xref: /petsc/src/ksp/pc/impls/tfs/comm.c (revision a501084f0ec6f311bc0b878b7b6f8a6508165e40)

/***********************************comm.c*************************************

Author: Henry M. Tufo III

e-mail: hmt@cs.brown.edu

snail-mail:
Division of Applied Mathematics
Brown University
Providence, RI 02912

Last Modification:
11.21.97
***********************************comm.c*************************************/

/***********************************comm.c*************************************
File Description:
-----------------

***********************************comm.c*************************************/
#include "petsc.h"

#include "const.h"
#include "types.h"
#include "error.h"
#include "ivec.h"
#include "comm.h"


/* global program control variables - explicitly exported */
int my_id            = 0;
int num_nodes        = 1;
int floor_num_nodes  = 0;
int i_log2_num_nodes = 0;

/* global program control variables */
static int p_init = 0;
static int modfl_num_nodes;
static int edge_not_pow_2;

static unsigned int edge_node[sizeof(PetscInt)*32];

/***********************************comm.c*************************************
Function: giop()

Input :
Output:
Return:
Description:
***********************************comm.c*************************************/
void
comm_init (void)
{

  if (p_init++) return;

#if PETSC_SIZEOF_INT != 4
  error_msg_warning("Int != 4 Bytes!");
#endif


  MPI_Comm_size(MPI_COMM_WORLD,&num_nodes);
  MPI_Comm_rank(MPI_COMM_WORLD,&my_id);

  if (num_nodes> (INT_MAX >> 1))
  {error_msg_fatal("Can't have more then MAX_INT/2 nodes!!!");}

  ivec_zero((int*)edge_node,sizeof(PetscInt)*32);

  floor_num_nodes = 1;
  i_log2_num_nodes = modfl_num_nodes = 0;
  while (floor_num_nodes <= num_nodes)
    {
      edge_node[i_log2_num_nodes] = my_id ^ floor_num_nodes;
      floor_num_nodes <<= 1;
      i_log2_num_nodes++;
    }

  i_log2_num_nodes--;
  floor_num_nodes >>= 1;
  modfl_num_nodes = (num_nodes - floor_num_nodes);

  if ((my_id > 0) && (my_id <= modfl_num_nodes))
    {edge_not_pow_2=((my_id|floor_num_nodes)-1);}
  else if (my_id >= floor_num_nodes)
    {edge_not_pow_2=((my_id^floor_num_nodes)+1);
    }
  else
    {edge_not_pow_2 = 0;}

}



/***********************************comm.c*************************************
Function: giop()

Input :
Output:
Return:
Description: fan-in/out version
***********************************comm.c*************************************/
void
giop(int *vals, int *work, int n, int *oprs)
{
   int mask, edge;
  int type, dest;
  vfp fp;
  MPI_Status  status;


#ifdef SAFE
  /* ok ... should have some data, work, and operator(s) */
  if (!vals||!work||!oprs)
    {error_msg_fatal("giop() :: vals=%D, work=%D, oprs=%D",vals,work,oprs);}

  /* non-uniform should have at least two entries */
  if ((oprs[0] == NON_UNIFORM)&&(n<2))
    {error_msg_fatal("giop() :: non_uniform and n=0,1?");}

  /* check to make sure comm package has been initialized */
  if (!p_init)
    {comm_init();}
#endif

  /* if there's nothing to do return */
  if ((num_nodes<2)||(!n))
    {
      return;
    }

  /* a negative number if items to send ==> fatal */
  if (n<0)
    {error_msg_fatal("giop() :: n=%D<0?",n);}

  /* advance to list of n operations for custom */
  if ((type=oprs[0])==NON_UNIFORM)
    {oprs++;}

  /* major league hack */
  if (!(fp = (vfp) ivec_fct_addr(type))) {
    error_msg_warning("giop() :: hope you passed in a rbfp!\n");
    fp = (vfp) oprs;
  }

  /* all msgs will be of the same length */
  /* if not a hypercube must colapse partial dim */
  if (edge_not_pow_2)
    {
      if (my_id >= floor_num_nodes)
	{MPI_Send(vals,n,MPI_INT,edge_not_pow_2,MSGTAG0+my_id,MPI_COMM_WORLD);}
      else
	{
	  MPI_Recv(work,n,MPI_INT,MPI_ANY_SOURCE,MSGTAG0+edge_not_pow_2,
		   MPI_COMM_WORLD,&status);
	  (*fp)(vals,work,n,oprs);
	}
    }

