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 "threads.h"
cannam@95: 
cannam@95: typedef struct {
cannam@95:      plan_dft super;
cannam@95:      plan *cld;
cannam@95:      plan **cldws;
cannam@95:      int nthr;
cannam@95:      INT r;
cannam@95: } P;
cannam@95: 
cannam@95: typedef struct {
cannam@95:      plan **cldws;
cannam@95:      R *r, *i;
cannam@95: } PD;
cannam@95: 
cannam@95: static void *spawn_apply(spawn_data *d)
cannam@95: {
cannam@95:      PD *ego = (PD *) d->data;
cannam@95:      INT thr_num = d->thr_num;
cannam@95: 
cannam@95:      plan_dftw *cldw = (plan_dftw *) (ego->cldws[thr_num]);
cannam@95:      cldw->apply((plan *) cldw, ego->r, ego->i);
cannam@95:      return 0;
cannam@95: }
cannam@95: 
cannam@95: static void apply_dit(const plan *ego_, R *ri, R *ii, R *ro, R *io)
cannam@95: {
cannam@95:      const P *ego = (const P *) ego_;
cannam@95:      plan_dft *cld;
cannam@95: 
cannam@95:      cld = (plan_dft *) ego->cld;
cannam@95:      cld->apply(ego->cld, ri, ii, ro, io);
cannam@95: 
cannam@95:      {
cannam@95: 	  PD d;
cannam@95: 
cannam@95: 	  d.r = ro; d.i = io;
cannam@95: 	  d.cldws = ego->cldws;
cannam@95: 
cannam@95: 	  X(spawn_loop)(ego->nthr, ego->nthr, spawn_apply, (void*)&d);
cannam@95:      }
cannam@95: }
cannam@95: 
cannam@95: static void apply_dif(const plan *ego_, R *ri, R *ii, R *ro, R *io)
cannam@95: {
cannam@95:      const P *ego = (const P *) ego_;
cannam@95:      plan_dft *cld;
cannam@95: 
cannam@95:      {
cannam@95: 	  PD d;
cannam@95: 
cannam@95: 	  d.r = ri; d.i = ii;
cannam@95: 	  d.cldws = ego->cldws;
cannam@95: 
cannam@95: 	  X(spawn_loop)(ego->nthr, ego->nthr, spawn_apply, (void*)&d);
cannam@95:      }
cannam@95: 
cannam@95:      cld = (plan_dft *) ego->cld;
cannam@95:      cld->apply(ego->cld, ri, ii, ro, io);
cannam@95: }
cannam@95: 
cannam@95: static void awake(plan *ego_, enum wakefulness wakefulness)
cannam@95: {
cannam@95:      P *ego = (P *) ego_;
cannam@95:      int i;
cannam@95:      X(plan_awake)(ego->cld, wakefulness);
cannam@95:      for (i = 0; i < ego->nthr; ++i)
cannam@95: 	  X(plan_awake)(ego->cldws[i], wakefulness);
cannam@95: }
cannam@95: 
cannam@95: static void destroy(plan *ego_)
cannam@95: {
cannam@95:      P *ego = (P *) ego_;
cannam@95:      int i;
cannam@95:      X(plan_destroy_internal)(ego->cld);
cannam@95:      for (i = 0; i < ego->nthr; ++i)
cannam@95: 	  X(plan_destroy_internal)(ego->cldws[i]);
cannam@95:      X(ifree)(ego->cldws);
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:      int i;
cannam@95:      p->print(p, "(dft-thr-ct-%s-x%d/%D",
cannam@95: 	      ego->super.apply == apply_dit ? "dit" : "dif",
cannam@95: 	      ego->nthr, ego->r);
cannam@95:      for (i = 0; i < ego->nthr; ++i)
cannam@95:           if (i == 0 || (ego->cldws[i] != ego->cldws[i-1] &&
cannam@95:                          (i <= 1 || ego->cldws[i] != ego->cldws[i-2])))
cannam@95:                p->print(p, "%(%p%)", ego->cldws[i]);
cannam@95:      p->print(p, "%(%p%))", ego->cld);
cannam@95: }
cannam@95: 
cannam@95: static plan *mkplan(const solver *ego_, const problem *p_, planner *plnr)
cannam@95: {
cannam@95:      const ct_solver *ego = (const ct_solver *) ego_;
cannam@95:      const problem_dft *p;
cannam@95:      P *pln = 0;
cannam@95:      plan *cld = 0, **cldws = 0;
