面临的挑战是编写最快的代码来计算矩阵的永久性。

n-by- nMatrix A=（ai,j）的永久变量定义为

在此S_n表示的所有排列的集合[1, n]。

作为一个例子（来自维基）：

在这个问题中，矩阵都是正方形的，只有值-1和1在其中。

例子

输入：

[[ 1 -1 -1  1]
 [-1 -1 -1  1]
 [-1  1 -1  1]
 [ 1 -1 -1  1]]

常驻：

-4

输入：

[[-1 -1 -1 -1]
 [-1  1 -1 -1]
 [ 1 -1 -1 -1]
 [ 1 -1  1 -1]]

常驻：

0

输入：

[[ 1 -1  1 -1 -1 -1 -1 -1]
 [-1 -1  1  1 -1  1  1 -1]
 [ 1 -1 -1 -1 -1  1  1  1]
 [-1 -1 -1  1 -1  1  1  1]
 [ 1 -1 -1  1  1  1  1 -1]
 [-1  1 -1  1 -1  1  1 -1]
 [ 1 -1  1 -1  1 -1  1 -1]
 [-1 -1  1 -1  1  1  1  1]]

常驻：

192

输入：

[[1, -1, 1, -1, -1, 1, 1, 1, -1, -1, -1, -1, 1, 1, 1, 1, -1, 1, 1, -1],
 [1, -1, 1, 1, 1, 1, 1, -1, 1, -1, -1, 1, 1, 1, -1, -1, 1, 1, 1, -1],
 [-1, -1, 1, 1, 1, -1, -1, -1, -1, 1, -1, 1, 1, 1, -1, -1, -1, 1, -1, -1],
 [-1, -1, -1, 1, 1, -1, 1, 1, 1, 1, 1, 1, -1, -1, -1, -1, -1, -1, 1, -1],
 [-1, 1, 1, 1, -1, 1, 1, 1, -1, -1, -1, 1, -1, 1, -1, 1, 1, 1, 1, 1],
 [1, -1, 1, 1, -1, -1, 1, -1, 1, 1, 1, 1, -1, 1, 1, -1, 1, -1, -1, -1],
 [1, -1, -1, 1, -1, -1, -1, 1, -1, 1, 1, 1, 1, -1, -1, -1, 1, 1, 1, -1],
 [1, -1, -1, 1, -1, 1, 1, -1, 1, 1, 1, -1, 1, -1, 1, 1, 1, -1, 1, 1],
 [1, -1, -1, -1, -1, -1, 1, 1, 1, -1, -1, -1, -1, -1, 1, 1, -1, 1, 1, -1],
 [-1, -1, 1, -1, 1, -1, 1, 1, -1, 1, -1, 1, 1, 1, 1, 1, 1, -1, 1, 1],
 [-1, -1, -1, -1, -1, -1, -1, 1, -1, -1, -1, -1, 1, 1, 1, 1, -1, -1, -1, -1],
 [1, 1, -1, -1, -1, 1, 1, -1, -1, 1, -1, 1, 1, -1, 1, 1, 1, 1, 1, 1],
 [-1, 1, 1, -1, -1, -1, -1, -1, 1, 1, 1, 1, -1, -1, -1, -1, -1, 1, -1, 1],
 [1, 1, -1, -1, -1, 1, -1, 1, -1, -1, -1, -1, 1, -1, 1, 1, -1, 1, -1, 1],
 [1, 1, 1, 1, 1, -1, -1, -1, 1, 1, 1, -1, 1, -1, 1, 1, 1, -1, 1, 1],
 [1, -1, -1, 1, -1, -1, -1, -1, 1, -1, -1, 1, 1, -1, 1, -1, -1, -1, -1, -1],
 [-1, 1, 1, 1, -1, 1, 1, -1, -1, 1, 1, 1, -1, -1, 1, 1, -1, -1, 1, 1],
 [1, 1, -1, -1, 1, 1, -1, 1, 1, -1, 1, 1, 1, -1, 1, 1, -1, 1, -1, 1],
 [1, 1, 1, -1, -1, -1, 1, -1, -1, 1, 1, -1, -1, -1, 1, -1, -1, -1, -1, 1],
 [-1, 1, 1, 1, -1, -1, -1, -1, -1, -1, -1, 1, 1, -1, 1, 1, -1, 1, -1, -1]]

