#line 1 "/home/maspy/compro/library/my_template.hpp"
#if defined(LOCAL)
#include <my_template_compiled.hpp>
#else

// https://codeforces.com/blog/entry/96344
#pragma GCC optimize("Ofast,unroll-loops")
// いまの CF だとこれ入れると動かない？
// #pragma GCC target("avx2,popcnt")

#include <bits/stdc++.h>

using namespace std;

using ll = long long;
using u8 = uint8_t;
using u16 = uint16_t;
using u32 = uint32_t;
using u64 = uint64_t;
using i128 = __int128;
using u128 = unsigned __int128;
using f128 = __float128;

template <class T>
constexpr T infty = 0;
template <>
constexpr int infty<int> = 1'010'000'000;
template <>
constexpr ll infty<ll> = 2'020'000'000'000'000'000;
template <>
constexpr u32 infty<u32> = infty<int>;
template <>
constexpr u64 infty<u64> = infty<ll>;
template <>
constexpr i128 infty<i128> = i128(infty<ll>) * 2'000'000'000'000'000'000;
template <>
constexpr double infty<double> = infty<ll>;
template <>
constexpr long double infty<long double> = infty<ll>;

using pi = pair<ll, ll>;
using vi = vector<ll>;
template <class T>
using vc = vector<T>;
template <class T>
using vvc = vector<vc<T>>;
template <class T>
using vvvc = vector<vvc<T>>;
template <class T>
using vvvvc = vector<vvvc<T>>;
template <class T>
using vvvvvc = vector<vvvvc<T>>;
template <class T>
using pq = priority_queue<T>;
template <class T>
using pqg = priority_queue<T, vector<T>, greater<T>>;

#define vv(type, name, h, ...) vector<vector<type>> name(h, vector<type>(__VA_ARGS__))
#define vvv(type, name, h, w, ...) vector<vector<vector<type>>> name(h, vector<vector<type>>(w, vector<type>(__VA_ARGS__)))
#define vvvv(type, name, a, b, c, ...) \
  vector<vector<vector<vector<type>>>> name(a, vector<vector<vector<type>>>(b, vector<vector<type>>(c, vector<type>(__VA_ARGS__))))

// https://trap.jp/post/1224/
#define FOR1(a) for (ll _ = 0; _ < ll(a); ++_)
#define FOR2(i, a) for (ll i = 0; i < ll(a); ++i)
#define FOR3(i, a, b) for (ll i = a; i < ll(b); ++i)
#define FOR4(i, a, b, c) for (ll i = a; i < ll(b); i += (c))
#define FOR1_R(a) for (ll i = (a)-1; i >= ll(0); --i)
#define FOR2_R(i, a) for (ll i = (a)-1; i >= ll(0); --i)
#define FOR3_R(i, a, b) for (ll i = (b)-1; i >= ll(a); --i)
#define overload4(a, b, c, d, e, ...) e
#define overload3(a, b, c, d, ...) d
#define FOR(...) overload4(__VA_ARGS__, FOR4, FOR3, FOR2, FOR1)(__VA_ARGS__)
#define FOR_R(...) overload3(__VA_ARGS__, FOR3_R, FOR2_R, FOR1_R)(__VA_ARGS__)

#define FOR_subset(t, s) for (ll t = (s); t >= 0; t = (t == 0 ? -1 : (t - 1) & (s)))
#define all(x) x.begin(), x.end()
#define len(x) ll(x.size())
#define elif else if

#define eb emplace_back
#define mp make_pair
#define mt make_tuple
#define fi first
#define se second

