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: #include "rdft.h"
cannam@95: 
cannam@95: typedef struct {
cannam@95:      solver super;
cannam@95: } S;
cannam@95: 
cannam@95: typedef struct {
cannam@95:      plan_rdft2 super;
cannam@95: 
cannam@95:      plan *cld, *cldrest;
cannam@95:      INT n, vl, nbuf, bufdist;
cannam@95:      INT cs, ivs, ovs;
cannam@95: } P;
cannam@95: 
cannam@95: /***************************************************************************/
cannam@95: 
cannam@95: /* FIXME: have alternate copy functions that push a vector loop inside
cannam@95:    the n loops? */
cannam@95: 
cannam@95: /* copy halfcomplex array r (contiguous) to complex (strided) array rio/iio. */
cannam@95: static void hc2c(INT n, R *r, R *rio, R *iio, INT os)
cannam@95: {
cannam@95:      INT i;
cannam@95: 
cannam@95:      rio[0] = r[0];
cannam@95:      iio[0] = 0;
cannam@95: 
cannam@95:      for (i = 1; i + i < n; ++i) {
cannam@95: 	  rio[i * os] = r[i];
cannam@95: 	  iio[i * os] = r[n - i];
cannam@95:      }
cannam@95: 
cannam@95:      if (i + i == n) {	/* store the Nyquist frequency */
cannam@95: 	  rio[i * os] = r[i];
cannam@95: 	  iio[i * os] = K(0.0);
cannam@95:      }
cannam@95: }
cannam@95: 
cannam@95: /* reverse of hc2c */
cannam@95: static void c2hc(INT n, R *rio, R *iio, INT is, R *r)
cannam@95: {
cannam@95:      INT i;
cannam@95: 
cannam@95:      r[0] = rio[0];
cannam@95: 
cannam@95:      for (i = 1; i + i < n; ++i) {
cannam@95: 	  r[i] = rio[i * is];
cannam@95: 	  r[n - i] = iio[i * is];
cannam@95:      }
cannam@95: 
cannam@95:      if (i + i == n)		/* store the Nyquist frequency */
cannam@95: 	  r[i] = rio[i * is];
cannam@95: }
cannam@95: 
cannam@95: /***************************************************************************/
cannam@95: 
cannam@95: static void apply_r2hc(const plan *ego_, R *r0, R *r1, R *cr, R *ci)
cannam@95: {
cannam@95:      const P *ego = (const P *) ego_;
cannam@95:      plan_rdft *cld = (plan_rdft *) ego->cld;
cannam@95:      INT i, j, vl = ego->vl, nbuf = ego->nbuf, bufdist = ego->bufdist;
cannam@95:      INT n = ego->n;
cannam@95:      INT ivs = ego->ivs, ovs = ego->ovs, os = ego->cs;
cannam@95:      R *bufs = (R *)MALLOC(sizeof(R) * nbuf * bufdist, BUFFERS);
cannam@95:      plan_rdft2 *cldrest;
cannam@95: 
cannam@95:      for (i = nbuf; i <= vl; i += nbuf) {
cannam@95:           /* transform to bufs: */
cannam@95:           cld->apply((plan *) cld, r0, bufs);
cannam@95: 	  r0 += ivs * nbuf; r1 += ivs * nbuf;
cannam@95: 
cannam@95:           /* copy back */
cannam@95: 	  for (j = 0; j < nbuf; ++j, cr += ovs, ci += ovs)
cannam@95: 	       hc2c(n, bufs + j*bufdist, cr, ci, os);
cannam@95:      }
cannam@95: 
cannam@95:      X(ifree)(bufs);
cannam@95: 
cannam@95:      /* Do the remaining transforms, if any: */