常驻：

1021509632

任务

您应该编写代码，给定一个nby n矩阵，并输出其永久性代码。

由于我将需要测试您的代码，因此，如果您能为我提供一种简单的方法将矩阵作为代码的输入，这将很有帮助，例如通过读取标准输入。

请注意，永久性可能很大（全1矩阵是极端情况）。

得分和联系

我将在大小增加的随机+ -1矩阵上测试您的代码，并在您的代码第一次在我的计算机上花费超过1分钟时停止。所有提交的评分矩阵都将保持一致，以确保公平。

如果两个人获得相同的分数，那么胜者就是那个值最快的人n。如果它们彼此相距1秒以内，则它是第一个发布的。

语言和图书馆

您可以使用任何喜欢的可用语言和库，但不能使用预先存在的函数来计算永久性对象。在可行的情况下，能够运行您的代码将是一个很好的选择，因此，请尽可能提供有关如何在Linux中运行/编译代码的完整说明。

参考实施

已经存在一个带有许多不同语言代码的代码高尔夫问题，用于计算小型矩阵的永久性。如果您可以访问Mathematica和Maple，它们也都具有永久实现。

我的机器计时将在我的64位机器上运行。这是带有8GB RAM，AMD FX-8350八核处理器和Radeon HD 4250的标准ubuntu安装。这也意味着我需要能够运行您的代码。

有关我的机器的低级信息

cat /proc/cpuinfo/|grep flags 给

国旗：FPU VME德PSE TSC MSR，PAE MCE CX8 APIC月MTRR PGE MCA CMOV拍拍PSE36 CLFLUSH MMX FXSR SSE SSE2 HT系统调用NX mmxext fxsr_opt pdpe1gb rdtscp流明constant_tsc rep_good nopl nonstop_tsc extd_apicid aperfmperf PNI pclmulqdq监控SSSE3 FMA CX16 sse4_1 sse4_2 POPCNT AES XSAVE AVX F16C lahf_lm cmp_legacy SVM extapic cr8_legacy ABM SSE4A misalignsse 3dnowprefetch osvw IBS XOP SKINIT WDT LWP FMA4 TCE nodeid_msr TBM topoext perfctr_core perfctr_nb CPB hw_pstate vmmcall BMI1 ARAT NPT lbrv svm_lock nrip_save tsc_scale vmcb_clean flushbyasid decodeassists pausefilter pfthreshold

我会问一个紧密相关的后续多语言问题，该问题不会遇到较大的Int问题，因此Scala，Nim，Julia，Rust，Bash的爱好者也可以炫耀自己的语言。

排行榜

• n = 33（45秒。n= 34为64秒）。吨Hospel在C ++使用g ++ 5.4.0。
• n = 32（32秒）。丹尼斯在C中使用Ton Hospel的gcc标志使用gcc 5.4.0。
• n = 31（54秒）。基督教西弗斯在哈斯克尔
• n = 31（60秒）。普里莫在rpython
• n = 30（26秒）。ezrast在拉斯特
• n = 28（49秒）。使用Python + pypy 5.4.1的xnor
• n = 22（25秒）。Shebang使用Python + pypy 5.4.1

注意。在实践中，出于神秘的原因，丹尼斯（Dennis）和唐·霍斯珀尔（Ton Hospel）的演出时间相差很大。例如，在我加载Web浏览器后，它们似乎更快！引用的时间是我完成的所有测试中最快的。

gcc C ++ n≈36（在我的系统上为57秒）

如果所有列的和都是偶数，则使用带有灰色代码的Glynn公式进行更新，否则使用Ryser的方法。线程化和向量化。针对AVX进行了优化，因此对于较旧的处理器不要期望太大。n>=35即使您的系统足够快，也不要担心只有+1的矩阵，因为有符号的128位累加器会溢出。对于随机矩阵，您可能不会遇到溢出问题。对于n>=37内部乘法器，所有1/-1矩阵将开始溢出。因此，仅将此程序用于n<=36。

只需给STDIN上的矩阵元素加上任何空白即可

permanent
1 2
3 4
^D

permanent.cpp：

/*
  Compile using something like:
    g++ -Wall -O3 -march=native -fstrict-aliasing -std=c++11 -pthread -s permanent.cpp -o permanent
*/

