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: #include "rdft.h"
Chris@10: 
Chris@10: typedef struct {
Chris@10:      solver super;
Chris@10: } S;
Chris@10: 
Chris@10: typedef struct {
Chris@10:      plan_rdft2 super;
Chris@10: 
Chris@10:      plan *cld, *cldrest;
Chris@10:      INT n, vl, nbuf, bufdist;
Chris@10:      INT cs, ivs, ovs;
Chris@10: } P;
Chris@10: 
Chris@10: /***************************************************************************/
Chris@10: 
Chris@10: /* FIXME: have alternate copy functions that push a vector loop inside
Chris@10:    the n loops? */
Chris@10: 
Chris@10: /* copy halfcomplex array r (contiguous) to complex (strided) array rio/iio. */
Chris@10: static void hc2c(INT n, R *r, R *rio, R *iio, INT os)
Chris@10: {
Chris@10:      INT i;
Chris@10: 
Chris@10:      rio[0] = r[0];
Chris@10:      iio[0] = 0;
Chris@10: 
Chris@10:      for (i = 1; i + i < n; ++i) {
Chris@10: 	  rio[i * os] = r[i];
Chris@10: 	  iio[i * os] = r[n - i];
Chris@10:      }
Chris@10: 
Chris@10:      if (i + i == n) {	/* store the Nyquist frequency */
Chris@10: 	  rio[i * os] = r[i];
Chris@10: 	  iio[i * os] = K(0.0);
Chris@10:      }
Chris@10: }
Chris@10: 
Chris@10: /* reverse of hc2c */
Chris@10: static void c2hc(INT n, R *rio, R *iio, INT is, R *r)
Chris@10: {
Chris@10:      INT i;
Chris@10: 
Chris@10:      r[0] = rio[0];
Chris@10: 
Chris@10:      for (i = 1; i + i < n; ++i) {
Chris@10: 	  r[i] = rio[i * is];
Chris@10: 	  r[n - i] = iio[i * is];
Chris@10:      }
Chris@10: 
Chris@10:      if (i + i == n)		/* store the Nyquist frequency */
Chris@10: 	  r[i] = rio[i * is];
Chris@10: }
Chris@10: 
Chris@10: /***************************************************************************/
Chris@10: 
Chris@10: static void apply_r2hc(const plan *ego_, R *r0, R *r1, R *cr, R *ci)
Chris@10: {
Chris@10:      const P *ego = (const P *) ego_;
Chris@10:      plan_rdft *cld = (plan_rdft *) ego->cld;
Chris@10:      INT i, j, vl = ego->vl, nbuf = ego->nbuf, bufdist = ego->bufdist;
Chris@10:      INT n = ego->n;
Chris@10:      INT ivs = ego->ivs, ovs = ego->ovs, os = ego->cs;
Chris@10:      R *bufs = (R *)MALLOC(sizeof(R) * nbuf * bufdist, BUFFERS);
Chris@10:      plan_rdft2 *cldrest;
Chris@10: 
Chris@10:      for (i = nbuf; i <= vl; i += nbuf) {
Chris@10:           /* transform to bufs: */
Chris@10:           cld->apply((plan *) cld, r0, bufs);
Chris@10: 	  r0 += ivs * nbuf; r1 += ivs * nbuf;
Chris@10: 
Chris@10:           /* copy back */
Chris@10: 	  for (j = 0; j < nbuf; ++j, cr += ovs, ci += ovs)
Chris@10: 	       hc2c(n, bufs + j*bufdist, cr, ci, os);
Chris@10:      }
Chris@10: 
Chris@10:      X(ifree)(bufs);
Chris@10: 
Chris@10:      /* Do the remaining transforms, if any: */
