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 "threads.h"
Chris@10: 
Chris@10: typedef struct {
Chris@10:      plan_dft super;
Chris@10:      plan *cld;
Chris@10:      plan **cldws;
Chris@10:      int nthr;
Chris@10:      INT r;
Chris@10: } P;
Chris@10: 
Chris@10: typedef struct {
Chris@10:      plan **cldws;
Chris@10:      R *r, *i;
Chris@10: } PD;
Chris@10: 
Chris@10: static void *spawn_apply(spawn_data *d)
Chris@10: {
Chris@10:      PD *ego = (PD *) d->data;
Chris@10:      INT thr_num = d->thr_num;
Chris@10: 
Chris@10:      plan_dftw *cldw = (plan_dftw *) (ego->cldws[thr_num]);
Chris@10:      cldw->apply((plan *) cldw, ego->r, ego->i);
Chris@10:      return 0;
Chris@10: }
Chris@10: 
Chris@10: static void apply_dit(const plan *ego_, R *ri, R *ii, R *ro, R *io)
Chris@10: {
Chris@10:      const P *ego = (const P *) ego_;
Chris@10:      plan_dft *cld;
Chris@10: 
Chris@10:      cld = (plan_dft *) ego->cld;
Chris@10:      cld->apply(ego->cld, ri, ii, ro, io);
Chris@10: 
Chris@10:      {
Chris@10: 	  PD d;
Chris@10: 
Chris@10: 	  d.r = ro; d.i = io;
Chris@10: 	  d.cldws = ego->cldws;
Chris@10: 
Chris@10: 	  X(spawn_loop)(ego->nthr, ego->nthr, spawn_apply, (void*)&d);
Chris@10:      }
Chris@10: }
Chris@10: 
Chris@10: static void apply_dif(const plan *ego_, R *ri, R *ii, R *ro, R *io)
Chris@10: {
Chris@10:      const P *ego = (const P *) ego_;
Chris@10:      plan_dft *cld;
Chris@10: 
Chris@10:      {
Chris@10: 	  PD d;
Chris@10: 
Chris@10: 	  d.r = ri; d.i = ii;
Chris@10: 	  d.cldws = ego->cldws;
Chris@10: 
Chris@10: 	  X(spawn_loop)(ego->nthr, ego->nthr, spawn_apply, (void*)&d);
Chris@10:      }
Chris@10: 
Chris@10:      cld = (plan_dft *) ego->cld;
Chris@10:      cld->apply(ego->cld, ri, ii, ro, io);
Chris@10: }
Chris@10: 
Chris@10: static void awake(plan *ego_, enum wakefulness wakefulness)
Chris@10: {
Chris@10:      P *ego = (P *) ego_;
Chris@10:      int i;
Chris@10:      X(plan_awake)(ego->cld, wakefulness);
Chris@10:      for (i = 0; i < ego->nthr; ++i)
Chris@10: 	  X(plan_awake)(ego->cldws[i], wakefulness);
Chris@10: }
Chris@10: 
Chris@10: static void destroy(plan *ego_)
Chris@10: {
Chris@10:      P *ego = (P *) ego_;
Chris@10:      int i;
Chris@10:      X(plan_destroy_internal)(ego->cld);
Chris@10:      for (i = 0; i < ego->nthr; ++i)
Chris@10: 	  X(plan_destroy_internal)(ego->cldws[i]);
Chris@10:      X(ifree)(ego->cldws);
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:      int i;
Chris@10:      p->print(p, "(dft-thr-ct-%s-x%d/%D",
Chris@10: 	      ego->super.apply == apply_dit ? "dit" : "dif",
Chris@10: 	      ego->nthr, ego->r);
Chris@10:      for (i = 0; i < ego->nthr; ++i)
Chris@10:           if (i == 0 || (ego->cldws[i] != ego->cldws[i-1] &&
Chris@10:                          (i <= 1 || ego->cldws[i] != ego->cldws[i-2])))
Chris@10:                p->print(p, "%(%p%)", ego->cldws[i]);
Chris@10:      p->print(p, "%(%p%))", ego->cld);
Chris@10: }
Chris@10: 
Chris@10: static plan *mkplan(const solver *ego_, const problem *p_, planner *plnr)
Chris@10: {
Chris@10:      const ct_solver *ego = (const ct_solver *) ego_;
Chris@10:      const problem_dft *p;
Chris@10:      P *pln = 0;
Chris@10:      plan *cld = 0, **cldws = 0;