cannam@95:      cldrest = (plan_rdft2 *) ego->cldrest;
cannam@95:      cldrest->apply((plan *) cldrest, r0, r1, cr, ci);
cannam@95: }
cannam@95: 
cannam@95: static void apply_hc2r(const plan *ego_, R *r0, R *r1, R *cr, R *ci)
cannam@95: {
cannam@95:      const P *ego = (const P *) ego_;
cannam@95:      plan_rdft *cld = (plan_rdft *) ego->cld;
cannam@95:      INT i, j, vl = ego->vl, nbuf = ego->nbuf, bufdist = ego->bufdist;
cannam@95:      INT n = ego->n;
cannam@95:      INT ivs = ego->ivs, ovs = ego->ovs, is = ego->cs;
cannam@95:      R *bufs = (R *)MALLOC(sizeof(R) * nbuf * bufdist, BUFFERS);
cannam@95:      plan_rdft2 *cldrest;
cannam@95: 
cannam@95:      for (i = nbuf; i <= vl; i += nbuf) {
cannam@95:           /* copy to bufs */
cannam@95: 	  for (j = 0; j < nbuf; ++j, cr += ivs, ci += ivs)
cannam@95: 	       c2hc(n, cr, ci, is, bufs + j*bufdist);
cannam@95: 
cannam@95:           /* transform back: */
cannam@95:           cld->apply((plan *) cld, bufs, r0);
cannam@95: 	  r0 += ovs * nbuf; r1 += ovs * nbuf;
cannam@95:      }
cannam@95: 
cannam@95:      X(ifree)(bufs);
cannam@95: 
cannam@95:      /* Do the remaining transforms, if any: */
cannam@95:      cldrest = (plan_rdft2 *) ego->cldrest;
cannam@95:      cldrest->apply((plan *) cldrest, r0, r1, cr, ci);
cannam@95: }
cannam@95: 
cannam@95: static void awake(plan *ego_, enum wakefulness wakefulness)
cannam@95: {
cannam@95:      P *ego = (P *) ego_;
cannam@95: 
cannam@95:      X(plan_awake)(ego->cld, wakefulness);
cannam@95:      X(plan_awake)(ego->cldrest, 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->cldrest);
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, "(rdft2-rdft-%s-%D%v/%D-%D%(%p%)%(%p%))",
cannam@95: 	      ego->super.apply == apply_r2hc ? "r2hc" : "hc2r",
cannam@95:               ego->n, ego->nbuf,
cannam@95:               ego->vl, ego->bufdist % ego->n,
cannam@95:               ego->cld, ego->cldrest);
cannam@95: }
cannam@95: 
cannam@95: static INT min_nbuf(const problem_rdft2 *p, INT n, INT vl)
cannam@95: {
cannam@95:      INT is, os, ivs, ovs;
cannam@95: 
cannam@95:      if (p->r0 != p->cr)
cannam@95: 	  return 1;
cannam@95:      if (X(rdft2_inplace_strides(p, RNK_MINFTY)))
cannam@95: 	  return 1;
cannam@95:      A(p->vecsz->rnk == 1); /*  rank 0 and MINFTY are inplace */
cannam@95: 
cannam@95:      X(rdft2_strides)(p->kind, p->sz->dims, &is, &os);
cannam@95:      X(rdft2_strides)(p->kind, p->vecsz->dims, &ivs, &ovs);
cannam@95:      
cannam@95:      /* handle one potentially common case: "contiguous" real and
cannam@95: 	complex arrays, which overlap because of the differing sizes. */
cannam@95:      if (n * X(iabs)(is) <= X(iabs)(ivs)
cannam@95: 	 && (n/2 + 1) * X(iabs)(os) <= X(iabs)(ovs)
cannam@95: 	 && ( ((p->cr - p->ci) <= X(iabs)(os)) || 
