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: 
Chris@10: /* Plans for handling vector transform loops.  These are *just* the
Chris@10:    loops, and rely on child plans for the actual RDFT2s.
Chris@10:  
Chris@10:    They form a wrapper around solvers that don't have apply functions
Chris@10:    for non-null vectors.
Chris@10:  
Chris@10:    vrank-geq1-rdft2 plans also recursively handle the case of
Chris@10:    multi-dimensional vectors, obviating the need for most solvers to
Chris@10:    deal with this.  We can also play games here, such as reordering
Chris@10:    the vector loops.
Chris@10:  
Chris@10:    Each vrank-geq1-rdft2 plan reduces the vector rank by 1, picking out a
Chris@10:    dimension determined by the vecloop_dim field of the solver. */
Chris@10: 
Chris@10: #include "rdft.h"
Chris@10: 
Chris@10: typedef struct {
Chris@10:      solver super;
Chris@10:      int vecloop_dim;
Chris@10:      const int *buddies;
Chris@10:      int nbuddies;
Chris@10: } S;
Chris@10: 
Chris@10: typedef struct {
Chris@10:      plan_rdft2 super;
Chris@10: 
Chris@10:      plan *cld;
Chris@10:      INT vl;
Chris@10:      INT rvs, cvs;
Chris@10:      const S *solver;
Chris@10: } P;
Chris@10: 
Chris@10: static void apply(const plan *ego_, R *r0, R *r1, R *cr, R *ci)
Chris@10: {
Chris@10:      const P *ego = (const P *) ego_;
Chris@10:      INT i, vl = ego->vl;
Chris@10:      INT rvs = ego->rvs, cvs = ego->cvs;
Chris@10:      rdft2apply cldapply = ((plan_rdft2 *) ego->cld)->apply;
Chris@10: 
Chris@10:      for (i = 0; i < vl; ++i) {
Chris@10:           cldapply(ego->cld, r0 + i * rvs, r1 + i * rvs,
Chris@10: 		   cr + i * cvs, ci + i * cvs);
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->cld, 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->cld);
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, "(rdft2-vrank>=1-x%D/%d%(%p%))",
Chris@10: 	      ego->vl, s->vecloop_dim, ego->cld);
Chris@10: }
Chris@10: 
Chris@10: static int pickdim(const S *ego, const tensor *vecsz, int oop, int *dp)
Chris@10: {
Chris@10:      return X(pickdim)(ego->vecloop_dim, ego->buddies, ego->nbuddies,
Chris@10: 		       vecsz, oop, dp);
Chris@10: }
Chris@10: 
Chris@10: static int applicable0(const solver *ego_, const problem *p_, int *dp)
Chris@10: {
Chris@10:      const S *ego = (const S *) ego_;
Chris@10:      const problem_rdft2 *p = (const problem_rdft2 *) p_;
Chris@10:      if (FINITE_RNK(p->vecsz->rnk)
Chris@10: 	 && p->vecsz->rnk > 0
Chris@10: 	 && pickdim(ego, p->vecsz, p->r0 != p->cr, dp)) {
Chris@10: 	  if (p->r0 != p->cr)
Chris@10: 	       return 1;  /* can always operate out-of-place */
Chris@10: 
Chris@10: 	  return(X(rdft2_inplace_strides)(p, *dp));
Chris@10:      }
Chris@10: 
Chris@10:      return 0;
Chris@10: }
Chris@10: 
Chris@10: 
Chris@10: static int applicable(const solver *ego_, const problem *p_,
Chris@10: 		      const planner *plnr, int *dp)
Chris@10: {
Chris@10:      const S *ego = (const S *)ego_;
Chris@10:      if (!applicable0(ego_, p_, dp)) return 0;
Chris@10: 
Chris@10:      /* fftw2 behavior */
Chris@10:      if (NO_VRANK_SPLITSP(plnr) && (ego->vecloop_dim != ego->buddies[0]))
Chris@10: 	  return 0;
Chris@10: 
Chris@10:      if (NO_UGLYP(plnr)) {
Chris@10: 	  const problem_rdft2 *p = (const problem_rdft2 *) p_;