  /* implement the mesh fan in/out exchange algorithm */
  if (my_id<floor_num_nodes)
    {
      for (mask=1,edge=0; edge<i_log2_num_nodes; edge++,mask<<=1)
	{
	  dest = my_id^mask;
	  if (my_id > dest)
	    {MPI_Send(vals,n,MPI_INT,dest,MSGTAG2+my_id,MPI_COMM_WORLD);}
	  else
	    {
	      MPI_Recv(work,n,MPI_INT,MPI_ANY_SOURCE,MSGTAG2+dest,
		       MPI_COMM_WORLD, &status);
	      (*fp)(vals, work, n, oprs);
	    }
	}

      mask=floor_num_nodes>>1;
      for (edge=0; edge<i_log2_num_nodes; edge++,mask>>=1)
	{
	  if (my_id%mask)
	    {continue;}

	  dest = my_id^mask;
	  if (my_id < dest)
	    {MPI_Send(vals,n,MPI_INT,dest,MSGTAG4+my_id,MPI_COMM_WORLD);}
	  else
	    {
	      MPI_Recv(vals,n,MPI_INT,MPI_ANY_SOURCE,MSGTAG4+dest,
		       MPI_COMM_WORLD, &status);
	    }
	}
    }

  /* if not a hypercube must expand to partial dim */
  if (edge_not_pow_2)
    {
      if (my_id >= floor_num_nodes)
	{
	  MPI_Recv(vals,n,MPI_INT,MPI_ANY_SOURCE,MSGTAG5+edge_not_pow_2,
		   MPI_COMM_WORLD,&status);
	}
      else
	{MPI_Send(vals,n,MPI_INT,edge_not_pow_2,MSGTAG5+my_id,MPI_COMM_WORLD);}
    }
}

/***********************************comm.c*************************************
Function: grop()

Input :
Output:
Return:
Description: fan-in/out version
***********************************comm.c*************************************/
void
grop(PetscScalar *vals, PetscScalar *work, int n, int *oprs)
{
   int mask, edge;
  int type, dest;
  vfp fp;
  MPI_Status  status;


#ifdef SAFE
  /* ok ... should have some data, work, and operator(s) */
  if (!vals||!work||!oprs)
    {error_msg_fatal("grop() :: vals=%D, work=%D, oprs=%D",vals,work,oprs);}

  /* non-uniform should have at least two entries */
  if ((oprs[0] == NON_UNIFORM)&&(n<2))
    {error_msg_fatal("grop() :: non_uniform and n=0,1?");}

  /* check to make sure comm package has been initialized */
  if (!p_init)
    {comm_init();}
#endif

  /* if there's nothing to do return */
  if ((num_nodes<2)||(!n))
    {return;}

  /* a negative number of items to send ==> fatal */
  if (n<0)
    {error_msg_fatal("gdop() :: n=%D<0?",n);}

  /* advance to list of n operations for custom */
  if ((type=oprs[0])==NON_UNIFORM)
    {oprs++;}

  if (!(fp = (vfp) rvec_fct_addr(type))) {
    error_msg_warning("grop() :: hope you passed in a rbfp!\n");
    fp = (vfp) oprs;
  }

  /* all msgs will be of the same length */
  /* if not a hypercube must colapse partial dim */
  if (edge_not_pow_2)
    {
      if (my_id >= floor_num_nodes)
	{MPI_Send(vals,n,MPIU_SCALAR,edge_not_pow_2,MSGTAG0+my_id,
		  MPI_COMM_WORLD);}
      else
	{
	  MPI_Recv(work,n,MPIU_SCALAR,MPI_ANY_SOURCE,MSGTAG0+edge_not_pow_2,
		   MPI_COMM_WORLD,&status);
	  (*fp)(vals,work,n,oprs);
	}
    }

