This documentation is automatically generated by online-judge-tools/verification-helper
#include "other/sliding_puzzle_solver.hpp"
#include "linalg/transpose.hpp"
#include "seq/inversion.hpp"
/*
O(HW(H+W))
空マスは -1 (unique)
同じ値が複数あってもよい
操作回数を K として、長さ K+1 の空マスの座標列をかえす
*/
struct Slinding_Puzzle_Solver {
using P = pair<int, int>;
vc<P> solve(vvc<int> A, vvc<int> B) {
int H = len(A), W = len(A[0]);
auto find = [&](vvc<int>& A, int k) -> P {
FOR(x, H) FOR(y, W) if (A[x][y] == k) return {x, y};
assert(0);
return {0, 0};
};
auto [ax, ay] = find(A, -1);
auto [bx, by] = find(B, -1);
vc<P> ANS_1, ANS_2;
while (ax > 0) { ANS_1.eb(ax, ay), swap(A[ax][ay], A[ax - 1][ay]), --ax; }
while (ay > 0) { ANS_1.eb(ax, ay), swap(A[ax][ay], A[ax][ay - 1]), --ay; }
while (bx > 0) { ANS_2.eb(bx, by), swap(B[bx][by], B[bx - 1][by]), --bx; }
while (by > 0) { ANS_2.eb(bx, by), swap(B[bx][by], B[bx][by - 1]), --by; }
vc<P> ANS = solve_00(A, B);
if (ANS.empty()) return {};
reverse(all(ANS_2));
return concat(ANS_1, ANS, ANS_2);
}
private:
vc<P> solve_00(vvc<int> A, vvc<int> B) {
assert(A[0][0] == -1 && B[0][0] == -1);
int H = len(A), W = len(A[0]);
if (H == 1 || W == 1) {
if (A != B) return {};
vc<P> ANS;
ANS.eb(0, 0);
return ANS;
}
vc<P> XYA, XYB;
FOR(x, H) FOR(y, W) XYA.eb(x, y), XYB.eb(x, y);
sort(all(XYA), [&](auto& a, auto& b) -> bool { return A[a.fi][a.se] < A[b.fi][b.se]; });
sort(all(XYB), [&](auto& a, auto& b) -> bool { return B[a.fi][a.se] < B[b.fi][b.se]; });
auto check = [&]() -> bool {
vc<int> S, T;
FOR(i, H * W) {
auto [x1, y1] = XYA[i];
auto [x2, y2] = XYB[i];
if (A[x1][y1] != B[x2][y2]) return 0;
S.eb(W * x1 + y1);
T.eb(W * x2 + y2);
}
ll x = inversion_between(S, T);
return x % 2 == 0;
};
if (!check()) {
FOR(i, H * W - 1) {
auto [x1, y1] = XYA[i];
auto [x2, y2] = XYA[i + 1];
if (A[x1][y1] != A[x2][y2]) continue;
swap(XYA[i], XYA[i + 1]);
break;
}
if (!check()) return {};
}
vv(P, X, H, W);
FOR(i, H * W) {
auto [x1, y1] = XYA[i];
auto [x2, y2] = XYB[i];
X[x1][y1] = {x2, y2};
}
vc<P> ANS;
ANS.eb(0, 0);
solve_sort(X, ANS, false);
return ANS;
}
// 移動先の座標の列を並べたグリッドを与える.
