static plan *mkplan(const solver *ego, const problem *p_, planner *plnr)
{
const problem_mpi_rdft2 *p = (const problem_mpi_rdft2 *) p_;
P *pln;
plan *cld;
int my_pe;
R *r0, *r1, *cr, *ci;
static const plan_adt padt = {
XM(rdft2_solve), awake, print, destroy
};
UNUSED(ego);
/* check whether applicable: */
if (!XM(rdft2_serial_applicable)(p))
return (plan *) 0;
if (p->kind == R2HC) {
r1 = (r0 = p->I) + p->vn;
ci = (cr = p->O) + 1;
}
else {
r1 = (r0 = p->O) + p->vn;
ci = (cr = p->I) + 1;
}
MPI_Comm_rank(p->comm, &my_pe);
if (my_pe == 0 && p->vn > 0) {
INT ivs = 1 + (p->kind == HC2R), ovs = 1 + (p->kind == R2HC);
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 / 2 + 1;
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;
}
sz->dims[rnk - 1].n = p->sz->dims[rnk - 1].n;
cld = X(mkplan_d)(plnr,
X(mkproblem_rdft2_d)(sz,
X(mktensor_1d)(p->vn,ivs,ovs),
r0, r1, cr, ci, p->kind));
}
else { /* idle process: make nop plan */
cld = X(mkplan_d)(plnr,
X(mkproblem_rdft2_d)(X(mktensor_0d)(),
X(mktensor_1d)(0,0,0),
cr, ci, cr, ci, HC2R));
}
if (XM(any_true)(!cld, p->comm)) return (plan *) 0;
pln = MKPLAN_MPI_RDFT2(P, &padt, p->kind == R2HC ? apply_r2c : apply_c2r);
pln->cld = cld;
pln->vn = p->vn;
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_rdft2 *p;
P *pln;
plan *cld1 = 0, *cldt = 0, *cld2 = 0;
R *r0, *r1, *cr, *ci, *ri, *ii, *ro, *io, *I, *O;
tensor *sz;
int i, my_pe, n_pes;
INT nrest, n1, b1;
static const plan_adt padt = {
XM(rdft2_solve), awake, print, destroy
};
block_kind k1, k2;
UNUSED(ego);
if (!applicable(ego, p_, plnr))
return (plan *) 0;
p = (const problem_mpi_rdft2 *) p_;
I = p->I; O = p->O;
if (p->kind == R2HC) {
k1 = IB; k2 = OB;
r1 = (r0 = I) + p->vn;
if (ego->preserve_input || NO_DESTROY_INPUTP(plnr)) {
ci = (cr = O) + 1;
I = O;
}
else
ci = (cr = I) + 1;
io = ii = (ro = ri = O) + 1;
}
else {
k1 = OB; k2 = IB;
r1 = (r0 = O) + p->vn;
ci = (cr = O) + 1;
if (ego->preserve_input || NO_DESTROY_INPUTP(plnr)) {
ri = (ii = I) + 1;
ro = (io = O) + 1;
I = O;
}
else
ro = ri = (io = ii = I) + 1;
}
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 / 2 + 1;
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 ivs = 1 + (p->kind == HC2R), ovs = 1 + (p->kind == R2HC);
INT is = sz->dims[0].n * sz->dims[0].is;
INT b = XM(block)(p->sz->dims[0].n, p->sz->dims[0].b[k1], my_pe);
sz->dims[p->sz->rnk - 2].n = p->sz->dims[p->sz->rnk - 1].n;
cld1 = X(mkplan_d)(plnr,
X(mkproblem_rdft2_d)(sz,
X(mktensor_2d)(b, is, is,
p->vn,ivs,ovs),
r0, r1, cr, ci, p->kind));
if (XM(any_true)(!cld1, p->comm)) goto nada;
}
nrest *= p->vn;
n1 = p->sz->dims[1].n;
b1 = p->sz->dims[1].b[k2];
if (p->sz->rnk == 2) { /* n1 dimension is cut in ~half */
n1 = n1 / 2 + 1;
b1 = b1 == p->sz->dims[1].n ? n1 : b1;
