Chris@42: /*
Chris@42:  * Copyright (c) 2003, 2007-14 Matteo Frigo
Chris@42:  * Copyright (c) 2003, 2007-14 Massachusetts Institute of Technology
Chris@42:  *
Chris@42:  * This program is free software; you can redistribute it and/or modify
Chris@42:  * it under the terms of the GNU General Public License as published by
Chris@42:  * the Free Software Foundation; either version 2 of the License, or
Chris@42:  * (at your option) any later version.
Chris@42:  *
Chris@42:  * This program is distributed in the hope that it will be useful,
Chris@42:  * but WITHOUT ANY WARRANTY; without even the implied warranty of
Chris@42:  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
Chris@42:  * GNU General Public License for more details.
Chris@42:  *
Chris@42:  * You should have received a copy of the GNU General Public License
Chris@42:  * along with this program; if not, write to the Free Software
Chris@42:  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA
Chris@42:  *
Chris@42:  */
Chris@42: 
Chris@42: /* Complex DFTs of rank >= 2, for the case where we are distributed
Chris@42:    across the first dimension only, and the output is transposed both
Chris@42:    in data distribution and in ordering (for the first 2 dimensions).
Chris@42: 
Chris@42:    (Note that we don't have to handle the case where the input is
Chris@42:    transposed, since this is equivalent to transposed output with the
Chris@42:    first two dimensions swapped, and is automatically canonicalized as
Chris@42:    such by dft-problem.c. */
Chris@42: 
Chris@42: #include "mpi-dft.h"
Chris@42: #include "mpi-transpose.h"
Chris@42: #include "dft.h"
Chris@42: 
Chris@42: typedef struct {
Chris@42:      solver super;
Chris@42:      int preserve_input; /* preserve input even if DESTROY_INPUT was passed */
Chris@42: } S;
Chris@42: 
Chris@42: typedef struct {
Chris@42:      plan_mpi_dft super;
Chris@42: 
Chris@42:      plan *cld1, *cldt, *cld2;
Chris@42:      INT roff, ioff;
Chris@42:      int preserve_input;
Chris@42: } P;
Chris@42: 
Chris@42: static void apply(const plan *ego_, R *I, R *O)
Chris@42: {
Chris@42:      const P *ego = (const P *) ego_;
Chris@42:      plan_dft *cld1, *cld2;
Chris@42:      plan_rdft *cldt;
Chris@42:      INT roff = ego->roff, ioff = ego->ioff;
Chris@42:      
Chris@42:      /* DFT local dimensions */
Chris@42:      cld1 = (plan_dft *) ego->cld1;
Chris@42:      if (ego->preserve_input) {
Chris@42: 	  cld1->apply(ego->cld1, I+roff, I+ioff, O+roff, O+ioff);
Chris@42: 	  I = O;
Chris@42:      }
Chris@42:      else
Chris@42: 	  cld1->apply(ego->cld1, I+roff, I+ioff, I+roff, I+ioff);
Chris@42: 
Chris@42:      /* global transpose */
Chris@42:      cldt = (plan_rdft *) ego->cldt;
Chris@42:      cldt->apply(ego->cldt, I, O);
Chris@42: 
Chris@42:      /* DFT final local dimension */
Chris@42:      cld2 = (plan_dft *) ego->cld2;
Chris@42:      cld2->apply(ego->cld2, O+roff, O+ioff, O+roff, O+ioff);
Chris@42: }
Chris@42: 
Chris@42: static int applicable(const S *ego, const problem *p_,
