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: /* solvers/plans for vectors of small DFT's that cannot be done
Chris@10:    in-place directly.  Use a rank-0 plan to rearrange the data
Chris@10:    before or after the transform.  Can also change an out-of-place
Chris@10:    plan into a copy + in-place (where the in-place transform
Chris@10:    is e.g. unit stride). */
Chris@10: 
Chris@10: /* FIXME: merge with rank-geq2.c(?), since this is just a special case
Chris@10:    of a rank split where the first/second transform has rank 0. */
Chris@10: 
Chris@10: #include "dft.h"
Chris@10: 
Chris@10: typedef problem *(*mkcld_t) (const problem_dft *p);
Chris@10: 
Chris@10: typedef struct {
Chris@10:      dftapply apply;
Chris@10:      problem *(*mkcld)(const problem_dft *p);
Chris@10:      const char *nam;
Chris@10: } ndrct_adt;
Chris@10: 
Chris@10: typedef struct {
Chris@10:      solver super;
Chris@10:      const ndrct_adt *adt;
Chris@10: } S;
Chris@10: 
Chris@10: typedef struct {
Chris@10:      plan_dft super;
Chris@10:      plan *cldcpy, *cld;
Chris@10:      const S *slv;
Chris@10: } P;
Chris@10: 
Chris@10: /*-----------------------------------------------------------------------*/
Chris@10: /* first rearrange, then transform */
Chris@10: static void apply_before(const plan *ego_, R *ri, R *ii, R *ro, R *io)
Chris@10: {
Chris@10:      const P *ego = (const P *) ego_;
Chris@10: 
Chris@10:      {
Chris@10:           plan_dft *cldcpy = (plan_dft *) ego->cldcpy;
Chris@10:           cldcpy->apply(ego->cldcpy, ri, ii, ro, io);
Chris@10:      }
Chris@10:      {
Chris@10:           plan_dft *cld = (plan_dft *) ego->cld;
Chris@10:           cld->apply(ego->cld, ro, io, ro, io);
Chris@10:      }
Chris@10: }
Chris@10: 
Chris@10: static problem *mkcld_before(const problem_dft *p)
Chris@10: {
Chris@10:      return X(mkproblem_dft_d)(X(tensor_copy_inplace)(p->sz, INPLACE_OS),
Chris@10: 			       X(tensor_copy_inplace)(p->vecsz, INPLACE_OS),
Chris@10: 			       p->ro, p->io, p->ro, p->io);
Chris@10: }
Chris@10: 
Chris@10: static const ndrct_adt adt_before =
Chris@10: {
Chris@10:      apply_before, mkcld_before, "dft-indirect-before"
Chris@10: };
Chris@10: 
Chris@10: /*-----------------------------------------------------------------------*/
Chris@10: /* first transform, then rearrange */
Chris@10: 
Chris@10: static void apply_after(const plan *ego_, R *ri, R *ii, R *ro, R *io)
Chris@10: {
Chris@10:      const P *ego = (const P *) ego_;
Chris@10: 
Chris@10:      {
Chris@10:           plan_dft *cld = (plan_dft *) ego->cld;
Chris@10:           cld->apply(ego->cld, ri, ii, ri, ii);
Chris@10:      }
Chris@10:      {
Chris@10:           plan_dft *cldcpy = (plan_dft *) ego->cldcpy;
Chris@10:           cldcpy->apply(ego->cldcpy, ri, ii, ro, io);
Chris@10:      }
Chris@10: }
Chris@10: 
Chris@10: static problem *mkcld_after(const problem_dft *p)
Chris@10: {
Chris@10:      return X(mkproblem_dft_d)(X(tensor_copy_inplace)(p->sz, INPLACE_IS),
Chris@10: 			       X(tensor_copy_inplace)(p->vecsz, INPLACE_IS),
Chris@10: 			       p->ri, p->ii, p->ri, p->ii);
Chris@10: }
Chris@10: 
Chris@10: static const ndrct_adt adt_after =
Chris@10: {
Chris@10:      apply_after, mkcld_after, "dft-indirect-after"
Chris@10: };
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:      X(plan_destroy_internal)(ego->cldcpy);
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->cldcpy, wakefulness);
Chris@10:      X(plan_awake)(ego->cld, wakefulness);
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->slv;
Chris@10:      p->print(p, "(%s%(%p%)%(%p%))", s->adt->nam, ego->cld, ego->cldcpy);
Chris@10: }
Chris@10: 
Chris@10: static int applicable0(const solver *ego_, const problem *p_,
Chris@10: 		       const planner *plnr)
Chris@10: {
Chris@10:      const S *ego = (const S *) ego_;
Chris@10:      const problem_dft *p = (const problem_dft *) p_;
Chris@10:      return (1
Chris@10: 	     && FINITE_RNK(p->vecsz->rnk)
Chris@10: 
Chris@10: 	     /* problem must be a nontrivial transform, not just a copy */
Chris@10: 	     && p->sz->rnk > 0
Chris@10: 
Chris@10: 	     && (0
Chris@10: 
Chris@10: 		 /* problem must be in-place & require some
Chris@10: 		    rearrangement of the data; to prevent
Chris@10: 		    infinite loops with indirect-transpose, we
Chris@10: 		    further require that at least some transform
Chris@10: 		    strides must decrease */
Chris@10: 		 || (p->ri == p->ro
Chris@10: 		     && !X(tensor_inplace_strides2)(p->sz, p->vecsz)
Chris@10: 		     && X(tensor_strides_decrease)(
Chris@10: 			  p->sz, p->vecsz,
Chris@10: 			  ego->adt->apply == apply_after ? 
