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: /* solvers/plans for vectors of DFTs corresponding to the columns
cannam@95:    of a matrix: first transpose the matrix so that the DFTs are
cannam@95:    contiguous, then do DFTs with transposed output.   In particular,
cannam@95:    we restrict ourselves to the case of a square transpose (or a
cannam@95:    sequence thereof). */
cannam@95: 
cannam@95: #include "dft.h"
cannam@95: 
cannam@95: typedef solver S;
cannam@95: 
cannam@95: typedef struct {
cannam@95:      plan_dft super;
cannam@95:      INT vl, ivs, ovs;
cannam@95:      plan *cldtrans, *cld, *cldrest;
cannam@95: } P;
cannam@95: 
cannam@95: /* initial transpose is out-of-place from input to output */
cannam@95: static void apply_op(const plan *ego_, R *ri, R *ii, R *ro, R *io)
cannam@95: {
cannam@95:      const P *ego = (const P *) ego_;
cannam@95:      INT vl = ego->vl, ivs = ego->ivs, ovs = ego->ovs, i;
cannam@95: 
cannam@95:      for (i = 0; i < vl; ++i) {
cannam@95: 	  {
cannam@95: 	       plan_dft *cldtrans = (plan_dft *) ego->cldtrans;
cannam@95: 	       cldtrans->apply(ego->cldtrans, 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: 	  ri += ivs; ii += ivs;
cannam@95: 	  ro += ovs; io += ovs;
cannam@95:      }
cannam@95:      {
cannam@95: 	  plan_dft *cldrest = (plan_dft *) ego->cldrest;
cannam@95: 	  cldrest->apply(ego->cldrest, ri, ii, ro, io);
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->cldrest);
cannam@95:      X(plan_destroy_internal)(ego->cld);
cannam@95:      X(plan_destroy_internal)(ego->cldtrans);
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->cldtrans, wakefulness);
cannam@95:      X(plan_awake)(ego->cld, wakefulness);
cannam@95:      X(plan_awake)(ego->cldrest, 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:      p->print(p, "(indirect-transpose%v%(%p%)%(%p%)%(%p%))", 
cannam@95: 	      ego->vl, ego->cldtrans, ego->cld, ego->cldrest);
cannam@95: }
cannam@95: 
cannam@95: static int pickdim(const tensor *vs, const tensor *s, int *pdim0, int *pdim1)
cannam@95: {
cannam@95:      int dim0, dim1;
cannam@95:      *pdim0 = *pdim1 = -1;
cannam@95:      for (dim0 = 0; dim0 < vs->rnk; ++dim0)
cannam@95:           for (dim1 = 0; dim1 < s->rnk; ++dim1) 
cannam@95: 	       if (vs->dims[dim0].n * X(iabs)(vs->dims[dim0].is) <= X(iabs)(s->dims[dim1].is)
cannam@95: 		   && vs->dims[dim0].n >= s->dims[dim1].n
cannam@95: 		   && (*pdim0 == -1 
cannam@95: 		       || (X(iabs)(vs->dims[dim0].is) <= X(iabs)(vs->dims[*pdim0].is)
cannam@95: 			   && X(iabs)(s->dims[dim1].is) >= X(iabs)(s->dims[*pdim1].is)))) {
cannam@95: 		    *pdim0 = dim0;
cannam@95: 		    *pdim1 = dim1;
cannam@95: 	       }
cannam@95:      return (*pdim0 != -1 && *pdim1 != -1);
cannam@95: }
cannam@95: 
cannam@95: static int applicable0(const solver *ego_, const problem *p_,
cannam@95: 		       const planner *plnr,
cannam@95: 		       int *pdim0, int *pdim1)
cannam@95: {
cannam@95:      const problem_dft *p = (const problem_dft *) p_;
cannam@95:      UNUSED(ego_); UNUSED(plnr);
cannam@95: 
cannam@95:      return (1
cannam@95: 	     && FINITE_RNK(p->vecsz->rnk) && FINITE_RNK(p->sz->rnk)
cannam@95: 
cannam@95: 	     /* FIXME: can/should we relax this constraint? */
cannam@95: 	     && X(tensor_inplace_strides2)(p->vecsz, p->sz)
cannam@95: 
cannam@95: 	     && pickdim(p->vecsz, p->sz, pdim0, pdim1)
cannam@95: 
cannam@95: 	     /* output should not *already* include the transpose
cannam@95: 		(in which case we duplicate the regular indirect.c) */
cannam@95: 	     && (p->sz->dims[*pdim1].os != p->vecsz->dims[*pdim0].is)
cannam@95: 	  );
cannam@95: }
cannam@95: 
cannam@95: static int applicable(const solver *ego_, const problem *p_,
cannam@95: 		      const planner *plnr,
cannam@95: 		      int *pdim0, int *pdim1)
cannam@95: {
cannam@95:      if (!applicable0(ego_, p_, plnr, pdim0, pdim1)) return 0;
cannam@95:      {
cannam@95:           const problem_dft *p = (const problem_dft *) p_;
cannam@95: 	  INT u = p->ri == p->ii + 1 || p->ii == p->ri + 1 ? (INT)2 : (INT)1;
