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: /* Complex DFTs of rank == 1 when the vector length vn is >= # processes.
Chris@10:    In this case, we don't need to use a six-step type algorithm, and can
Chris@10:    instead transpose the DFT dimension with the vector dimension to 
Chris@10:    make the DFT local. */
Chris@10: 
Chris@10: #include "mpi-dft.h"
Chris@10: #include "mpi-transpose.h"
Chris@10: #include "dft.h"
Chris@10: 
Chris@10: typedef struct {
Chris@10:      solver super;
Chris@10:      int preserve_input; /* preserve input even if DESTROY_INPUT was passed */
Chris@10:      rearrangement rearrange;
Chris@10: } S;
Chris@10: 
Chris@10: typedef struct {
Chris@10:      plan_mpi_dft super;
Chris@10: 
Chris@10:      plan *cldt_before, *cld, *cldt_after;
Chris@10:      INT roff, ioff;
Chris@10:      int preserve_input;
Chris@10:      rearrangement rearrange;
Chris@10: } P;
Chris@10: 
Chris@10: static void apply(const plan *ego_, R *I, R *O)
Chris@10: {
Chris@10:      const P *ego = (const P *) ego_;
Chris@10:      plan_dft *cld;
Chris@10:      plan_rdft *cldt_before, *cldt_after;
Chris@10:      INT roff = ego->roff, ioff = ego->ioff;
Chris@10:      
Chris@10:      /* global transpose */
Chris@10:      cldt_before = (plan_rdft *) ego->cldt_before;
Chris@10:      cldt_before->apply(ego->cldt_before, I, O);
Chris@10:      
Chris@10:      if (ego->preserve_input) I = O;
Chris@10: 	  
Chris@10:      /* 1d DFT(s) */
Chris@10:      cld = (plan_dft *) ego->cld;
Chris@10:      cld->apply(ego->cld, O+roff, O+ioff, I+roff, I+ioff);
Chris@10:      
Chris@10:      /* global transpose */
Chris@10:      cldt_after = (plan_rdft *) ego->cldt_after;
Chris@10:      cldt_after->apply(ego->cldt_after, I, O);
Chris@10: }
Chris@10: 
Chris@10: static int applicable(const S *ego, const problem *p_,
Chris@10: 		      const planner *plnr)
Chris@10: {
Chris@10:      const problem_mpi_dft *p = (const problem_mpi_dft *) p_;
Chris@10:      int n_pes;
Chris@10:      MPI_Comm_size(p->comm, &n_pes);
Chris@10:      return (1
Chris@10: 	     && p->sz->rnk == 1
Chris@10: 	     && !(p->flags & ~RANK1_BIGVEC_ONLY)
Chris@10: 	     && (!ego->preserve_input || (!NO_DESTROY_INPUTP(plnr)
Chris@10: 					  && p->I != p->O))
Chris@10: 	     && (p->vn >= n_pes /* TODO: relax this, using more memory? */
Chris@10: 		 || (p->flags & RANK1_BIGVEC_ONLY))
Chris@10: 
Chris@10: 	     && XM(rearrange_applicable)(ego->rearrange,
Chris@10: 					 p->sz->dims[0], p->vn, n_pes)
Chris@10: 
Chris@10: 	     && (!NO_SLOWP(plnr) /* slow if dft-serial is applicable */
Chris@10:                  || !XM(dft_serial_applicable)(p))
Chris@10: 	  );
Chris@10: }
Chris@10: 
Chris@10: static void awake(plan *ego_, enum wakefulness wakefulness)
Chris@10: {
Chris@10:      P *ego = (P *) ego_;
Chris@10:      X(plan_awake)(ego->cldt_before, wakefulness);
Chris@10:      X(plan_awake)(ego->cld, wakefulness);
Chris@10:      X(plan_awake)(ego->cldt_after, wakefulness);
Chris@10: }
Chris@10: 
Chris@10: static void destroy(plan *ego_)
Chris@10: {
Chris@10:      P *ego = (P *) ego_;
Chris@10:      X(plan_destroy_internal)(ego->cldt_after);
Chris@10:      X(plan_destroy_internal)(ego->cld);
Chris@10:      X(plan_destroy_internal)(ego->cldt_before);
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 char descrip[][16] = { "contig", "discontig", "square-after",
Chris@10: 				  "square-middle", "square-before" };
Chris@10:      p->print(p, "(mpi-dft-rank1-bigvec/%s%s %(%p%) %(%p%) %(%p%))",
Chris@10: 	      descrip[ego->rearrange], ego->preserve_input==2 ?"/p":"",
Chris@10: 	      ego->cldt_before, ego->cld, ego->cldt_after);
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_mpi_dft *p;
