cannam@167: /*
cannam@167:  * Copyright (c) 2003, 2007-14 Matteo Frigo
cannam@167:  * Copyright (c) 2003, 2007-14 Massachusetts Institute of Technology
cannam@167:  *
cannam@167:  * This program is free software; you can redistribute it and/or modify
cannam@167:  * it under the terms of the GNU General Public License as published by
cannam@167:  * the Free Software Foundation; either version 2 of the License, or
cannam@167:  * (at your option) any later version.
cannam@167:  *
cannam@167:  * This program is distributed in the hope that it will be useful,
cannam@167:  * but WITHOUT ANY WARRANTY; without even the implied warranty of
cannam@167:  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
cannam@167:  * GNU General Public License for more details.
cannam@167:  *
cannam@167:  * You should have received a copy of the GNU General Public License
cannam@167:  * along with this program; if not, write to the Free Software
cannam@167:  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA
cannam@167:  *
cannam@167:  */
cannam@167: 
cannam@167: 
cannam@167: #include "threads/threads.h"
cannam@167: 
cannam@167: typedef struct {
cannam@167:      solver super;
cannam@167:      int vecloop_dim;
cannam@167:      const int *buddies;
cannam@167:      size_t nbuddies;
cannam@167: } S;
cannam@167: 
cannam@167: typedef struct {
cannam@167:      plan_dft super;
cannam@167:      plan **cldrn;
cannam@167:      INT its, ots;
cannam@167:      int nthr;
cannam@167:      const S *solver;
cannam@167: } P;
cannam@167: 
cannam@167: typedef struct {
cannam@167:      INT its, ots;
cannam@167:      R *ri, *ii, *ro, *io;
cannam@167:      plan **cldrn;
cannam@167: } PD;
cannam@167: 
cannam@167: static void *spawn_apply(spawn_data *d)
cannam@167: {
cannam@167:      PD *ego = (PD *) d->data;
cannam@167:      INT its = ego->its;
cannam@167:      INT ots = ego->ots;
cannam@167:      int thr_num = d->thr_num;
cannam@167:      plan_dft *cld = (plan_dft *) ego->cldrn[thr_num];
cannam@167: 
cannam@167:      cld->apply((plan *) cld,
cannam@167: 		ego->ri + thr_num * its, ego->ii + thr_num * its,
cannam@167: 		ego->ro + thr_num * ots, ego->io + thr_num * ots);
cannam@167:      return 0;
cannam@167: }
cannam@167: 
cannam@167: static void apply(const plan *ego_, R *ri, R *ii, R *ro, R *io)
cannam@167: {
cannam@167:      const P *ego = (const P *) ego_;
cannam@167:      PD d;
cannam@167: 
cannam@167:      d.its = ego->its;
cannam@167:      d.ots = ego->ots;
cannam@167:      d.cldrn = ego->cldrn;
cannam@167:      d.ri = ri; d.ii = ii; d.ro = ro; d.io = io;
cannam@167: 
cannam@167:      X(spawn_loop)(ego->nthr, ego->nthr, spawn_apply, (void*) &d);
cannam@167: }
cannam@167: 
cannam@167: static void awake(plan *ego_, enum wakefulness wakefulness)
cannam@167: {
cannam@167:      P *ego = (P *) ego_;
cannam@167:      int i;
cannam@167:      for (i = 0; i < ego->nthr; ++i)
cannam@167: 	  X(plan_awake)(ego->cldrn[i], wakefulness);
cannam@167: }
cannam@167: 
cannam@167: static void destroy(plan *ego_)
cannam@167: {
cannam@167:      P *ego = (P *) ego_;
cannam@167:      int i;
cannam@167:      for (i = 0; i < ego->nthr; ++i)
cannam@167: 	  X(plan_destroy_internal)(ego->cldrn[i]);
cannam@167:      X(ifree)(ego->cldrn);
cannam@167: }
cannam@167: 
cannam@167: static void print(const plan *ego_, printer *p)
cannam@167: {
cannam@167:      const P *ego = (const P *) ego_;
cannam@167:      const S *s = ego->solver;
cannam@167:      int i;
cannam@167:      p->print(p, "(dft-thr-vrank>=1-x%d/%d", ego->nthr, s->vecloop_dim);
cannam@167:      for (i = 0; i < ego->nthr; ++i)
cannam@167: 	  if (i == 0 || (ego->cldrn[i] != ego->cldrn[i-1] &&
cannam@167: 			 (i <= 1 || ego->cldrn[i] != ego->cldrn[i-2])))
cannam@167: 	       p->print(p, "%(%p%)", ego->cldrn[i]);
cannam@167:      p->putchr(p, ')');
cannam@167: }
cannam@167: 
cannam@167: static int pickdim(const S *ego, const tensor *vecsz, int oop, int *dp)
cannam@167: {
cannam@167:      return X(pickdim)(ego->vecloop_dim, ego->buddies, ego->nbuddies,
cannam@167:                        vecsz, oop, dp);
cannam@167: }
cannam@167: 
cannam@167: static int applicable0(const solver *ego_, const problem *p_,
cannam@167: 		       const planner *plnr, int *dp)
cannam@167: {
cannam@167:      const S *ego = (const S *) ego_;
cannam@167:      const problem_dft *p = (const problem_dft *) p_;