#define stoi stoll

int popcnt(int x) { return __builtin_popcount(x); }
int popcnt(u32 x) { return __builtin_popcount(x); }
int popcnt(ll x) { return __builtin_popcountll(x); }
int popcnt(u64 x) { return __builtin_popcountll(x); }
int popcnt_mod_2(int x) { return __builtin_parity(x); }
int popcnt_mod_2(u32 x) { return __builtin_parity(x); }
int popcnt_mod_2(ll x) { return __builtin_parityll(x); }
int popcnt_mod_2(u64 x) { return __builtin_parityll(x); }
// (0, 1, 2, 3, 4) -> (-1, 0, 1, 1, 2)
int topbit(int x) { return (x == 0 ? -1 : 31 - __builtin_clz(x)); }
int topbit(u32 x) { return (x == 0 ? -1 : 31 - __builtin_clz(x)); }
int topbit(ll x) { return (x == 0 ? -1 : 63 - __builtin_clzll(x)); }
int topbit(u64 x) { return (x == 0 ? -1 : 63 - __builtin_clzll(x)); }
// (0, 1, 2, 3, 4) -> (-1, 0, 1, 0, 2)
int lowbit(int x) { return (x == 0 ? -1 : __builtin_ctz(x)); }
int lowbit(u32 x) { return (x == 0 ? -1 : __builtin_ctz(x)); }
int lowbit(ll x) { return (x == 0 ? -1 : __builtin_ctzll(x)); }
int lowbit(u64 x) { return (x == 0 ? -1 : __builtin_ctzll(x)); }

template <typename T>
T floor(T a, T b) {
  return a / b - (a % b && (a ^ b) < 0);
}
template <typename T>
T ceil(T x, T y) {
  return floor(x + y - 1, y);
}
template <typename T>
T bmod(T x, T y) {
  return x - y * floor(x, y);
}
template <typename T>
pair<T, T> divmod(T x, T y) {
  T q = floor(x, y);
  return {q, x - q * y};
}

template <typename T, typename U>
T SUM(const vector<U> &A) {
  T sm = 0;
  for (auto &&a: A) sm += a;
  return sm;
}

#define MIN(v) *min_element(all(v))
#define MAX(v) *max_element(all(v))
#define LB(c, x) distance((c).begin(), lower_bound(all(c), (x)))
#define UB(c, x) distance((c).begin(), upper_bound(all(c), (x)))
#define UNIQUE(x) sort(all(x)), x.erase(unique(all(x)), x.end()), x.shrink_to_fit()

template <typename T>
T POP(deque<T> &que) {
  T a = que.front();
  que.pop_front();
  return a;
}
template <typename T>
T POP(pq<T> &que) {
  T a = que.top();
  que.pop();
  return a;
}
template <typename T>
T POP(pqg<T> &que) {
  T a = que.top();
  que.pop();
  return a;
}
template <typename T>
T POP(vc<T> &que) {
  T a = que.back();
  que.pop_back();
  return a;
}

template <typename F>
ll binary_search(F check, ll ok, ll ng, bool check_ok = true) {
  if (check_ok) assert(check(ok));
  while (abs(ok - ng) > 1) {
    auto x = (ng + ok) / 2;
    (check(x) ? ok : ng) = x;
  }
  return ok;
}
template <typename F>
double binary_search_real(F check, double ok, double ng, int iter = 100) {
  FOR(iter) {
    double x = (ok + ng) / 2;
    (check(x) ? ok : ng) = x;
  }
  return (ok + ng) / 2;
}

template <class T, class S>
inline bool chmax(T &a, const S &b) {
  return (a < b ? a = b, 1 : 0);
}
template <class T, class S>
inline bool chmin(T &a, const S &b) {
  return (a > b ? a = b, 1 : 0);
}

// ? は -1
vc<int> s_to_vi(const string &S, char first_char) {
  vc<int> A(S.size());
  FOR(i, S.size()) { A[i] = (S[i] != '?' ? S[i] - first_char : -1); }
  return A;
}

template <typename T, typename U>
vector<T> cumsum(vector<U> &A, int off = 1) {
  int N = A.size();
  vector<T> B(N + 1);
  FOR(i, N) { B[i + 1] = B[i] + A[i]; }
  if (off == 0) B.erase(B.begin());
  return B;
}

// stable sort
template <typename T>
vector<int> argsort(const vector<T> &A) {
  vector<int> ids(len(A));
  iota(all(ids), 0);
  sort(all(ids), [&](int i, int j) { return (A[i] == A[j] ? i < j : A[i] < A[j]); });
  return ids;
}

// A[I[0]], A[I[1]], ...
template <typename T>
vc<T> rearrange(const vc<T> &A, const vc<int> &I) {
  vc<T> B(len(I));
  FOR(i, len(I)) B[i] = A[I[i]];
  return B;
}

template <typename T, typename... Vectors>
void concat(vc<T> &first, const Vectors &... others) {
  vc<T> &res = first;
  (res.insert(res.end(), others.begin(), others.end()), ...);
}
#endif
#line 1 "/home/maspy/compro/library/other/io.hpp"
#define FASTIO
#include <unistd.h>

