csci5451/assignments/03/lpa.cpp

#include <algorithm>
#include <array>
#include <cstring>
#include <functional>
#include <limits>
#include <map>
#include <set>
#include <vector>

#include <mpi.h>
#include <stdio.h>
#include <stdlib.h>
#include <time.h>
#include <unistd.h>
#include <utility>

// #include <fmt/format.h>
// #include <fmt/ranges.h>

#define TAG_SEND_NUM_EDGES 1001
#define TAG_SEND_EDGES 1002
#define TAG_SEND_FINAL_RESULT 1003

#define min(a, b)                                                              \
  ({                                                                           \
    __typeof__(a) _a = (a);                                                    \
    __typeof__(b) _b = (b);                                                    \
    _a < _b ? _a : _b;                                                         \
  })

typedef struct {
  int fst;
  int snd;
} pair;
void init_pair_type(MPI_Datatype *out);

struct pair_vector {
  pair *ptr;
  int cap;
  int len;
};
void pair_vector_init(struct pair_vector *);
void pair_vector_clear(struct pair_vector *);
void pair_vector_push(struct pair_vector *v, int fst, int snd);

pair compute_node_range(int p, int total_num_nodes, int each_num_nodes,
                        int process);

int main(int argc, char **argv) {
  MPI_Init(&argc, &argv);
  int rank, p;
  MPI_Comm_rank(MPI_COMM_WORLD, &rank);
  MPI_Comm_size(MPI_COMM_WORLD, &p);

  MPI_Datatype IntPairType;
  init_pair_type(&IntPairType);

// One process reads the file and distributes the data to the other processes
// using a 1D decomposition (each rank gets approx same number of vertices).
#pragma region
  FILE *fp;
  char *line = NULL;
  size_t len;
  ssize_t read;
  pair params;

  if (rank == 0) {
    fp = fopen(argv[1], "r");

    // Read the first line
    if (getline(&line, &len, fp) != -1)
      sscanf(line, "%d %d", &params.fst, &params.snd);
  }

  // Send the params
  MPI_Bcast(&params, 1, IntPairType, 0, MPI_COMM_WORLD);
  int total_num_nodes = params.fst;
  int total_num_edges = params.snd;
  int each_num_nodes = total_num_nodes / p;

  // Calculate exactly how many nodes my current process holds
  int num_my_nodes =
      rank == p - 1 ? total_num_nodes - rank * each_num_nodes : each_num_nodes;
  int my_nodes[num_my_nodes];

  // Read the edges
  int num_my_edges;
  pair *my_edges;
  int counts[p], displs[p];
  if (rank == 0) {
    line = NULL;
    // pair all_edges[total_num_edges];
    struct pair_vector all_edges;
    pair_vector_init(&all_edges);

    // For the current process, what's the last node we're expecting to see?
    int current_process = 0;
    pair current_node_range =
        compute_node_range(p, total_num_nodes, each_num_nodes, current_process);
    int edge_counter = 0;

    for (int i = 0; i < total_num_edges; ++i) {
      if (getline(&line, &len, fp) == -1)
        break;

      int fst, snd;
      sscanf(line, "%d %d", &fst, &snd);

      if (fst >= current_node_range.snd) {
        if (current_process == 0) {
          num_my_edges = edge_counter;
          my_edges = (pair *)calloc(num_my_edges, sizeof(pair));
          memcpy(my_edges, all_edges.ptr, edge_counter * sizeof(pair));
        } else {
          MPI_Send(&edge_counter, 1, MPI_INT, current_process,
                   TAG_SEND_NUM_EDGES, MPI_COMM_WORLD);
          MPI_Send(all_edges.ptr, edge_counter, IntPairType, current_process,
                   TAG_SEND_EDGES, MPI_COMM_WORLD);
        }

        // We're starting on the next process
        current_process += 1;
        current_node_range = compute_node_range(
            p, total_num_nodes, each_num_nodes, current_process);
        edge_counter = 0;
        pair_vector_clear(&all_edges);
      }

      pair_vector_push(&all_edges, fst, snd);
      edge_counter += 1;
    }

    // We have to send the last one again here, since it didn't get caught in
    // the loop above
    MPI_Send(&edge_counter, 1, MPI_INT, current_process, TAG_SEND_NUM_EDGES,
             MPI_COMM_WORLD);
    MPI_Send(all_edges.ptr, edge_counter, IntPairType, current_process,
             TAG_SEND_EDGES, MPI_COMM_WORLD);

