static int applicable(const S *ego, const problem *p_,
const planner *plnr)
{
const problem_mpi_transpose *p = (const problem_mpi_transpose *) p_;
/* Note: this is *not* UGLY for out-of-place, destroy-input plans;
the planner often prefers transpose-pairwise to transpose-alltoall,
at least with LAM MPI on my machine. */
return (1
&& (!ego->preserve_input || (!NO_DESTROY_INPUTP(plnr)
&& p->I != p->O))
&& ONLY_TRANSPOSEDP(p->flags));
}
static int applicable(const S *ego, const problem *p_,
const planner *plnr)
{
const problem_mpi_transpose *p = (const problem_mpi_transpose *) p_;
return (1
&& p->I != p->O
&& (!NO_DESTROY_INPUTP(plnr) ||
((p->flags & TRANSPOSED_IN) && !ego->copy_transposed_in))
&& ((p->flags & TRANSPOSED_IN) || !ego->copy_transposed_in)
&& ONLY_TRANSPOSEDP(p->flags)
);
}
static int applicable(const S *ego, const problem *p_,
const planner *plnr, int *r)
{
const problem_mpi_transpose *p = (const problem_mpi_transpose *) p_;
int n_pes;
MPI_Comm_size(p->comm, &n_pes);
return (1
&& p->tblock * n_pes == p->ny
&& (!ego->preserve_input || (!NO_DESTROY_INPUTP(plnr)
&& p->I != p->O))
&& (*r = ego->radix(n_pes)) && *r < n_pes && *r > 1
&& enough_space(p->nx, p->ny, p->block, p->tblock, *r, n_pes)
&& (!CONSERVE_MEMORYP(plnr) || *r > 8
|| !X(toobig)((p->nx * (p->ny / n_pes) * p->vn) / *r))
&& (!NO_SLOWP(plnr) ||
(p->nx * (p->ny / n_pes) * p->vn) / n_pes <= SMALL_MESSAGE)
&& ONLY_TRANSPOSEDP(p->flags)
);
}