  /* implement the mesh fan in/out exchange algorithm */
  if (my_id<floor_num_nodes)
    {
      for (mask=1,edge=0; edge<i_log2_num_nodes; edge++,mask<<=1)
	{
	  dest = my_id^mask;
	  if (my_id > dest)
	    {MPI_Send(vals,n,MPIU_SCALAR,dest,MSGTAG2+my_id,MPI_COMM_WORLD);}
	  else
	    {
	      MPI_Recv(work,n,MPIU_SCALAR,MPI_ANY_SOURCE,MSGTAG2+dest,
		       MPI_COMM_WORLD, &status);
	      (*fp)(vals, work, n, oprs);
	    }
	}

      mask=floor_num_nodes>>1;
      for (edge=0; edge<i_log2_num_nodes; edge++,mask>>=1)
	{
	  if (my_id%mask)
	    {continue;}

	  dest = my_id^mask;
	  if (my_id < dest)
	    {MPI_Send(vals,n,MPIU_SCALAR,dest,MSGTAG4+my_id,MPI_COMM_WORLD);}
	  else
	    {
	      MPI_Recv(vals,n,MPIU_SCALAR,MPI_ANY_SOURCE,MSGTAG4+dest,
		       MPI_COMM_WORLD, &status);
	    }
	}
    }

  /* if not a hypercube must expand to partial dim */
  if (edge_not_pow_2)
    {
      if (my_id >= floor_num_nodes)
	{
	  MPI_Recv(vals,n,MPIU_SCALAR,MPI_ANY_SOURCE,MSGTAG5+edge_not_pow_2,
		   MPI_COMM_WORLD,&status);
	}
      else
	{MPI_Send(vals,n,MPIU_SCALAR,edge_not_pow_2,MSGTAG5+my_id,
		  MPI_COMM_WORLD);}
    }
}


/***********************************comm.c*************************************
Function: grop()

Input :
Output:
Return:
Description: fan-in/out version

note good only for num_nodes=2^k!!!

***********************************comm.c*************************************/
void
grop_hc(PetscScalar *vals, PetscScalar *work, int n, int *oprs, int dim)
{
   int mask, edge;
  int type, dest;
  vfp fp;
  MPI_Status  status;

#ifdef SAFE
  /* ok ... should have some data, work, and operator(s) */
  if (!vals||!work||!oprs)
    {error_msg_fatal("grop_hc() :: vals=%D, work=%D, oprs=%D",vals,work,oprs);}

  /* non-uniform should have at least two entries */
  if ((oprs[0] == NON_UNIFORM)&&(n<2))
    {error_msg_fatal("grop_hc() :: non_uniform and n=0,1?");}

  /* check to make sure comm package has been initialized */
  if (!p_init)
    {comm_init();}
#endif

  /* if there's nothing to do return */
  if ((num_nodes<2)||(!n)||(dim<=0))
    {return;}

  /* the error msg says it all!!! */
  if (modfl_num_nodes)
    {error_msg_fatal("grop_hc() :: num_nodes not a power of 2!?!");}

  /* a negative number of items to send ==> fatal */
  if (n<0)
    {error_msg_fatal("grop_hc() :: n=%D<0?",n);}

  /* can't do more dimensions then exist */
  dim = PetscMin(dim,i_log2_num_nodes);

  /* advance to list of n operations for custom */
  if ((type=oprs[0])==NON_UNIFORM)
    {oprs++;}

  if (!(fp = (vfp) rvec_fct_addr(type))) {
    error_msg_warning("grop_hc() :: hope you passed in a rbfp!\n");
    fp = (vfp) oprs;
  }

  for (mask=1,edge=0; edge<dim; edge++,mask<<=1)
    {
      dest = my_id^mask;
      if (my_id > dest)
	{MPI_Send(vals,n,MPIU_SCALAR,dest,MSGTAG2+my_id,MPI_COMM_WORLD);}
      else
	{
	  MPI_Recv(work,n,MPIU_SCALAR,MPI_ANY_SOURCE,MSGTAG2+dest,MPI_COMM_WORLD,
		   &status);
	  (*fp)(vals, work, n, oprs);
	}
    }

  if (edge==dim)
    {mask>>=1;}
  else
    {while (++edge<dim) {mask<<=1;}}

  for (edge=0; edge<dim; edge++,mask>>=1)
    {
      if (my_id%mask)
	{continue;}

      dest = my_id^mask;
      if (my_id < dest)
	{MPI_Send(vals,n,MPIU_SCALAR,dest,MSGTAG4+my_id,MPI_COMM_WORLD);}
      else
	{
	  MPI_Recv(vals,n,MPIU_SCALAR,MPI_ANY_SOURCE,MSGTAG4+dest,MPI_COMM_WORLD,
		   &status);
	}
    }
}


/***********************************comm.c*************************************
Function: gop()

Input :
Output:
Return:
Description: fan-in/out version
***********************************comm.c*************************************/
void gfop(void *vals, void *work, int n, vbfp fp, MPI_Datatype dt, int comm_type)
{
   int mask, edge;
  int dest;
  MPI_Status  status;
  MPI_Op op;

