Chris@42: /*
Chris@42:  * Copyright (c) 2003, 2007-14 Matteo Frigo
Chris@42:  * Copyright (c) 2003, 2007-14 Massachusetts Institute of Technology
Chris@42:  *
Chris@42:  * This program is free software; you can redistribute it and/or modify
Chris@42:  * it under the terms of the GNU General Public License as published by
Chris@42:  * the Free Software Foundation; either version 2 of the License, or
Chris@42:  * (at your option) any later version.
Chris@42:  *
Chris@42:  * This program is distributed in the hope that it will be useful,
Chris@42:  * but WITHOUT ANY WARRANTY; without even the implied warranty of
Chris@42:  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
Chris@42:  * GNU General Public License for more details.
Chris@42:  *
Chris@42:  * You should have received a copy of the GNU General Public License
Chris@42:  * along with this program; if not, write to the Free Software
Chris@42:  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA
Chris@42:  *
Chris@42:  */
Chris@42: 
Chris@42: #include "mpi-transpose.h"
Chris@42: 
Chris@42: static void destroy(problem *ego_)
Chris@42: {
Chris@42:      problem_mpi_transpose *ego = (problem_mpi_transpose *) ego_;
Chris@42:      MPI_Comm_free(&ego->comm);
Chris@42:      X(ifree)(ego_);
Chris@42: }
Chris@42: 
Chris@42: static void hash(const problem *p_, md5 *m)
Chris@42: {
Chris@42:      const problem_mpi_transpose *p = (const problem_mpi_transpose *) p_;
Chris@42:      int i;
Chris@42:      X(md5puts)(m, "mpi-transpose");
Chris@42:      X(md5int)(m, p->I == p->O);
Chris@42:      /* don't include alignment -- may differ between processes
Chris@42: 	X(md5int)(m, X(ialignment_of)(p->I));
Chris@42: 	X(md5int)(m, X(ialignment_of)(p->O));
Chris@42: 	... note that applicability of MPI plans does not depend
Chris@42: 	    on alignment (although optimality may, in principle). */
Chris@42:      X(md5INT)(m, p->vn);
Chris@42:      X(md5INT)(m, p->nx);
Chris@42:      X(md5INT)(m, p->ny);
Chris@42:      X(md5INT)(m, p->block);
Chris@42:      X(md5INT)(m, p->tblock);
Chris@42:      MPI_Comm_size(p->comm, &i); X(md5int)(m, i);
Chris@42:      A(XM(md5_equal)(*m, p->comm));
Chris@42: }
Chris@42: 
Chris@42: static void print(const problem *ego_, printer *p)
Chris@42: {
Chris@42:      const problem_mpi_transpose *ego = (const problem_mpi_transpose *) ego_;
Chris@42:      int i;
Chris@42:      MPI_Comm_size(ego->comm, &i);
Chris@42:      p->print(p, "(mpi-transpose %d %d %d %D %D %D %D %D %d)", 
Chris@42: 	      ego->I == ego->O,
Chris@42: 	      X(ialignment_of)(ego->I),
Chris@42: 	      X(ialignment_of)(ego->O),
Chris@42: 	      ego->vn,
Chris@42: 	      ego->nx, ego->ny,
Chris@42: 	      ego->block, ego->tblock,
Chris@42: 	      i);
Chris@42: }
Chris@42: 
Chris@42: static void zero(const problem *ego_)
Chris@42: {
Chris@42:      const problem_mpi_transpose *ego = (const problem_mpi_transpose *) ego_;
Chris@42:      R *I = ego->I;
Chris@42:      INT i, N = ego->vn * ego->ny;
Chris@42:      int my_pe;
Chris@42: 
Chris@42:      MPI_Comm_rank(ego->comm, &my_pe);
Chris@42:      N *= XM(block)(ego->nx, ego->block, my_pe);
Chris@42: 
Chris@42:      for (i = 0; i < N; ++i) I[i] = K(0.0);
Chris@42: }
Chris@42: 
Chris@42: static const problem_adt padt =
Chris@42: {
Chris@42:      PROBLEM_MPI_TRANSPOSE,
Chris@42:      hash,
Chris@42:      zero,
Chris@42:      print,
Chris@42:      destroy
Chris@42: };
Chris@42: 
Chris@42: problem *XM(mkproblem_transpose)(INT nx, INT ny, INT vn,
Chris@42: 				 R *I, R *O,
Chris@42: 				 INT block, INT tblock,
Chris@42: 				 MPI_Comm comm,
Chris@42: 				 unsigned flags)
Chris@42: {
Chris@42:      problem_mpi_transpose *ego =
Chris@42:           (problem_mpi_transpose *)X(mkproblem)(sizeof(problem_mpi_transpose), &padt);
Chris@42: 
Chris@42:      A(nx > 0 && ny > 0 && vn > 0);
Chris@42:      A(block > 0 && XM(num_blocks_ok)(nx, block, comm)
Chris@42:        && tblock > 0 && XM(num_blocks_ok)(ny, tblock, comm));
Chris@42: 
Chris@42:      /* enforce pointer equality if untainted pointers are equal */
Chris@42:      if (UNTAINT(I) == UNTAINT(O))
Chris@42: 	  I = O = JOIN_TAINT(I, O);
Chris@42: 
Chris@42:      ego->nx = nx;
Chris@42:      ego->ny = ny;
Chris@42:      ego->vn = vn;
Chris@42:      ego->I = I;
Chris@42:      ego->O = O;
Chris@42:      ego->block = block > nx ? nx : block;
Chris@42:      ego->tblock = tblock > ny ? ny : tblock;
Chris@42: 
Chris@42:      /* canonicalize flags: we can freely assume that the data is
Chris@42: 	"transposed" if one of the dimensions is 1. */
Chris@42:      if (ego->block == 1)
Chris@42: 	  flags |= TRANSPOSED_IN;
Chris@42:      if (ego->tblock == 1)
Chris@42: 	  flags |= TRANSPOSED_OUT;
Chris@42:      ego->flags = flags;
Chris@42: 
Chris@42:      MPI_Comm_dup(comm, &ego->comm);
Chris@42: 
Chris@42:      return &(ego->super);
Chris@42: }