cannam@127: /*
cannam@127:  * Copyright (c) 2003, 2007-14 Matteo Frigo
cannam@127:  * Copyright (c) 2003, 2007-14 Massachusetts Institute of Technology
cannam@127:  *
cannam@127:  * This program is free software; you can redistribute it and/or modify
cannam@127:  * it under the terms of the GNU General Public License as published by
cannam@127:  * the Free Software Foundation; either version 2 of the License, or
cannam@127:  * (at your option) any later version.
cannam@127:  *
cannam@127:  * This program is distributed in the hope that it will be useful,
cannam@127:  * but WITHOUT ANY WARRANTY; without even the implied warranty of
cannam@127:  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
cannam@127:  * GNU General Public License for more details.
cannam@127:  *
cannam@127:  * You should have received a copy of the GNU General Public License
cannam@127:  * along with this program; if not, write to the Free Software
cannam@127:  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA
cannam@127:  *
cannam@127:  */
cannam@127: 
cannam@127: 
cannam@127: /* plans for RDFT2 of rank >= 2 (multidimensional) */
cannam@127: 
cannam@127: #include "rdft.h"
cannam@127: #include "dft.h"
cannam@127: 
cannam@127: typedef struct {
cannam@127:      solver super;
cannam@127:      int spltrnk;
cannam@127:      const int *buddies;
cannam@127:      size_t nbuddies;
cannam@127: } S;
cannam@127: 
cannam@127: typedef struct {
cannam@127:      plan_dft super;
cannam@127:      plan *cldr, *cldc;
cannam@127:      const S *solver;
cannam@127: } P;
cannam@127: 
cannam@127: static void apply_r2hc(const plan *ego_, R *r0, R *r1, R *cr, R *ci)
cannam@127: {
cannam@127:      const P *ego = (const P *) ego_;
cannam@127: 
cannam@127:      {
cannam@127: 	  plan_rdft2 *cldr = (plan_rdft2 *) ego->cldr;
cannam@127: 	  cldr->apply((plan *) cldr, r0, r1, cr, ci);
cannam@127:      }
cannam@127:      
cannam@127:      {
cannam@127: 	  plan_dft *cldc = (plan_dft *) ego->cldc;
cannam@127: 	  cldc->apply((plan *) cldc, cr, ci, cr, ci);
cannam@127:      }
cannam@127: }
cannam@127: 
cannam@127: static void apply_hc2r(const plan *ego_, R *r0, R *r1, R *cr, R *ci)
cannam@127: {
cannam@127:      const P *ego = (const P *) ego_;
cannam@127: 
cannam@127:      {
cannam@127: 	  plan_dft *cldc = (plan_dft *) ego->cldc;
cannam@127: 	  cldc->apply((plan *) cldc, ci, cr, ci, cr);
cannam@127:      }
cannam@127: 
cannam@127:      {
cannam@127: 	  plan_rdft2 *cldr = (plan_rdft2 *) ego->cldr;
cannam@127: 	  cldr->apply((plan *) cldr, r0, r1, cr, ci);
cannam@127:      }
cannam@127:      
cannam@127: }
cannam@127: 
cannam@127: static void awake(plan *ego_, enum wakefulness wakefulness)
cannam@127: {
cannam@127:      P *ego = (P *) ego_;
cannam@127:      X(plan_awake)(ego->cldr, wakefulness);
cannam@127:      X(plan_awake)(ego->cldc, wakefulness);
cannam@127: }
cannam@127: 
cannam@127: static void destroy(plan *ego_)
cannam@127: {
cannam@127:      P *ego = (P *) ego_;
cannam@127:      X(plan_destroy_internal)(ego->cldr);
cannam@127:      X(plan_destroy_internal)(ego->cldc);
cannam@127: }
cannam@127: 
cannam@127: static void print(const plan *ego_, printer *p)
