cannam@127: /*
cannam@127:  * Copyright (c) 2003, 2007-14 Matteo Frigo
cannam@127:  * Copyright (c) 2003, 2007-14 Massachusetts Institute of Technology
cannam@127:  *
cannam@127:  * This program is free software; you can redistribute it and/or modify
cannam@127:  * it under the terms of the GNU General Public License as published by
cannam@127:  * the Free Software Foundation; either version 2 of the License, or
cannam@127:  * (at your option) any later version.
cannam@127:  *
cannam@127:  * This program is distributed in the hope that it will be useful,
cannam@127:  * but WITHOUT ANY WARRANTY; without even the implied warranty of
cannam@127:  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
cannam@127:  * GNU General Public License for more details.
cannam@127:  *
cannam@127:  * You should have received a copy of the GNU General Public License
cannam@127:  * along with this program; if not, write to the Free Software
cannam@127:  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA
cannam@127:  *
cannam@127:  */
cannam@127: 
cannam@127: /* Real-input (r2c) DFTs of rank >= 2, for the case where we are distributed
cannam@127:    across the first dimension only, and the output is transposed both
cannam@127:    in data distribution and in ordering (for the first 2 dimensions).
cannam@127: 
cannam@127:    Conversely, real-output (c2r) DFTs where the input is transposed.
cannam@127: 
cannam@127:    We don't currently support transposed-input r2c or transposed-output
cannam@127:    c2r transforms. */
cannam@127: 
cannam@127: #include "mpi-rdft2.h"
cannam@127: #include "mpi-transpose.h"
cannam@127: #include "rdft.h"
cannam@127: #include "dft.h"
cannam@127: 
cannam@127: typedef struct {
cannam@127:      solver super;
cannam@127:      int preserve_input; /* preserve input even if DESTROY_INPUT was passed */
cannam@127: } S;
cannam@127: 
cannam@127: typedef struct {
cannam@127:      plan_mpi_rdft2 super;
cannam@127: 
cannam@127:      plan *cld1, *cldt, *cld2;
cannam@127:      INT vn;
cannam@127:      int preserve_input;
cannam@127: } P;
cannam@127: 
cannam@127: static void apply_r2c(const plan *ego_, R *I, R *O)
cannam@127: {
cannam@127:      const P *ego = (const P *) ego_;
cannam@127:      plan_rdft2 *cld1;
cannam@127:      plan_dft *cld2;
cannam@127:      plan_rdft *cldt;
cannam@127:      
cannam@127:      /* RDFT2 local dimensions */
cannam@127:      cld1 = (plan_rdft2 *) ego->cld1;
cannam@127:      if (ego->preserve_input) {
cannam@127: 	  cld1->apply(ego->cld1, I, I+ego->vn, O, O+1);
cannam@127: 	  I = O;
cannam@127:      }
cannam@127:      else
cannam@127: 	  cld1->apply(ego->cld1, I, I+ego->vn, I, I+1);
cannam@127: 
cannam@127:      /* global transpose */
cannam@127:      cldt = (plan_rdft *) ego->cldt;
cannam@127:      cldt->apply(ego->cldt, I, O);
cannam@127: 
cannam@127:      /* DFT final local dimension */
cannam@127:      cld2 = (plan_dft *) ego->cld2;
