#pragma region Macros
#ifdef noimi
#include "my_template.hpp"
#else
#pragma GCC target("avx2")
#pragma GCC optimize("O3")
#pragma GCC optimize("unroll-loops")
#include <immintrin.h>
#include <algorithm>
#include <array>
#include <bitset>
#include <cassert>
#include <cctype>
#include <cfenv>
#include <cfloat>
#include <chrono>
#include <cinttypes>
#include <climits>
#include <cmath>
#include <complex>
#include <cstdarg>
#include <cstddef>
#include <cstdint>
#include <cstdio>
#include <cstdlib>
#include <cstring>
#include <deque>
#include <fstream>
#include <functional>
#include <immintrin.h>
#include <initializer_list>
#include <iomanip>
#include <ios>
#include <iostream>
#include <istream>
#include <iterator>
#include <limits>
#include <list>
#include <map>
#include <memory>
#include <new>
#include <numeric>
#include <ostream>
#include <queue>
#include <random>
#include <set>
#include <sstream>
#include <stack>
#include <streambuf>
#include <string>
#include <tuple>
#include <type_traits>
#include <utility>
#include <variant>
#ifdef noimi
#define oj_local(a, b) b
#else
#define oj_local(a, b) a
#endif
#define LOCAL if(oj_local(0, 1))
#define OJ if(oj_local(1, 0))
using namespace std;
using ll = long long;
using ull = unsigned long long int;
using i128 = __int128_t;
using pii = pair<int, int>;
using pll = pair<ll, ll>;
using ld = long double;
template <typename T> using vc = vector<T>;
template <typename T> using vvc = vector<vc<T>>;
template <typename T> using vvvc = vector<vvc<T>>;
using vi = vc<int>;
using vl = vc<ll>;
using vpi = vc<pii>;
using vpl = vc<pll>;
template <class T> using pq = priority_queue<T>;
template <class T> using pqg = priority_queue<T, vector<T>, greater<T>>;
template <typename T> int si(const T &x) { return x.size(); }
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); }
vi iota(int n) {
vi a(n);
return iota(a.begin(), a.end(), 0), a;
}
template <typename T> vi iota(const vector<T> &a, bool greater = false) {
vi res(a.size());
iota(res.begin(), res.end(), 0);
sort(res.begin(), res.end(), [&](int i, int j) {
if(greater) return a[i] > a[j];
return a[i] < a[j];
});
return res;
}
// macros
#define overload5(a, b, c, d, e, name, ...) name
#define overload4(a, b, c, d, name, ...) name
#define endl '\n'
#define REP0(n) for(ll jidlsjf = 0; jidlsjf < n; ++jidlsjf)
#define REP1(i, n) for(ll i = 0; i < (n); ++i)
#define REP2(i, a, b) for(ll i = (a); i < (b); ++i)
#define REP3(i, a, b, c) for(ll i = (a); i < (b); i += (c))
#define rep(...) overload4(__VA_ARGS__, REP3, REP2, REP1, REP0)(__VA_ARGS__)
#define per0(n) for(int jidlsjf = 0; jidlsjf < (n); ++jidlsjf)
#define per1(i, n) for(ll i = (n)-1; i >= 0; --i)
#define per2(i, a, b) for(ll i = (a)-1; i >= b; --i)
#define per3(i, a, b, c) for(ll i = (a)-1; i >= (b); i -= (c))
#define per(...) overload4(__VA_ARGS__, per3, per2, per1, per0)(__VA_ARGS__)
#define fore0(a) rep(a.size())
#define fore1(i, a) for(auto &&i : a)
#define fore2(a, b, v) for(auto &&[a, b] : v)
#define fore3(a, b, c, v) for(auto &&[a, b, c] : v)
#define fore4(a, b, c, d, v) for(auto &&[a, b, c, d] : v)
#define fore(...) overload5(__VA_ARGS__, fore4, fore3, fore2, fore1, fore0)(__VA_ARGS__)
#define setbits(j, n) for(ll iiiii = (n), j = lowbit(iiiii); iiiii; iiiii ^= 1 << j, j = lowbit(iiiii))
#define perm(v) for(bool permrepflag = true; (permrepflag ? exchange(permrepflag, false) : next_permutation(all(v)));)
#define fi first
#define se second
#define pb push_back
#define ppb pop_back
#define ppf pop_front
#define eb emplace_back
#define drop(s) cout << #s << endl, exit(0)
#define si(c) (int)(c).size()
#define lb(c, x) distance((c).begin(), lower_bound(all(c), (x)))
#define lbg(c, x) distance((c).begin(), lower_bound(all(c), (x), greater{}))
#define ub(c, x) distance((c).begin(), upper_bound(all(c), (x)))
#define ubg(c, x) distance((c).begin(), upper_bound(all(c), (x), greater{}))
#define rng(v, l, r) v.begin() + (l), v.begin() + (r)
#define all(c) begin(c), end(c)
#define rall(c) rbegin(c), rend(c)
#define SORT(v) sort(all(v))
