-- */
if (request_ptr) MPID_Request_release(request_ptr);
if (mpi_errno && !*errflag) {
if (MPIX_ERR_PROC_FAILED == MPIR_ERR_GET_CLASS(mpi_errno)) {
*errflag = MPIR_ERR_PROC_FAILED;
} else {
*errflag = MPIR_ERR_OTHER;
}
}
goto fn_exit;
/* --END ERROR HANDLING-- */
}
int MPIC_Recv(void *buf, MPI_Aint count, MPI_Datatype datatype, int source, int tag,
MPID_Comm *comm_ptr, MPI_Status *status, mpir_errflag_t *errflag)
{
int mpi_errno = MPI_SUCCESS;
int context_id;
MPI_Status mystatus;
MPID_Request *request_ptr = NULL;
MPIDI_STATE_DECL(MPID_STATE_MPIC_RECV);
MPIDI_FUNC_ENTER(MPID_STATE_MPIC_RECV);
MPIU_DBG_MSG_D(PT2PT, TYPICAL, "IN: errflag = %d", *errflag);
MPIU_ERR_CHKANDJUMP1((count < 0), mpi_errno, MPI_ERR_COUNT,
"**countneg", "**countneg %d", count);
context_id = (comm_ptr->comm_kind == MPID_INTRACOMM) ?
MPID_CONTEXT_INTRA_COLL : MPID_CONTEXT_INTER_COLL;
if (status == MPI_STATUS_IGNORE)
status = &mystatus;
mpi_errno = MPID_Recv(buf, count, datatype, source, tag, comm_ptr,
context_id, status, &request_ptr);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
if (request_ptr) {
mpi_errno = MPIC_Wait(request_ptr, errflag);
if (mpi_errno != MPI_SUCCESS)
MPIU_ERR_POP(mpi_errno);
*status = request_ptr->status;
mpi_errno = status->MPI_ERROR;
MPID_Request_release(request_ptr);
} else {
MPIR_Process_status(status, errflag);
MPIR_TAG_CLEAR_ERROR_BITS(status->MPI_TAG);
}
if (MPI_SUCCESS == MPIR_ERR_GET_CLASS(status->MPI_ERROR)) {
MPIU_Assert(status->MPI_TAG == tag);
}
fn_exit:
MPIU_DBG_MSG_D(PT2PT, TYPICAL, "OUT: errflag = %d", *errflag);
MPIDI_FUNC_EXIT(MPID_STATE_MPIC_RECV);
return mpi_errno;
fn_fail:
/* --BEGIN ERROR HANDLING-- */
if (request_ptr) MPID_Request_release(request_ptr);
goto fn_exit;
/* --END ERROR HANDLING-- */
}
int MPIR_Bcast_intra_auto(void *buffer,
int count,
MPI_Datatype datatype,
int root, MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
{
int mpi_errno = MPI_SUCCESS;
int mpi_errno_ret = MPI_SUCCESS;
int comm_size;
MPI_Aint nbytes = 0;
MPI_Aint type_size;
MPIR_FUNC_TERSE_STATE_DECL(MPID_STATE_MPIR_BCAST);
MPIR_FUNC_TERSE_ENTER(MPID_STATE_MPIR_BCAST);
if (count == 0)
goto fn_exit;
MPIR_Datatype_get_size_macro(datatype, type_size);
nbytes = MPIR_CVAR_MAX_SMP_BCAST_MSG_SIZE ? type_size * count : 0;
if (MPIR_CVAR_ENABLE_SMP_COLLECTIVES && MPIR_CVAR_ENABLE_SMP_BCAST &&
nbytes <= MPIR_CVAR_MAX_SMP_BCAST_MSG_SIZE && MPIR_Comm_is_node_aware(comm_ptr)) {
mpi_errno = MPIR_Bcast_intra_smp(buffer, count, datatype, root, comm_ptr, errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag =
MPIX_ERR_PROC_FAILED ==
MPIR_ERR_GET_CLASS(mpi_errno) ? MPIR_ERR_PROC_FAILED : MPIR_ERR_OTHER;
MPIR_ERR_SET(mpi_errno, *errflag, "**fail");
MPIR_ERR_ADD(mpi_errno_ret, mpi_errno);
}
goto fn_exit;
}
comm_size = comm_ptr->local_size;
MPIR_Datatype_get_size_macro(datatype, type_size);
nbytes = type_size * count;
if (nbytes == 0)
goto fn_exit; /* nothing to do */
if ((nbytes < MPIR_CVAR_BCAST_SHORT_MSG_SIZE) || (comm_size < MPIR_CVAR_BCAST_MIN_PROCS)) {
mpi_errno = MPIR_Bcast_intra_binomial(buffer, count, datatype, root, comm_ptr, errflag);
} else { /* (nbytes >= MPIR_CVAR_BCAST_SHORT_MSG_SIZE) && (comm_size >= MPIR_CVAR_BCAST_MIN_PROCS) */
if ((nbytes < MPIR_CVAR_BCAST_LONG_MSG_SIZE) && (MPL_is_pof2(comm_size, NULL))) {
mpi_errno =
MPIR_Bcast_intra_scatter_recursive_doubling_allgather(buffer, count, datatype, root,
comm_ptr, errflag);
} else { /* (nbytes >= MPIR_CVAR_BCAST_LONG_MSG_SIZE) || !(comm_size_is_pof2) */
mpi_errno =
MPIR_Bcast_intra_scatter_ring_allgather(buffer, count, datatype, root, comm_ptr,
errflag);
}
}
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag =
MPIX_ERR_PROC_FAILED ==
MPIR_ERR_GET_CLASS(mpi_errno) ? MPIR_ERR_PROC_FAILED : MPIR_ERR_OTHER;
MPIR_ERR_SET(mpi_errno, *errflag, "**fail");
MPIR_ERR_ADD(mpi_errno_ret, mpi_errno);
}
fn_exit:
MPIR_FUNC_TERSE_EXIT(MPID_STATE_MPIR_BCAST);
/* --BEGIN ERROR HANDLING-- */
if (mpi_errno_ret)
mpi_errno = mpi_errno_ret;
else if (*errflag != MPIR_ERR_NONE)
MPIR_ERR_SET(mpi_errno, *errflag, "**coll_fail");
/* --END ERROR HANDLING-- */
return mpi_errno;
}
/* comm shrink impl; assumes that standard error checking has already taken
* place in the calling function */
int MPIR_Comm_shrink(MPIR_Comm * comm_ptr, MPIR_Comm ** newcomm_ptr)
{
int mpi_errno = MPI_SUCCESS;
MPIR_Group *global_failed = NULL, *comm_grp = NULL, *new_group_ptr = NULL;
int attempts = 0;
MPIR_Errflag_t errflag = MPIR_ERR_NONE;
MPIR_FUNC_TERSE_STATE_DECL(MPID_STATE_MPIR_COMM_SHRINK);
MPIR_FUNC_TERSE_ENTER(MPID_STATE_MPIR_COMM_SHRINK);
/* TODO - Implement this function for intercommunicators */
MPIR_Comm_group_impl(comm_ptr, &comm_grp);
do {
errflag = MPIR_ERR_NONE;
MPID_Comm_get_all_failed_procs(comm_ptr, &global_failed, MPIR_SHRINK_TAG);
/* Ignore the mpi_errno value here as it will definitely communicate
* with failed procs */
mpi_errno = MPIR_Group_difference_impl(comm_grp, global_failed, &new_group_ptr);
if (mpi_errno)
MPIR_ERR_POP(mpi_errno);
if (MPIR_Group_empty != global_failed)
MPIR_Group_release(global_failed);
mpi_errno = MPIR_Comm_create_group(comm_ptr, new_group_ptr, MPIR_SHRINK_TAG, newcomm_ptr);
if (*newcomm_ptr == NULL) {
errflag = MPIR_ERR_PROC_FAILED;
} else if (mpi_errno) {
errflag =
MPIX_ERR_PROC_FAILED ==
MPIR_ERR_GET_CLASS(mpi_errno) ? MPIR_ERR_PROC_FAILED : MPIR_ERR_OTHER;
MPIR_Comm_release(*newcomm_ptr);
}
mpi_errno = MPII_Allreduce_group(MPI_IN_PLACE, &errflag, 1, MPI_INT, MPI_MAX, comm_ptr,
new_group_ptr, MPIR_SHRINK_TAG, &errflag);
MPIR_Group_release(new_group_ptr);
if (errflag) {
if (*newcomm_ptr != NULL && MPIR_Object_get_ref(*newcomm_ptr) > 0) {
MPIR_Object_set_ref(*newcomm_ptr, 1);
MPIR_Comm_release(*newcomm_ptr);
}
if (MPIR_Object_get_ref(new_group_ptr) > 0) {
MPIR_Object_set_ref(new_group_ptr, 1);
MPIR_Group_release(new_group_ptr);
}
}
} while (errflag && ++attempts < 5);
if (errflag && attempts >= 5)
goto fn_fail;
else
mpi_errno = MPI_SUCCESS;
fn_exit:
MPIR_Group_release(comm_grp);
MPIR_FUNC_TERSE_EXIT(MPID_STATE_MPIR_COMM_SHRINK);
return mpi_errno;
fn_fail:
if (*newcomm_ptr)
MPIR_Object_set_ref(*newcomm_ptr, 0);
MPIR_Object_set_ref(global_failed, 0);
MPIR_Object_set_ref(new_group_ptr, 0);
goto fn_exit;
}
int MPIR_Allgatherv_intra (
const void *sendbuf,
int sendcount,
MPI_Datatype sendtype,
void *recvbuf,
const int *recvcounts,
const int *displs,
MPI_Datatype recvtype,
MPID_Comm *comm_ptr,
mpir_errflag_t *errflag )
{
int comm_size, rank, j, i, left, right;
int mpi_errno = MPI_SUCCESS;
int mpi_errno_ret = MPI_SUCCESS;
MPI_Status status;
MPI_Aint recvbuf_extent, recvtype_extent, recvtype_true_extent,
recvtype_true_lb;
int curr_cnt, send_cnt, dst, total_count, recvtype_size, pof2, src, rem;
int recv_cnt;
void *tmp_buf;
int mask, dst_tree_root, my_tree_root, is_homogeneous, position,
send_offset, recv_offset, last_recv_cnt, nprocs_completed, k,
offset, tmp_mask, tree_root;
#ifdef MPID_HAS_HETERO
int tmp_buf_size, nbytes;
#endif
MPIU_CHKLMEM_DECL(1);
/* check if multiple threads are calling this collective function */
MPIDU_ERR_CHECK_MULTIPLE_THREADS_ENTER( comm_ptr );
comm_size = comm_ptr->local_size;
rank = comm_ptr->rank;
total_count = 0;
for (i=0; i<comm_size; i++)
total_count += recvcounts[i];
if (total_count == 0) goto fn_exit;
MPID_Datatype_get_extent_macro( recvtype, recvtype_extent );
MPID_Datatype_get_size_macro(recvtype, recvtype_size);
if ((total_count*recvtype_size < MPIR_CVAR_ALLGATHER_LONG_MSG_SIZE) &&
!(comm_size & (comm_size - 1))) {
/* Short or medium size message and power-of-two no. of processes. Use
* recursive doubling algorithm */
is_homogeneous = 1;
#ifdef MPID_HAS_HETERO
if (comm_ptr->is_hetero)
is_homogeneous = 0;
#endif
if (is_homogeneous) {
/* need to receive contiguously into tmp_buf because
displs could make the recvbuf noncontiguous */
MPIR_Type_get_true_extent_impl(recvtype, &recvtype_true_lb, &recvtype_true_extent);
MPID_Ensure_Aint_fits_in_pointer(total_count *
(MPIR_MAX(recvtype_true_extent, recvtype_extent)));
MPIU_CHKLMEM_MALLOC(tmp_buf, void *, total_count*(MPIR_MAX(recvtype_true_extent,recvtype_extent)), mpi_errno, "tmp_buf");
/* adjust for potential negative lower bound in datatype */
tmp_buf = (void *)((char*)tmp_buf - recvtype_true_lb);
/* copy local data into right location in tmp_buf */
position = 0;
for (i=0; i<rank; i++) position += recvcounts[i];
if (sendbuf != MPI_IN_PLACE)
{
mpi_errno = MPIR_Localcopy(sendbuf, sendcount, sendtype,
((char *)tmp_buf + position*
recvtype_extent),
recvcounts[rank], recvtype);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
}
else
{
/* if in_place specified, local data is found in recvbuf */
mpi_errno = MPIR_Localcopy(((char *)recvbuf +
displs[rank]*recvtype_extent),
recvcounts[rank], recvtype,
((char *)tmp_buf + position*
recvtype_extent),
recvcounts[rank], recvtype);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
}
curr_cnt = recvcounts[rank];
mask = 0x1;
i = 0;
while (mask < comm_size) {
dst = rank ^ mask;
/* find offset into send and recv buffers. zero out
the least significant "i" bits of rank and dst to
find root of src and dst subtrees. Use ranks of
roots as index to send from and recv into buffer */
dst_tree_root = dst >> i;
dst_tree_root <<= i;
my_tree_root = rank >> i;
my_tree_root <<= i;
if (dst < comm_size) {
send_offset = 0;
for (j=0; j<my_tree_root; j++)
send_offset += recvcounts[j];
recv_offset = 0;
for (j=0; j<dst_tree_root; j++)
recv_offset += recvcounts[j];
mpi_errno = MPIC_Sendrecv(((char *)tmp_buf + send_offset * recvtype_extent),
curr_cnt, recvtype, dst,
MPIR_ALLGATHERV_TAG,
((char *)tmp_buf + recv_offset * recvtype_extent),
total_count - recv_offset, recvtype, dst,
MPIR_ALLGATHERV_TAG,
comm_ptr, &status, errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag = MPIR_ERR_GET_CLASS(mpi_errno);
MPIU_ERR_SET(mpi_errno, *errflag, "**fail");
MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
last_recv_cnt = 0;
} else
/* for convenience, recv is posted for a bigger amount
than will be sent */
MPIR_Get_count_impl(&status, recvtype, &last_recv_cnt);
curr_cnt += last_recv_cnt;
}
/* if some processes in this process's subtree in this step
did not have any destination process to communicate with
because of non-power-of-two, we need to send them the
data that they would normally have received from those
processes. That is, the haves in this subtree must send to
the havenots. We use a logarithmic
recursive-halfing algorithm for this. */
/* This part of the code will not currently be
executed because we are not using recursive
doubling for non power of two. Mark it as experimental
so that it doesn't show up as red in the coverage
tests. */
/* --BEGIN EXPERIMENTAL-- */
if (dst_tree_root + mask > comm_size) {
nprocs_completed = comm_size - my_tree_root - mask;
/* nprocs_completed is the number of processes in this
subtree that have all the data. Send data to others
in a tree fashion. First find root of current tree
that is being divided into two. k is the number of
least-significant bits in this process's rank that
must be zeroed out to find the rank of the root */
j = mask;
k = 0;
while (j) {
j >>= 1;
k++;
}
k--;
tmp_mask = mask >> 1;
while (tmp_mask) {
dst = rank ^ tmp_mask;
tree_root = rank >> k;
tree_root <<= k;
/* send only if this proc has data and destination
doesn't have data. at any step, multiple processes
can send if they have the data */
if ((dst > rank) &&
(rank < tree_root + nprocs_completed)
&& (dst >= tree_root + nprocs_completed)) {
offset = 0;
for (j=0; j<(my_tree_root+mask); j++)
offset += recvcounts[j];
offset *= recvtype_extent;
mpi_errno = MPIC_Send(((char *)tmp_buf + offset),
last_recv_cnt,
recvtype, dst,
MPIR_ALLGATHERV_TAG, comm_ptr, errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag = MPIR_ERR_GET_CLASS(mpi_errno);
MPIU_ERR_SET(mpi_errno, *errflag, "**fail");
MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
}
/* last_recv_cnt was set in the previous
receive. that's the amount of data to be
sent now. */
}
/* recv only if this proc. doesn't have data and sender
has data */
else if ((dst < rank) &&
(dst < tree_root + nprocs_completed) &&
(rank >= tree_root + nprocs_completed)) {
offset = 0;
for (j=0; j<(my_tree_root+mask); j++)
offset += recvcounts[j];
mpi_errno = MPIC_Recv(((char *)tmp_buf + offset * recvtype_extent),
total_count - offset, recvtype,
dst, MPIR_ALLGATHERV_TAG,
comm_ptr, &status, errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag = MPIR_ERR_GET_CLASS(mpi_errno);
MPIU_ERR_SET(mpi_errno, *errflag, "**fail");
MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
last_recv_cnt = 0;
} else
/* for convenience, recv is posted for a
bigger amount than will be sent */
MPIR_Get_count_impl(&status, recvtype, &last_recv_cnt);
curr_cnt += last_recv_cnt;
}
tmp_mask >>= 1;
k--;
}
}
/* --END EXPERIMENTAL-- */
mask <<= 1;
i++;
}
/* copy data from tmp_buf to recvbuf */
position = 0;
for (j=0; j<comm_size; j++) {
if ((sendbuf != MPI_IN_PLACE) || (j != rank)) {
/* not necessary to copy if in_place and
j==rank. otherwise copy. */
mpi_errno = MPIR_Localcopy(((char *)tmp_buf + position*recvtype_extent),
recvcounts[j], recvtype,
((char *)recvbuf + displs[j]*recvtype_extent),
recvcounts[j], recvtype);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
}
position += recvcounts[j];
}
}
int MPIR_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;
}
/* This function produces topology aware trees for reduction and broadcasts, with different
* K values. This is a heavy-weight function as it allocates shared memory, generates topology
* information, builds a package-level tree (for package leaders), and a per-package tree.
