EXODUS/packages/seacas/applications/nem_slice/elb_groups.C

/*
 * Copyright(C) 1999-2021, 2023 National Technology & Engineering Solutions
 * of Sandia, LLC (NTESS).  Under the terms of Contract DE-NA0003525 with
 * NTESS, the U.S. Government retains certain rights in this software.
 *
 * See packages/seacas/LICENSE for details
 */

/*++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
 *----------------------------------------------------------------------------
 * Functions contained in this file:
 *+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++*/

#include "elb.h"      // for Problem_Description, etc
#include "elb_elem.h" // for elem_name_from_enum
#include "elb_err.h"  // for Gen_Error
#include "elb_groups.h"
#include "elb_util.h"
#include <cstdio>  // for sscanf, nullptr
#include <cstdlib> // for free, malloc
#include <cstring> // for strchr, strlen
#include <fmt/format.h>
#include <vector>  // for vector

/*****************************************************************************/
namespace {
  template <typename INT>
  void scandescriptor(const char *d, INT *blkids, int n, int nblks, Problem_Description *prob);
  template <typename INT>
  void chgrp(int grp_id, size_t blk, INT *blkids, int nblks, Problem_Description *prob);
} // namespace
extern int ilog2i(size_t n);

/*****************************************************************************/

/*****************************************************************************/
/*****************************************************************************/
/*****************************************************************************/
/* Function parse_groups() begins:
 *----------------------------------------------------------------------------
 * This function will parse the group designation string and set up which
 * element blocks will be in which groups.
 *
 * the group designator prob->groups follows these rules:
 *  - Blocks are grouped using the slash "/" character.
 *  - Ids are separated with white space, comma, or by the hyphen "-" character.
 *  - Any blocks not included in the list, are added to a separate group.
 *  - Duplicates in the list are permitted, but the last group to which a
 *    block is placed is where the block will go.
 *  - Block IDs not in the exodus file are quietly ignored.
 *
 * Examples.
 *   Assume block IDs= 1-20 31-45
 *
 *     descriptor              group1          group2     group3
 *   - "1-20"                  1-20             31-45
 *   - "30-45 3/ 10-12"        3, 30-45         10,11,12   1,2,4-20
 *   - "1-20/40-45/5-10 21-41   1-4,11-20       42,43,45   5-10 31-41
 *
 *****************************************************************************/
template int parse_groups(Mesh_Description<int> *mesh, Problem_Description *prob);
template int parse_groups(Mesh_Description<int64_t> *mesh, Problem_Description *prob);

template <typename INT> int parse_groups(Mesh_Description<INT> *mesh, Problem_Description *prob)
{
  /* prepare the group number array, and copy the element block counts */
  prob->group_no.resize(mesh->num_el_blks, -1);

  /* convert any comma's to blank spaces in the designator string */
  for (size_t i = 0; i < strlen(prob->groups); i++) {
    if (prob->groups[i] == ',') {
      prob->groups[i] = ' ';

      /* fill in the group identifier for each block */
    }
  }
  char  *id = prob->groups;
  size_t i  = 0;
  do {
    if (*id == '/') {
      id++;
    }
    scandescriptor(id, mesh->eb_ids.data(), i, mesh->num_el_blks, prob);
    id = strchr(id, '/');
    i++;
  } while (id != nullptr);
  int last = i;

  /* set any remaining blocks to new group */
  int found = 0;
  for (i = 0; i < mesh->num_el_blks; i++) {
    if (prob->group_no[i] < 0) {
      prob->group_no[i] = last;
      found             = 1;
    }
  }

  if (found) {
    last++;
  }

  prob->num_groups = last;

  {
    size_t first_el = 0;
    fmt::print("\nNumber of blocks: {}\n", mesh->num_el_blks);
    fmt::print("Block ID and associated groups:\n");
    fmt::print("   block   #elems  group   type\n");
    for (i = 0; i < mesh->num_el_blks; i++) {
      fmt::print("{:8d}{:8d}{:8d}{:8s}\n", (size_t)mesh->eb_ids[i], (size_t)mesh->eb_cnts[i],
                 prob->group_no[i], elem_name_from_enum(mesh->elem_type[first_el]));
      first_el += mesh->eb_cnts[i];
    }
    fmt::print("There are {} groups of blocks\n", prob->num_groups);
  }

  /* finished with the group designator string */
  delete[] prob->groups;

  return 1;
}

/*****************************************************************************/
/*****************************************************************************/
/*****************************************************************************/
/* Function get_group_info() begins:
 *----------------------------------------------------------------------------
 * This function will get the information that is needed to break up the
 * groups before passing them to Chaco.
 * It generates the following:
 *   - an array of with the group number for each element
 *   - an array of elements per group
 *   - an array of processors to be used per group
 *   - the max number of vertecies and adjacencies for all of the groups
 *     this value is needed so that arrays used to pass information
 *     to Chaco can be allocated
 *****************************************************************************/
template int get_group_info(Machine_Description *machine, Problem_Description *prob,
                            Mesh_Description<int> *mesh, Graph_Description<int> *graph,
                            int elem2grp[], int nprocg[], int nelemg[], size_t *max_vtx,
                            size_t *max_adj);
template int get_group_info(Machine_Description *machine, Problem_Description *prob,
                            Mesh_Description<int64_t> *mesh, Graph_Description<int64_t> *graph,
                            int elem2grp[], int nprocg[], int nelemg[], size_t *max_vtx,
                            size_t *max_adj);

template <typename INT>
int get_group_info(Machine_Description *machine, Problem_Description *prob,
                   Mesh_Description<INT> *mesh, Graph_Description<INT> *graph, int elem2grp[],
                   int nprocg[], int nelemg[], size_t *max_vtx, size_t *max_adj)
{
  int              nproc = 0;
  std::vector<int> nadj_per_grp;