// https://judge.yosupo.jp/submission/21623
namespace fastio {
static constexpr uint32_t SZ = 1 << 17;
char ibuf[SZ];
char obuf[SZ];
char out[100];
// pointer of ibuf, obuf
uint32_t pil = 0, pir = 0, por = 0;

struct Pre {
  char num[10000][4];
  constexpr Pre() : num() {
    for (int i = 0; i < 10000; i++) {
      int n = i;
      for (int j = 3; j >= 0; j--) {
        num[i][j] = n % 10 | '0';
        n /= 10;
      }
    }
  }
} constexpr pre;

inline void load() {
  memcpy(ibuf, ibuf + pil, pir - pil);
  pir = pir - pil + fread(ibuf + pir - pil, 1, SZ - pir + pil, stdin);
  pil = 0;
  if (pir < SZ) ibuf[pir++] = '\n';
}

inline void flush() {
  fwrite(obuf, 1, por, stdout);
  por = 0;
}

void rd(char &c) {
  do {
    if (pil + 1 > pir) load();
    c = ibuf[pil++];
  } while (isspace(c));
}

void rd(string &x) {
  x.clear();
  char c;
  do {
    if (pil + 1 > pir) load();
    c = ibuf[pil++];
  } while (isspace(c));
  do {
    x += c;
    if (pil == pir) load();
    c = ibuf[pil++];
  } while (!isspace(c));
}

template <typename T>
void rd_real(T &x) {
  string s;
  rd(s);
  x = stod(s);
}

template <typename T>
void rd_integer(T &x) {
  if (pil + 100 > pir) load();
  char c;
  do
    c = ibuf[pil++];
  while (c < '-');
  bool minus = 0;
  if constexpr (is_signed<T>::value || is_same_v<T, i128>) {
    if (c == '-') { minus = 1, c = ibuf[pil++]; }
  }
  x = 0;
  while ('0' <= c) { x = x * 10 + (c & 15), c = ibuf[pil++]; }
  if constexpr (is_signed<T>::value || is_same_v<T, i128>) {
    if (minus) x = -x;
  }
}

void rd(int &x) { rd_integer(x); }
void rd(ll &x) { rd_integer(x); }
void rd(i128 &x) { rd_integer(x); }
void rd(u32 &x) { rd_integer(x); }
void rd(u64 &x) { rd_integer(x); }
void rd(u128 &x) { rd_integer(x); }
void rd(double &x) { rd_real(x); }
void rd(long double &x) { rd_real(x); }
void rd(f128 &x) { rd_real(x); }

template <class T, class U>
void rd(pair<T, U> &p) {
  return rd(p.first), rd(p.second);
}
template <size_t N = 0, typename T>
void rd_tuple(T &t) {
  if constexpr (N < std::tuple_size<T>::value) {
    auto &x = std::get<N>(t);
    rd(x);
    rd_tuple<N + 1>(t);
  }
}
template <class... T>
void rd(tuple<T...> &tpl) {
  rd_tuple(tpl);
}

template <size_t N = 0, typename T>
void rd(array<T, N> &x) {
  for (auto &d: x) rd(d);
}
template <class T>
void rd(vc<T> &x) {
  for (auto &d: x) rd(d);
}

void read() {}
template <class H, class... T>
void read(H &h, T &... t) {
  rd(h), read(t...);
}

void wt(const char c) {
  if (por == SZ) flush();
  obuf[por++] = c;
}
void wt(const string s) {
  for (char c: s) wt(c);
}
void wt(const char *s) {
  size_t len = strlen(s);
  for (size_t i = 0; i < len; i++) wt(s[i]);
}

template <typename T>
void wt_integer(T x) {
  if (por > SZ - 100) flush();
  if (x < 0) { obuf[por++] = '-', x = -x; }
  int outi;
  for (outi = 96; x >= 10000; outi -= 4) {
    memcpy(out + outi, pre.num[x % 10000], 4);
    x /= 10000;
  }
  if (x >= 1000) {
    memcpy(obuf + por, pre.num[x], 4);
    por += 4;
  } else if (x >= 100) {
    memcpy(obuf + por, pre.num[x] + 1, 3);
    por += 3;
  } else if (x >= 10) {
    int q = (x * 103) >> 10;
    obuf[por] = q | '0';
    obuf[por + 1] = (x - q * 10) | '0';
    por += 2;
  } else
    obuf[por++] = x | '0';
  memcpy(obuf + por, out + outi + 4, 96 - outi);
  por += 96 - outi;
}