cannam@95: 
cannam@95: 	  /* UGLY if does not result in contiguous transforms or
cannam@95: 	     transforms of contiguous vectors (since the latter at
cannam@95: 	     least have efficient transpositions) */
cannam@95: 	  if (NO_UGLYP(plnr)
cannam@95: 	      && p->vecsz->dims[*pdim0].is != u
cannam@95: 	      && !(p->vecsz->rnk == 2
cannam@95: 		   && p->vecsz->dims[1-*pdim0].is == u
cannam@95: 		   && p->vecsz->dims[*pdim0].is
cannam@95: 		      == u * p->vecsz->dims[1-*pdim0].n))
cannam@95: 	       return 0;
cannam@95: 
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:      P *pln;
cannam@95:      plan *cld = 0, *cldtrans = 0, *cldrest = 0;
cannam@95:      int pdim0, pdim1;
cannam@95:      tensor *ts, *tv;
cannam@95:      INT vl, ivs, ovs;
cannam@95:      R *rit, *iit, *rot, *iot;
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, &pdim0, &pdim1))
cannam@95:           return (plan *) 0;
cannam@95: 
cannam@95:      vl = p->vecsz->dims[pdim0].n / p->sz->dims[pdim1].n;
cannam@95:      A(vl >= 1);
cannam@95:      ivs = p->sz->dims[pdim1].n * p->vecsz->dims[pdim0].is;
cannam@95:      ovs = p->sz->dims[pdim1].n * p->vecsz->dims[pdim0].os;
cannam@95:      rit = TAINT(p->ri, vl == 1 ? 0 : ivs);
cannam@95:      iit = TAINT(p->ii, vl == 1 ? 0 : ivs);
cannam@95:      rot = TAINT(p->ro, vl == 1 ? 0 : ovs);
cannam@95:      iot = TAINT(p->io, vl == 1 ? 0 : ovs);
cannam@95: 
cannam@95:      ts = X(tensor_copy_inplace)(p->sz, INPLACE_IS);
cannam@95:      ts->dims[pdim1].os = p->vecsz->dims[pdim0].is;
cannam@95:      tv = X(tensor_copy_inplace)(p->vecsz, INPLACE_IS);
cannam@95:      tv->dims[pdim0].os = p->sz->dims[pdim1].is;
cannam@95:      tv->dims[pdim0].n = p->sz->dims[pdim1].n;
cannam@95:      cldtrans = X(mkplan_d)(plnr, 
cannam@95: 			    X(mkproblem_dft_d)(X(mktensor_0d)(),
cannam@95: 					       X(tensor_append)(tv, ts),
cannam@95: 					       rit, iit, 
cannam@95: 					       rot, iot));
cannam@95:      X(tensor_destroy2)(ts, tv);
cannam@95:      if (!cldtrans) goto nada;
cannam@95: 
cannam@95:      ts = X(tensor_copy)(p->sz);
cannam@95:      ts->dims[pdim1].is = p->vecsz->dims[pdim0].is;
cannam@95:      tv = X(tensor_copy)(p->vecsz);
cannam@95:      tv->dims[pdim0].is = p->sz->dims[pdim1].is;
cannam@95:      tv->dims[pdim0].n = p->sz->dims[pdim1].n;
cannam@95:      cld = X(mkplan_d)(plnr, X(mkproblem_dft_d)(ts, tv,
cannam@95: 						rot, iot,
cannam@95: 						rot, iot));
cannam@95:      if (!cld) goto nada;
cannam@95: 
cannam@95:      tv = X(tensor_copy)(p->vecsz);
cannam@95:      tv->dims[pdim0].n -= vl * p->sz->dims[pdim1].n;
cannam@95:      cldrest = X(mkplan_d)(plnr, X(mkproblem_dft_d)(X(tensor_copy)(p->sz), tv,
cannam@95: 						    p->ri + ivs * vl,
cannam@95: 						    p->ii + ivs * vl,
cannam@95: 						    p->ro + ovs * vl,
cannam@95: 						    p->io + ovs * vl));
cannam@95:      if (!cldrest) goto nada;
cannam@95: 
cannam@95:      pln = MKPLAN_DFT(P, &padt, apply_op);
cannam@95:      pln->cldtrans = cldtrans;
cannam@95:      pln->cld = cld;
cannam@95:      pln->cldrest = cldrest;
cannam@95:      pln->vl = vl;
cannam@95:      pln->ivs = ivs;
cannam@95:      pln->ovs = ovs;
cannam@95:      X(ops_cpy)(&cldrest->ops, &pln->super.super.ops);
cannam@95:      X(ops_madd2)(vl, &cld->ops, &pln->super.super.ops);
cannam@95:      X(ops_madd2)(vl, &cldtrans->ops, &pln->super.super.ops);
cannam@95:      return &(pln->super.super);
cannam@95: 
cannam@95:  nada:
cannam@95:      X(plan_destroy_internal)(cldrest);
cannam@95:      X(plan_destroy_internal)(cld);
cannam@95:      X(plan_destroy_internal)(cldtrans);
cannam@95:      return (plan *)0;
cannam@95: }
cannam@95: 
cannam@95: static solver *mksolver(void)
cannam@95: {
cannam@95:      static const solver_adt sadt = { PROBLEM_DFT, mkplan, 0 };
cannam@95:      S *slv = MKSOLVER(S, &sadt);
cannam@95:      return slv;
cannam@95: }
cannam@95: 
cannam@95: void X(dft_indirect_transpose_register)(planner *p)
cannam@95: {
cannam@95:      REGISTER_SOLVER(p, mksolver());
cannam@95: }