cannam@127: /*
cannam@127:  * Copyright (c) 2003, 2007-14 Matteo Frigo
cannam@127:  * Copyright (c) 2003, 2007-14 Massachusetts Institute of Technology
cannam@127:  *
cannam@127:  * This program is free software; you can redistribute it and/or modify
cannam@127:  * it under the terms of the GNU General Public License as published by
cannam@127:  * the Free Software Foundation; either version 2 of the License, or
cannam@127:  * (at your option) any later version.
cannam@127:  *
cannam@127:  * This program is distributed in the hope that it will be useful,
cannam@127:  * but WITHOUT ANY WARRANTY; without even the implied warranty of
cannam@127:  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
cannam@127:  * GNU General Public License for more details.
cannam@127:  *
cannam@127:  * You should have received a copy of the GNU General Public License
cannam@127:  * along with this program; if not, write to the Free Software
cannam@127:  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA
cannam@127:  *
cannam@127:  */
cannam@127: 
cannam@127: 
cannam@127: #include "dft.h"
cannam@127: 
cannam@127: typedef struct {
cannam@127:      solver super;
cannam@127:      size_t maxnbuf_ndx;
cannam@127: } S;
cannam@127: 
cannam@127: static const INT maxnbufs[] = { 8, 256 };
cannam@127: 
cannam@127: typedef struct {
cannam@127:      plan_dft super;
cannam@127: 
cannam@127:      plan *cld, *cldcpy, *cldrest;
cannam@127:      INT n, vl, nbuf, bufdist;
cannam@127:      INT ivs_by_nbuf, ovs_by_nbuf;
cannam@127:      INT roffset, ioffset;
cannam@127: } P;
cannam@127: 
cannam@127: /* transform a vector input with the help of bufs */
cannam@127: static void apply(const plan *ego_, R *ri, R *ii, R *ro, R *io)
cannam@127: {
cannam@127:      const P *ego = (const P *) ego_;
cannam@127:      INT nbuf = ego->nbuf;
cannam@127:      R *bufs = (R *)MALLOC(sizeof(R) * nbuf * ego->bufdist * 2, BUFFERS);
cannam@127: 
cannam@127:      plan_dft *cld = (plan_dft *) ego->cld;
cannam@127:      plan_dft *cldcpy = (plan_dft *) ego->cldcpy;
cannam@127:      plan_dft *cldrest;
cannam@127:      INT i, vl = ego->vl;
cannam@127:      INT ivs_by_nbuf = ego->ivs_by_nbuf, ovs_by_nbuf = ego->ovs_by_nbuf;
cannam@127:      INT roffset = ego->roffset, ioffset = ego->ioffset;
cannam@127: 
cannam@127:      for (i = nbuf; i <= vl; i += nbuf) {
cannam@127:           /* transform to bufs: */
cannam@127:           cld->apply((plan *) cld, ri, ii, bufs + roffset, bufs + ioffset);
cannam@127: 	  ri += ivs_by_nbuf; ii += ivs_by_nbuf;
cannam@127: 
cannam@127:           /* copy back */
cannam@127:           cldcpy->apply((plan *) cldcpy, bufs+roffset, bufs+ioffset, ro, io);
cannam@127: 	  ro += ovs_by_nbuf; io += ovs_by_nbuf;
cannam@127:      }
cannam@127: 
cannam@127:      X(ifree)(bufs);
cannam@127: 
cannam@127:      /* Do the remaining transforms, if any: */
cannam@127:      cldrest = (plan_dft *) ego->cldrest;
cannam@127:      cldrest->apply((plan *) cldrest, ri, ii, ro, io);
cannam@127: }
cannam@127: 
cannam@127: 
cannam@127: static void awake(plan *ego_, enum wakefulness wakefulness)
cannam@127: {
cannam@127:      P *ego = (P *) ego_;
cannam@127: 
cannam@127:      X(plan_awake)(ego->cld, wakefulness);
cannam@127:      X(plan_awake)(ego->cldcpy, wakefulness);
cannam@127:      X(plan_awake)(ego->cldrest, wakefulness);
cannam@127: }
cannam@127: 
cannam@127: static void destroy(plan *ego_)
cannam@127: {
cannam@127:      P *ego = (P *) ego_;
cannam@127:      X(plan_destroy_internal)(ego->cldrest);
cannam@127:      X(plan_destroy_internal)(ego->cldcpy);
cannam@127:      X(plan_destroy_internal)(ego->cld);
cannam@127: }
cannam@127: 
cannam@127: static void print(const plan *ego_, printer *p)
cannam@127: {
cannam@127:      const P *ego = (const P *) ego_;
