/* try to get a small message out on to the wire quickly */
static inline int mca_pml_ob1_send_inline (const void *buf, size_t count,
ompi_datatype_t * datatype,
int dst, int tag, int16_t seqn,
ompi_proc_t *dst_proc, mca_bml_base_endpoint_t* endpoint,
ompi_communicator_t * comm)
{
mca_pml_ob1_match_hdr_t match;
mca_bml_base_btl_t *bml_btl;
opal_convertor_t convertor;
size_t size;
int rc;
bml_btl = mca_bml_base_btl_array_get_next(&endpoint->btl_eager);
if( NULL == bml_btl->btl->btl_sendi)
return OMPI_ERR_NOT_AVAILABLE;
ompi_datatype_type_size (datatype, &size);
if ((size * count) > 256) { /* some random number */
return OMPI_ERR_NOT_AVAILABLE;
}
if (count > 0) {
/* initialize just enough of the convertor to avoid a SEGV in opal_convertor_cleanup */
OBJ_CONSTRUCT(&convertor, opal_convertor_t);
/* We will create a convertor specialized for the */
/* remote architecture and prepared with the datatype. */
opal_convertor_copy_and_prepare_for_send (dst_proc->super.proc_convertor,
(const struct opal_datatype_t *) datatype,
count, buf, 0, &convertor);
opal_convertor_get_packed_size (&convertor, &size);
} else {
size = 0;
}
mca_pml_ob1_match_hdr_prepare (&match, MCA_PML_OB1_HDR_TYPE_MATCH, 0,
comm->c_contextid, comm->c_my_rank,
tag, seqn);
ob1_hdr_hton(&match, MCA_PML_OB1_HDR_TYPE_MATCH, dst_proc);
/* try to send immediately */
rc = mca_bml_base_sendi (bml_btl, &convertor, &match, OMPI_PML_OB1_MATCH_HDR_LEN,
size, MCA_BTL_NO_ORDER, MCA_BTL_DES_FLAGS_PRIORITY | MCA_BTL_DES_FLAGS_BTL_OWNERSHIP,
MCA_PML_OB1_HDR_TYPE_MATCH, NULL);
if (count > 0) {
opal_convertor_cleanup (&convertor);
}
if (OPAL_UNLIKELY(OMPI_SUCCESS != rc)) {
return rc;
}
return (int) size;
}
static mca_pml_yalla_convertor_t *mca_pml_yalla_get_send_convertor(void *buf, size_t count,
ompi_datatype_t *datatype)
{
mca_pml_yalla_convertor_t *convertor = PML_YALLA_FREELIST_GET(&ompi_pml_yalla.convs);
convertor->datatype = datatype;
OBJ_RETAIN(datatype);
opal_convertor_copy_and_prepare_for_send(ompi_proc_local_proc->proc_convertor,
&datatype->super, count, buf, 0,
&convertor->convertor);
return convertor;
}
static inline int create_iov_list(const void *addr, int count, ompi_datatype_t *datatype,
ucx_iovec_t **ucx_iov, uint32_t *ucx_iov_count) {
int ret = OMPI_SUCCESS;
size_t size;
bool done = false;
opal_convertor_t convertor;
uint32_t iov_count, iov_idx;
struct iovec iov[OSC_UCX_IOVEC_MAX];
uint32_t ucx_iov_idx;
OBJ_CONSTRUCT(&convertor, opal_convertor_t);
ret = opal_convertor_copy_and_prepare_for_send(ompi_mpi_local_convertor,
&datatype->super, count,
addr, 0, &convertor);
if (ret != OMPI_SUCCESS) {
return ret;
}
(*ucx_iov_count) = 0;
ucx_iov_idx = 0;
do {
iov_count = OSC_UCX_IOVEC_MAX;
iov_idx = 0;
done = opal_convertor_raw(&convertor, iov, &iov_count, &size);
(*ucx_iov_count) += iov_count;
(*ucx_iov) = (ucx_iovec_t *)realloc((*ucx_iov), (*ucx_iov_count) * sizeof(ucx_iovec_t));
if (*ucx_iov == NULL) {
return OMPI_ERR_TEMP_OUT_OF_RESOURCE;
}
while (iov_idx != iov_count) {
(*ucx_iov)[ucx_iov_idx].addr = iov[iov_idx].iov_base;
(*ucx_iov)[ucx_iov_idx].len = iov[iov_idx].iov_len;
ucx_iov_idx++;
iov_idx++;
}
assert((*ucx_iov_count) == ucx_iov_idx);
} while (!done);
opal_convertor_cleanup(&convertor);
OBJ_DESTRUCT(&convertor);
return ret;
}
