#826522 (C++17) No.5010 Better Mo's Algorithm is Needed!! (Unweighted)

提出ソース

結果

問題	No.5010 Better Mo's Algorithm is Needed!! (Unweighted)
ユーザー	komori3
提出日時	2022-12-17 17:27:25
言語	C++17 (gcc 12.3.0 + boost 1.83.0)
結果	AC
実行時間	4,922 ms / 5,000 ms
コード長	12,791 bytes
コンパイル時間	5,291 ms
実行使用メモリ	6,952 KB
スコア	19,246,189,272
最終ジャッジ日時	2022-12-17 17:38:09
合計ジャッジ時間	644,153 ms
ジャッジサーバーID （参考情報）	judge12 / judge13
	純コード判定しない問題か言語

このコードへのチャレンジ
（要ログイン）

テストケース

テストケース表示

入力	結果	実行時間実行使用メモリ
testcase_00	AC	4,918 ms 6,672 KB
testcase_01	AC	4,917 ms 6,588 KB
testcase_02	AC	4,918 ms 6,580 KB
testcase_03	AC	4,917 ms 6,592 KB
testcase_04	AC	4,918 ms 6,584 KB
testcase_05	AC	4,916 ms 6,528 KB
testcase_06	AC	4,917 ms 6,584 KB
testcase_07	AC	4,917 ms 6,584 KB
testcase_08	AC	4,917 ms 6,524 KB
testcase_09	AC	4,916 ms 6,528 KB
testcase_10	AC	4,917 ms 6,512 KB
testcase_11	AC	4,918 ms 6,524 KB
testcase_12	AC	4,917 ms 6,528 KB
testcase_13	AC	4,917 ms 6,676 KB
testcase_14	AC	4,917 ms 6,596 KB
testcase_15	AC	4,916 ms 6,588 KB
testcase_16	AC	4,917 ms 6,528 KB
testcase_17	AC	4,918 ms 6,528 KB
testcase_18	AC	4,917 ms 6,948 KB
testcase_19	AC	4,917 ms 6,628 KB
testcase_20	AC	4,916 ms 6,528 KB
testcase_21	AC	4,917 ms 6,596 KB
testcase_22	AC	4,918 ms 6,524 KB
testcase_23	AC	4,919 ms 6,672 KB
testcase_24	AC	4,917 ms 6,724 KB
testcase_25	AC	4,918 ms 6,824 KB
testcase_26	AC	4,916 ms 6,844 KB
testcase_27	AC	4,922 ms 6,588 KB
testcase_28	AC	4,917 ms 6,840 KB
testcase_29	AC	4,918 ms 6,680 KB
testcase_30	AC	4,918 ms 6,952 KB
testcase_31	AC	4,916 ms 6,840 KB
testcase_32	AC	4,917 ms 6,948 KB
testcase_33	AC	4,917 ms 6,676 KB
testcase_34	AC	4,918 ms 6,672 KB
testcase_35	AC	4,917 ms 6,524 KB
testcase_36	AC	4,916 ms 6,680 KB
testcase_37	AC	4,917 ms 6,524 KB
testcase_38	AC	4,917 ms 6,948 KB
testcase_39	AC	4,917 ms 6,516 KB
testcase_40	AC	4,918 ms 6,560 KB
testcase_41	AC	4,916 ms 6,528 KB
testcase_42	AC	4,916 ms 6,948 KB
testcase_43	AC	4,917 ms 6,588 KB
testcase_44	AC	4,917 ms 6,832 KB
testcase_45	AC	4,917 ms 6,724 KB
testcase_46	AC	4,917 ms 6,952 KB
testcase_47	AC	4,917 ms 6,676 KB
testcase_48	AC	4,917 ms 6,948 KB
testcase_49	AC	4,917 ms 6,520 KB
testcase_50	AC	4,918 ms 6,620 KB
testcase_51	AC	4,918 ms 6,524 KB
testcase_52	AC	4,917 ms 6,592 KB
testcase_53	AC	4,916 ms 6,528 KB
testcase_54	AC	4,917 ms 6,948 KB
testcase_55	AC	4,916 ms 6,592 KB
testcase_56	AC	4,918 ms 6,676 KB
testcase_57	AC	4,917 ms 6,592 KB
testcase_58	AC	4,918 ms 6,832 KB
testcase_59	AC	4,917 ms 6,952 KB
testcase_60	AC	4,918 ms 6,680 KB
testcase_61	AC	4,916 ms 6,620 KB
testcase_62	AC	4,917 ms 6,680 KB
testcase_63	AC	4,917 ms 6,948 KB
testcase_64	AC	4,918 ms 6,676 KB
testcase_65	AC	4,918 ms 6,680 KB
testcase_66	AC	4,917 ms 6,528 KB
testcase_67	AC	4,918 ms 6,524 KB
testcase_68	AC	4,916 ms 6,948 KB
testcase_69	AC	4,919 ms 6,844 KB
testcase_70	AC	4,917 ms 6,948 KB
testcase_71	AC	4,917 ms 6,952 KB
testcase_72	AC	4,917 ms 6,508 KB
testcase_73	AC	4,918 ms 6,596 KB
testcase_74	AC	4,916 ms 6,668 KB
testcase_75	AC	4,917 ms 6,528 KB
testcase_76	AC	4,917 ms 6,528 KB
testcase_77	AC	4,917 ms 6,724 KB
testcase_78	AC	4,916 ms 6,592 KB
testcase_79	AC	4,916 ms 6,948 KB
testcase_80	AC	4,917 ms 6,948 KB
testcase_81	AC	4,916 ms 6,672 KB
testcase_82	AC	4,917 ms 6,952 KB
testcase_83	AC	4,917 ms 6,952 KB
testcase_84	AC	4,917 ms 6,948 KB
testcase_85	AC	4,918 ms 6,948 KB
testcase_86	AC	4,918 ms 6,724 KB
testcase_87	AC	4,917 ms 6,952 KB
testcase_88	AC	4,916 ms 6,948 KB
testcase_89	AC	4,918 ms 6,532 KB
testcase_90	AC	4,918 ms 6,592 KB
testcase_91	AC	4,917 ms 6,948 KB
testcase_92	AC	4,917 ms 6,948 KB
testcase_93	AC	4,917 ms 6,948 KB
testcase_94	AC	4,919 ms 6,948 KB
testcase_95	AC	4,917 ms 6,948 KB
testcase_96	AC	4,916 ms 6,948 KB
testcase_97	AC	4,917 ms 6,528 KB
testcase_98	AC	4,918 ms 6,676 KB
testcase_99	AC	4,918 ms 6,952 KB
testcase_100	AC	4,917 ms 6,832 KB
testcase_101	AC	4,917 ms 6,596 KB
testcase_102	AC	4,917 ms 6,724 KB
testcase_103	AC	4,917 ms 6,532 KB
testcase_104	AC	4,917 ms 6,508 KB
testcase_105	AC	4,918 ms 6,840 KB
testcase_106	AC	4,917 ms 6,668 KB
testcase_107	AC	4,916 ms 6,528 KB
testcase_108	AC	4,918 ms 6,668 KB
testcase_109	AC	4,917 ms 6,680 KB
testcase_110	AC	4,918 ms 6,596 KB
testcase_111	AC	4,917 ms 6,840 KB
testcase_112	AC	4,917 ms 6,628 KB
testcase_113	AC	4,917 ms 6,528 KB
testcase_114	AC	4,916 ms 6,828 KB
testcase_115	AC	4,918 ms 6,524 KB
testcase_116	AC	4,917 ms 6,596 KB
testcase_117	AC	4,918 ms 6,624 KB
testcase_118	AC	4,917 ms 6,624 KB
testcase_119	AC	4,917 ms 6,952 KB

