cannam@95: /*
cannam@95:  * Copyright (c) 2003, 2007-11 Matteo Frigo
cannam@95:  * Copyright (c) 2003, 2007-11 Massachusetts Institute of Technology
cannam@95:  *
cannam@95:  * This program is free software; you can redistribute it and/or modify
cannam@95:  * it under the terms of the GNU General Public License as published by
cannam@95:  * the Free Software Foundation; either version 2 of the License, or
cannam@95:  * (at your option) any later version.
cannam@95:  *
cannam@95:  * This program is distributed in the hope that it will be useful,
cannam@95:  * but WITHOUT ANY WARRANTY; without even the implied warranty of
cannam@95:  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
cannam@95:  * GNU General Public License for more details.
cannam@95:  *
cannam@95:  * You should have received a copy of the GNU General Public License
cannam@95:  * along with this program; if not, write to the Free Software
cannam@95:  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA
cannam@95:  *
cannam@95:  */
cannam@95: 
cannam@95: 
cannam@95: 
cannam@95: /* solvers/plans for vectors of small DFT's that cannot be done
cannam@95:    in-place directly.  Use a rank-0 plan to rearrange the data
cannam@95:    before or after the transform.  Can also change an out-of-place
cannam@95:    plan into a copy + in-place (where the in-place transform
cannam@95:    is e.g. unit stride). */
cannam@95: 
cannam@95: /* FIXME: merge with rank-geq2.c(?), since this is just a special case
cannam@95:    of a rank split where the first/second transform has rank 0. */
cannam@95: 
cannam@95: #include "dft.h"
cannam@95: 
cannam@95: typedef problem *(*mkcld_t) (const problem_dft *p);
cannam@95: 
cannam@95: typedef struct {
cannam@95:      dftapply apply;
cannam@95:      problem *(*mkcld)(const problem_dft *p);
cannam@95:      const char *nam;
cannam@95: } ndrct_adt;
cannam@95: 
cannam@95: typedef struct {
cannam@95:      solver super;
cannam@95:      const ndrct_adt *adt;
cannam@95: } S;
cannam@95: 
cannam@95: typedef struct {
cannam@95:      plan_dft super;
cannam@95:      plan *cldcpy, *cld;
cannam@95:      const S *slv;
cannam@95: } P;
cannam@95: 
cannam@95: /*-----------------------------------------------------------------------*/
cannam@95: /* first rearrange, then transform */
cannam@95: static void apply_before(const plan *ego_, R *ri, R *ii, R *ro, R *io)
cannam@95: {
cannam@95:      const P *ego = (const P *) ego_;
cannam@95: 
cannam@95:      {
cannam@95:           plan_dft *cldcpy = (plan_dft *) ego->cldcpy;
cannam@95:           cldcpy->apply(ego->cldcpy, ri, ii, ro, io);
cannam@95:      }
cannam@95:      {
cannam@95:           plan_dft *cld = (plan_dft *) ego->cld;
cannam@95:           cld->apply(ego->cld, ro, io, ro, io);
cannam@95:      }
cannam@95: }
cannam@95: 
cannam@95: static problem *mkcld_before(const problem_dft *p)
cannam@95: {
cannam@95:      return X(mkproblem_dft_d)(X(tensor_copy_inplace)(p->sz, INPLACE_OS),
cannam@95: 			       X(tensor_copy_inplace)(p->vecsz, INPLACE_OS),
cannam@95: 			       p->ro, p->io, p->ro, p->io);
cannam@95: }