/*
* opal_datatype_sndrcv
*
* Function: - copy MPI message from buffer into another
* - send/recv done if cannot optimize
* Accepts: - send buffer
* - send count
* - send datatype
* - receive buffer
* - receive count
* - receive datatype
* - tag
* - communicator
* Returns: - MPI_SUCCESS or error code
*/
int32_t ompi_datatype_sndrcv( void *sbuf, int32_t scount, const ompi_datatype_t* sdtype,
void *rbuf, int32_t rcount, const ompi_datatype_t* rdtype)
{
opal_convertor_t send_convertor, recv_convertor;
struct iovec iov;
int length, completed;
uint32_t iov_count;
size_t max_data;
/* First check if we really have something to do */
if (0 == rcount || 0 == rdtype->super.size) {
return ((0 == scount || 0 == sdtype->super.size) ? MPI_SUCCESS : MPI_ERR_TRUNCATE);
}
/* If same datatypes used, just copy. */
if (sdtype == rdtype) {
int32_t count = ( scount < rcount ? scount : rcount );
opal_datatype_copy_content_same_ddt(&(rdtype->super), count, (char*)rbuf, (char*)sbuf);
return ((scount > rcount) ? MPI_ERR_TRUNCATE : MPI_SUCCESS);
}
/* If receive packed. */
if (rdtype->id == OMPI_DATATYPE_MPI_PACKED) {
OBJ_CONSTRUCT( &send_convertor, opal_convertor_t );
opal_convertor_copy_and_prepare_for_send( ompi_mpi_local_convertor,
&(sdtype->super), scount, sbuf, 0,
&send_convertor );
iov_count = 1;
iov.iov_base = (IOVBASE_TYPE*)rbuf;
iov.iov_len = scount * sdtype->super.size;
if( (int32_t)iov.iov_len > rcount ) iov.iov_len = rcount;
opal_convertor_pack( &send_convertor, &iov, &iov_count, &max_data );
OBJ_DESTRUCT( &send_convertor );
return ((max_data < (size_t)rcount) ? MPI_ERR_TRUNCATE : MPI_SUCCESS);
}
/* If send packed. */
if (sdtype->id == OMPI_DATATYPE_MPI_PACKED) {
OBJ_CONSTRUCT( &recv_convertor, opal_convertor_t );
opal_convertor_copy_and_prepare_for_recv( ompi_mpi_local_convertor,
&(rdtype->super), rcount, rbuf, 0,
&recv_convertor );
iov_count = 1;
iov.iov_base = (IOVBASE_TYPE*)sbuf;
iov.iov_len = rcount * rdtype->super.size;
if( (int32_t)iov.iov_len > scount ) iov.iov_len = scount;
opal_convertor_unpack( &recv_convertor, &iov, &iov_count, &max_data );
OBJ_DESTRUCT( &recv_convertor );
return (((size_t)scount > max_data) ? MPI_ERR_TRUNCATE : MPI_SUCCESS);
}
iov.iov_len = length = 64 * 1024;
iov.iov_base = (IOVBASE_TYPE*)malloc( length * sizeof(char) );
OBJ_CONSTRUCT( &send_convertor, opal_convertor_t );
opal_convertor_copy_and_prepare_for_send( ompi_mpi_local_convertor,
&(sdtype->super), scount, sbuf, 0,
&send_convertor );
OBJ_CONSTRUCT( &recv_convertor, opal_convertor_t );
opal_convertor_copy_and_prepare_for_recv( ompi_mpi_local_convertor,
&(rdtype->super), rcount, rbuf, 0,
&recv_convertor );
completed = 0;
while( !completed ) {
iov.iov_len = length;
iov_count = 1;
max_data = length;
completed |= opal_convertor_pack( &send_convertor, &iov, &iov_count, &max_data );
completed |= opal_convertor_unpack( &recv_convertor, &iov, &iov_count, &max_data );
}
free( iov.iov_base );
OBJ_DESTRUCT( &send_convertor );
OBJ_DESTRUCT( &recv_convertor );
return ( (scount * sdtype->super.size) <= (rcount * rdtype->super.size) ? MPI_SUCCESS : MPI_ERR_TRUNCATE );
}
/* try to get a small message out on to the wire quickly */
static inline int mca_pml_ob1_send_inline (void *buf, size_t count,
ompi_datatype_t * datatype,