#ifdef SAFE
  /* check to make sure comm package has been initialized */
  if (!p_init)
    {comm_init();}

  /* ok ... should have some data, work, and operator(s) */
  if (!vals||!work||!fp)
    {error_msg_fatal("gop() :: v=%D, w=%D, f=%D",vals,work,fp);}
#endif

  /* if there's nothing to do return */
  if ((num_nodes<2)||(!n))
    {return;}

  /* a negative number of items to send ==> fatal */
  if (n<0)
    {error_msg_fatal("gop() :: n=%D<0?",n);}

  if (comm_type==MPI)
    {
      MPI_Op_create(fp,TRUE,&op);
      MPI_Allreduce (vals, work, n, dt, op, MPI_COMM_WORLD);
      MPI_Op_free(&op);
      return;
    }

  /* if not a hypercube must colapse partial dim */
  if (edge_not_pow_2)
    {
      if (my_id >= floor_num_nodes)
	{MPI_Send(vals,n,dt,edge_not_pow_2,MSGTAG0+my_id,
		  MPI_COMM_WORLD);}
      else
	{
	  MPI_Recv(work,n,dt,MPI_ANY_SOURCE,MSGTAG0+edge_not_pow_2,
		   MPI_COMM_WORLD,&status);
	  (*fp)(vals,work,&n,&dt);
	}
    }

  /* implement the mesh fan in/out exchange algorithm */
  if (my_id<floor_num_nodes)
    {
      for (mask=1,edge=0; edge<i_log2_num_nodes; edge++,mask<<=1)
	{
	  dest = my_id^mask;
	  if (my_id > dest)
	    {MPI_Send(vals,n,dt,dest,MSGTAG2+my_id,MPI_COMM_WORLD);}
	  else
	    {
	      MPI_Recv(work,n,dt,MPI_ANY_SOURCE,MSGTAG2+dest,
		       MPI_COMM_WORLD, &status);
	      (*fp)(vals, work, &n, &dt);
	    }
	}

      mask=floor_num_nodes>>1;
      for (edge=0; edge<i_log2_num_nodes; edge++,mask>>=1)
	{
	  if (my_id%mask)
	    {continue;}

	  dest = my_id^mask;
	  if (my_id < dest)
	    {MPI_Send(vals,n,dt,dest,MSGTAG4+my_id,MPI_COMM_WORLD);}
	  else
	    {
	      MPI_Recv(vals,n,dt,MPI_ANY_SOURCE,MSGTAG4+dest,
		       MPI_COMM_WORLD, &status);
	    }
	}
    }

  /* if not a hypercube must expand to partial dim */
  if (edge_not_pow_2)
    {
      if (my_id >= floor_num_nodes)
	{
	  MPI_Recv(vals,n,dt,MPI_ANY_SOURCE,MSGTAG5+edge_not_pow_2,
		   MPI_COMM_WORLD,&status);
	}
      else
	{MPI_Send(vals,n,dt,edge_not_pow_2,MSGTAG5+my_id,
		  MPI_COMM_WORLD);}
    }
}






/******************************************************************************
Function: giop()

Input :
Output:
Return:
Description:

ii+1 entries in seg :: 0 .. ii

******************************************************************************/
void
ssgl_radd( PetscScalar *vals,  PetscScalar *work,  int level,
	   int *segs)
{
   int edge, type, dest, mask;
   int stage_n;
  MPI_Status  status;

#ifdef SAFE
  /* check to make sure comm package has been initialized */
  if (!p_init)
    {comm_init();}
#endif