// (0,0) が空マス
void solve_sort(vvc<pair<int, int>>& A, vc<P>& ANS, bool tr) {
int H = len(A), W = len(A[0]);
vv(P, pos, H, W);
FOR(x, H) FOR(y, W) {
P p = A[x][y];
pos[p.fi][p.se] = {x, y};
}
auto [px, py] = pos[0][0];
auto ope = [&](int x, int y) -> void {
assert(abs(px - x) + abs(py - y) == 1);
swap(A[px][py], A[x][y]);
if (!tr) ANS.eb(x, y);
if (tr) ANS.eb(y, x);
pos[A[px][py].fi][A[px][py].se] = {px, py};
px = x, py = y;
pos[A[px][py].fi][A[px][py].se] = {px, py};
};
if (H == 2 && W == 2) {
auto check = [&]() -> bool {
FOR(x, H) FOR(y, W) if (A[x][y].fi != x || A[x][y].se != y) return 0;
return 1;
};
while (!check()) {
if (px == 0 && py == 0) ope(1, 0);
if (px == 1 && py == 0) ope(1, 1);
if (px == 1 && py == 1) ope(0, 1);
if (px == 0 && py == 1) ope(0, 0);
}
return;
}
if (H < W) {
FOR(x, H) FOR(y, W) { swap(A[x][y].fi, A[x][y].se); }
A = transpose(A, H, W);
solve_sort(A, ANS, !tr);
return;
}
assert(H >= 3 && W >= 2);
// 最後の行をそろえる
FOR_R(y, 1, W) {
auto& [tx, ty] = pos[H - 1][y];
if (px == H - 1) ope(px - 1, py);
while (px - 1 > tx) ope(px - 1, py), tie(tx, ty) = pos[H - 1][y];
while (px + 1 < tx) ope(px + 1, py), tie(tx, ty) = pos[H - 1][y];
if (px == tx) {
if (px == 0)
ope(px + 1, py);
else
ope(px - 1, py);
}
assert(abs(px - tx) == 1);
while (py < ty) ope(px, py + 1);
while (py > ty) ope(px, py - 1);
if (px == tx + 1) ope(px - 1, py), tie(tx, ty) = pos[H - 1][y];
assert(px == tx - 1 && py == ty);
while (ty < y) { ope(px, py + 1), ope(px + 1, py), ope(px, py - 1), ope(px - 1, py), ope(px, py + 1); }
while (ty > y) { ope(px, py - 1), ope(px + 1, py), ope(px, py + 1), ope(px - 1, py), ope(px, py - 1); }
tie(tx, ty) = pos[H - 1][y];
while (tx < H - 1) {
ope(px, py - 1), ope(px + 1, py), ope(px + 1, py), ope(px, py + 1), ope(px - 1, py);
tie(tx, ty) = pos[H - 1][y];
}
assert(A[H - 1][y] == (pair<int, int>{H - 1, y}));
}
auto& [tx, ty] = pos[H - 1][0];
if (px == H - 1) ope(px - 1, py);
while (px - 1 > tx) ope(px - 1, py), tie(tx, ty) = pos[H - 1][0];
while (px + 1 < tx) ope(px + 1, py), tie(tx, ty) = pos[H - 1][0];
if (px == tx) {
if (px == 0)
ope(px + 1, py);
else
ope(px - 1, py);
}
tie(tx, ty) = pos[H - 1][0];
assert(abs(px - tx) == 1);
while (py < ty) ope(px, py + 1);
while (py > ty) ope(px, py - 1);
if (px == tx + 1) ope(px - 1, py);
tie(tx, ty) = pos[H - 1][0];
if (tx < H - 1) {
assert(px == tx - 1 && py == ty);
while (ty > 0) { ope(px, py - 1), ope(px + 1, py), ope(px, py + 1), ope(px - 1, py), ope(px, py - 1); }
tie(tx, ty) = pos[H - 1][0];
while (tx < H - 2) { ope(px, py + 1), ope(px + 1, py), ope(px + 1, py), ope(px, py - 1), ope(px - 1, py); }
ope(px + 1, py), ope(px + 1, py), ope(px, py + 1);
ope(px - 1, py), ope(px, py - 1), ope(px - 1, py);
ope(px, py + 1), ope(px + 1, py), ope(px + 1, py);
ope(px, py - 1), ope(px - 1, py);
}
FOR(y, W) assert(A[H - 1][y] == (pair<int, int>{H - 1, y}));
POP(A);
solve_sort(A, ANS, tr);
}
};
#line 1 "linalg/transpose.hpp"
template <typename VC>
vc<VC> transpose(const vc<VC>& A, int H = -1, int W = -1) {
if (H == -1) { H = len(A), W = (len(A) == 0 ? 0 : len(A[0])); }
vc<VC> B(W, VC(H, 0));
FOR(x, H) FOR(y, W) B[y][x] = A[x][y];
return B;
