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:      solver super;
Chris@10:      int vecloop_dim;
Chris@10:      const int *buddies;
Chris@10:      int nbuddies;
Chris@10: } S;
Chris@10: 
Chris@10: typedef struct {
Chris@10:      plan_dft super;
Chris@10:      plan **cldrn;
Chris@10:      INT its, ots;
Chris@10:      int nthr;
Chris@10:      const S *solver;
Chris@10: } P;
Chris@10: 
Chris@10: typedef struct {
Chris@10:      INT its, ots;
Chris@10:      R *ri, *ii, *ro, *io;
Chris@10:      plan **cldrn;
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 its = ego->its;
Chris@10:      INT ots = ego->ots;
Chris@10:      int thr_num = d->thr_num;
Chris@10:      plan_dft *cld = (plan_dft *) ego->cldrn[thr_num];
Chris@10: 
Chris@10:      cld->apply((plan *) cld,
Chris@10: 		ego->ri + thr_num * its, ego->ii + thr_num * its,
Chris@10: 		ego->ro + thr_num * ots, ego->io + thr_num * ots);
Chris@10:      return 0;
Chris@10: }
Chris@10: 
Chris@10: static void apply(const plan *ego_, R *ri, R *ii, R *ro, R *io)
Chris@10: {
Chris@10:      const P *ego = (const P *) ego_;
Chris@10:      PD d;
Chris@10: 
Chris@10:      d.its = ego->its;
Chris@10:      d.ots = ego->ots;
Chris@10:      d.cldrn = ego->cldrn;
Chris@10:      d.ri = ri; d.ii = ii; d.ro = ro; d.io = io;
Chris@10: 
Chris@10:      X(spawn_loop)(ego->nthr, ego->nthr, spawn_apply, (void*) &d);
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:      for (i = 0; i < ego->nthr; ++i)
Chris@10: 	  X(plan_awake)(ego->cldrn[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:      for (i = 0; i < ego->nthr; ++i)
Chris@10: 	  X(plan_destroy_internal)(ego->cldrn[i]);
Chris@10:      X(ifree)(ego->cldrn);
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:      const S *s = ego->solver;
Chris@10:      int i;
Chris@10:      p->print(p, "(dft-thr-vrank>=1-x%d/%d", ego->nthr, s->vecloop_dim);
Chris@10:      for (i = 0; i < ego->nthr; ++i)
Chris@10: 	  if (i == 0 || (ego->cldrn[i] != ego->cldrn[i-1] &&
Chris@10: 			 (i <= 1 || ego->cldrn[i] != ego->cldrn[i-2])))
Chris@10: 	       p->print(p, "%(%p%)", ego->cldrn[i]);
Chris@10:      p->putchr(p, ')');
Chris@10: }
Chris@10: 
Chris@10: static int pickdim(const S *ego, const tensor *vecsz, int oop, int *dp)
Chris@10: {
Chris@10:      return X(pickdim)(ego->vecloop_dim, ego->buddies, ego->nbuddies,
Chris@10:                        vecsz, oop, dp);
Chris@10: }
Chris@10: 
Chris@10: static int applicable0(const solver *ego_, const problem *p_,
Chris@10: 		       const planner *plnr, int *dp)
Chris@10: {
Chris@10:      const S *ego = (const S *) ego_;
Chris@10:      const problem_dft *p = (const problem_dft *) p_;
Chris@10: 
Chris@10:      return (1
Chris@10: 	     && plnr->nthr > 1
Chris@10: 	     && FINITE_RNK(p->vecsz->rnk)
Chris@10: 	     && p->vecsz->rnk > 0
Chris@10: 	     && pickdim(ego, p->vecsz, p->ri != p->ro, dp)
Chris@10: 	  );
Chris@10: }
Chris@10: 
Chris@10: static int applicable(const solver *ego_, const problem *p_,
Chris@10: 		      const planner *plnr, int *dp)
Chris@10: {
Chris@10:      const S *ego = (const S *)ego_;
Chris@10: 
Chris@10:      if (!applicable0(ego_, p_, plnr, dp)) return 0;
Chris@10: 
Chris@10:      /* fftw2 behavior */
Chris@10:      if (NO_VRANK_SPLITSP(plnr) && (ego->vecloop_dim != ego->buddies[0]))
Chris@10: 	  return 0;
