int ompi_coll_tuned_reduce_intra_binary( void *sendbuf, void *recvbuf,
int count, ompi_datatype_t* datatype,
ompi_op_t* op, int root,
ompi_communicator_t* comm,
mca_coll_base_module_t *module,
uint32_t segsize,
int max_outstanding_reqs )
{
int segcount = count;
size_t typelng;
mca_coll_tuned_module_t *tuned_module = (mca_coll_tuned_module_t*) module;
mca_coll_tuned_comm_t *data = tuned_module->tuned_data;
OPAL_OUTPUT((ompi_coll_tuned_stream,"coll:tuned:reduce_intra_binary rank %d ss %5d",
ompi_comm_rank(comm), segsize));
COLL_TUNED_UPDATE_BINTREE( comm, tuned_module, root );
/**
* Determine number of segments and number of elements
* sent per operation
*/
ompi_datatype_type_size( datatype, &typelng );
COLL_TUNED_COMPUTED_SEGCOUNT( segsize, typelng, segcount );
return ompi_coll_tuned_reduce_generic( sendbuf, recvbuf, count, datatype,
op, root, comm, module,
data->cached_bintree,
segcount, max_outstanding_reqs );
}
int ompi_coll_tuned_reduce_intra_chain( void *sendbuf, void *recvbuf, int count,
ompi_datatype_t* datatype, ompi_op_t* op,
int root, ompi_communicator_t* comm, uint32_t segsize,
int fanout)
{
int segcount = count;
size_t typelng;
OPAL_OUTPUT((ompi_coll_tuned_stream,"coll:tuned:reduce_intra_chain rank %d fo %d ss %5d", ompi_comm_rank(comm), fanout, segsize));
COLL_TUNED_UPDATE_CHAIN( comm, root, fanout );
/**
* Determine number of segments and number of elements
* sent per operation
*/
ompi_ddt_type_size( datatype, &typelng );
COLL_TUNED_COMPUTED_SEGCOUNT( segsize, typelng, segcount );
return ompi_coll_tuned_reduce_generic( sendbuf, recvbuf, count, datatype, op, root, comm,
comm->c_coll_selected_data->cached_chain, segcount );
}
/*
* reduce_intra_in_order_binary
*
* Function: Logarithmic reduce operation for non-commutative operations.
* Acecpts: same as MPI_Reduce()
* Returns: MPI_SUCCESS or error code
*/
int ompi_coll_tuned_reduce_intra_in_order_binary( void *sendbuf, void *recvbuf,
int count,
ompi_datatype_t* datatype,
ompi_op_t* op, int root,
ompi_communicator_t* comm,
mca_coll_base_module_t *module,
uint32_t segsize,
int max_outstanding_reqs )
{
int ret, rank, size, io_root, segcount = count;
void *use_this_sendbuf = NULL, *use_this_recvbuf = NULL;
size_t typelng;
mca_coll_tuned_module_t *tuned_module = (mca_coll_tuned_module_t*) module;
mca_coll_tuned_comm_t *data = tuned_module->tuned_data;
rank = ompi_comm_rank(comm);
size = ompi_comm_size(comm);
OPAL_OUTPUT((ompi_coll_tuned_stream,"coll:tuned:reduce_intra_in_order_binary rank %d ss %5d",
rank, segsize));
COLL_TUNED_UPDATE_IN_ORDER_BINTREE( comm, tuned_module );
/**
* Determine number of segments and number of elements
* sent per operation
*/
ompi_datatype_type_size( datatype, &typelng );
COLL_TUNED_COMPUTED_SEGCOUNT( segsize, typelng, segcount );
/* An in-order binary tree must use root (size-1) to preserve the order of
operations. Thus, if root is not rank (size - 1), then we must handle
1. MPI_IN_PLACE option on real root, and
2. we must allocate temporary recvbuf on rank (size - 1).
Note that generic function must be careful not to switch order of
operations for non-commutative ops.
