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: /* solvers/plans for vectors of DFTs corresponding to the columns
Chris@10:    of a matrix: first transpose the matrix so that the DFTs are
Chris@10:    contiguous, then do DFTs with transposed output.   In particular,
Chris@10:    we restrict ourselves to the case of a square transpose (or a
Chris@10:    sequence thereof). */
Chris@10: 
Chris@10: #include "dft.h"
Chris@10: 
Chris@10: typedef solver S;
Chris@10: 
Chris@10: typedef struct {
Chris@10:      plan_dft super;
Chris@10:      INT vl, ivs, ovs;
Chris@10:      plan *cldtrans, *cld, *cldrest;
Chris@10: } P;
Chris@10: 
Chris@10: /* initial transpose is out-of-place from input to output */
Chris@10: static void apply_op(const plan *ego_, R *ri, R *ii, R *ro, R *io)
Chris@10: {
Chris@10:      const P *ego = (const P *) ego_;
Chris@10:      INT vl = ego->vl, ivs = ego->ivs, ovs = ego->ovs, i;
Chris@10: 
Chris@10:      for (i = 0; i < vl; ++i) {
Chris@10: 	  {
Chris@10: 	       plan_dft *cldtrans = (plan_dft *) ego->cldtrans;
Chris@10: 	       cldtrans->apply(ego->cldtrans, 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: 	  ri += ivs; ii += ivs;
Chris@10: 	  ro += ovs; io += ovs;
Chris@10:      }
Chris@10:      {
Chris@10: 	  plan_dft *cldrest = (plan_dft *) ego->cldrest;
Chris@10: 	  cldrest->apply(ego->cldrest, ri, ii, ro, io);
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->cldrest);
Chris@10:      X(plan_destroy_internal)(ego->cld);
Chris@10:      X(plan_destroy_internal)(ego->cldtrans);
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->cldtrans, wakefulness);
Chris@10:      X(plan_awake)(ego->cld, wakefulness);
Chris@10:      X(plan_awake)(ego->cldrest, 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:      p->print(p, "(indirect-transpose%v%(%p%)%(%p%)%(%p%))", 
Chris@10: 	      ego->vl, ego->cldtrans, ego->cld, ego->cldrest);
Chris@10: }
Chris@10: 
Chris@10: static int pickdim(const tensor *vs, const tensor *s, int *pdim0, int *pdim1)
Chris@10: {
Chris@10:      int dim0, dim1;
Chris@10:      *pdim0 = *pdim1 = -1;
Chris@10:      for (dim0 = 0; dim0 < vs->rnk; ++dim0)
Chris@10:           for (dim1 = 0; dim1 < s->rnk; ++dim1) 
Chris@10: 	       if (vs->dims[dim0].n * X(iabs)(vs->dims[dim0].is) <= X(iabs)(s->dims[dim1].is)
Chris@10: 		   && vs->dims[dim0].n >= s->dims[dim1].n
Chris@10: 		   && (*pdim0 == -1 
Chris@10: 		       || (X(iabs)(vs->dims[dim0].is) <= X(iabs)(vs->dims[*pdim0].is)
Chris@10: 			   && X(iabs)(s->dims[dim1].is) >= X(iabs)(s->dims[*pdim1].is)))) {
Chris@10: 		    *pdim0 = dim0;
Chris@10: 		    *pdim1 = dim1;
Chris@10: 	       }
Chris@10:      return (*pdim0 != -1 && *pdim1 != -1);
Chris@10: }
Chris@10: 
Chris@10: static int applicable0(const solver *ego_, const problem *p_,
Chris@10: 		       const planner *plnr,
Chris@10: 		       int *pdim0, int *pdim1)
Chris@10: {
Chris@10:      const problem_dft *p = (const problem_dft *) p_;
Chris@10:      UNUSED(ego_); UNUSED(plnr);
Chris@10: 
Chris@10:      return (1
Chris@10: 	     && FINITE_RNK(p->vecsz->rnk) && FINITE_RNK(p->sz->rnk)
Chris@10: 
Chris@10: 	     /* FIXME: can/should we relax this constraint? */
Chris@10: 	     && X(tensor_inplace_strides2)(p->vecsz, p->sz)
Chris@10: 
Chris@10: 	     && pickdim(p->vecsz, p->sz, pdim0, pdim1)
Chris@10: 
Chris@10: 	     /* output should not *already* include the transpose
Chris@10: 		(in which case we duplicate the regular indirect.c) */
Chris@10: 	     && (p->sz->dims[*pdim1].os != p->vecsz->dims[*pdim0].is)
Chris@10: 	  );
Chris@10: }
Chris@10: 
Chris@10: static int applicable(const solver *ego_, const problem *p_,
Chris@10: 		      const planner *plnr,
Chris@10: 		      int *pdim0, int *pdim1)
Chris@10: {
Chris@10:      if (!applicable0(ego_, p_, plnr, pdim0, pdim1)) return 0;
Chris@10:      {
Chris@10:           const problem_dft *p = (const problem_dft *) p_;
Chris@10: 	  INT u = p->ri == p->ii + 1 || p->ii == p->ri + 1 ? (INT)2 : (INT)1;