}
#line 2 "ds/fenwicktree/fenwicktree_01.hpp"
#line 2 "alg/monoid/add.hpp"
template <typename E>
struct Monoid_Add {
using X = E;
using value_type = X;
static constexpr X op(const X &x, const X &y) noexcept { return x + y; }
static constexpr X inverse(const X &x) noexcept { return -x; }
static constexpr X power(const X &x, ll n) noexcept { return X(n) * x; }
static constexpr X unit() { return X(0); }
static constexpr bool commute = true;
};
#line 3 "ds/fenwicktree/fenwicktree.hpp"
template <typename Monoid>
struct FenwickTree {
using G = Monoid;
using MX = Monoid;
using E = typename G::value_type;
int n;
vector<E> dat;
E total;
FenwickTree() {}
FenwickTree(int n) { build(n); }
template <typename F>
FenwickTree(int n, F f) {
build(n, f);
}
FenwickTree(const vc<E>& v) { build(v); }
void build(int m) {
n = m;
dat.assign(m, G::unit());
total = G::unit();
}
void build(const vc<E>& v) {
build(len(v), [&](int i) -> E { return v[i]; });
}
template <typename F>
void build(int m, F f) {
n = m;
dat.clear();
dat.reserve(n);
total = G::unit();
FOR(i, n) { dat.eb(f(i)); }
for (int i = 1; i <= n; ++i) {
int j = i + (i & -i);
if (j <= n) dat[j - 1] = G::op(dat[i - 1], dat[j - 1]);
}
total = prefix_sum(m);
}
E prod_all() { return total; }
E sum_all() { return total; }
E sum(int k) { return prefix_sum(k); }
E prod(int k) { return prefix_prod(k); }
E prefix_sum(int k) { return prefix_prod(k); }
E prefix_prod(int k) {
chmin(k, n);
E ret = G::unit();
for (; k > 0; k -= k & -k) ret = G::op(ret, dat[k - 1]);
return ret;
}
E sum(int L, int R) { return prod(L, R); }
E prod(int L, int R) {
chmax(L, 0), chmin(R, n);
if (L == 0) return prefix_prod(R);
assert(0 <= L && L <= R && R <= n);
E pos = G::unit(), neg = G::unit();
while (L < R) { pos = G::op(pos, dat[R - 1]), R -= R & -R; }
while (R < L) { neg = G::op(neg, dat[L - 1]), L -= L & -L; }
return G::op(pos, G::inverse(neg));
}
vc<E> get_all() {
vc<E> res(n);
FOR(i, n) res[i] = prod(i, i + 1);
return res;
}
void add(int k, E x) { multiply(k, x); }
void multiply(int k, E x) {
static_assert(G::commute);
total = G::op(total, x);
for (++k; k <= n; k += k & -k) dat[k - 1] = G::op(dat[k - 1], x);
}
void set(int k, E x) { add(k, G::op(G::inverse(prod(k, k + 1)), x)); }
template <class F>
int max_right(const F check, int L = 0) {
assert(check(G::unit()));
E s = G::unit();
int i = L;
// 2^k 進むとダメ
int k = [&]() {
while (1) {
if (i % 2 == 1) { s = G::op(s, G::inverse(dat[i - 1])), i -= 1; }
if (i == 0) { return topbit(n) + 1; }
int k = lowbit(i) - 1;
if (i + (1 << k) > n) return k;
E t = G::op(s, dat[i + (1 << k) - 1]);
if (!check(t)) { return k; }
s = G::op(s, G::inverse(dat[i - 1])), i -= i & -i;
}
}();
while (k) {
--k;
if (i + (1 << k) - 1 < len(dat)) {
E t = G::op(s, dat[i + (1 << k) - 1]);
if (check(t)) { i += (1 << k), s = t; }
}
}
return i;
}
// check(i, x)
template <class F>
int max_right_with_index(const F check, int L = 0) {
assert(check(L, G::unit()));
E s = G::unit();
int i = L;
// 2^k 進むとダメ
int k = [&]() {
while (1) {
if (i % 2 == 1) { s = G::op(s, G::inverse(dat[i - 1])), i -= 1; }
if (i == 0) { return topbit(n) + 1; }
int k = lowbit(i) - 1;
