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: /* buffering of rdft2.  We always buffer the complex array */
Chris@10: 
Chris@10: #include "rdft.h"
Chris@10: #include "dft.h"
Chris@10: 
Chris@10: typedef struct {
Chris@10:      solver super;
Chris@10:      int maxnbuf_ndx;
Chris@10: } S;
Chris@10: 
Chris@10: static const INT maxnbufs[] = { 8, 256 };
Chris@10: 
Chris@10: typedef struct {
Chris@10:      plan_rdft2 super;
Chris@10: 
Chris@10:      plan *cld, *cldcpy, *cldrest;
Chris@10:      INT n, vl, nbuf, bufdist;
Chris@10:      INT ivs_by_nbuf, ovs_by_nbuf;
Chris@10:      INT ioffset, roffset;
Chris@10: } P;
Chris@10: 
Chris@10: /* transform a vector input with the help of bufs */
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_rdft2 *cld = (plan_rdft2 *) ego->cld;
Chris@10:      plan_dft *cldcpy = (plan_dft *) ego->cldcpy;
Chris@10:      INT i, vl = ego->vl, nbuf = ego->nbuf;
Chris@10:      INT ivs_by_nbuf = ego->ivs_by_nbuf, ovs_by_nbuf = ego->ovs_by_nbuf;
Chris@10:      R *bufs = (R *)MALLOC(sizeof(R) * nbuf * ego->bufdist, BUFFERS);
Chris@10:      R *bufr = bufs + ego->roffset;
Chris@10:      R *bufi = bufs + ego->ioffset;
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, r1, bufr, bufi);
Chris@10: 	  r0 += ivs_by_nbuf; r1 += ivs_by_nbuf;
Chris@10: 
Chris@10:           /* copy back */
Chris@10:           cldcpy->apply((plan *) cldcpy, bufr, bufi, cr, ci);
Chris@10: 	  cr += ovs_by_nbuf; ci += ovs_by_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: /* for hc2r problems, copy the input into buffer, and then
Chris@10:    transform buffer->output, which allows for destruction of the
Chris@10:    buffer */
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_rdft2 *cld = (plan_rdft2 *) ego->cld;
Chris@10:      plan_dft *cldcpy = (plan_dft *) ego->cldcpy;
Chris@10:      INT i, vl = ego->vl, nbuf = ego->nbuf;
Chris@10:      INT ivs_by_nbuf = ego->ivs_by_nbuf, ovs_by_nbuf = ego->ovs_by_nbuf;
Chris@10:      R *bufs = (R *)MALLOC(sizeof(R) * nbuf * ego->bufdist, BUFFERS);
Chris@10:      R *bufr = bufs + ego->roffset;
Chris@10:      R *bufi = bufs + ego->ioffset;
Chris@10:      plan_rdft2 *cldrest;
Chris@10: 
Chris@10:      for (i = nbuf; i <= vl; i += nbuf) {
Chris@10:           /* copy input into bufs: */
Chris@10:           cldcpy->apply((plan *) cldcpy, cr, ci, bufr, bufi);
Chris@10: 	  cr += ivs_by_nbuf; ci += ivs_by_nbuf;
Chris@10: 
Chris@10:           /* transform to output */
Chris@10:           cld->apply((plan *) cld, r0, r1, bufr, bufi);
Chris@10: 	  r0 += ovs_by_nbuf; r1 += ovs_by_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: 
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->cldcpy, 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->cldcpy);
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-buffered-%D%v/%D-%D%(%p%)%(%p%)%(%p%))",
Chris@10:               ego->n, ego->nbuf,
Chris@10:               ego->vl, ego->bufdist % ego->n,
Chris@10:               ego->cld, ego->cldcpy, ego->cldrest);
Chris@10: }
Chris@10: 
Chris@10: static int applicable0(const S *ego, const problem *p_, const planner *plnr)
Chris@10: {
Chris@10:      const problem_rdft2 *p = (const problem_rdft2 *) p_;
Chris@10:      iodim *d = p->sz->dims;
Chris@10: 
Chris@10:      if (1
Chris@10: 	 && p->vecsz->rnk <= 1
Chris@10: 	 && p->sz->rnk == 1
Chris@10: 
Chris@10: 	 /* we assume even n throughout */
Chris@10: 	 && (d[0].n % 2) == 0
Chris@10: 
Chris@10: 	 /* and we only consider these two cases */
Chris@10: 	 && (p->kind == R2HC || p->kind == HC2R)
Chris@10: 
Chris@10: 	  ) {
Chris@10: 	  INT vl, ivs, ovs;
Chris@10: 	  X(tensor_tornk1)(p->vecsz, &vl, &ivs, &ovs);
Chris@10: 
Chris@10: 	  if (X(toobig)(d[0].n) && CONSERVE_MEMORYP(plnr))
Chris@10: 	       return 0;
Chris@10: 
Chris@10: 	  /* if this solver is redundant, in the sense that a solver
Chris@10: 	     of lower index generates the same plan, then prune this
Chris@10: 	     solver */
Chris@10: 	  if (X(nbuf_redundant)(d[0].n, vl,
Chris@10: 				ego->maxnbuf_ndx,
Chris@10: 				maxnbufs, NELEM(maxnbufs)))
Chris@10: 	       return 0;
Chris@10: 
Chris@10: 	  if (p->r0 != p->cr) {
Chris@10: 	       if (p->kind == HC2R) {
Chris@10: 		    /* Allow HC2R problems only if the input is to be
Chris@10: 		       preserved.  This solver sets NO_DESTROY_INPUT,
Chris@10: 		       which prevents infinite loops */
Chris@10: 		    return (NO_DESTROY_INPUTP(plnr));
Chris@10: 	       } else {
Chris@10: 		    /*
Chris@10: 		      In principle, the buffered transforms might be useful
Chris@10: 		      when working out of place.  However, in order to
Chris@10: 		      prevent infinite loops in the planner, we require
Chris@10: 		      that the output stride of the buffered transforms be
Chris@10: 		      greater than 2.