* These are combined in shared memory for other ranks to read out from.
* */
int MPIDI_SHM_topology_tree_init(MPIR_Comm * comm_ptr, int root, int bcast_k,
MPIR_Treealgo_tree_t * bcast_tree, int *bcast_topotree_fail,
int reduce_k, MPIR_Treealgo_tree_t * reduce_tree,
int *reduce_topotree_fail, MPIR_Errflag_t * errflag)
{
int *shared_region;
MPL_shm_hnd_t fd;
int num_ranks, rank;
int mpi_errno = MPI_SUCCESS, mpi_errno_ret = MPI_SUCCESS;
size_t shm_size;
int **bind_map = NULL;
int *max_entries_per_level = NULL;
int **ranks_per_package = NULL;
int *package_ctr = NULL;
size_t topo_depth = 0;
int package_level = 0, i, max_ranks_per_package = 0;
bool mapfail_flag = false;
MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_SHM_TOPOLOGY_TREE_INIT);
MPIR_FUNC_VERBOSE_ENTER(MPID_STATE_SHM_TOPOLOGY_TREE_INIT);
num_ranks = MPIR_Comm_size(comm_ptr);
rank = MPIR_Comm_rank(comm_ptr);
/* Calculate the size of shared memory that would be needed */
shm_size = sizeof(int) * 5 * num_ranks + num_ranks * sizeof(cpu_set_t);
/* STEP 1. Create shared memory region for exchanging topology information (root only) */
mpi_errno = MPIDIU_allocate_shm_segment(comm_ptr, shm_size, &fd, (void **) &shared_region,
&mapfail_flag);
if (mpi_errno || mapfail_flag) {
/* for communication errors, just record the error but continue */
*errflag =
MPIX_ERR_PROC_FAILED ==
MPIR_ERR_GET_CLASS(mpi_errno) ? MPIR_ERR_PROC_FAILED : MPIR_ERR_OTHER;
MPIR_ERR_SET(mpi_errno, *errflag, "**fail");
MPIR_ERR_ADD(mpi_errno_ret, mpi_errno);
}
/* STEP 2. Collect cpu_sets for each rank at the root */
cpu_set_t my_cpu_set;
CPU_ZERO(&my_cpu_set);
sched_getaffinity(0, sizeof(my_cpu_set), &my_cpu_set);
((cpu_set_t *) (shared_region))[rank] = my_cpu_set;
mpi_errno = MPIR_Barrier_impl(comm_ptr, errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag =
MPIX_ERR_PROC_FAILED ==
MPIR_ERR_GET_CLASS(mpi_errno) ? MPIR_ERR_PROC_FAILED : MPIR_ERR_OTHER;
MPIR_ERR_SET(mpi_errno, *errflag, "**fail");
MPIR_ERR_ADD(mpi_errno_ret, mpi_errno);
}
/* STEP 3. Root has all the cpu_set information, now build tree */
if (rank == root) {
topo_depth = hwloc_topology_get_depth(MPIR_Process.hwloc_topology);
bind_map = (int **) MPL_malloc(num_ranks * sizeof(int *), MPL_MEM_OTHER);
MPIR_ERR_CHKANDJUMP(!bind_map, mpi_errno, MPI_ERR_OTHER, "**nomem");
for (i = 0; i < num_ranks; ++i) {
bind_map[i] = (int *) MPL_calloc(topo_depth, sizeof(int), MPL_MEM_OTHER);
MPIR_ERR_CHKANDJUMP(!bind_map[i], mpi_errno, MPI_ERR_OTHER, "**nomem");
}
MPIDI_SHM_hwloc_init_bindmap(num_ranks, topo_depth, shared_region, bind_map);
/* Done building the topology information */
/* STEP 3.1. Count the maximum entries at each level - used for breaking the tree into
* intra/inter socket */
max_entries_per_level = (int *) MPL_calloc(topo_depth, sizeof(size_t), MPL_MEM_OTHER);
MPIR_ERR_CHKANDJUMP(!max_entries_per_level, mpi_errno, MPI_ERR_OTHER, "**nomem");
package_level =
MPIDI_SHM_topotree_get_package_level(topo_depth, max_entries_per_level, num_ranks,
bind_map);
if (MPIDI_SHM_TOPOTREE_DEBUG)
fprintf(stderr, "Breaking topology at :: %d (default= %d)\n", package_level,
MPIDI_SHM_TOPOTREE_CUTOFF);
/* STEP 3.2. allocate space for the entries that go in each package based on hwloc info */
ranks_per_package =
(int
**) MPL_malloc(max_entries_per_level[package_level] * sizeof(int *), MPL_MEM_OTHER);
MPIR_ERR_CHKANDJUMP(!ranks_per_package, mpi_errno, MPI_ERR_OTHER, "**nomem");
package_ctr =
(int *) MPL_calloc(max_entries_per_level[package_level], sizeof(int), MPL_MEM_OTHER);
MPIR_ERR_CHKANDJUMP(!package_ctr, mpi_errno, MPI_ERR_OTHER, "**nomem");
for (i = 0; i < max_entries_per_level[package_level]; ++i) {
package_ctr[i] = 0;
ranks_per_package[i] = (int *) MPL_calloc(num_ranks, sizeof(int), MPL_MEM_OTHER);
MPIR_ERR_CHKANDJUMP(!ranks_per_package[i], mpi_errno, MPI_ERR_OTHER, "**nomem");
}
/* sort the ranks into packages based on the binding information */
for (i = 0; i < num_ranks; ++i) {
int package = bind_map[i][package_level];
ranks_per_package[package][package_ctr[package]++] = i;
}
max_ranks_per_package = 0;
for (i = 0; i < max_entries_per_level[package_level]; ++i) {
max_ranks_per_package = MPL_MAX(max_ranks_per_package, package_ctr[i]);
}
/* At this point we have done the common work in extracting topology information
* and restructuring it to our needs. Now we generate the tree. */
/* For Bcast, package leaders are added before the package local ranks, and the per_package
* tree is left_skewed */
mpi_errno = MPIDI_SHM_gen_tree(bcast_k, shared_region, max_entries_per_level,
ranks_per_package, max_ranks_per_package, package_ctr,
package_level, num_ranks, 1 /*package_leaders_first */ ,
0 /*left_skewed */ , errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag =
MPIX_ERR_PROC_FAILED ==
MPIR_ERR_GET_CLASS(mpi_errno) ? MPIR_ERR_PROC_FAILED : MPIR_ERR_OTHER;
MPIR_ERR_SET(mpi_errno, *errflag, "**fail");
MPIR_ERR_ADD(mpi_errno_ret, mpi_errno);
}
}
mpi_errno = MPIR_Barrier_impl(comm_ptr, errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag =
MPIX_ERR_PROC_FAILED ==
MPIR_ERR_GET_CLASS(mpi_errno) ? MPIR_ERR_PROC_FAILED : MPIR_ERR_OTHER;
MPIR_ERR_SET(mpi_errno, *errflag, "**fail");
MPIR_ERR_ADD(mpi_errno_ret, mpi_errno);
}
/* Every rank copies their tree out from shared memory */
MPIDI_SHM_copy_tree(shared_region, num_ranks, rank, bcast_tree, bcast_topotree_fail);
if (MPIDI_SHM_TOPOTREE_DEBUG)
MPIDI_SHM_print_topotree_file("BCAST", comm_ptr->context_id, rank, bcast_tree);
/* Wait until shared memory is available */
mpi_errno = MPIR_Barrier_impl(comm_ptr, errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag =
MPIX_ERR_PROC_FAILED ==
MPIR_ERR_GET_CLASS(mpi_errno) ? MPIR_ERR_PROC_FAILED : MPIR_ERR_OTHER;
MPIR_ERR_SET(mpi_errno, *errflag, "**fail");
MPIR_ERR_ADD(mpi_errno_ret, mpi_errno);
}
/* Generate the reduce tree */
/* For Reduce, package leaders are added after the package local ranks, and the per_package
* tree is right_skewed (children are added in the reverse order */
if (rank == root) {
memset(shared_region, 0, shm_size);
mpi_errno = MPIDI_SHM_gen_tree(reduce_k, shared_region, max_entries_per_level,
ranks_per_package, max_ranks_per_package, package_ctr,
package_level, num_ranks, 0 /*package_leaders_last */ ,
1 /*right_skewed */ , errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag =
MPIX_ERR_PROC_FAILED ==
MPIR_ERR_GET_CLASS(mpi_errno) ? MPIR_ERR_PROC_FAILED : MPIR_ERR_OTHER;
MPIR_ERR_SET(mpi_errno, *errflag, "**fail");
MPIR_ERR_ADD(mpi_errno_ret, mpi_errno);
}
}
mpi_errno = MPIR_Barrier_impl(comm_ptr, errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag =
MPIX_ERR_PROC_FAILED ==
MPIR_ERR_GET_CLASS(mpi_errno) ? MPIR_ERR_PROC_FAILED : MPIR_ERR_OTHER;
MPIR_ERR_SET(mpi_errno, *errflag, "**fail");
MPIR_ERR_ADD(mpi_errno_ret, mpi_errno);
}
/* each rank copy the reduce tree out */
MPIDI_SHM_copy_tree(shared_region, num_ranks, rank, reduce_tree, reduce_topotree_fail);
if (MPIDI_SHM_TOPOTREE_DEBUG)
MPIDI_SHM_print_topotree_file("REDUCE", comm_ptr->context_id, rank, reduce_tree);
/* Wait for all ranks to copy out the tree */
mpi_errno = MPIR_Barrier_impl(comm_ptr, errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag =
MPIX_ERR_PROC_FAILED ==
MPIR_ERR_GET_CLASS(mpi_errno) ? MPIR_ERR_PROC_FAILED : MPIR_ERR_OTHER;
MPIR_ERR_SET(mpi_errno, *errflag, "**fail");
MPIR_ERR_ADD(mpi_errno_ret, mpi_errno);
}
/* Cleanup */
if (rank == root) {
for (i = 0; i < max_entries_per_level[package_level]; ++i) {
MPL_free(ranks_per_package[i]);
}
MPL_free(ranks_per_package);
MPL_free(package_ctr);
if (MPIDI_SHM_TOPOTREE_DEBUG)
for (i = 0; i < topo_depth; ++i) {
fprintf(stderr, "Level :: %d, Max :: %d\n", i, max_entries_per_level[i]);
}
for (i = 0; i < num_ranks; ++i) {
MPL_free(bind_map[i]);
}
MPL_free(max_entries_per_level);
MPL_free(bind_map);
}
MPIDIU_destroy_shm_segment(shm_size, &fd, (void **) &shared_region);
fn_exit:
if (rank == root && MPIDI_SHM_TOPOTREE_DEBUG)
fprintf(stderr, "Done creating tree for %d\n", num_ranks);
MPIR_FUNC_VERBOSE_EXIT(MPID_STATE_SHM_TOPOLOGY_TREE_INIT);
return mpi_errno;
fn_fail:
goto fn_exit;
}
int MPIR_Alltoall_intra_scattered(
const void *sendbuf,
int sendcount,
MPI_Datatype sendtype,
void *recvbuf,
int recvcount,
MPI_Datatype recvtype,
MPIR_Comm *comm_ptr,
MPIR_Errflag_t *errflag )
{
int comm_size, i, j;
MPI_Aint sendtype_extent, recvtype_extent;
int mpi_errno=MPI_SUCCESS, dst, rank;
int mpi_errno_ret = MPI_SUCCESS;
MPI_Datatype newtype = MPI_DATATYPE_NULL;
MPIR_Request **reqarray;
MPI_Status *starray;
MPIR_CHKLMEM_DECL(6);
if (recvcount == 0) return MPI_SUCCESS;
comm_size = comm_ptr->local_size;
rank = comm_ptr->rank;
#ifdef HAVE_ERROR_CHECKING