Chris@10: 			  INPLACE_IS : INPLACE_OS))
Chris@10: 
Chris@10: 		 /* or problem must be out of place, transforming
Chris@10: 		    from stride 1/2 to bigger stride, for apply_after */
Chris@10: 		 || (p->ri != p->ro && ego->adt->apply == apply_after
Chris@10: 		     && !NO_DESTROY_INPUTP(plnr)
Chris@10: 		     && X(tensor_min_istride)(p->sz) <= 2
Chris@10: 		     && X(tensor_min_ostride)(p->sz) > 2)
Chris@10: 			  
Chris@10: 		 /* or problem must be out of place, transforming
Chris@10: 		    to stride 1/2 from bigger stride, for apply_before */
Chris@10: 		 || (p->ri != p->ro && ego->adt->apply == apply_before
Chris@10: 		     && X(tensor_min_ostride)(p->sz) <= 2
Chris@10: 		     && X(tensor_min_istride)(p->sz) > 2)
Chris@10: 		  )
Chris@10: 	  );
Chris@10: }
Chris@10: 
Chris@10: static int applicable(const solver *ego_, const problem *p_,
Chris@10: 		      const planner *plnr)
Chris@10: {
Chris@10:      if (!applicable0(ego_, p_, plnr)) return 0;
Chris@10:      {
Chris@10:           const problem_dft *p = (const problem_dft *) p_;
Chris@10: 	  if (NO_INDIRECT_OP_P(plnr) && p->ri != p->ro) 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 problem_dft *p = (const problem_dft *) p_;
Chris@10:      const S *ego = (const S *) ego_;
Chris@10:      P *pln;
Chris@10:      plan *cld = 0, *cldcpy = 0;
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))
Chris@10:           return (plan *) 0;
Chris@10: 
Chris@10:      cldcpy =
Chris@10: 	  X(mkplan_d)(plnr, 
Chris@10: 		      X(mkproblem_dft_d)(X(mktensor_0d)(),
Chris@10: 					 X(tensor_append)(p->vecsz, p->sz),
Chris@10: 					 p->ri, p->ii, p->ro, p->io));
Chris@10: 
Chris@10:      if (!cldcpy) goto nada;
Chris@10: 
Chris@10:      cld = X(mkplan_f_d)(plnr, ego->adt->mkcld(p), NO_BUFFERING, 0, 0);
Chris@10:      if (!cld) goto nada;
Chris@10: 
Chris@10:      pln = MKPLAN_DFT(P, &padt, ego->adt->apply);
Chris@10:      pln->cld = cld;
Chris@10:      pln->cldcpy = cldcpy;
Chris@10:      pln->slv = ego;
Chris@10:      X(ops_add)(&cld->ops, &cldcpy->ops, &pln->super.super.ops);
Chris@10: 
Chris@10:      return &(pln->super.super);
Chris@10: 
Chris@10:  nada:
Chris@10:      X(plan_destroy_internal)(cld);
Chris@10:      X(plan_destroy_internal)(cldcpy);
Chris@10:      return (plan *)0;
Chris@10: }
Chris@10: 
Chris@10: static solver *mksolver(const ndrct_adt *adt)
Chris@10: {
Chris@10:      static const solver_adt sadt = { PROBLEM_DFT, mkplan, 0 };
Chris@10:      S *slv = MKSOLVER(S, &sadt);
Chris@10:      slv->adt = adt;
Chris@10:      return &(slv->super);
Chris@10: }
Chris@10: 
Chris@10: void X(dft_indirect_register)(planner *p)
Chris@10: {
Chris@10:      unsigned i;
Chris@10:      static const ndrct_adt *const adts[] = {
Chris@10: 	  &adt_before, &adt_after
Chris@10:      };
Chris@10: 
Chris@10:      for (i = 0; i < sizeof(adts) / sizeof(adts[0]); ++i)
Chris@10:           REGISTER_SOLVER(p, mksolver(adts[i]));
Chris@10: }