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: /* Compute the complex DFT by combining R2HC RDFTs on the real
Chris@10:    and imaginary parts.   This could be useful for people just wanting
Chris@10:    to link to the real codelets and not the complex ones.  It could
Chris@10:    also even be faster than the complex algorithms for split (as opposed
Chris@10:    to interleaved) real/imag complex data. */
Chris@10: 
Chris@10: #include "rdft.h"
Chris@10: #include "dft.h"
Chris@10: 
Chris@10: typedef struct {
Chris@10:      solver super;
Chris@10: } S;
Chris@10: 
Chris@10: typedef struct {
Chris@10:      plan_dft super;
Chris@10:      plan *cld;
Chris@10:      INT ishift, oshift;
Chris@10:      INT os;
Chris@10:      INT n;
Chris@10: } P;
Chris@10: 
Chris@10: static void apply(const plan *ego_, R *ri, R *ii, R *ro, R *io)
Chris@10: {
Chris@10:      const P *ego = (const P *) ego_;
Chris@10:      INT n;
Chris@10: 
Chris@10:      UNUSED(ii);
Chris@10: 
Chris@10:      { /* transform vector of real & imag parts: */
Chris@10: 	  plan_rdft *cld = (plan_rdft *) ego->cld;
Chris@10: 	  cld->apply((plan *) cld, ri + ego->ishift, ro + ego->oshift);
Chris@10:      }
Chris@10: 
Chris@10:      n = ego->n;
Chris@10:      if (n > 1) {
Chris@10: 	  INT i, os = ego->os;
Chris@10: 	  for (i = 1; i < (n + 1)/2; ++i) {
Chris@10: 	       E rop, iop, iom, rom;
Chris@10: 	       rop = ro[os * i];
Chris@10: 	       iop = io[os * i];
Chris@10: 	       rom = ro[os * (n - i)];
Chris@10: 	       iom = io[os * (n - i)];
Chris@10: 	       ro[os * i] = rop - iom;
Chris@10: 	       io[os * i] = iop + rom;
Chris@10: 	       ro[os * (n - i)] = rop + iom;
Chris@10: 	       io[os * (n - i)] = iop - rom;
Chris@10: 	  }
Chris@10:      }
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->cld, wakefulness);
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: }
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, "(dft-r2hc-%D%(%p%))", ego->n, ego->cld);
Chris@10: }
Chris@10: 
Chris@10: 
Chris@10: static int applicable0(const problem *p_)
Chris@10: {
Chris@10:      const problem_dft *p = (const problem_dft *) p_;
Chris@10:      return ((p->sz->rnk == 1 && p->vecsz->rnk == 0)
Chris@10: 	     || (p->sz->rnk == 0 && FINITE_RNK(p->vecsz->rnk))
Chris@10: 	  );
Chris@10: }
Chris@10: 
Chris@10: static int splitp(R *r, R *i, INT n, INT s)
Chris@10: {
Chris@10:      return ((r > i ? (r - i) : (i - r)) >= n * (s > 0 ? s : 0-s));
Chris@10: }
Chris@10: 
Chris@10: static int applicable(const problem *p_, const planner *plnr)
Chris@10: {
Chris@10:      if (!applicable0(p_)) return 0;
Chris@10: 
Chris@10:      {
Chris@10: 	  const problem_dft *p = (const problem_dft *) p_;
Chris@10: 
Chris@10: 	  /* rank-0 problems are always OK */
Chris@10: 	  if (p->sz->rnk == 0) return 1;
Chris@10: 
Chris@10: 	  /* this solver is ok for split arrays */
Chris@10: 	  if (p->sz->rnk == 1 &&
Chris@10: 	      splitp(p->ri, p->ii, p->sz->dims[0].n, p->sz->dims[0].is) &&
Chris@10: 	      splitp(p->ro, p->io, p->sz->dims[0].n, p->sz->dims[0].os))
Chris@10: 	       return 1;
Chris@10: 
Chris@10: 	  return !(NO_DFT_R2HCP(plnr));
Chris@10:      }
Chris@10: }
Chris@10: 
Chris@10: static plan *mkplan(const solver *ego_, const problem *p_, planner *plnr)
Chris@10: {
Chris@10:      P *pln;
Chris@10:      const problem_dft *p;
Chris@10:      plan *cld;
Chris@10:      INT ishift = 0, oshift = 0;
Chris@10: 
Chris@10:      static const plan_adt padt = {
Chris@10: 	  X(dft_solve), awake, print, destroy
Chris@10:      };
Chris@10: 
Chris@10:      UNUSED(ego_);
Chris@10:      if (!applicable(p_, plnr))
Chris@10:           return (plan *)0;
Chris@10: 
Chris@10:      p = (const problem_dft *) p_;
Chris@10: 
Chris@10:      {
Chris@10: 	  tensor *ri_vec = X(mktensor_1d)(2, p->ii - p->ri, p->io - p->ro);
Chris@10: 	  tensor *cld_vec = X(tensor_append)(ri_vec, p->vecsz);
Chris@10: 	  int i;
Chris@10: 	  for (i = 0; i < cld_vec->rnk; ++i) { /* make all istrides > 0 */
Chris@10: 	       if (cld_vec->dims[i].is < 0) {
Chris@10: 		    INT nm1 = cld_vec->dims[i].n - 1;
Chris@10: 		    ishift -= nm1 * (cld_vec->dims[i].is *= -1);
Chris@10: 		    oshift -= nm1 * (cld_vec->dims[i].os *= -1);
Chris@10: 	       }
Chris@10: 	  }
Chris@10: 	  cld = X(mkplan_d)(plnr, 
Chris@10: 			    X(mkproblem_rdft_1)(p->sz, cld_vec, 
Chris@10: 						p->ri + ishift, 
Chris@10: 						p->ro + oshift, R2HC));
Chris@10: 	  X(tensor_destroy2)(ri_vec, cld_vec);
Chris@10:      }
Chris@10:      if (!cld) return (plan *)0;
Chris@10: 
Chris@10:      pln = MKPLAN_DFT(P, &padt, apply);
Chris@10: 
Chris@10:      if (p->sz->rnk == 0) {
Chris@10: 	  pln->n = 1;
Chris@10: 	  pln->os = 0;
Chris@10:      }
Chris@10:      else {
Chris@10: 	  pln->n = p->sz->dims[0].n;
Chris@10: 	  pln->os = p->sz->dims[0].os;
Chris@10:      }
Chris@10:      pln->ishift = ishift;
Chris@10:      pln->oshift = oshift;
Chris@10: 
Chris@10:      pln->cld = cld;
Chris@10:      
Chris@10:      pln->super.super.ops = cld->ops;
Chris@10:      pln->super.super.ops.other += 8 * ((pln->n - 1)/2);
Chris@10:      pln->super.super.ops.add += 4 * ((pln->n - 1)/2);
Chris@10:      pln->super.super.ops.other += 1; /* estimator hack for nop plans */
Chris@10: 
Chris@10:      return &(pln->super.super);
Chris@10: }
Chris@10: 
Chris@10: /* constructor */
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->super);
Chris@10: }
Chris@10: 
Chris@10: void X(dft_r2hc_register)(planner *p)
Chris@10: {
Chris@10:      REGISTER_SOLVER(p, mksolver());
Chris@10: }