#define REV(v) reverse(all(v))
#define UNIQUE(x) SORT(x), x.erase(unique(all(x)), x.end())
template <typename T = ll, typename S> T SUM(const S &v) { return accumulate(all(v), T(0)); }
#define MIN(v) *min_element(all(v))
#define MAX(v) *max_element(all(v))
#define overload2(_1, _2, name, ...) name
#define vec(type, name, ...) vector<type> name(__VA_ARGS__)
#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__))))
constexpr pii dx4[4] = {pii{1, 0}, pii{0, 1}, pii{-1, 0}, pii{0, -1}};
constexpr pii dx8[8] = {{1, 0}, {1, 1}, {0, 1}, {-1, 1}, {-1, 0}, {-1, -1}, {0, -1}, {1, -1}};
namespace yesno_impl {
const string YESNO[2] = {"NO", "YES"};
const string YesNo[2] = {"No", "Yes"};
const string yesno[2] = {"no", "yes"};
const string firstsecond[2] = {"second", "first"};
const string FirstSecond[2] = {"Second", "First"};
const string possiblestr[2] = {"impossible", "possible"};
const string Possiblestr[2] = {"Impossible", "Possible"};
void YES(bool t = 1) { cout << YESNO[t] << endl; }
void NO(bool t = 1) { YES(!t); }
void Yes(bool t = 1) { cout << YesNo[t] << endl; }
void No(bool t = 1) { Yes(!t); }
void yes(bool t = 1) { cout << yesno[t] << endl; }
void no(bool t = 1) { yes(!t); }
void first(bool t = 1) { cout << firstsecond[t] << endl; }
void First(bool t = 1) { cout << FirstSecond[t] << endl; }
void possible(bool t = 1) { cout << possiblestr[t] << endl; }
void Possible(bool t = 1) { cout << Possiblestr[t] << endl; }
}; // namespace yesno_impl
using namespace yesno_impl;
#define INT(...) \
int __VA_ARGS__; \
IN(__VA_ARGS__)
#define INTd(...) \
int __VA_ARGS__; \
IN2(__VA_ARGS__)
#define LL(...) \
ll __VA_ARGS__; \
IN(__VA_ARGS__)
#define LLd(...) \
ll __VA_ARGS__; \
IN2(__VA_ARGS__)
#define STR(...) \
string __VA_ARGS__; \
IN(__VA_ARGS__)
#define CHR(...) \
char __VA_ARGS__; \
IN(__VA_ARGS__)
#define DBL(...) \
double __VA_ARGS__; \
IN(__VA_ARGS__)
#define VEC(type, name, size) \
vector<type> name(size); \
IN(name)
#define VECd(type, name, size) \
vector<type> name(size); \
IN2(name)
#define VEC2(type, name1, name2, size) \
vector<type> name1(size), name2(size); \
for(int i = 0; i < size; i++) IN(name1[i], name2[i])
#define VEC2d(type, name1, name2, size) \
vector<type> name1(size), name2(size); \
for(int i = 0; i < size; i++) IN2(name1[i], name2[i])
#define VEC3(type, name1, name2, name3, size) \
vector<type> name1(size), name2(size), name3(size); \
for(int i = 0; i < size; i++) IN(name1[i], name2[i], name3[i])
#define VEC3d(type, name1, name2, name3, size) \
vector<type> name1(size), name2(size), name3(size); \
for(int i = 0; i < size; i++) IN2(name1[i], name2[i], name3[i])
#define VEC4(type, name1, name2, name3, name4, size) \
vector<type> name1(size), name2(size), name3(size), name4(size); \
for(int i = 0; i < size; i++) IN(name1[i], name2[i], name3[i], name4[i]);
#define VEC4d(type, name1, name2, name3, name4, size) \
vector<type> name1(size), name2(size), name3(size), name4(size); \
for(int i = 0; i < size; i++) IN2(name1[i], name2[i], name3[i], name4[i]);
#define VV(type, name, h, w) \
vector<vector<type>> name(h, vector<type>(w)); \
IN(name)
#define VVd(type, name, h, w) \
vector<vector<type>> name(h, vector<type>(w)); \
IN2(name)
int scan() { return getchar(); }
void scan(int &a) { cin >> a; }
void scan(long long &a) { cin >> a; }
void scan(char &a) { cin >> a; }
void scan(double &a) { cin >> a; }
void scan(string &a) { cin >> a; }
template <class T, class S> void scan(pair<T, S> &p) { scan(p.first), scan(p.second); }
template <class T> void scan(vector<T> &);
template <class T> void scan(vector<T> &a) {
for(auto &i : a) scan(i);
}
template <class T> void scan(T &a) { cin >> a; }
void IN() {}
void IN2() {}
template <class Head, class... Tail> void IN(Head &head, Tail &...tail) {
scan(head);
IN(tail...);
}
template <class Head, class... Tail> void IN2(Head &head, Tail &...tail) {
scan(head);
--head;
IN2(tail...);
}
template <int p = -1> void pat() {}
template <int p = -1, class Head, class... Tail> void pat(Head &h, Tail &...tail) {
h += p;