  /* all msgs are *NOT* the same length */
  /* implement the mesh fan in/out exchange algorithm */
  for (mask=0, edge=0; edge<level; edge++, mask++)
    {
      stage_n = (segs[level] - segs[edge]);
      if (stage_n && !(my_id & mask))
	{
	  dest = edge_node[edge];
	  type = MSGTAG3 + my_id + (num_nodes*edge);
	  if (my_id>dest)
          {MPI_Send(vals+segs[edge],stage_n,MPIU_SCALAR,dest,type,
                      MPI_COMM_WORLD);}
	  else
	    {
	      type =  type - my_id + dest;
              MPI_Recv(work,stage_n,MPIU_SCALAR,MPI_ANY_SOURCE,type,
                       MPI_COMM_WORLD,&status);
	      rvec_add(vals+segs[edge], work, stage_n);
	    }
	}
      mask <<= 1;
    }
  mask>>=1;
  for (edge=0; edge<level; edge++)
    {
      stage_n = (segs[level] - segs[level-1-edge]);
      if (stage_n && !(my_id & mask))
	{
	  dest = edge_node[level-edge-1];
	  type = MSGTAG6 + my_id + (num_nodes*edge);
	  if (my_id<dest)
            {MPI_Send(vals+segs[level-1-edge],stage_n,MPIU_SCALAR,dest,type,
                      MPI_COMM_WORLD);}
	  else
	    {
	      type =  type - my_id + dest;
              MPI_Recv(vals+segs[level-1-edge],stage_n,MPIU_SCALAR,
                       MPI_ANY_SOURCE,type,MPI_COMM_WORLD,&status);
	    }
	}
      mask >>= 1;
    }
}



/***********************************comm.c*************************************
Function: grop_hc_vvl()

Input :
Output:
Return:
Description: fan-in/out version

note good only for num_nodes=2^k!!!

***********************************comm.c*************************************/
void
grop_hc_vvl(PetscScalar *vals, PetscScalar *work, int *segs, int *oprs, int dim)
{
   int mask, edge, n;
  int type, dest;
  vfp fp;
  MPI_Status  status;

  error_msg_fatal("grop_hc_vvl() :: is not working!\n");

#ifdef SAFE
  /* ok ... should have some data, work, and operator(s) */
  if (!vals||!work||!oprs||!segs)
    {error_msg_fatal("grop_hc() :: vals=%D, work=%D, oprs=%D",vals,work,oprs);}

  /* non-uniform should have at least two entries */

  /* check to make sure comm package has been initialized */
  if (!p_init)
    {comm_init();}
#endif

  /* if there's nothing to do return */
  if ((num_nodes<2)||(dim<=0))
    {return;}

  /* the error msg says it all!!! */
  if (modfl_num_nodes)
    {error_msg_fatal("grop_hc() :: num_nodes not a power of 2!?!");}

  /* can't do more dimensions then exist */
  dim = PetscMin(dim,i_log2_num_nodes);

  /* advance to list of n operations for custom */
  if ((type=oprs[0])==NON_UNIFORM)
    {oprs++;}

  if (!(fp = (vfp) rvec_fct_addr(type))){
    error_msg_warning("grop_hc() :: hope you passed in a rbfp!\n");
    fp = (vfp) oprs;
  }

  for (mask=1,edge=0; edge<dim; edge++,mask<<=1)
    {
      n = segs[dim]-segs[edge];
      dest = my_id^mask;
      if (my_id > dest)
	{MPI_Send(vals+segs[edge],n,MPIU_SCALAR,dest,MSGTAG2+my_id,MPI_COMM_WORLD);}
      else
	{
	  MPI_Recv(work,n,MPIU_SCALAR,MPI_ANY_SOURCE,MSGTAG2+dest,
		   MPI_COMM_WORLD, &status);
	  (*fp)(vals+segs[edge], work, n, oprs);
	}
    }

  if (edge==dim)
    {mask>>=1;}
  else
    {while (++edge<dim) {mask<<=1;}}

  for (edge=0; edge<dim; edge++,mask>>=1)
    {
      if (my_id%mask)
	{continue;}

      n = (segs[dim]-segs[dim-1-edge]);

      dest = my_id^mask;
      if (my_id < dest)
	{MPI_Send(vals+segs[dim-1-edge],n,MPIU_SCALAR,dest,MSGTAG4+my_id,
		  MPI_COMM_WORLD);}
      else
	{
	  MPI_Recv(vals+segs[dim-1-edge],n,MPIU_SCALAR,MPI_ANY_SOURCE,
		   MSGTAG4+dest,MPI_COMM_WORLD, &status);
	}
    }
}

/******************************************************************************
Function: giop()

Input :
Output:
Return:
Description:

ii+1 entries in seg :: 0 .. ii

******************************************************************************/
void new_ssgl_radd( PetscScalar *vals,  PetscScalar *work,  int level, int *segs)
{
   int edge, type, dest, mask;
   int stage_n;
  MPI_Status  status;

#ifdef SAFE
  /* check to make sure comm package has been initialized */
  if (!p_init)
    {comm_init();}
#endif