cannam@127:      cld2->apply(ego->cld2, O, O+1, O, O+1);
cannam@127: }
cannam@127: 
cannam@127: static void apply_c2r(const plan *ego_, R *I, R *O)
cannam@127: {
cannam@127:      const P *ego = (const P *) ego_;
cannam@127:      plan_rdft2 *cld1;
cannam@127:      plan_dft *cld2;
cannam@127:      plan_rdft *cldt;
cannam@127:      
cannam@127:      /* IDFT local dimensions */
cannam@127:      cld2 = (plan_dft *) ego->cld2;
cannam@127:      if (ego->preserve_input) {
cannam@127: 	  cld2->apply(ego->cld2, I+1, I, O+1, O);
cannam@127: 	  I = O;
cannam@127:      }
cannam@127:      else
cannam@127: 	  cld2->apply(ego->cld2, I+1, I, I+1, I);
cannam@127: 
cannam@127:      /* global transpose */
cannam@127:      cldt = (plan_rdft *) ego->cldt;
cannam@127:      cldt->apply(ego->cldt, I, O);
cannam@127: 
cannam@127:      /* RDFT2 final local dimension */
cannam@127:      cld1 = (plan_rdft2 *) ego->cld1;
cannam@127:      cld1->apply(ego->cld1, O, O+ego->vn, O, O+1);
cannam@127: }
cannam@127: 
cannam@127: static int applicable(const S *ego, const problem *p_,
cannam@127: 		      const planner *plnr)
cannam@127: {
cannam@127:      const problem_mpi_rdft2 *p = (const problem_mpi_rdft2 *) p_;
cannam@127:      return (1
cannam@127: 	     && p->sz->rnk > 1
cannam@127: 	     && (!ego->preserve_input || (!NO_DESTROY_INPUTP(plnr)
cannam@127: 					  && p->I != p->O))
cannam@127: 	     && ((p->flags == TRANSPOSED_OUT && p->kind == R2HC
cannam@127: 		  && XM(is_local_after)(1, p->sz, IB)
cannam@127: 		  && XM(is_local_after)(2, p->sz, OB)
cannam@127: 		  && XM(num_blocks)(p->sz->dims[0].n, 
cannam@127: 				    p->sz->dims[0].b[OB]) == 1)
cannam@127: 		 || 
cannam@127: 		 (p->flags == TRANSPOSED_IN && p->kind == HC2R
cannam@127: 		  && XM(is_local_after)(1, p->sz, OB)
cannam@127: 		  && XM(is_local_after)(2, p->sz, IB)
cannam@127: 		  && XM(num_blocks)(p->sz->dims[0].n, 
cannam@127: 				    p->sz->dims[0].b[IB]) == 1))
cannam@127: 	     && (!NO_SLOWP(plnr) /* slow if rdft2-serial is applicable */
cannam@127: 		 || !XM(rdft2_serial_applicable)(p))
cannam@127: 	  );
cannam@127: }
cannam@127: 
cannam@127: static void awake(plan *ego_, enum wakefulness wakefulness)
cannam@127: {
cannam@127:      P *ego = (P *) ego_;
cannam@127:      X(plan_awake)(ego->cld1, wakefulness);
cannam@127:      X(plan_awake)(ego->cldt, wakefulness);
cannam@127:      X(plan_awake)(ego->cld2, wakefulness);
cannam@127: }
cannam@127: 
cannam@127: static void destroy(plan *ego_)
cannam@127: {
cannam@127:      P *ego = (P *) ego_;
cannam@127:      X(plan_destroy_internal)(ego->cld2);
cannam@127:      X(plan_destroy_internal)(ego->cldt);
cannam@127:      X(plan_destroy_internal)(ego->cld1);
cannam@127: }
cannam@127: 
cannam@127: static void print(const plan *ego_, printer *p)
cannam@127: {
cannam@127:      const P *ego = (const P *) ego_;
cannam@127:      p->print(p, "(mpi-rdft2-rank-geq2-transposed%s%(%p%)%(%p%)%(%p%))", 
cannam@127: 	      ego->preserve_input==2 ?"/p":"",