pat<p>(tail...);
}
template <typename T, typename S> T ceil(T x, S y) {
assert(y);
return (y < 0 ? ceil(-x, -y) : (x > 0 ? (x + y - 1) / y : x / y));
}
template <typename T, typename S> T floor(T x, S y) {
assert(y);
return (y < 0 ? floor(-x, -y) : (x > 0 ? x / y : x / y - (x % y == 0 ? 0 : 1)));
}
template <typename T, typename S, typename U> U bigmul(const T &x, const S &y, const U &lim) { // clamp(x * y, -lim, lim)
if(x < 0 and y < 0) return bigmul(-x, -y, lim);
if(x < 0) return -bigmul(-x, y, lim);
if(y < 0) return -bigmul(x, -y, lim);
return y == 0 or x <= lim / y ? x * y : lim;
}
template <class T> T POW(T x, int n) {
T res = 1;
for(; n; n >>= 1, x *= x)
if(n & 1) res *= x;
return res;
}
template <class T, class S> T POW(T x, S n, const ll &mod) {
T res = 1;
x %= mod;
for(; n; n >>= 1, x = x * x % mod)
if(n & 1) res = res * x % mod;
return res;
}
vector<pll> factor(ll x) {
vector<pll> ans;
for(ll i = 2; i * i <= x; i++)
if(x % i == 0) {
ans.push_back({i, 1});
while((x /= i) % i == 0) ans.back().second++;
}
if(x != 1) ans.push_back({x, 1});
return ans;
}
template <class T> vector<T> divisor(T x) {
vector<T> ans;
for(T i = 1; i * i <= x; i++)
if(x % i == 0) {
ans.pb(i);
if(i * i != x) ans.pb(x / i);
}
return ans;
}
template <typename T> void zip(vector<T> &x) {
vector<T> y = x;
UNIQUE(y);
for(int i = 0; i < x.size(); ++i) { x[i] = lb(y, x[i]); }
}
template <class S> void fold_in(vector<S> &v) {}
template <typename Head, typename... Tail, class S> void fold_in(vector<S> &v, Head &&a, Tail &&...tail) {
for(auto e : a) v.emplace_back(e);
fold_in(v, tail...);
}
template <class S> void renumber(vector<S> &v) {}
template <typename Head, typename... Tail, class S> void renumber(vector<S> &v, Head &&a, Tail &&...tail) {
for(auto &&e : a) e = lb(v, e);
renumber(v, tail...);
}
template <class S, class... Args> vector<S> zip(vector<S> &head, Args &&...args) {
vector<S> v;
fold_in(v, head, args...);
sort(all(v)), v.erase(unique(all(v)), v.end());
renumber(v, head, args...);
return v;
}
template <typename S> void rearrange(const vector<S> &id) {}
template <typename S, typename T> void rearrange_exec(const vector<S> &id, vector<T> &v) {
vector<T> w(v.size());
rep(i, si(id)) w[i] = v[id[i]];
v.swap(w);
}
// 並び替える順番, 並び替える vector 達
template <typename S, typename Head, typename... Tail> void rearrange(const vector<S> &id, Head &a, Tail &...tail) {
rearrange_exec(id, a);
rearrange(id, tail...);
}
template <typename T> vector<T> RUI(const vector<T> &v) {
vector<T> res(v.size() + 1);
for(int i = 0; i < v.size(); i++) res[i + 1] = res[i] + v[i];
return res;
}
template <typename T> void zeta_supersetsum(vector<T> &f) {
int n = f.size();
for(int i = 1; i < n; i <<= 1) rep(b, n) if(!(i & b)) f[b] += f[b | i];
}
template <typename T> void zeta_subsetsum(vector<T> &f) {
int n = f.size();
for(int i = 1; i < n; i <<= 1) rep(b, n) if(!(i & b)) f[b | i] += f[b];
}
template <typename T> void mobius_subset(vector<T> &f) {
int n = f.size();
for(int i = 1; i < n; i <<= 1) rep(b, n) if(!(i & b)) f[b] -= f[b | i];
}
template <typename T> void mobius_superset(vector<T> &f) {
int n = f.size();
for(int i = 1; i < n; i <<= 1) rep(b, n) if(!(i & b)) f[b | i] -= f[b];
}
// 反時計周りに 90 度回転
template <typename T> void rot(vector<vector<T>> &v) {
if(empty(v)) return;
int n = v.size(), m = v[0].size();
vector<vector<T>> res(m, vector<T>(n));
rep(i, n) rep(j, m) res[m - 1 - j][i] = v[i][j];
v.swap(res);
}
vector<int> counter(const vector<int> &v, int max_num = -1) {
if(max_num == -1) max_num = MAX(v);
vector<int> res(max_num + 1);
fore(e, v) res[e]++;
return res;
}
// x in [l, r)
template <class T, class S> bool inc(const T &x, const S &l, const S &r) { return l <= x and x < r; }
template <class T, class S> bool inc(const T &x, const pair<S, S> &p) { return p.first <= x and x < p.second; }
// 便利関数
constexpr ll ten(int n) { return n == 0 ? 1 : ten(n - 1) * 10; }
constexpr ll tri(ll n) { return n * (n + 1) / 2; }
// l + ... + r
constexpr ll tri(ll l, ll r) { return (l + r) * (r - l + 1) / 2; }
