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 "fftw3-mpi.h"
cannam@127: #include "ifftw-mpi.h"
cannam@127: #include <string.h>
cannam@127: 
cannam@127: #if SIZEOF_SIZE_T == SIZEOF_UNSIGNED_INT
cannam@127: #  define FFTW_MPI_SIZE_T MPI_UNSIGNED
cannam@127: #elif SIZEOF_SIZE_T == SIZEOF_UNSIGNED_LONG
cannam@127: #  define FFTW_MPI_SIZE_T MPI_UNSIGNED_LONG
cannam@127: #elif SIZEOF_SIZE_T == SIZEOF_UNSIGNED_LONG_LONG
cannam@127: #  define FFTW_MPI_SIZE_T MPI_UNSIGNED_LONG_LONG
cannam@127: #else
cannam@127: #  error MPI type for size_t is unknown
cannam@127: #  define FFTW_MPI_SIZE_T MPI_UNSIGNED_LONG
cannam@127: #endif
cannam@127: 
cannam@127: /* Import wisdom from all processes to process 0, as prelude to
cannam@127:    exporting a single wisdom file (this is convenient when we are
cannam@127:    running on identical processors, to avoid the annoyance of having
cannam@127:    per-process wisdom files).  In order to make the time for this
cannam@127:    operation logarithmic in the number of processors (rather than
cannam@127:    linear), we employ a tree reduction algorithm.  This means that the
cannam@127:    wisdom is modified on processes other than root, which shouldn't
cannam@127:    matter in practice. */
cannam@127: void XM(gather_wisdom)(MPI_Comm comm_)
cannam@127: {
cannam@127:      MPI_Comm comm, comm2;
cannam@127:      int my_pe, n_pes;
cannam@127:      char *wis;
cannam@127:      size_t wislen;
cannam@127:      MPI_Status status;
cannam@127: 
cannam@127:      MPI_Comm_dup(comm_, &comm);
cannam@127:      MPI_Comm_rank(comm, &my_pe);
cannam@127:      MPI_Comm_size(comm, &n_pes);
cannam@127: 
cannam@127:      if (n_pes > 2) { /* recursively split into even/odd processes */
cannam@127: 	  MPI_Comm_split(comm, my_pe % 2, my_pe, &comm2);
cannam@127: 	  XM(gather_wisdom)(comm2);
cannam@127: 	  MPI_Comm_free(&comm2);
cannam@127:      }
cannam@127:      if (n_pes > 1 && my_pe < 2) { /* import process 1 -> 0 */
cannam@127: 	  if (my_pe == 1) {
cannam@127: 	       wis = X(export_wisdom_to_string)();
cannam@127: 	       wislen = strlen(wis) + 1;
cannam@127: 	       MPI_Send(&wislen, 1, FFTW_MPI_SIZE_T, 0, 111, comm);
cannam@127: 	       MPI_Send(wis, wislen, MPI_CHAR, 0, 222, comm);
cannam@127: 	       free(wis);
cannam@127: 	  }
cannam@127: 	  else /* my_pe == 0 */ {
cannam@127: 	       MPI_Recv(&wislen, 1, FFTW_MPI_SIZE_T, 1, 111, comm, &status);
cannam@127: 	       wis = (char *) MALLOC(wislen * sizeof(char), OTHER);
cannam@127: 	       MPI_Recv(wis, wislen, MPI_CHAR, 1, 222, comm, &status);
cannam@127: 	       if (!X(import_wisdom_from_string)(wis))
cannam@127: 		    MPI_Abort(comm, 1);
cannam@127: 	       X(ifree)(wis);
cannam@127: 	  }
cannam@127:      }
cannam@127:      MPI_Comm_free(&comm);
cannam@127: }
cannam@127: 
cannam@127: /* broadcast wisdom from process 0 to all other processes; this
cannam@127:    is useful so that we can import wisdom once and not worry
cannam@127:    about parallel I/O or process-specific wisdom, although of
cannam@127:    course it assumes that all the processes have identical
cannam@127:    performance characteristics (i.e. identical hardware). */
cannam@127: void XM(broadcast_wisdom)(MPI_Comm comm_)
cannam@127: {
cannam@127:      MPI_Comm comm;
cannam@127:      int my_pe;
cannam@127:      char *wis;
cannam@127:      size_t wislen;
cannam@127: 
cannam@127:      MPI_Comm_dup(comm_, &comm);
cannam@127:      MPI_Comm_rank(comm, &my_pe);
cannam@127: 
cannam@127:      if (my_pe != 0) {
cannam@127: 	  MPI_Bcast(&wislen, 1, FFTW_MPI_SIZE_T, 0, comm);
cannam@127: 	  wis = (char *) MALLOC(wislen * sizeof(char), OTHER);
cannam@127: 	  MPI_Bcast(wis, wislen, MPI_CHAR, 0, comm);
cannam@127: 	  if (!X(import_wisdom_from_string)(wis))
cannam@127: 	       MPI_Abort(comm, 1);
cannam@127: 	  X(ifree)(wis);
cannam@127:      }
cannam@127:      else /* my_pe == 0 */ {
cannam@127: 	  wis = X(export_wisdom_to_string)();
cannam@127: 	  wislen = strlen(wis) + 1;
cannam@127: 	  MPI_Bcast(&wislen, 1, FFTW_MPI_SIZE_T, 0, comm);
cannam@127: 	  MPI_Bcast(wis, wislen, MPI_CHAR, 0, comm);
cannam@127: 	  X(free)(wis);
cannam@127:      }
cannam@127:      MPI_Comm_free(&comm);
cannam@127: }