Chris@10: 		    */
Chris@10: 		    return (d[0].os > 2);
Chris@10: 	       }
Chris@10: 	  }
Chris@10: 
Chris@10: 	  /*
Chris@10: 	   * If the problem is in place, the input/output strides must
Chris@10: 	   * be the same or the whole thing must fit in the buffer.
Chris@10: 	   */
Chris@10: 	  if (X(rdft2_inplace_strides(p, RNK_MINFTY)))
Chris@10: 	       return 1;
Chris@10: 
Chris@10: 	  if (/* fits into buffer: */
Chris@10: 	       ((p->vecsz->rnk == 0)
Chris@10: 		||
Chris@10: 		(X(nbuf)(d[0].n, p->vecsz->dims[0].n,
Chris@10: 			 maxnbufs[ego->maxnbuf_ndx])
Chris@10: 		 == p->vecsz->dims[0].n)))
Chris@10: 	       return 1;
Chris@10:      }
Chris@10: 
Chris@10:      return 0;
Chris@10: }
Chris@10: 
Chris@10: static int applicable(const S *ego, const problem *p_, 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(ego, p_, plnr)) return 0;
Chris@10: 
Chris@10:      p = (const problem_rdft2 *) p_;
Chris@10:      if (p->kind == HC2R) {
Chris@10: 	  if (NO_UGLYP(plnr)) {
Chris@10: 	       /* UGLY if in-place and too big, since the problem
Chris@10: 		  could be solved via transpositions */
Chris@10: 	       if (p->r0 == p->cr && X(toobig)(p->sz->dims[0].n)) 
Chris@10: 		    return 0;
Chris@10: 	  }
Chris@10:      } else {
Chris@10: 	  if (NO_UGLYP(plnr)) {
Chris@10: 	       if (p->r0 != p->cr || X(toobig)(p->sz->dims[0].n))
Chris@10: 		    return 0;
Chris@10: 	  }
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:      P *pln;
Chris@10:      const S *ego = (const S *)ego_;
Chris@10:      plan *cld = (plan *) 0;
Chris@10:      plan *cldcpy = (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, ioffset, roffset, 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(ego, p_, plnr))
Chris@10:           goto nada;
Chris@10: 
Chris@10:      n = X(tensor_sz)(p->sz);
Chris@10:      X(tensor_tornk1)(p->vecsz, &vl, &ivs, &ovs);
Chris@10: 
Chris@10:      nbuf = X(nbuf)(n, vl, maxnbufs[ego->maxnbuf_ndx]);
Chris@10:      bufdist = X(bufdist)(n + 2, vl); /* complex-side rdft2 stores N+2
Chris@10: 					 real numbers */
Chris@10:      A(nbuf > 0);
Chris@10: 
Chris@10:      /* attempt to keep real and imaginary part in the same order,
Chris@10: 	so as to allow optimizations in the the copy plan */
Chris@10:      roffset = (p->cr - p->ci > 0) ? (INT)1 : (INT)0;
Chris@10:      ioffset = 1 - roffset;
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: 	  /* allow destruction of input if problem is in place */
Chris@10: 	  cld = X(mkplan_f_d)(
Chris@10: 	       plnr, 
Chris@10: 	       X(mkproblem_rdft2_d)(
Chris@10: 		    X(mktensor_1d)(n, p->sz->dims[0].is, 2),
Chris@10: 		    X(mktensor_1d)(nbuf, ivs, bufdist),
Chris@10: 		    TAINT(p->r0, ivs * nbuf), TAINT(p->r1, ivs * nbuf),
Chris@10: 		    bufs + roffset, bufs + ioffset, p->kind),
Chris@10: 	       0, 0, (p->r0 == p->cr) ? NO_DESTROY_INPUT : 0);
Chris@10: 	  if (!cld) goto nada;
Chris@10: 