Chris@10:      INT n, r, m, v, ivs, ovs;
Chris@10:      INT block_size;
Chris@10:      int i, nthr, plnr_nthr_save;
Chris@10:      iodim *d;
Chris@10: 
Chris@10:      static const plan_adt padt = {
Chris@10: 	  X(dft_solve), awake, print, destroy
Chris@10:      };
Chris@10: 
Chris@10:      if (plnr->nthr <= 1 || !X(ct_applicable)(ego, p_, plnr))
Chris@10:           return (plan *) 0;
Chris@10: 
Chris@10:      p = (const problem_dft *) p_;
Chris@10:      d = p->sz->dims;
Chris@10:      n = d[0].n;
Chris@10:      r = X(choose_radix)(ego->r, n);
Chris@10:      m = n / r;
Chris@10: 
Chris@10:      X(tensor_tornk1)(p->vecsz, &v, &ivs, &ovs);
Chris@10: 
Chris@10:      block_size = (m + plnr->nthr - 1) / plnr->nthr;
Chris@10:      nthr = (int)((m + block_size - 1) / block_size);
Chris@10:      plnr_nthr_save = plnr->nthr;
Chris@10:      plnr->nthr = (plnr->nthr + nthr - 1) / nthr;
Chris@10: 
Chris@10:      cldws = (plan **) MALLOC(sizeof(plan *) * nthr, PLANS);
Chris@10:      for (i = 0; i < nthr; ++i) cldws[i] = (plan *) 0;
Chris@10: 
Chris@10:      switch (ego->dec) {
Chris@10: 	 case DECDIT:
Chris@10: 	 {
Chris@10: 	      for (i = 0; i < nthr; ++i) {
Chris@10: 		   cldws[i] = ego->mkcldw(ego,
Chris@10: 					  r, m * d[0].os, m * d[0].os,
Chris@10: 					  m, d[0].os,
Chris@10: 					  v, ovs, ovs,
Chris@10: 					  i*block_size,
Chris@10: 					  (i == nthr - 1) ?
Chris@10: 					  (m - i*block_size) : block_size,
Chris@10: 					  p->ro, p->io, plnr);
Chris@10: 		   if (!cldws[i]) goto nada;
Chris@10: 	      }
Chris@10: 
Chris@10: 	      plnr->nthr = plnr_nthr_save;
Chris@10: 
Chris@10: 	      cld = X(mkplan_d)(plnr,
Chris@10: 				X(mkproblem_dft_d)(
Chris@10: 				     X(mktensor_1d)(m, r * d[0].is, d[0].os),
Chris@10: 				     X(mktensor_2d)(r, d[0].is, m * d[0].os,
Chris@10: 						    v, ivs, ovs),
Chris@10: 				     p->ri, p->ii, p->ro, p->io)
Chris@10: 		   );
Chris@10: 	      if (!cld) goto nada;
Chris@10: 
Chris@10: 	      pln = MKPLAN_DFT(P, &padt, apply_dit);
Chris@10: 	      break;
Chris@10: 	 }
Chris@10: 	 case DECDIF:
Chris@10: 	 case DECDIF+TRANSPOSE:
Chris@10: 	 {
Chris@10: 	      INT cors, covs; /* cldw ors, ovs */
Chris@10: 	      if (ego->dec == DECDIF+TRANSPOSE) {
Chris@10: 		   cors = ivs;
Chris@10: 		   covs = m * d[0].is;
Chris@10: 		   /* ensure that we generate well-formed dftw subproblems */
Chris@10: 		   /* FIXME: too conservative */
Chris@10: 		   if (!(1
Chris@10: 			 && r == v
Chris@10: 			 && d[0].is == r * cors))
Chris@10: 			goto nada;
Chris@10: 
Chris@10: 		   /* FIXME: allow in-place only for now, like in
Chris@10: 		      fftw-3.[01] */
Chris@10: 		   if (!(1
Chris@10: 			 && p->ri == p->ro
Chris@10: 			 && d[0].is == r * d[0].os
Chris@10: 			 && cors == d[0].os
Chris@10: 			 && covs == ovs
Chris@10: 			    ))
Chris@10: 			goto nada;
Chris@10: 	      } else {
Chris@10: 		   cors = m * d[0].is;
Chris@10: 		   covs = ivs;
Chris@10: 	      }
Chris@10: 
Chris@10: 	      for (i = 0; i < nthr; ++i) {
Chris@10: 		   cldws[i] = ego->mkcldw(ego,
Chris@10: 					  r, m * d[0].is, cors,
Chris@10: 					  m, d[0].is,
Chris@10: 					  v, ivs, covs,
Chris@10: 					  i*block_size,
Chris@10: 					  (i == nthr - 1) ?
