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: /* Do a REDFT00 problem via an R2HC problem, padded symmetrically to
cannam@95:    twice the size.  This is asymptotically a factor of ~2 worse than
cannam@95:    redft00e-r2hc.c (the algorithm used in e.g. FFTPACK and Numerical
cannam@95:    Recipes), but we abandoned the latter after we discovered that it
cannam@95:    has intrinsic accuracy problems. */
cannam@95: 
cannam@95: #include "reodft.h"
cannam@95: 
cannam@95: typedef struct {
cannam@95:      solver super;
cannam@95: } S;
cannam@95: 
cannam@95: typedef struct {
cannam@95:      plan_rdft super;
cannam@95:      plan *cld, *cldcpy;
cannam@95:      INT is;
cannam@95:      INT n;
cannam@95:      INT vl;
cannam@95:      INT ivs, ovs;
cannam@95: } P;
cannam@95: 
cannam@95: static void apply(const plan *ego_, R *I, R *O)
cannam@95: {
cannam@95:      const P *ego = (const P *) ego_;
cannam@95:      INT is = ego->is;
cannam@95:      INT i, n = ego->n;
cannam@95:      INT iv, vl = ego->vl;
cannam@95:      INT ivs = ego->ivs, ovs = ego->ovs;
cannam@95:      R *buf;
cannam@95: 
cannam@95:      buf = (R *) MALLOC(sizeof(R) * (2*n), BUFFERS);
cannam@95: 
cannam@95:      for (iv = 0; iv < vl; ++iv, I += ivs, O += ovs) {
cannam@95: 	  buf[0] = I[0];
cannam@95: 	  for (i = 1; i < n; ++i) {
cannam@95: 	       R a = I[i * is];
cannam@95: 	       buf[i] = a;
cannam@95: 	       buf[2*n - i] = a;
cannam@95: 	  }
cannam@95: 	  buf[i] = I[i * is]; /* i == n, Nyquist */
cannam@95: 	  
cannam@95: 	  /* r2hc transform of size 2*n */
cannam@95: 	  {
cannam@95: 	       plan_rdft *cld = (plan_rdft *) ego->cld;
cannam@95: 	       cld->apply((plan *) cld, buf, buf);
cannam@95: 	  }
cannam@95: 	  
cannam@95: 	  /* copy n+1 real numbers (real parts of hc array) from buf to O */
cannam@95: 	  {
cannam@95: 	       plan_rdft *cldcpy = (plan_rdft *) ego->cldcpy;
cannam@95: 	       cldcpy->apply((plan *) cldcpy, buf, O);
cannam@95: 	  }
cannam@95:      }
cannam@95: 
cannam@95:      X(ifree)(buf);
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->cld, wakefulness);
cannam@95:      X(plan_awake)(ego->cldcpy, wakefulness);
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->cldcpy);
cannam@95:      X(plan_destroy_internal)(ego->cld);
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, "(redft00e-r2hc-pad-%D%v%(%p%)%(%p%))", 
cannam@95: 	      ego->n + 1, ego->vl, ego->cld, ego->cldcpy);
cannam@95: }
cannam@95: 
cannam@95: static int applicable0(const solver *ego_, const problem *p_)
cannam@95: {
cannam@95:      const problem_rdft *p = (const problem_rdft *) p_;
cannam@95:      UNUSED(ego_);
cannam@95: 
cannam@95:      return (1
cannam@95: 	     && p->sz->rnk == 1
cannam@95: 	     && p->vecsz->rnk <= 1
cannam@95: 	     && p->kind[0] == REDFT00
cannam@95: 	     && p->sz->dims[0].n > 1  /* n == 1 is not well-defined */
cannam@95: 	  );
cannam@95: }
cannam@95: 
