cannam@167: /*
cannam@167:  * Copyright (c) 2003, 2007-14 Matteo Frigo
cannam@167:  * Copyright (c) 2003, 2007-14 Massachusetts Institute of Technology
cannam@167:  *
cannam@167:  * This program is free software; you can redistribute it and/or modify
cannam@167:  * it under the terms of the GNU General Public License as published by
cannam@167:  * the Free Software Foundation; either version 2 of the License, or
cannam@167:  * (at your option) any later version.
cannam@167:  *
cannam@167:  * This program is distributed in the hope that it will be useful,
cannam@167:  * but WITHOUT ANY WARRANTY; without even the implied warranty of
cannam@167:  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
cannam@167:  * GNU General Public License for more details.
cannam@167:  *
cannam@167:  * You should have received a copy of the GNU General Public License
cannam@167:  * along with this program; if not, write to the Free Software
cannam@167:  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA
cannam@167:  *
cannam@167:  */
cannam@167: 
cannam@167: 
cannam@167: 
cannam@167: /* solvers/plans for vectors of small RDFT's that cannot be done
cannam@167:    in-place directly.  Use a rank-0 plan to rearrange the data
cannam@167:    before or after the transform.  Can also change an out-of-place
cannam@167:    plan into a copy + in-place (where the in-place transform
cannam@167:    is e.g. unit stride). */
cannam@167: 
cannam@167: /* FIXME: merge with rank-geq2.c(?), since this is just a special case
cannam@167:    of a rank split where the first/second transform has rank 0. */
cannam@167: 
cannam@167: #include "rdft/rdft.h"
cannam@167: 
cannam@167: typedef problem *(*mkcld_t) (const problem_rdft *p);
cannam@167: 
cannam@167: typedef struct {
cannam@167:      rdftapply apply;
cannam@167:      problem *(*mkcld)(const problem_rdft *p);
cannam@167:      const char *nam;
cannam@167: } ndrct_adt;
cannam@167: 
cannam@167: typedef struct {
cannam@167:      solver super;
cannam@167:      const ndrct_adt *adt;
cannam@167: } S;
cannam@167: 
cannam@167: typedef struct {
cannam@167:      plan_rdft super;
cannam@167:      plan *cldcpy, *cld;
cannam@167:      const S *slv;
cannam@167: } P;
cannam@167: 
cannam@167: /*-----------------------------------------------------------------------*/
cannam@167: /* first rearrange, then transform */
cannam@167: static void apply_before(const plan *ego_, R *I, R *O)
cannam@167: {
cannam@167:      const P *ego = (const P *) ego_;
cannam@167: 
cannam@167:      {
cannam@167:           plan_rdft *cldcpy = (plan_rdft *) ego->cldcpy;
cannam@167:           cldcpy->apply(ego->cldcpy, I, O);
cannam@167:      }
cannam@167:      {
cannam@167:           plan_rdft *cld = (plan_rdft *) ego->cld;
cannam@167:           cld->apply(ego->cld, O, O);
cannam@167:      }
cannam@167: }
cannam@167: 
cannam@167: static problem *mkcld_before(const problem_rdft *p)
cannam@167: {
cannam@167:      return X(mkproblem_rdft_d)(X(tensor_copy_inplace)(p->sz, INPLACE_OS),
cannam@167: 				X(tensor_copy_inplace)(p->vecsz, INPLACE_OS),
cannam@167: 				p->O, p->O, p->kind);
cannam@167: }
cannam@167: 
cannam@167: static const ndrct_adt adt_before =
cannam@167: {
cannam@167:      apply_before, mkcld_before, "rdft-indirect-before"
cannam@167: };
cannam@167: 
cannam@167: /*-----------------------------------------------------------------------*/
cannam@167: /* first transform, then rearrange */
cannam@167: 
cannam@167: static void apply_after(const plan *ego_, R *I, R *O)
cannam@167: {
cannam@167:      const P *ego = (const P *) ego_;
cannam@167: 
cannam@167:      {
cannam@167:           plan_rdft *cld = (plan_rdft *) ego->cld;
cannam@167:           cld->apply(ego->cld, I, I);
cannam@167:      }
cannam@167:      {
cannam@167:           plan_rdft *cldcpy = (plan_rdft *) ego->cldcpy;
cannam@167:           cldcpy->apply(ego->cldcpy, I, O);
cannam@167:      }
cannam@167: }
cannam@167: 
cannam@167: static problem *mkcld_after(const problem_rdft *p)
cannam@167: {
cannam@167:      return X(mkproblem_rdft_d)(X(tensor_copy_inplace)(p->sz, INPLACE_IS),
cannam@167: 				X(tensor_copy_inplace)(p->vecsz, INPLACE_IS),
cannam@167: 				p->I, p->I, p->kind);
cannam@167: }
cannam@167: 
cannam@167: static const ndrct_adt adt_after =
cannam@167: {
cannam@167:      apply_after, mkcld_after, "rdft-indirect-after"
cannam@167: };
cannam@167: 
cannam@167: /*-----------------------------------------------------------------------*/
cannam@167: static void destroy(plan *ego_)
cannam@167: {
cannam@167:      P *ego = (P *) ego_;
cannam@167:      X(plan_destroy_internal)(ego->cld);