権限があれば一括ダウンロードができます

ソースコード

raw source code

#define _CRT_NONSTDC_NO_WARNINGS
#define _SILENCE_CXX17_ITERATOR_BASE_CLASS_DEPRECATION_WARNING
#include <bits/stdc++.h>
#include <random>
#include <unordered_set>
#include <array>
#include <optional>
#ifdef _MSC_VER
#include <opencv2/core.hpp>
#include <opencv2/core/utils/logger.hpp>
#include <opencv2/imgproc.hpp>
#include <opencv2/highgui.hpp>
#include <conio.h>
#include <ppl.h>
#include <filesystem>
#include <intrin.h>
/* g++ functions */
int __builtin_clz(unsigned int n) { unsigned long index; _BitScanReverse(&index, n); return 31 - index; }
int __builtin_ctz(unsigned int n) { unsigned long index; _BitScanForward(&index, n); return index; }
namespace std { inline int __lg(int __n) { return sizeof(int) * 8 - 1 - __builtin_clz(__n); } }
/* enable __uint128_t in MSVC */
//#include <boost/multiprecision/cpp_int.hpp>
//using __uint128_t = boost::multiprecision::uint128_t;
#else
#pragma GCC target("avx2")
#pragma GCC optimize("O3")
#pragma GCC optimize("unroll-loops")
#endif