cannam@127: {
cannam@127:      const P *ego = (const P *) ego_;
cannam@127:      const S *s = ego->solver;
cannam@127:      p->print(p, "(rdft2-rank>=2/%d%(%p%)%(%p%))", 
cannam@127: 	      s->spltrnk, ego->cldr, ego->cldc);
cannam@127: }
cannam@127:  
cannam@127: static int picksplit(const S *ego, const tensor *sz, int *rp)
cannam@127: {
cannam@127:      A(sz->rnk > 1); /* cannot split rnk <= 1 */
cannam@127:      if (!X(pickdim)(ego->spltrnk, ego->buddies, ego->nbuddies, sz, 1, rp))
cannam@127:           return 0;
cannam@127:      *rp += 1; /* convert from dim. index to rank */
cannam@127:      if (*rp >= sz->rnk) /* split must reduce rank */
cannam@127:           return 0;
cannam@127:      return 1;
cannam@127: }
cannam@127: 
cannam@127: static int applicable0(const solver *ego_, const problem *p_, int *rp,
cannam@127: 		       const planner *plnr)
cannam@127: {
cannam@127:      const problem_rdft2 *p = (const problem_rdft2 *) p_;
cannam@127:      const S *ego = (const S *)ego_;
cannam@127:      return (1
cannam@127: 	     && FINITE_RNK(p->sz->rnk) && FINITE_RNK(p->vecsz->rnk)
cannam@127: 
cannam@127: 	     /* FIXME: multidimensional R2HCII ? */
cannam@127: 	     && (p->kind == R2HC || p->kind == HC2R)
cannam@127: 
cannam@127: 	     && p->sz->rnk >= 2
cannam@127: 	     && picksplit(ego, p->sz, rp)
cannam@127: 	     && (0
cannam@127: 
cannam@127: 		 /* can work out-of-place, but HC2R destroys input */
cannam@127: 		 || (p->r0 != p->cr && 
cannam@127: 		     (p->kind == R2HC || !NO_DESTROY_INPUTP(plnr)))
cannam@127: 
cannam@127: 		 /* FIXME: what are sufficient conditions for inplace? */
cannam@127: 		 || (p->r0 == p->cr))
cannam@127: 	  );
cannam@127: }
cannam@127: 
cannam@127: /* TODO: revise this. */
cannam@127: static int applicable(const solver *ego_, const problem *p_, 
cannam@127: 		      const planner *plnr, int *rp)
cannam@127: {
cannam@127:      const S *ego = (const S *)ego_;
cannam@127: 
cannam@127:      if (!applicable0(ego_, p_, rp, plnr)) return 0;
cannam@127: 
cannam@127:      if (NO_RANK_SPLITSP(plnr) && (ego->spltrnk != ego->buddies[0]))
cannam@127:           return 0;
cannam@127: 
cannam@127:      if (NO_UGLYP(plnr)) {
cannam@127: 	  const problem_rdft2 *p = (const problem_rdft2 *) p_;
cannam@127: 
cannam@127: 	  /* Heuristic: if the vector stride is greater than the transform
cannam@127: 	     size, don't use (prefer to do the vector loop first with a
cannam@127: 	     vrank-geq1 plan). */
cannam@127: 	  if (p->vecsz->rnk > 0 &&
cannam@127: 	      X(tensor_min_stride)(p->vecsz) 
cannam@127: 	      > X(rdft2_tensor_max_index)(p->sz, p->kind))
cannam@127: 	       return 0;
cannam@127:      }
cannam@127: 
cannam@127:      return 1;
cannam@127: }
cannam@127: 
cannam@127: static plan *mkplan(const solver *ego_, const problem *p_, planner *plnr)
cannam@127: {
cannam@127:      const S *ego = (const S *) ego_;
cannam@127:      const problem_rdft2 *p;
cannam@127:      P *pln;
cannam@127:      plan *cldr = 0, *cldc = 0;
cannam@127:      tensor *sz1, *sz2, *vecszi, *sz2i;
cannam@127:      int spltrnk;
cannam@127:      inplace_kind k;