Chris@10: 					  (m - i*block_size) : block_size,
Chris@10: 					  p->ri, p->ii, plnr);
Chris@10: 		   if (!cldws[i]) goto nada;
Chris@10: 	      }
Chris@10: 
Chris@10: 	      plnr->nthr = plnr_nthr_save;
Chris@10: 
Chris@10: 	      cld = X(mkplan_d)(plnr,
Chris@10: 				X(mkproblem_dft_d)(
Chris@10: 				     X(mktensor_1d)(m, d[0].is, r * d[0].os),
Chris@10: 				     X(mktensor_2d)(r, cors, d[0].os,
Chris@10: 						    v, covs, ovs),
Chris@10: 				     p->ri, p->ii, p->ro, p->io)
Chris@10: 		   );
Chris@10: 	      if (!cld) goto nada;
Chris@10: 
Chris@10: 	      pln = MKPLAN_DFT(P, &padt, apply_dif);
Chris@10: 	      break;
Chris@10: 	 }
Chris@10: 
Chris@10: 	 default: A(0);
Chris@10: 
Chris@10:      }
Chris@10: 
Chris@10:      pln->cld = cld;
Chris@10:      pln->cldws = cldws;
Chris@10:      pln->nthr = nthr;
Chris@10:      pln->r = r;
Chris@10:      X(ops_zero)(&pln->super.super.ops);
Chris@10:      for (i = 0; i < nthr; ++i) {
Chris@10:           X(ops_add2)(&cldws[i]->ops, &pln->super.super.ops);
Chris@10: 	  pln->super.super.could_prune_now_p |= cldws[i]->could_prune_now_p;
Chris@10:      }
Chris@10:      X(ops_add2)(&cld->ops, &pln->super.super.ops);
Chris@10:      return &(pln->super.super);
Chris@10: 
Chris@10:  nada:
Chris@10:      if (cldws) {
Chris@10: 	  for (i = 0; i < nthr; ++i)
Chris@10: 	       X(plan_destroy_internal)(cldws[i]);
Chris@10: 	  X(ifree)(cldws);
Chris@10:      }
Chris@10:      X(plan_destroy_internal)(cld);
Chris@10:      return (plan *) 0;
Chris@10: }
Chris@10: 
Chris@10: ct_solver *X(mksolver_ct_threads)(size_t size, INT r, int dec,
Chris@10: 				  ct_mkinferior mkcldw,
Chris@10: 				  ct_force_vrecursion force_vrecursionp)
Chris@10: {
Chris@10:      static const solver_adt sadt = { PROBLEM_DFT, mkplan, 0 };
Chris@10:      ct_solver *slv = (ct_solver *) X(mksolver)(size, &sadt);
Chris@10:      slv->r = r;
Chris@10:      slv->dec = dec;
Chris@10:      slv->mkcldw = mkcldw;
Chris@10:      slv->force_vrecursionp = force_vrecursionp;
Chris@10:      return slv;
Chris@10: }