if (i + (1 << k) > n) return k;
E t = G::op(s, dat[i + (1 << k) - 1]);
if (!check(i + (1 << k), t)) { return k; }
s = G::op(s, G::inverse(dat[i - 1])), i -= i & -i;
}
}();
while (k) {
--k;
if (i + (1 << k) - 1 < len(dat)) {
E t = G::op(s, dat[i + (1 << k) - 1]);
if (check(i + (1 << k), t)) { i += (1 << k), s = t; }
}
}
return i;
}
template <class F>
int min_left(const F check, int R) {
assert(check(G::unit()));
E s = G::unit();
int i = R;
// false になるところまで戻る
int k = 0;
while (i > 0 && check(s)) {
s = G::op(s, dat[i - 1]);
k = lowbit(i);
i -= i & -i;
}
if (check(s)) {
assert(i == 0);
return 0;
}
// 2^k 進むと ok になる
// false を維持して進む
while (k) {
--k;
E t = G::op(s, G::inverse(dat[i + (1 << k) - 1]));
if (!check(t)) { i += (1 << k), s = t; }
}
return i + 1;
}
int kth(E k, int L = 0) {
return max_right([&k](E x) -> bool { return x <= k; }, L);
}
};
#line 4 "ds/fenwicktree/fenwicktree_01.hpp"
struct FenwickTree_01 {
int N, n;
vc<u64> dat;
FenwickTree<Monoid_Add<int>> bit;
FenwickTree_01() {}
FenwickTree_01(int n) { build(n); }
template <typename F>
FenwickTree_01(int n, F f) {
build(n, f);
}
void build(int m) {
N = m;
n = ceil<int>(N + 1, 64);
dat.assign(n, u64(0));
bit.build(n);
}
template <typename F>
void build(int m, F f) {
N = m;
n = ceil<int>(N + 1, 64);
dat.assign(n, u64(0));
FOR(i, N) { dat[i / 64] |= u64(f(i)) << (i % 64); }
bit.build(n, [&](int i) -> int { return popcnt(dat[i]); });
}
int sum_all() { return bit.sum_all(); }
int sum(int k) { return prefix_sum(k); }
int prefix_sum(int k) {
int ans = bit.sum(k / 64);
ans += popcnt(dat[k / 64] & ((u64(1) << (k % 64)) - 1));
return ans;
}
int sum(int L, int R) {
if (L == 0) return prefix_sum(R);
int ans = 0;
ans -= popcnt(dat[L / 64] & ((u64(1) << (L % 64)) - 1));
ans += popcnt(dat[R / 64] & ((u64(1) << (R % 64)) - 1));
ans += bit.sum(L / 64, R / 64);
return ans;
}
void add(int k, int x) {
if (x == 1) add(k);
elif (x == -1) remove(k);
else assert(0);
}
void add(int k) {
dat[k / 64] |= u64(1) << (k % 64);
bit.add(k / 64, 1);
}
void remove(int k) {
dat[k / 64] &= ~(u64(1) << (k % 64));
bit.add(k / 64, -1);
}
int kth(int k, int L = 0) {
if (k >= sum_all()) return N;
k += popcnt(dat[L / 64] & ((u64(1) << (L % 64)) - 1));
L /= 64;
int mid = 0;
auto check = [&](auto e) -> bool {
if (e <= k) chmax(mid, e);
return e <= k;
};
int idx = bit.max_right(check, L);
if (idx == n) return N;
k -= mid;
u64 x = dat[idx];
int p = popcnt(x);
if (p <= k) return N;
k = binary_search([&](int n) -> bool { return (p - popcnt(x >> n)) <= k; }, 0, 64, 0);
return 64 * idx + k;
}
int next(int k) {
int idx = k / 64;
k %= 64;
u64 x = dat[idx] & ~((u64(1) << k) - 1);
if (x) return 64 * idx + lowbit(x);
idx = bit.kth(0, idx + 1);
if (idx == n || !dat[idx]) return N;
return 64 * idx + lowbit(dat[idx]);
}
int prev(int k) {
if (k == N) --k;
int idx = k / 64;
k %= 64;
u64 x = dat[idx];
if (k < 63) x &= (u64(1) << (k + 1)) - 1;
if (x) return 64 * idx + topbit(x);
idx = bit.min_left([&](auto e) -> bool { return e <= 0; }, idx) - 1;
if (idx == -1) return -1;
return 64 * idx + topbit(dat[idx]);
}
};
#line 3 "seq/inversion.hpp"
template <typename T>
ll inversion(vc<T> A) {