Chris@42: 		      const planner *plnr)
Chris@42: {
Chris@42:      const problem_mpi_dft *p = (const problem_mpi_dft *) p_;
Chris@42:      return (1
Chris@42: 	     && p->sz->rnk > 1
Chris@42: 	     && p->flags == TRANSPOSED_OUT
Chris@42: 	     && (!ego->preserve_input || (!NO_DESTROY_INPUTP(plnr)
Chris@42: 					  && p->I != p->O))
Chris@42: 	     && XM(is_local_after)(1, p->sz, IB)
Chris@42: 	     && XM(is_local_after)(2, p->sz, OB)
Chris@42: 	     && XM(num_blocks)(p->sz->dims[0].n, p->sz->dims[0].b[OB]) == 1
Chris@42: 	     && (!NO_SLOWP(plnr) /* slow if dft-serial is applicable */
Chris@42: 		 || !XM(dft_serial_applicable)(p))
Chris@42: 	  );
Chris@42: }
Chris@42: 
Chris@42: static void awake(plan *ego_, enum wakefulness wakefulness)
Chris@42: {
Chris@42:      P *ego = (P *) ego_;
Chris@42:      X(plan_awake)(ego->cld1, wakefulness);
Chris@42:      X(plan_awake)(ego->cldt, wakefulness);
Chris@42:      X(plan_awake)(ego->cld2, wakefulness);
Chris@42: }
Chris@42: 
Chris@42: static void destroy(plan *ego_)
Chris@42: {
Chris@42:      P *ego = (P *) ego_;
Chris@42:      X(plan_destroy_internal)(ego->cld2);
Chris@42:      X(plan_destroy_internal)(ego->cldt);
Chris@42:      X(plan_destroy_internal)(ego->cld1);
Chris@42: }
Chris@42: 
Chris@42: static void print(const plan *ego_, printer *p)
Chris@42: {
Chris@42:      const P *ego = (const P *) ego_;
Chris@42:      p->print(p, "(mpi-dft-rank-geq2-transposed%s%(%p%)%(%p%)%(%p%))", 
Chris@42: 	      ego->preserve_input==2 ?"/p":"",
Chris@42: 	      ego->cld1, ego->cldt, ego->cld2);
Chris@42: }
Chris@42: 
Chris@42: static plan *mkplan(const solver *ego_, const problem *p_, planner *plnr)
Chris@42: {
Chris@42:      const S *ego = (const S *) ego_;
Chris@42:      const problem_mpi_dft *p;
Chris@42:      P *pln;
Chris@42:      plan *cld1 = 0, *cldt = 0, *cld2 = 0;
Chris@42:      R *ri, *ii, *ro, *io, *I, *O;
Chris@42:      tensor *sz;
Chris@42:      int i, my_pe, n_pes;
Chris@42:      INT nrest;
Chris@42:      static const plan_adt padt = {
Chris@42:           XM(dft_solve), awake, print, destroy
Chris@42:      };
Chris@42: 
Chris@42:      UNUSED(ego);
Chris@42: 
Chris@42:      if (!applicable(ego, p_, plnr))
Chris@42:           return (plan *) 0;
Chris@42: 
Chris@42:      p = (const problem_mpi_dft *) p_;
Chris@42: 
Chris@42:      X(extract_reim)(p->sign, I = p->I, &ri, &ii);
Chris@42:      X(extract_reim)(p->sign, O = p->O, &ro, &io);
Chris@42:      if (ego->preserve_input || NO_DESTROY_INPUTP(plnr)) 
Chris@42: 	  I = O; 
Chris@42:      else { 
Chris@42: 	  ro = ri;
Chris@42: 	  io = ii;
Chris@42:      }
Chris@42:      MPI_Comm_rank(p->comm, &my_pe);
Chris@42:      MPI_Comm_size(p->comm, &n_pes);
Chris@42: 
Chris@42:      sz = X(mktensor)(p->sz->rnk - 1); /* tensor of last rnk-1 dimensions */
Chris@42:      i = p->sz->rnk - 2; A(i >= 0);
Chris@42:      sz->dims[i].n = p->sz->dims[i+1].n;
Chris@42:      sz->dims[i].is = sz->dims[i].os = 2 * p->vn;
Chris@42:      for (--i; i >= 0; --i) {
Chris@42: 	  sz->dims[i].n = p->sz->dims[i+1].n;