cannam@95: static int applicable(const solver *ego, const problem *p, const planner *plnr)
cannam@95: {
cannam@95:      return (!NO_SLOWP(plnr) && applicable0(ego, p));
cannam@95: }
cannam@95: 
cannam@95: static plan *mkplan(const solver *ego_, const problem *p_, planner *plnr)
cannam@95: {
cannam@95:      P *pln;
cannam@95:      const problem_rdft *p;
cannam@95:      plan *cld = (plan *) 0, *cldcpy;
cannam@95:      R *buf = (R *) 0;
cannam@95:      INT n;
cannam@95:      INT vl, ivs, ovs;
cannam@95:      opcnt ops;
cannam@95: 
cannam@95:      static const plan_adt padt = {
cannam@95: 	  X(rdft_solve), awake, print, destroy
cannam@95:      };
cannam@95: 
cannam@95:      if (!applicable(ego_, p_, plnr))
cannam@95: 	  goto nada;
cannam@95: 
cannam@95:      p = (const problem_rdft *) p_;
cannam@95: 
cannam@95:      n = p->sz->dims[0].n - 1;
cannam@95:      A(n > 0);
cannam@95:      buf = (R *) MALLOC(sizeof(R) * (2*n), BUFFERS);
cannam@95: 
cannam@95:      cld = X(mkplan_d)(plnr,X(mkproblem_rdft_1_d)(X(mktensor_1d)(2*n,1,1), 
cannam@95: 						  X(mktensor_0d)(), 
cannam@95: 						  buf, buf, R2HC));
cannam@95:      if (!cld)
cannam@95: 	  goto nada;
cannam@95: 
cannam@95:      X(tensor_tornk1)(p->vecsz, &vl, &ivs, &ovs);
cannam@95:      cldcpy =
cannam@95: 	  X(mkplan_d)(plnr,
cannam@95: 		      X(mkproblem_rdft_1_d)(X(mktensor_0d)(),
cannam@95: 					    X(mktensor_1d)(n+1,1,
cannam@95: 							   p->sz->dims[0].os), 
cannam@95: 					    buf, TAINT(p->O, ovs), R2HC));
cannam@95:      if (!cldcpy)
cannam@95: 	  goto nada;
cannam@95: 
cannam@95:      X(ifree)(buf);
cannam@95: 
cannam@95:      pln = MKPLAN_RDFT(P, &padt, apply);
cannam@95: 
cannam@95:      pln->n = n;
cannam@95:      pln->is = p->sz->dims[0].is;
cannam@95:      pln->cld = cld;
cannam@95:      pln->cldcpy = cldcpy;
cannam@95:      pln->vl = vl;
cannam@95:      pln->ivs = ivs;
cannam@95:      pln->ovs = ovs;
cannam@95:      
cannam@95:      X(ops_zero)(&ops);
cannam@95:      ops.other = n + 2*n; /* loads + stores (input -> buf) */
cannam@95: 
cannam@95:      X(ops_zero)(&pln->super.super.ops);
cannam@95:      X(ops_madd2)(pln->vl, &ops, &pln->super.super.ops);
cannam@95:      X(ops_madd2)(pln->vl, &cld->ops, &pln->super.super.ops);
cannam@95:      X(ops_madd2)(pln->vl, &cldcpy->ops, &pln->super.super.ops);
cannam@95: 
cannam@95:      return &(pln->super.super);
cannam@95: 
cannam@95:  nada:
cannam@95:      X(ifree0)(buf);
cannam@95:      if (cld)
cannam@95: 	  X(plan_destroy_internal)(cld);  
cannam@95:      return (plan *)0;
cannam@95: }
cannam@95: 
cannam@95: /* constructor */
cannam@95: static solver *mksolver(void)
cannam@95: {
cannam@95:      static const solver_adt sadt = { PROBLEM_RDFT, mkplan, 0 };
cannam@95:      S *slv = MKSOLVER(S, &sadt);
cannam@95:      return &(slv->super);
cannam@95: }
cannam@95: 
cannam@95: void X(redft00e_r2hc_pad_register)(planner *p)
cannam@95: {
cannam@95:      REGISTER_SOLVER(p, mksolver());
cannam@95: }