int N = len(A);
if (A.empty()) return 0;
ll ANS = 0;
FenwickTree_01 bit(N);
auto I = argsort(A);
for (auto& i: I) {
ANS += bit.sum_all() - bit.sum(i);
bit.add(i, 1);
}
return ANS;
}
// i 番目:A_i が先頭になるように rotate したときの転倒数
template <typename T, bool SMALL = false>
vi inversion_rotate(vc<T>& A) {
const int N = len(A);
if (!SMALL) {
auto key = A;
UNIQUE(key);
for (auto&& x: A) x = LB(key, x);
}
ll K = MAX(A) + 1;
ll ANS = 0;
FenwickTree<Monoid_Add<int>> bit(K);
for (auto&& x: A) {
ANS += bit.sum(x + 1, K);
bit.add(x, 1);
}
vi res(N);
FOR(i, N) {
res[i] = ANS;
ll x = A[i];
ANS = ANS + bit.sum(x + 1, K) - bit.prefix_sum(x);
}
return res;
}
// inv[i][j] = inversion A[i:j) であるような (N+1, N+1) テーブル
template <typename T>
vvc<int> all_range_inversion(vc<T>& A) {
int N = len(A);
vv(int, dp, N + 1, N + 1);
FOR_R(L, N + 1) FOR(R, L + 2, N + 1) {
dp[L][R] = dp[L][R - 1] + dp[L + 1][R] - dp[L + 1][R - 1];
if (A[L] > A[R - 1]) ++dp[L][R];
}
return dp;
}
template <typename T>
ll inversion_between(vc<T> A, vc<T> B) {
int N = len(A);
map<T, vc<int>> MP;
FOR(i, N) MP[B[i]].eb(i);
vc<int> TO(N);
FOR_R(i, N) {
auto& I = MP[A[i]];
if (I.empty()) return -1;
TO[i] = POP(I);
}
return inversion(TO);
}
#line 3 "other/sliding_puzzle_solver.hpp"
/*
O(HW(H+W))
空マスは -1 (unique)
同じ値が複数あってもよい
操作回数を K として、長さ K+1 の空マスの座標列をかえす
*/
struct Slinding_Puzzle_Solver {
using P = pair<int, int>;
vc<P> solve(vvc<int> A, vvc<int> B) {
int H = len(A), W = len(A[0]);
auto find = [&](vvc<int>& A, int k) -> P {
FOR(x, H) FOR(y, W) if (A[x][y] == k) return {x, y};
assert(0);
return {0, 0};
};
auto [ax, ay] = find(A, -1);
auto [bx, by] = find(B, -1);
vc<P> ANS_1, ANS_2;
while (ax > 0) { ANS_1.eb(ax, ay), swap(A[ax][ay], A[ax - 1][ay]), --ax; }
while (ay > 0) { ANS_1.eb(ax, ay), swap(A[ax][ay], A[ax][ay - 1]), --ay; }
while (bx > 0) { ANS_2.eb(bx, by), swap(B[bx][by], B[bx - 1][by]), --bx; }
while (by > 0) { ANS_2.eb(bx, by), swap(B[bx][by], B[bx][by - 1]), --by; }
vc<P> ANS = solve_00(A, B);
if (ANS.empty()) return {};
reverse(all(ANS_2));
return concat(ANS_1, ANS, ANS_2);
}
private:
vc<P> solve_00(vvc<int> A, vvc<int> B) {
assert(A[0][0] == -1 && B[0][0] == -1);
int H = len(A), W = len(A[0]);
if (H == 1 || W == 1) {
if (A != B) return {};
vc<P> ANS;
ANS.eb(0, 0);
return ANS;
}
vc<P> XYA, XYB;
FOR(x, H) FOR(y, W) XYA.eb(x, y), XYB.eb(x, y);
sort(all(XYA), [&](auto& a, auto& b) -> bool { return A[a.fi][a.se] < A[b.fi][b.se]; });
sort(all(XYB), [&](auto& a, auto& b) -> bool { return B[a.fi][a.se] < B[b.fi][b.se]; });
auto check = [&]() -> bool {
vc<int> S, T;
FOR(i, H * W) {
auto [x1, y1] = XYA[i];
auto [x2, y2] = XYB[i];
if (A[x1][y1] != B[x2][y2]) return 0;
S.eb(W * x1 + y1);
T.eb(W * x2 + y2);
}
ll x = inversion_between(S, T);
return x % 2 == 0;
};
if (!check()) {
FOR(i, H * W - 1) {
auto [x1, y1] = XYA[i];
auto [x2, y2] = XYA[i + 1];
if (A[x1][y1] != A[x2][y2]) continue;
swap(XYA[i], XYA[i + 1]);
break;
}
if (!check()) return {};
}
vv(P, X, H, W);
FOR(i, H * W) {
auto [x1, y1] = XYA[i];
auto [x2, y2] = XYB[i];
X[x1][y1] = {x2, y2};
}
vc<P> ANS;
ANS.eb(0, 0);
solve_sort(X, ANS, false);
return ANS;
}
// 移動先の座標の列を並べたグリッドを与える.