Chris@42: 	  sz->dims[i].is = sz->dims[i].os = sz->dims[i+1].n * sz->dims[i+1].is;
Chris@42:      }
Chris@42:      nrest = 1; for (i = 1; i < sz->rnk; ++i) nrest *= sz->dims[i].n;
Chris@42:      {
Chris@42:           INT is = sz->dims[0].n * sz->dims[0].is;
Chris@42:           INT b = XM(block)(p->sz->dims[0].n, p->sz->dims[0].b[IB], my_pe);
Chris@42: 	  cld1 = X(mkplan_d)(plnr,
Chris@42:                              X(mkproblem_dft_d)(sz,
Chris@42:                                                 X(mktensor_2d)(b, is, is,
Chris@42:                                                                p->vn, 2, 2),
Chris@42:                                                 ri, ii, ro, io));
Chris@42: 	  if (XM(any_true)(!cld1, p->comm)) goto nada;
Chris@42:      }
Chris@42: 
Chris@42:      nrest *= p->vn;
Chris@42:      cldt = X(mkplan_d)(plnr,
Chris@42: 			XM(mkproblem_transpose)(
Chris@42: 			     p->sz->dims[0].n, p->sz->dims[1].n, nrest * 2,
Chris@42: 			     I, O,
Chris@42: 			     p->sz->dims[0].b[IB], p->sz->dims[1].b[OB], 
Chris@42: 			     p->comm, 0));
Chris@42:      if (XM(any_true)(!cldt, p->comm)) goto nada;
Chris@42: 
Chris@42:      X(extract_reim)(p->sign, O, &ro, &io);
Chris@42:      {
Chris@42: 	  INT is = p->sz->dims[0].n * nrest * 2;
Chris@42: 	  INT b = XM(block)(p->sz->dims[1].n, p->sz->dims[1].b[OB], my_pe);
Chris@42: 	  cld2 = X(mkplan_d)(plnr,
Chris@42: 			     X(mkproblem_dft_d)(X(mktensor_1d)(
Chris@42: 						     p->sz->dims[0].n,
Chris@42: 						     nrest * 2, nrest * 2),
Chris@42: 						X(mktensor_2d)(b, is, is,
Chris@42: 							       nrest, 2, 2),
Chris@42: 						ro, io, ro, io));
Chris@42: 	  if (XM(any_true)(!cld2, p->comm)) goto nada;
Chris@42:      }
Chris@42: 
Chris@42:      pln = MKPLAN_MPI_DFT(P, &padt, apply);
Chris@42:      pln->cld1 = cld1;
Chris@42:      pln->cldt = cldt;
Chris@42:      pln->cld2 = cld2;
Chris@42:      pln->preserve_input = ego->preserve_input ? 2 : NO_DESTROY_INPUTP(plnr);
Chris@42:      pln->roff = ri - p->I;
Chris@42:      pln->ioff = ii - p->I;
Chris@42: 
Chris@42:      X(ops_add)(&cld1->ops, &cld2->ops, &pln->super.super.ops);
Chris@42:      X(ops_add2)(&cldt->ops, &pln->super.super.ops);
Chris@42: 
Chris@42:      return &(pln->super.super);
Chris@42: 
Chris@42:  nada:
Chris@42:      X(plan_destroy_internal)(cld2);
Chris@42:      X(plan_destroy_internal)(cldt);
Chris@42:      X(plan_destroy_internal)(cld1);
Chris@42:      return (plan *) 0;
Chris@42: }
Chris@42: 
Chris@42: static solver *mksolver(int preserve_input)
Chris@42: {
Chris@42:      static const solver_adt sadt = { PROBLEM_MPI_DFT, mkplan, 0 };
Chris@42:      S *slv = MKSOLVER(S, &sadt);
Chris@42:      slv->preserve_input = preserve_input;
Chris@42:      return &(slv->super);
Chris@42: }
Chris@42: 
Chris@42: void XM(dft_rank_geq2_transposed_register)(planner *p)
Chris@42: {
Chris@42:      int preserve_input;
Chris@42:      for (preserve_input = 0; preserve_input <= 1; ++preserve_input)
Chris@42: 	  REGISTER_SOLVER(p, mksolver(preserve_input));
Chris@42: }