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: #include "mpi-rdft2.h"
cannam@127: 
cannam@127: static void destroy(problem *ego_)
cannam@127: {
cannam@127:      problem_mpi_rdft2 *ego = (problem_mpi_rdft2 *) ego_;
cannam@127:      XM(dtensor_destroy)(ego->sz);
cannam@127:      MPI_Comm_free(&ego->comm);
cannam@127:      X(ifree)(ego_);
cannam@127: }
cannam@127: 
cannam@127: static void hash(const problem *p_, md5 *m)
cannam@127: {
cannam@127:      const problem_mpi_rdft2 *p = (const problem_mpi_rdft2 *) p_;
cannam@127:      int i;
cannam@127:      X(md5puts)(m, "mpi-rdft2");
cannam@127:      X(md5int)(m, p->I == p->O);
cannam@127:      /* don't include alignment -- may differ between processes
cannam@127: 	X(md5int)(m, X(ialignment_of)(p->I));
cannam@127: 	X(md5int)(m, X(ialignment_of)(p->O));
cannam@127: 	... note that applicability of MPI plans does not depend
cannam@127: 	    on alignment (although optimality may, in principle). */
cannam@127:      XM(dtensor_md5)(m, p->sz);
cannam@127:      X(md5INT)(m, p->vn);
cannam@127:      X(md5int)(m, p->kind);
cannam@127:      X(md5int)(m, p->flags);
cannam@127:      MPI_Comm_size(p->comm, &i); X(md5int)(m, i);
cannam@127:      A(XM(md5_equal)(*m, p->comm));
cannam@127: }
cannam@127: 
cannam@127: static void print(const problem *ego_, printer *p)
cannam@127: {
cannam@127:      const problem_mpi_rdft2 *ego = (const problem_mpi_rdft2 *) ego_;
cannam@127:      int i;
cannam@127:      p->print(p, "(mpi-rdft2 %d %d %d ", 
cannam@127: 	      ego->I == ego->O,
cannam@127: 	      X(ialignment_of)(ego->I),
cannam@127: 	      X(ialignment_of)(ego->O));
cannam@127:      XM(dtensor_print)(ego->sz, p);
cannam@127:      p->print(p, " %D %d %d", ego->vn, (int) ego->kind, ego->flags);
cannam@127:      MPI_Comm_size(ego->comm, &i); p->print(p, " %d)", i);
cannam@127: }
cannam@127: 
cannam@127: static void zero(const problem *ego_)
cannam@127: {
cannam@127:      const problem_mpi_rdft2 *ego = (const problem_mpi_rdft2 *) ego_;
cannam@127:      R *I = ego->I;
cannam@127:      dtensor *sz;
cannam@127:      INT i, N;
cannam@127:      int my_pe;
cannam@127: 
cannam@127:      sz = XM(dtensor_copy)(ego->sz);
cannam@127:      sz->dims[sz->rnk - 1].n = sz->dims[sz->rnk - 1].n / 2 + 1;
cannam@127:      MPI_Comm_rank(ego->comm, &my_pe);
cannam@127:      N = 2 * ego->vn * XM(total_block)(sz, IB, my_pe);
cannam@127:      XM(dtensor_destroy)(sz);
cannam@127:      for (i = 0; i < N; ++i) I[i] = K(0.0);
cannam@127: }
cannam@127: 
cannam@127: static const problem_adt padt =
cannam@127: {
cannam@127:      PROBLEM_MPI_RDFT2,
cannam@127:      hash,
cannam@127:      zero,
cannam@127:      print,
cannam@127:      destroy
cannam@127: };
cannam@127: 
cannam@127: problem *XM(mkproblem_rdft2)(const dtensor *sz, INT vn,
cannam@127: 			   R *I, R *O,
cannam@127: 			   MPI_Comm comm,
cannam@127: 			   rdft_kind kind,
cannam@127: 			   unsigned flags)
cannam@127: {
cannam@127:      problem_mpi_rdft2 *ego =
cannam@127:           (problem_mpi_rdft2 *)X(mkproblem)(sizeof(problem_mpi_rdft2), &padt);
cannam@127:      int n_pes;
cannam@127: 
cannam@127:      A(XM(dtensor_validp)(sz) && FINITE_RNK(sz->rnk) && sz->rnk > 1);
cannam@127:      MPI_Comm_size(comm, &n_pes);
cannam@127:      A(vn >= 0);
cannam@127:      A(kind == R2HC || kind == HC2R);
cannam@127: 
cannam@127:      /* enforce pointer equality if untainted pointers are equal */
cannam@127:      if (UNTAINT(I) == UNTAINT(O))
cannam@127: 	  I = O = JOIN_TAINT(I, O);
cannam@127: 
cannam@127:      ego->sz = XM(dtensor_canonical)(sz, 0);
cannam@127: #ifdef FFTW_DEBUG
cannam@127:      ego->sz->dims[sz->rnk - 1].n = sz->dims[sz->rnk - 1].n / 2 + 1;
cannam@127:      A(n_pes >= XM(num_blocks_total)(ego->sz, IB)
cannam@127:        && n_pes >= XM(num_blocks_total)(ego->sz, OB));
cannam@127:      ego->sz->dims[sz->rnk - 1].n = sz->dims[sz->rnk - 1].n;
cannam@127: #endif
cannam@127: 
cannam@127:      ego->vn = vn;
cannam@127:      ego->I = I;
cannam@127:      ego->O = O;
cannam@127:      ego->kind = kind;
cannam@127: 
cannam@127:      /* We only support TRANSPOSED_OUT for r2c and TRANSPOSED_IN for
cannam@127: 	c2r transforms. */
cannam@127: 
cannam@127:      ego->flags = flags;
cannam@127: 
cannam@127:      MPI_Comm_dup(comm, &ego->comm);
cannam@127: 
cannam@127:      return &(ego->super);
cannam@127: }
cannam@127: 
cannam@127: problem *XM(mkproblem_rdft2_d)(dtensor *sz, INT vn,
cannam@127: 			     R *I, R *O,
cannam@127: 			     MPI_Comm comm,
cannam@127: 			     rdft_kind kind,
cannam@127: 			     unsigned flags)
cannam@127: {
cannam@127:      problem *p = XM(mkproblem_rdft2)(sz, vn, I, O, comm, kind, flags);
cannam@127:      XM(dtensor_destroy)(sz);
cannam@127:      return p;
cannam@127: }