}
if (p->kind == R2HC)
cldt = X(mkplan_d)(plnr,
XM(mkproblem_transpose)(
p->sz->dims[0].n, n1, nrest * 2,
I, O,
p->sz->dims[0].b[IB], b1,
p->comm, 0));
else
cldt = X(mkplan_d)(plnr,
XM(mkproblem_transpose)(
n1, p->sz->dims[0].n, nrest * 2,
I, O,
b1, p->sz->dims[0].b[OB],
p->comm, 0));
if (XM(any_true)(!cldt, p->comm)) goto nada;
{
INT is = p->sz->dims[0].n * nrest * 2;
INT b = XM(block)(n1, b1, 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),
ri, ii, ro, io));
if (XM(any_true)(!cld2, p->comm)) goto nada;
}
pln = MKPLAN_MPI_RDFT2(P, &padt, p->kind == R2HC ? apply_r2c : apply_c2r);
pln->cld1 = cld1;
pln->cldt = cldt;
pln->cld2 = cld2;
pln->preserve_input = ego->preserve_input ? 2 : NO_DESTROY_INPUTP(plnr);
pln->vn = p->vn;
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;
}
static plan *mkplan(const solver *ego_, const problem *p_, planner *plnr)
{
const S *ego = (const S *) ego_;
const problem_mpi_rdft2 *p;
P *pln;
plan *cld1 = 0, *cld2 = 0;
R *r0, *r1, *cr, *ci, *I, *O;
tensor *sz;
dtensor *sz2;
int i, my_pe, n_pes;
INT nrest;
static const plan_adt padt = {
XM(rdft2_solve), awake, print, destroy
};
UNUSED(ego);
if (!applicable(ego, p_, plnr))
return (plan *) 0;
p = (const problem_mpi_rdft2 *) p_;
I = p->I; O = p->O;
if (p->kind == R2HC) {
r1 = (r0 = p->I) + p->vn;
if (ego->preserve_input || NO_DESTROY_INPUTP(plnr)) {
ci = (cr = p->O) + 1;
I = O;
}
else
ci = (cr = p->I) + 1;
}
else {
r1 = (r0 = p->O) + p->vn;
ci = (cr = p->O) + 1;
}
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].is = sz->dims[i].os = 2 * p->vn;
sz->dims[i].n = p->sz->dims[i+1].n / 2 + 1;
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 = X(tensor_sz)(sz);
{
INT ivs = 1 + (p->kind == HC2R), ovs = 1 + (p->kind == R2HC);
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);
sz->dims[p->sz->rnk - 2].n = p->sz->dims[p->sz->rnk - 1].n;
cld1 = X(mkplan_d)(plnr,
X(mkproblem_rdft2_d)(sz,
X(mktensor_2d)(b, is, is,
p->vn,ivs,ovs),
r0, r1, cr, ci, p->kind));
if (XM(any_true)(!cld1, p->comm)) goto nada;
}
sz2 = XM(mkdtensor)(1); /* tensor for first (distributed) dimension */
sz2->dims[0] = p->sz->dims[0];
cld2 = X(mkplan_d)(plnr, XM(mkproblem_dft_d)(sz2, nrest * p->vn,
I, O, p->comm,
p->kind == R2HC ?
FFT_SIGN : -FFT_SIGN,
RANK1_BIGVEC_ONLY));
if (XM(any_true)(!cld2, p->comm)) goto nada;
pln = MKPLAN_MPI_RDFT2(P, &padt, p->kind == R2HC ? apply_r2c : apply_c2r);
pln->cld1 = cld1;
pln->cld2 = cld2;
pln->preserve_input = ego->preserve_input ? 2 : NO_DESTROY_INPUTP(plnr);
pln->vn = p->vn;
X(ops_add)(&cld1->ops, &cld2->ops, &pln->super.super.ops);
return &(pln->super.super);
nada:
X(plan_destroy_internal)(cld2);
X(plan_destroy_internal)(cld1);
return (plan *) 0;
}
