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: /* Complex DFTs of rank >= 2, for the case where we are distributed
cannam@95:    across the first dimension only, and the output is transposed both
cannam@95:    in data distribution and in ordering (for the first 2 dimensions).
cannam@95: 
cannam@95:    (Note that we don't have to handle the case where the input is
cannam@95:    transposed, since this is equivalent to transposed output with the
cannam@95:    first two dimensions swapped, and is automatically canonicalized as
cannam@95:    such by dft-problem.c. */
cannam@95: 
cannam@95: #include "mpi-dft.h"
cannam@95: #include "mpi-transpose.h"
cannam@95: #include "dft.h"
cannam@95: 
cannam@95: typedef struct {
cannam@95:      solver super;
cannam@95:      int preserve_input; /* preserve input even if DESTROY_INPUT was passed */
cannam@95: } S;
cannam@95: 
cannam@95: typedef struct {
cannam@95:      plan_mpi_dft super;
cannam@95: 
cannam@95:      plan *cld1, *cldt, *cld2;
cannam@95:      INT roff, ioff;
cannam@95:      int preserve_input;
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 *cld1, *cld2;
cannam@95:      plan_rdft *cldt;
cannam@95:      INT roff = ego->roff, ioff = ego->ioff;
cannam@95:      
cannam@95:      /* DFT local dimensions */
cannam@95:      cld1 = (plan_dft *) ego->cld1;
cannam@95:      if (ego->preserve_input) {
cannam@95: 	  cld1->apply(ego->cld1, I+roff, I+ioff, O+roff, O+ioff);
cannam@95: 	  I = O;
cannam@95:      }
cannam@95:      else
cannam@95: 	  cld1->apply(ego->cld1, I+roff, I+ioff, I+roff, I+ioff);
cannam@95: 
cannam@95:      /* global transpose */
cannam@95:      cldt = (plan_rdft *) ego->cldt;
cannam@95:      cldt->apply(ego->cldt, I, O);
cannam@95: 
cannam@95:      /* DFT final local dimension */
cannam@95:      cld2 = (plan_dft *) ego->cld2;
cannam@95:      cld2->apply(ego->cld2, O+roff, O+ioff, O+roff, O+ioff);
cannam@95: }
cannam@95: 
cannam@95: static int applicable(const S *ego, const problem *p_,
cannam@95: 		      const planner *plnr)
cannam@95: {
cannam@95:      const problem_mpi_dft *p = (const problem_mpi_dft *) p_;
cannam@95:      return (1
cannam@95: 	     && p->sz->rnk > 1
cannam@95: 	     && p->flags == TRANSPOSED_OUT
cannam@95: 	     && (!ego->preserve_input || (!NO_DESTROY_INPUTP(plnr)
cannam@95: 					  && p->I != p->O))
cannam@95: 	     && XM(is_local_after)(1, p->sz, IB)
cannam@95: 	     && XM(is_local_after)(2, p->sz, OB)
cannam@95: 	     && XM(num_blocks)(p->sz->dims[0].n, p->sz->dims[0].b[OB]) == 1
cannam@95: 	     && (!NO_SLOWP(plnr) /* slow if dft-serial is applicable */
cannam@95: 		 || !XM(dft_serial_applicable)(p))
cannam@95: 	  );
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->cld1, wakefulness);
cannam@95:      X(plan_awake)(ego->cldt, wakefulness);
cannam@95:      X(plan_awake)(ego->cld2, 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->cld2);
cannam@95:      X(plan_destroy_internal)(ego->cldt);
cannam@95:      X(plan_destroy_internal)(ego->cld1);
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-rank-geq2-transposed%s%(%p%)%(%p%)%(%p%))", 
cannam@95: 	      ego->preserve_input==2 ?"/p":"",
cannam@95: 	      ego->cld1, ego->cldt, ego->cld2);
cannam@95: }
cannam@95: 
cannam@95: static plan *mkplan(const solver *ego_, const problem *p_, planner *plnr)
cannam@95: {
cannam@95:      const S *ego = (const S *) ego_;
cannam@95:      const problem_mpi_dft *p;
cannam@95:      P *pln;
cannam@95:      plan *cld1 = 0, *cldt = 0, *cld2 = 0;
cannam@95:      R *ri, *ii, *ro, *io, *I, *O;
cannam@95:      tensor *sz;
cannam@95:      int i, my_pe, n_pes;
cannam@95:      INT nrest;
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:      if (!applicable(ego, p_, plnr))
cannam@95:           return (plan *) 0;
cannam@95: 
cannam@95:      p = (const problem_mpi_dft *) p_;
cannam@95: 