MPIR_Assert(sendbuf != MPI_IN_PLACE);
#endif /* HAVE_ERROR_CHECKING */
/* Get extent of send and recv types */
MPIR_Datatype_get_extent_macro(recvtype, recvtype_extent);
MPIR_Datatype_get_extent_macro(sendtype, sendtype_extent);
int ii, ss, bblock;
bblock = MPIR_CVAR_ALLTOALL_THROTTLE;
if (bblock == 0) bblock = comm_size;
MPIR_CHKLMEM_MALLOC(reqarray, MPIR_Request **, 2*bblock*sizeof(MPIR_Request*), mpi_errno, "reqarray", MPL_MEM_BUFFER);
MPIR_CHKLMEM_MALLOC(starray, MPI_Status *, 2*bblock*sizeof(MPI_Status), mpi_errno, "starray", MPL_MEM_BUFFER);
for (ii=0; ii<comm_size; ii+=bblock) {
ss = comm_size-ii < bblock ? comm_size-ii : bblock;
/* do the communication -- post ss sends and receives: */
for ( i=0; i<ss; i++ ) {
dst = (rank+i+ii) % comm_size;
mpi_errno = MPIC_Irecv((char *)recvbuf +
dst*recvcount*recvtype_extent,
recvcount, recvtype, dst,
MPIR_ALLTOALL_TAG, comm_ptr,
&reqarray[i]);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
}
for ( i=0; i<ss; i++ ) {
dst = (rank-i-ii+comm_size) % comm_size;
mpi_errno = MPIC_Isend((char *)sendbuf +
dst*sendcount*sendtype_extent,
sendcount, sendtype, dst,
MPIR_ALLTOALL_TAG, comm_ptr,
&reqarray[i+ss], errflag);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
}
/* ... then wait for them to finish: */
mpi_errno = MPIC_Waitall(2*ss,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 (j=0; j<2*ss; j++) {
if (starray[j].MPI_ERROR != MPI_SUCCESS) {
mpi_errno = starray[j].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:
MPIR_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:
if (newtype != MPI_DATATYPE_NULL)
MPIR_Type_free_impl(&newtype);
goto fn_exit;
}
static int MPIDI_CH3I_Progress_handle_sock_event(MPIDU_Sock_event_t * event)
{
int mpi_errno = MPI_SUCCESS;
MPIDI_STATE_DECL(MPID_STATE_MPIDI_CH3I_PROGRESS_HANDLE_SOCK_EVENT);
MPIDI_FUNC_ENTER(MPID_STATE_MPIDI_CH3I_PROGRESS_HANDLE_SOCK_EVENT);
MPIU_DBG_MSG_D(CH3_OTHER,VERBOSE,"Socket event of type %d", event->op_type );
switch (event->op_type)
{
case MPIDU_SOCK_OP_READ:
{
MPIDI_CH3I_Connection_t * conn =
(MPIDI_CH3I_Connection_t *) event->user_ptr;
MPID_Request * rreq = conn->recv_active;
/* --BEGIN ERROR HANDLING-- */
if (event->error != MPI_SUCCESS)
{
/* FIXME: the following should be handled by the close
protocol */
if (MPIR_ERR_GET_CLASS(event->error) != MPIDU_SOCK_ERR_CONN_CLOSED) {
mpi_errno = event->error;
MPIU_ERR_POP(mpi_errno);
}
break;
}
/* --END ERROR HANDLING-- */
if (conn->state == CONN_STATE_CONNECTED)
{
if (conn->recv_active == NULL)
{
MPIDI_msg_sz_t buflen = sizeof (MPIDI_CH3_Pkt_t);
MPIU_Assert(conn->pkt.type < MPIDI_CH3_PKT_END_CH3);
mpi_errno = pktArray[conn->pkt.type]( conn->vc, &conn->pkt,
&buflen, &rreq );
if (mpi_errno != MPI_SUCCESS) {
MPIU_ERR_POP(mpi_errno);
}
MPIU_Assert(buflen == sizeof (MPIDI_CH3_Pkt_t));
if (rreq == NULL)
{
if (conn->state != CONN_STATE_CLOSING)
{
/* conn->recv_active = NULL; --
already set to NULL */
mpi_errno = connection_post_recv_pkt(conn);
if (mpi_errno != MPI_SUCCESS) {
MPIU_ERR_POP(mpi_errno);
}
}
}
else
{
mpi_errno = ReadMoreData( conn, rreq );
if (mpi_errno) { MPIU_ERR_POP(mpi_errno); }
}
}
else /* incoming data */
{
int (*reqFn)(MPIDI_VC_t *, MPID_Request *, int *);
int complete;
reqFn = rreq->dev.OnDataAvail;
if (!reqFn) {
MPIU_Assert(MPIDI_Request_get_type(rreq)!=MPIDI_REQUEST_TYPE_GET_RESP);
MPIDI_CH3U_Request_complete(rreq);
complete = TRUE;
}
else {
mpi_errno = reqFn( conn->vc, rreq, &complete );
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
}
if (complete)
{
conn->recv_active = NULL;
mpi_errno = connection_post_recv_pkt(conn);
if (mpi_errno != MPI_SUCCESS) {
MPIU_ERR_POP(mpi_errno);
}
}
else /* more data to be read */
{
mpi_errno = ReadMoreData( conn, rreq );
if (mpi_errno) { MPIU_ERR_POP(mpi_errno); }
}
}
}
else if (conn->state == CONN_STATE_OPEN_LRECV_DATA)
{
mpi_errno = MPIDI_CH3_Sockconn_handle_connopen_event( conn );
if (mpi_errno) { MPIU_ERR_POP( mpi_errno ); }
}
else /* Handling some internal connection establishment or
tear down packet */
{
mpi_errno = MPIDI_CH3_Sockconn_handle_conn_event( conn );
if (mpi_errno) { MPIU_ERR_POP(mpi_errno); }
}
break;
}
/* END OF SOCK_OP_READ */
case MPIDU_SOCK_OP_WRITE:
{
MPIDI_CH3I_Connection_t * conn =
(MPIDI_CH3I_Connection_t *) event->user_ptr;
/* --BEGIN ERROR HANDLING-- */
if (event->error != MPI_SUCCESS) {
mpi_errno = event->error;
MPIU_ERR_POP(mpi_errno);
}
/* --END ERROR HANDLING-- */
if (conn->send_active)
{
MPID_Request * sreq = conn->send_active;
int (*reqFn)(MPIDI_VC_t *, MPID_Request *, int *);
int complete;
reqFn = sreq->dev.OnDataAvail;
if (!reqFn) {
MPIU_Assert(MPIDI_Request_get_type(sreq)!=MPIDI_REQUEST_TYPE_GET_RESP);
MPIDI_CH3U_Request_complete(sreq);
complete = TRUE;
}
else {
mpi_errno = reqFn( conn->vc, sreq, &complete );
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
}
if (complete)
{
mpi_errno = connection_pop_sendq_req(conn);
if (mpi_errno != MPI_SUCCESS) {
MPIU_ERR_POP(mpi_errno);
}
}
else /* more data to send */
{
for(;;)
{
MPID_IOV * iovp;
MPIU_Size_t nb;
iovp = sreq->dev.iov;
mpi_errno = MPIDU_Sock_writev(conn->sock, iovp, sreq->dev.iov_count, &nb);
/* --BEGIN ERROR HANDLING-- */
if (mpi_errno != MPI_SUCCESS)
{
mpi_errno = MPIR_Err_create_code(mpi_errno, MPIR_ERR_FATAL, FCNAME, __LINE__, MPI_ERR_OTHER,
"**ch3|sock|immedwrite", "ch3|sock|immedwrite %p %p %p",
sreq, conn, conn->vc);
goto fn_fail;
}
/* --END ERROR HANDLING-- */
MPIU_DBG_MSG_FMT(CH3_CHANNEL,VERBOSE,
(MPIU_DBG_FDEST,"immediate writev, vc=%p, sreq=0x%08x, nb=" MPIDI_MSG_SZ_FMT,
conn->vc, sreq->handle, nb));
if (nb > 0 && adjust_iov(&iovp, &sreq->dev.iov_count, nb))
{
reqFn = sreq->dev.OnDataAvail;
if (!reqFn) {
MPIU_Assert(MPIDI_Request_get_type(sreq)!=MPIDI_REQUEST_TYPE_GET_RESP);
MPIDI_CH3U_Request_complete(sreq);
complete = TRUE;
}
else {
mpi_errno = reqFn( conn->vc, sreq, &complete );
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
}
if (complete)
{
mpi_errno = connection_pop_sendq_req(conn);
if (mpi_errno != MPI_SUCCESS) {
MPIU_ERR_POP(mpi_errno);
}
break;
}
}
else
{
MPIU_DBG_MSG_FMT(CH3_CHANNEL,VERBOSE,
(MPIU_DBG_FDEST,"posting writev, vc=%p, conn=%p, sreq=0x%08x",
conn->vc, conn, sreq->handle));
mpi_errno = MPIDU_Sock_post_writev(conn->sock, iovp, sreq->dev.iov_count, NULL);
/* --BEGIN ERROR HANDLING-- */
if (mpi_errno != MPI_SUCCESS)
{
mpi_errno = MPIR_Err_create_code(
mpi_errno, MPIR_ERR_FATAL, FCNAME, __LINE__, MPI_ERR_OTHER, "**ch3|sock|postwrite",
"ch3|sock|postwrite %p %p %p", sreq, conn, conn->vc);
goto fn_fail;
}
/* --END ERROR HANDLING-- */
break;
}
}
}
}
else /* finished writing internal packet header */
{
/* the connection is not active yet */
mpi_errno = MPIDI_CH3_Sockconn_handle_connwrite( conn );
if (mpi_errno) { MPIU_ERR_POP( mpi_errno ); }
}
break;
}
/* END OF SOCK_OP_WRITE */
case MPIDU_SOCK_OP_ACCEPT:
{
mpi_errno = MPIDI_CH3_Sockconn_handle_accept_event();
if (mpi_errno) { MPIU_ERR_POP(mpi_errno); }
break;
}
case MPIDU_SOCK_OP_CONNECT:
{
mpi_errno = MPIDI_CH3_Sockconn_handle_connect_event(
(MPIDI_CH3I_Connection_t *) event->user_ptr,
event->error );
if (mpi_errno) { MPIU_ERR_POP(mpi_errno); }
break;
}
case MPIDU_SOCK_OP_CLOSE:
{
mpi_errno = MPIDI_CH3_Sockconn_handle_close_event(
(MPIDI_CH3I_Connection_t *) event->user_ptr );
if (mpi_errno) { MPIU_ERR_POP(mpi_errno); }
break;
}
case MPIDU_SOCK_OP_WAKEUP:
{
MPIDI_CH3_Progress_signal_completion();
/* MPIDI_CH3I_progress_completion_count++; */
break;
}
}
fn_exit:
MPIDI_FUNC_EXIT(MPID_STATE_MPIDI_CH3I_PROGRESS_HANDLE_SOCK_EVENT);
return mpi_errno;
fn_fail:
goto fn_exit;
}
int MPIR_Reduce_intra_auto (
const void *sendbuf,
void *recvbuf,
int count,
MPI_Datatype datatype,
MPI_Op op,
int root,
MPIR_Comm *comm_ptr,
MPIR_Errflag_t *errflag )
{
int mpi_errno = MPI_SUCCESS;
int mpi_errno_ret = MPI_SUCCESS;
int is_commutative, type_size, pof2;
int nbytes = 0;
if (count == 0) return MPI_SUCCESS;
/* is the op commutative? We do SMP optimizations only if it is. */
is_commutative = MPIR_Op_is_commutative(op);
MPIR_Datatype_get_size_macro(datatype, type_size);
nbytes = MPIR_CVAR_MAX_SMP_REDUCE_MSG_SIZE ? type_size*count : 0;
if (MPIR_CVAR_ENABLE_SMP_COLLECTIVES &&
MPIR_CVAR_ENABLE_SMP_REDUCE &&
MPIR_Comm_is_node_aware(comm_ptr) &&
is_commutative &&
nbytes <= MPIR_CVAR_MAX_SMP_REDUCE_MSG_SIZE) {
mpi_errno = MPIR_Reduce_intra_smp(sendbuf, recvbuf, count, datatype,
op, root, comm_ptr, 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);
}
goto fn_exit;
}
MPIR_Datatype_get_size_macro(datatype, type_size);
/* get nearest power-of-two less than or equal to comm_size */
pof2 = comm_ptr->pof2;
if ((count*type_size > MPIR_CVAR_REDUCE_SHORT_MSG_SIZE) &&
