mirror of
https://github.com/tildearrow/furnace.git
synced 2024-12-18 14:30:15 +00:00
234 lines
7.2 KiB
C
234 lines
7.2 KiB
C
/*
|
|
* Copyright (c) 2003, 2007-14 Matteo Frigo
|
|
* Copyright (c) 2003, 2007-14 Massachusetts Institute of Technology
|
|
*
|
|
* This program is free software; you can redistribute it and/or modify
|
|
* it under the terms of the GNU General Public License as published by
|
|
* the Free Software Foundation; either version 2 of the License, or
|
|
* (at your option) any later version.
|
|
*
|
|
* This program is distributed in the hope that it will be useful,
|
|
* but WITHOUT ANY WARRANTY; without even the implied warranty of
|
|
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
|
|
* GNU General Public License for more details.
|
|
*
|
|
* You should have received a copy of the GNU General Public License
|
|
* along with this program; if not, write to the Free Software
|
|
* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
|
|
*
|
|
*/
|
|
|
|
/* solvers/plans for vectors of DFTs corresponding to the columns
|
|
of a matrix: first transpose the matrix so that the DFTs are
|
|
contiguous, then do DFTs with transposed output. In particular,
|
|
we restrict ourselves to the case of a square transpose (or a
|
|
sequence thereof). */
|
|
|
|
#include "dft/dft.h"
|
|
|
|
typedef solver S;
|
|
|
|
typedef struct {
|
|
plan_dft super;
|
|
INT vl, ivs, ovs;
|
|
plan *cldtrans, *cld, *cldrest;
|
|
} P;
|
|
|
|
/* initial transpose is out-of-place from input to output */
|
|
static void apply_op(const plan *ego_, R *ri, R *ii, R *ro, R *io)
|
|
{
|
|
const P *ego = (const P *) ego_;
|
|
INT vl = ego->vl, ivs = ego->ivs, ovs = ego->ovs, i;
|
|
|
|
for (i = 0; i < vl; ++i) {
|
|
{
|
|
plan_dft *cldtrans = (plan_dft *) ego->cldtrans;
|
|
cldtrans->apply(ego->cldtrans, ri, ii, ro, io);
|
|
}
|
|
{
|
|
plan_dft *cld = (plan_dft *) ego->cld;
|
|
cld->apply(ego->cld, ro, io, ro, io);
|
|
}
|
|
ri += ivs; ii += ivs;
|
|
ro += ovs; io += ovs;
|
|
}
|
|
{
|
|
plan_dft *cldrest = (plan_dft *) ego->cldrest;
|
|
cldrest->apply(ego->cldrest, ri, ii, ro, io);
|
|
}
|
|
}
|
|
|
|
static void destroy(plan *ego_)
|
|
{
|
|
P *ego = (P *) ego_;
|
|
X(plan_destroy_internal)(ego->cldrest);
|
|
X(plan_destroy_internal)(ego->cld);
|
|
X(plan_destroy_internal)(ego->cldtrans);
|
|
}
|
|
|
|
static void awake(plan *ego_, enum wakefulness wakefulness)
|
|
{
|
|
P *ego = (P *) ego_;
|
|
X(plan_awake)(ego->cldtrans, wakefulness);
|
|
X(plan_awake)(ego->cld, wakefulness);
|
|
X(plan_awake)(ego->cldrest, wakefulness);
|
|
}
|
|
|
|
static void print(const plan *ego_, printer *p)
|
|
{
|
|
const P *ego = (const P *) ego_;
|
|
p->print(p, "(indirect-transpose%v%(%p%)%(%p%)%(%p%))",
|
|
ego->vl, ego->cldtrans, ego->cld, ego->cldrest);
|
|
}
|
|
|
|
static int pickdim(const tensor *vs, const tensor *s, int *pdim0, int *pdim1)
|
|
{
|
|
int dim0, dim1;
|
|
*pdim0 = *pdim1 = -1;
|
|
for (dim0 = 0; dim0 < vs->rnk; ++dim0)
|
|
for (dim1 = 0; dim1 < s->rnk; ++dim1)
|
|
if (vs->dims[dim0].n * X(iabs)(vs->dims[dim0].is) <= X(iabs)(s->dims[dim1].is)
|
|
&& vs->dims[dim0].n >= s->dims[dim1].n
|
|
&& (*pdim0 == -1
|
|
|| (X(iabs)(vs->dims[dim0].is) <= X(iabs)(vs->dims[*pdim0].is)
|
|
&& X(iabs)(s->dims[dim1].is) >= X(iabs)(s->dims[*pdim1].is)))) {
|
|
*pdim0 = dim0;
|
|
*pdim1 = dim1;
|
|
}
|
|
return (*pdim0 != -1 && *pdim1 != -1);
|
|
}
|
|
|
|
static int applicable0(const solver *ego_, const problem *p_,
|
|
const planner *plnr,
|
|
int *pdim0, int *pdim1)
|
|
{
|
|
const problem_dft *p = (const problem_dft *) p_;
|
|
UNUSED(ego_); UNUSED(plnr);
|
|
|
|
return (1
|
|
&& FINITE_RNK(p->vecsz->rnk) && FINITE_RNK(p->sz->rnk)
|
|
|
|
/* FIXME: can/should we relax this constraint? */
|
|
&& X(tensor_inplace_strides2)(p->vecsz, p->sz)