cannam@95:      X(extract_reim)(p->sign, I = p->I, &ri, &ii);
cannam@95:      X(extract_reim)(p->sign, O = p->O, &ro, &io);
cannam@95:      if (ego->preserve_input || NO_DESTROY_INPUTP(plnr)) 
cannam@95: 	  I = O; 
cannam@95:      else { 
cannam@95: 	  ro = ri;
cannam@95: 	  io = ii;
cannam@95:      }
cannam@95:      MPI_Comm_rank(p->comm, &my_pe);
cannam@95:      MPI_Comm_size(p->comm, &n_pes);
cannam@95: 
cannam@95:      sz = X(mktensor)(p->sz->rnk - 1); /* tensor of last rnk-1 dimensions */
cannam@95:      i = p->sz->rnk - 2; A(i >= 0);
cannam@95:      sz->dims[i].n = p->sz->dims[i+1].n;
cannam@95:      sz->dims[i].is = sz->dims[i].os = 2 * p->vn;
cannam@95:      for (--i; i >= 0; --i) {
cannam@95: 	  sz->dims[i].n = p->sz->dims[i+1].n;
cannam@95: 	  sz->dims[i].is = sz->dims[i].os = sz->dims[i+1].n * sz->dims[i+1].is;
cannam@95:      }
cannam@95:      nrest = 1; for (i = 1; i < sz->rnk; ++i) nrest *= sz->dims[i].n;
cannam@95:      {
cannam@95:           INT is = sz->dims[0].n * sz->dims[0].is;
cannam@95:           INT b = XM(block)(p->sz->dims[0].n, p->sz->dims[0].b[IB], my_pe);
cannam@95: 	  cld1 = X(mkplan_d)(plnr,
cannam@95:                              X(mkproblem_dft_d)(sz,
cannam@95:                                                 X(mktensor_2d)(b, is, is,
cannam@95:                                                                p->vn, 2, 2),
cannam@95:                                                 ri, ii, ro, io));
cannam@95: 	  if (XM(any_true)(!cld1, p->comm)) goto nada;
cannam@95:      }
cannam@95: 
cannam@95:      nrest *= p->vn;
cannam@95:      cldt = X(mkplan_d)(plnr,
cannam@95: 			XM(mkproblem_transpose)(
cannam@95: 			     p->sz->dims[0].n, p->sz->dims[1].n, nrest * 2,
cannam@95: 			     I, O,
cannam@95: 			     p->sz->dims[0].b[IB], p->sz->dims[1].b[OB], 
cannam@95: 			     p->comm, 0));
cannam@95:      if (XM(any_true)(!cldt, p->comm)) goto nada;
cannam@95: 
cannam@95:      X(extract_reim)(p->sign, O, &ro, &io);
cannam@95:      {
cannam@95: 	  INT is = p->sz->dims[0].n * nrest * 2;
cannam@95: 	  INT b = XM(block)(p->sz->dims[1].n, p->sz->dims[1].b[OB], my_pe);
cannam@95: 	  cld2 = X(mkplan_d)(plnr,
cannam@95: 			     X(mkproblem_dft_d)(X(mktensor_1d)(
cannam@95: 						     p->sz->dims[0].n,
cannam@95: 						     nrest * 2, nrest * 2),
cannam@95: 						X(mktensor_2d)(b, is, is,
cannam@95: 							       nrest, 2, 2),
cannam@95: 						ro, io, ro, io));
cannam@95: 	  if (XM(any_true)(!cld2, p->comm)) goto nada;
cannam@95:      }
cannam@95: 
cannam@95:      pln = MKPLAN_MPI_DFT(P, &padt, apply);
cannam@95:      pln->cld1 = cld1;
cannam@95:      pln->cldt = cldt;
cannam@95:      pln->cld2 = cld2;
cannam@95:      pln->preserve_input = ego->preserve_input ? 2 : NO_DESTROY_INPUTP(plnr);
cannam@95:      pln->roff = ri - p->I;
cannam@95:      pln->ioff = ii - p->I;
cannam@95: 
cannam@95:      X(ops_add)(&cld1->ops, &cld2->ops, &pln->super.super.ops);
cannam@95:      X(ops_add2)(&cldt->ops, &pln->super.super.ops);
cannam@95: 
cannam@95:      return &(pln->super.super);
cannam@95: 
cannam@95:  nada:
cannam@95:      X(plan_destroy_internal)(cld2);
cannam@95:      X(plan_destroy_internal)(cldt);
cannam@95:      X(plan_destroy_internal)(cld1);
cannam@95:      return (plan *) 0;
cannam@95: }
cannam@95: 
cannam@95: static solver *mksolver(int preserve_input)
cannam@95: {
cannam@95:      static const solver_adt sadt = { PROBLEM_MPI_DFT, mkplan, 0 };
cannam@95:      S *slv = MKSOLVER(S, &sadt);
cannam@95:      slv->preserve_input = preserve_input;
cannam@95:      return &(slv->super);
cannam@95: }
cannam@95: 
cannam@95: void XM(dft_rank_geq2_transposed_register)(planner *p)
cannam@95: {
cannam@95:      int preserve_input;
cannam@95:      for (preserve_input = 0; preserve_input <= 1; ++preserve_input)
cannam@95: 	  REGISTER_SOLVER(p, mksolver(preserve_input));
cannam@95: }