/** compro io **/
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) << ", "; 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); } };
}
template<typename... Ts> std::ostream& operator<<(std::ostream& os, const std::tuple<Ts...>& t) { os << '['; aux::tp<std::tuple<Ts...>, 0, sizeof...(Ts) - 1>::output(os, t); return os << ']'; } // tuple out
template<class Ch, class Tr, class Container> std::basic_ostream<Ch, Tr>& operator<<(std::basic_ostream<Ch, Tr>& os, const Container& x); // container out (fwd decl)
template<class S, class T> std::ostream& operator<<(std::ostream& os, const std::pair<S, T>& p) { return os << "[" << p.first << ", " << p.second << "]"; } // pair out
template<class S, class T> std::istream& operator>>(std::istream& is, std::pair<S, T>& p) { return is >> p.first >> p.second; } // pair in
std::ostream& operator<<(std::ostream& os, const std::vector<bool>::reference& v) { os << (v ? '1' : '0'); return os; } // bool (vector) out
std::ostream& operator<<(std::ostream& os, const std::vector<bool>& v) { bool f = true; os << "["; for (const auto& x : v) { os << (f ? "" : ", ") << x; f = false; } os << "]"; return os; } // vector<bool> out
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; os << "["; for (auto& y : x) { os << (f ? "" : ", ") << y; f = false; } return os << "]"; } // container out
template<class T, class = decltype(std::begin(std::declval<T&>())), class = typename std::enable_if<!std::is_same<T, std::string>::value>::type> std::istream& operator>>(std::istream& is, T& a) { for (auto& x : a) is >> x; return is; } // container in
template<typename T> auto operator<<(std::ostream& out, const T& t) -> decltype(out << t.stringify()) { out << t.stringify(); return out; } // struct (has stringify() func) out
/** io setup **/
struct IOSetup { IOSetup(bool f) { if (f) { std::cin.tie(nullptr); std::ios::sync_with_stdio(false); } std::cout << std::fixed << std::setprecision(15); } }
iosetup(true); // set false when solving interective problems
/** string formatter **/
template<typename... Ts> std::string format(const std::string& f, Ts... t) { size_t l = std::snprintf(nullptr, 0, f.c_str(), t...); std::vector<char> b(l + 1); std::snprintf(&b[0], l + 1, f.c_str(), t...); return std::string(&b[0], &b[0] + l); }
/** dump **/
#define DUMPOUT std::cerr
std::ostringstream DUMPBUF;
#define dump(...) do{DUMPBUF<<"  ";DUMPBUF<<#__VA_ARGS__<<" :[DUMP - "<<__LINE__<<":"<<__FUNCTION__<<"]"<<std::endl;DUMPBUF<<"    ";dump_func(__VA_ARGS__);DUMPOUT<<DUMPBUF.str();DUMPBUF.str("");DUMPBUF.clear();}while(0);
void dump_func() { DUMPBUF << std::endl; }
template <class Head, class... Tail> void dump_func(Head&& head, Tail&&... tail) { DUMPBUF << head; if (sizeof...(Tail) == 0) { DUMPBUF << " "; } else { DUMPBUF << ", "; } dump_func(std::move(tail)...); }
/** timer **/
class Timer {
    double t = 0, paused = 0, tmp;
public:
    Timer() { reset(); }
    static double time() {
#ifdef _MSC_VER
        return __rdtsc() / 2.3e9;
#else
        unsigned long long a, d;
        __asm__ volatile("rdtsc"
            : "=a"(a), "=d"(d));
        return (d << 32 | a) / 2.3e9;
#endif
    }
    void reset() { t = time(); }
    void pause() { tmp = time(); }
    void restart() { paused += time() - tmp; }
    double elapsed_ms() const { return (time() - t - paused) * 1000.0; }
};
/** rand **/
struct Xorshift {
    static constexpr uint64_t M = INT_MAX;
    static constexpr double e = 1.0 / M;
    uint64_t x = 88172645463325252LL;
    Xorshift() {}
    Xorshift(uint64_t seed) { reseed(seed); }
    inline void reseed(uint64_t seed) { x = 0x498b3bc5 ^ seed; for (int i = 0; i < 20; i++) next(); }
    inline uint64_t next() { x = x ^ (x << 7); return x = x ^ (x >> 9); }
    inline int next_int() { return next() & M; }
    inline int next_int(int mod) { return next() % mod; }
    inline int next_int(int l, int r) { return l + next_int(r - l + 1); }
    inline double next_double() { return next_int() * e; }
};
/** shuffle **/
template<typename T> void shuffle_vector(std::vector<T>& v, Xorshift& rnd) { int n = v.size(); for (int i = n - 1; i >= 1; i--) { int r = rnd.next_int(i); std::swap(v[i], v[r]); } }
/** split **/
std::vector<std::string> split(std::string str, const std::string& delim) { for (char& c : str) if (delim.find(c) != std::string::npos) c = ' '; std::istringstream iss(str); std::vector<std::string> parsed; std::string buf; while (iss >> buf) parsed.push_back(buf); return parsed; }
/** misc **/
template<typename A, size_t N, typename T> inline void Fill(A(&array)[N], const T& val) { std::fill((T*)array, (T*)(array + N), val); } // fill array
template<typename T, typename ...Args> auto make_vector(T x, int arg, Args ...args) { if constexpr (sizeof...(args) == 0)return std::vector<T>(arg, x); else return std::vector(arg, make_vector<T>(x, args...)); }
template<typename T> bool chmax(T& a, const T& b) { if (a < b) { a = b; return true; } return false; }
template<typename T> bool chmin(T& a, const T& b) { if (a > b) { a = b; return true; } return false; }
/** using **/
using ll = long long;
using ld = long double;
//using ld = boost::multiprecision::cpp_bin_float_quad;
using pii = std::pair<int, int>;
using pll = std::pair<ll, ll>;
using std::cin, std::cout, std::cerr, std::endl, std::string, std::vector, std::array;