*/
io_root = size - 1;
use_this_sendbuf = sendbuf;
use_this_recvbuf = recvbuf;
if (io_root != root) {
ptrdiff_t tlb, text, lb, ext;
char *tmpbuf = NULL;
ompi_datatype_get_extent(datatype, &lb, &ext);
ompi_datatype_get_true_extent(datatype, &tlb, &text);
if ((root == rank) && (MPI_IN_PLACE == sendbuf)) {
tmpbuf = (char *) malloc(text + (ptrdiff_t)(count - 1) * ext);
if (NULL == tmpbuf) {
return MPI_ERR_INTERN;
}
ompi_datatype_copy_content_same_ddt(datatype, count,
(char*)tmpbuf,
(char*)recvbuf);
use_this_sendbuf = tmpbuf;
} else if (io_root == rank) {
tmpbuf = (char *) malloc(text + (ptrdiff_t)(count - 1) * ext);
if (NULL == tmpbuf) {
return MPI_ERR_INTERN;
}
use_this_recvbuf = tmpbuf;
}
}
/* Use generic reduce with in-order binary tree topology and io_root */
ret = ompi_coll_tuned_reduce_generic( use_this_sendbuf, use_this_recvbuf, count, datatype,
op, io_root, comm, module,
data->cached_in_order_bintree,
segcount, max_outstanding_reqs );
if (MPI_SUCCESS != ret) { return ret; }
/* Clean up */
if (io_root != root) {
if (root == rank) {
/* Receive result from rank io_root to recvbuf */
ret = MCA_PML_CALL(recv(recvbuf, count, datatype, io_root,
MCA_COLL_BASE_TAG_REDUCE, comm,
MPI_STATUS_IGNORE));
if (MPI_SUCCESS != ret) { return ret; }
if (MPI_IN_PLACE == sendbuf) {
free(use_this_sendbuf);
}
} else if (io_root == rank) {
/* Send result from use_this_recvbuf to root */
ret = MCA_PML_CALL(send(use_this_recvbuf, count, datatype, root,
MCA_COLL_BASE_TAG_REDUCE,
MCA_PML_BASE_SEND_STANDARD, comm));
if (MPI_SUCCESS != ret) { return ret; }
free(use_this_recvbuf);
}
}
return MPI_SUCCESS;
}
int ompi_coll_tuned_reduce_intra_binomial( void *sendbuf, void *recvbuf,
int count, ompi_datatype_t* datatype,
ompi_op_t* op, int root,
ompi_communicator_t* comm,
mca_coll_base_module_t *module,
uint32_t segsize,
int max_outstanding_reqs )
{
int segcount = count;
size_t typelng;
mca_coll_tuned_module_t *tuned_module = (mca_coll_tuned_module_t*) module;
mca_coll_tuned_comm_t *data = tuned_module->tuned_data;
OPAL_OUTPUT((ompi_coll_tuned_stream,"coll:tuned:reduce_intra_binomial rank %d ss %5d",
ompi_comm_rank(comm), segsize));
//printf ("USE BINOMIAL REDUCTION\n");
if (sdn_comp_enable > 1) {
assert(false);
ompi_coll_tree_t *tree = (ompi_coll_tree_t*)malloc(sizeof(ompi_coll_tree_t));
tree->tree_bmtree = 1;
tree->tree_root = MPI_UNDEFINED;
tree->tree_nextsize = MPI_UNDEFINED;
tree->tree_root = root;
switch (ompi_comm_rank(comm)) {
case 0:
tree->tree_nextsize = 0;
tree->tree_prev = 1;
break;
case 1:
tree->tree_nextsize = 1;
tree->tree_next[0] = 0;
tree->tree_prev = 3;
break;
case 2:
tree->tree_nextsize = 0;
tree->tree_prev = 3;
break;
case 3:
tree->tree_nextsize = 2;
tree->tree_next[0] = 1;
tree->tree_next[1] = 2;
tree->tree_prev = 7;
break;
case 4:
tree->tree_nextsize = 1;
tree->tree_next[0] = 5;
tree->tree_prev = 7;
break;
case 5:
tree->tree_nextsize = 0;
tree->tree_prev = 4;
break;
case 6:
tree->tree_nextsize = 0;
tree->tree_prev = 7;
break;
case 7:
tree->tree_nextsize = 3;
tree->tree_next[0] = 3;
tree->tree_next[1] = 4;
tree->tree_next[2] = 6;
tree->tree_prev = 7;
break;
}
/**
* Determine number of segments and number of elements
* sent per operation
*/
ompi_datatype_type_size( datatype, &typelng );
COLL_TUNED_COMPUTED_SEGCOUNT( segsize, typelng, segcount );
return ompi_coll_tuned_reduce_generic( sendbuf, recvbuf, count, datatype,
op, root, comm, module, tree,
segcount, max_outstanding_reqs );
} else if (sdn_comp_enable > 0) {
/**
* Determine number of segments and number of elements
* sent per operation
*/
ompi_datatype_type_size( datatype, &typelng );
COLL_TUNED_COMPUTED_SEGCOUNT( segsize, typelng, segcount );
return ompi_coll_tuned_reduce_generic( sendbuf, recvbuf, count, datatype,
op, root, comm, module,
sdn_shortest_bmtree[root],
segcount, max_outstanding_reqs );
} else {
COLL_TUNED_UPDATE_IN_ORDER_BMTREE( comm, tuned_module, root );
/**
* Determine number of segments and number of elements
* sent per operation
*/
ompi_datatype_type_size( datatype, &typelng );
COLL_TUNED_COMPUTED_SEGCOUNT( segsize, typelng, segcount );
return ompi_coll_tuned_reduce_generic( sendbuf, recvbuf, count, datatype,
op, root, comm, module,
data->cached_in_order_bmtree,
segcount, max_outstanding_reqs );
} /* end if sdn_comp_enable */
}