template <typename T>
void wt_real(T x) {
  ostringstream oss;
  oss << fixed << setprecision(15) << double(x);
  string s = oss.str();
  wt(s);
}

void wt(int x) { wt_integer(x); }
void wt(ll x) { wt_integer(x); }
void wt(i128 x) { wt_integer(x); }
void wt(u32 x) { wt_integer(x); }
void wt(u64 x) { wt_integer(x); }
void wt(u128 x) { wt_integer(x); }
void wt(double x) { wt_real(x); }
void wt(long double x) { wt_real(x); }
void wt(f128 x) { wt_real(x); }

template <class T, class U>
void wt(const pair<T, U> val) {
  wt(val.first);
  wt(' ');
  wt(val.second);
}
template <size_t N = 0, typename T>
void wt_tuple(const T t) {
  if constexpr (N < std::tuple_size<T>::value) {
    if constexpr (N > 0) { wt(' '); }
    const auto x = std::get<N>(t);
    wt(x);
    wt_tuple<N + 1>(t);
  }
}
template <class... T>
void wt(tuple<T...> tpl) {
  wt_tuple(tpl);
}
template <class T, size_t S>
void wt(const array<T, S> val) {
  auto n = val.size();
  for (size_t i = 0; i < n; i++) {
    if (i) wt(' ');
    wt(val[i]);
  }
}
template <class T>
void wt(const vector<T> val) {
  auto n = val.size();
  for (size_t i = 0; i < n; i++) {
    if (i) wt(' ');
    wt(val[i]);
  }
}

void print() { wt('\n'); }
template <class Head, class... Tail>
void print(Head &&head, Tail &&... tail) {
  wt(head);
  if (sizeof...(Tail)) wt(' ');
  print(forward<Tail>(tail)...);
}

// gcc expansion. called automaticall after main.
void __attribute__((destructor)) _d() { flush(); }
} // namespace fastio
using fastio::read;
using fastio::print;
using fastio::flush;

#if defined(LOCAL)
#define SHOW(...) SHOW_IMPL(__VA_ARGS__, SHOW6, SHOW5, SHOW4, SHOW3, SHOW2, SHOW1)(__VA_ARGS__)
#define SHOW_IMPL(_1, _2, _3, _4, _5, _6, NAME, ...) NAME
#define SHOW1(x) print(#x, "=", (x)), flush()
#define SHOW2(x, y) print(#x, "=", (x), #y, "=", (y)), flush()
#define SHOW3(x, y, z) print(#x, "=", (x), #y, "=", (y), #z, "=", (z)), flush()
#define SHOW4(x, y, z, w) print(#x, "=", (x), #y, "=", (y), #z, "=", (z), #w, "=", (w)), flush()
#define SHOW5(x, y, z, w, v) print(#x, "=", (x), #y, "=", (y), #z, "=", (z), #w, "=", (w), #v, "=", (v)), flush()
#define SHOW6(x, y, z, w, v, u) print(#x, "=", (x), #y, "=", (y), #z, "=", (z), #w, "=", (w), #v, "=", (v), #u, "=", (u)), flush()
#else
#define SHOW(...)
#endif

#define INT(...)   \
  int __VA_ARGS__; \
  read(__VA_ARGS__)
#define LL(...)   \
  ll __VA_ARGS__; \
  read(__VA_ARGS__)
#define U32(...)   \
  u32 __VA_ARGS__; \
  read(__VA_ARGS__)
#define U64(...)   \
  u64 __VA_ARGS__; \
  read(__VA_ARGS__)
#define STR(...)      \
  string __VA_ARGS__; \
  read(__VA_ARGS__)
#define CHAR(...)   \
  char __VA_ARGS__; \
  read(__VA_ARGS__)
#define DBL(...)      \
  double __VA_ARGS__; \
  read(__VA_ARGS__)