int dst, int tag, int16_t seqn,
ompi_proc_t *dst_proc, mca_bml_base_endpoint_t* endpoint,
ompi_communicator_t * comm)
{
mca_btl_base_descriptor_t *des = NULL;
mca_pml_ob1_match_hdr_t match;
mca_bml_base_btl_t *bml_btl;
OPAL_PTRDIFF_TYPE lb, extent;
opal_convertor_t convertor;
size_t size = 0;
int rc;
bml_btl = mca_bml_base_btl_array_get_next(&endpoint->btl_eager);
ompi_datatype_get_extent (datatype, &lb, &extent);
if (OPAL_UNLIKELY((extent * count) > 256 || !bml_btl->btl->btl_sendi)) {
return OMPI_ERR_NOT_AVAILABLE;
}
if (count > 0) {
/* initialize just enough of the convertor to avoid a SEGV in opal_convertor_cleanup */
OBJ_CONSTRUCT(&convertor, opal_convertor_t);
/* We will create a convertor specialized for the */
/* remote architecture and prepared with the datatype. */
opal_convertor_copy_and_prepare_for_send (dst_proc->proc_convertor,
(const struct opal_datatype_t *) datatype,
count, buf, 0, &convertor);
opal_convertor_get_packed_size (&convertor, &size);
}
match.hdr_common.hdr_flags = 0;
match.hdr_common.hdr_type = MCA_PML_OB1_HDR_TYPE_MATCH;
match.hdr_ctx = comm->c_contextid;
match.hdr_src = comm->c_my_rank;
match.hdr_tag = tag;
match.hdr_seq = seqn;
ob1_hdr_hton(&match, MCA_PML_OB1_HDR_TYPE_MATCH, dst_proc);
/* try to send immediately */
rc = mca_bml_base_sendi (bml_btl, &convertor, &match, OMPI_PML_OB1_MATCH_HDR_LEN,
size, MCA_BTL_NO_ORDER, MCA_BTL_DES_FLAGS_PRIORITY | MCA_BTL_DES_FLAGS_BTL_OWNERSHIP,
MCA_PML_OB1_HDR_TYPE_MATCH, &des);
if (count > 0) {
opal_convertor_cleanup (&convertor);
}
if (OPAL_UNLIKELY(OMPI_SUCCESS != rc)) {
if (des) {
mca_bml_base_free (bml_btl, des);
}
return rc;
}
return (int) size;
}
static inline int ompi_osc_rdma_gacc_master (ompi_osc_rdma_sync_t *sync, const void *source_buffer, int source_count,
ompi_datatype_t *source_datatype, void *result_buffer, int result_count,
ompi_datatype_t *result_datatype, ompi_osc_rdma_peer_t *peer, uint64_t target_address,
mca_btl_base_registration_handle_t *target_handle, int target_count,
ompi_datatype_t *target_datatype, ompi_op_t *op, ompi_osc_rdma_request_t *request)
{
ompi_osc_rdma_module_t *module = sync->module;
struct iovec source_iovec[OMPI_OSC_RDMA_DECODE_MAX], target_iovec[OMPI_OSC_RDMA_DECODE_MAX];
const size_t acc_limit = (mca_osc_rdma_component.buffer_size >> 3);
uint32_t source_primitive_count, target_primitive_count;
opal_convertor_t source_convertor, target_convertor;
uint32_t source_iov_count, target_iov_count;
uint32_t source_iov_index, target_iov_index;
ompi_datatype_t *source_primitive, *target_primitive;
/* needed for opal_convertor_raw but not used */
size_t source_size, target_size;
ompi_osc_rdma_request_t *subreq;
size_t result_position;
ptrdiff_t lb, extent;
int ret, acc_len;
bool done;
(void) ompi_datatype_get_extent (target_datatype, &lb, &extent);
target_address += lb;
/* fast path for accumulate on built-in types */
if (OPAL_LIKELY((!source_count || ompi_datatype_is_predefined (source_datatype)) &&
ompi_datatype_is_predefined (target_datatype) &&
(!result_count || ompi_datatype_is_predefined (result_datatype)) &&
(target_datatype->super.size * target_count <= acc_limit))) {
if (NULL == request) {
OMPI_OSC_RDMA_REQUEST_ALLOC(module, peer, request);
request->internal = true;