(HANDLE_GET_KIND(op) == HANDLE_KIND_BUILTIN) && (count >= pof2)) {
/* do a reduce-scatter followed by gather to root. */
mpi_errno = MPIR_Reduce_intra_reduce_scatter_gather(sendbuf, recvbuf, count, datatype, op, root, comm_ptr, errflag);
}
else {
/* use a binomial tree algorithm */
mpi_errno = MPIR_Reduce_intra_binomial(sendbuf, recvbuf, count, datatype, op, root, comm_ptr, 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);
}
fn_exit:
if (mpi_errno_ret)
mpi_errno = mpi_errno_ret;
else if (*errflag != MPIR_ERR_NONE)
MPIR_ERR_SET(mpi_errno, *errflag, "**coll_fail");
return mpi_errno;
}
int MPIR_Alltoallv_intra_pairwise_sendrecv_replace(const void *sendbuf, const int *sendcounts,
const int *sdispls, MPI_Datatype sendtype,
void *recvbuf, const int *recvcounts,
const int *rdispls, MPI_Datatype recvtype,
MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
{
int comm_size, i, j;
MPI_Aint recv_extent;
int mpi_errno = MPI_SUCCESS;
int mpi_errno_ret = MPI_SUCCESS;
MPI_Status status;
int rank;
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) */
MPIR_Datatype_get_extent_macro(recvtype, recv_extent);
#ifdef HAVE_ERROR_CHECKING
MPIR_Assert(sendbuf == MPI_IN_PLACE);
#endif
/* 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 =
MPIX_ERR_PROC_FAILED ==
MPIR_ERR_GET_CLASS(mpi_errno) ? MPIR_ERR_PROC_FAILED : MPIR_ERR_OTHER;
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 =
MPIX_ERR_PROC_FAILED ==
MPIR_ERR_GET_CLASS(mpi_errno) ? MPIR_ERR_PROC_FAILED : MPIR_ERR_OTHER;
MPIR_ERR_SET(mpi_errno, *errflag, "**fail");
MPIR_ERR_ADD(mpi_errno_ret, mpi_errno);
}
}
}
}
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;
}
int MPIR_Allreduce_intra_recursive_doubling(
const void *sendbuf,
void *recvbuf,
int count,
MPI_Datatype datatype,
MPI_Op op,
MPIR_Comm * comm_ptr,
MPIR_Errflag_t * errflag)
{
MPIR_CHKLMEM_DECL(1);
#ifdef MPID_HAS_HETERO
int is_homogeneous;
int rc;
#endif
int comm_size, rank;
int mpi_errno = MPI_SUCCESS;
int mpi_errno_ret = MPI_SUCCESS;
int mask, dst, is_commutative, pof2, newrank, rem, newdst;
MPI_Aint true_extent, true_lb, extent;
void *tmp_buf;
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);
MPIR_Datatype_get_extent_macro(datatype, extent);
MPIR_Ensure_Aint_fits_in_pointer(count * MPL_MAX(extent, true_extent));
MPIR_CHKLMEM_MALLOC(tmp_buf, void *, count*(MPL_MAX(extent,true_extent)), mpi_errno, "temporary buffer", MPL_MEM_BUFFER);
/* adjust for potential negative lower bound in datatype */
tmp_buf = (void *)((char*)tmp_buf - true_lb);
/* copy local data into recvbuf */
if (sendbuf != MPI_IN_PLACE) {
mpi_errno = MPIR_Localcopy(sendbuf, count, datatype, recvbuf,
count, datatype);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
}
/* get nearest power-of-two less than or equal to comm_size */
pof2 = comm_ptr->pof2;
rem = comm_size - pof2;
/* In the non-power-of-two case, all even-numbered
processes of rank < 2*rem send their data to
(rank+1). These even-numbered processes no longer
participate in the algorithm until the very end. The
remaining processes form a nice power-of-two. */
if (rank < 2*rem) {
if (rank % 2 == 0) { /* even */
mpi_errno = MPIC_Send(recvbuf, count,
datatype, rank+1,
MPIR_ALLREDUCE_TAG, comm_ptr, 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);
}
/* temporarily set the rank to -1 so that this
process does not pariticipate in recursive
doubling */
newrank = -1;
}
else { /* odd */
mpi_errno = MPIC_Recv(tmp_buf, count,
datatype, rank-1,
MPIR_ALLREDUCE_TAG, comm_ptr,
MPI_STATUS_IGNORE, errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag = MPIR_ERR_GET_CLASS(mpi_errno);
MPIR_ERR_SET(mpi_errno, *errflag, "**fail");
MPIR_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(tmp_buf, recvbuf, count, datatype, op);
if (mpi_errno) MPIR_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) {
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(recvbuf, count, datatype,
dst, MPIR_ALLREDUCE_TAG, tmp_buf,
count, datatype, dst,
MPIR_ALLREDUCE_TAG, comm_ptr,
MPI_STATUS_IGNORE, errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag = MPIR_ERR_GET_CLASS(mpi_errno);
MPIR_ERR_SET(mpi_errno, *errflag, "**fail");
MPIR_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(tmp_buf, recvbuf, count, datatype, op);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
}
else {
/* op is noncommutative and the order is not right */
mpi_errno = MPIR_Reduce_local(recvbuf, tmp_buf, count, datatype, op);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
/* copy result back into recvbuf */
mpi_errno = MPIR_Localcopy(tmp_buf, count, datatype,
recvbuf, count, datatype);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
}
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(recvbuf, count,
datatype, rank-1,
MPIR_ALLREDUCE_TAG, comm_ptr, errflag);
else /* even */
mpi_errno = MPIC_Recv(recvbuf, count,
datatype, rank+1,
MPIR_ALLREDUCE_TAG, comm_ptr,
MPI_STATUS_IGNORE, errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag = MPIR_ERR_GET_CLASS(mpi_errno);
MPIR_ERR_SET(mpi_errno, *errflag, "**fail");
MPIR_ERR_ADD(mpi_errno_ret, mpi_errno);
}
}
fn_exit:
MPIR_CHKLMEM_FREEALL();
return mpi_errno;
fn_fail:
goto fn_exit;
}
int MPIR_Gatherv (
const void *sendbuf,
int sendcount,
MPI_Datatype sendtype,
void *recvbuf,
const int *recvcounts,
const int *displs,
MPI_Datatype recvtype,
int root,
MPID_Comm *comm_ptr,
MPIR_Errflag_t *errflag )
{
int comm_size, rank;
int mpi_errno = MPI_SUCCESS;
int mpi_errno_ret = MPI_SUCCESS;
MPI_Aint extent;
int i, reqs;
int min_procs;
MPID_Request **reqarray;
MPI_Status *starray;
MPIU_CHKLMEM_DECL(2);
rank = comm_ptr->rank;
/* check if multiple threads are calling this collective function */
MPIDU_ERR_CHECK_MULTIPLE_THREADS_ENTER( comm_ptr );
/* If rank == root, then I recv lots, otherwise I send */
if (((comm_ptr->comm_kind == MPID_INTRACOMM) && (root == rank)) ||
((comm_ptr->comm_kind == MPID_INTERCOMM) && (root == MPI_ROOT))) {
if (comm_ptr->comm_kind == MPID_INTRACOMM)
comm_size = comm_ptr->local_size;
else
comm_size = comm_ptr->remote_size;
MPID_Datatype_get_extent_macro(recvtype, extent);
/* each node can make sure it is not going to overflow aint */
MPIU_Ensure_Aint_fits_in_pointer(MPIU_VOID_PTR_CAST_TO_MPI_AINT recvbuf +
displs[rank] * extent);
MPIU_CHKLMEM_MALLOC(reqarray, MPID_Request **, comm_size * sizeof(MPID_Request *), mpi_errno, "reqarray");
MPIU_CHKLMEM_MALLOC(starray, MPI_Status *, comm_size * sizeof(MPI_Status), mpi_errno, "starray");
reqs = 0;
for (i = 0; i < comm_size; i++) {
if (recvcounts[i]) {
if ((comm_ptr->comm_kind == MPID_INTRACOMM) && (i == rank)) {
if (sendbuf != MPI_IN_PLACE) {
mpi_errno = MPIR_Localcopy(sendbuf, sendcount, sendtype,
((char *)recvbuf+displs[rank]*extent),
recvcounts[rank], recvtype);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
}
}
else {
mpi_errno = MPIC_Irecv(((char *)recvbuf+displs[i]*extent),
recvcounts[i], recvtype, i,
MPIR_GATHERV_TAG, comm_ptr,
&reqarray[reqs++]);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
}
}
}
/* ... then wait for *all* of them to finish: */
mpi_errno = MPIC_Waitall(reqs, 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 < reqs; i++) {
if (starray[i].MPI_ERROR != MPI_SUCCESS) {
mpi_errno = starray[i].MPI_ERROR;
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag = MPIR_ERR_GET_CLASS(mpi_errno);
MPIR_ERR_SET(mpi_errno, *errflag, "**fail");
MPIR_ERR_ADD(mpi_errno_ret, mpi_errno);
}
}
}
}
/* --END ERROR HANDLING-- */
}
else if (root != MPI_PROC_NULL) { /* non-root nodes, and in the intercomm. case, non-root nodes on remote side */
/*
* Recursive Doubling Algorithm:
*
* Restrictions: power-of-two no. of processes
*
* Cost = lgp.alpha + n.((p-1)/p).beta
*
* TODO: On TCP, we may want to use recursive doubling instead of the
* Bruck's algorithm in all cases because of the pairwise-exchange
* property of recursive doubling (see Benson et al paper in Euro
* PVM/MPI 2003).
*/
int MPIR_Allgather_intra_recursive_doubling(const void *sendbuf,
int sendcount,
MPI_Datatype sendtype,
void *recvbuf,
int recvcount,
MPI_Datatype recvtype,
MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
{
int comm_size, rank;
int mpi_errno = MPI_SUCCESS;
int mpi_errno_ret = MPI_SUCCESS;
MPI_Aint recvtype_extent;
int j, i;
MPI_Aint curr_cnt, last_recv_cnt = 0;
int dst;
MPI_Status status;
int mask, dst_tree_root, my_tree_root,
send_offset, recv_offset, nprocs_completed, k, offset, tmp_mask, tree_root;
if (((sendcount == 0) && (sendbuf != MPI_IN_PLACE)) || (recvcount == 0))
return MPI_SUCCESS;
comm_size = comm_ptr->local_size;
rank = comm_ptr->rank;
#ifdef HAVE_ERROR_CHECKING
/* Currently this algorithm can only handle power-of-2 comm_size.
