static plan *mkplan(const solver *ego, const problem *p_, planner *plnr)
{
const problem_mpi_dft *p = (const problem_mpi_dft *) p_;
P *pln;
plan *cld;
int my_pe;
R *ri, *ii, *ro, *io;
static const plan_adt padt = {
XM(dft_solve), awake, print, destroy
};
UNUSED(ego);
/* check whether applicable: */
if (!XM(dft_serial_applicable)(p))
return (plan *) 0;
X(extract_reim)(p->sign, p->I, &ri, &ii);
X(extract_reim)(p->sign, p->O, &ro, &io);
MPI_Comm_rank(p->comm, &my_pe);
if (my_pe == 0 && p->vn > 0) {
int i, rnk = p->sz->rnk;
tensor *sz = X(mktensor)(p->sz->rnk);
sz->dims[rnk - 1].is = sz->dims[rnk - 1].os = 2 * p->vn;
sz->dims[rnk - 1].n = p->sz->dims[rnk - 1].n;
for (i = rnk - 1; i > 0; --i) {
sz->dims[i - 1].is = sz->dims[i - 1].os =
sz->dims[i].is * sz->dims[i].n;
sz->dims[i - 1].n = p->sz->dims[i - 1].n;
}
cld = X(mkplan_d)(plnr,
X(mkproblem_dft_d)(sz,
X(mktensor_1d)(p->vn, 2, 2),
ri, ii, ro, io));
}
else { /* idle process: make nop plan */
cld = X(mkplan_d)(plnr,
X(mkproblem_dft_d)(X(mktensor_0d)(),
X(mktensor_1d)(0,0,0),
ri, ii, ro, io));
}
if (XM(any_true)(!cld, p->comm)) return (plan *) 0;
pln = MKPLAN_MPI_DFT(P, &padt, apply);
pln->cld = cld;
pln->roff = ro - p->O;
pln->ioff = io - p->O;
X(ops_cpy)(&cld->ops, &pln->super.super.ops);
return &(pln->super.super);
}
static plan *mkplan(const solver *ego_, const problem *p_, planner *plnr)
{
const S *ego = (const S *) ego_;
const problem_mpi_dft *p;
P *pln;
plan *cld = 0, *cldt_before = 0, *cldt_after = 0;
R *ri, *ii, *ro, *io, *I, *O;
INT yblock, yb, nx, ny, vn;
int my_pe, n_pes;
static const plan_adt padt = {
XM(dft_solve), awake, print, destroy
};
UNUSED(ego);
if (!applicable(ego, p_, plnr))
return (plan *) 0;
p = (const problem_mpi_dft *) p_;
MPI_Comm_rank(p->comm, &my_pe);
MPI_Comm_size(p->comm, &n_pes);
nx = p->sz->dims[0].n;
if (!(ny = XM(rearrange_ny)(ego->rearrange, p->sz->dims[0],p->vn,n_pes)))
return (plan *) 0;
vn = p->vn / ny;
A(ny * vn == p->vn);
yblock = XM(default_block)(ny, n_pes);
cldt_before = X(mkplan_d)(plnr,
XM(mkproblem_transpose)(
nx, ny, vn*2,
I = p->I, O = p->O,
p->sz->dims[0].b[IB], yblock,
p->comm, 0));
if (XM(any_true)(!cldt_before, p->comm)) goto nada;
if (ego->preserve_input || NO_DESTROY_INPUTP(plnr)) { I = O; }
X(extract_reim)(p->sign, I, &ri, &ii);
X(extract_reim)(p->sign, O, &ro, &io);
yb = XM(block)(ny, yblock, my_pe);
cld = X(mkplan_d)(plnr,
X(mkproblem_dft_d)(X(mktensor_1d)(nx, vn*2, vn*2),
X(mktensor_2d)(yb, vn*2*nx, vn*2*nx,
vn, 2, 2),
ro, io, ri, ii));
if (XM(any_true)(!cld, p->comm)) goto nada;
cldt_after = X(mkplan_d)(plnr,
XM(mkproblem_transpose)(
ny, nx, vn*2,
I, O,
yblock, p->sz->dims[0].b[OB],
p->comm, 0));
if (XM(any_true)(!cldt_after, p->comm)) goto nada;
pln = MKPLAN_MPI_DFT(P, &padt, apply);
pln->cldt_before = cldt_before;
pln->cld = cld;