#define VEC(type, name, size) \
  vector<type> name(size);    \
  read(name)
#define VV(type, name, h, w)                     \
  vector<vector<type>> name(h, vector<type>(w)); \
  read(name)

void YES(bool t = 1) { print(t ? "YES" : "NO"); }
void NO(bool t = 1) { YES(!t); }
void Yes(bool t = 1) { print(t ? "Yes" : "No"); }
void No(bool t = 1) { Yes(!t); }
void yes(bool t = 1) { print(t ? "yes" : "no"); }
void no(bool t = 1) { yes(!t); }
#line 3 "main.cpp"

#line 1 "/home/maspy/compro/library/flow/k_ary_optimization.hpp"

#line 1 "/home/maspy/compro/library/flow/maxflow.hpp"
// incremental に辺を追加してよい
// 辺の容量の変更が可能
// 変更する capacity が F のとき、O((N+M)|F|) 時間で更新
template <typename Cap>
struct MaxFlow {
  struct Edge {
    int to, rev;
    Cap cap; // 残っている容量. したがって cap+flow が定数.
    Cap flow = 0;
  };

  const int N, source, sink;
  vvc<Edge> edges;
  vc<pair<int, int>> pos;
  vc<int> prog, level;
  vc<int> que;
  bool calculated;

  MaxFlow(int N, int source, int sink)
      : N(N),
        source(source),
        sink(sink),
        edges(N),
        calculated(0),
        flow_ans(0) {}

  void add(int frm, int to, Cap cap, Cap rev_cap = 0) {
    calculated = 0;
    assert(0 <= frm && frm < N);
    assert(0 <= to && to < N);
    assert(Cap(0) <= cap);
    int a = len(edges[frm]);
    int b = (frm == to ? a + 1 : len(edges[to]));
    pos.eb(frm, a);
    edges[frm].eb(Edge{to, b, cap, 0});
    edges[to].eb(Edge{frm, a, rev_cap, 0});
  }

  void change_capacity(int i, Cap after) {
    auto [frm, idx] = pos[i];
    auto& e = edges[frm][idx];
    Cap before = e.cap + e.flow;
    if (before < after) {
      calculated = (e.cap > 0);
      e.cap += after - before;
      return;
    }
    e.cap = after - e.flow;
    // 差分を押し戻す処理発生
    if (e.cap < 0) flow_push_back(e);
  }

  void flow_push_back(Edge& e0) {
    auto& re0 = edges[e0.to][e0.rev];
    int a = re0.to;
    int b = e0.to;
    /*
    辺 e0 の容量が正になるように戻す
    path-cycle 分解を考えれば、
    - uv 辺を含むサイクルを消す
    - suvt パスを消す
    前者は残余グラフで ab パス（flow_ans が変わらない）
    後者は残余グラフで tb, as パス
    */

    auto find_path = [&](int s, int t, Cap lim) -> Cap {
      vc<bool> vis(N);
      prog.assign(N, 0);
      auto dfs = [&](auto& dfs, int v, Cap f) -> Cap {
        if (v == t) return f;
        for (int& i = prog[v]; i < len(edges[v]); ++i) {
          auto& e = edges[v][i];
          if (vis[e.to] || e.cap <= Cap(0)) continue;
          vis[e.to] = 1;
          Cap a = dfs(dfs, e.to, min(f, e.cap));
          assert(a >= 0);
          if (a == Cap(0)) continue;
          e.cap -= a, e.flow += a;
          edges[e.to][e.rev].cap += a, edges[e.to][e.rev].flow -= a;
          return a;
        }
        return 0;
      };
      return dfs(dfs, s, lim);
    };

    while (e0.cap < 0) {
      Cap x = find_path(a, b, -e0.cap);
      if (x == Cap(0)) break;
      e0.cap += x, e0.flow -= x;
      re0.cap -= x, re0.flow += x;
    }
    Cap c = -e0.cap;
    while (c > 0 && a != source) {
      Cap x = find_path(a, source, c);
      assert(x > 0);
      c -= x;
    }
    c = -e0.cap;
    while (c > 0 && b != sink) {
      Cap x = find_path(sink, b, c);
      assert(x > 0);
      c -= x;
    }
    c = -e0.cap;
    e0.cap += c, e0.flow -= c;
    re0.cap -= c, re0.flow += c;
    flow_ans -= c;
  }