* Non power-of-2 comm_size is still experimental */
MPIR_Assert(!(comm_size & (comm_size - 1)));
#endif /* HAVE_ERROR_CHECKING */
MPIR_Datatype_get_extent_macro(recvtype, recvtype_extent);
if (sendbuf != MPI_IN_PLACE) {
mpi_errno = MPIR_Localcopy(sendbuf, sendcount, sendtype,
((char *) recvbuf +
rank * recvcount * recvtype_extent), recvcount, recvtype);
if (mpi_errno) {
MPIR_ERR_POP(mpi_errno);
}
}
curr_cnt = recvcount;
mask = 0x1;
i = 0;
while (mask < comm_size) {
dst = rank ^ mask;
/* find offset into send and recv buffers. zero out
* the least significant "i" bits of rank and dst to
* find root of src and dst subtrees. Use ranks of
* roots as index to send from and recv into buffer */
dst_tree_root = dst >> i;
dst_tree_root <<= i;
my_tree_root = rank >> i;
my_tree_root <<= i;
/* FIXME: saving an MPI_Aint into an int */
send_offset = my_tree_root * recvcount * recvtype_extent;
recv_offset = dst_tree_root * recvcount * recvtype_extent;
if (dst < comm_size) {
mpi_errno = MPIC_Sendrecv(((char *) recvbuf + send_offset),
curr_cnt, recvtype, dst,
MPIR_ALLGATHER_TAG,
((char *) recvbuf + recv_offset),
(comm_size - dst_tree_root) * recvcount,
recvtype, dst,
MPIR_ALLGATHER_TAG, comm_ptr, &status, errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag =
MPIX_ERR_PROC_FAILED ==
MPIR_ERR_GET_CLASS(mpi_errno) ? MPIR_ERR_PROC_FAILED : MPIR_ERR_OTHER;
MPIR_ERR_SET(mpi_errno, *errflag, "**fail");
MPIR_ERR_ADD(mpi_errno_ret, mpi_errno);
last_recv_cnt = 0;
} else {
MPIR_Get_count_impl(&status, recvtype, &last_recv_cnt);
}
curr_cnt += last_recv_cnt;
}
/* if some processes in this process's subtree in this step
* did not have any destination process to communicate with
* because of non-power-of-two, we need to send them the
* data that they would normally have received from those
* processes. That is, the haves in this subtree must send to
* the havenots. We use a logarithmic recursive-halfing algorithm
* for this. */
/* This part of the code will not currently be
* executed because we are not using recursive
* doubling for non power of two. Mark it as experimental
* so that it doesn't show up as red in the coverage
* tests. */
/* --BEGIN EXPERIMENTAL-- */
if (dst_tree_root + mask > comm_size) {
nprocs_completed = comm_size - my_tree_root - mask;
/* nprocs_completed is the number of processes in this
* subtree that have all the data. Send data to others
* in a tree fashion. First find root of current tree
* that is being divided into two. k is the number of
* least-significant bits in this process's rank that
* must be zeroed out to find the rank of the root */
j = mask;
k = 0;
while (j) {
j >>= 1;
k++;
}
k--;
/* FIXME: saving an MPI_Aint into an int */
offset = recvcount * (my_tree_root + mask) * recvtype_extent;
tmp_mask = mask >> 1;
while (tmp_mask) {
dst = rank ^ tmp_mask;
tree_root = rank >> k;
tree_root <<= k;
/* send only if this proc has data and destination
* doesn't have data. at any step, multiple processes
* can send if they have the data */
if ((dst > rank) && (rank < tree_root + nprocs_completed)
&& (dst >= tree_root + nprocs_completed)) {
mpi_errno = MPIC_Send(((char *) recvbuf + offset),
last_recv_cnt,
recvtype, dst, MPIR_ALLGATHER_TAG, comm_ptr, errflag);
/* last_recv_cnt was set in the previous
* receive. that's the amount of data to be
* sent now. */
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag =
MPIX_ERR_PROC_FAILED ==
MPIR_ERR_GET_CLASS(mpi_errno) ? MPIR_ERR_PROC_FAILED : MPIR_ERR_OTHER;
MPIR_ERR_SET(mpi_errno, *errflag, "**fail");
MPIR_ERR_ADD(mpi_errno_ret, mpi_errno);
}
}
/* recv only if this proc. doesn't have data and sender
* has data */
else if ((dst < rank) &&
(dst < tree_root + nprocs_completed) &&
(rank >= tree_root + nprocs_completed)) {
mpi_errno = MPIC_Recv(((char *) recvbuf + offset),
(comm_size - (my_tree_root + mask)) * recvcount,
recvtype, dst,
MPIR_ALLGATHER_TAG, comm_ptr, &status, errflag);
/* nprocs_completed is also equal to the
* no. of processes whose data we don't have */
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag =
MPIX_ERR_PROC_FAILED ==
MPIR_ERR_GET_CLASS(mpi_errno) ? MPIR_ERR_PROC_FAILED : MPIR_ERR_OTHER;
MPIR_ERR_SET(mpi_errno, *errflag, "**fail");
MPIR_ERR_ADD(mpi_errno_ret, mpi_errno);
last_recv_cnt = 0;
} else {
MPIR_Get_count_impl(&status, recvtype, &last_recv_cnt);
}
curr_cnt += last_recv_cnt;
}
tmp_mask >>= 1;
k--;
}
}
/* --END EXPERIMENTAL-- */
mask <<= 1;
i++;
}
static int MPIDI_CH3i_Progress_wait(MPID_Progress_state * progress_state)
{
MPIDU_Sock_event_t event;
int mpi_errno = MPI_SUCCESS;
MPIDI_STATE_DECL(MPID_STATE_MPIDI_CH3I_PROGRESS_WAIT);
MPIDI_FUNC_ENTER(MPID_STATE_MPIDI_CH3I_PROGRESS_WAIT);
/*
* MT: the following code will be needed if progress can occur between
* MPIDI_CH3_Progress_start() and
* MPIDI_CH3_Progress_wait(), or iterations of MPIDI_CH3_Progress_wait().
*
* This is presently not possible, and thus the code is commented out.
*/
# if 0
/* FIXME: Was (USE_THREAD_IMPL == MPICH_THREAD_IMPL_NOT_IMPLEMENTED),
which really meant not-using-global-mutex-thread model . This
was true for the single threaded case, but was probably not intended
for that case*/
{
if (progress_state->ch.completion_count != MPIDI_CH3I_progress_completion_count)
{
goto fn_exit;
}
}
# endif
# ifdef MPICH_IS_THREADED
MPIU_THREAD_CHECK_BEGIN
{
if (MPIDI_CH3I_progress_blocked == TRUE)
{
/*
* Another thread is already blocking in the progress engine.
*
* MT: Another thread is already blocking in poll. Right now,
* calls to MPIDI_CH3_Progress_wait() are effectively
* serialized by the device. The only way another thread may
* enter this function is if MPIDU_Sock_wait() blocks. If
* this changes, a flag other than MPIDI_CH3I_Progress_blocked
* may be required to determine if another thread is in
* the progress engine.
*/
MPIDI_CH3I_Progress_delay(MPIDI_CH3I_progress_completion_count);
goto fn_exit;
}
}
MPIU_THREAD_CHECK_END
# endif
do
{
int made_progress = FALSE;
/* make progress on NBC schedules, must come before we block on sock_wait */
mpi_errno = MPIDU_Sched_progress(&made_progress);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
if (made_progress) {
MPIDI_CH3_Progress_signal_completion();
break;
}
# ifdef MPICH_IS_THREADED
/* The logic for this case is just complicated enough that
we write separate code for each possibility */
# ifdef HAVE_RUNTIME_THREADCHECK
if (MPIR_ThreadInfo.isThreaded) {
MPIDI_CH3I_progress_blocked = TRUE;
mpi_errno = MPIDU_Sock_wait(MPIDI_CH3I_sock_set,
MPIDU_SOCK_INFINITE_TIME, &event);
MPIDI_CH3I_progress_blocked = FALSE;
MPIDI_CH3I_progress_wakeup_signalled = FALSE;
}
else {
mpi_errno = MPIDU_Sock_wait(MPIDI_CH3I_sock_set,
MPIDU_SOCK_INFINITE_TIME, &event);
}
# else
MPIDI_CH3I_progress_blocked = TRUE;
mpi_errno = MPIDU_Sock_wait(MPIDI_CH3I_sock_set,
MPIDU_SOCK_INFINITE_TIME, &event);
MPIDI_CH3I_progress_blocked = FALSE;
MPIDI_CH3I_progress_wakeup_signalled = FALSE;
# endif /* HAVE_RUNTIME_THREADCHECK */
# else
mpi_errno = MPIDU_Sock_wait(MPIDI_CH3I_sock_set,
MPIDU_SOCK_INFINITE_TIME, &event);
# endif
/* --BEGIN ERROR HANDLING-- */
if (mpi_errno != MPI_SUCCESS)
{
MPIU_Assert(MPIR_ERR_GET_CLASS(mpi_errno) != MPIDU_SOCK_ERR_TIMEOUT);
MPIU_ERR_SET(mpi_errno,MPI_ERR_OTHER,"**progress_sock_wait");
goto fn_fail;
}
/* --END ERROR HANDLING-- */
mpi_errno = MPIDI_CH3I_Progress_handle_sock_event(&event);
if (mpi_errno != MPI_SUCCESS) {
MPIU_ERR_SETANDJUMP(mpi_errno,MPI_ERR_OTHER,
"**ch3|sock|handle_sock_event");
}
}
while (progress_state->ch.completion_count == MPIDI_CH3I_progress_completion_count);
/*
* We could continue to call MPIU_Sock_wait in a non-blocking fashion
* and process any other events; however, this would not
* give the application a chance to post new receives, and thus could
* result in an increased number of unexpected messages
* that would need to be buffered.
*/
# if MPICH_IS_THREADED
{
/*
* Awaken any threads which are waiting for the progress that just
* occurred
*/
MPIDI_CH3I_Progress_continue(MPIDI_CH3I_progress_completion_count);
}
# endif
fn_exit:
/*
* Reset the progress state so it is fresh for the next iteration
*/
progress_state->ch.completion_count = MPIDI_CH3I_progress_completion_count;
MPIDI_FUNC_EXIT(MPID_STATE_MPIDI_CH3I_PROGRESS_WAIT);
return mpi_errno;
fn_fail:
goto fn_exit;
}
int MPIR_Barrier_inter( MPID_Comm *comm_ptr, MPIR_Errflag_t *errflag )
{
int rank, mpi_errno = MPI_SUCCESS, root;
int mpi_errno_ret = MPI_SUCCESS;
int i = 0;
MPID_Comm *newcomm_ptr = NULL;
rank = comm_ptr->rank;
/* Get the local intracommunicator */
if (!comm_ptr->local_comm) {
mpi_errno = MPIR_Setup_intercomm_localcomm( comm_ptr );
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
}
newcomm_ptr = comm_ptr->local_comm;
/* do a barrier on the local intracommunicator */
mpi_errno = MPIR_Barrier_intra(newcomm_ptr, errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag = MPIR_ERR_GET_CLASS(mpi_errno);
MPIR_ERR_SET(mpi_errno, *errflag, "**fail");
MPIR_ERR_ADD(mpi_errno_ret, mpi_errno);
}
/* rank 0 on each group does an intercommunicator broadcast to the
remote group to indicate that all processes in the local group
have reached the barrier. We do a 1-byte bcast because a 0-byte
bcast will just return without doing anything. */
/* first broadcast from left to right group, then from right to
left group */
if (comm_ptr->is_low_group) {
/* bcast to right*/
root = (rank == 0) ? MPI_ROOT : MPI_PROC_NULL;
mpi_errno = MPIR_Bcast_inter(&i, 1, MPI_BYTE, root, comm_ptr, errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag = MPIR_ERR_GET_CLASS(mpi_errno);
MPIR_ERR_SET(mpi_errno, *errflag, "**fail");
MPIR_ERR_ADD(mpi_errno_ret, mpi_errno);
}
/* receive bcast from right */
root = 0;
mpi_errno = MPIR_Bcast_inter(&i, 1, MPI_BYTE, root, comm_ptr, errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag = MPIR_ERR_GET_CLASS(mpi_errno);
MPIR_ERR_SET(mpi_errno, *errflag, "**fail");
MPIR_ERR_ADD(mpi_errno_ret, mpi_errno);
}
}
else {
/* receive bcast from left */
root = 0;
mpi_errno = MPIR_Bcast_inter(&i, 1, MPI_BYTE, root, comm_ptr, errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag = MPIR_ERR_GET_CLASS(mpi_errno);
MPIR_ERR_SET(mpi_errno, *errflag, "**fail");
MPIR_ERR_ADD(mpi_errno_ret, mpi_errno);
}
/* bcast to left */
root = (rank == 0) ? MPI_ROOT : MPI_PROC_NULL;
mpi_errno = MPIR_Bcast_inter(&i, 1, MPI_BYTE, root, comm_ptr, errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag = MPIR_ERR_GET_CLASS(mpi_errno);
MPIR_ERR_SET(mpi_errno, *errflag, "**fail");
MPIR_ERR_ADD(mpi_errno_ret, mpi_errno);
}
}
fn_exit:
if (mpi_errno_ret)
mpi_errno = mpi_errno_ret;
else if (*errflag != MPIR_ERR_NONE)
MPIR_ERR_SET(mpi_errno, *errflag, "**coll_fail");
return mpi_errno;
fn_fail:
goto fn_exit;
}
static int MPIDI_CH3i_Progress_test(void)
{
MPIDU_Sock_event_t event;
int mpi_errno = MPI_SUCCESS;
int made_progress;
MPIDI_STATE_DECL(MPID_STATE_MPIDI_CH3I_PROGRESS_TEST);
MPIDI_FUNC_ENTER(MPID_STATE_MPIDI_CH3I_PROGRESS_TEST);
# ifdef MPICH_IS_THREADED
{
/* We don't bother testing whether threads are enabled in the
runtime-checking case because this simple test will always be false
if threads are not enabled. */
if (MPIDI_CH3I_progress_blocked == TRUE)
{
/*
* Another thread is already blocking in the progress engine.
* We are not going to block waiting for progress, so we
* simply return. It might make sense to yield before * returning,
* giving the PE thread a change to make progress.
*
* MT: Another thread is already blocking in poll. Right now,
* calls to the progress routines are effectively
* serialized by the device. The only way another thread may
* enter this function is if MPIDU_Sock_wait() blocks. If
* this changes, a flag other than MPIDI_CH3I_Progress_blocked
* may be required to determine if another thread is in
* the progress engine.