cannam@127: 	      ego->cld1, ego->cldt, ego->cld2);
cannam@127: }
cannam@127: 
cannam@127: static plan *mkplan(const solver *ego_, const problem *p_, planner *plnr)
cannam@127: {
cannam@127:      const S *ego = (const S *) ego_;
cannam@127:      const problem_mpi_rdft2 *p;
cannam@127:      P *pln;
cannam@127:      plan *cld1 = 0, *cldt = 0, *cld2 = 0;
cannam@127:      R *r0, *r1, *cr, *ci, *ri, *ii, *ro, *io, *I, *O;
cannam@127:      tensor *sz;
cannam@127:      int i, my_pe, n_pes;
cannam@127:      INT nrest, n1, b1;
cannam@127:      static const plan_adt padt = {
cannam@127:           XM(rdft2_solve), awake, print, destroy
cannam@127:      };
cannam@127:      block_kind k1, k2;
cannam@127: 
cannam@127:      UNUSED(ego);
cannam@127: 
cannam@127:      if (!applicable(ego, p_, plnr))
cannam@127:           return (plan *) 0;
cannam@127: 
cannam@127:      p = (const problem_mpi_rdft2 *) p_;
cannam@127: 
cannam@127:      I = p->I; O = p->O;
cannam@127:      if (p->kind == R2HC) {
cannam@127: 	  k1 = IB; k2 = OB;
cannam@127:           r1 = (r0 = I) + p->vn;
cannam@127: 	  if (ego->preserve_input || NO_DESTROY_INPUTP(plnr)) {
cannam@127: 	       ci = (cr = O) + 1;
cannam@127: 	       I = O; 
cannam@127: 	  }
cannam@127: 	  else 
cannam@127: 	       ci = (cr = I) + 1;
cannam@127: 	  io = ii = (ro = ri = O) + 1;
cannam@127:      }
cannam@127:      else {
cannam@127: 	  k1 = OB; k2 = IB;
cannam@127: 	  r1 = (r0 = O) + p->vn;
cannam@127: 	  ci = (cr = O) + 1;
cannam@127: 	  if (ego->preserve_input || NO_DESTROY_INPUTP(plnr)) {
cannam@127: 	       ri = (ii = I) + 1;
cannam@127: 	       ro = (io = O) + 1;
cannam@127: 	       I = O;
cannam@127: 	  }
cannam@127: 	  else
cannam@127: 	       ro = ri = (io = ii = I) + 1;
cannam@127:      }
cannam@127: 
cannam@127:      MPI_Comm_rank(p->comm, &my_pe);
cannam@127:      MPI_Comm_size(p->comm, &n_pes);
cannam@127: 
cannam@127:      sz = X(mktensor)(p->sz->rnk - 1); /* tensor of last rnk-1 dimensions */
cannam@127:      i = p->sz->rnk - 2; A(i >= 0);
cannam@127:      sz->dims[i].n = p->sz->dims[i+1].n / 2 + 1;
cannam@127:      sz->dims[i].is = sz->dims[i].os = 2 * p->vn;
cannam@127:      for (--i; i >= 0; --i) {
cannam@127: 	  sz->dims[i].n = p->sz->dims[i+1].n;
cannam@127: 	  sz->dims[i].is = sz->dims[i].os = sz->dims[i+1].n * sz->dims[i+1].is;
cannam@127:      }
cannam@127:      nrest = 1; for (i = 1; i < sz->rnk; ++i) nrest *= sz->dims[i].n;
cannam@127:      {
cannam@127: 	  INT ivs = 1 + (p->kind == HC2R), ovs = 1 + (p->kind == R2HC);
cannam@127:           INT is = sz->dims[0].n * sz->dims[0].is;
cannam@127:           INT b = XM(block)(p->sz->dims[0].n, p->sz->dims[0].b[k1], my_pe);
cannam@127: 	  sz->dims[p->sz->rnk - 2].n = p->sz->dims[p->sz->rnk - 1].n;
cannam@127: 	  cld1 = X(mkplan_d)(plnr,
cannam@127:                              X(mkproblem_rdft2_d)(sz,
cannam@127: 						  X(mktensor_2d)(b, is, is,
cannam@127: 								p->vn,ivs,ovs),