request->type = result_datatype ? OMPI_OSC_RDMA_TYPE_GET_ACC : OMPI_OSC_RDMA_TYPE_ACC;
}
if (source_datatype) {
(void) ompi_datatype_get_extent (source_datatype, &lb, &extent);
source_buffer = (void *)((intptr_t) source_buffer + lb);
}
if (result_datatype) {
(void) ompi_datatype_get_extent (result_datatype, &lb, &extent);
result_buffer = (void *)((intptr_t) result_buffer + lb);
}
ret = ompi_osc_rdma_gacc_contig (sync, source_buffer, source_count, source_datatype, result_buffer,
result_count, result_datatype, peer, target_address,
target_handle, target_count, target_datatype, op,
request);
if (OPAL_LIKELY(OMPI_SUCCESS == ret)) {
return OMPI_SUCCESS;
}
if (source_datatype) {
/* the convertors will handle the lb */
(void) ompi_datatype_get_extent (source_datatype, &lb, &extent);
source_buffer = (void *)((intptr_t) source_buffer - lb);
}
if (result_datatype) {
(void) ompi_datatype_get_extent (result_datatype, &lb, &extent);
result_buffer = (void *)((intptr_t) result_buffer - lb);
}
}
/* the convertor will handle lb from here */
(void) ompi_datatype_get_extent (target_datatype, &lb, &extent);
target_address -= lb;
/* get the primitive datatype info */
ret = ompi_osc_base_get_primitive_type_info (target_datatype, &target_primitive, &target_primitive_count);
if (OPAL_UNLIKELY(OMPI_SUCCESS != ret)) {
/* target datatype is not made up of a single basic datatype */
return ret;
}
if (source_datatype) {
ret = ompi_osc_base_get_primitive_type_info (source_datatype, &source_primitive, &source_primitive_count);
if (OPAL_UNLIKELY(OMPI_SUCCESS != ret)) {
/* target datatype is not made up of a single basic datatype */
return ret;
}
if (OPAL_UNLIKELY(source_primitive != target_primitive)) {
return MPI_ERR_TYPE;
}
}
/* prepare convertors for the source and target. these convertors will be used to determine the
* contiguous segments within the source and target. */
/* the source may be NULL if using MPI_OP_NO_OP with MPI_Get_accumulate */
if (source_datatype) {
OBJ_CONSTRUCT(&source_convertor, opal_convertor_t);
ret = opal_convertor_copy_and_prepare_for_send (ompi_mpi_local_convertor, &source_datatype->super, source_count, source_buffer,
0, &source_convertor);
if (OPAL_UNLIKELY(OMPI_SUCCESS != ret)) {
return ret;
}
}
/* target_datatype can never be NULL */
OBJ_CONSTRUCT(&target_convertor, opal_convertor_t);
ret = opal_convertor_copy_and_prepare_for_send (ompi_mpi_local_convertor, &target_datatype->super, target_count,
(void *) (intptr_t) target_address, 0, &target_convertor);
if (OPAL_UNLIKELY(OMPI_SUCCESS != ret)) {
return ret;
}
if (request) {
/* keep the request from completing until all the transfers have started */
request->outstanding_requests = 1;
}
target_iov_index = 0;
target_iov_count = 0;
result_position = 0;
do {
/* decode segments of the source data */
source_iov_count = OMPI_OSC_RDMA_DECODE_MAX;
source_iov_index = 0;
/* opal_convertor_raw returns done when it has reached the end of the data */
if (!source_datatype) {
done = true;
source_iovec[0].iov_len = (size_t) -1;
source_iovec[0].iov_base = NULL;
source_iov_count = 1;
} else {
done = opal_convertor_raw (&source_convertor, source_iovec, &source_iov_count, &source_size);
}
/* loop on the target segments until we have exhaused the decoded source data */