*/
goto fn_exit;
}
}
# endif
/* make progress on NBC schedules */
mpi_errno = MPIDU_Sched_progress(&made_progress);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
mpi_errno = MPIDU_Sock_wait(MPIDI_CH3I_sock_set, 0, &event);
if (mpi_errno == MPI_SUCCESS)
{
mpi_errno = MPIDI_CH3I_Progress_handle_sock_event(&event);
if (mpi_errno != MPI_SUCCESS) {
MPIU_ERR_SETANDJUMP(mpi_errno,MPI_ERR_OTHER,
"**ch3|sock|handle_sock_event");
}
}
else if (MPIR_ERR_GET_CLASS(mpi_errno) == MPIDU_SOCK_ERR_TIMEOUT)
{
mpi_errno = MPI_SUCCESS;
goto fn_exit;
}
else {
MPIU_ERR_SETANDJUMP(mpi_errno,MPI_ERR_OTHER, "**progress_sock_wait");
}
fn_exit:
MPIDI_FUNC_EXIT(MPID_STATE_MPIDI_CH3I_PROGRESS_TEST);
return mpi_errno;
fn_fail:
goto fn_exit;
}
int MPIR_Allgather_intra_brucks (
const void *sendbuf,
int sendcount,
MPI_Datatype sendtype,
void *recvbuf,
int recvcount,
MPI_Datatype recvtype,
MPIR_Comm *comm_ptr,
MPIR_Errflag_t *errflag )
{
int comm_size, rank;
int mpi_errno = MPI_SUCCESS;
int mpi_errno_ret = MPI_SUCCESS;
MPI_Aint recvtype_extent;
MPI_Aint recvtype_true_extent, recvbuf_extent, recvtype_true_lb;
int pof2, src, rem;
void *tmp_buf = NULL;
int curr_cnt, dst;
MPIR_CHKLMEM_DECL(1);
if (((sendcount == 0) && (sendbuf != MPI_IN_PLACE)) || (recvcount == 0))
return MPI_SUCCESS;
comm_size = comm_ptr->local_size;
rank = comm_ptr->rank;
MPIR_Datatype_get_extent_macro( recvtype, recvtype_extent );
/* This is the largest offset we add to recvbuf */
MPIR_Ensure_Aint_fits_in_pointer(MPIR_VOID_PTR_CAST_TO_MPI_AINT recvbuf +
(comm_size * recvcount * recvtype_extent));
/* allocate 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 * (MPL_MAX(recvtype_true_extent, recvtype_extent));
MPIR_CHKLMEM_MALLOC(tmp_buf, void*, recvbuf_extent, mpi_errno, "tmp_buf", MPL_MEM_BUFFER);
/* adjust for potential negative lower bound in datatype */
tmp_buf = (void *)((char*)tmp_buf - recvtype_true_lb);
/* copy local data to the top of tmp_buf */
if (sendbuf != MPI_IN_PLACE) {
mpi_errno = MPIR_Localcopy (sendbuf, sendcount, sendtype,
tmp_buf, recvcount, recvtype);
if (mpi_errno) {
MPIR_ERR_POP(mpi_errno);
}
}
else {
mpi_errno = MPIR_Localcopy (((char *)recvbuf +
rank * recvcount * recvtype_extent),
recvcount, recvtype, tmp_buf,
recvcount, recvtype);
if (mpi_errno) {
MPIR_ERR_POP(mpi_errno);
}
}
/* do the first \floor(\lg p) steps */
curr_cnt = recvcount;
pof2 = 1;
while (pof2 <= comm_size/2) {
src = (rank + pof2) % comm_size;
dst = (rank - pof2 + comm_size) % comm_size;
mpi_errno = MPIC_Sendrecv(tmp_buf, curr_cnt, recvtype, dst,
MPIR_ALLGATHER_TAG,
((char *)tmp_buf + curr_cnt*recvtype_extent),
curr_cnt, recvtype,
src, MPIR_ALLGATHER_TAG, comm_ptr,
MPI_STATUS_IGNORE, errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag = MPIR_ERR_GET_CLASS(mpi_errno);
MPIR_ERR_SET(mpi_errno, *errflag, "**fail");
MPIR_ERR_ADD(mpi_errno_ret, mpi_errno);
}
curr_cnt *= 2;
pof2 *= 2;
}
/* if comm_size is not a power of two, one more step is needed */
rem = comm_size - pof2;
if (rem) {
src = (rank + pof2) % comm_size;
dst = (rank - pof2 + comm_size) % comm_size;
mpi_errno = MPIC_Sendrecv(tmp_buf, rem * recvcount, recvtype,
dst, MPIR_ALLGATHER_TAG,
((char *)tmp_buf + curr_cnt*recvtype_extent),
rem * recvcount, recvtype,
src, MPIR_ALLGATHER_TAG, comm_ptr,
MPI_STATUS_IGNORE, errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag = MPIR_ERR_GET_CLASS(mpi_errno);
MPIR_ERR_SET(mpi_errno, *errflag, "**fail");
MPIR_ERR_ADD(mpi_errno_ret, mpi_errno);
}
}
/* Rotate blocks in tmp_buf down by (rank) blocks and store
* result in recvbuf. */
mpi_errno = MPIR_Localcopy(tmp_buf, (comm_size-rank)*recvcount,
recvtype, (char *) recvbuf + rank*recvcount*recvtype_extent,
(comm_size-rank)*recvcount, recvtype);
if (mpi_errno) {
MPIR_ERR_POP(mpi_errno);
}
if (rank) {
mpi_errno = MPIR_Localcopy((char *) tmp_buf +
(comm_size-rank)*recvcount*recvtype_extent,
rank*recvcount, recvtype, recvbuf,
rank*recvcount, recvtype);
if (mpi_errno) {
MPIR_ERR_POP(mpi_errno);
}
}
fn_exit:
MPIR_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;
}
static int MPIR_Scan_generic (
const void *sendbuf,
void *recvbuf,
int count,
MPI_Datatype datatype,
MPI_Op op,
MPID_Comm *comm_ptr,
MPIR_Errflag_t *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;
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_size = comm_ptr->local_size;
rank = comm_ptr->rank;
/* set op_errno to 0. stored in perthread structure */
{
MPIR_Per_thread_t *per_thread = NULL;
int err = 0;
MPID_THREADPRIV_KEY_GET_ADDR(MPIR_ThreadInfo.isThreaded, MPIR_Per_thread_key,
MPIR_Per_thread, per_thread, &err);
MPIU_Assert(err == 0);
per_thread->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*(MPL_MAX(extent,true_extent)), mpi_errno, "partial_scan");
/* This eventually gets malloc()ed as a temp buffer, not added to
* any user buffers */
MPIU_Ensure_Aint_fits_in_pointer(count * MPL_MAX(extent, true_extent));
/* adjust for potential negative lower bound in datatype */
partial_scan = (void *)((char*)partial_scan - true_lb);
/* need to allocate temporary buffer to store incoming data*/
MPIU_CHKLMEM_MALLOC(tmp_buf, void *, count*(MPL_MAX(extent,true_extent)), mpi_errno, "tmp_buf");
/* adjust for potential negative lower bound in datatype */
tmp_buf = (void *)((char*)tmp_buf - true_lb);
/* Since this is an inclusive scan, copy local contribution into
recvbuf. */
if (sendbuf != MPI_IN_PLACE) {
mpi_errno = MPIR_Localcopy(sendbuf, count, datatype,
recvbuf, count, datatype);
if (mpi_errno) MPIR_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) MPIR_ERR_POP(mpi_errno);
mask = 0x1;
while (mask < comm_size) {
dst = rank ^ mask;
if (dst < comm_size) {
/* Send partial_scan to dst. Recv into tmp_buf */
mpi_errno = MPIC_Sendrecv(partial_scan, count, datatype,
dst, MPIR_SCAN_TAG, tmp_buf,
count, datatype, dst,
MPIR_SCAN_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);
}
if (rank > dst) {
mpi_errno = MPIR_Reduce_local_impl(
tmp_buf, partial_scan, count, datatype, op);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
mpi_errno = MPIR_Reduce_local_impl(
tmp_buf, recvbuf, count, datatype, op);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
}
else {
if (is_commutative) {
mpi_errno = MPIR_Reduce_local_impl(
tmp_buf, partial_scan, count, datatype, op);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
}
else {
mpi_errno = MPIR_Reduce_local_impl(
partial_scan, tmp_buf, count, datatype, op);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
mpi_errno = MPIR_Localcopy(tmp_buf, count, datatype,
partial_scan,
count, datatype);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
}
}
}
mask <<= 1;
}
{
MPIR_Per_thread_t *per_thread = NULL;
int err = 0;
MPID_THREADPRIV_KEY_GET_ADDR(MPIR_ThreadInfo.isThreaded, MPIR_Per_thread_key,
MPIR_Per_thread, per_thread, &err);
MPIU_Assert(err == 0);
if (per_thread->op_errno) {
mpi_errno = per_thread->op_errno;
if (mpi_errno) MPIR_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 != MPIR_ERR_NONE)
MPIR_ERR_SET(mpi_errno, *errflag, "**coll_fail");
return mpi_errno;
fn_fail:
goto fn_exit;
}
/* This is the main function which generates a tree in shared memory. The tree is parameterized
* over the different data-structures:
* k_val : the tree K-value
* shared_region : the shared memory region where the tree will be generated
* max_entries_per_level : the maximum number of ranks per level
* ranks_per_package : the different ranks at each level
* max_ranks_per_package : the maximum ranks in any package
* package_ctr : number of ranks in each package
* package_level : the topology level where we cutoff the tree
* num_ranks : the number of ranks
* */
int MPIDI_SHM_gen_tree(int k_val, int *shared_region, int *max_entries_per_level,
int **ranks_per_package, int max_ranks_per_package, int *package_ctr,
int package_level, int num_ranks, bool package_leaders_first,
bool right_skewed, MPIR_Errflag_t * errflag)
{
int mpi_errno = MPI_SUCCESS, mpi_errno_ret = MPI_SUCCESS;
int i, j, p, r, rank, idx;
int num_packages = max_entries_per_level[package_level];
int package_count = 0;
MPIDI_SHM_topotree_t package_tree, tree, template_tree;
const int package_tree_sz = num_packages > num_ranks ? num_packages : num_ranks;
int *package_leaders = NULL;
MPIR_CHKPMEM_DECL(1);
mpi_errno = MPIDI_SHM_topotree_allocate(&tree, num_ranks, k_val);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag =
MPIX_ERR_PROC_FAILED ==
MPIR_ERR_GET_CLASS(mpi_errno) ? MPIR_ERR_PROC_FAILED : MPIR_ERR_OTHER;
MPIR_ERR_SET(mpi_errno, *errflag, "**fail");
MPIR_ERR_ADD(mpi_errno_ret, mpi_errno);
}
mpi_errno = MPIDI_SHM_topotree_allocate(&package_tree, package_tree_sz, k_val);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag =
MPIX_ERR_PROC_FAILED ==
MPIR_ERR_GET_CLASS(mpi_errno) ? MPIR_ERR_PROC_FAILED : MPIR_ERR_OTHER;
MPIR_ERR_SET(mpi_errno, *errflag, "**fail");
MPIR_ERR_ADD(mpi_errno_ret, mpi_errno);
}
MPIR_CHKPMEM_CALLOC(package_leaders, int *, num_packages * sizeof(int), mpi_errno,
"intra_node_package_leaders", MPL_MEM_OTHER);
/* We pick package leaders as the first rank in each package */
for (p = 0; p < max_entries_per_level[package_level]; ++p) {
package_leaders[p] = -1;
if (package_ctr[p] > 0) {
package_leaders[package_count++] = ranks_per_package[p][0];
}
}
num_packages = package_count;
/* STEP 4. Now use the template tree to generate the top level tree */
MPIDI_SHM_gen_package_tree(num_packages, k_val, &package_tree, package_leaders);
/* STEP 5. Create a template tree for the ranks */
mpi_errno =
MPIDI_SHM_create_template_tree(&template_tree, k_val, right_skewed,
max_ranks_per_package, errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag =
MPIX_ERR_PROC_FAILED ==
MPIR_ERR_GET_CLASS(mpi_errno) ? MPIR_ERR_PROC_FAILED : MPIR_ERR_OTHER;
MPIR_ERR_SET(mpi_errno, *errflag, "**fail");
MPIR_ERR_ADD(mpi_errno_ret, mpi_errno);
}
if (MPIDI_SHM_TOPOTREE_DEBUG) {
for (i = 0; i < max_entries_per_level[package_level]; ++i) {
fprintf(stderr, "pre-Rank %d, parent %d, children=%d [", i,
MPIDI_SHM_TOPOTREE_PARENT(&tree, i), MPIDI_SHM_TOPOTREE_NUM_CHILD(&tree, i));
for (j = 0; j < MPIDI_SHM_TOPOTREE_NUM_CHILD(&tree, i); ++j) {
fprintf(stderr, "%d, ", MPIDI_SHM_TOPOTREE_CHILD(&tree, i, j));
}
fprintf(stderr, "]\n");
}
}
/* use the template tree to generate the tree for each rank */
for (p = 0; p < max_entries_per_level[package_level]; ++p) {
for (r = 0; r < package_ctr[p]; ++r) {
rank = ranks_per_package[p][r];
if (MPIDI_SHM_TOPOTREE_DEBUG)
fprintf(stderr, "Rank=%d, p=%d, r=%d, opt1=%d, opt2=%d\n", rank, p, r,
MPIDI_SHM_TOPOTREE_PARENT(&template_tree, r),
ranks_per_package[p][MPIDI_SHM_TOPOTREE_PARENT(&template_tree, r)]);
if (MPIDI_SHM_TOPOTREE_PARENT(&template_tree, r) == -1) {
MPIDI_SHM_TOPOTREE_PARENT(&tree, rank) = -1;
} else {
MPIDI_SHM_TOPOTREE_PARENT(&tree, rank) =
ranks_per_package[p][MPIDI_SHM_TOPOTREE_PARENT(&template_tree, r)];
}
for (j = 0; j < MPIDI_SHM_TOPOTREE_NUM_CHILD(&template_tree, r); ++j) {