#include <iostream>
#include <iomanip>
#include <cstdlib>
#include <cstdint>
#include <climits>
#include <array>
#include <vector>
#include <thread>
#include <future>
#include <ctgmath>
#include <immintrin.h>

using namespace std;

bool const DEBUG = false;
int const CACHE = 64;

using Index  = int_fast32_t;
Index glynn;
// Number of elements in our vectors
Index const POW   = 3;
Index const ELEMS = 1 << POW;
// Over how many floats we distribute each row
Index const WIDTH = 9;
// Number of bits in the fraction part of a floating point number
int const FLOAT_MANTISSA = 23;
// Type to use for the first add/multiply phase
using Sum  = float;
using SumN = __restrict__ Sum __attribute__((vector_size(ELEMS*sizeof(Sum))));
// Type to convert to between the first and second phase
using ProdN = __restrict__ int32_t __attribute__((vector_size(ELEMS*sizeof(int32_t))));
// Type to use for the third and last multiply phase.
// Also used for the final accumulator
using Value = __int128;
using UValue = unsigned __int128;

// Wrap Value so C++ doesn't really see it and we can put it in vectors etc.
// Needed since C++ doesn't fully support __int128
struct Number {
    Number& operator+=(Number const& right) {
        value += right.value;
        return *this;
    }
    // Output the value
    void print(ostream& os, bool dbl = false) const;
    friend ostream& operator<<(ostream& os, Number const& number) {
        number.print(os);
        return os;
    }

    Value value;
};

using ms = chrono::milliseconds;

auto nr_threads = thread::hardware_concurrency();
vector<Sum> input;

// Allocate cache aligned datastructures
template<typename T>
T* alloc(size_t n) {
    T* mem = static_cast<T*>(aligned_alloc(CACHE, sizeof(T) * n));
    if (mem == nullptr) throw(bad_alloc());
    return mem;
}

// Work assigned to thread k of nr_threads threads
Number permanent_part(Index n, Index k, SumN** more) {
    uint64_t loops = (UINT64_C(1) << n) / nr_threads;
    if (glynn) loops /= 2;
    Index l = loops < ELEMS ? loops : ELEMS;
    loops /= l;
    auto from = loops * k;
    auto to   = loops * (k+1);

    if (DEBUG) cout << "From=" << from << "\n";
    uint64_t old_gray = from ^ from/2;
    uint64_t bit = 1;
    bool bits = (to-from) & 1;

    Index nn = (n+WIDTH-1)/WIDTH;
    Index ww = nn * WIDTH;
    auto column = alloc<SumN>(ww);
    for (Index i=0; i<n; ++i)
        for (Index j=0; j<ELEMS; ++j) column[i][j] = 0;
    for (Index i=n; i<ww; ++i)
        for (Index j=0; j<ELEMS; ++j) column[i][j] = 1;
    Index b;
    if (glynn) {
        b = n > POW+1 ? n - POW - 1: 0;
        auto c = n-1-b;
        for (Index k=0; k<l; k++) {
            Index gray = k ^ k/2;
            for (Index j=0; j< c; ++j)
                if (gray & 1 << j)
                    for (Index i=0; i<n; ++i)
                        column[i][k] -= input[(b+j)*n+i];
                else
                    for (Index i=0; i<n; ++i)
                        column[i][k] += input[(b+j)*n+i];
        }
        for (Index i=0; i<n; ++i)
            for (Index k=0; k<l; k++)
                column[i][k] += input[n*(n-1)+i];

        for (Index k=1; k<l; k+=2)
            column[0][k] = -column[0][k];

        for (Index i=0; i<b; ++i, bit <<= 1) {
            if (old_gray & bit) {
                bits = bits ^ 1;
                for (Index j=0; j<ww; ++j)
                    column[j] -= more[i][j];
            } else {
                for (Index j=0; j<ww; ++j)
                    column[j] += more[i][j];
            }
        }