while (source_iov_index != source_iov_count) {
if (target_iov_index == target_iov_count) {
/* decode segments of the target buffer */
target_iov_count = OMPI_OSC_RDMA_DECODE_MAX;
target_iov_index = 0;
(void) opal_convertor_raw (&target_convertor, target_iovec, &target_iov_count, &target_size);
}
/* we already checked that the target was large enough. this should be impossible */
assert (0 != target_iov_count);
/* determine how much to put in this operation */
acc_len = min(target_iovec[target_iov_index].iov_len, source_iovec[source_iov_index].iov_len);
acc_len = min((size_t) acc_len, acc_limit);
/* execute the get */
OMPI_OSC_RDMA_REQUEST_ALLOC(module, peer, subreq);
subreq->internal = true;
subreq->parent_request = request;
if (request) {
(void) OPAL_THREAD_ADD32 (&request->outstanding_requests, 1);
}
if (result_datatype) {
/* prepare a convertor for this part of the result */
opal_convertor_copy_and_prepare_for_recv (ompi_mpi_local_convertor, &result_datatype->super, result_count,
result_buffer, 0, &subreq->convertor);
opal_convertor_set_position (&subreq->convertor, &result_position);
subreq->type = OMPI_OSC_RDMA_TYPE_GET_ACC;
} else {
subreq->type = OMPI_OSC_RDMA_TYPE_ACC;
}
OPAL_OUTPUT_VERBOSE((60, ompi_osc_base_framework.framework_output,
"target index = %d, target = {%p, %lu}, source_index = %d, source = {%p, %lu}, result = %p, result position = %lu, "
"acc_len = %d, count = %lu",
target_iov_index, target_iovec[target_iov_index].iov_base, (unsigned long) target_iovec[target_iov_index].iov_len,
source_iov_index, source_iovec[source_iov_index].iov_base, (unsigned long) source_iovec[source_iov_index].iov_len,
result_buffer, (unsigned long) result_position, acc_len, (unsigned long)(acc_len / target_primitive->super.size)));
ret = ompi_osc_rdma_gacc_contig (sync, source_iovec[source_iov_index].iov_base, acc_len / target_primitive->super.size,
target_primitive, NULL, 0, NULL, peer, (uint64_t) (intptr_t) target_iovec[target_iov_index].iov_base,
target_handle, acc_len / target_primitive->super.size, target_primitive, op, subreq);
if (OPAL_UNLIKELY(OMPI_SUCCESS != ret)) {
if (OPAL_UNLIKELY(OMPI_ERR_OUT_OF_RESOURCE != ret)) {
/* something bad happened. need to figure out how to handle these errors */
return ret;
}
/* progress and try again */
ompi_osc_rdma_progress (module);
continue;
}
/* adjust io vectors */
target_iovec[target_iov_index].iov_len -= acc_len;
source_iovec[source_iov_index].iov_len -= acc_len;
target_iovec[target_iov_index].iov_base = (void *)((intptr_t) target_iovec[target_iov_index].iov_base + acc_len);
source_iovec[source_iov_index].iov_base = (void *)((intptr_t) source_iovec[source_iov_index].iov_base + acc_len);
result_position += acc_len;
source_iov_index += !source_datatype || (0 == source_iovec[source_iov_index].iov_len);
target_iov_index += (0 == target_iovec[target_iov_index].iov_len);
}
} while (!done);
if (request) {
/* release our reference so the request can complete */
(void) OPAL_THREAD_ADD32 (&request->outstanding_requests, -1);
}
if (source_datatype) {
opal_convertor_cleanup (&source_convertor);
OBJ_DESTRUCT(&source_convertor);
}
opal_convertor_cleanup (&target_convertor);
OBJ_DESTRUCT(&target_convertor);
return OMPI_SUCCESS;
}
/**
* Shared memory broadcast.