cannam@95: 	      ((p->ci - p->cr) <= X(iabs)(os)) )
cannam@95: 	 && ivs > 0 && ovs > 0) {
cannam@95: 	  INT vsmin = X(imin)(ivs, ovs);
cannam@95: 	  INT vsmax = X(imax)(ivs, ovs);
cannam@95: 	  return(((vsmax - vsmin) * vl + vsmin - 1) / vsmin);
cannam@95:      }
cannam@95: 
cannam@95:      return vl; /* punt: just buffer the whole vector */
cannam@95: }
cannam@95: 
cannam@95: static int applicable0(const problem *p_, const S *ego, const planner *plnr)
cannam@95: {
cannam@95:      const problem_rdft2 *p = (const problem_rdft2 *) p_;
cannam@95:      UNUSED(ego);
cannam@95:      return(1
cannam@95: 	    && p->vecsz->rnk <= 1
cannam@95: 	    && p->sz->rnk == 1
cannam@95: 
cannam@95: 	    /* FIXME: does it make sense to do R2HCII ? */
cannam@95: 	    && (p->kind == R2HC || p->kind == HC2R)
cannam@95: 
cannam@95: 	    /* real strides must allow for reduction to rdft */
cannam@95: 	    && (2 * (p->r1 - p->r0) ==
cannam@95: 		(((p->kind == R2HC) ? p->sz->dims[0].is : p->sz->dims[0].os)))
cannam@95: 
cannam@95: 	    && !(X(toobig)(p->sz->dims[0].n) && CONSERVE_MEMORYP(plnr))
cannam@95: 	  );
cannam@95: }
cannam@95: 
cannam@95: static int applicable(const problem *p_, const S *ego, const planner *plnr)
cannam@95: {
cannam@95:      const problem_rdft2 *p;
cannam@95: 
cannam@95:      if (NO_BUFFERINGP(plnr)) return 0;
cannam@95: 
cannam@95:      if (!applicable0(p_, ego, plnr)) return 0;
cannam@95: 
cannam@95:      p = (const problem_rdft2 *) p_;
cannam@95:      if (NO_UGLYP(plnr)) {
cannam@95: 	  if (p->r0 != p->cr) return 0;
cannam@95: 	  if (X(toobig)(p->sz->dims[0].n)) 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 S *ego = (const S *) ego_;
cannam@95:      P *pln;
cannam@95:      plan *cld = (plan *) 0;
cannam@95:      plan *cldrest = (plan *) 0;
cannam@95:      const problem_rdft2 *p = (const problem_rdft2 *) p_;
cannam@95:      R *bufs = (R *) 0;
cannam@95:      INT nbuf = 0, bufdist, n, vl;
cannam@95:      INT ivs, ovs, rs, id, od;
cannam@95: 
cannam@95:      static const plan_adt padt = {
cannam@95: 	  X(rdft2_solve), awake, print, destroy
cannam@95:      };
cannam@95: 
cannam@95:      if (!applicable(p_, ego, plnr))
cannam@95:           goto nada;
cannam@95: 
cannam@95:      n = p->sz->dims[0].n;
cannam@95:      X(tensor_tornk1)(p->vecsz, &vl, &ivs, &ovs);
cannam@95: 
cannam@95:      nbuf = X(imax)(X(nbuf)(n, vl, 0), min_nbuf(p, n, vl));
cannam@95:      bufdist = X(bufdist)(n, vl);
cannam@95:      A(nbuf > 0);
cannam@95: 
cannam@95:      /* initial allocation for the purpose of planning */
cannam@95:      bufs = (R *) MALLOC(sizeof(R) * nbuf * bufdist, BUFFERS);
cannam@95: 
cannam@95:      id = ivs * (nbuf * (vl / nbuf));
cannam@95:      od = ovs * (nbuf * (vl / nbuf));
cannam@95: 
cannam@95:      if (p->kind == R2HC) {
cannam@95: 	  cld = X(mkplan_f_d)(
