Chris@10: /*
Chris@10:  * Copyright (c) 2003, 2007-11 Matteo Frigo
Chris@10:  * Copyright (c) 2003, 2007-11 Massachusetts Institute of Technology
Chris@10:  *
Chris@10:  * This program is free software; you can redistribute it and/or modify
Chris@10:  * it under the terms of the GNU General Public License as published by
Chris@10:  * the Free Software Foundation; either version 2 of the License, or
Chris@10:  * (at your option) any later version.
Chris@10:  *
Chris@10:  * This program is distributed in the hope that it will be useful,
Chris@10:  * but WITHOUT ANY WARRANTY; without even the implied warranty of
Chris@10:  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
Chris@10:  * GNU General Public License for more details.
Chris@10:  *
Chris@10:  * You should have received a copy of the GNU General Public License
Chris@10:  * along with this program; if not, write to the Free Software
Chris@10:  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA
Chris@10:  *
Chris@10:  */
Chris@10: 
Chris@10: 
Chris@10: /* plans for DFT of rank >= 2 (multidimensional) */
Chris@10: 
Chris@10: #include "dft.h"
Chris@10: 
Chris@10: typedef struct {
Chris@10:      solver super;
Chris@10:      int spltrnk;
Chris@10:      const int *buddies;
Chris@10:      int nbuddies;
Chris@10: } S;
Chris@10: 
Chris@10: typedef struct {
Chris@10:      plan_dft super;
Chris@10: 
Chris@10:      plan *cld1, *cld2;
Chris@10:      const S *solver;
Chris@10: } P;
Chris@10: 
Chris@10: /* Compute multi-dimensional DFT by applying the two cld plans
Chris@10:    (lower-rnk DFTs). */
Chris@10: static void apply(const plan *ego_, R *ri, R *ii, R *ro, R *io)
Chris@10: {
Chris@10:      const P *ego = (const P *) ego_;
Chris@10:      plan_dft *cld1, *cld2;
Chris@10: 
Chris@10:      cld1 = (plan_dft *) ego->cld1;
Chris@10:      cld1->apply(ego->cld1, ri, ii, ro, io);
Chris@10: 
Chris@10:      cld2 = (plan_dft *) ego->cld2;
Chris@10:      cld2->apply(ego->cld2, ro, io, ro, io);
Chris@10: }
Chris@10: 
Chris@10: 
Chris@10: static void awake(plan *ego_, enum wakefulness wakefulness)
Chris@10: {
Chris@10:      P *ego = (P *) ego_;
Chris@10:      X(plan_awake)(ego->cld1, wakefulness);
Chris@10:      X(plan_awake)(ego->cld2, wakefulness);
Chris@10: }
Chris@10: 
Chris@10: static void destroy(plan *ego_)
Chris@10: {
Chris@10:      P *ego = (P *) ego_;
Chris@10:      X(plan_destroy_internal)(ego->cld2);
Chris@10:      X(plan_destroy_internal)(ego->cld1);
Chris@10: }
Chris@10: 
Chris@10: static void print(const plan *ego_, printer *p)
Chris@10: {
Chris@10:      const P *ego = (const P *) ego_;
Chris@10:      const S *s = ego->solver;
Chris@10:      p->print(p, "(dft-rank>=2/%d%(%p%)%(%p%))",
Chris@10: 	      s->spltrnk, ego->cld1, ego->cld2);
Chris@10: }
Chris@10: 
Chris@10: static int picksplit(const S *ego, const tensor *sz, int *rp)
Chris@10: {
Chris@10:      A(sz->rnk > 1); /* cannot split rnk <= 1 */
Chris@10:      if (!X(pickdim)(ego->spltrnk, ego->buddies, ego->nbuddies, sz, 1, rp))
Chris@10: 	  return 0;
Chris@10:      *rp += 1; /* convert from dim. index to rank */
Chris@10:      if (*rp >= sz->rnk) /* split must reduce rank */
Chris@10: 	  return 0;
Chris@10:      return 1;
Chris@10: }
Chris@10: 
Chris@10: static int applicable0(const solver *ego_, const problem *p_, int *rp)
Chris@10: {
Chris@10:      const problem_dft *p = (const problem_dft *) p_;
Chris@10:      const S *ego = (const S *)ego_;
Chris@10:      return (1
Chris@10: 	     && FINITE_RNK(p->sz->rnk) && FINITE_RNK(p->vecsz->rnk)
Chris@10: 	     && p->sz->rnk >= 2
Chris@10: 	     && picksplit(ego, p->sz, rp)
Chris@10: 	  );
Chris@10: }
Chris@10: 
Chris@10: /* TODO: revise this. */
Chris@10: static int applicable(const solver *ego_, const problem *p_, 
Chris@10: 		      const planner *plnr, int *rp)
Chris@10: {
Chris@10:      const S *ego = (const S *)ego_;