        for (Index i=0; i<n; ++i)
            for (Index k=0; k<l; k++)
                column[i][k] /= 2;
    } else {
        b = n > POW ? n - POW : 0;
        auto c = n-b;
        for (Index k=0; k<l; k++) {
            Index gray = k ^ k/2;
            for (Index j=0; j<c; ++j)
                if (gray & 1 << j)
                    for (Index i=0; i<n; ++i)
                        column[i][k] -= input[(b+j)*n+i];
        }

        for (Index k=1; k<l; k+=2)
            column[0][k] = -column[0][k];

        for (Index i=0; i<b; ++i, bit <<= 1) {
            if (old_gray & bit) {
                bits = bits ^ 1;
                for (Index j=0; j<ww; ++j)
                    column[j] -= more[i][j];
            }
        }
    }

    if (DEBUG) {
        for (Index i=0; i<ww; ++i) {
            cout << "Column[" << i << "]=";
            for (Index j=0; j<ELEMS; ++j) cout << " " << column[i][j];
            cout << "\n";
        }
    }

    --more;
    old_gray = (from ^ from/2) | UINT64_C(1) << b;
    Value total = 0;
    SumN accu[WIDTH];
    for (auto p=from; p<to; ++p) {
        uint64_t new_gray = p ^ p/2;
        uint64_t bit = old_gray ^ new_gray;
        Index i = __builtin_ffsl(bit);
        auto diff = more[i];
        auto c = column;
        if (new_gray > old_gray) {
            // Phase 1 add/multiply.
            // Uses floats until just before loss of precision
            for (Index i=0; i<WIDTH; ++i) accu[i] = *c++ -= *diff++;

            for (Index j=1; j < nn; ++j)
                for (Index i=0; i<WIDTH; ++i) accu[i] *= *c++ -= *diff++;
        } else {
            // Phase 1 add/multiply.
            // Uses floats until just before loss of precision
            for (Index i=0; i<WIDTH; ++i) accu[i] = *c++ += *diff++;

            for (Index j=1; j < nn; ++j)
                for (Index i=0; i<WIDTH; ++i) accu[i] *= *c++ += *diff++;
        }

        if (DEBUG) {
            cout << "p=" << p << "\n";
            for (Index i=0; i<ww; ++i) {
                cout << "Column[" << i << "]=";
                for (Index j=0; j<ELEMS; ++j) cout << " " << column[i][j];
                cout << "\n";
            }
        }

        // Convert floats to int32_t
        ProdN prod32[WIDTH] __attribute__((aligned (32)));
        for (Index i=0; i<WIDTH; ++i)
            // Unfortunately gcc doesn't recognize the static_cast<int32_t>
            // as a vector pattern, so force it with an intrinsic
#ifdef __AVX__
            //prod32[i] = static_cast<ProdN>(accu[i]);
            reinterpret_cast<__m256i&>(prod32[i]) = _mm256_cvttps_epi32(accu[i]);
#else   // __AVX__
            for (Index j=0; j<ELEMS; ++j)
                prod32[i][j] = static_cast<int32_t>(accu[i][j]);
#endif  // __AVX__

        // Phase 2 multiply. Uses int64_t until just before overflow
        int64_t prod64[3][ELEMS];
        for (Index i=0; i<3; ++i) {
            for (Index j=0; j<ELEMS; ++j)
                prod64[i][j] = static_cast<int64_t>(prod32[i][j]) * prod32[i+3][j] * prod32[i+6][j];
        }
        // Phase 3 multiply. Collect into __int128. For large matrices this will
        // actually overflow but that's ok as long as all 128 low bits are
        // correct. Terms will cancel and the final sum can fit into 128 bits
        // (This will start to fail at n=35 for the all 1 matrix)
        // Strictly speaking this needs the -fwrapv gcc option
        for (Index j=0; j<ELEMS; ++j) {
            auto value = static_cast<Value>(prod64[0][j]) * prod64[1][j] * prod64[2][j];
            if (DEBUG) cout << "value[" << j << "]=" << static_cast<double>(value) << "\n";
            total += value;
        }
        total = -total;

        old_gray = new_gray;
    }

    return bits ? Number{-total} : Number{total};
}

// Prepare datastructures, Assign work to threads
Number permanent(Index n) {
    Index nn = (n+WIDTH-1)/WIDTH;
    Index ww = nn*WIDTH;