*
* For the root, the general algorithm is to wait for a set of
* segments to become available. Once it is, the root claims the set
* by writing the current operation number and the number of processes
* using the set to the flag. The root then loops over the set of
* segments; for each segment, it copies a fragment of the user's
* buffer into the shared data segment and then writes the data size
* into its childrens' control buffers. The process is repeated until
* all fragments have been written.
*
* For non-roots, for each set of buffers, they wait until the current
* operation number appears in the in-use flag (i.e., written by the
* root). Then for each segment, they wait for a nonzero to appear
* into their control buffers. If they have children, they copy the
* data from their parent's shared data segment into their shared data
* segment, and write the data size into each of their childrens'
* control buffers. They then copy the data from their shared [local]
* data segment into the user's output buffer. The process is
* repeated until all fragments have been received. If they do not
* have children, they copy the data directly from the parent's shared
* data segment into the user's output buffer.
*/
int mca_coll_sm_bcast_intra(void *buff, int count,
struct ompi_datatype_t *datatype, int root,
struct ompi_communicator_t *comm,
mca_coll_base_module_t *module)
{
struct iovec iov;
mca_coll_sm_module_t *sm_module = (mca_coll_sm_module_t*) module;
mca_coll_sm_comm_t *data;
int i, ret, rank, size, num_children, src_rank;
int flag_num, segment_num, max_segment_num;
int parent_rank;
size_t total_size, max_data, bytes;
mca_coll_sm_in_use_flag_t *flag;
opal_convertor_t convertor;
mca_coll_sm_tree_node_t *me, *parent, **children;
mca_coll_sm_data_index_t *index;
/* Lazily enable the module the first time we invoke a collective
on it */
if (!sm_module->enabled) {
if (OMPI_SUCCESS != (ret = ompi_coll_sm_lazy_enable(module, comm))) {
return ret;
}
}
data = sm_module->sm_comm_data;
/* Setup some identities */
rank = ompi_comm_rank(comm);
size = ompi_comm_size(comm);
OBJ_CONSTRUCT(&convertor, opal_convertor_t);
iov.iov_len = mca_coll_sm_component.sm_fragment_size;
bytes = 0;
me = &data->mcb_tree[(rank + size - root) % size];
parent = me->mcstn_parent;
children = me->mcstn_children;
num_children = me->mcstn_num_children;
/* Only have one top-level decision as to whether I'm the root or
not. Do this at the slight expense of repeating a little logic
-- but it's better than a conditional branch in every loop
iteration. */
/*********************************************************************
* Root
*********************************************************************/
if (root == rank) {
/* The root needs a send convertor to pack from the user's
buffer to shared memory */
if (OMPI_SUCCESS !=
(ret =
opal_convertor_copy_and_prepare_for_send(ompi_mpi_local_convertor,
&(datatype->super),
count,
buff,
0,
&convertor))) {
return ret;
}
opal_convertor_get_packed_size(&convertor, &total_size);
/* Main loop over sending fragments */
do {
flag_num = (data->mcb_operation_count++ %
mca_coll_sm_component.sm_comm_num_in_use_flags);
FLAG_SETUP(flag_num, flag, data);
FLAG_WAIT_FOR_IDLE(flag, bcast_root_label);
FLAG_RETAIN(flag, size - 1, data->mcb_operation_count - 1);