Chris@10: 	  /* copying back from the buffer is a rank-0 DFT: */
Chris@10: 	  cldcpy = X(mkplan_d)(
Chris@10: 	       plnr, 
Chris@10: 	       X(mkproblem_dft_d)(
Chris@10: 		    X(mktensor_0d)(),
Chris@10: 		    X(mktensor_2d)(nbuf, bufdist, ovs,
Chris@10: 				   n/2+1, 2, p->sz->dims[0].os),
Chris@10: 		    bufs + roffset, bufs + ioffset,
Chris@10: 		    TAINT(p->cr, ovs * nbuf), TAINT(p->ci, ovs * nbuf) ));
Chris@10: 	  if (!cldcpy) goto nada;
Chris@10: 
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: 	  pln = MKPLAN_RDFT2(P, &padt, apply_r2hc);
Chris@10:      } else {
Chris@10: 	  /* allow destruction of buffer */
Chris@10: 	  cld = X(mkplan_f_d)(
Chris@10: 	       plnr, 
Chris@10: 	       X(mkproblem_rdft2_d)(
Chris@10: 		    X(mktensor_1d)(n, 2, p->sz->dims[0].os),
Chris@10: 		    X(mktensor_1d)(nbuf, bufdist, ovs),
Chris@10: 		    TAINT(p->r0, ovs * nbuf), TAINT(p->r1, ovs * nbuf),
Chris@10: 		    bufs + roffset, bufs + ioffset, p->kind),
Chris@10: 	       0, 0, NO_DESTROY_INPUT);
Chris@10: 	  if (!cld) goto nada;
Chris@10: 
Chris@10: 	  /* copying input into buffer is a rank-0 DFT: */
Chris@10: 	  cldcpy = X(mkplan_d)(
Chris@10: 	       plnr, 
Chris@10: 	       X(mkproblem_dft_d)(
Chris@10: 		    X(mktensor_0d)(),
Chris@10: 		    X(mktensor_2d)(nbuf, ivs, bufdist,
Chris@10: 				   n/2+1, p->sz->dims[0].is, 2),
Chris@10: 		    TAINT(p->cr, ivs * nbuf), TAINT(p->ci, ivs * nbuf), 
Chris@10: 		    bufs + roffset, bufs + ioffset));
Chris@10: 	  if (!cldcpy) goto nada;
Chris@10: 
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: 
Chris@10: 	  pln = MKPLAN_RDFT2(P, &padt, apply_hc2r);
Chris@10:      }
Chris@10: 
Chris@10:      pln->cld = cld;
Chris@10:      pln->cldcpy = cldcpy;
Chris@10:      pln->cldrest = cldrest;
Chris@10:      pln->n = n;
Chris@10:      pln->vl = vl;
Chris@10:      pln->ivs_by_nbuf = ivs * nbuf;
Chris@10:      pln->ovs_by_nbuf = ovs * nbuf;
Chris@10:      pln->roffset = roffset;
Chris@10:      pln->ioffset = ioffset;
Chris@10: 
Chris@10:      pln->nbuf = nbuf;
Chris@10:      pln->bufdist = bufdist;
Chris@10: 
Chris@10:      {
Chris@10: 	  opcnt t;
Chris@10: 	  X(ops_add)(&cld->ops, &cldcpy->ops, &t);
Chris@10: 	  X(ops_madd)(vl / nbuf, &t, &cldrest->ops, &pln->super.super.ops);
Chris@10:      }
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)(cldcpy);
Chris@10:      X(plan_destroy_internal)(cld);
Chris@10:      return (plan *) 0;
Chris@10: }
Chris@10: 
Chris@10: static solver *mksolver(int maxnbuf_ndx)
Chris@10: {
Chris@10:      static const solver_adt sadt = { PROBLEM_RDFT2, mkplan, 0 };
Chris@10:      S *slv = MKSOLVER(S, &sadt);
Chris@10:      slv->maxnbuf_ndx = maxnbuf_ndx;
Chris@10:      return &(slv->super);
Chris@10: }
Chris@10: 
Chris@10: void X(rdft2_buffered_register)(planner *p)
Chris@10: {
Chris@10:      size_t i;
Chris@10:      for (i = 0; i < NELEM(maxnbufs); ++i)
Chris@10: 	  REGISTER_SOLVER(p, mksolver(i));
Chris@10: }