ll max(int x, ll y) { return max((ll)x, y); }
ll max(ll x, int y) { return max(x, (ll)y); }
int min(int x, ll y) { return min((ll)x, y); }
int min(ll x, int y) { return min(x, (ll)y); }
// bit 演算系
#define bit(i) (1LL << i) // (1 << i)
#define test(b, i) (b >> i & 1) // b の i bit 目が立っているか
ll pow2(int i) { return 1LL << i; }
int topbit(signed t) { return t == 0 ? -1 : 31 - __builtin_clz(t); }
int topbit(ll t) { return t == 0 ? -1 : 63 - __builtin_clzll(t); }
int lowbit(signed a) { return a == 0 ? 32 : __builtin_ctz(a); }
int lowbit(ll a) { return a == 0 ? 64 : __builtin_ctzll(a); }
// int allbit(int n) { return (1 << n) - 1; }
constexpr ll mask(int n) { return (1LL << n) - 1; }
// int popcount(signed t) { return __builtin_popcount(t); }
// int popcount(ll t) { return __builtin_popcountll(t); }
int popcount(uint64_t t) { return __builtin_popcountll(t); }
static inline uint64_t popcount64(uint64_t x) {
uint64_t m1 = 0x5555555555555555ll;
uint64_t m2 = 0x3333333333333333ll;
uint64_t m4 = 0x0F0F0F0F0F0F0F0Fll;
uint64_t h01 = 0x0101010101010101ll;
x -= (x >> 1) & m1;
x = (x & m2) + ((x >> 2) & m2);
x = (x + (x >> 4)) & m4;
return (x * h01) >> 56;
}
bool ispow2(int i) { return i && (i & -i) == i; }
ll rnd(ll l, ll r) { //[l, r)
#ifdef noimi
static mt19937_64 gen;
#else
static mt19937_64 gen(chrono::steady_clock::now().time_since_epoch().count());
#endif
return uniform_int_distribution<ll>(l, r - 1)(gen);
}
ll rnd(ll n) { return rnd(0, n); }
template <class t> void random_shuffle(vc<t> &a) { rep(i, si(a)) swap(a[i], a[rnd(0, i + 1)]); }
int in() {
int x;
cin >> x;
return x;
}
ll lin() {
unsigned long long x;
cin >> x;
return x;
}
template <class T, class S> pair<T, S> operator-(const pair<T, S> &x) { return pair<T, S>(-x.first, -x.second); }
template <class T, class S> pair<T, S> operator-(const pair<T, S> &x, const pair<T, S> &y) { return pair<T, S>(x.fi - y.fi, x.se - y.se); }
template <class T, class S> pair<T, S> operator+(const pair<T, S> &x, const pair<T, S> &y) { return pair<T, S>(x.fi + y.fi, x.se + y.se); }
template <class T> pair<T, T> operator&(const pair<T, T> &l, const pair<T, T> &r) { return pair<T, T>(max(l.fi, r.fi), min(l.se, r.se)); }
template <class T, class S> pair<T, S> operator+=(pair<T, S> &l, const pair<T, S> &r) { return l = l + r; }
template <class T, class S> pair<T, S> operator-=(pair<T, S> &l, const pair<T, S> &r) { return l = l - r; }
template <class T> bool intersect(const pair<T, T> &l, const pair<T, T> &r) { return (l.se < r.se ? r.fi < l.se : l.fi < r.se); }
template <class T> vector<T> &operator++(vector<T> &v) {
fore(e, v) e++;
return v;
}
template <class T> vector<T> operator++(vector<T> &v, int) {
auto res = v;
fore(e, v) e++;
return res;
}
template <class T> vector<T> &operator--(vector<T> &v) {
fore(e, v) e--;
return v;
}
template <class T> vector<T> operator--(vector<T> &v, int) {
auto res = v;
fore(e, v) e--;
return res;
}
template <class T> void connect(vector<T> &l, const vector<T> &r) { fore(e, r) l.eb(e); }
template <class T> vector<T> operator+(const vector<T> &l, const vector<T> &r) {
vector<T> res(max(si(l), si(r)));
rep(i, si(l)) res[i] += l[i];
rep(i, si(r)) res[i] += r[i];
return res;
}
template <class T> vector<T> operator-(const vector<T> &l, const vector<T> &r) {
vector<T> res(max(si(l), si(r)));
rep(i, si(l)) res[i] += l[i];
rep(i, si(r)) res[i] -= r[i];
return res;
}
template <class T> vector<T> &operator+=(const vector<T> &l, const vector<T> &r) {
if(si(l) < si(r)) l.resize(si(r));
rep(i, si(r)) l[i] += r[i];
return l;
}
template <class T> vector<T> &operator-=(const vector<T> &l, const vector<T> &r) {
if(si(l) < si(r)) l.resize(si(r));
rep(i, si(r)) l[i] -= r[i];
return l;
}
template <class T> vector<T> &operator+=(vector<T> &v, const T &x) {
fore(e, v) e += x;
return v;
}
template <class T> vector<T> &operator-=(vector<T> &v, const T &x) {
fore(e, v) e -= x;
return v;
}
template <typename T> struct edge {
int from, to;
T cost;
int id;
edge(int to, T cost) : from(-1), to(to), cost(cost) {}