pln->cldt_after = cldt_after;
pln->preserve_input = ego->preserve_input ? 2 : NO_DESTROY_INPUTP(plnr);
pln->roff = ro - p->O;
pln->ioff = io - p->O;
pln->rearrange = ego->rearrange;
X(ops_add)(&cldt_before->ops, &cld->ops, &pln->super.super.ops);
X(ops_add2)(&cldt_after->ops, &pln->super.super.ops);
return &(pln->super.super);
nada:
X(plan_destroy_internal)(cldt_after);
X(plan_destroy_internal)(cld);
X(plan_destroy_internal)(cldt_before);
return (plan *) 0;
}
static plan *mkplan(const solver *ego_, const problem *p_, planner *plnr)
{
const S *ego = (const S *) ego_;
const problem_mpi_dft *p;
P *pln;
plan *cld1 = 0, *cldt = 0, *cld2 = 0;
R *ri, *ii, *ro, *io, *I, *O;
tensor *sz;
int i, my_pe, n_pes;
INT nrest;
static const plan_adt padt = {
XM(dft_solve), awake, print, destroy
};
UNUSED(ego);
if (!applicable(ego, p_, plnr))
return (plan *) 0;
p = (const problem_mpi_dft *) p_;
X(extract_reim)(p->sign, I = p->I, &ri, &ii);
X(extract_reim)(p->sign, O = p->O, &ro, &io);
if (ego->preserve_input || NO_DESTROY_INPUTP(plnr))
I = O;
else {
ro = ri;
io = ii;
}
MPI_Comm_rank(p->comm, &my_pe);
MPI_Comm_size(p->comm, &n_pes);
sz = X(mktensor)(p->sz->rnk - 1); /* tensor of last rnk-1 dimensions */
i = p->sz->rnk - 2; A(i >= 0);
sz->dims[i].n = p->sz->dims[i+1].n;
sz->dims[i].is = sz->dims[i].os = 2 * p->vn;
for (--i; i >= 0; --i) {
sz->dims[i].n = p->sz->dims[i+1].n;
sz->dims[i].is = sz->dims[i].os = sz->dims[i+1].n * sz->dims[i+1].is;
}
nrest = 1; for (i = 1; i < sz->rnk; ++i) nrest *= sz->dims[i].n;
{
INT is = sz->dims[0].n * sz->dims[0].is;
INT b = XM(block)(p->sz->dims[0].n, p->sz->dims[0].b[IB], my_pe);
cld1 = X(mkplan_d)(plnr,
X(mkproblem_dft_d)(sz,
X(mktensor_2d)(b, is, is,
p->vn, 2, 2),
ri, ii, ro, io));
if (XM(any_true)(!cld1, p->comm)) goto nada;
}
nrest *= p->vn;
cldt = X(mkplan_d)(plnr,
XM(mkproblem_transpose)(
p->sz->dims[0].n, p->sz->dims[1].n, nrest * 2,
I, O,
p->sz->dims[0].b[IB], p->sz->dims[1].b[OB],
p->comm, 0));
if (XM(any_true)(!cldt, p->comm)) goto nada;
X(extract_reim)(p->sign, O, &ro, &io);
{
INT is = p->sz->dims[0].n * nrest * 2;
INT b = XM(block)(p->sz->dims[1].n, p->sz->dims[1].b[OB], my_pe);
cld2 = X(mkplan_d)(plnr,
X(mkproblem_dft_d)(X(mktensor_1d)(
p->sz->dims[0].n,
nrest * 2, nrest * 2),
X(mktensor_2d)(b, is, is,
nrest, 2, 2),
ro, io, ro, io));
if (XM(any_true)(!cld2, p->comm)) goto nada;
}
pln = MKPLAN_MPI_DFT(P, &padt, apply);
pln->cld1 = cld1;
pln->cldt = cldt;
pln->cld2 = cld2;
pln->preserve_input = ego->preserve_input ? 2 : NO_DESTROY_INPUTP(plnr);
pln->roff = ri - p->I;
pln->ioff = ii - p->I;
X(ops_add)(&cld1->ops, &cld2->ops, &pln->super.super.ops);
X(ops_add2)(&cldt->ops, &pln->super.super.ops);
return &(pln->super.super);
nada:
X(plan_destroy_internal)(cld2);
X(plan_destroy_internal)(cldt);
X(plan_destroy_internal)(cld1);
return (plan *) 0;
}