    Index rows  = n > (POW+glynn) ? n-POW-glynn : 0;
    auto data = alloc<SumN>(ww*(rows+1));
    auto pointers = alloc<SumN *>(rows+1);
    auto more = &pointers[0];
    for (Index i=0; i<rows; ++i)
        more[i] = &data[ww*i];
    more[rows] = &data[ww*rows];
    for (Index j=0; j<ww; ++j)
        for (Index i=0; i<ELEMS; ++i)
            more[rows][j][i] = 0;

    Index loops = n >= POW+glynn ? ELEMS : 1 << (n-glynn);
    auto a = &input[0];
    for (Index r=0; r<rows; ++r) {
        for (Index j=0; j<n; ++j) {
            for (Index i=0; i<loops; ++i)
                more[r][j][i] = j == 0 && i %2 ? -*a : *a;
            for (Index i=loops; i<ELEMS; ++i)
                more[r][j][i] = 0;
            ++a;
        }
        for (Index j=n; j<ww; ++j)
            for (Index i=0; i<ELEMS; ++i)
                more[r][j][i] = 0;
    }

    if (DEBUG)
        for (Index r=0; r<=rows; ++r)
            for (Index j=0; j<ww; ++j) {
                cout << "more[" << r << "][" << j << "]=";
                for (Index i=0; i<ELEMS; ++i)
                    cout << " " << more[r][j][i];
                cout << "\n";
            }

    // Send work to threads...
    vector<future<Number>> results;
    for (auto i=1U; i < nr_threads; ++i)
        results.emplace_back(async(DEBUG ? launch::deferred: launch::async, permanent_part, n, i, more));
    // And collect results
    auto r = permanent_part(n, 0, more);
    for (auto& result: results)
        r += result.get();

    free(data);
    free(pointers);

    // For glynn we should double the result, but we will only do this during
    // the final print. This allows n=34 for an all 1 matrix to work
    // if (glynn) r *= 2;
    return r;
}

// Print 128 bit number
void Number::print(ostream& os, bool dbl) const {
    const UValue BILLION = 1000000000;

    UValue val;
    if (value < 0) {
        os << "-";
        val = -value;
    } else
        val = value;
    if (dbl) val *= 2;

    uint32_t output[5];
    for (int i=0; i<5; ++i) {
        output[i] = val % BILLION;
        val /= BILLION;
    }
    bool print = false;
    for (int i=4; i>=0; --i) {
        if (print) {
            os << setfill('0') << setw(9) << output[i];
        } else if (output[i] || i == 0) {
            print = true;
            os << output[i];
        }
    }
}

// Read matrix, check for sanity
void my_main() {
    Sum a;
    while (cin >> a)
        input.push_back(a);

    size_t n = sqrt(input.size());
    if (input.size() != n*n)
        throw(logic_error("Read " + to_string(input.size()) +
                          " elements which does not make a square matrix"));

    vector<double> columns_pos(n, 0);
    vector<double> columns_neg(n, 0);
    Sum *p = &input[0];
    for (size_t i=0; i<n; ++i)
        for (size_t j=0; j<n; ++j, ++p) {
            if (*p >= 0) columns_pos[j] += *p;
            else         columns_neg[j] -= *p;
        }
    std::array<double,WIDTH> prod;
    prod.fill(1);

    int32_t odd = 0;
    for (size_t j=0; j<n; ++j) {
        prod[j%WIDTH] *= max(columns_pos[j], columns_neg[j]);
        auto sum = static_cast<int32_t>(columns_pos[j] - columns_neg[j]);
        odd |= sum;
    }
    glynn = (odd & 1) ^ 1;
    for (Index i=0; i<WIDTH; ++i)
        // A float has an implicit 1. in front of the fraction so it can
        // represent 1 bit more than the mantissa size. And 1 << (mantissa+1)
        // itself is in fact representable
        if (prod[i] && log2(prod[i]) > FLOAT_MANTISSA+1)
            throw(range_error("Values in matrix are too large. A subproduct reaches " + to_string(prod[i]) + " which doesn't fit in a float without loss of precision"));

    for (Index i=0; i<3; ++i) {
        auto prod3 = prod[i] * prod[i+3] * prod[i+6];
        if (log2(prod3) >= CHAR_BIT*sizeof(int64_t)-1)
            throw(range_error("Values in matrix are too large. A subproduct reaches " + to_string(prod3) + " which doesn't fit in an int64"));
    }

    nr_threads = pow(2, ceil(log2(static_cast<float>(nr_threads))));
    uint64_t loops = UINT64_C(1) << n;
    if (glynn) loops /= 2;
    if (nr_threads * ELEMS > loops)
        nr_threads = max(loops / ELEMS, UINT64_C(1));
    // if (DEBUG) nr_threads = 1;

    cout << n << " x " << n << " matrix, method " << (glynn ? "Glynn" : "Ryser") << ", " << nr_threads << " threads" << endl;