cannam@95: 
cannam@95: static const ndrct_adt adt_before =
cannam@95: {
cannam@95:      apply_before, mkcld_before, "dft-indirect-before"
cannam@95: };
cannam@95: 
cannam@95: /*-----------------------------------------------------------------------*/
cannam@95: /* first transform, then rearrange */
cannam@95: 
cannam@95: static void apply_after(const plan *ego_, R *ri, R *ii, R *ro, R *io)
cannam@95: {
cannam@95:      const P *ego = (const P *) ego_;
cannam@95: 
cannam@95:      {
cannam@95:           plan_dft *cld = (plan_dft *) ego->cld;
cannam@95:           cld->apply(ego->cld, ri, ii, ri, ii);
cannam@95:      }
cannam@95:      {
cannam@95:           plan_dft *cldcpy = (plan_dft *) ego->cldcpy;
cannam@95:           cldcpy->apply(ego->cldcpy, ri, ii, ro, io);
cannam@95:      }
cannam@95: }
cannam@95: 
cannam@95: static problem *mkcld_after(const problem_dft *p)
cannam@95: {
cannam@95:      return X(mkproblem_dft_d)(X(tensor_copy_inplace)(p->sz, INPLACE_IS),
cannam@95: 			       X(tensor_copy_inplace)(p->vecsz, INPLACE_IS),
cannam@95: 			       p->ri, p->ii, p->ri, p->ii);
cannam@95: }
cannam@95: 
cannam@95: static const ndrct_adt adt_after =
cannam@95: {
cannam@95:      apply_after, mkcld_after, "dft-indirect-after"
cannam@95: };
cannam@95: 
cannam@95: /*-----------------------------------------------------------------------*/
cannam@95: static void destroy(plan *ego_)
cannam@95: {
cannam@95:      P *ego = (P *) ego_;
cannam@95:      X(plan_destroy_internal)(ego->cld);
cannam@95:      X(plan_destroy_internal)(ego->cldcpy);
cannam@95: }
cannam@95: 
cannam@95: static void awake(plan *ego_, enum wakefulness wakefulness)
cannam@95: {
cannam@95:      P *ego = (P *) ego_;
cannam@95:      X(plan_awake)(ego->cldcpy, wakefulness);
cannam@95:      X(plan_awake)(ego->cld, wakefulness);
cannam@95: }
cannam@95: 
cannam@95: static void print(const plan *ego_, printer *p)
cannam@95: {
cannam@95:      const P *ego = (const P *) ego_;
cannam@95:      const S *s = ego->slv;
cannam@95:      p->print(p, "(%s%(%p%)%(%p%))", s->adt->nam, ego->cld, ego->cldcpy);
cannam@95: }
cannam@95: 
cannam@95: static int applicable0(const solver *ego_, const problem *p_,
cannam@95: 		       const planner *plnr)
cannam@95: {
cannam@95:      const S *ego = (const S *) ego_;
cannam@95:      const problem_dft *p = (const problem_dft *) p_;
cannam@95:      return (1
cannam@95: 	     && FINITE_RNK(p->vecsz->rnk)
cannam@95: 
cannam@95: 	     /* problem must be a nontrivial transform, not just a copy */
cannam@95: 	     && p->sz->rnk > 0
cannam@95: 
cannam@95: 	     && (0
cannam@95: 
cannam@95: 		 /* problem must be in-place & require some
cannam@95: 		    rearrangement of the data; to prevent
cannam@95: 		    infinite loops with indirect-transpose, we
cannam@95: 		    further require that at least some transform
cannam@95: 		    strides must decrease */
cannam@95: 		 || (p->ri == p->ro
cannam@95: 		     && !X(tensor_inplace_strides2)(p->sz, p->vecsz)
cannam@95: 		     && X(tensor_strides_decrease)(
cannam@95: 			  p->sz, p->vecsz,
cannam@95: 			  ego->adt->apply == apply_after ? 