  /* allocate array to hold adjacency counts, if necessary */
  if (prob->alloc_graph == ELB_TRUE) {
    nadj_per_grp.resize(prob->num_groups);
  }

  /* initialize the group counts arrays */
  for (int i = 0; i < prob->num_groups; i++) {
    nelemg[i] = 0;
  }

  /*
   * fill the vertex2proc array with the group number of the individual
   * elements, calculate how many elements in each group and determine
   * how many adjacencies are in each group (if necessary).
   */
  INT sum  = 0;
  int iblk = 0;
  for (size_t i = 0; i < prob->num_vertices; i++) {

    /* figure out which element block this is */
    if (sum == mesh->eb_cnts[iblk]) {
      sum = 0;
      iblk++;
    }
    sum++;

    /*
     * use negative numbers to specify the groups, in order to
     * avoid having a problem with group 0, add 1 to the group
     * number
     */
    elem2grp[i] = -(prob->group_no[iblk] + 1);

    nelemg[prob->group_no[iblk]]++;

    if (prob->alloc_graph == ELB_TRUE) {
      nadj_per_grp[prob->group_no[iblk]] += graph->start[i + 1] - graph->start[i];
    }
  }

  /*
   * calculate how many processors to use for each group
   *   using method from the materials group, haven't really checked it
   */
  if (machine->type == MESH) {
    nproc = machine->procs_per_box;
  }
  else if (machine->type == HCUBE) {
    nproc = ilog2i(machine->procs_per_box);
  }
  for (int i = 0; i < prob->num_groups; i++) {
    nprocg[i] = int((nproc * (nelemg[i] + 0.5F)) / static_cast<float>(prob->num_vertices));
    if (nelemg[i] && !nprocg[i]) {
      nprocg[i] = 1;
    }
  }

  /*
   * check to see if correct number of processors have been allocated
   * and get the maximum number of vertices
   */
  sum      = 0;
  size_t j = 0;
  *max_vtx = 0;
  *max_adj = 0;
  for (int i = 0; i < prob->num_groups; i++) {
    sum += nprocg[i];
    if (nprocg[i] > nprocg[j]) {
      j        = i;
      *max_vtx = nelemg[j]; /* most processors implies most elements */
    }

    /* determine how large to make temporary arrays */
    if (static_cast<size_t>(nelemg[i]) > *max_vtx) {
      *max_vtx = nelemg[i];
    }
    if (prob->alloc_graph == ELB_TRUE) {
      if (static_cast<size_t>(nadj_per_grp[i]) > *max_adj) {
        *max_adj = nadj_per_grp[i];
      }
    }
  }
  if (sum != nproc) {
    /* correct group with most processors (j determined above) */
    nprocg[j] -= (sum - nproc);
    if (nprocg[j] <= 0) {
      Gen_Error(0, "Unable to balance # processors in get_group_info().");
      return 0;
    }
  }
  fmt::print("Load balance information\n");
  for (int i = 0; i < prob->num_groups; i++) {
    fmt::print("group[{}]  #elements={:10d}  #proc={}\n", i, nelemg[i], nprocg[i]);
  }

  return 1;
}

/**********************************************************************/
namespace {
  template <typename INT>
  void scandescriptor(const char *d, INT *blkids, int n, int nblks, Problem_Description *prob)
  {
    /* reads the descriptor up to the next "/" character. interprets the
       descriptor. The ranges specified in the descriptor are then stored
       in the grp array. */
    int i;
    int last = 0; /* last integer read */
    int stop;     /* stop value in a string range */
    int qn;       /* number of bytes read */
    int c;        /* integer index when spanning a range */

    const char *p = d;
    while (*p != '/' && *p != 0) {
      int q = sscanf(p, "%d%n", &i, &qn);
      if (q == 0 || i < 0) {
        if (p[qn - 1] == '/') {
          return;
        }
        if (i < 0) {
          stop = -i;
          for (c = last; c <= stop; c++) {
            chgrp(n, c, blkids, nblks, prob);
          }
        }
        else if (p[qn - 1] == '-') {
          p += qn;
          sscanf(p, "%d%n", &stop, &qn);
          for (c = last; c <= stop; c++) {
            chgrp(n, c, blkids, nblks, prob);
          }
        }
        else {
          /* check for minus sign */
          for (c = 0; c < qn; c++) {
            if (p[c] == '-') {
              break;
            }
          }
          if (c < qn) {
            p += qn;
            sscanf(p, "%d%n", &stop, &qn);
            for (c = last; c <= stop; c++) {
              chgrp(n, c, blkids, nblks, prob);
            }
          }
          else {
            fmt::print("Error reading descriptor '{}'\n", d);
            fmt::print("                          ");
            for (c = 0; c < qn; c++) {
              fmt::print(" ");
            }
            fmt::print("^\n");
            return;
          }
        }
      }
      else {
        last = i;
      }
      chgrp(n, i, blkids, nblks, prob);
      p += qn;
    }
  }

  /**********************************************************************/
  /* changes the grp[] entry corresponding to block=blk to the value grp_id
   * It uses the extern variables "nblks", "grp" and "blkid". only grp
   * is altered.
   *
   * Not very efficient. To find the blk, it loops through all blks.
   */
  template <typename INT>
  void chgrp(int grp_id, size_t blk, INT *blkids, int nblks, Problem_Description *prob)
  {
    for (int j = 0; j < nblks; j++) {
      if (blkids[j] == (INT)blk) {
        prob->group_no[j] = grp_id;
        return;
      }
    }
  }
} // namespace