Chris@10:      P *pln;
Chris@10:      plan *cld = 0, *cldt_before = 0, *cldt_after = 0;
Chris@10:      R *ri, *ii, *ro, *io, *I, *O;
Chris@10:      INT yblock, yb, nx, ny, vn;
Chris@10:      int my_pe, n_pes;
Chris@10:      static const plan_adt padt = {
Chris@10:           XM(dft_solve), awake, print, destroy
Chris@10:      };
Chris@10: 
Chris@10:      UNUSED(ego);
Chris@10: 
Chris@10:      if (!applicable(ego, p_, plnr))
Chris@10:           return (plan *) 0;
Chris@10: 
Chris@10:      p = (const problem_mpi_dft *) p_;
Chris@10: 
Chris@10:      MPI_Comm_rank(p->comm, &my_pe);
Chris@10:      MPI_Comm_size(p->comm, &n_pes);
Chris@10:      
Chris@10:      nx = p->sz->dims[0].n;
Chris@10:      if (!(ny = XM(rearrange_ny)(ego->rearrange, p->sz->dims[0],p->vn,n_pes)))
Chris@10: 	  return (plan *) 0;
Chris@10:      vn = p->vn / ny;
Chris@10:      A(ny * vn == p->vn);
Chris@10: 
Chris@10:      yblock = XM(default_block)(ny, n_pes);
Chris@10:      cldt_before = X(mkplan_d)(plnr,
Chris@10: 			       XM(mkproblem_transpose)(
Chris@10: 				    nx, ny, vn*2,
Chris@10: 				    I = p->I, O = p->O,
Chris@10: 				    p->sz->dims[0].b[IB], yblock,
Chris@10: 				    p->comm, 0));
Chris@10:      if (XM(any_true)(!cldt_before, p->comm)) goto nada;	  
Chris@10:      if (ego->preserve_input || NO_DESTROY_INPUTP(plnr)) { I = O; }
Chris@10:      
Chris@10:      X(extract_reim)(p->sign, I, &ri, &ii);
Chris@10:      X(extract_reim)(p->sign, O, &ro, &io);
Chris@10: 
Chris@10:      yb = XM(block)(ny, yblock, my_pe);
Chris@10:      cld = X(mkplan_d)(plnr,
Chris@10: 		       X(mkproblem_dft_d)(X(mktensor_1d)(nx, vn*2, vn*2),
Chris@10: 					  X(mktensor_2d)(yb, vn*2*nx, vn*2*nx,
Chris@10: 							 vn, 2, 2),
Chris@10: 					  ro, io, ri, ii));
Chris@10:      if (XM(any_true)(!cld, p->comm)) goto nada;	  
Chris@10:      
Chris@10:      cldt_after = X(mkplan_d)(plnr,
Chris@10: 			      XM(mkproblem_transpose)(
Chris@10: 				   ny, nx, vn*2,
Chris@10: 				   I, O,
Chris@10: 				   yblock, p->sz->dims[0].b[OB], 
Chris@10: 				   p->comm, 0));
Chris@10:      if (XM(any_true)(!cldt_after, p->comm)) goto nada;	  
Chris@10: 
Chris@10:      pln = MKPLAN_MPI_DFT(P, &padt, apply);
Chris@10: 
Chris@10:      pln->cldt_before = cldt_before;
Chris@10:      pln->cld = cld;
Chris@10:      pln->cldt_after = cldt_after;
Chris@10:      pln->preserve_input = ego->preserve_input ? 2 : NO_DESTROY_INPUTP(plnr);
Chris@10:      pln->roff = ro - p->O;
Chris@10:      pln->ioff = io - p->O;
Chris@10:      pln->rearrange = ego->rearrange;
Chris@10: 
Chris@10:      X(ops_add)(&cldt_before->ops, &cld->ops, &pln->super.super.ops);
Chris@10:      X(ops_add2)(&cldt_after->ops, &pln->super.super.ops);
Chris@10: 
Chris@10:      return &(pln->super.super);
Chris@10: 
Chris@10:  nada:
Chris@10:      X(plan_destroy_internal)(cldt_after);
Chris@10:      X(plan_destroy_internal)(cld);
Chris@10:      X(plan_destroy_internal)(cldt_before);
Chris@10:      return (plan *) 0;
Chris@10: }
Chris@10: 
Chris@10: static solver *mksolver(rearrangement rearrange, int preserve_input)
Chris@10: {
Chris@10:      static const solver_adt sadt = { PROBLEM_MPI_DFT, mkplan, 0 };
Chris@10:      S *slv = MKSOLVER(S, &sadt);
Chris@10:      slv->rearrange = rearrange;
Chris@10:      slv->preserve_input = preserve_input;
Chris@10:      return &(slv->super);
Chris@10: }
Chris@10: 
Chris@10: void XM(dft_rank1_bigvec_register)(planner *p)
Chris@10: {
Chris@10:      rearrangement rearrange;
Chris@10:      int preserve_input;
Chris@10:      FORALL_REARRANGE(rearrange)
Chris@10: 	  for (preserve_input = 0; preserve_input <= 1; ++preserve_input)
Chris@10: 	       REGISTER_SOLVER(p, mksolver(rearrange, preserve_input));
Chris@10: }