Chris@10:      cldrest = (plan_rdft2 *) ego->cldrest;
Chris@10:      cldrest->apply((plan *) cldrest, r0, r1, cr, ci);
Chris@10: }
Chris@10: 
Chris@10: static void apply_hc2r(const plan *ego_, R *r0, R *r1, R *cr, R *ci)
Chris@10: {
Chris@10:      const P *ego = (const P *) ego_;
Chris@10:      plan_rdft *cld = (plan_rdft *) ego->cld;
Chris@10:      INT i, j, vl = ego->vl, nbuf = ego->nbuf, bufdist = ego->bufdist;
Chris@10:      INT n = ego->n;
Chris@10:      INT ivs = ego->ivs, ovs = ego->ovs, is = ego->cs;
Chris@10:      R *bufs = (R *)MALLOC(sizeof(R) * nbuf * bufdist, BUFFERS);
Chris@10:      plan_rdft2 *cldrest;
Chris@10: 
Chris@10:      for (i = nbuf; i <= vl; i += nbuf) {
Chris@10:           /* copy to bufs */
Chris@10: 	  for (j = 0; j < nbuf; ++j, cr += ivs, ci += ivs)
Chris@10: 	       c2hc(n, cr, ci, is, bufs + j*bufdist);
Chris@10: 
Chris@10:           /* transform back: */
Chris@10:           cld->apply((plan *) cld, bufs, r0);
Chris@10: 	  r0 += ovs * nbuf; r1 += ovs * nbuf;
Chris@10:      }
Chris@10: 
Chris@10:      X(ifree)(bufs);
Chris@10: 
Chris@10:      /* Do the remaining transforms, if any: */
Chris@10:      cldrest = (plan_rdft2 *) ego->cldrest;
Chris@10:      cldrest->apply((plan *) cldrest, r0, r1, cr, ci);
Chris@10: }
Chris@10: 
Chris@10: static void awake(plan *ego_, enum wakefulness wakefulness)
Chris@10: {
Chris@10:      P *ego = (P *) ego_;
Chris@10: 
Chris@10:      X(plan_awake)(ego->cld, wakefulness);
Chris@10:      X(plan_awake)(ego->cldrest, 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->cldrest);
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, "(rdft2-rdft-%s-%D%v/%D-%D%(%p%)%(%p%))",
Chris@10: 	      ego->super.apply == apply_r2hc ? "r2hc" : "hc2r",
Chris@10:               ego->n, ego->nbuf,
Chris@10:               ego->vl, ego->bufdist % ego->n,
Chris@10:               ego->cld, ego->cldrest);
Chris@10: }
Chris@10: 
Chris@10: static INT min_nbuf(const problem_rdft2 *p, INT n, INT vl)
Chris@10: {
Chris@10:      INT is, os, ivs, ovs;
Chris@10: 
Chris@10:      if (p->r0 != p->cr)
Chris@10: 	  return 1;
Chris@10:      if (X(rdft2_inplace_strides(p, RNK_MINFTY)))
Chris@10: 	  return 1;
Chris@10:      A(p->vecsz->rnk == 1); /*  rank 0 and MINFTY are inplace */
Chris@10: 
Chris@10:      X(rdft2_strides)(p->kind, p->sz->dims, &is, &os);
Chris@10:      X(rdft2_strides)(p->kind, p->vecsz->dims, &ivs, &ovs);
Chris@10:      
Chris@10:      /* handle one potentially common case: "contiguous" real and
Chris@10: 	complex arrays, which overlap because of the differing sizes. */
Chris@10:      if (n * X(iabs)(is) <= X(iabs)(ivs)
Chris@10: 	 && (n/2 + 1) * X(iabs)(os) <= X(iabs)(ovs)
Chris@10: 	 && ( ((p->cr - p->ci) <= X(iabs)(os)) || 