// (0,0) が空マス
void solve_sort(vvc<pair<int, int>>& A, vc<P>& ANS, bool tr) {
int H = len(A), W = len(A[0]);
vv(P, pos, H, W);
FOR(x, H) FOR(y, W) {
P p = A[x][y];
pos[p.fi][p.se] = {x, y};
}
auto [px, py] = pos[0][0];
auto ope = [&](int x, int y) -> void {
assert(abs(px - x) + abs(py - y) == 1);
swap(A[px][py], A[x][y]);
if (!tr) ANS.eb(x, y);
if (tr) ANS.eb(y, x);
pos[A[px][py].fi][A[px][py].se] = {px, py};
px = x, py = y;
pos[A[px][py].fi][A[px][py].se] = {px, py};
};
if (H == 2 && W == 2) {
auto check = [&]() -> bool {
FOR(x, H) FOR(y, W) if (A[x][y].fi != x || A[x][y].se != y) return 0;
return 1;
};
while (!check()) {
if (px == 0 && py == 0) ope(1, 0);
if (px == 1 && py == 0) ope(1, 1);
if (px == 1 && py == 1) ope(0, 1);
if (px == 0 && py == 1) ope(0, 0);
}
return;
}
if (H < W) {
FOR(x, H) FOR(y, W) { swap(A[x][y].fi, A[x][y].se); }
A = transpose(A, H, W);
solve_sort(A, ANS, !tr);
return;
}
assert(H >= 3 && W >= 2);
// 最後の行をそろえる
FOR_R(y, 1, W) {
auto& [tx, ty] = pos[H - 1][y];
if (px == H - 1) ope(px - 1, py);
while (px - 1 > tx) ope(px - 1, py), tie(tx, ty) = pos[H - 1][y];
while (px + 1 < tx) ope(px + 1, py), tie(tx, ty) = pos[H - 1][y];
if (px == tx) {
if (px == 0)
ope(px + 1, py);
else
ope(px - 1, py);
}
assert(abs(px - tx) == 1);
while (py < ty) ope(px, py + 1);
while (py > ty) ope(px, py - 1);
if (px == tx + 1) ope(px - 1, py), tie(tx, ty) = pos[H - 1][y];
assert(px == tx - 1 && py == ty);
while (ty < y) { ope(px, py + 1), ope(px + 1, py), ope(px, py - 1), ope(px - 1, py), ope(px, py + 1); }
while (ty > y) { ope(px, py - 1), ope(px + 1, py), ope(px, py + 1), ope(px - 1, py), ope(px, py - 1); }
tie(tx, ty) = pos[H - 1][y];
while (tx < H - 1) {
ope(px, py - 1), ope(px + 1, py), ope(px + 1, py), ope(px, py + 1), ope(px - 1, py);
tie(tx, ty) = pos[H - 1][y];
}
assert(A[H - 1][y] == (pair<int, int>{H - 1, y}));
}
auto& [tx, ty] = pos[H - 1][0];
if (px == H - 1) ope(px - 1, py);
while (px - 1 > tx) ope(px - 1, py), tie(tx, ty) = pos[H - 1][0];
while (px + 1 < tx) ope(px + 1, py), tie(tx, ty) = pos[H - 1][0];
if (px == tx) {
if (px == 0)
ope(px + 1, py);
else
ope(px - 1, py);
}
tie(tx, ty) = pos[H - 1][0];
assert(abs(px - tx) == 1);
while (py < ty) ope(px, py + 1);
while (py > ty) ope(px, py - 1);
if (px == tx + 1) ope(px - 1, py);
tie(tx, ty) = pos[H - 1][0];
if (tx < H - 1) {
assert(px == tx - 1 && py == ty);
while (ty > 0) { ope(px, py - 1), ope(px + 1, py), ope(px, py + 1), ope(px - 1, py), ope(px, py - 1); }
tie(tx, ty) = pos[H - 1][0];
while (tx < H - 2) { ope(px, py + 1), ope(px + 1, py), ope(px + 1, py), ope(px, py - 1), ope(px - 1, py); }
ope(px + 1, py), ope(px + 1, py), ope(px, py + 1);
ope(px - 1, py), ope(px, py - 1), ope(px - 1, py);
ope(px, py + 1), ope(px + 1, py), ope(px + 1, py);
ope(px, py - 1), ope(px - 1, py);
}
FOR(y, W) assert(A[H - 1][y] == (pair<int, int>{H - 1, y}));
POP(A);
solve_sort(A, ANS, tr);
}
};