cannam@127:      p->print(p, "(dft-buffered-%D%v/%D-%D%(%p%)%(%p%)%(%p%))",
cannam@127:               ego->n, ego->nbuf,
cannam@127:               ego->vl, ego->bufdist % ego->n,
cannam@127:               ego->cld, ego->cldcpy, ego->cldrest);
cannam@127: }
cannam@127: 
cannam@127: static int applicable0(const S *ego, const problem *p_, const planner *plnr)
cannam@127: {
cannam@127:      const problem_dft *p = (const problem_dft *) p_;
cannam@127:      const iodim *d = p->sz->dims;
cannam@127: 
cannam@127:      if (1
cannam@127: 	 && p->vecsz->rnk <= 1
cannam@127: 	 && p->sz->rnk == 1
cannam@127: 	  ) {
cannam@127: 	  INT vl, ivs, ovs;
cannam@127: 	  X(tensor_tornk1)(p->vecsz, &vl, &ivs, &ovs);
cannam@127: 
cannam@127: 	  if (X(toobig)(p->sz->dims[0].n) && CONSERVE_MEMORYP(plnr))
cannam@127: 	       return 0;
cannam@127: 
cannam@127: 	  /* if this solver is redundant, in the sense that a solver
cannam@127: 	     of lower index generates the same plan, then prune this
cannam@127: 	     solver */
cannam@127: 	  if (X(nbuf_redundant)(d[0].n, vl, 
cannam@127: 				ego->maxnbuf_ndx,
cannam@127: 				maxnbufs, NELEM(maxnbufs)))
cannam@127: 	       return 0;
cannam@127: 
cannam@127: 	  /*
cannam@127: 	    In principle, the buffered transforms might be useful
cannam@127: 	    when working out of place.  However, in order to
cannam@127: 	    prevent infinite loops in the planner, we require
cannam@127: 	    that the output stride of the buffered transforms be
cannam@127: 	    greater than 2.
cannam@127: 	  */
cannam@127: 	  if (p->ri != p->ro)
cannam@127: 	       return (d[0].os > 2);
cannam@127: 
cannam@127: 	  /*
cannam@127: 	   * If the problem is in place, the input/output strides must
cannam@127: 	   * be the same or the whole thing must fit in the buffer.
cannam@127: 	   */
cannam@127: 	  if (X(tensor_inplace_strides2)(p->sz, p->vecsz))
cannam@127: 	       return 1;
cannam@127: 
cannam@127: 	  if (/* fits into buffer: */
cannam@127: 	       ((p->vecsz->rnk == 0)
cannam@127: 		||
cannam@127: 		(X(nbuf)(d[0].n, p->vecsz->dims[0].n, 
cannam@127: 			 maxnbufs[ego->maxnbuf_ndx]) 
cannam@127: 		 == p->vecsz->dims[0].n)))
cannam@127: 	       return 1;
cannam@127:      }
cannam@127: 
cannam@127:      return 0;
cannam@127: }
cannam@127: 
cannam@127: static int applicable(const S *ego, const problem *p_, const planner *plnr)
cannam@127: {
cannam@127:      if (NO_BUFFERINGP(plnr)) return 0;
cannam@127:      if (!applicable0(ego, p_, plnr)) return 0;
cannam@127: 
cannam@127:      if (NO_UGLYP(plnr)) {
cannam@127: 	  const problem_dft *p = (const problem_dft *) p_;
cannam@127: 	  if (p->ri != p->ro) return 0;
cannam@127: 	  if (X(toobig)(p->sz->dims[0].n)) return 0;
cannam@127:      }
cannam@127:      return 1;
cannam@127: }
cannam@127: 
cannam@127: static plan *mkplan(const solver *ego_, const problem *p_, planner *plnr)
cannam@127: {
cannam@127:      P *pln;
cannam@127:      const S *ego = (const S *)ego_;
cannam@127:      plan *cld = (plan *) 0;
cannam@127:      plan *cldcpy = (plan *) 0;
cannam@127:      plan *cldrest = (plan *) 0;
cannam@127:      const problem_dft *p = (const problem_dft *) p_;
cannam@127:      R *bufs = (R *) 0;
cannam@127:      INT nbuf = 0, bufdist, n, vl;
cannam@127:      INT ivs, ovs, roffset, ioffset;
cannam@127: 
cannam@127:      static const plan_adt padt = {
cannam@127: 	  X(dft_solve), awake, print, destroy
cannam@127:      };
cannam@127: 
cannam@127:      if (!applicable(ego, p_, plnr))
cannam@127:           goto nada;
cannam@127: 
cannam@127:      n = X(tensor_sz)(p->sz);
cannam@127: 
cannam@127:      X(tensor_tornk1)(p->vecsz, &vl, &ivs, &ovs);
cannam@127: 
cannam@127:      nbuf = X(nbuf)(n, vl, maxnbufs[ego->maxnbuf_ndx]);
cannam@127:      bufdist = X(bufdist)(n, vl);
cannam@127:      A(nbuf > 0);
cannam@127: 