Chris@10: 	      ((p->ci - p->cr) <= X(iabs)(os)) )
Chris@10: 	 && ivs > 0 && ovs > 0) {
Chris@10: 	  INT vsmin = X(imin)(ivs, ovs);
Chris@10: 	  INT vsmax = X(imax)(ivs, ovs);
Chris@10: 	  return(((vsmax - vsmin) * vl + vsmin - 1) / vsmin);
Chris@10:      }
Chris@10: 
Chris@10:      return vl; /* punt: just buffer the whole vector */
Chris@10: }
Chris@10: 
Chris@10: static int applicable0(const problem *p_, const S *ego, const planner *plnr)
Chris@10: {
Chris@10:      const problem_rdft2 *p = (const problem_rdft2 *) p_;
Chris@10:      UNUSED(ego);
Chris@10:      return(1
Chris@10: 	    && p->vecsz->rnk <= 1
Chris@10: 	    && p->sz->rnk == 1
Chris@10: 
Chris@10: 	    /* FIXME: does it make sense to do R2HCII ? */
Chris@10: 	    && (p->kind == R2HC || p->kind == HC2R)
Chris@10: 
Chris@10: 	    /* real strides must allow for reduction to rdft */
Chris@10: 	    && (2 * (p->r1 - p->r0) ==
Chris@10: 		(((p->kind == R2HC) ? p->sz->dims[0].is : p->sz->dims[0].os)))
Chris@10: 
Chris@10: 	    && !(X(toobig)(p->sz->dims[0].n) && CONSERVE_MEMORYP(plnr))
Chris@10: 	  );
Chris@10: }
Chris@10: 
Chris@10: static int applicable(const problem *p_, const S *ego, const planner *plnr)
Chris@10: {
Chris@10:      const problem_rdft2 *p;
Chris@10: 
Chris@10:      if (NO_BUFFERINGP(plnr)) return 0;
Chris@10: 
Chris@10:      if (!applicable0(p_, ego, plnr)) return 0;
Chris@10: 
Chris@10:      p = (const problem_rdft2 *) p_;
Chris@10:      if (NO_UGLYP(plnr)) {
Chris@10: 	  if (p->r0 != p->cr) return 0;
Chris@10: 	  if (X(toobig)(p->sz->dims[0].n)) 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 S *ego = (const S *) ego_;
Chris@10:      P *pln;
Chris@10:      plan *cld = (plan *) 0;
Chris@10:      plan *cldrest = (plan *) 0;
Chris@10:      const problem_rdft2 *p = (const problem_rdft2 *) p_;
Chris@10:      R *bufs = (R *) 0;
Chris@10:      INT nbuf = 0, bufdist, n, vl;
Chris@10:      INT ivs, ovs, rs, id, od;
Chris@10: 
Chris@10:      static const plan_adt padt = {
Chris@10: 	  X(rdft2_solve), awake, print, destroy
Chris@10:      };
Chris@10: 
Chris@10:      if (!applicable(p_, ego, plnr))
Chris@10:           goto nada;
Chris@10: 
Chris@10:      n = p->sz->dims[0].n;
Chris@10:      X(tensor_tornk1)(p->vecsz, &vl, &ivs, &ovs);
Chris@10: 
Chris@10:      nbuf = X(imax)(X(nbuf)(n, vl, 0), min_nbuf(p, n, vl));
Chris@10:      bufdist = X(bufdist)(n, vl);
Chris@10:      A(nbuf > 0);
Chris@10: 
Chris@10:      /* initial allocation for the purpose of planning */
Chris@10:      bufs = (R *) MALLOC(sizeof(R) * nbuf * bufdist, BUFFERS);
Chris@10: 
Chris@10:      id = ivs * (nbuf * (vl / nbuf));
Chris@10:      od = ovs * (nbuf * (vl / nbuf));
Chris@10: 
Chris@10:      if (p->kind == R2HC) {
Chris@10: 	  cld = X(mkplan_f_d)(