cannam@95:      INT n, r, m, v, ivs, ovs;
cannam@95:      INT block_size;
cannam@95:      int i, nthr, plnr_nthr_save;
cannam@95:      iodim *d;
cannam@95: 
cannam@95:      static const plan_adt padt = {
cannam@95: 	  X(dft_solve), awake, print, destroy
cannam@95:      };
cannam@95: 
cannam@95:      if (plnr->nthr <= 1 || !X(ct_applicable)(ego, p_, plnr))
cannam@95:           return (plan *) 0;
cannam@95: 
cannam@95:      p = (const problem_dft *) p_;
cannam@95:      d = p->sz->dims;
cannam@95:      n = d[0].n;
cannam@95:      r = X(choose_radix)(ego->r, n);
cannam@95:      m = n / r;
cannam@95: 
cannam@95:      X(tensor_tornk1)(p->vecsz, &v, &ivs, &ovs);
cannam@95: 
cannam@95:      block_size = (m + plnr->nthr - 1) / plnr->nthr;
cannam@95:      nthr = (int)((m + block_size - 1) / block_size);
cannam@95:      plnr_nthr_save = plnr->nthr;
cannam@95:      plnr->nthr = (plnr->nthr + nthr - 1) / nthr;
cannam@95: 
cannam@95:      cldws = (plan **) MALLOC(sizeof(plan *) * nthr, PLANS);
cannam@95:      for (i = 0; i < nthr; ++i) cldws[i] = (plan *) 0;
cannam@95: 
cannam@95:      switch (ego->dec) {
cannam@95: 	 case DECDIT:
cannam@95: 	 {
cannam@95: 	      for (i = 0; i < nthr; ++i) {
cannam@95: 		   cldws[i] = ego->mkcldw(ego,
cannam@95: 					  r, m * d[0].os, m * d[0].os,
cannam@95: 					  m, d[0].os,
cannam@95: 					  v, ovs, ovs,
cannam@95: 					  i*block_size,
cannam@95: 					  (i == nthr - 1) ?
cannam@95: 					  (m - i*block_size) : block_size,
cannam@95: 					  p->ro, p->io, plnr);
cannam@95: 		   if (!cldws[i]) goto nada;
cannam@95: 	      }
cannam@95: 
cannam@95: 	      plnr->nthr = plnr_nthr_save;
cannam@95: 
cannam@95: 	      cld = X(mkplan_d)(plnr,
cannam@95: 				X(mkproblem_dft_d)(
cannam@95: 				     X(mktensor_1d)(m, r * d[0].is, d[0].os),
cannam@95: 				     X(mktensor_2d)(r, d[0].is, m * d[0].os,
cannam@95: 						    v, ivs, ovs),
cannam@95: 				     p->ri, p->ii, p->ro, p->io)
cannam@95: 		   );
cannam@95: 	      if (!cld) goto nada;
cannam@95: 
cannam@95: 	      pln = MKPLAN_DFT(P, &padt, apply_dit);
cannam@95: 	      break;
cannam@95: 	 }
cannam@95: 	 case DECDIF:
cannam@95: 	 case DECDIF+TRANSPOSE:
cannam@95: 	 {
cannam@95: 	      INT cors, covs; /* cldw ors, ovs */
cannam@95: 	      if (ego->dec == DECDIF+TRANSPOSE) {
cannam@95: 		   cors = ivs;
cannam@95: 		   covs = m * d[0].is;
cannam@95: 		   /* ensure that we generate well-formed dftw subproblems */
cannam@95: 		   /* FIXME: too conservative */
cannam@95: 		   if (!(1
cannam@95: 			 && r == v
cannam@95: 			 && d[0].is == r * cors))
cannam@95: 			goto nada;
cannam@95: 
cannam@95: 		   /* FIXME: allow in-place only for now, like in
cannam@95: 		      fftw-3.[01] */
cannam@95: 		   if (!(1
cannam@95: 			 && p->ri == p->ro
cannam@95: 			 && d[0].is == r * d[0].os
cannam@95: 			 && cors == d[0].os
cannam@95: 			 && covs == ovs
cannam@95: 			    ))
cannam@95: 			goto nada;
cannam@95: 	      } else {
cannam@95: 		   cors = m * d[0].is;
cannam@95: 		   covs = ivs;
cannam@95: 	      }
cannam@95: 
cannam@95: 	      for (i = 0; i < nthr; ++i) {
cannam@95: 		   cldws[i] = ego->mkcldw(ego,
cannam@95: 					  r, m * d[0].is, cors,
cannam@95: 					  m, d[0].is,
cannam@95: 					  v, ivs, covs,
cannam@95: 					  i*block_size,
cannam@95: 					  (i == nthr - 1) ?