cannam@127:      problem *cldp;
cannam@127: 
cannam@127:      static const plan_adt padt = {
cannam@127: 	  X(rdft2_solve), awake, print, destroy
cannam@127:      };
cannam@127: 
cannam@127:      if (!applicable(ego_, p_, plnr, &spltrnk))
cannam@127:           return (plan *) 0;
cannam@127: 
cannam@127:      p = (const problem_rdft2 *) p_;
cannam@127:      X(tensor_split)(p->sz, &sz1, spltrnk, &sz2);
cannam@127: 
cannam@127:      k = p->kind == R2HC ? INPLACE_OS : INPLACE_IS;
cannam@127:      vecszi = X(tensor_copy_inplace)(p->vecsz, k);
cannam@127:      sz2i = X(tensor_copy_inplace)(sz2, k);
cannam@127: 
cannam@127:      /* complex data is ~half of real */
cannam@127:      sz2i->dims[sz2i->rnk - 1].n = sz2i->dims[sz2i->rnk - 1].n/2 + 1;
cannam@127: 
cannam@127:      cldr = X(mkplan_d)(plnr, 
cannam@127: 		       X(mkproblem_rdft2_d)(X(tensor_copy)(sz2),
cannam@127: 					    X(tensor_append)(p->vecsz, sz1),
cannam@127: 					    p->r0, p->r1,
cannam@127: 					    p->cr, p->ci, p->kind));
cannam@127:      if (!cldr) goto nada;
cannam@127: 
cannam@127:      if (p->kind == R2HC)
cannam@127: 	  cldp = X(mkproblem_dft_d)(X(tensor_copy_inplace)(sz1, k),
cannam@127: 				    X(tensor_append)(vecszi, sz2i),
cannam@127: 				    p->cr, p->ci, p->cr, p->ci);
cannam@127:      else /* HC2R must swap re/im parts to get IDFT */
cannam@127: 	  cldp = X(mkproblem_dft_d)(X(tensor_copy_inplace)(sz1, k),
cannam@127: 				    X(tensor_append)(vecszi, sz2i),
cannam@127: 				    p->ci, p->cr, p->ci, p->cr);
cannam@127:      cldc = X(mkplan_d)(plnr, cldp);
cannam@127:      if (!cldc) goto nada;
cannam@127: 
cannam@127:      pln = MKPLAN_RDFT2(P, &padt, p->kind == R2HC ? apply_r2hc : apply_hc2r);
cannam@127: 
cannam@127:      pln->cldr = cldr;
cannam@127:      pln->cldc = cldc;
cannam@127: 
cannam@127:      pln->solver = ego;
cannam@127:      X(ops_add)(&cldr->ops, &cldc->ops, &pln->super.super.ops);
cannam@127: 
cannam@127:      X(tensor_destroy4)(sz2i, vecszi, sz2, sz1);
cannam@127: 
cannam@127:      return &(pln->super.super);
cannam@127: 
cannam@127:  nada:
cannam@127:      X(plan_destroy_internal)(cldr);
cannam@127:      X(plan_destroy_internal)(cldc);
cannam@127:      X(tensor_destroy4)(sz2i, vecszi, sz2, sz1);
cannam@127:      return (plan *) 0;
cannam@127: }
cannam@127: 
cannam@127: static solver *mksolver(int spltrnk, const int *buddies, size_t nbuddies)
cannam@127: {
cannam@127:      static const solver_adt sadt = { PROBLEM_RDFT2, mkplan, 0 };
cannam@127:      S *slv = MKSOLVER(S, &sadt);
cannam@127:      slv->spltrnk = spltrnk;
cannam@127:      slv->buddies = buddies;
cannam@127:      slv->nbuddies = nbuddies;
cannam@127:      return &(slv->super);
cannam@127: }
cannam@127: 
cannam@127: void X(rdft2_rank_geq2_register)(planner *p)
cannam@127: {
cannam@127:      static const int buddies[] = { 1, 0, -2 };
cannam@127:      size_t i;
cannam@127: 
cannam@127:      for (i = 0; i < NELEM(buddies); ++i)
cannam@127:           REGISTER_SOLVER(p, mksolver(buddies[i], buddies, NELEM(buddies)));
cannam@127: 
cannam@127:      /* FIXME: Should we try more buddies?  See also dft/rank-geq2. */
cannam@127: }