edge(int from, int to, T cost) : from(from), to(to), cost(cost) {}
edge(int from, int to, T cost, int id) : from(from), to(to), cost(cost), id(id) {}
constexpr bool operator<(const edge<T> &rhs) const noexcept { return cost < rhs.cost; }
edge &operator=(const int &x) {
to = x;
return *this;
}
operator int() const { return to; }
friend ostream operator<<(ostream &os, const edge &e) { return os << e.to; }
};
template <typename T> using Edges = vector<edge<T>>;
template <typename T = int> Edges<T> read_edges(int m, bool weighted = false) {
Edges<T> res;
res.reserve(m);
for(int i = 0; i < m; i++) {
int u, v, c = 0;
scan(u), scan(v), u--, v--;
if(weighted) scan(c);
res.eb(u, v, c, i);
}
return res;
}
using Tree = vector<vector<int>>;
using Graph = vector<vector<int>>;
template <class T> using Wgraph = vector<vector<edge<T>>>;
Graph getG(int n, int m = -1, bool directed = false, int margin = 1) {
Tree res(n);
if(m == -1) m = n - 1;
while(m--) {
int a, b;
cin >> a >> b;
a -= margin, b -= margin;
res[a].emplace_back(b);
if(!directed) res[b].emplace_back(a);
}
return res;
}
Graph getTreeFromPar(int n, int margin = 1) {
Graph res(n);
for(int i = 1; i < n; i++) {
int a;
cin >> a;
res[a - margin].emplace_back(i);
}
return res;
}
template <class T> Wgraph<T> getWg(int n, int m = -1, bool directed = false, int margin = 1) {
Wgraph<T> res(n);
if(m == -1) m = n - 1;
while(m--) {
int a, b;
T c;
scan(a), scan(b), scan(c);
a -= margin, b -= margin;
res[a].emplace_back(b, c);
if(!directed) res[b].emplace_back(a, c);
}
return res;
}
void add(Graph &G, int x, int y) { G[x].eb(y), G[y].eb(x); }
template <class S, class T> void add(Wgraph<S> &G, int x, int y, T c) { G[x].eb(y, c), G[y].eb(x, c); }
#define TEST \
INT(testcases); \
while(testcases--)
i128 abs(const i128 &x) { return x > 0 ? x : -x; }
istream &operator>>(istream &is, i128 &v) {
string s;
is >> s;
v = 0;
for(int i = 0; i < (int)s.size(); i++) {
if(isdigit(s[i])) { v = v * 10 + s[i] - '0'; }
}
if(s[0] == '-') { v *= -1; }
return is;
}
ostream &operator<<(ostream &os, const i128 &v) {
if(v == 0) { return (os << "0"); }
i128 num = v;
if(v < 0) {
os << '-';
num = -num;
}
string s;
for(; num > 0; num /= 10) { s.push_back((char)(num % 10) + '0'); }
reverse(s.begin(), s.end());
return (os << s);
}
namespace aux {
template <typename T, unsigned N, unsigned L> struct tp {
static void output(std::ostream &os, const T &v) {
os << std::get<N>(v) << (&os == &cerr ? ", " : " ");
tp<T, N + 1, L>::output(os, v);
}
};
template <typename T, unsigned N> struct tp<T, N, N> {
static void output(std::ostream &os, const T &v) { os << std::get<N>(v); }
};
} // namespace aux
template <typename... Ts> std::ostream &operator<<(std::ostream &os, const std::tuple<Ts...> &t) {
if(&os == &cerr) { os << '('; }
aux::tp<std::tuple<Ts...>, 0, sizeof...(Ts) - 1>::output(os, t);
if(&os == &cerr) { os << ')'; }
return os;
}
template <typename T, typename S, typename U> std::ostream &operator<<(std::ostream &os, const priority_queue<T, S, U> &_pq) {
auto pq = _pq;
vector<T> res;
while(!empty(pq)) res.emplace_back(pq.top()), pq.pop();
return os << res;
}
template <class T, class S> ostream &operator<<(ostream &os, const pair<T, S> &p) {
if(&os == &cerr) { return os << "(" << p.first << ", " << p.second << ")"; }
return os << p.first << " " << p.second;
}
template <class Ch, class Tr, class Container> std::basic_ostream<Ch, Tr> &operator<<(std::basic_ostream<Ch, Tr> &os, const Container &x) {
bool f = true;
if(&os == &cerr) os << "[";
for(auto &y : x) {
if(&os == &cerr)
os << (f ? "" : ", ") << y;
else
os << (f ? "" : " ") << y;
f = false;
}
if(&os == &cerr) os << "]";
return os;
}
#define dump(...) static_cast<void>(0)
#define dbg(...) static_cast<void>(0)
void OUT() { cout << endl; }
template <class Head, class... Tail> void OUT(const Head &head, const Tail &...tail) {
cout << head;
if(sizeof...(tail)) cout << ' ';
OUT(tail...);
}
template <typename T> static constexpr T inf = numeric_limits<T>::max() / 2;
template <class T, class S> constexpr pair<T, S> inf<pair<T, S>> = {inf<T>, inf<S>};