cannam@95: 					  (m - i*block_size) : block_size,
cannam@95: 					  p->ri, p->ii, plnr);
cannam@95: 		   if (!cldws[i]) goto nada;
cannam@95: 	      }
cannam@95: 
cannam@95: 	      plnr->nthr = plnr_nthr_save;
cannam@95: 
cannam@95: 	      cld = X(mkplan_d)(plnr,
cannam@95: 				X(mkproblem_dft_d)(
cannam@95: 				     X(mktensor_1d)(m, d[0].is, r * d[0].os),
cannam@95: 				     X(mktensor_2d)(r, cors, d[0].os,
cannam@95: 						    v, covs, ovs),
cannam@95: 				     p->ri, p->ii, p->ro, p->io)
cannam@95: 		   );
cannam@95: 	      if (!cld) goto nada;
cannam@95: 
cannam@95: 	      pln = MKPLAN_DFT(P, &padt, apply_dif);
cannam@95: 	      break;
cannam@95: 	 }
cannam@95: 
cannam@95: 	 default: A(0);
cannam@95: 
cannam@95:      }
cannam@95: 
cannam@95:      pln->cld = cld;
cannam@95:      pln->cldws = cldws;
cannam@95:      pln->nthr = nthr;
cannam@95:      pln->r = r;
cannam@95:      X(ops_zero)(&pln->super.super.ops);
cannam@95:      for (i = 0; i < nthr; ++i) {
cannam@95:           X(ops_add2)(&cldws[i]->ops, &pln->super.super.ops);
cannam@95: 	  pln->super.super.could_prune_now_p |= cldws[i]->could_prune_now_p;
cannam@95:      }
cannam@95:      X(ops_add2)(&cld->ops, &pln->super.super.ops);
cannam@95:      return &(pln->super.super);
cannam@95: 
cannam@95:  nada:
cannam@95:      if (cldws) {
cannam@95: 	  for (i = 0; i < nthr; ++i)
cannam@95: 	       X(plan_destroy_internal)(cldws[i]);
cannam@95: 	  X(ifree)(cldws);
cannam@95:      }
cannam@95:      X(plan_destroy_internal)(cld);
cannam@95:      return (plan *) 0;
cannam@95: }
cannam@95: 
cannam@95: ct_solver *X(mksolver_ct_threads)(size_t size, INT r, int dec,
cannam@95: 				  ct_mkinferior mkcldw,
cannam@95: 				  ct_force_vrecursion force_vrecursionp)
cannam@95: {
cannam@95:      static const solver_adt sadt = { PROBLEM_DFT, mkplan, 0 };
cannam@95:      ct_solver *slv = (ct_solver *) X(mksolver)(size, &sadt);
cannam@95:      slv->r = r;
cannam@95:      slv->dec = dec;
cannam@95:      slv->mkcldw = mkcldw;
cannam@95:      slv->force_vrecursionp = force_vrecursionp;
cannam@95:      return slv;
cannam@95: }