Chris@10: 	       plnr,
Chris@10: 	       X(mkproblem_rdft_d)(
Chris@10: 		    X(mktensor_1d)(n, p->sz->dims[0].is/2, 1),
Chris@10: 		    X(mktensor_1d)(nbuf, ivs, bufdist),
Chris@10: 		    TAINT(p->r0, ivs * nbuf), bufs, &p->kind),
Chris@10: 	       0, 0, (p->r0 == p->cr) ? NO_DESTROY_INPUT : 0);
Chris@10: 	  if (!cld) goto nada;
Chris@10: 	  X(ifree)(bufs); bufs = 0;
Chris@10: 
Chris@10: 	  cldrest = X(mkplan_d)(plnr, 
Chris@10: 				X(mkproblem_rdft2_d)(
Chris@10: 				     X(tensor_copy)(p->sz),
Chris@10: 				     X(mktensor_1d)(vl % nbuf, ivs, ovs),
Chris@10: 				     p->r0 + id, p->r1 + id, 
Chris@10: 				     p->cr + od, p->ci + od,
Chris@10: 				     p->kind));
Chris@10: 	  if (!cldrest) goto nada;
Chris@10: 
Chris@10: 	  pln = MKPLAN_RDFT2(P, &padt, apply_r2hc);
Chris@10:      } else {
Chris@10: 	  A(p->kind == HC2R);
Chris@10: 	  cld = X(mkplan_f_d)(
Chris@10: 	       plnr,
Chris@10: 	       X(mkproblem_rdft_d)(
Chris@10: 		    X(mktensor_1d)(n, 1, p->sz->dims[0].os/2),
Chris@10: 		    X(mktensor_1d)(nbuf, bufdist, ovs),
Chris@10: 		    bufs, TAINT(p->r0, ovs * nbuf), &p->kind),
Chris@10: 	       0, 0, NO_DESTROY_INPUT); /* always ok to destroy bufs */
Chris@10: 	  if (!cld) goto nada;
Chris@10: 	  X(ifree)(bufs); bufs = 0;
Chris@10: 
Chris@10: 	  cldrest = X(mkplan_d)(plnr, 
Chris@10: 				X(mkproblem_rdft2_d)(
Chris@10: 				     X(tensor_copy)(p->sz),
Chris@10: 				     X(mktensor_1d)(vl % nbuf, ivs, ovs),
Chris@10: 				     p->r0 + od, p->r1 + od, 
Chris@10: 				     p->cr + id, p->ci + id,
Chris@10: 				     p->kind));
Chris@10: 	  if (!cldrest) goto nada;
Chris@10: 	  pln = MKPLAN_RDFT2(P, &padt, apply_hc2r);
Chris@10:      }
Chris@10: 
Chris@10:      pln->cld = cld;
Chris@10:      pln->cldrest = cldrest;
Chris@10:      pln->n = n;
Chris@10:      pln->vl = vl;
Chris@10:      pln->ivs = ivs;
Chris@10:      pln->ovs = ovs;
Chris@10:      X(rdft2_strides)(p->kind, &p->sz->dims[0], &rs, &pln->cs);
Chris@10:      pln->nbuf = nbuf;
Chris@10:      pln->bufdist = bufdist;
Chris@10: 
Chris@10:      X(ops_madd)(vl / nbuf, &cld->ops, &cldrest->ops,
Chris@10: 		 &pln->super.super.ops);
Chris@10:      pln->super.super.ops.other += (p->kind == R2HC ? (n + 2) : n) * vl;
Chris@10: 
Chris@10:      return &(pln->super.super);
Chris@10: 
Chris@10:  nada:
Chris@10:      X(ifree0)(bufs);
Chris@10:      X(plan_destroy_internal)(cldrest);
Chris@10:      X(plan_destroy_internal)(cld);
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_RDFT2, mkplan, 0 };
Chris@10:      S *slv = MKSOLVER(S, &sadt);
Chris@10:      return &(slv->super);
Chris@10: }
Chris@10: 
Chris@10: void X(rdft2_rdft_register)(planner *p)
Chris@10: {
Chris@10:      REGISTER_SOLVER(p, mksolver());
Chris@10: }