  // frm, to, flow
  vc<tuple<int, int, Cap>> get_flow_edges() {
    vc<tuple<int, int, Cap>> res;
    FOR(frm, N) {
      for (auto&& e: edges[frm]) {
        if (e.flow <= 0) continue;
        res.eb(frm, e.to, e.flow);
      }
    }
    return res;
  }

  vc<bool> vis;

  // 差分ではなくこれまでの総量
  Cap flow() {
    if (calculated) return flow_ans;
    calculated = true;
    while (set_level()) {
      prog.assign(N, 0);
      while (1) {
        Cap x = flow_dfs(source, infty<Cap>);
        if (x == 0) break;
        flow_ans += x;
        chmin(flow_ans, infty<Cap>);
        if (flow_ans == infty<Cap>) return flow_ans;
      }
    }
    return flow_ans;
  }

  // 最小カットの値および、カットを表す 01 列を返す
  pair<Cap, vc<int>> cut() {
    flow();
    vc<int> res(N);
    FOR(v, N) res[v] = (level[v] >= 0 ? 0 : 1);
    return {flow_ans, res};
  }

  // O(F(N+M)) くらい使って経路復元
  // simple path になる
  vvc<int> path_decomposition() {
    flow();
    auto edges = get_flow_edges();
    vvc<int> TO(N);
    for (auto&& [frm, to, flow]: edges) { FOR(flow) TO[frm].eb(to); }
    vvc<int> res;
    vc<int> vis(N);

    FOR(flow_ans) {
      vc<int> path = {source};
      vis[source] = 1;
      while (path.back() != sink) {
        int to = POP(TO[path.back()]);
        while (vis[to]) { vis[POP(path)] = 0; }
        path.eb(to), vis[to] = 1;
      }
      for (auto&& v: path) vis[v] = 0;
      res.eb(path);
    }
    return res;
  }

  void debug() {
    print("source", source);
    print("sink", sink);
    print("edges (frm, to, cap, flow)");
    FOR(v, N) {
      for (auto& e: edges[v]) {
        if (e.cap == 0 && e.flow == 0) continue;
        print(v, e.to, e.cap, e.flow);
      }
    }
  }

private:
  Cap flow_ans;

  bool set_level() {
    que.resize(N);
    level.assign(N, -1);
    level[source] = 0;
    int l = 0, r = 0;
    que[r++] = source;
    while (l < r) {
      int v = que[l++];
      for (auto&& e: edges[v]) {
        if (e.cap > 0 && level[e.to] == -1) {
          level[e.to] = level[v] + 1;
          if (e.to == sink) return true;
          que[r++] = e.to;
        }
      }
    }
    return false;
  }

  Cap flow_dfs(int v, Cap lim) {
    if (v == sink) return lim;
    Cap res = 0;
    for (int& i = prog[v]; i < len(edges[v]); ++i) {
      auto& e = edges[v][i];
      if (e.cap > 0 && level[e.to] == level[v] + 1) {
        Cap a = flow_dfs(e.to, min(lim, e.cap));
        if (a > 0) {
          e.cap -= a, e.flow += a;
          edges[e.to][e.rev].cap += a, edges[e.to][e.rev].flow -= a;
          res += a;
          lim -= a;
          if (lim == 0) break;
        }
      }
    }
    return res;
  }
};
#line 3 "/home/maspy/compro/library/flow/k_ary_optimization.hpp"

// ABC347G
template <typename T, bool MINIMIZE>
struct K_ary_Optimization {
  int n;
  vc<int> ks;
  vvc<int> idx;
  map<pair<int, int>, T> edges;
  int source, sink, nxt;
  T base_cost;

  // idx[i][j] が cut の source 側：val[i]>=j
  K_ary_Optimization(vc<int> ks) : n(len(ks)), ks(ks), base_cost(0) {
    source = 0, sink = 1, nxt = 2;
    for (auto& k: ks) {
      assert(k >= 1);
      vc<int> I(k + 1);
      I[0] = source, I[k] = sink;
      FOR(i, 1, k) { I[i] = nxt++; }
      idx.eb(I);
      if (k >= 3) { FOR(j, 1, k - 1) add_edge(I[j + 1], I[j], infty<T>); }
    }
  }