idx = MPIDI_SHM_TOPOTREE_NUM_CHILD(&tree, rank);
if (MPIDI_SHM_TOPOTREE_CHILD(&template_tree, r, j) < package_ctr[p]) {
MPIDI_SHM_TOPOTREE_NUM_CHILD(&tree, rank)++;
MPIDI_SHM_TOPOTREE_CHILD(&tree, rank, idx) =
ranks_per_package[p][MPIDI_SHM_TOPOTREE_CHILD(&template_tree, r, j)];
}
}
}
}
if (MPIDI_SHM_TOPOTREE_DEBUG) {
char str[1024], tmp[128];
for (i = 0; i < num_ranks; ++i) {
sprintf(str, "*BaseTreeRank %d, parent %d, children=%d [", i,
MPIDI_SHM_TOPOTREE_PARENT(&tree, i), MPIDI_SHM_TOPOTREE_NUM_CHILD(&tree, i));
for (j = 0; j < MPIDI_SHM_TOPOTREE_NUM_CHILD(&tree, i); ++j) {
sprintf(tmp, "%d, ", MPIDI_SHM_TOPOTREE_CHILD(&tree, i, j));
strcat(str, tmp);
}
fprintf(stderr, "%s]\n", str);
}
}
/* Assemble the per package tree package leaders tree and copy it to shared memory region */
MPIDI_SHM_gen_tree_sharedmemory(shared_region, &tree, &package_tree, package_leaders,
num_packages, num_ranks, k_val, package_leaders_first);
MPL_free(tree.base);
MPL_free(package_tree.base);
MPL_free(template_tree.base);
fn_exit:
MPIR_CHKPMEM_REAP();
return mpi_errno;
fn_fail:
goto fn_exit;
}
static int barrier_smp_intra(MPID_Comm *comm_ptr, MPIR_Errflag_t *errflag)
{
int mpi_errno=MPI_SUCCESS;
int mpi_errno_ret = MPI_SUCCESS;
MPIU_Assert(MPIR_CVAR_ENABLE_SMP_COLLECTIVES && MPIR_CVAR_ENABLE_SMP_BARRIER &&
MPIR_Comm_is_node_aware(comm_ptr));
#if defined(FINEGRAIN_MPI)
int colocated_size = -1;
int colocated_sense = -1;
/* do barrier on osproc_colocated_comm */
if (comm_ptr->osproc_colocated_comm != NULL)
{
colocated_size = comm_ptr->osproc_colocated_comm->local_size;
MPIU_Assert( (comm_ptr->osproc_colocated_comm->co_shared_vars != NULL) && (comm_ptr->osproc_colocated_comm->co_shared_vars->co_barrier_vars != NULL) );
MPIU_Assert(colocated_size > 1 );
colocated_sense = comm_ptr->osproc_colocated_comm->co_shared_vars->co_barrier_vars->coproclet_signal;
if( comm_ptr->osproc_colocated_comm->rank != 0 ) { /* non-leader */
(comm_ptr->osproc_colocated_comm->co_shared_vars->co_barrier_vars->coproclet_counter)++;
if (comm_ptr->osproc_colocated_comm->co_shared_vars->co_barrier_vars->coproclet_counter == (colocated_size-1)){ /* excluding the leader */
comm_ptr->osproc_colocated_comm->co_shared_vars->co_barrier_vars->leader_signal = 1;
}
while(comm_ptr->osproc_colocated_comm->co_shared_vars->co_barrier_vars->coproclet_signal == colocated_sense) {
FG_Yield();
}
}
else { /* leader */
while(comm_ptr->osproc_colocated_comm->co_shared_vars->co_barrier_vars->leader_signal == 0) {
FG_Yield();
}
}
#if 0 /* Non-optimized version */
mpi_errno = MPIR_Barrier_impl(comm_ptr->osproc_colocated_comm, errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag = MPIR_ERR_GET_CLASS(mpi_errno);
MPIU_ERR_SET(mpi_errno, *errflag, "**fail");
MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
}
#endif
}
#endif
/* do the intranode barrier on all nodes */
if (comm_ptr->node_comm != NULL)
{
mpi_errno = MPIR_Barrier_impl(comm_ptr->node_comm, errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag = MPIR_ERR_GET_CLASS(mpi_errno);
MPIR_ERR_SET(mpi_errno, *errflag, "**fail");
MPIR_ERR_ADD(mpi_errno_ret, mpi_errno);
}
}
/* do the barrier across roots of all nodes */
if (comm_ptr->node_roots_comm != NULL) {
mpi_errno = MPIR_Barrier_impl(comm_ptr->node_roots_comm, errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag = MPIR_ERR_GET_CLASS(mpi_errno);
MPIR_ERR_SET(mpi_errno, *errflag, "**fail");
MPIR_ERR_ADD(mpi_errno_ret, mpi_errno);
}
}
/* release the local processes on each node with a 1-byte
broadcast (0-byte broadcast just returns without doing
anything) */
if (comm_ptr->node_comm != NULL)
{
int i=0;
mpi_errno = MPIR_Bcast_impl(&i, 1, MPI_BYTE, 0, comm_ptr->node_comm, errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag = MPIR_ERR_GET_CLASS(mpi_errno);
MPIR_ERR_SET(mpi_errno, *errflag, "**fail");
MPIR_ERR_ADD(mpi_errno_ret, mpi_errno);
}
}
#if defined(FINEGRAIN_MPI)
if (comm_ptr->osproc_colocated_comm != NULL)
{
if (comm_ptr->osproc_colocated_comm->rank == 0) { /* leader */
comm_ptr->osproc_colocated_comm->co_shared_vars->co_barrier_vars->leader_signal = 0;
comm_ptr->osproc_colocated_comm->co_shared_vars->co_barrier_vars->coproclet_counter = 0;
comm_ptr->osproc_colocated_comm->co_shared_vars->co_barrier_vars->coproclet_signal = 1 - comm_ptr->osproc_colocated_comm->co_shared_vars->co_barrier_vars->coproclet_signal;
}
#if 0 /* Non-optimized version */
/* release the colocated processes in each OS-process with a 1-byte
broadcast (0-byte broadcast just returns without doing
anything) */
int i=0;
mpi_errno = MPIR_Bcast_impl(&i, 1, MPI_BYTE, 0, comm_ptr->osproc_colocated_comm, errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag = MPIR_ERR_GET_CLASS(mpi_errno);
MPIU_ERR_SET(mpi_errno, *errflag, "**fail");
MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
}
#endif
}
#endif
fn_exit:
if (mpi_errno_ret)
mpi_errno = mpi_errno_ret;
else if (*errflag != MPIR_ERR_NONE)
MPIR_ERR_SET(mpi_errno, *errflag, "**coll_fail");
return mpi_errno;
fn_fail:
goto fn_exit;
}
int MPIR_Bcast_inter_remote_send_local_bcast (
void *buffer,
int count,
MPI_Datatype datatype,
int root,
MPIR_Comm *comm_ptr,
MPIR_Errflag_t *errflag)
{
int rank, mpi_errno;
int mpi_errno_ret = MPI_SUCCESS;
MPI_Status status;
MPIR_Comm *newcomm_ptr = NULL;
MPIR_FUNC_TERSE_STATE_DECL(MPID_STATE_MPIR_BCAST_INTER);
MPIR_FUNC_TERSE_ENTER(MPID_STATE_MPIR_BCAST_INTER);
if (root == MPI_PROC_NULL)
{
/* local processes other than root do nothing */
mpi_errno = MPI_SUCCESS;
}
else if (root == MPI_ROOT)
{
/* root sends to rank 0 on remote group and returns */
mpi_errno = MPIC_Send(buffer, count, datatype, 0,
MPIR_BCAST_TAG, comm_ptr, 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
{
/* remote group. rank 0 on remote group receives from root */
rank = comm_ptr->rank;
if (rank == 0)
{
mpi_errno = MPIC_Recv(buffer, count, datatype, root,
MPIR_BCAST_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);
}
}
/* Get the local intracommunicator */
if (!comm_ptr->local_comm)
MPII_Setup_intercomm_localcomm( comm_ptr );
newcomm_ptr = comm_ptr->local_comm;
/* now do the usual broadcast on this intracommunicator
with rank 0 as root. */
mpi_errno = MPIR_Bcast_intra_auto(buffer, count, datatype, 0, newcomm_ptr, 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);
}
}
MPIR_FUNC_TERSE_EXIT(MPID_STATE_MPIR_BCAST_INTER);
/* --BEGIN ERROR HANDLING-- */
if (mpi_errno_ret)
mpi_errno = mpi_errno_ret;
else if (*errflag != MPIR_ERR_NONE)
MPIR_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**coll_fail");
/* --END ERROR HANDLING-- */
return mpi_errno;
}
int MPIR_Alltoallv_inter(const void *sendbuf, const int *sendcounts, const int *sdispls,
MPI_Datatype sendtype, void *recvbuf, const int *recvcounts,
const int *rdispls, MPI_Datatype recvtype, MPID_Comm *comm_ptr,
MPIR_Errflag_t *errflag)
{
/* Intercommunicator alltoallv. We use a pairwise exchange algorithm
similar to the one used in intracommunicator alltoallv. Since the
local and remote groups can be of different
sizes, we first compute the max of local_group_size,
remote_group_size. At step i, 0 <= i < max_size, each process
receives from src = (rank - i + max_size) % max_size if src <
remote_size, and sends to dst = (rank + i) % max_size if dst <
remote_size.
FIXME: change algorithm to match intracommunicator alltoallv
*/
int local_size, remote_size, max_size, i;
MPI_Aint send_extent, recv_extent;
int mpi_errno = MPI_SUCCESS;
int mpi_errno_ret = MPI_SUCCESS;
MPI_Status status;
int src, dst, rank, sendcount, recvcount;
char *sendaddr, *recvaddr;
local_size = comm_ptr->local_size;
remote_size = comm_ptr->remote_size;
rank = comm_ptr->rank;
/* Get extent of send and recv types */
MPID_Datatype_get_extent_macro(sendtype, send_extent);
MPID_Datatype_get_extent_macro(recvtype, recv_extent);
/* check if multiple threads are calling this collective function */
MPIDU_ERR_CHECK_MULTIPLE_THREADS_ENTER( comm_ptr );
/* Use pairwise exchange algorithm. */
max_size = MPIR_MAX(local_size, remote_size);
for (i=0; i<max_size; i++) {
src = (rank - i + max_size) % max_size;
dst = (rank + i) % max_size;
if (src >= remote_size) {
src = MPI_PROC_NULL;
recvaddr = NULL;
recvcount = 0;
}
else {
MPIU_Ensure_Aint_fits_in_pointer(MPIU_VOID_PTR_CAST_TO_MPI_AINT recvbuf +
rdispls[src]*recv_extent);
recvaddr = (char *)recvbuf + rdispls[src]*recv_extent;
recvcount = recvcounts[src];
}
if (dst >= remote_size) {
dst = MPI_PROC_NULL;
sendaddr = NULL;
sendcount = 0;
}
else {
MPIU_Ensure_Aint_fits_in_pointer(MPIU_VOID_PTR_CAST_TO_MPI_AINT sendbuf +
sdispls[dst]*send_extent);
sendaddr = (char *)sendbuf + sdispls[dst]*send_extent;
sendcount = sendcounts[dst];
}
mpi_errno = MPIC_Sendrecv(sendaddr, sendcount, sendtype, dst,
MPIR_ALLTOALLV_TAG, recvaddr, recvcount,
recvtype, src, 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);
}
}
fn_exit:
/* check if multiple threads are calling this collective function */
MPIDU_ERR_CHECK_MULTIPLE_THREADS_EXIT( comm_ptr );
if (mpi_errno_ret)
mpi_errno = mpi_errno_ret;
else if (*errflag != MPIR_ERR_NONE)
MPIR_ERR_SET(mpi_errno, *errflag, "**coll_fail");
return mpi_errno;
fn_fail:
goto fn_exit;
}
int MPIDI_CH3_iSend(MPIDI_VC_t * vc, MPIR_Request * sreq, void * hdr,
intptr_t hdr_sz)
{
int mpi_errno = MPI_SUCCESS;
int (*reqFn)(MPIDI_VC_t *, MPIR_Request *, int *);
MPIDI_CH3I_VC *vcch = &vc->ch;
MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_MPIDI_CH3_ISEND);
MPIR_FUNC_VERBOSE_ENTER(MPID_STATE_MPIDI_CH3_ISEND);
MPIR_Assert( hdr_sz <= sizeof(MPIDI_CH3_Pkt_t) );
/* The sock channel uses a fixed length header, the size of which is the
maximum of all possible packet headers */
hdr_sz = sizeof(MPIDI_CH3_Pkt_t);
MPL_DBG_STMT(MPIDI_CH3_DBG_CHANNEL,VERBOSE,
MPIDI_DBG_Print_packet((MPIDI_CH3_Pkt_t*)hdr));
if (vcch->state == MPIDI_CH3I_VC_STATE_CONNECTED) /* MT */
{
/* Connection already formed. If send queue is empty attempt to send
data, queuing any unsent data. */
if (MPIDI_CH3I_SendQ_empty(vcch)) /* MT */
{
size_t nb;
int rc;
MPL_DBG_MSG(MPIDI_CH3_DBG_CHANNEL,VERBOSE,
"send queue empty, attempting to write");
MPL_DBG_PKT(vcch->conn,hdr,"isend");
/* MT: need some signalling to lock down our right to use the
channel, thus insuring that the progress engine does
also try to write */
rc = MPIDI_CH3I_Sock_write(vcch->sock, hdr, hdr_sz, &nb);
if (rc == MPI_SUCCESS)
{
MPL_DBG_MSG_D(MPIDI_CH3_DBG_CHANNEL,VERBOSE,
"wrote %ld bytes", (unsigned long) nb);
if (nb == hdr_sz)
{
MPL_DBG_MSG_D(MPIDI_CH3_DBG_CHANNEL,VERBOSE,
"write complete %" PRIdPTR " bytes, calling OnDataAvail fcn", nb);
reqFn = sreq->dev.OnDataAvail;
if (!reqFn) {
MPIR_Assert(MPIDI_Request_get_type(sreq)!=MPIDI_REQUEST_TYPE_GET_RESP);
mpi_errno = MPID_Request_complete(sreq);
if (mpi_errno != MPI_SUCCESS) {
MPIR_ERR_POP(mpi_errno);
}
}
else {
int complete;
mpi_errno = reqFn( vc, sreq, &complete );