Chris@10: 
Chris@10: 	  /* UGLY if does not result in contiguous transforms or
Chris@10: 	     transforms of contiguous vectors (since the latter at
Chris@10: 	     least have efficient transpositions) */
Chris@10: 	  if (NO_UGLYP(plnr)
Chris@10: 	      && p->vecsz->dims[*pdim0].is != u
Chris@10: 	      && !(p->vecsz->rnk == 2
Chris@10: 		   && p->vecsz->dims[1-*pdim0].is == u
Chris@10: 		   && p->vecsz->dims[*pdim0].is
Chris@10: 		      == u * p->vecsz->dims[1-*pdim0].n))
Chris@10: 	       return 0;
Chris@10: 
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:      P *pln;
Chris@10:      plan *cld = 0, *cldtrans = 0, *cldrest = 0;
Chris@10:      int pdim0, pdim1;
Chris@10:      tensor *ts, *tv;
Chris@10:      INT vl, ivs, ovs;
Chris@10:      R *rit, *iit, *rot, *iot;
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, &pdim0, &pdim1))
Chris@10:           return (plan *) 0;
Chris@10: 
Chris@10:      vl = p->vecsz->dims[pdim0].n / p->sz->dims[pdim1].n;
Chris@10:      A(vl >= 1);
Chris@10:      ivs = p->sz->dims[pdim1].n * p->vecsz->dims[pdim0].is;
Chris@10:      ovs = p->sz->dims[pdim1].n * p->vecsz->dims[pdim0].os;
Chris@10:      rit = TAINT(p->ri, vl == 1 ? 0 : ivs);
Chris@10:      iit = TAINT(p->ii, vl == 1 ? 0 : ivs);
Chris@10:      rot = TAINT(p->ro, vl == 1 ? 0 : ovs);
Chris@10:      iot = TAINT(p->io, vl == 1 ? 0 : ovs);
Chris@10: 
Chris@10:      ts = X(tensor_copy_inplace)(p->sz, INPLACE_IS);
Chris@10:      ts->dims[pdim1].os = p->vecsz->dims[pdim0].is;
Chris@10:      tv = X(tensor_copy_inplace)(p->vecsz, INPLACE_IS);
Chris@10:      tv->dims[pdim0].os = p->sz->dims[pdim1].is;
Chris@10:      tv->dims[pdim0].n = p->sz->dims[pdim1].n;
Chris@10:      cldtrans = X(mkplan_d)(plnr, 
Chris@10: 			    X(mkproblem_dft_d)(X(mktensor_0d)(),
Chris@10: 					       X(tensor_append)(tv, ts),
Chris@10: 					       rit, iit, 
Chris@10: 					       rot, iot));
Chris@10:      X(tensor_destroy2)(ts, tv);
Chris@10:      if (!cldtrans) goto nada;
Chris@10: 
Chris@10:      ts = X(tensor_copy)(p->sz);
Chris@10:      ts->dims[pdim1].is = p->vecsz->dims[pdim0].is;
Chris@10:      tv = X(tensor_copy)(p->vecsz);
Chris@10:      tv->dims[pdim0].is = p->sz->dims[pdim1].is;
Chris@10:      tv->dims[pdim0].n = p->sz->dims[pdim1].n;
Chris@10:      cld = X(mkplan_d)(plnr, X(mkproblem_dft_d)(ts, tv,
Chris@10: 						rot, iot,
Chris@10: 						rot, iot));
Chris@10:      if (!cld) goto nada;
Chris@10: 
Chris@10:      tv = X(tensor_copy)(p->vecsz);
Chris@10:      tv->dims[pdim0].n -= vl * p->sz->dims[pdim1].n;
Chris@10:      cldrest = X(mkplan_d)(plnr, X(mkproblem_dft_d)(X(tensor_copy)(p->sz), tv,
Chris@10: 						    p->ri + ivs * vl,
Chris@10: 						    p->ii + ivs * vl,
Chris@10: 						    p->ro + ovs * vl,
Chris@10: 						    p->io + ovs * vl));
Chris@10:      if (!cldrest) goto nada;
Chris@10: 
Chris@10:      pln = MKPLAN_DFT(P, &padt, apply_op);
Chris@10:      pln->cldtrans = cldtrans;
Chris@10:      pln->cld = cld;
Chris@10:      pln->cldrest = cldrest;
Chris@10:      pln->vl = vl;
Chris@10:      pln->ivs = ivs;
Chris@10:      pln->ovs = ovs;
Chris@10:      X(ops_cpy)(&cldrest->ops, &pln->super.super.ops);
Chris@10:      X(ops_madd2)(vl, &cld->ops, &pln->super.super.ops);
Chris@10:      X(ops_madd2)(vl, &cldtrans->ops, &pln->super.super.ops);
Chris@10:      return &(pln->super.super);
Chris@10: 
Chris@10:  nada:
Chris@10:      X(plan_destroy_internal)(cldrest);
Chris@10:      X(plan_destroy_internal)(cld);
Chris@10:      X(plan_destroy_internal)(cldtrans);
Chris@10:      return (plan *)0;
Chris@10: }
Chris@10: 
Chris@10: static solver *mksolver(void)
Chris@10: {
Chris@10:      static const solver_adt sadt = { PROBLEM_DFT, mkplan, 0 };
Chris@10:      S *slv = MKSOLVER(S, &sadt);
Chris@10:      return slv;
Chris@10: }
Chris@10: 
Chris@10: void X(dft_indirect_transpose_register)(planner *p)
Chris@10: {
Chris@10:      REGISTER_SOLVER(p, mksolver());
Chris@10: }