cannam@95: 	       plnr,
cannam@95: 	       X(mkproblem_rdft_d)(
cannam@95: 		    X(mktensor_1d)(n, p->sz->dims[0].is/2, 1),
cannam@95: 		    X(mktensor_1d)(nbuf, ivs, bufdist),
cannam@95: 		    TAINT(p->r0, ivs * nbuf), bufs, &p->kind),
cannam@95: 	       0, 0, (p->r0 == p->cr) ? NO_DESTROY_INPUT : 0);
cannam@95: 	  if (!cld) goto nada;
cannam@95: 	  X(ifree)(bufs); bufs = 0;
cannam@95: 
cannam@95: 	  cldrest = X(mkplan_d)(plnr, 
cannam@95: 				X(mkproblem_rdft2_d)(
cannam@95: 				     X(tensor_copy)(p->sz),
cannam@95: 				     X(mktensor_1d)(vl % nbuf, ivs, ovs),
cannam@95: 				     p->r0 + id, p->r1 + id, 
cannam@95: 				     p->cr + od, p->ci + od,
cannam@95: 				     p->kind));
cannam@95: 	  if (!cldrest) goto nada;
cannam@95: 
cannam@95: 	  pln = MKPLAN_RDFT2(P, &padt, apply_r2hc);
cannam@95:      } else {
cannam@95: 	  A(p->kind == HC2R);
cannam@95: 	  cld = X(mkplan_f_d)(
cannam@95: 	       plnr,
cannam@95: 	       X(mkproblem_rdft_d)(
cannam@95: 		    X(mktensor_1d)(n, 1, p->sz->dims[0].os/2),
cannam@95: 		    X(mktensor_1d)(nbuf, bufdist, ovs),
cannam@95: 		    bufs, TAINT(p->r0, ovs * nbuf), &p->kind),
cannam@95: 	       0, 0, NO_DESTROY_INPUT); /* always ok to destroy bufs */
cannam@95: 	  if (!cld) goto nada;
cannam@95: 	  X(ifree)(bufs); bufs = 0;
cannam@95: 
cannam@95: 	  cldrest = X(mkplan_d)(plnr, 
cannam@95: 				X(mkproblem_rdft2_d)(
cannam@95: 				     X(tensor_copy)(p->sz),
cannam@95: 				     X(mktensor_1d)(vl % nbuf, ivs, ovs),
cannam@95: 				     p->r0 + od, p->r1 + od, 
cannam@95: 				     p->cr + id, p->ci + id,
cannam@95: 				     p->kind));
cannam@95: 	  if (!cldrest) goto nada;
cannam@95: 	  pln = MKPLAN_RDFT2(P, &padt, apply_hc2r);
cannam@95:      }
cannam@95: 
cannam@95:      pln->cld = cld;
cannam@95:      pln->cldrest = cldrest;
cannam@95:      pln->n = n;
cannam@95:      pln->vl = vl;
cannam@95:      pln->ivs = ivs;
cannam@95:      pln->ovs = ovs;
cannam@95:      X(rdft2_strides)(p->kind, &p->sz->dims[0], &rs, &pln->cs);
cannam@95:      pln->nbuf = nbuf;
cannam@95:      pln->bufdist = bufdist;
cannam@95: 
cannam@95:      X(ops_madd)(vl / nbuf, &cld->ops, &cldrest->ops,
cannam@95: 		 &pln->super.super.ops);
cannam@95:      pln->super.super.ops.other += (p->kind == R2HC ? (n + 2) : n) * vl;
cannam@95: 
cannam@95:      return &(pln->super.super);
cannam@95: 
cannam@95:  nada:
cannam@95:      X(ifree0)(bufs);
cannam@95:      X(plan_destroy_internal)(cldrest);
cannam@95:      X(plan_destroy_internal)(cld);
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_RDFT2, mkplan, 0 };
cannam@95:      S *slv = MKSOLVER(S, &sadt);
cannam@95:      return &(slv->super);
cannam@95: }
cannam@95: 
cannam@95: void X(rdft2_rdft_register)(planner *p)
cannam@95: {
cannam@95:      REGISTER_SOLVER(p, mksolver());
cannam@95: }