template <class T> void OUT2(const T &t, T INF = inf<T>, T res = -1) { OUT(t != INF ? t : res); }
template <class T> void OUT2(vector<T> &v, T INF = inf<T>, T res = -1) {
fore(e, v) if(e == INF) e = res;
OUT(v);
fore(e, v) if(e == res) e = INF;
}
template <class F> struct REC {
F f;
REC(F &&f_) : f(forward<F>(f_)) {}
template <class... Args> auto operator()(Args &&...args) const { return f(*this, forward<Args>(args)...); }
};
template <class S> vector<pair<S, int>> runLength(const vector<S> &v) {
vector<pair<S, int>> res;
for(auto &e : v) {
if(res.empty() or res.back().fi != e)
res.eb(e, 1);
else
res.back().se++;
}
return res;
}
vector<pair<char, int>> runLength(const string &v) {
vector<pair<char, int>> res;
for(auto &e : v) {
if(res.empty() or res.back().fi != e)
res.eb(e, 1);
else
res.back().se++;
}
return res;
}
struct string_converter {
char start = 0;
char type(const char &c) const { return (islower(c) ? 'a' : isupper(c) ? 'A' : isdigit(c) ? '0' : 0); }
int convert(const char &c) {
if(!start) start = type(c);
return c - start;
}
int convert(const char &c, const string &chars) { return chars.find(c); }
template <typename T> auto convert(const T &v) {
vector<decltype(convert(v[0]))> ret;
ret.reserve(size(v));
for(auto &&e : v) ret.emplace_back(convert(e));
return ret;
}
template <typename T> auto convert(const T &v, const string &chars) {
vector<decltype(convert(v[0], chars))> ret;
ret.reserve(size(v));
for(auto &&e : v) ret.emplace_back(convert(e, chars));
return ret;
}
int operator()(const char &v, char s = 0) {
start = s;
return convert(v);
}
int operator()(const char &v, const string &chars) { return convert(v, chars); }
template <typename T> auto operator()(const T &v, char s = 0) {
start = s;
return convert(v);
}
template <typename T> auto operator()(const T &v, const string &chars) { return convert(v, chars); }
} toint;
template <class T, class F> T bin_search(T ok, T ng, const F &f) {
while((ok > ng ? ok - ng : ng - ok) > 1) {
T mid = ok + ng >> 1;
(f(mid) ? ok : ng) = mid;
}
return ok;
}
template <class T, class F> T bin_search_double(T ok, T ng, const F &f, int iter = 80) {
while(iter--) {
// T mid = (ok + ng) / 2;
T mid = sqrtl(ok * ng);
(f(mid) ? ok : ng) = mid;
}
return ok;
}
struct Setup_io {
Setup_io() {
ios_base::sync_with_stdio(0), cin.tie(0), cout.tie(0);
cout << fixed << setprecision(11);
}
} setup_io;
#endif
#pragma endregion
#include <atcoder/convolution>
#include <vector>
namespace suisen {
template <typename mint> class inv_mods {
public:
inv_mods() = default;
inv_mods(int n) { ensure(n); }
const mint &operator[](int i) const {
ensure(i);
return invs[i];
}
static void ensure(int n) {
int sz = invs.size();
if(sz < 2) invs = {0, 1}, sz = 2;
if(sz < n + 1) {
invs.resize(n + 1);
for(int i = sz; i <= n; ++i) invs[i] = mint(mod - mod / i) * invs[mod % i];
}
}
private:
static std::vector<mint> invs;
static constexpr int mod = mint::mod();
};
template <typename mint> std::vector<mint> inv_mods<mint>::invs{};
template <typename mint> std::vector<mint> get_invs(const std::vector<mint> &vs) {
const int n = vs.size();
mint p = 1;
for(auto &e : vs) {
p *= e;
assert(e != 0);
}
mint ip = p.inv();
std::vector<mint> rp(n + 1);
rp[n] = 1;
for(int i = n - 1; i >= 0; --i) { rp[i] = rp[i + 1] * vs[i]; }
std::vector<mint> res(n);
for(int i = 0; i < n; ++i) {
res[i] = ip * rp[i + 1];
ip *= vs[i];
}
return res;
}
} // namespace suisen
#line 1 "library/convolution/relaxed_convolution_ntt.hpp"
#line 5 "library/convolution/relaxed_convolution_ntt.hpp"
namespace suisen {
// reference: https://qiita.com/Kiri8128/items/1738d5403764a0e26b4c
template <typename mint> struct RelaxedConvolutionNTT {
RelaxedConvolutionNTT() : _n(0), _f{}, _g{}, _h{} {}
mint append(const mint &fi, const mint &gi) {
static constexpr int threshold_log = 6;
static constexpr int threshold = 1 << threshold_log;
static constexpr int threshold_mask = threshold - 1;
++_n;
_f.push_back(fi), _g.push_back(gi);
const int q = _n >> threshold_log, r = _n & threshold_mask;