    free(all_edges.ptr);
  } else {
    MPI_Recv(&num_my_edges, 1, MPI_INT, 0, TAG_SEND_NUM_EDGES, MPI_COMM_WORLD,
             NULL);
    my_edges = (pair *)calloc(num_my_edges, sizeof(pair));
    MPI_Recv(my_edges, num_my_edges, IntPairType, 0, TAG_SEND_EDGES,
             MPI_COMM_WORLD, NULL);
  }

  char *buf = (char *)calloc(sizeof(char), 1000);
  int offset = 0; // Keep track of the current position in the buffer
  for (int i = 0; i < min(num_my_edges, 5); i++) {
    offset +=
        sprintf(buf + offset, "(%d, %d)", my_edges[i].fst, my_edges[i].snd);
    if (i < len - 1) {
      // Add a separator (e.g., comma or space) if it's not the last
      offset += sprintf(buf + offset, " ");
    }
  }

  if (rank == 0) {
    fclose(fp);
    if (line)
      free(line);
  }
#pragma endregion

  // STEP 2 TIMER STARTS HERE
  MPI_Barrier(MPI_COMM_WORLD);
  double step_2_start_time = MPI_Wtime();

// Each process analyzes the non-local edges that are contained in its portion
// of the graph.
#pragma region
  std::map<int, int> node_label_assignment;
  pair my_node_range =
      compute_node_range(p, total_num_nodes, each_num_nodes, rank);

  // Initial node assignment
  for (int i = my_node_range.fst; i < my_node_range.snd; ++i) {
    node_label_assignment[i] = i;
  }

  std::map<int, std::set<int>> adj;
  std::set<int> non_local_nodes;
  std::set<std::pair<int, int>> non_local_edges;

  for (int i = 0; i < num_my_edges; ++i) {
    pair edge = my_edges[i];
    adj[edge.fst].insert(edge.snd);

    if (!(my_node_range.fst <= edge.fst && edge.fst < my_node_range.snd)) {
      non_local_nodes.insert(edge.fst);
      non_local_edges.insert(std::make_pair(edge.snd, edge.fst));
    }

    if (!(my_node_range.fst <= edge.snd && edge.snd < my_node_range.snd)) {
      non_local_nodes.insert(edge.snd);
      non_local_edges.insert(std::make_pair(edge.fst, edge.snd));
    }
  }
#pragma endregion

// Each process determines which processors stores the non-local vertices
// corresponding to the non-local edges.
#pragma region
  std::map<int, std::set<int>> send_map;
  std::map<int, std::set<int>> recv_map;

  for (auto entry : non_local_edges) {
    int local_node = entry.first, remote_node = entry.second;

    int corresponding_process = remote_node / each_num_nodes;
    // The last process gets some extra nodes
    if (corresponding_process >= p)
      corresponding_process = p - 1;

    send_map[corresponding_process].insert(local_node);
    recv_map[corresponding_process].insert(remote_node);
  }
#pragma endregion

// All the processes are communicating to figure out which process needs to
// send what data to the other processes.
#pragma region
#pragma endregion

  // STEP 5 TIMER STARTS HERE
  MPI_Barrier(MPI_COMM_WORLD);
  double step_5_start_time = MPI_Wtime();

// The processes perform the transfers of non-local labels and  updates of
// local labels until convergence.
#pragma region
  while (true) {
    // First, exchange the data that needs to be exchanged
    std::vector<int> sendbuf;
    std::vector<int> send_counts;
    std::vector<int> send_displs;
    std::vector<int> recv_counts;
    std::vector<int> recv_displs;

    int recv_total;
    {
      int offset = 0;
      for (int i = 0; i < p; ++i) {
        int count = send_map[i].size();
        // std::sort(send_map[i].begin(), send_map[i].end());
        for (auto k : send_map[i]) {
          sendbuf.push_back(node_label_assignment[k]);
        }
        send_counts.push_back(count);
        send_displs.push_back(offset);
        offset += count;
      }

      offset = 0;
      for (int i = 0; i < p; ++i) {
        int count = recv_map[i].size();
        // std::sort(recv_map[i].begin(), recv_map[i].end());
        recv_counts.push_back(count);
        recv_displs.push_back(offset);
        offset += count;
      }
      recv_total = offset;
    }

    std::vector<int> recvbuf(recv_total, 0);
    MPI_Alltoallv(sendbuf.data(), send_counts.data(), send_displs.data(),
                  MPI_INT, recvbuf.data(), recv_counts.data(),
                  recv_displs.data(), MPI_INT, MPI_COMM_WORLD);

    std::map<int, int> total_node_label_assignment(node_label_assignment);
    for (int i = 0; i < p; ++i) {
      std::vector<int> ouais(recv_map[i].begin(), recv_map[i].end());
      for (int j = 0; j < recv_counts[i]; ++j) {
        int remote_node = ouais[j];
        int remote_value = recvbuf[recv_displs[i] + j];
        total_node_label_assignment[remote_node] = remote_value;
      }
    }