  /* all msgs are *NOT* the same length */
  /* implement the mesh fan in/out exchange algorithm */
  for (mask=0, edge=0; edge<level; edge++, mask++)
    {
      stage_n = (segs[level] - segs[edge]);
      if (stage_n && !(my_id & mask))
	{
	  dest = edge_node[edge];
	  type = MSGTAG3 + my_id + (num_nodes*edge);
	  if (my_id>dest)
          {MPI_Send(vals+segs[edge],stage_n,MPIU_SCALAR,dest,type,
                      MPI_COMM_WORLD);}
	  else
	    {
	      type =  type - my_id + dest;
              MPI_Recv(work,stage_n,MPIU_SCALAR,MPI_ANY_SOURCE,type,
                       MPI_COMM_WORLD,&status);
	      rvec_add(vals+segs[edge], work, stage_n);
	    }
	}
      mask <<= 1;
    }
  mask>>=1;
  for (edge=0; edge<level; edge++)
    {
      stage_n = (segs[level] - segs[level-1-edge]);
      if (stage_n && !(my_id & mask))
	{
	  dest = edge_node[level-edge-1];
	  type = MSGTAG6 + my_id + (num_nodes*edge);
	  if (my_id<dest)
            {MPI_Send(vals+segs[level-1-edge],stage_n,MPIU_SCALAR,dest,type,
                      MPI_COMM_WORLD);}
	  else
	    {
	      type =  type - my_id + dest;
              MPI_Recv(vals+segs[level-1-edge],stage_n,MPIU_SCALAR,
                       MPI_ANY_SOURCE,type,MPI_COMM_WORLD,&status);
	    }
	}
      mask >>= 1;
    }
}



/***********************************comm.c*************************************
Function: giop()

Input :
Output:
Return:
Description: fan-in/out version

note good only for num_nodes=2^k!!!

***********************************comm.c*************************************/
void giop_hc(int *vals, int *work, int n, int *oprs, int dim)
{
   int mask, edge;
  int type, dest;
  vfp fp;
  MPI_Status  status;

#ifdef SAFE
  /* ok ... should have some data, work, and operator(s) */
  if (!vals||!work||!oprs)
    {error_msg_fatal("giop_hc() :: vals=%D, work=%D, oprs=%D",vals,work,oprs);}

  /* non-uniform should have at least two entries */
  if ((oprs[0] == NON_UNIFORM)&&(n<2))
    {error_msg_fatal("giop_hc() :: non_uniform and n=0,1?");}

  /* check to make sure comm package has been initialized */
  if (!p_init)
    {comm_init();}
#endif

  /* if there's nothing to do return */
  if ((num_nodes<2)||(!n)||(dim<=0))
    {return;}

  /* the error msg says it all!!! */
  if (modfl_num_nodes)
    {error_msg_fatal("giop_hc() :: num_nodes not a power of 2!?!");}

  /* a negative number of items to send ==> fatal */
  if (n<0)
    {error_msg_fatal("giop_hc() :: n=%D<0?",n);}

  /* can't do more dimensions then exist */
  dim = PetscMin(dim,i_log2_num_nodes);

  /* advance to list of n operations for custom */
  if ((type=oprs[0])==NON_UNIFORM)
    {oprs++;}

  if (!(fp = (vfp) ivec_fct_addr(type))){
    error_msg_warning("giop_hc() :: hope you passed in a rbfp!\n");
    fp = (vfp) oprs;
  }

  for (mask=1,edge=0; edge<dim; edge++,mask<<=1)
    {
      dest = my_id^mask;
      if (my_id > dest)
	{MPI_Send(vals,n,MPIU_INT,dest,MSGTAG2+my_id,MPI_COMM_WORLD);}
      else
	{
	  MPI_Recv(work,n,MPIU_INT,MPI_ANY_SOURCE,MSGTAG2+dest,MPI_COMM_WORLD,
		   &status);
	  (*fp)(vals, work, n, oprs);
	}
    }

  if (edge==dim)
    {mask>>=1;}
  else
    {while (++edge<dim) {mask<<=1;}}

  for (edge=0; edge<dim; edge++,mask>>=1)
    {
      if (my_id%mask)
	{continue;}

      dest = my_id^mask;
      if (my_id < dest)
	{MPI_Send(vals,n,MPIU_INT,dest,MSGTAG4+my_id,MPI_COMM_WORLD);}
      else
	{
	  MPI_Recv(vals,n,MPIU_INT,MPI_ANY_SOURCE,MSGTAG4+dest,MPI_COMM_WORLD,
		   &status);
	}
    }
}