|
|
|
|
&& pickdim(p->vecsz, p->sz, pdim0, pdim1)
|
|
|
|
/* output should not *already* include the transpose
|
|
(in which case we duplicate the regular indirect.c) */
|
|
&& (p->sz->dims[*pdim1].os != p->vecsz->dims[*pdim0].is)
|
|
);
|
|
}
|
|
|
|
static int applicable(const solver *ego_, const problem *p_,
|
|
const planner *plnr,
|
|
int *pdim0, int *pdim1)
|
|
{
|
|
if (!applicable0(ego_, p_, plnr, pdim0, pdim1)) return 0;
|
|
{
|
|
const problem_dft *p = (const problem_dft *) p_;
|
|
INT u = p->ri == p->ii + 1 || p->ii == p->ri + 1 ? (INT)2 : (INT)1;
|
|
|
|
/* UGLY if does not result in contiguous transforms or
|
|
transforms of contiguous vectors (since the latter at
|
|
least have efficient transpositions) */
|
|
if (NO_UGLYP(plnr)
|
|
&& p->vecsz->dims[*pdim0].is != u
|
|
&& !(p->vecsz->rnk == 2
|
|
&& p->vecsz->dims[1-*pdim0].is == u
|
|
&& p->vecsz->dims[*pdim0].is
|
|
== u * p->vecsz->dims[1-*pdim0].n))
|
|
return 0;
|
|
|
|
if (NO_INDIRECT_OP_P(plnr) && p->ri != p->ro) return 0;
|
|
}
|
|
return 1;
|
|
}
|
|
|
|
static plan *mkplan(const solver *ego_, const problem *p_, planner *plnr)
|
|
{
|
|
const problem_dft *p = (const problem_dft *) p_;
|
|
P *pln;
|
|
plan *cld = 0, *cldtrans = 0, *cldrest = 0;
|
|
int pdim0, pdim1;
|
|
tensor *ts, *tv;
|
|
INT vl, ivs, ovs;
|
|
R *rit, *iit, *rot, *iot;
|
|
|
|
static const plan_adt padt = {
|
|
X(dft_solve), awake, print, destroy
|
|
};
|
|
|
|
if (!applicable(ego_, p_, plnr, &pdim0, &pdim1))
|
|
return (plan *) 0;
|
|
|
|
vl = p->vecsz->dims[pdim0].n / p->sz->dims[pdim1].n;
|
|
A(vl >= 1);
|
|
ivs = p->sz->dims[pdim1].n * p->vecsz->dims[pdim0].is;
|
|
ovs = p->sz->dims[pdim1].n * p->vecsz->dims[pdim0].os;
|
|
rit = TAINT(p->ri, vl == 1 ? 0 : ivs);
|
|
iit = TAINT(p->ii, vl == 1 ? 0 : ivs);
|
|
rot = TAINT(p->ro, vl == 1 ? 0 : ovs);
|
|
iot = TAINT(p->io, vl == 1 ? 0 : ovs);
|
|
|
|
ts = X(tensor_copy_inplace)(p->sz, INPLACE_IS);
|
|
ts->dims[pdim1].os = p->vecsz->dims[pdim0].is;
|
|
tv = X(tensor_copy_inplace)(p->vecsz, INPLACE_IS);
|
|
tv->dims[pdim0].os = p->sz->dims[pdim1].is;
|
|
tv->dims[pdim0].n = p->sz->dims[pdim1].n;
|
|
cldtrans = X(mkplan_d)(plnr,
|
|
X(mkproblem_dft_d)(X(mktensor_0d)(),
|
|
X(tensor_append)(tv, ts),
|
|
rit, iit,
|
|
rot, iot));
|
|
X(tensor_destroy2)(ts, tv);
|
|
if (!cldtrans) goto nada;
|
|
|
|
ts = X(tensor_copy)(p->sz);
|
|
ts->dims[pdim1].is = p->vecsz->dims[pdim0].is;
|
|
tv = X(tensor_copy)(p->vecsz);
|
|
tv->dims[pdim0].is = p->sz->dims[pdim1].is;
|
|
tv->dims[pdim0].n = p->sz->dims[pdim1].n;
|
|
cld = X(mkplan_d)(plnr, X(mkproblem_dft_d)(ts, tv,
|
|
rot, iot,
|
|
rot, iot));
|
|
if (!cld) goto nada;
|
|
|
|
tv = X(tensor_copy)(p->vecsz);
|
|
tv->dims[pdim0].n -= vl * p->sz->dims[pdim1].n;
|
|
cldrest = X(mkplan_d)(plnr, X(mkproblem_dft_d)(X(tensor_copy)(p->sz), tv,
|
|
p->ri + ivs * vl,
|
|
p->ii + ivs * vl,
|
|
p->ro + ovs * vl,
|
|
p->io + ovs * vl));
|
|
if (!cldrest) goto nada;
|
|
|
|
pln = MKPLAN_DFT(P, &padt, apply_op);
|
|
pln->cldtrans = cldtrans;
|
|
pln->cld = cld;
|
|
pln->cldrest = cldrest;
|
|
pln->vl = vl;
|
|
pln->ivs = ivs;
|
|
pln->ovs = ovs;
|
|
X(ops_cpy)(&cldrest->ops, &pln->super.super.ops);
|
|
X(ops_madd2)(vl, &cld->ops, &pln->super.super.ops);
|
|
X(ops_madd2)(vl, &cldtrans->ops, &pln->super.super.ops);
|
|
return &(pln->super.super);
|
|
|
|
nada:
|
|
X(plan_destroy_internal)(cldrest);
|
|
X(plan_destroy_internal)(cld);
|
|
X(plan_destroy_internal)(cldtrans);
|
|
return (plan *)0;
|
|
}
|
|
|
|
static solver *mksolver(void)
|
|
{
|
|
static const solver_adt sadt = { PROBLEM_DFT, mkplan, 0 };
|
|
S *slv = MKSOLVER(S, &sadt);
|
|
return slv;
|
|
}
|
|
|
|
void X(dft_indirect_transpose_register)(planner *p)
|
|
{
|
|
REGISTER_SOLVER(p, mksolver());
|
|
}
|