cannam@95: 			  INPLACE_IS : INPLACE_OS))
cannam@95: 
cannam@95: 		 /* or problem must be out of place, transforming
cannam@95: 		    from stride 1/2 to bigger stride, for apply_after */
cannam@95: 		 || (p->ri != p->ro && ego->adt->apply == apply_after
cannam@95: 		     && !NO_DESTROY_INPUTP(plnr)
cannam@95: 		     && X(tensor_min_istride)(p->sz) <= 2
cannam@95: 		     && X(tensor_min_ostride)(p->sz) > 2)
cannam@95: 			  
cannam@95: 		 /* or problem must be out of place, transforming
cannam@95: 		    to stride 1/2 from bigger stride, for apply_before */
cannam@95: 		 || (p->ri != p->ro && ego->adt->apply == apply_before
cannam@95: 		     && X(tensor_min_ostride)(p->sz) <= 2
cannam@95: 		     && X(tensor_min_istride)(p->sz) > 2)
cannam@95: 		  )
cannam@95: 	  );
cannam@95: }
cannam@95: 
cannam@95: static int applicable(const solver *ego_, const problem *p_,
cannam@95: 		      const planner *plnr)
cannam@95: {
cannam@95:      if (!applicable0(ego_, p_, plnr)) return 0;
cannam@95:      {
cannam@95:           const problem_dft *p = (const problem_dft *) p_;
cannam@95: 	  if (NO_INDIRECT_OP_P(plnr) && p->ri != p->ro) return 0;
cannam@95:      }
cannam@95:      return 1;
cannam@95: }
cannam@95: 
cannam@95: static plan *mkplan(const solver *ego_, const problem *p_, planner *plnr)
cannam@95: {
cannam@95:      const problem_dft *p = (const problem_dft *) p_;
cannam@95:      const S *ego = (const S *) ego_;
cannam@95:      P *pln;
cannam@95:      plan *cld = 0, *cldcpy = 0;
cannam@95: 
cannam@95:      static const plan_adt padt = {
cannam@95: 	  X(dft_solve), awake, print, destroy
cannam@95:      };
cannam@95: 
cannam@95:      if (!applicable(ego_, p_, plnr))
cannam@95:           return (plan *) 0;
cannam@95: 
cannam@95:      cldcpy =
cannam@95: 	  X(mkplan_d)(plnr, 
cannam@95: 		      X(mkproblem_dft_d)(X(mktensor_0d)(),
cannam@95: 					 X(tensor_append)(p->vecsz, p->sz),
cannam@95: 					 p->ri, p->ii, p->ro, p->io));
cannam@95: 
cannam@95:      if (!cldcpy) goto nada;
cannam@95: 
cannam@95:      cld = X(mkplan_f_d)(plnr, ego->adt->mkcld(p), NO_BUFFERING, 0, 0);
cannam@95:      if (!cld) goto nada;
cannam@95: 
cannam@95:      pln = MKPLAN_DFT(P, &padt, ego->adt->apply);
cannam@95:      pln->cld = cld;
cannam@95:      pln->cldcpy = cldcpy;
cannam@95:      pln->slv = ego;
cannam@95:      X(ops_add)(&cld->ops, &cldcpy->ops, &pln->super.super.ops);
cannam@95: 
cannam@95:      return &(pln->super.super);
cannam@95: 
cannam@95:  nada:
cannam@95:      X(plan_destroy_internal)(cld);
cannam@95:      X(plan_destroy_internal)(cldcpy);
cannam@95:      return (plan *)0;
cannam@95: }
cannam@95: 
cannam@95: static solver *mksolver(const ndrct_adt *adt)
cannam@95: {
cannam@95:      static const solver_adt sadt = { PROBLEM_DFT, mkplan, 0 };
cannam@95:      S *slv = MKSOLVER(S, &sadt);
cannam@95:      slv->adt = adt;
cannam@95:      return &(slv->super);
cannam@95: }
cannam@95: 
cannam@95: void X(dft_indirect_register)(planner *p)
cannam@95: {
cannam@95:      unsigned i;
cannam@95:      static const ndrct_adt *const adts[] = {
cannam@95: 	  &adt_before, &adt_after
cannam@95:      };
cannam@95: 
cannam@95:      for (i = 0; i < sizeof(adts) / sizeof(adts[0]); ++i)
cannam@95:           REGISTER_SOLVER(p, mksolver(adts[i]));
cannam@95: }