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 RDFTs of rank >= 2, for the case where we are distributed
Chris@10:    across the first dimension only, and the output is transposed both
Chris@10:    in data distribution and in ordering (for the first 2 dimensions).
Chris@10: 
Chris@10:    (Note that we don't have to handle the case where the input is
Chris@10:    transposed, since this is equivalent to transposed output with the
Chris@10:    first two dimensions swapped, and is automatically canonicalized as
Chris@10:    such by rdft-problem.c. */
Chris@10: 
Chris@10: #include "mpi-rdft.h"
Chris@10: #include "mpi-transpose.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: } S;
Chris@10: 
Chris@10: typedef struct {
Chris@10:      plan_mpi_rdft super;
Chris@10: 
Chris@10:      plan *cld1, *cldt, *cld2;
Chris@10:      INT roff, ioff;
Chris@10:      int preserve_input;
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_rdft *cld1, *cld2, *cldt;
Chris@10:      
Chris@10:      /* RDFT local dimensions */
Chris@10:      cld1 = (plan_rdft *) ego->cld1;
Chris@10:      if (ego->preserve_input) {
Chris@10: 	  cld1->apply(ego->cld1, I, O);
Chris@10: 	  I = O;
Chris@10:      }
Chris@10:      else
Chris@10: 	  cld1->apply(ego->cld1, I, I);
Chris@10: 
Chris@10:      /* global transpose */
Chris@10:      cldt = (plan_rdft *) ego->cldt;
Chris@10:      cldt->apply(ego->cldt, I, O);
Chris@10: 
Chris@10:      /* RDFT final local dimension */
Chris@10:      cld2 = (plan_rdft *) ego->cld2;
Chris@10:      cld2->apply(ego->cld2, O, 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_rdft *p = (const problem_mpi_rdft *) p_;
Chris@10:      return (1
Chris@10: 	     && p->sz->rnk > 1
Chris@10: 	     && p->flags == TRANSPOSED_OUT
Chris@10: 	     && (!ego->preserve_input || (!NO_DESTROY_INPUTP(plnr)
Chris@10: 					  && p->I != p->O))
Chris@10: 	     && XM(is_local_after)(1, p->sz, IB)
Chris@10: 	     && XM(is_local_after)(2, p->sz, OB)
Chris@10: 	     && XM(num_blocks)(p->sz->dims[0].n, p->sz->dims[0].b[OB]) == 1
Chris@10: 	     && (!NO_SLOWP(plnr) /* slow if rdft-serial is applicable */
Chris@10: 		 || !XM(rdft_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->cld1, wakefulness);
Chris@10:      X(plan_awake)(ego->cldt, wakefulness);
Chris@10:      X(plan_awake)(ego->cld2, 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->cld2);
Chris@10:      X(plan_destroy_internal)(ego->cldt);
Chris@10:      X(plan_destroy_internal)(ego->cld1);
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:      p->print(p, "(mpi-rdft-rank-geq2-transposed%s%(%p%)%(%p%)%(%p%))", 
Chris@10: 	      ego->preserve_input==2 ?"/p":"",
Chris@10: 	      ego->cld1, ego->cldt, ego->cld2);
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_rdft *p;
Chris@10:      P *pln;
Chris@10:      plan *cld1 = 0, *cldt = 0, *cld2 = 0;
Chris@10:      R *I, *O, *I2;
Chris@10:      tensor *sz;
Chris@10:      int i, my_pe, n_pes;
Chris@10:      INT nrest;
Chris@10:      static const plan_adt padt = {
Chris@10:           XM(rdft_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_rdft *) p_;
Chris@10: 
Chris@10:      I2 = I = p->I;
Chris@10:      O = p->O;
Chris@10:      if (ego->preserve_input || NO_DESTROY_INPUTP(plnr)) 