if(r == 0) {
if(q == (-q & q)) {
std::vector<mint> f_fft = _f;
std::vector<mint> g_fft = _g;
f_fft.resize(2 * _n);
g_fft.resize(2 * _n);
atcoder::internal::butterfly(f_fft);
atcoder::internal::butterfly(g_fft);
std::vector<mint> h(2 * _n);
for(int i = 0; i < 2 * _n; ++i) { h[i] = f_fft[i] * g_fft[i]; }
atcoder::internal::butterfly_inv(h);
ensure(2 * _n);
const mint z = mint(2 * _n).inv();
for(int i = _n - 1; i < 2 * _n; ++i) { _h[i] += h[i] * z; }
_f_fft.push_back(std::move(f_fft));
_g_fft.push_back(std::move(g_fft));
} else {
const int log_q = __builtin_ctz(q);
const int k = (-q & q) << threshold_log;
std::vector<mint> f_fft(_f.end() - k, _f.end());
std::vector<mint> g_fft(_g.end() - k, _g.end());
f_fft.resize(2 * k);
g_fft.resize(2 * k);
atcoder::internal::butterfly(f_fft);
atcoder::internal::butterfly(g_fft);
std::vector<mint> h(2 * k);
for(int i = 0; i < 2 * k; ++i) { h[i] = _f_fft[log_q + 1][i] * g_fft[i] + f_fft[i] * _g_fft[log_q + 1][i]; }
atcoder::internal::butterfly_inv(h);
const mint z = mint(2 * k).inv();
for(int i = 0; i < k; ++i) { _h[_n - 1 + i] += h[k - 1 + i] * z; }
}
} else {
// naive convolve
ensure(_n);
for(int i = 0; i < r; ++i) { _h[_n - 1] += _f[i] * _g[_n - 1 - i]; }
if(_n != r) {
for(int i = 0; i < r; ++i) { _h[_n - 1] += _f[_n - i - 1] * _g[i]; }
}
}
return _h[_n - 1];
}
const mint &operator[](int i) const { return _h[i]; }
std::vector<mint> get() const { return _h; }
private:
int _n;
std::vector<mint> _f, _g, _h;
std::vector<std::vector<mint>> _f_fft, _g_fft;
void ensure(std::size_t n) {
if(_h.size() < n) _h.resize(n);
}
};
} // namespace suisen
#line 1 "library/math/modint_extension.hpp"
#include <cassert>
#include <optional>
/**
* refernce: https://37zigen.com/tonelli-shanks-algorithm/
* calculates x s.t. x^2 = a mod p in O((log p)^2).
*/
template <typename mint> std::optional<mint> safe_sqrt(mint a) {
static int p = mint::mod();
if(a == 0) return std::make_optional(0);
if(p == 2) return std::make_optional(a);
if(a.pow((p - 1) / 2) != 1) return std::nullopt;
mint b = 1;
while(b.pow((p - 1) / 2) == 1) ++b;
static int tlz = __builtin_ctz(p - 1), q = (p - 1) >> tlz;
mint x = a.pow((q + 1) / 2);
b = b.pow(q);
for(int shift = 2; x * x != a; ++shift) {
mint e = a.inv() * x * x;
if(e.pow(1 << (tlz - shift)) != 1) x *= b;
b *= b;
}
return std::make_optional(x);
}
/**
* calculates x s.t. x^2 = a mod p in O((log p)^2).
* if not exists, raises runtime error.
*/
template <typename mint> auto sqrt(mint a) -> decltype(mint::mod(), mint()) { return *safe_sqrt(a); }
template <typename mint> auto log(mint a) -> decltype(mint::mod(), mint()) {
assert(a == 1);
return 0;
}
template <typename mint> auto exp(mint a) -> decltype(mint::mod(), mint()) {
assert(a == 0);
return 1;
}
template <typename mint, typename T> auto pow(mint a, T b) -> decltype(mint::mod(), mint()) { return a.pow(b); }
template <typename mint> auto inv(mint a) -> decltype(mint::mod(), mint()) { return a.inv(); }
#line 8 "library/polynomial/formal_power_series_relaxed.hpp"
namespace suisen {
template <typename mint> struct RelaxedInv {
mint append(const mint &fi) {
const int i = g.size();
if(i == 0) {
assert(fi != 0);
g.push_back(fi.inv());
} else {
g.push_back(-g[0] * fg.append(fi, g[i - 1]));
}
return g.back();
}
mint operator[](int i) const { return g[i]; }
private:
std::vector<mint> g;
RelaxedConvolutionNTT<mint> fg;
};
template <typename mint> struct RelaxedExp {
mint append(const mint &fi) {
static inv_mods<mint> invs;
const int i = g.size();
if(i == 0) {
assert(fi == 0);
g.push_back(1);
} else {
g.push_back(df_g.append(i * fi, g[i - 1]) * invs[i]);
}
return g.back();
}
mint operator[](int i) const { return g[i]; }
private:
std::vector<mint> g;
RelaxedConvolutionNTT<mint> df_g;
};
template <typename mint> struct RelaxedLog {
mint append(const mint &fi) {
static inv_mods<mint> invs;
f.push_back(fi);
const int i = g.size();
if(i == 0) {
assert(f[i] == 1);
g.push_back(0);
} else if(i == 1) {
g.push_back(f[i]);
} else {
g.push_back(f[i] - fg.append((i - 1) * g[i - 1], f[i - 1]) * invs[i]);