constexpr int N = 200000;
constexpr int Q = 200000;

double ltbl[0x10000];

// hyper params
bool optuna_mode = false;
int weight_type = -1;
int bucket_size = 1718;
double start_temp = 156.71493463839815;
double end_temp = 0;
int range_max = 35;

int bucket_size_opt[3] = { 1718, 1299, 1638 };
double start_temp_opt[3] = { 156.71493463839815, 107.94021671154113, 130.36303804130475 };
int range_max_opt[3] = { 35, 30, 32 };


void initialize(int argc, char** argv) {

    Xorshift rnd;
    for(int i = 0; i < 1000; i++) rnd.next_int();
    for(int i = 0; i < 0x10000; i++) ltbl[i] = log(rnd.next_double());

    if (argc > 1 && string(argv[1]) == "optuna") {
        optuna_mode = true;
        weight_type = std::stoi(argv[2]); // 1,2,3
        bucket_size = std::stoi(argv[3]);
        start_temp = std::stod(argv[4]);
        range_max = std::stoi(argv[5]);
    }

}

struct Result {
    int64_t score;
    int64_t loop;
};

struct State {

    Xorshift rnd;
    
    int ST;
    int L[Q], R[Q];

    int P[Q];

    int64_t cost;

    State(std::istream& in) {

        in >> ST >> ST >> ST >> ST;

        if (!optuna_mode) {
            bucket_size = bucket_size_opt[ST - 1];
            start_temp = start_temp_opt[ST - 1];
            range_max = range_max_opt[ST - 1];
        }

        for (int i = 0; i < N; i++) in >> L[0]; // ignore weights
        for (int i = 0; i < Q; i++) in >> L[i] >> R[i];

        std::iota(P, P + Q, 0);
        std::sort(P, P + Q, [&](int a, int b) {
            int ablock = L[a] / bucket_size, bblock = L[b] / bucket_size;
            if (ablock != bblock) return ablock < bblock;
            return (ablock & 1) ? R[a] > R[b] : R[a] < R[b];
            });

        cost = R[P[0]] - L[P[0]] + 1;
        for (int k = 0; k < Q - 1; k++) cost += calc_adj_cost(k);

