int smpi_coll_tuned_reduce_ompi_binary( void *sendbuf, void *recvbuf,
int count, MPI_Datatype datatype,
MPI_Op op, int root,
MPI_Comm comm)
{
uint32_t segsize;
int segcount = count;
size_t typelng;
/**
* Determine number of segments and number of elements
* sent per operation
*/
typelng=smpi_datatype_size( datatype );
// Binary_32K
segsize = 32*1024;
XBT_DEBUG("coll:tuned:reduce_intra_binary rank %d ss %5d",
smpi_comm_rank(comm), segsize);
COLL_TUNED_COMPUTED_SEGCOUNT( segsize, typelng, segcount );
return smpi_coll_tuned_ompi_reduce_generic( sendbuf, recvbuf, count, datatype,
op, root, comm,
ompi_coll_tuned_topo_build_tree(2, comm, root),
segcount, 0);
}
/*
* Init module on the communicator
*/
const struct mca_coll_base_module_1_0_0_t *
ompi_coll_tuned_module_init(struct ompi_communicator_t *comm)
{
int size, rank;
struct mca_coll_base_comm_t *data;
/* fanout parameters */
int rc=0;
int i;
OPAL_OUTPUT((ompi_coll_tuned_stream,"coll:tuned:module_init called."));
/* This routine will become more complex and might have to be */
/* broken into more sections/function calls */
/* Order of operations:
* alloc memory for nb reqs (in case we fall through)
* add decision rules if using dynamic rules
* compact rules using communicator size info etc
* build first guess cached topologies (might depend on the rules from above)
*
* then attach all to the communicator and return base module funct ptrs
*/
/* Allocate the data that hangs off the communicator */
if (OMPI_COMM_IS_INTER(comm)) {
size = ompi_comm_remote_size(comm);
} else {
size = ompi_comm_size(comm);
}
/*
* we still malloc data as it is used by the TUNED modules
* if we don't allocate it and fall back to a BASIC module routine then confuses debuggers
* we place any special info after the default data
*
* BUT on very large systems we might not be able to allocate all this memory so
* we do check a MCA parameter to see if if we should allocate this memory
*
* The default is set very high
*
*/
/* if we within the memory/size limit, allow preallocated data */
if (size<=ompi_coll_tuned_preallocate_memory_comm_size_limit) {
data = (mca_coll_base_comm_t*)malloc(sizeof(struct mca_coll_base_comm_t) +
(sizeof(ompi_request_t *) * size * 2));
if (NULL == data) {
return NULL;
}
data->mcct_reqs = (ompi_request_t **) (data + 1);
data->mcct_num_reqs = size * 2;
}
else {
data = (mca_coll_base_comm_t*)malloc(sizeof(struct mca_coll_base_comm_t));
if (NULL == data) {
return NULL;
}
data->mcct_reqs = (ompi_request_t **) NULL;
data->mcct_num_reqs = 0;
}
/*
* If using dynamic and you are MPI_COMM_WORLD and you want to use a parameter file..
* then this effects how much storage space you need
* (This is a basic version of what will go into V2)
*
*/
size = ompi_comm_size(comm); /* find size so we can (A) decide if to access the file directly */
/* (B) so we can get our very own customised ompi_coll_com_rule_t ptr */
/* which only has rules in it for our com size */
rank = ompi_comm_rank(comm); /* find rank as only MCW:0 opens any tuned conf files */
/* actually if they are below a threadhold, they all open it */
/* have to build a collective in here.. but just for MCW.. */
/* but we have to make sure we have the same rules everywhere :( */
/* if using dynamic rules make sure all overrides are NULL before we start override anything accidently */
if (ompi_coll_tuned_use_dynamic_rules) {
/* base rules */
data->all_base_rules = (ompi_coll_alg_rule_t*) NULL;
/* each collective rule for my com size */
for (i=0;i<COLLCOUNT;i++) {
data->com_rules[i] = (ompi_coll_com_rule_t*) NULL;
}
}
/* next dynamic state, recheck all forced rules as well */
/* warning, we should check to make sure this is really an INTRA comm here... */
if (ompi_coll_tuned_use_dynamic_rules) {
ompi_coll_tuned_forced_getvalues (ompi_coll_tuned_forced_params[ALLREDUCE], &(data->user_forced[ALLREDUCE]));
