long long mod_inv(long long a, long long mod){
// a^{-1} = 1/a mod p (拡張Euclidの互除法)
long long b = mod, u = 1, v = 0;
while (b) {
long long t = a / b;
a -= t * b; swap(a, b);
u -= t * v; swap(u, v);
}
u %= mod;
if (u < 0) u += mod;
return u;
}
long long mod_pow(long long a, long long n, long long mod){
long long ret = 1;
long long p = a % mod;
while(n){
if(n&1) ret = ret * p % mod;
p = p * p % mod;
n >>= 1;
}
return ret;
}
template<int mod, int primitive_root>
class NTT {
public:
int get_mod() const { return mod; }
void _ntt(vector<long long>& a, int sign) {
const int n = a.size();
assert((n ^ (n&-n)) == 0); //n = 2^k
const int g = 3; //g is primitive root of mod
int h = (int)mod_pow(g, (mod - 1) / n, mod); // h^n = 1
if (sign == -1) h = (int)mod_inv(h, mod); //h = h^-1 % mod
//bit reverse
int i = 0;
for (int j=1; j<n-1; ++j) {
for (int k = n >> 1; k >(i ^= k); k >>= 1);
if (j < i) swap(a[i], a[j]);
}
for (int m=1; m<n; m*=2) {
const int m2 = 2 * m;
const long long base = mod_pow(h, n / m2, mod);
long long w = 1;
for(int x=0; x<m; x++) {
for (int s = x; s < n; s += m2) {
long long u = a[s];
long long d = a[s + m] * w % mod;
a[s] = u + d;
if (a[s] >= mod) a[s] -= mod;
a[s + m] = u - d;
if (a[s + m] < 0) a[s + m] += mod;
}
w = w * base % mod;
}
}
for (auto& x : a) if (x < 0) x += mod;
}
void ntt(vector<long long>& input) {
_ntt(input, 1);
}
void intt(vector<long long>& input) {
_ntt(input, -1);
const int n_inv = mod_inv(input.size(), mod);
for (auto& x : input) x = x * n_inv % mod;
}
// 畳み込み演算を行う
vector<long long> convolution(const vector<long long>& a, const vector<long long>& b){
int ntt_size = 1;
while (ntt_size < (int)(a.size()+b.size())) ntt_size *= 2;
vector<long long> _a = a, _b = b;
_a.resize(ntt_size); _b.resize(ntt_size);
ntt(_a);
ntt(_b);
for(int i=0; i<ntt_size; i++){
(_a[i] *= _b[i]) %= mod;
}
intt(_a);
return _a;
}
};
typedef NTT<167772161,3> NTT_1;
typedef NTT<469762049,3> NTT_2;
typedef NTT<1224736769,3> NTT_3;
// garnerのアルゴリズムを直書きしたversion,速い
vector<long long> int32mod_convolution(vector<long long> a, vector<long long> b,int mod){
for (auto& x : a) x %= mod;
for (auto& x : b) x %= mod;
NTT_1 ntt1; NTT_2 ntt2; NTT_3 ntt3;
assert(ntt1.get_mod() < ntt2.get_mod() && ntt2.get_mod() < ntt3.get_mod());
auto x = ntt1.convolution(a, b);
auto y = ntt2.convolution(a, b);
auto z = ntt3.convolution(a, b);
// garnerのアルゴリズムを極力高速化した
const long long m1 = ntt1.get_mod(), m2 = ntt2.get_mod(), m3 = ntt3.get_mod();
const long long m1_inv_m2 = mod_inv(m1, m2);
const long long m12_inv_m3 = mod_inv(m1 * m2, m3);
const long long m12_mod = m1 * m2 % mod;
vector<long long> ret(x.size());
for(int i=0; i<(int)x.size(); i++){
long long v1 = (y[i] - x[i]) * m1_inv_m2 % m2;
if (v1 < 0) v1 += m2;
long long v2 = (z[i] - (x[i] + m1 * v1) % m3) * m12_inv_m3 % m3;
if (v2 < 0) v2 += m3;
long long constants3 = (x[i] + m1 * v1 + m12_mod * v2) % mod;
if (constants3 < 0) constants3 += mod;
ret[i] = constants3;
}
return ret;
}
//2^23より大きく,primitive rootに3を持つもの
// const int mods[] = { 1224736769, 469762049, 167772161, 595591169, 645922817, 897581057, 998244353 };