    }

    Result run(double duration) {
        
        Timer timer;
        int64_t loop = 0;

        auto get_temp = [](double start_temp, double end_temp, double now_time, double end_time) {
            return end_temp + (start_temp - end_temp) * (end_time - now_time) / end_time;
        };

        double now_time = timer.elapsed_ms();
        double temp = get_temp(start_temp, end_temp, now_time, duration);
        while(true) {
            loop++;
            int l = loop % (Q - 1), r = std::min(Q - 1, l + rnd.next_int(1, range_max));

            int diff = calc_rev_cost(l, r);
            if (-diff > temp * ltbl[loop & 0xFFFF]) {
                std::reverse(P + l, P + r + 1);
                cost += diff;
            }

            //if (!(loop & 0xFFFFFF)) {
            //    dump(loop, temp, cost);
            //}

            if (!(loop & 0xFFFF)) {
                now_time = timer.elapsed_ms();
                if (now_time > duration) break;
                temp = get_temp(start_temp, end_temp, now_time, duration);
            }
        }

        return { eval(), loop };
    }

    void run(int num_loop) {

        auto get_temp = [](double start_temp, double end_temp, double now_time, double end_time) {
            return end_temp + (start_temp - end_temp) * (end_time - now_time) / end_time;
        };

        double temp = get_temp(start_temp, end_temp, 0, num_loop);
        for (int loop = 0; loop < num_loop; loop++) {
            int l = loop % (Q - 1), r = std::min(Q - 1, l + rnd.next_int(1, range_max));

            int diff = calc_rev_cost(l, r);
            if (-diff > temp * ltbl[loop & 0xFFFF]) {
                std::reverse(P + l, P + r + 1);
                cost += diff;
            }
            if (!(loop & 0xFFFF)) temp = get_temp(start_temp, end_temp, loop, num_loop);
        }
    }

    inline int calc_adj_cost(int k) const {
        return abs(L[P[k]] - L[P[k + 1]]) + abs(R[P[k]] - R[P[k + 1]]);
    }

    int calc_rev_cost(int l, int r) {
        int diff = 0;
        diff -= l ? calc_adj_cost(l - 1) : (R[P[0]] - L[P[0]] + 1);
        diff -= (r + 1 < Q) ? calc_adj_cost(r) : 0;
        std::swap(P[l], P[r]);
        diff += l ? calc_adj_cost(l - 1) : (R[P[0]] - L[P[0]] + 1);
        diff += (r + 1 < Q) ? calc_adj_cost(r) : 0;
        std::swap(P[l], P[r]);
        return diff;
    }

    int64_t eval() const {
        return std::min((ll)std::round(1e16 / cost), (ll)1e16);
    }

    void output(std::ostream& out) const {
        out << P[0] + 1;
        for (int i = 1; i < Q; i++) out << ' ' << P[i] + 1;
        out << '\n';
    }

};

int main([[maybe_unused]] int argc, [[maybe_unused]] char** argv) {

    Timer timer;

#ifdef HAVE_OPENCV_HIGHGUI
    cv::utils::logging::setLogLevel(cv::utils::logging::LogLevel::LOG_LEVEL_SILENT);
#endif

#ifdef _MSC_VER
    std::ifstream ifs("../tools/input1/0002.txt");
    std::istream& in = ifs;
    std::ofstream ofs("../tools/output1/0002.txt");
    std::ostream& out = ofs;
#else
    std::istream& in = cin;
    std::ostream& out = cout;
#endif

    initialize(argc, argv);

    if (optuna_mode) {
#ifdef _MSC_VER
        vector<string> paths;
        for (int s = weight_type; s <= 30; s += 3) {
            paths.push_back(format("G:\\dev\\heuristic\\yukicoder5010\\tools\\input1\\%04d.txt", s));
        }
        vector<ll> scores(paths.size());
        int batch_size = 5;
        for (int begin = 0; begin < paths.size(); begin += batch_size) {
            int end = std::min((int)paths.size(), begin + batch_size);
            concurrency::critical_section mtx;
            concurrency::parallel_for(begin, end, [&](int i) {
                std::ifstream ifs(paths[i]);
                State state(ifs);
                auto [score, loop] = state.run(4900.0);
                scores[i] = score;
                {
                    mtx.lock();
                    cerr << format("case %2d: score=%lld, loop=%lld\n", i, scores[i], loop);
                    mtx.unlock();
                }
            });
        }
        cout << (double)std::accumulate(scores.begin(), scores.end(), 0LL) / scores.size() << '\n';
#endif
    }
    else {
        State state(in);
        dump(state.cost, state.eval());
        state.run(4900 - timer.elapsed_ms());
        dump(state.cost, state.eval());
        state.output(out);
    }

    return 0;
}