Chris@10: 	  I = O; 
Chris@10:      MPI_Comm_rank(p->comm, &my_pe);
Chris@10:      MPI_Comm_size(p->comm, &n_pes);
Chris@10: 
Chris@10:      sz = X(mktensor)(p->sz->rnk - 1); /* tensor of last rnk-1 dimensions */
Chris@10:      i = p->sz->rnk - 2; A(i >= 0);
Chris@10:      sz->dims[i].n = p->sz->dims[i+1].n;
Chris@10:      sz->dims[i].is = sz->dims[i].os = p->vn;
Chris@10:      for (--i; i >= 0; --i) {
Chris@10: 	  sz->dims[i].n = p->sz->dims[i+1].n;
Chris@10: 	  sz->dims[i].is = sz->dims[i].os = sz->dims[i+1].n * sz->dims[i+1].is;
Chris@10:      }
Chris@10:      nrest = 1; for (i = 1; i < sz->rnk; ++i) nrest *= sz->dims[i].n;
Chris@10:      {
Chris@10:           INT is = sz->dims[0].n * sz->dims[0].is;
Chris@10:           INT b = XM(block)(p->sz->dims[0].n, p->sz->dims[0].b[IB], my_pe);
Chris@10: 	  cld1 = X(mkplan_d)(plnr,
Chris@10:                              X(mkproblem_rdft_d)(sz,
Chris@10: 						 X(mktensor_2d)(b, is, is,
Chris@10: 								p->vn, 1, 1),
Chris@10: 						 I2, I, p->kind + 1));
Chris@10: 	  if (XM(any_true)(!cld1, p->comm)) goto nada;
Chris@10:      }
Chris@10: 
Chris@10:      nrest *= p->vn;
Chris@10:      cldt = X(mkplan_d)(plnr,
Chris@10: 			XM(mkproblem_transpose)(
Chris@10: 			     p->sz->dims[0].n, p->sz->dims[1].n, nrest,
Chris@10: 			     I, O,
Chris@10: 			     p->sz->dims[0].b[IB], p->sz->dims[1].b[OB], 
Chris@10: 			     p->comm, 0));
Chris@10:      if (XM(any_true)(!cldt, p->comm)) goto nada;
Chris@10: 
Chris@10:      {
Chris@10: 	  INT is = p->sz->dims[0].n * nrest;
Chris@10: 	  INT b = XM(block)(p->sz->dims[1].n, p->sz->dims[1].b[OB], my_pe);
Chris@10: 	  cld2 = X(mkplan_d)(plnr,
Chris@10: 			     X(mkproblem_rdft_1_d)(X(mktensor_1d)(
Chris@10: 							p->sz->dims[0].n,
Chris@10: 							nrest, nrest),
Chris@10: 						   X(mktensor_2d)(b, is, is,
Chris@10: 								  nrest, 1, 1),
Chris@10: 						   O, O, p->kind[0]));
Chris@10: 	  if (XM(any_true)(!cld2, p->comm)) goto nada;
Chris@10:      }
Chris@10: 
Chris@10:      pln = MKPLAN_MPI_RDFT(P, &padt, apply);
Chris@10:      pln->cld1 = cld1;
Chris@10:      pln->cldt = cldt;
Chris@10:      pln->cld2 = cld2;
Chris@10:      pln->preserve_input = ego->preserve_input ? 2 : NO_DESTROY_INPUTP(plnr);
Chris@10: 
Chris@10:      X(ops_add)(&cld1->ops, &cld2->ops, &pln->super.super.ops);
Chris@10:      X(ops_add2)(&cldt->ops, &pln->super.super.ops);
Chris@10: 
Chris@10:      return &(pln->super.super);
Chris@10: 
Chris@10:  nada:
Chris@10:      X(plan_destroy_internal)(cld2);
Chris@10:      X(plan_destroy_internal)(cldt);
Chris@10:      X(plan_destroy_internal)(cld1);
Chris@10:      return (plan *) 0;
Chris@10: }
Chris@10: 
Chris@10: static solver *mksolver(int preserve_input)
Chris@10: {
Chris@10:      static const solver_adt sadt = { PROBLEM_MPI_RDFT, mkplan, 0 };
Chris@10:      S *slv = MKSOLVER(S, &sadt);
Chris@10:      slv->preserve_input = preserve_input;
Chris@10:      return &(slv->super);
Chris@10: }
Chris@10: 
Chris@10: void XM(rdft_rank_geq2_transposed_register)(planner *p)
Chris@10: {
Chris@10:      int preserve_input;
Chris@10:      for (preserve_input = 0; preserve_input <= 1; ++preserve_input)
Chris@10: 	  REGISTER_SOLVER(p, mksolver(preserve_input));
Chris@10: }