ompi_coll_tuned_forced_getvalues (ompi_coll_tuned_forced_params[ALLTOALL], &(data->user_forced[ALLTOALL]));
ompi_coll_tuned_forced_getvalues (ompi_coll_tuned_forced_params[ALLGATHER], &(data->user_forced[ALLGATHER]));
/* ompi_coll_tuned_forced_getvalues (ompi_coll_tuned_forced_params[ALLTOALLV], &(data->user_forced[ALLTOALLV])); */
ompi_coll_tuned_forced_getvalues_barrier (ompi_coll_tuned_forced_params[BARRIER], &(data->user_forced[BARRIER]));
ompi_coll_tuned_forced_getvalues (ompi_coll_tuned_forced_params[BCAST], &(data->user_forced[BCAST]));
ompi_coll_tuned_forced_getvalues (ompi_coll_tuned_forced_params[REDUCE], &(data->user_forced[REDUCE]));
}
if (&ompi_mpi_comm_world==comm) {
if (ompi_coll_tuned_use_dynamic_rules) {
OPAL_OUTPUT((ompi_coll_tuned_stream,"coll:tuned:module_init MCW & Dynamic"));
if (ompi_coll_tuned_dynamic_rules_filename) {
OPAL_OUTPUT((ompi_coll_tuned_stream,"coll:tuned:module_init Opening [%s]",
ompi_coll_tuned_dynamic_rules_filename));
rc = ompi_coll_tuned_read_rules_config_file (ompi_coll_tuned_dynamic_rules_filename,
&(data->all_base_rules), COLLCOUNT);
if (rc>=0) {
OPAL_OUTPUT((ompi_coll_tuned_stream,"coll:tuned:module_init Read %d valid rules\n", rc));
/* at this point we all have a base set of rules */
/* now we can get our customized communicator sized rule set, for each collective */
for (i=0;i<COLLCOUNT;i++) {
data->com_rules[i] = ompi_coll_tuned_get_com_rule_ptr (data->all_base_rules, i, size);
}
}
else { /* failed to read config file, thus make sure its a NULL... */
data->all_base_rules = (ompi_coll_alg_rule_t*) NULL;
}
} /* end if a config filename exists */
} /* end if dynamic_rules */
} /* end if MCW */
/* ok, if using dynamic rules, not MCW and we are just any rank and a base set of rules exist.. ref them */
/* order of eval is important here, if we are MCW ompi_mpi_comm_world.c_coll_selected_data is NULL still.. */
if ((ompi_coll_tuned_use_dynamic_rules)&&(!(&ompi_mpi_comm_world==comm))&&
((ompi_mpi_comm_world.c_coll_selected_data)->all_base_rules)) {
OPAL_OUTPUT((ompi_coll_tuned_stream,"coll:tuned:module_init NOT MCW & Dynamic"));
/* this will, erm fail if MCW doesn't exist which it should! */
data->all_base_rules = (ompi_mpi_comm_world.c_coll_selected_data)->all_base_rules;
/* at this point we all have a base set of rules if they exist atall */
/* now we can get our customized communicator sized rule set, for each collective */
for (i=0;i<COLLCOUNT;i++) {
data->com_rules[i] = ompi_coll_tuned_get_com_rule_ptr (data->all_base_rules, i, size);
}
}
/*
* now for the cached topo functions
* guess the initial topologies to use rank 0 as root
*/
/* general n fan out tree */
data->cached_ntree = ompi_coll_tuned_topo_build_tree (ompi_coll_tuned_init_tree_fanout, comm, 0);
data->cached_ntree_root = 0;
data->cached_ntree_fanout = ompi_coll_tuned_init_tree_fanout;
/* binary tree */
data->cached_bintree = ompi_coll_tuned_topo_build_tree (2, comm, 0);
data->cached_bintree_root = 0;
/* binomial tree */
data->cached_bmtree = ompi_coll_tuned_topo_build_bmtree (comm, 0);
data->cached_bmtree_root = 0;
/*
* chains (fanout followed by pipelines)
* are more difficuilt as the fan out really really depends on message size [sometimes]..
* as size gets larger fan-out gets smaller [usually]
*
* will probably change how we cache this later, for now a midsize
* GEF
*/
data->cached_chain = ompi_coll_tuned_topo_build_chain (ompi_coll_tuned_init_chain_fanout, comm, 0);
data->cached_chain_root = 0;
data->cached_chain_fanout = ompi_coll_tuned_init_chain_fanout;
/* standard pipeline */
data->cached_pipeline = ompi_coll_tuned_topo_build_chain (1, comm, 0);
data->cached_pipeline_root = 0;
/* All done */
comm->c_coll_selected_data = data;
OPAL_OUTPUT((ompi_coll_tuned_stream,"coll:tuned:module_init Tuned is in use"));
return to_use;
}