cannam@167: 
cannam@167:      return (1
cannam@167: 	     && plnr->nthr > 1
cannam@167: 	     && FINITE_RNK(p->vecsz->rnk)
cannam@167: 	     && p->vecsz->rnk > 0
cannam@167: 	     && pickdim(ego, p->vecsz, p->ri != p->ro, dp)
cannam@167: 	  );
cannam@167: }
cannam@167: 
cannam@167: static int applicable(const solver *ego_, const problem *p_,
cannam@167: 		      const planner *plnr, int *dp)
cannam@167: {
cannam@167:      const S *ego = (const S *)ego_;
cannam@167: 
cannam@167:      if (!applicable0(ego_, p_, plnr, dp)) return 0;
cannam@167: 
cannam@167:      /* fftw2 behavior */
cannam@167:      if (NO_VRANK_SPLITSP(plnr) && (ego->vecloop_dim != ego->buddies[0]))
cannam@167: 	  return 0;
cannam@167: 
cannam@167:      return 1;
cannam@167: }
cannam@167: 
cannam@167: static plan *mkplan(const solver *ego_, const problem *p_, planner *plnr)
cannam@167: {
cannam@167:      const S *ego = (const S *) ego_;
cannam@167:      const problem_dft *p;
cannam@167:      P *pln;
cannam@167:      problem *cldp;
cannam@167:      int vdim;
cannam@167:      iodim *d;
cannam@167:      plan **cldrn = (plan **) 0;
cannam@167:      int i, nthr;
cannam@167:      INT its, ots, block_size;
cannam@167:      tensor *vecsz = 0;
cannam@167: 
cannam@167:      static const plan_adt padt = {
cannam@167: 	  X(dft_solve), awake, print, destroy
cannam@167:      };
cannam@167: 
cannam@167:      if (!applicable(ego_, p_, plnr, &vdim))
cannam@167:           return (plan *) 0;
cannam@167:      p = (const problem_dft *) p_;
cannam@167:      d = p->vecsz->dims + vdim;
cannam@167: 
cannam@167:      block_size = (d->n + plnr->nthr - 1) / plnr->nthr;
cannam@167:      nthr = (int)((d->n + block_size - 1) / block_size);
cannam@167:      plnr->nthr = (plnr->nthr + nthr - 1) / nthr;
cannam@167:      its = d->is * block_size;
cannam@167:      ots = d->os * block_size;
cannam@167: 
cannam@167:      cldrn = (plan **)MALLOC(sizeof(plan *) * nthr, PLANS);
cannam@167:      for (i = 0; i < nthr; ++i) cldrn[i] = (plan *) 0;
cannam@167:      
cannam@167:      vecsz = X(tensor_copy)(p->vecsz);
cannam@167:      for (i = 0; i < nthr; ++i) {
cannam@167: 	  vecsz->dims[vdim].n =
cannam@167: 	       (i == nthr - 1) ? (d->n - i*block_size) : block_size;
cannam@167: 	  cldp = X(mkproblem_dft)(p->sz, vecsz,
cannam@167: 				  p->ri + i*its, p->ii + i*its, 
cannam@167: 				  p->ro + i*ots, p->io + i*ots);
cannam@167: 	  cldrn[i] = X(mkplan_d)(plnr, cldp);
cannam@167: 	  if (!cldrn[i]) goto nada;
cannam@167:      }
cannam@167:      X(tensor_destroy)(vecsz);
cannam@167: 
cannam@167:      pln = MKPLAN_DFT(P, &padt, apply);
cannam@167: 
cannam@167:      pln->cldrn = cldrn;
cannam@167:      pln->its = its;
cannam@167:      pln->ots = ots;
cannam@167:      pln->nthr = nthr;
cannam@167: 
cannam@167:      pln->solver = ego;
cannam@167:      X(ops_zero)(&pln->super.super.ops);
cannam@167:      pln->super.super.pcost = 0;
cannam@167:      for (i = 0; i < nthr; ++i) {
cannam@167: 	  X(ops_add2)(&cldrn[i]->ops, &pln->super.super.ops);
cannam@167: 	  pln->super.super.pcost += cldrn[i]->pcost;
cannam@167:      }
cannam@167: 
cannam@167:      return &(pln->super.super);
cannam@167: 
cannam@167:  nada:
cannam@167:      if (cldrn) {
cannam@167: 	  for (i = 0; i < nthr; ++i)
cannam@167: 	       X(plan_destroy_internal)(cldrn[i]);
cannam@167: 	  X(ifree)(cldrn);
cannam@167:      }
cannam@167:      X(tensor_destroy)(vecsz);
cannam@167:      return (plan *) 0;
cannam@167: }
cannam@167: 
cannam@167: static solver *mksolver(int vecloop_dim, const int *buddies, size_t nbuddies)
cannam@167: {
cannam@167:      static const solver_adt sadt = { PROBLEM_DFT, mkplan, 0 };
cannam@167:      S *slv = MKSOLVER(S, &sadt);
cannam@167:      slv->vecloop_dim = vecloop_dim;
cannam@167:      slv->buddies = buddies;
cannam@167:      slv->nbuddies = nbuddies;
cannam@167:      return &(slv->super);
cannam@167: }
cannam@167: 
cannam@167: void X(dft_thr_vrank_geq1_register)(planner *p)
cannam@167: {
cannam@167:      /* FIXME: Should we try other vecloop_dim values? */
cannam@167:      static const int buddies[] = { 1, -1 };
cannam@167:      size_t i;
cannam@167: 
cannam@167:      for (i = 0; i < NELEM(buddies); ++i)
cannam@167:           REGISTER_SOLVER(p, mksolver(buddies[i], buddies, NELEM(buddies)));
cannam@167: }