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
if (!complete) {
MPIDI_CH3I_SendQ_enqueue_head(vcch, sreq);
MPL_DBG_MSG_FMT(MPIDI_CH3_DBG_CHANNEL,VERBOSE,
(MPL_DBG_FDEST,
"posting writev, vc=0x%p, sreq=0x%08x", vc, sreq->handle));
vcch->conn->send_active = sreq;
mpi_errno = MPIDI_CH3I_Sock_post_writev(
vcch->conn->sock, sreq->dev.iov,
sreq->dev.iov_count, NULL);
/* --BEGIN ERROR HANDLING-- */
if (mpi_errno != MPI_SUCCESS)
{
mpi_errno = MPIR_Err_create_code(mpi_errno, MPIR_ERR_FATAL, FCNAME, __LINE__, MPI_ERR_OTHER,
"**ch3|sock|postwrite", "ch3|sock|postwrite %p %p %p",
sreq, vcch->conn, vc);
}
/* --END ERROR HANDLING-- */
}
}
}
else
{
MPL_DBG_MSG_D(MPIDI_CH3_DBG_CHANNEL,VERBOSE,
"partial write of %" PRIdPTR " bytes, request enqueued at head", nb);
update_request(sreq, hdr, hdr_sz, nb);
MPIDI_CH3I_SendQ_enqueue_head(vcch, sreq);
MPL_DBG_MSG_FMT(MPIDI_CH3_DBG_CHANNEL,VERBOSE,
(MPL_DBG_FDEST,"posting write, vc=0x%p, sreq=0x%08x", vc, sreq->handle));
vcch->conn->send_active = sreq;
mpi_errno = MPIDI_CH3I_Sock_post_write(vcch->conn->sock,
sreq->dev.iov[0].MPL_IOV_BUF,
sreq->dev.iov[0].MPL_IOV_LEN,
sreq->dev.iov[0].MPL_IOV_LEN, NULL);
/* --BEGIN ERROR HANDLING-- */
if (mpi_errno != MPI_SUCCESS)
{
mpi_errno = MPIR_Err_create_code(mpi_errno, MPIR_ERR_FATAL, FCNAME, __LINE__, MPI_ERR_OTHER,
"**ch3|sock|postwrite", "ch3|sock|postwrite %p %p %p",
sreq, vcch->conn, vc);
}
/* --END ERROR HANDLING-- */
}
}
/* --BEGIN ERROR HANDLING-- */
else if (MPIR_ERR_GET_CLASS(rc) == MPIDI_CH3I_SOCK_ERR_NOMEM)
{
MPL_DBG_MSG(MPIDI_CH3_DBG_CHANNEL,TYPICAL,
"MPIDI_CH3I_Sock_write failed, out of memory");
sreq->status.MPI_ERROR = MPIR_ERR_MEMALLOCFAILED;
}
else
{
MPL_DBG_MSG_D(MPIDI_CH3_DBG_CHANNEL,TYPICAL,
"MPIDI_CH3I_Sock_write failed, rc=%d", rc);
/* Connection just failed. Mark the request complete and
return an error. */
MPL_DBG_VCCHSTATECHANGE(vc,VC_STATE_FAILED);
/* FIXME: Shouldn't the vc->state also change? */
vcch->state = MPIDI_CH3I_VC_STATE_FAILED;
sreq->status.MPI_ERROR = MPIR_Err_create_code( rc,
MPIR_ERR_RECOVERABLE, FCNAME, __LINE__,
MPI_ERR_INTERN, "**ch3|sock|writefailed",
"**ch3|sock|writefailed %d", rc );
/* MT -CH3U_Request_complete() performs write barrier */
MPID_Request_complete(sreq);
/* Make sure that the caller sees this error */
mpi_errno = sreq->status.MPI_ERROR;
}
/* --END ERROR HANDLING-- */
}
else
{
MPL_DBG_MSG(MPIDI_CH3_DBG_CHANNEL,VERBOSE,"send queue not empty, enqueuing");
update_request(sreq, hdr, hdr_sz, 0);
MPIDI_CH3I_SendQ_enqueue(vcch, sreq);
}
}
else if (vcch->state == MPIDI_CH3I_VC_STATE_CONNECTING) /* MT */
{
/* queuing the data so it can be sent later. */
MPL_DBG_VCUSE(vc,"connecting. enqueuing request");
update_request(sreq, hdr, hdr_sz, 0);
MPIDI_CH3I_SendQ_enqueue(vcch, sreq);
}
else if (vcch->state == MPIDI_CH3I_VC_STATE_UNCONNECTED) /* MT */
{
/* Form a new connection, queuing the data so it can be sent later. */
MPL_DBG_VCUSE(vc,"unconnected. enqueuing request");
update_request(sreq, hdr, hdr_sz, 0);
MPIDI_CH3I_SendQ_enqueue(vcch, sreq);
mpi_errno = MPIDI_CH3I_VC_post_connect(vc);
if (mpi_errno) {
MPIR_ERR_POP(mpi_errno);
}
}
else if (vcch->state != MPIDI_CH3I_VC_STATE_FAILED)
{
/* Unable to send data at the moment, so queue it for later */
MPL_DBG_VCUSE(vc,"still connecting. Enqueuing request");
update_request(sreq, hdr, hdr_sz, 0);
MPIDI_CH3I_SendQ_enqueue(vcch, sreq);
}
/* --BEGIN ERROR HANDLING-- */
else
{
/* Connection failed. Mark the request complete and return an error. */
/* TODO: Create an appropriate error message */
sreq->status.MPI_ERROR = MPI_ERR_INTERN;
/* MT - CH3U_Request_complete() performs write barrier */
MPID_Request_complete(sreq);
}
/* --END ERROR HANDLING-- */
fn_fail:
MPIR_FUNC_VERBOSE_EXIT(MPID_STATE_MPIDI_CH3_ISEND);
return mpi_errno;
}
int MPIR_Allreduce_intra_reduce_scatter_allgather(
const void *sendbuf,
void *recvbuf,
int count,
MPI_Datatype datatype,
MPI_Op op,
MPIR_Comm * comm_ptr,
MPIR_Errflag_t * errflag)
{
MPIR_CHKLMEM_DECL(3);
#ifdef MPID_HAS_HETERO
int is_homogeneous;
int rc;
#endif
int comm_size, rank;
int mpi_errno = MPI_SUCCESS;
int mpi_errno_ret = MPI_SUCCESS;
int mask, dst, 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;
comm_size = comm_ptr->local_size;
rank = comm_ptr->rank;
/* need to allocate temporary buffer to store incoming data*/
MPIR_Type_get_true_extent_impl(datatype, &true_lb, &true_extent);
MPIR_Datatype_get_extent_macro(datatype, extent);
MPIR_Ensure_Aint_fits_in_pointer(count * MPL_MAX(extent, true_extent));
MPIR_CHKLMEM_MALLOC(tmp_buf, void *, count*(MPL_MAX(extent,true_extent)), mpi_errno, "temporary buffer", MPL_MEM_BUFFER);
/* adjust for potential negative lower bound in datatype */
tmp_buf = (void *)((char*)tmp_buf - true_lb);
/* copy local data into recvbuf */
if (sendbuf != MPI_IN_PLACE) {
mpi_errno = MPIR_Localcopy(sendbuf, count, datatype, recvbuf,
count, datatype);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
}
/* get nearest power-of-two less than or equal to comm_size */
pof2 = comm_ptr->pof2;
rem = comm_size - pof2;
/* In the non-power-of-two case, all even-numbered
processes of rank < 2*rem send their data to
(rank+1). These even-numbered processes no longer
participate in the algorithm until the very end. The
remaining processes form a nice power-of-two. */
if (rank < 2*rem) {
if (rank % 2 == 0) { /* even */
mpi_errno = MPIC_Send(recvbuf, count,
datatype, rank+1,
MPIR_ALLREDUCE_TAG, comm_ptr, 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);
}
/* temporarily set the rank to -1 so that this
process does not pariticipate in recursive
doubling */
newrank = -1;
}
else { /* odd */
mpi_errno = MPIC_Recv(tmp_buf, count,
datatype, rank-1,
MPIR_ALLREDUCE_TAG, comm_ptr,
MPI_STATUS_IGNORE, errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag = MPIR_ERR_GET_CLASS(mpi_errno);
MPIR_ERR_SET(mpi_errno, *errflag, "**fail");
MPIR_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(tmp_buf, recvbuf, count, datatype, op);
if (mpi_errno) MPIR_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.) */
#ifdef HAVE_ERROR_CHECKING
MPIR_Assert(HANDLE_GET_KIND(op)==HANDLE_KIND_BUILTIN);
MPIR_Assert(count >= pof2);
#endif /* HAVE_ERROR_CHECKING */
if (newrank != -1) {
MPIR_CHKLMEM_MALLOC(cnts, int *, pof2*sizeof(int), mpi_errno, "counts", MPL_MEM_BUFFER);
MPIR_CHKLMEM_MALLOC(disps, int *, pof2*sizeof(int), mpi_errno, "displacements", MPL_MEM_BUFFER);
for (i=0; i<pof2; i++)
cnts[i] = count/pof2;
if ((count % pof2) > 0) {
for (i=0; i<(count % pof2); i++)
cnts[i] += 1;
}
disps[0] = 0;
for (i=1; i<pof2; i++)
disps[i] = disps[i-1] + cnts[i-1];
mask = 0x1;
send_idx = recv_idx = 0;
last_idx = pof2;
while (mask < pof2) {
newdst = newrank ^ mask;
/* find real rank of dest */
dst = (newdst < rem) ? newdst*2 + 1 : newdst + rem;
send_cnt = recv_cnt = 0;
if (newrank < newdst) {
send_idx = recv_idx + pof2/(mask*2);
for (i=send_idx; i<last_idx; i++)
send_cnt += cnts[i];
for (i=recv_idx; i<send_idx; i++)
recv_cnt += cnts[i];
}
else {
recv_idx = send_idx + pof2/(mask*2);
for (i=send_idx; i<recv_idx; i++)
send_cnt += cnts[i];
for (i=recv_idx; i<last_idx; i++)
recv_cnt += cnts[i];
}
/* Send data from recvbuf. Recv into tmp_buf */
mpi_errno = MPIC_Sendrecv((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_ptr,
MPI_STATUS_IGNORE, 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);
}
/* 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(((char *) tmp_buf + disps[recv_idx]*extent),
((char *) recvbuf + disps[recv_idx]*extent),
recv_cnt, datatype, op);
if (mpi_errno) MPIR_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((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_ptr,
MPI_STATUS_IGNORE, errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag = MPIR_ERR_GET_CLASS(mpi_errno);
MPIR_ERR_SET(mpi_errno, *errflag, "**fail");
MPIR_ERR_ADD(mpi_errno_ret, mpi_errno);
}
if (newrank > newdst) send_idx = recv_idx;
mask >>= 1;
}
}
int MPIR_Exscan (
const void *sendbuf,
void *recvbuf,
int count,
MPI_Datatype datatype,
MPI_Op op,
MPIR_Comm *comm_ptr,
MPIR_Errflag_t *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;
MPIR_Op *op_ptr;
MPIR_CHKLMEM_DECL(2);
if (count == 0) return MPI_SUCCESS;
comm_size = comm_ptr->local_size;
rank = comm_ptr->rank;
/* set op_errno to 0. stored in perthread structure */
{
MPIR_Per_thread_t *per_thread = NULL;
int err = 0;
MPID_THREADPRIV_KEY_GET_ADDR(MPIR_ThreadInfo.isThreaded, MPIR_Per_thread_key,
MPIR_Per_thread, per_thread, &err);
MPIR_Assert(err == 0);
per_thread->op_errno = 0;
}
if (HANDLE_GET_KIND(op) == HANDLE_KIND_BUILTIN) {
is_commutative = 1;
}
else {
MPIR_Op_get_ptr(op, op_ptr);
if (op_ptr->kind == MPIR_OP_KIND__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 );
MPIR_CHKLMEM_MALLOC(partial_scan, void *, (count*(MPL_MAX(true_extent,extent))), mpi_errno, "partial_scan");
/* adjust for potential negative lower bound in datatype */
partial_scan = (void *)((char*)partial_scan - true_lb);
/* need to allocate temporary buffer to store incoming data*/
MPIR_CHKLMEM_MALLOC(tmp_buf, void *, (count*(MPL_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 ? (const void *)recvbuf : sendbuf), count, datatype,
partial_scan, count, datatype);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
flag = 0;
mask = 0x1;
while (mask < comm_size) {
dst = rank ^ mask;
if (dst < comm_size) {
/* Send partial_scan to dst. Recv into tmp_buf */
mpi_errno = MPIC_Sendrecv(partial_scan, count, datatype,
dst, MPIR_EXSCAN_TAG, tmp_buf,
count, datatype, dst,
MPIR_EXSCAN_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);
}
if (rank > dst) {
mpi_errno = MPIR_Reduce_local_impl( tmp_buf, partial_scan,
count, datatype, op );
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
/* 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) MPIR_ERR_POP(mpi_errno);
flag = 1;
}
else {
mpi_errno = MPIR_Reduce_local_impl( tmp_buf,
recvbuf, count, datatype, op );
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
}
}
}
else {
if (is_commutative) {
mpi_errno = MPIR_Reduce_local_impl( tmp_buf, partial_scan,
count, datatype, op );
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
}
else {
mpi_errno = MPIR_Reduce_local_impl( partial_scan,
tmp_buf, count, datatype, op );
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
mpi_errno = MPIR_Localcopy(tmp_buf, count, datatype,
partial_scan,
count, datatype);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
}
}
}
mask <<= 1;
}
{
MPIR_Per_thread_t *per_thread = NULL;
int err = 0;
MPID_THREADPRIV_KEY_GET_ADDR(MPIR_ThreadInfo.isThreaded, MPIR_Per_thread_key,
MPIR_Per_thread, per_thread, &err);
MPIR_Assert(err == 0);
if (per_thread->op_errno)
mpi_errno = per_thread->op_errno;
}
fn_exit:
MPIR_CHKLMEM_FREEALL();
if (mpi_errno_ret)
mpi_errno = mpi_errno_ret;
else if (*errflag != MPIR_ERR_NONE)
MPIR_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**coll_fail");
return mpi_errno;
fn_fail:
goto fn_exit;
}
