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-transpose.h"
cannam@127: 
cannam@127: static void destroy(problem *ego_)
cannam@127: {
cannam@127:      problem_mpi_transpose *ego = (problem_mpi_transpose *) ego_;
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_transpose *p = (const problem_mpi_transpose *) p_;
cannam@127:      int i;
cannam@127:      X(md5puts)(m, "mpi-transpose");
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:      X(md5INT)(m, p->vn);
cannam@127:      X(md5INT)(m, p->nx);
cannam@127:      X(md5INT)(m, p->ny);
cannam@127:      X(md5INT)(m, p->block);
cannam@127:      X(md5INT)(m, p->tblock);
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_transpose *ego = (const problem_mpi_transpose *) ego_;
cannam@127:      int i;
cannam@127:      MPI_Comm_size(ego->comm, &i);
cannam@127:      p->print(p, "(mpi-transpose %d %d %d %D %D %D %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: 	      ego->vn,
cannam@127: 	      ego->nx, ego->ny,
cannam@127: 	      ego->block, ego->tblock,
cannam@127: 	      i);
cannam@127: }
cannam@127: 
cannam@127: static void zero(const problem *ego_)
cannam@127: {
cannam@127:      const problem_mpi_transpose *ego = (const problem_mpi_transpose *) ego_;
cannam@127:      R *I = ego->I;
cannam@127:      INT i, N = ego->vn * ego->ny;
cannam@127:      int my_pe;
cannam@127: 
cannam@127:      MPI_Comm_rank(ego->comm, &my_pe);
cannam@127:      N *= XM(block)(ego->nx, ego->block, my_pe);
cannam@127: 
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_TRANSPOSE,
cannam@127:      hash,
cannam@127:      zero,
cannam@127:      print,
cannam@127:      destroy
cannam@127: };
cannam@127: 
cannam@127: problem *XM(mkproblem_transpose)(INT nx, INT ny, INT vn,
cannam@127: 				 R *I, R *O,
cannam@127: 				 INT block, INT tblock,
cannam@127: 				 MPI_Comm comm,
cannam@127: 				 unsigned flags)
cannam@127: {
cannam@127:      problem_mpi_transpose *ego =
cannam@127:           (problem_mpi_transpose *)X(mkproblem)(sizeof(problem_mpi_transpose), &padt);
cannam@127: 
cannam@127:      A(nx > 0 && ny > 0 && vn > 0);
cannam@127:      A(block > 0 && XM(num_blocks_ok)(nx, block, comm)
cannam@127:        && tblock > 0 && XM(num_blocks_ok)(ny, tblock, comm));
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->nx = nx;
cannam@127:      ego->ny = ny;
cannam@127:      ego->vn = vn;
cannam@127:      ego->I = I;
cannam@127:      ego->O = O;
cannam@127:      ego->block = block > nx ? nx : block;
cannam@127:      ego->tblock = tblock > ny ? ny : tblock;
cannam@127: 
cannam@127:      /* canonicalize flags: we can freely assume that the data is
cannam@127: 	"transposed" if one of the dimensions is 1. */
cannam@127:      if (ego->block == 1)
cannam@127: 	  flags |= TRANSPOSED_IN;
cannam@127:      if (ego->tblock == 1)
cannam@127: 	  flags |= TRANSPOSED_OUT;
cannam@127:      ego->flags = flags;
cannam@127: 
cannam@127:      MPI_Comm_dup(comm, &ego->comm);
cannam@127: 
cannam@127:      return &(ego->super);
cannam@127: }