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 RDFTs of rank >= 2, for the case where we are distributed
cannam@95:    across the first dimension only, and the output is not transposed. */
cannam@95: 
cannam@95: #include "mpi-rdft.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_rdft super;
cannam@95: 
cannam@95:      plan *cld1, *cld2;
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_rdft *cld1, *cld2;
cannam@95:      
cannam@95:      /* RDFT local dimensions */
cannam@95:      cld1 = (plan_rdft *) ego->cld1;
cannam@95:      if (ego->preserve_input) {
cannam@95: 	  cld1->apply(ego->cld1, I, O);
cannam@95: 	  I = O;
cannam@95:      }
cannam@95:      else
cannam@95: 	  cld1->apply(ego->cld1, I, I);
cannam@95: 
cannam@95:      /* RDFT non-local dimension (via rdft-rank1-bigvec, usually): */
cannam@95:      cld2 = (plan_rdft *) ego->cld2;
cannam@95:      cld2->apply(ego->cld2, I, O);
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_rdft *p = (const problem_mpi_rdft *) p_;
cannam@95:      return (1
cannam@95: 	     && p->sz->rnk > 1
cannam@95: 	     && p->flags == 0 /* TRANSPOSED/SCRAMBLED_IN/OUT not supported */
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)(1, p->sz, OB)
cannam@95: 	     && (!NO_SLOWP(plnr) /* slow if rdft-serial is applicable */
cannam@95: 		 || !XM(rdft_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->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->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-rdft-rank-geq2%s%(%p%)%(%p%))", 
cannam@95: 	      ego->preserve_input==2 ?"/p":"", ego->cld1, 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_rdft *p;
cannam@95:      P *pln;
cannam@95:      plan *cld1 = 0, *cld2 = 0;
cannam@95:      R *I, *O, *I2;
cannam@95:      tensor *sz;
cannam@95:      dtensor *sz2;
cannam@95:      int i, my_pe, n_pes;
cannam@95:      INT nrest;
cannam@95:      static const plan_adt padt = {
cannam@95:           XM(rdft_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_rdft *) p_;
cannam@95: 
cannam@95:      I2 = I = p->I;
cannam@95:      O = p->O;
cannam@95:      if (ego->preserve_input || NO_DESTROY_INPUTP(plnr)) 
cannam@95: 	  I = O; 
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 = 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 = X(tensor_sz)(sz);
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_rdft_d)(sz,
cannam@95: 						 X(mktensor_2d)(b, is, is,
cannam@95: 								p->vn, 1, 1),
cannam@95: 						 I2, I, p->kind + 1));
cannam@95: 	  if (XM(any_true)(!cld1, p->comm)) goto nada;
cannam@95:      }
cannam@95: 
cannam@95:      sz2 = XM(mkdtensor)(1); /* tensor for first (distributed) dimension */
cannam@95:      sz2->dims[0] = p->sz->dims[0];
cannam@95:      cld2 = X(mkplan_d)(plnr, XM(mkproblem_rdft_d)(sz2, nrest * p->vn,
cannam@95: 						   I, O,
cannam@95: 						   p->comm, p->kind,
cannam@95: 						   RANK1_BIGVEC_ONLY));
cannam@95:      if (XM(any_true)(!cld2, p->comm)) goto nada;
cannam@95: 
cannam@95:      pln = MKPLAN_MPI_RDFT(P, &padt, apply);
cannam@95:      pln->cld1 = cld1;
cannam@95:      pln->cld2 = cld2;
cannam@95:      pln->preserve_input = ego->preserve_input ? 2 : NO_DESTROY_INPUTP(plnr);
cannam@95: 
cannam@95:      X(ops_add)(&cld1->ops, &cld2->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)(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_RDFT, 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(rdft_rank_geq2_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: }