cannam@167:      X(plan_destroy_internal)(ego->cldcpy);
cannam@167: }
cannam@167: 
cannam@167: static void awake(plan *ego_, enum wakefulness wakefulness)
cannam@167: {
cannam@167:      P *ego = (P *) ego_;
cannam@167:      X(plan_awake)(ego->cldcpy, wakefulness);
cannam@167:      X(plan_awake)(ego->cld, wakefulness);
cannam@167: }
cannam@167: 
cannam@167: static void print(const plan *ego_, printer *p)
cannam@167: {
cannam@167:      const P *ego = (const P *) ego_;
cannam@167:      const S *s = ego->slv;
cannam@167:      p->print(p, "(%s%(%p%)%(%p%))", s->adt->nam, ego->cld, ego->cldcpy);
cannam@167: }
cannam@167: 
cannam@167: static int applicable0(const solver *ego_, const problem *p_,
cannam@167: 		       const planner *plnr)
cannam@167: {
cannam@167:      const S *ego = (const S *) ego_;
cannam@167:      const problem_rdft *p = (const problem_rdft *) p_;
cannam@167:      return (1
cannam@167: 	     && FINITE_RNK(p->vecsz->rnk)
cannam@167: 
cannam@167: 	     /* problem must be a nontrivial transform, not just a copy */
cannam@167: 	     && p->sz->rnk > 0
cannam@167: 
cannam@167: 	     && (0
cannam@167: 
cannam@167: 		 /* problem must be in-place & require some
cannam@167: 		    rearrangement of the data */
cannam@167: 		 || (p->I == p->O
cannam@167: 		     && !(X(tensor_inplace_strides2)(p->sz, p->vecsz)))
cannam@167: 
cannam@167: 		 /* or problem must be out of place, transforming
cannam@167: 		    from stride 1/2 to bigger stride, for apply_after */
cannam@167: 		 || (p->I != p->O && ego->adt->apply == apply_after
cannam@167: 		     && !NO_DESTROY_INPUTP(plnr)
cannam@167: 		     && X(tensor_min_istride)(p->sz) <= 2
cannam@167: 		     && X(tensor_min_ostride)(p->sz) > 2)
cannam@167: 			  
cannam@167: 		 /* or problem must be out of place, transforming
cannam@167: 		    to stride 1/2 from bigger stride, for apply_before */
cannam@167: 		 || (p->I != p->O && ego->adt->apply == apply_before
cannam@167: 		     && X(tensor_min_ostride)(p->sz) <= 2
cannam@167: 		     && X(tensor_min_istride)(p->sz) > 2)
cannam@167: 			  
cannam@167: 		  )
cannam@167: 	  );
cannam@167: }
cannam@167: 
cannam@167: static int applicable(const solver *ego_, const problem *p_,
cannam@167: 		      const planner *plnr)
cannam@167: {
cannam@167:      if (!applicable0(ego_, p_, plnr)) return 0;
cannam@167: 	  
cannam@167:      if (NO_INDIRECT_OP_P(plnr)) {
cannam@167: 	  const problem_rdft *p = (const problem_rdft *)p_;
cannam@167: 	  if (p->I != p->O) return 0;
cannam@167:      }
cannam@167: 
cannam@167:      return 1;
cannam@167: }
cannam@167: 
cannam@167: static plan *mkplan(const solver *ego_, const problem *p_, planner *plnr)
cannam@167: {
cannam@167:      const problem_rdft *p = (const problem_rdft *) p_;
cannam@167:      const S *ego = (const S *) ego_;
cannam@167:      P *pln;
cannam@167:      plan *cld = 0, *cldcpy = 0;
cannam@167: 
cannam@167:      static const plan_adt padt = {
cannam@167: 	  X(rdft_solve), awake, print, destroy
cannam@167:      };
cannam@167: 
cannam@167:      if (!applicable(ego_, p_, plnr))
cannam@167:           return (plan *) 0;
cannam@167: 
cannam@167:      cldcpy = X(mkplan_d)(plnr,
cannam@167: 			  X(mkproblem_rdft_0_d)(
cannam@167: 			       X(tensor_append)(p->vecsz, p->sz),
cannam@167: 			       p->I, p->O));
cannam@167:      if (!cldcpy) goto nada;
cannam@167: 
cannam@167:      cld = X(mkplan_f_d)(plnr, ego->adt->mkcld(p), NO_BUFFERING, 0, 0);
cannam@167:      if (!cld) goto nada;
cannam@167: 
cannam@167:      pln = MKPLAN_RDFT(P, &padt, ego->adt->apply);
cannam@167:      pln->cld = cld;
cannam@167:      pln->cldcpy = cldcpy;
cannam@167:      pln->slv = ego;
cannam@167:      X(ops_add)(&cld->ops, &cldcpy->ops, &pln->super.super.ops);
cannam@167: 
cannam@167:      return &(pln->super.super);
cannam@167: 
cannam@167:  nada:
cannam@167:      X(plan_destroy_internal)(cld);
cannam@167:      X(plan_destroy_internal)(cldcpy);
cannam@167:      return (plan *)0;
cannam@167: }
cannam@167: 
cannam@167: static solver *mksolver(const ndrct_adt *adt)
cannam@167: {
cannam@167:      static const solver_adt sadt = { PROBLEM_RDFT, mkplan, 0 };
cannam@167:      S *slv = MKSOLVER(S, &sadt);
cannam@167:      slv->adt = adt;
cannam@167:      return &(slv->super);
cannam@167: }
cannam@167: 
cannam@167: void X(rdft_indirect_register)(planner *p)
cannam@167: {
cannam@167:      unsigned i;
cannam@167:      static const ndrct_adt *const adts[] = {
cannam@167: 	  &adt_before, &adt_after
cannam@167:      };
cannam@167: 
cannam@167:      for (i = 0; i < sizeof(adts) / sizeof(adts[0]); ++i)
cannam@167:           REGISTER_SOLVER(p, mksolver(adts[i]));
cannam@167: }