cannam@127: 						  r0, r1, cr, ci, p->kind));
cannam@127: 	  if (XM(any_true)(!cld1, p->comm)) goto nada;
cannam@127:      }
cannam@127: 
cannam@127:      nrest *= p->vn;
cannam@127:      n1 = p->sz->dims[1].n;
cannam@127:      b1 = p->sz->dims[1].b[k2];
cannam@127:      if (p->sz->rnk == 2) { /* n1 dimension is cut in ~half */
cannam@127: 	  n1 = n1 / 2 + 1;
cannam@127: 	  b1 = b1 == p->sz->dims[1].n ? n1 : b1;
cannam@127:      }
cannam@127: 
cannam@127:      if (p->kind == R2HC)
cannam@127: 	  cldt = X(mkplan_d)(plnr,
cannam@127: 			     XM(mkproblem_transpose)(
cannam@127: 				  p->sz->dims[0].n, n1, nrest * 2,
cannam@127: 				  I, O,
cannam@127: 				  p->sz->dims[0].b[IB], b1,
cannam@127: 				  p->comm, 0));
cannam@127:      else
cannam@127: 	  cldt = X(mkplan_d)(plnr,
cannam@127: 			     XM(mkproblem_transpose)(
cannam@127: 				  n1, p->sz->dims[0].n, nrest * 2,
cannam@127: 				  I, O,
cannam@127: 				  b1, p->sz->dims[0].b[OB], 
cannam@127: 				  p->comm, 0));
cannam@127:      if (XM(any_true)(!cldt, p->comm)) goto nada;
cannam@127: 
cannam@127:      {
cannam@127: 	  INT is = p->sz->dims[0].n * nrest * 2;
cannam@127: 	  INT b = XM(block)(n1, b1, my_pe);
cannam@127: 	  cld2 = X(mkplan_d)(plnr,
cannam@127: 			     X(mkproblem_dft_d)(X(mktensor_1d)(
cannam@127: 						     p->sz->dims[0].n,
cannam@127: 						     nrest * 2, nrest * 2),
cannam@127: 						X(mktensor_2d)(b, is, is,
cannam@127: 							       nrest, 2, 2),
cannam@127: 						ri, ii, ro, io));
cannam@127: 	  if (XM(any_true)(!cld2, p->comm)) goto nada;
cannam@127:      }
cannam@127: 
cannam@127:      pln = MKPLAN_MPI_RDFT2(P, &padt, p->kind == R2HC ? apply_r2c : apply_c2r);
cannam@127:      pln->cld1 = cld1;
cannam@127:      pln->cldt = cldt;
cannam@127:      pln->cld2 = cld2;
cannam@127:      pln->preserve_input = ego->preserve_input ? 2 : NO_DESTROY_INPUTP(plnr);
cannam@127:      pln->vn = p->vn;
cannam@127: 
cannam@127:      X(ops_add)(&cld1->ops, &cld2->ops, &pln->super.super.ops);
cannam@127:      X(ops_add2)(&cldt->ops, &pln->super.super.ops);
cannam@127: 
cannam@127:      return &(pln->super.super);
cannam@127: 
cannam@127:  nada:
cannam@127:      X(plan_destroy_internal)(cld2);
cannam@127:      X(plan_destroy_internal)(cldt);
cannam@127:      X(plan_destroy_internal)(cld1);
cannam@127:      return (plan *) 0;
cannam@127: }
cannam@127: 
cannam@127: static solver *mksolver(int preserve_input)
cannam@127: {
cannam@127:      static const solver_adt sadt = { PROBLEM_MPI_RDFT2, mkplan, 0 };
cannam@127:      S *slv = MKSOLVER(S, &sadt);
cannam@127:      slv->preserve_input = preserve_input;
cannam@127:      return &(slv->super);
cannam@127: }
cannam@127: 
cannam@127: void XM(rdft2_rank_geq2_transposed_register)(planner *p)
cannam@127: {
cannam@127:      int preserve_input;
cannam@127:      for (preserve_input = 0; preserve_input <= 1; ++preserve_input)
cannam@127: 	  REGISTER_SOLVER(p, mksolver(preserve_input));
cannam@127: }