Chris@10:      const problem_dft *p = (const problem_dft *) p_;
Chris@10: 
Chris@10:      if (!applicable0(ego_, p_, rp)) return 0;
Chris@10: 
Chris@10:      if (NO_RANK_SPLITSP(plnr) && (ego->spltrnk != ego->buddies[0])) return 0;
Chris@10: 
Chris@10:      /* Heuristic: if the vector stride is greater than the transform
Chris@10:         sz, don't use (prefer to do the vector loop first with a
Chris@10:         vrank-geq1 plan). */
Chris@10:      if (NO_UGLYP(plnr))
Chris@10: 	  if (p->vecsz->rnk > 0 &&
Chris@10: 	      X(tensor_min_stride)(p->vecsz) > X(tensor_max_index)(p->sz))
Chris@10: 	       return 0;
Chris@10: 
Chris@10:      return 1;
Chris@10: }
Chris@10: 
Chris@10: static plan *mkplan(const solver *ego_, const problem *p_, planner *plnr)
Chris@10: {
Chris@10:      const S *ego = (const S *) ego_;
Chris@10:      const problem_dft *p;
Chris@10:      P *pln;
Chris@10:      plan *cld1 = 0, *cld2 = 0;
Chris@10:      tensor *sz1, *sz2, *vecszi, *sz2i;
Chris@10:      int spltrnk;
Chris@10: 
Chris@10:      static const plan_adt padt = {
Chris@10: 	  X(dft_solve), awake, print, destroy
Chris@10:      };
Chris@10: 
Chris@10:      if (!applicable(ego_, p_, plnr, &spltrnk))
Chris@10:           return (plan *) 0;
Chris@10: 
Chris@10:      p = (const problem_dft *) p_;
Chris@10:      X(tensor_split)(p->sz, &sz1, spltrnk, &sz2);
Chris@10:      vecszi = X(tensor_copy_inplace)(p->vecsz, INPLACE_OS);
Chris@10:      sz2i = X(tensor_copy_inplace)(sz2, INPLACE_OS);
Chris@10: 
Chris@10:      cld1 = X(mkplan_d)(plnr, 
Chris@10: 			X(mkproblem_dft_d)(X(tensor_copy)(sz2),
Chris@10: 					   X(tensor_append)(p->vecsz, sz1),
Chris@10: 					   p->ri, p->ii, p->ro, p->io));
Chris@10:      if (!cld1) goto nada;
Chris@10: 
Chris@10:      cld2 = X(mkplan_d)(plnr, 
Chris@10: 			X(mkproblem_dft_d)(
Chris@10: 			     X(tensor_copy_inplace)(sz1, INPLACE_OS),
Chris@10: 			     X(tensor_append)(vecszi, sz2i),
Chris@10: 			     p->ro, p->io, p->ro, p->io));
Chris@10:      if (!cld2) goto nada;
Chris@10: 
Chris@10:      pln = MKPLAN_DFT(P, &padt, apply);
Chris@10: 
Chris@10:      pln->cld1 = cld1;
Chris@10:      pln->cld2 = cld2;
Chris@10: 
Chris@10:      pln->solver = ego;
Chris@10:      X(ops_add)(&cld1->ops, &cld2->ops, &pln->super.super.ops);
Chris@10: 
Chris@10:      X(tensor_destroy4)(sz1, sz2, vecszi, sz2i);
Chris@10: 
Chris@10:      return &(pln->super.super);
Chris@10: 
Chris@10:  nada:
Chris@10:      X(plan_destroy_internal)(cld2);
Chris@10:      X(plan_destroy_internal)(cld1);
Chris@10:      X(tensor_destroy4)(sz1, sz2, vecszi, sz2i);
Chris@10:      return (plan *) 0;
Chris@10: }
Chris@10: 
Chris@10: static solver *mksolver(int spltrnk, const int *buddies, int nbuddies)
Chris@10: {
Chris@10:      static const solver_adt sadt = { PROBLEM_DFT, mkplan, 0 };
Chris@10:      S *slv = MKSOLVER(S, &sadt);
Chris@10:      slv->spltrnk = spltrnk;
Chris@10:      slv->buddies = buddies;
Chris@10:      slv->nbuddies = nbuddies;
Chris@10:      return &(slv->super);
Chris@10: }
Chris@10: 
Chris@10: void X(dft_rank_geq2_register)(planner *p)
Chris@10: {
Chris@10:      int i;
Chris@10:      static const int buddies[] = { 1, 0, -2 };
Chris@10: 
Chris@10:      const int nbuddies = (int)(sizeof(buddies) / sizeof(buddies[0]));
Chris@10: 
Chris@10:      for (i = 0; i < nbuddies; ++i)
Chris@10:           REGISTER_SOLVER(p, mksolver(buddies[i], buddies, nbuddies));
Chris@10: 
Chris@10:      /* FIXME:
Chris@10: 
Chris@10:         Should we try more buddies? 
Chris@10: 
Chris@10:         Another possible variant is to swap cld1 and cld2 (or rather,
Chris@10:         to swap their problems; they are not interchangeable because
Chris@10:         cld2 must be in-place).  In past versions of FFTW, however, I
Chris@10:         seem to recall that such rearrangements have made little or no
Chris@10:         difference.
Chris@10:      */
Chris@10: }