cannam@127:      /* attempt to keep real and imaginary part in the same order,
cannam@127: 	so as to allow optimizations in the the copy plan */
cannam@127:      roffset = (p->ri - p->ii > 0) ? (INT)1 : (INT)0;
cannam@127:      ioffset = 1 - roffset;
cannam@127: 
cannam@127:      /* initial allocation for the purpose of planning */
cannam@127:      bufs = (R *) MALLOC(sizeof(R) * nbuf * bufdist * 2, BUFFERS);
cannam@127: 
cannam@127:      /* allow destruction of input if problem is in place */
cannam@127:      cld = X(mkplan_f_d)(plnr,
cannam@127: 			 X(mkproblem_dft_d)(
cannam@127: 			      X(mktensor_1d)(n, p->sz->dims[0].is, 2),
cannam@127: 			      X(mktensor_1d)(nbuf, ivs, bufdist * 2),
cannam@127: 			      TAINT(p->ri, ivs * nbuf),
cannam@127: 			      TAINT(p->ii, ivs * nbuf),
cannam@127: 			      bufs + roffset, 
cannam@127: 			      bufs + ioffset),
cannam@127: 			 0, 0, (p->ri == p->ro) ? NO_DESTROY_INPUT : 0);
cannam@127:      if (!cld)
cannam@127:           goto nada;
cannam@127: 
cannam@127:      /* copying back from the buffer is a rank-0 transform: */
cannam@127:      cldcpy = X(mkplan_d)(plnr,
cannam@127: 			  X(mkproblem_dft_d)(
cannam@127: 			       X(mktensor_0d)(),
cannam@127: 			       X(mktensor_2d)(nbuf, bufdist * 2, ovs,
cannam@127: 					      n, 2, p->sz->dims[0].os),
cannam@127: 			       bufs + roffset, 
cannam@127: 			       bufs + ioffset, 
cannam@127: 			       TAINT(p->ro, ovs * nbuf), 
cannam@127: 			       TAINT(p->io, ovs * nbuf)));
cannam@127:      if (!cldcpy)
cannam@127:           goto nada;
cannam@127: 
cannam@127:      /* deallocate buffers, let apply() allocate them for real */
cannam@127:      X(ifree)(bufs);
cannam@127:      bufs = 0;
cannam@127: 
cannam@127:      /* plan the leftover transforms (cldrest): */
cannam@127:      {
cannam@127: 	  INT id = ivs * (nbuf * (vl / nbuf));
cannam@127: 	  INT od = ovs * (nbuf * (vl / nbuf));
cannam@127: 	  cldrest = X(mkplan_d)(plnr, 
cannam@127: 				X(mkproblem_dft_d)(
cannam@127: 				     X(tensor_copy)(p->sz),
cannam@127: 				     X(mktensor_1d)(vl % nbuf, ivs, ovs),
cannam@127: 				     p->ri+id, p->ii+id, p->ro+od, p->io+od));
cannam@127:      }
cannam@127:      if (!cldrest)
cannam@127:           goto nada;
cannam@127: 
cannam@127:      pln = MKPLAN_DFT(P, &padt, apply);
cannam@127:      pln->cld = cld;
cannam@127:      pln->cldcpy = cldcpy;
cannam@127:      pln->cldrest = cldrest;
cannam@127:      pln->n = n;
cannam@127:      pln->vl = vl;
cannam@127:      pln->ivs_by_nbuf = ivs * nbuf;
cannam@127:      pln->ovs_by_nbuf = ovs * nbuf;
cannam@127:      pln->roffset = roffset;
cannam@127:      pln->ioffset = ioffset;
cannam@127: 
cannam@127:      pln->nbuf = nbuf;
cannam@127:      pln->bufdist = bufdist;
cannam@127: 
cannam@127:      {
cannam@127: 	  opcnt t;
cannam@127: 	  X(ops_add)(&cld->ops, &cldcpy->ops, &t);
cannam@127: 	  X(ops_madd)(vl / nbuf, &t, &cldrest->ops, &pln->super.super.ops);
cannam@127:      }
cannam@127: 
cannam@127:      return &(pln->super.super);
cannam@127: 
cannam@127:  nada:
cannam@127:      X(ifree0)(bufs);
cannam@127:      X(plan_destroy_internal)(cldrest);
cannam@127:      X(plan_destroy_internal)(cldcpy);
cannam@127:      X(plan_destroy_internal)(cld);
cannam@127:      return (plan *) 0;
cannam@127: }
cannam@127: 
cannam@127: static solver *mksolver(size_t maxnbuf_ndx)
cannam@127: {
cannam@127:      static const solver_adt sadt = { PROBLEM_DFT, mkplan, 0 };
cannam@127:      S *slv = MKSOLVER(S, &sadt);
cannam@127:      slv->maxnbuf_ndx = maxnbuf_ndx;
cannam@127:      return &(slv->super);
cannam@127: }
cannam@127: 
cannam@127: void X(dft_buffered_register)(planner *p)
cannam@127: {
cannam@127:      size_t i;
cannam@127:      for (i = 0; i < NELEM(maxnbufs); ++i)
cannam@127: 	  REGISTER_SOLVER(p, mksolver(i));
cannam@127: }