    // For each local node, determine the minimum label out of its neighbors
    std::map<int, int> new_labels;
    for (int i = my_node_range.fst; i < my_node_range.snd; ++i) {
      int current_value = total_node_label_assignment[i];
      int min = current_value;

      for (auto neighbor : adj[i]) {
        if (total_node_label_assignment[neighbor] < min)
          min = total_node_label_assignment[neighbor];
      }

      if (min < current_value) {
        new_labels[i] = min;
      }
    }

    // std::cout << fmt::format("[{}] Helloge {}", rank,
    //                          fmt::join(new_labels, ", "))
    //           << std::endl;

    // Have there been any changes in the labels?
    int num_changes = new_labels.size();
    int total_changes;
    MPI_Allreduce(&num_changes, &total_changes, 1, MPI_INT, MPI_SUM,
                  MPI_COMM_WORLD);

    if (total_changes == 0) {
      break;
    }

    // Update the original node assignment
    for (auto entry : new_labels) {
      node_label_assignment[entry.first] = entry.second;
    }
  }
#pragma endregion

  // END TIMERS
  MPI_Barrier(MPI_COMM_WORLD);
  double end_time = MPI_Wtime();

  if (rank == 0) {
    printf("2-5 Time: %0.04fs\n", end_time - step_2_start_time);
    printf("5 Time: %0.04fs\n", end_time - step_5_start_time);
  }

// The results are gathered to a single process, which writes them to the
// disk.
#pragma region
  if (rank == 0) {
    std::vector<int> all_assignments(total_num_nodes);
    std::map<int, int> label_count;
    int ctr = 0;
    for (int i = 0; i < p; ++i) {
      pair this_node_range =
          compute_node_range(p, total_num_nodes, each_num_nodes, i);
      int count = this_node_range.snd - this_node_range.fst;
      if (i == 0) {
        for (int j = 0; j < count; ++j) {
          all_assignments[this_node_range.fst + j] =
              node_label_assignment[this_node_range.fst + j];
          label_count[all_assignments[this_node_range.fst + j]]++;
        }
      } else {
        MPI_Recv(&all_assignments[this_node_range.fst], count, MPI_INT, i,
                 TAG_SEND_FINAL_RESULT, MPI_COMM_WORLD, NULL);
        for (int j = 0; j < count; ++j) {
          label_count[all_assignments[this_node_range.fst + j]]++;
        }
      }
    }

    std::cout << "Done! " << label_count.size() << std::endl;
  } else {
    std::vector<int> flat_assignments;
    for (int i = my_node_range.fst; i < my_node_range.snd; ++i) {
      flat_assignments.push_back(node_label_assignment[i]);
    }
    MPI_Send(flat_assignments.data(), flat_assignments.size(), MPI_INT, 0,
             TAG_SEND_FINAL_RESULT, MPI_COMM_WORLD);
  }
#pragma endregion

  MPI_Finalize();
  return 0;
}

void init_pair_type(MPI_Datatype *out) {
  int blocklengths[2] = {1, 1};
  MPI_Datatype types[2] = {MPI_INT, MPI_INT};
  MPI_Aint offsets[2];

  offsets[0] = offsetof(pair, fst);
  offsets[1] = offsetof(pair, snd);

  MPI_Type_create_struct(2, blocklengths, offsets, types, out);
  MPI_Type_commit(out);
}

void pair_vector_init(struct pair_vector *v) {
  const int INITIAL = 100;
  v->ptr = (pair *)malloc(INITIAL * sizeof(pair));
  v->cap = INITIAL;
  v->len = 0;
}

void pair_vector_clear(struct pair_vector *v) { v->len = 0; }

void pair_vector_push(struct pair_vector *v, int fst, int snd) {
  if (v->len == v->cap) {
    v->cap *= 2;
    pair *new_loc = (pair *)malloc(v->cap * sizeof(pair));
    for (int i = 0; i < v->len; ++i) {
      new_loc[i].fst = v->ptr[i].fst;
      new_loc[i].snd = v->ptr[i].snd;
    }
    free(v->ptr);
    v->ptr = new_loc;
  }

  v->ptr[v->len].fst = fst;
  v->ptr[v->len].snd = snd;
  v->len++;
}

pair compute_node_range(int p, int total_num_nodes, int each_num_nodes,
                        int process) {
  int start = process * each_num_nodes;
  int end = process == p - 1 ? total_num_nodes : start + each_num_nodes;
  return {.fst = start, .snd = end};
}