/* Loop over all the segments in this set */
segment_num =
flag_num * mca_coll_sm_component.sm_segs_per_inuse_flag;
max_segment_num =
(flag_num + 1) * mca_coll_sm_component.sm_segs_per_inuse_flag;
do {
index = &(data->mcb_data_index[segment_num]);
/* Copy the fragment from the user buffer to my fragment
in the current segment */
max_data = mca_coll_sm_component.sm_fragment_size;
COPY_FRAGMENT_IN(convertor, index, rank, iov, max_data);
bytes += max_data;
/* Wait for the write to absolutely complete */
opal_atomic_wmb();
/* Tell my children that this fragment is ready */
PARENT_NOTIFY_CHILDREN(children, num_children, index,
max_data);
++segment_num;
} while (bytes < total_size && segment_num < max_segment_num);
} while (bytes < total_size);
}
/*********************************************************************
* Non-root
*********************************************************************/
else {
/* Non-root processes need a receive convertor to unpack from
shared mmory to the user's buffer */
if (OMPI_SUCCESS !=
(ret =
opal_convertor_copy_and_prepare_for_recv(ompi_mpi_local_convertor,
&(datatype->super),
count,
buff,
0,
&convertor))) {
return ret;
}
opal_convertor_get_packed_size(&convertor, &total_size);
/* Loop over receiving (and possibly re-sending) the
fragments */
do {
flag_num = (data->mcb_operation_count %
mca_coll_sm_component.sm_comm_num_in_use_flags);
/* Wait for the root to mark this set of segments as
ours */
FLAG_SETUP(flag_num, flag, data);
FLAG_WAIT_FOR_OP(flag, data->mcb_operation_count, bcast_nonroot_label1);
++data->mcb_operation_count;
/* Loop over all the segments in this set */
segment_num =
flag_num * mca_coll_sm_component.sm_segs_per_inuse_flag;
max_segment_num =
(flag_num + 1) * mca_coll_sm_component.sm_segs_per_inuse_flag;
do {
/* Pre-calculate some values */
parent_rank = (parent->mcstn_id + root) % size;
index = &(data->mcb_data_index[segment_num]);
/* Wait for my parent to tell me that the segment is ready */
CHILD_WAIT_FOR_NOTIFY(rank, index, max_data, bcast_nonroot_label2);
/* If I have children, send the data to them */
if (num_children > 0) {
/* Copy the fragment from the parent's portion in
the segment to my portion in the segment. */
COPY_FRAGMENT_BETWEEN(parent_rank, rank, index, max_data);
/* Wait for the write to absolutely complete */
opal_atomic_wmb();
/* Tell my children that this fragment is ready */
PARENT_NOTIFY_CHILDREN(children, num_children, index,
max_data);
/* Set the "copy from buffer" to be my local
segment buffer so that we don't potentially
incur a non-local memory copy from the parent's
fan out data segment [again] when copying to
the user's buffer */
src_rank = rank;
}
/* If I don't have any children, set the "copy from
buffer" to be my parent's fan out segment to copy
directly from my parent */
else {
src_rank = parent_rank;
}
/* Copy to my output buffer */
COPY_FRAGMENT_OUT(convertor, src_rank, index, iov, max_data);
bytes += max_data;
++segment_num;
} while (bytes < total_size && segment_num < max_segment_num);
/* Wait for all copy-out writes to complete before I say
I'm done with the segments */
opal_atomic_wmb();
/* We're finished with this set of segments */
FLAG_RELEASE(flag);
} while (bytes < total_size);
}
/* Kill the convertor */
OBJ_DESTRUCT(&convertor);
/* All done */
return OMPI_SUCCESS;
}