  // xi を 0, 1, ..., k-1 にするときにかかるコストを追加する。
  void add_1(int i, vc<T> cost) {
    assert(0 <= i && i < n && len(cost) == ks[i]);
    if (!MINIMIZE) {
      for (auto& x: cost) x = -x;
    }
    _add_1(i, cost);
  }

  void add_2(int i, int j, vvc<T> cost) {
    assert(0 <= i && i < n && 0 <= j && j < n && i != j);
    int H = ks[i], W = ks[j];
    assert(len(cost) == H);
    FOR(a, H) assert(len(cost[a]) == W);
    if (!MINIMIZE) { FOR(a, H) FOR(b, W) cost[a][b] = -cost[a][b]; }
    _add_2(i, j, cost);
  }

  // 最小値および、[0,k) 列を返す
  pair<T, vc<int>> calc() {
    MaxFlow<T> G(nxt, source, sink);
    for (auto&& [key, cap]: edges) {
      auto [frm, to] = key;
      G.add(frm, to, cap);
    }

    auto [val, cut] = G.cut();
    val += base_cost;
    vc<int> ANS(n);
    FOR(i, n) {
      FOR(j, 1, ks[i]) { ANS[i] += 1 - cut[idx[i][j]]; }
    }
    if (!MINIMIZE) val = -val;
    return {val, ANS};
  }

private:
  void add_base(T x) {
    base_cost += x;
    assert(-infty<T> < base_cost && base_cost < infty<T>);
  }

  void add_edge(int i, int j, T t) {
    assert(t >= 0);
    if (t == 0) return;
    pair<int, int> key = mp(i, j);
    edges[key] += t;
    assert(edges[key] <= infty<T>);
  }

  void _add_1(int i, vc<T> cost) {
    add_base(cost[0]);
    FOR_R(j, ks[i]) cost[j] -= cost[0];
    FOR(j, 1, ks[i]) {
      T x = cost[j] - cost[j - 1];
      // j 以上にすると x
      if (x > 0) add_edge(idx[i][j], sink, x);
      if (x < 0) add_base(x), add_edge(source, idx[i][j], -x);
    }
  }
  void _add_2(int i, int j, vvc<T> cost) {
    int H = ks[i], W = ks[j];
    _add_1(j, cost[0]);
    FOR_R(a, H) FOR(b, W) cost[a][b] -= cost[0][b];
    vc<T> tmp(H);
    FOR(a, H) tmp[a] = cost[a][W - 1];
    _add_1(i, tmp);
    FOR(a, H) FOR(b, W) cost[a][b] -= tmp[a];
    FOR(a, 1, H) FOR(b, W - 1) {
      T x = cost[a][b] + cost[a - 1][b + 1] - cost[a - 1][b] - cost[a][b + 1];
      assert(x >= 0); // monge
      add_edge(idx[i][a], idx[j][b + 1], x);
    }
  }
};
#line 5 "main.cpp"

void solve() {
  LL(H, W);
  VEC(string, G, H);
  K_ary_Optimization<ll, true> X(vc<int>(H * W, 4));
  auto idx = [&](int x, int y) -> int { return W * x + y; };
  FOR(x, H) FOR(y, W) {
    vi cost(4);
    if (x == 0 || y == W - 1) { cost[0] = infty<int>; }
    if (x == H - 1 || y == 0) { cost[3] = infty<int>; }
    if (G[x][y] == '#') cost[1] += 1, cost[2] += 1;
    X.add_1(idx(x, y), cost);
  }
  auto isin = [&](int x, int y) -> bool { return (0 <= x && x < H && 0 <= y && y < W); };
  int dx[] = {1, 1, 0, -1};
  int dy[] = {0, 1, 1, 1};
  FOR(x, H) FOR(y, W) {
    FOR(d, 4) {
      int x1 = x + dx[d], y1 = y + dy[d];
      if (!isin(x1, y1)) continue;
      int a = idx(x, y), b = idx(x1, y1);
      vv(ll, mat, 4, 4);
      mat[0][2] = infty<int>;
      mat[0][3] = 2 * infty<int>;
      mat[1][3] = infty<int>;
      mat[2][0] = infty<int>;
      mat[3][0] = 2 * infty<int>;
      mat[3][1] = infty<int>;
      X.add_2(a, b, mat);
    }
  }
  auto [val, ks] = X.calc();
  vv(int, A, H, W);
  FOR(x, H) FOR(y, W) A[x][y] = ks[idx(x, y)];
  FOR(x, H) SHOW(A[x]);
  print(val);
}

signed main() { solve(); }