}
return g.back();
}
mint operator[](int i) const { return g[i]; }
private:
std::vector<mint> f, g;
RelaxedConvolutionNTT<mint> fg;
};
template <typename mint> struct RelaxedPow {
RelaxedPow(long long k = 0) : k(k) {}
mint append(const mint &fi) {
if(k == 0) { return g.emplace_back(g.empty() ? 1 : 0); }
static inv_mods<mint> invs;
if(is_zero) {
if(fi == 0) {
z = std::min(z + k, 1000000000LL);
} else {
is_zero = false;
inv_base = fi.inv();
}
}
if(not is_zero) { f.push_back(fi); }
if(index < z) {
g.push_back(0);
} else if(index == z) {
g.push_back(f[0].pow(k));
} else {
int i = index - z;
mint v1 = fg1.append(mint(k - (i - 1)) * g[z + i - 1], f[i]);
mint v2 = fg2.append(g[z + i - 1], mint(k) * (i - 1) * f[i]);
g.push_back((v1 + v2) * inv_base * invs[i]);
}
++index;
return g.back();
}
mint operator[](int i) const { return g[i]; }
private:
long long k;
long long z = 0;
long long index = 0;
bool is_zero = true;
mint inv_base = 0;
std::vector<mint> f, g;
RelaxedConvolutionNTT<mint> fg1;
RelaxedConvolutionNTT<mint> fg2;
};
template <typename mint> struct RelaxedSqrt {
std::optional<mint> append(const mint &fi) {
if(g.empty()) {
auto opt_g0 = safe_sqrt(fi);
if(not opt_g0) return std::nullopt;
mint g0 = *opt_g0;
c = (2 * g0).inv();
return g.emplace_back(g0);
} else if(g.size() == 1) {
return g.emplace_back(c * fi);
} else {
mint gi = c * (fi - gg.append(g.back(), g.back()));
return g.emplace_back(gi);
}
}
mint operator[](int i) const { return g[i]; }
private:
mint c = 0;
std::vector<mint> g;
RelaxedConvolutionNTT<mint> gg;
};
} // namespace suisen
#include <vector>
namespace suisen {
// reference: https://qiita.com/Kiri8128/items/1738d5403764a0e26b4c
template <typename T> struct RelaxedConvolution {
using value_type = T;
using polynomial_type = std::vector<value_type>;
using convolution_type = polynomial_type (*)(const polynomial_type &, const polynomial_type &);
RelaxedConvolution() = default;
RelaxedConvolution(const convolution_type &convolve) : _convolve(convolve), _n(0), _f{}, _g{}, _h{} {}
void set_convolve_function(const convolution_type &convolve) { _convolve = convolve; }
value_type append(const value_type &fi, const value_type &gi) {
++_n;
_f.push_back(fi), _g.push_back(gi);
for(int p = 1;; p <<= 1) {
int l1 = _n - p, r1 = _n, l2 = p - 1, r2 = l2 + p;
add(l1 + l2, range_convolve(l1, r1, l2, r2));
if(l1 == l2) break;
add(l1 + l2, range_convolve(l2, r2, l1, r1));
if(not(_n & p)) break;
}
return _h[_n - 1];
}
const value_type &operator[](int i) const { return _h[i]; }
polynomial_type get() const { return _h; }
private:
convolution_type _convolve = [](const polynomial_type &, const polynomial_type &) -> polynomial_type { assert(false); };
int _n;
polynomial_type _f, _g, _h;
polynomial_type range_convolve(int l1, int r1, int l2, int r2) {
return _convolve(polynomial_type(_f.begin() + l1, _f.begin() + r1), polynomial_type(_g.begin() + l2, _g.begin() + r2));
}
void add(std::size_t bias, const polynomial_type &h) {
if(_h.size() < bias + h.size()) _h.resize(bias + h.size());
for(std::size_t i = 0; i < h.size(); ++i) _h[bias + i] += h[i];
}
};
} // namespace suisen
int main() {
INT(n);
using mint = atcoder::modint998244353;
vector<mint> f{0}, g{0};
auto conv = [](const auto &a, const auto &b) { return atcoder::convolution(a, b); };
suisen::RelaxedConvolution<mint> fg{conv}, ff{conv}, expfgf2{conv}, get_f{conv}, get_g{conv};
suisen::RelaxedExp<mint> expg, expfg;
suisen::RelaxedInv<mint> invone_f;
fg.append(0, 0);
ff.append(0, 0);
expfgf2.append(1, 0);
expg.append(0);
expfg.append(0);
invone_f.append(1);
get_f.append(expg[0] - expfg[0], 1);
get_g.append(expg[0] - expfgf2[0], 1);
rep(i, 1, n + 1) {
f.eb(get_f[i - 1]);
g.eb(get_g[i - 1]);
fg.append(f[i], g[i]);
ff.append(f[i], f[i]);
expfg.append(fg[i]);
expg.append(g[i]);
expfgf2.append(expfg[i], ff[i]);
invone_f.append(-f[i]);
get_f.append((expg[i] - expfg[i]), invone_f[i]);
get_g.append((expg[i] - expfgf2[i]), invone_f[i]);
}
mint res = f[n] + g[n];
rep(i, 1, n + 1) res *= i;
OUT(res.val());
}