Chris@10: 	  iodim *d = p->vecsz->dims + *dp;
Chris@10: 	       
Chris@10: 	  /* Heuristic: if the transform is multi-dimensional, and the
Chris@10: 	     vector stride is less than the transform size, then we
Chris@10: 	     probably want to use a rank>=2 plan first in order to combine
Chris@10: 	     this vector with the transform-dimension vectors. */
Chris@10: 	  if (p->sz->rnk > 1
Chris@10: 	      && X(imin)(X(iabs)(d->is), X(iabs)(d->os))
Chris@10: 	      < X(rdft2_tensor_max_index)(p->sz, p->kind)
Chris@10: 	       )
Chris@10: 	       return 0;
Chris@10: 
Chris@10: 	  /* Heuristic: don't use a vrank-geq1 for rank-0 vrank-1
Chris@10: 	     transforms, since this case is better handled by rank-0
Chris@10: 	     solvers. */
Chris@10: 	  if (p->sz->rnk == 0 && p->vecsz->rnk == 1) return 0;
Chris@10: 
Chris@10: 	  if (NO_NONTHREADEDP(plnr)) 
Chris@10: 	       return 0; /* prefer threaded version */
Chris@10:      }
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_rdft2 *p;
Chris@10:      P *pln;
Chris@10:      plan *cld;
Chris@10:      int vdim;
Chris@10:      iodim *d;
Chris@10:      INT rvs, cvs;
Chris@10: 
Chris@10:      static const plan_adt padt = {
Chris@10: 	  X(rdft2_solve), awake, print, destroy
Chris@10:      };
Chris@10: 
Chris@10:      if (!applicable(ego_, p_, plnr, &vdim))
Chris@10:           return (plan *) 0;
Chris@10:      p = (const problem_rdft2 *) p_;
Chris@10: 
Chris@10:      d = p->vecsz->dims + vdim;
Chris@10: 
Chris@10:      A(d->n > 1);  /* or else, p->ri + d->is etc. are invalid */
Chris@10: 
Chris@10:      X(rdft2_strides)(p->kind, d, &rvs, &cvs);
Chris@10: 
Chris@10:      cld = X(mkplan_d)(plnr, 
Chris@10: 		       X(mkproblem_rdft2_d)(
Chris@10: 			    X(tensor_copy)(p->sz),
Chris@10: 			    X(tensor_copy_except)(p->vecsz, vdim),
Chris@10: 			    TAINT(p->r0, rvs), TAINT(p->r1, rvs), 
Chris@10: 			    TAINT(p->cr, cvs), TAINT(p->ci, cvs),
Chris@10: 			    p->kind));
Chris@10:      if (!cld) return (plan *) 0;
Chris@10: 
Chris@10:      pln = MKPLAN_RDFT2(P, &padt, apply);
Chris@10: 
Chris@10:      pln->cld = cld;
Chris@10:      pln->vl = d->n;
Chris@10:      pln->rvs = rvs;
Chris@10:      pln->cvs = cvs;
Chris@10: 
Chris@10:      pln->solver = ego;
Chris@10:      X(ops_zero)(&pln->super.super.ops);
Chris@10:      pln->super.super.ops.other = 3.14159; /* magic to prefer codelet loops */
Chris@10:      X(ops_madd2)(pln->vl, &cld->ops, &pln->super.super.ops);
Chris@10: 
Chris@10:      if (p->sz->rnk != 1 || (p->sz->dims[0].n > 128))
Chris@10: 	  pln->super.super.pcost = pln->vl * cld->pcost;
Chris@10: 
Chris@10:      return &(pln->super.super);
Chris@10: }
Chris@10: 
Chris@10: static solver *mksolver(int vecloop_dim, const int *buddies, int nbuddies)
Chris@10: {
Chris@10:      static const solver_adt sadt = { PROBLEM_RDFT2, mkplan, 0 };
Chris@10:      S *slv = MKSOLVER(S, &sadt);
Chris@10:      slv->vecloop_dim = vecloop_dim;
Chris@10:      slv->buddies = buddies;
Chris@10:      slv->nbuddies = nbuddies;
Chris@10:      return &(slv->super);
Chris@10: }
Chris@10: 
Chris@10: void X(rdft2_vrank_geq1_register)(planner *p)
Chris@10: {
Chris@10:      int i;
Chris@10: 
Chris@10:      /* FIXME: Should we try other vecloop_dim values? */
Chris@10:      static const int buddies[] = { 1, -1 };
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: }