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:      const S *ego = (const S *) ego_;
Chris@10:      const problem_dft *p;
Chris@10:      P *pln;
Chris@10:      problem *cldp;
Chris@10:      int vdim;
Chris@10:      iodim *d;
Chris@10:      plan **cldrn = (plan **) 0;
Chris@10:      int i, nthr;
Chris@10:      INT its, ots, block_size;
Chris@10:      tensor *vecsz = 0;
Chris@10: 
Chris@10:      static const plan_adt padt = {
Chris@10: 	  X(dft_solve), awake, print, destroy
Chris@10:      };
Chris@10: 
Chris@10:      if (!applicable(ego_, p_, plnr, &vdim))
Chris@10:           return (plan *) 0;
Chris@10:      p = (const problem_dft *) p_;
Chris@10:      d = p->vecsz->dims + vdim;
Chris@10: 
Chris@10:      block_size = (d->n + plnr->nthr - 1) / plnr->nthr;
Chris@10:      nthr = (int)((d->n + block_size - 1) / block_size);
Chris@10:      plnr->nthr = (plnr->nthr + nthr - 1) / nthr;
Chris@10:      its = d->is * block_size;
Chris@10:      ots = d->os * block_size;
Chris@10: 
Chris@10:      cldrn = (plan **)MALLOC(sizeof(plan *) * nthr, PLANS);
Chris@10:      for (i = 0; i < nthr; ++i) cldrn[i] = (plan *) 0;
Chris@10:      
Chris@10:      vecsz = X(tensor_copy)(p->vecsz);
Chris@10:      for (i = 0; i < nthr; ++i) {
Chris@10: 	  vecsz->dims[vdim].n =
Chris@10: 	       (i == nthr - 1) ? (d->n - i*block_size) : block_size;
Chris@10: 	  cldp = X(mkproblem_dft)(p->sz, vecsz,
Chris@10: 				  p->ri + i*its, p->ii + i*its, 
Chris@10: 				  p->ro + i*ots, p->io + i*ots);
Chris@10: 	  cldrn[i] = X(mkplan_d)(plnr, cldp);
Chris@10: 	  if (!cldrn[i]) goto nada;
Chris@10:      }
Chris@10:      X(tensor_destroy)(vecsz);
Chris@10: 
Chris@10:      pln = MKPLAN_DFT(P, &padt, apply);
Chris@10: 
Chris@10:      pln->cldrn = cldrn;
Chris@10:      pln->its = its;
Chris@10:      pln->ots = ots;
Chris@10:      pln->nthr = nthr;
Chris@10: 
Chris@10:      pln->solver = ego;
Chris@10:      X(ops_zero)(&pln->super.super.ops);
Chris@10:      pln->super.super.pcost = 0;
Chris@10:      for (i = 0; i < nthr; ++i) {
Chris@10: 	  X(ops_add2)(&cldrn[i]->ops, &pln->super.super.ops);
Chris@10: 	  pln->super.super.pcost += cldrn[i]->pcost;
Chris@10:      }
Chris@10: 
Chris@10:      return &(pln->super.super);
Chris@10: 
Chris@10:  nada:
Chris@10:      if (cldrn) {
Chris@10: 	  for (i = 0; i < nthr; ++i)
Chris@10: 	       X(plan_destroy_internal)(cldrn[i]);
Chris@10: 	  X(ifree)(cldrn);
Chris@10:      }
Chris@10:      X(tensor_destroy)(vecsz);
Chris@10:      return (plan *) 0;
Chris@10: }
Chris@10: 
Chris@10: static solver *mksolver(int vecloop_dim, const int *buddies, int nbuddies)
Chris@10: {
Chris@10:      static const solver_adt sadt = { PROBLEM_DFT, mkplan, 0 };
Chris@10:      S *slv = MKSOLVER(S, &sadt);
Chris@10:      slv->vecloop_dim = vecloop_dim;
Chris@10:      slv->buddies = buddies;
Chris@10:      slv->nbuddies = nbuddies;
Chris@10:      return &(slv->super);
Chris@10: }
Chris@10: 
Chris@10: void X(dft_thr_vrank_geq1_register)(planner *p)
Chris@10: {
Chris@10:      int i;
Chris@10: 
Chris@10:      /* FIXME: Should we try other vecloop_dim values? */
Chris@10:      static const int buddies[] = { 1, -1 };
Chris@10: 
Chris@10:      const int nbuddies = (int)(sizeof(buddies) / sizeof(buddies[0]));
Chris@10: 
Chris@10:      for (i = 0; i < nbuddies; ++i)
Chris@10:           REGISTER_SOLVER(p, mksolver(buddies[i], buddies, nbuddies));
Chris@10: }