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: /* "MPI" DFTs where all of the data is on one processor...just
cannam@95:    call through to serial API. */
cannam@95: 
cannam@95: #include "mpi-dft.h"
cannam@95: #include "dft.h"
cannam@95: 
cannam@95: typedef struct {
cannam@95:      plan_mpi_dft super;
cannam@95:      plan *cld;
cannam@95:      INT roff, ioff;
cannam@95: } P;
cannam@95: 
cannam@95: static void apply(const plan *ego_, R *I, R *O)
cannam@95: {
cannam@95:      const P *ego = (const P *) ego_;
cannam@95:      plan_dft *cld;
cannam@95:      INT roff = ego->roff, ioff = ego->ioff;
cannam@95:      cld = (plan_dft *) ego->cld;
cannam@95:      cld->apply(ego->cld, I+roff, I+ioff, O+roff, O+ioff);
cannam@95: }
cannam@95: 
cannam@95: static void awake(plan *ego_, enum wakefulness wakefulness)
cannam@95: {
cannam@95:      P *ego = (P *) ego_;
cannam@95:      X(plan_awake)(ego->cld, wakefulness);
cannam@95: }
cannam@95: 
cannam@95: static void destroy(plan *ego_)
cannam@95: {
cannam@95:      P *ego = (P *) ego_;
cannam@95:      X(plan_destroy_internal)(ego->cld);
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:      p->print(p, "(mpi-dft-serial %(%p%))", ego->cld);
cannam@95: }
cannam@95: 
cannam@95: int XM(dft_serial_applicable)(const problem_mpi_dft *p)
cannam@95: {
cannam@95:      return (1
cannam@95: 	     && p->flags == 0 /* TRANSPOSED/SCRAMBLED_IN/OUT not supported */
cannam@95: 	     && ((XM(is_local)(p->sz, IB) && XM(is_local)(p->sz, OB))
cannam@95: 		 || p->vn == 0));
cannam@95: }
cannam@95: 
cannam@95: static plan *mkplan(const solver *ego, const problem *p_, planner *plnr)
cannam@95: {
cannam@95:      const problem_mpi_dft *p = (const problem_mpi_dft *) p_;
cannam@95:      P *pln;
cannam@95:      plan *cld;
cannam@95:      int my_pe;
cannam@95:      R *ri, *ii, *ro, *io;
cannam@95:      static const plan_adt padt = {
cannam@95:           XM(dft_solve), awake, print, destroy
cannam@95:      };
cannam@95: 
cannam@95:      UNUSED(ego);
cannam@95: 
cannam@95:      /* check whether applicable: */
cannam@95:      if (!XM(dft_serial_applicable)(p))
cannam@95:           return (plan *) 0;
cannam@95: 
cannam@95:      X(extract_reim)(p->sign, p->I, &ri, &ii);
cannam@95:      X(extract_reim)(p->sign, p->O, &ro, &io);
cannam@95: 
cannam@95:      MPI_Comm_rank(p->comm, &my_pe);
cannam@95:      if (my_pe == 0 && p->vn > 0) {
cannam@95: 	  int i, rnk = p->sz->rnk;
cannam@95: 	  tensor *sz = X(mktensor)(p->sz->rnk);
cannam@95: 	  sz->dims[rnk - 1].is = sz->dims[rnk - 1].os = 2 * p->vn;
cannam@95: 	  sz->dims[rnk - 1].n = p->sz->dims[rnk - 1].n;
cannam@95: 	  for (i = rnk - 1; i > 0; --i) {
cannam@95: 	       sz->dims[i - 1].is = sz->dims[i - 1].os = 
cannam@95: 		    sz->dims[i].is * sz->dims[i].n;
cannam@95: 	       sz->dims[i - 1].n = p->sz->dims[i - 1].n;
cannam@95: 	  }
cannam@95: 	  
cannam@95: 	  cld = X(mkplan_d)(plnr,
cannam@95: 			    X(mkproblem_dft_d)(sz,
cannam@95: 					       X(mktensor_1d)(p->vn, 2, 2),
cannam@95: 					       ri, ii, ro, io));
cannam@95:      }
cannam@95:      else { /* idle process: make nop plan */
cannam@95: 	  cld = X(mkplan_d)(plnr,
cannam@95: 			    X(mkproblem_dft_d)(X(mktensor_0d)(),
cannam@95: 					       X(mktensor_1d)(0,0,0),
cannam@95: 					       ri, ii, ro, io));
cannam@95:      }
cannam@95:      if (XM(any_true)(!cld, p->comm)) return (plan *) 0;
cannam@95: 
cannam@95:      pln = MKPLAN_MPI_DFT(P, &padt, apply);
cannam@95:      pln->cld = cld;
cannam@95:      pln->roff = ro - p->O;
cannam@95:      pln->ioff = io - p->O;
cannam@95:      X(ops_cpy)(&cld->ops, &pln->super.super.ops);
cannam@95:      return &(pln->super.super);
cannam@95: }
cannam@95: 
cannam@95: static solver *mksolver(void)
cannam@95: {
cannam@95:      static const solver_adt sadt = { PROBLEM_MPI_DFT, mkplan, 0 };
cannam@95:      return MKSOLVER(solver, &sadt);
cannam@95: }
cannam@95: 
cannam@95: void XM(dft_serial_register)(planner *p)
cannam@95: {
cannam@95:      REGISTER_SOLVER(p, mksolver());
cannam@95: }