int ompi_coll_tuned_alltoall_intra_do_forced(void *sbuf, int scount,
struct ompi_datatype_t *sdtype,
void* rbuf, int rcount,
struct ompi_datatype_t *rdtype,
struct ompi_communicator_t *comm,
mca_coll_base_module_t *module)
{
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:alltoall_intra_do_forced selected algorithm %d",
data->user_forced[ALLTOALL].algorithm));
switch (data->user_forced[ALLTOALL].algorithm) {
case (0): return ompi_coll_tuned_alltoall_intra_dec_fixed (sbuf, scount, sdtype, rbuf, rcount, rdtype, comm, module);
case (1): return ompi_coll_tuned_alltoall_intra_basic_linear (sbuf, scount, sdtype, rbuf, rcount, rdtype, comm, module);
case (2): return ompi_coll_tuned_alltoall_intra_pairwise (sbuf, scount, sdtype, rbuf, rcount, rdtype, comm, module);
case (3): return ompi_coll_tuned_alltoall_intra_bruck (sbuf, scount, sdtype, rbuf, rcount, rdtype, comm, module);
case (4): return ompi_coll_tuned_alltoall_intra_linear_sync (sbuf, scount, sdtype, rbuf, rcount, rdtype, comm, module, data->user_forced[ALLTOALL].max_requests);
case (5): return ompi_coll_tuned_alltoall_intra_two_procs (sbuf, scount, sdtype, rbuf, rcount, rdtype, comm, module);
default:
OPAL_OUTPUT((ompi_coll_tuned_stream,"coll:tuned:alltoall_intra_do_forced attempt to select algorithm %d when only 0-%d is valid?",
data->user_forced[ALLTOALL].algorithm, ompi_coll_tuned_forced_max_algorithms[ALLTOALL]));
return (MPI_ERR_ARG);
} /* switch */
}
int ompi_coll_tuned_alltoall_intra_do_this(void *sbuf, int scount,
struct ompi_datatype_t *sdtype,
void* rbuf, int rcount,
struct ompi_datatype_t *rdtype,
struct ompi_communicator_t *comm,
mca_coll_base_module_t *module,
int algorithm, int faninout, int segsize,
int max_requests)
{
OPAL_OUTPUT((ompi_coll_tuned_stream,"coll:tuned:alltoall_intra_do_this selected algorithm %d topo faninout %d segsize %d",
algorithm, faninout, segsize));
switch (algorithm) {
case (0): return ompi_coll_tuned_alltoall_intra_dec_fixed (sbuf, scount, sdtype, rbuf, rcount, rdtype, comm, module);
case (1): return ompi_coll_tuned_alltoall_intra_basic_linear (sbuf, scount, sdtype, rbuf, rcount, rdtype, comm, module);
case (2): return ompi_coll_tuned_alltoall_intra_pairwise (sbuf, scount, sdtype, rbuf, rcount, rdtype, comm, module);
case (3): return ompi_coll_tuned_alltoall_intra_bruck (sbuf, scount, sdtype, rbuf, rcount, rdtype, comm, module);
case (4): return ompi_coll_tuned_alltoall_intra_linear_sync (sbuf, scount, sdtype, rbuf, rcount, rdtype, comm, module, max_requests);
case (5): return ompi_coll_tuned_alltoall_intra_two_procs (sbuf, scount, sdtype, rbuf, rcount, rdtype, comm, module);
default:
OPAL_OUTPUT((ompi_coll_tuned_stream,"coll:tuned:alltoall_intra_do_this attempt to select algorithm %d when only 0-%d is valid?",
algorithm, ompi_coll_tuned_forced_max_algorithms[ALLTOALL]));
return (MPI_ERR_ARG);
} /* switch */
}
int ompi_coll_tuned_alltoall_intra_dec_dynamic(const void *sbuf, int scount,
struct ompi_datatype_t *sdtype,
void* rbuf, int rcount,
struct ompi_datatype_t *rdtype,
struct ompi_communicator_t *comm,
mca_coll_base_module_t *module)
{
mca_coll_tuned_module_t *tuned_module = (mca_coll_tuned_module_t*) module;
OPAL_OUTPUT((ompi_coll_tuned_stream, "ompi_coll_tuned_alltoall_intra_dec_dynamic"));
/* check to see if we have some filebased rules */
if (tuned_module->com_rules[ALLTOALL]) {
/* we do, so calc the message size or what ever we need and use this for the evaluation */
int comsize;
int alg, faninout, segsize, max_requests;
size_t dsize;
ompi_datatype_type_size (sdtype, &dsize);
comsize = ompi_comm_size(comm);
dsize *= (ptrdiff_t)comsize * (ptrdiff_t)scount;
alg = ompi_coll_tuned_get_target_method_params (tuned_module->com_rules[ALLTOALL],
dsize, &faninout, &segsize, &max_requests);
if (alg) {
/* we have found a valid choice from the file based rules for this message size */
return ompi_coll_tuned_alltoall_intra_do_this (sbuf, scount, sdtype,
rbuf, rcount, rdtype,
comm, module,
alg, faninout, segsize, max_requests);
} /* found a method */
} /*end if any com rules to check */
if (tuned_module->user_forced[ALLTOALL].algorithm) {
return ompi_coll_tuned_alltoall_intra_do_this(sbuf, scount, sdtype,
rbuf, rcount, rdtype,
comm, module,
tuned_module->user_forced[ALLTOALL].algorithm,
tuned_module->user_forced[ALLTOALL].tree_fanout,
tuned_module->user_forced[ALLTOALL].segsize,
tuned_module->user_forced[ALLTOALL].max_requests);
}
return ompi_coll_tuned_alltoall_intra_dec_fixed (sbuf, scount, sdtype,
rbuf, rcount, rdtype,
comm, module);
}
int ompi_coll_tuned_alltoall_intra_do_forced(void *sbuf, int scount,
struct ompi_datatype_t *sdtype,
void* rbuf, int rcount,
struct ompi_datatype_t *rdtype,
struct ompi_communicator_t *comm)
{
OPAL_OUTPUT((ompi_coll_tuned_stream,"coll:tuned:alltoall_intra_do_forced selected algorithm %d",
comm->c_coll_selected_data->user_forced[ALLTOALL].algorithm));
switch (comm->c_coll_selected_data->user_forced[ALLTOALL].algorithm) {
case (0): return ompi_coll_tuned_alltoall_intra_dec_fixed (sbuf, scount, sdtype, rbuf, rcount, rdtype, comm);
case (1): return ompi_coll_tuned_alltoall_intra_basic_linear (sbuf, scount, sdtype, rbuf, rcount, rdtype, comm);
case (2): return ompi_coll_tuned_alltoall_intra_pairwise (sbuf, scount, sdtype, rbuf, rcount, rdtype, comm);
case (3): return ompi_coll_tuned_alltoall_intra_bruck (sbuf, scount, sdtype, rbuf, rcount, rdtype, comm);
case (4): return ompi_coll_tuned_alltoall_intra_two_procs (sbuf, scount, sdtype, rbuf, rcount, rdtype, comm);
default:
OPAL_OUTPUT((ompi_coll_tuned_stream,"coll:tuned:alltoall_intra_do_forced attempt to select algorithm %d when only 0-%d is valid?",
comm->c_coll_selected_data->user_forced[ALLTOALL].algorithm, ompi_coll_tuned_forced_max_algorithms[ALLTOALL]));
return (MPI_ERR_ARG);
} /* switch */
}