    // Go for the actual calculation
    auto start = chrono::steady_clock::now();
    auto perm = permanent(n);
    auto end = chrono::steady_clock::now();
    auto elapsed = chrono::duration_cast<ms>(end-start).count();

    cout << "Permanent=";
    perm.print(cout, glynn);
    cout << " (" << elapsed / 1000. << " s)" << endl;
}

// Wrapper to print any exceptions
int main() {
    try {
        my_main();
    } catch(exception& e) {
        cerr << "Error: " << e.what() << endl;
        exit(EXIT_FAILURE);
    }
    exit(EXIT_SUCCESS);
}

C99，n≈33（35秒）

#include <stdint.h>
#include <stdio.h>

#define CHUNK_SIZE 12
#define NUM_THREADS 8

#define popcnt __builtin_popcountll
#define BILLION (1000 * 1000 * 1000)
#define UPDATE_ROW_PPROD() \
    update_row_pprod(row_pprod, row, rows, row_sums, mask, mask_popcnt)

typedef __int128 int128_t;

static inline int64_t update_row_pprod
(
    int64_t* row_pprod, int64_t row, int64_t* rows,
    int64_t* row_sums, int64_t mask, int64_t mask_popcnt
)
{
    int64_t temp = 2 * popcnt(rows[row] & mask) - mask_popcnt;

    row_pprod[0] *= temp;
    temp -= 1;
    row_pprod[1] *= temp;
    temp -= row_sums[row];
    row_pprod[2] *= temp;
    temp += 1;
    row_pprod[3] *= temp;

    return row + 1;
}

int main(int argc, char* argv[])
{
    int64_t size = argc - 1, rows[argc - 1];
    int64_t row_sums[argc - 1];
    int128_t permanent = 0, sign = size & 1 ? -1 : 1;

    if (argc == 2)
    {
        printf("%d\n", argv[1][0] == '-' ? -1 : 1);
        return 0;
    }

    for (int64_t row = 0; row < size; row++)
    {
        char positive = argv[row + 1][0] == '+' ? '-' : '+';

        sign *= ',' - positive;
        rows[row] = row_sums[row] = 0;

        for (char* p = &argv[row + 1][1]; *p; p++)
        {
            rows[row] <<= 1;
            rows[row] |= *p == positive;
            row_sums[row] += *p == positive;
        }

        row_sums[row] = 2 * row_sums[row] - size;
    }

    #pragma omp parallel for reduction(+:permanent) num_threads(NUM_THREADS)
    for (int64_t mask = 1; mask < 1LL << (size - 1); mask += 2)
    {
        int64_t mask_popcnt = popcnt(mask);
        int64_t row = 0;
        int128_t row_prod = 1 - 2 * (mask_popcnt & 1);
        int128_t row_prod_high = -row_prod;
        int128_t row_prod_inv = row_prod;
        int128_t row_prod_inv_high = -row_prod;

        for (int64_t chunk = 0; chunk < size / CHUNK_SIZE; chunk++)
        {
            int64_t row_pprod[4] = {1, 1, 1, 1};

            for (int64_t i = 0; i < CHUNK_SIZE; i++)
                row = UPDATE_ROW_PPROD();

            row_prod *= row_pprod[0], row_prod_high *= row_pprod[1];
            row_prod_inv *= row_pprod[3], row_prod_inv_high *= row_pprod[2];
        }

        int64_t row_pprod[4] = {1, 1, 1, 1};

        while (row < size)
            row = UPDATE_ROW_PPROD();

        row_prod *= row_pprod[0], row_prod_high *= row_pprod[1];
        row_prod_inv *= row_pprod[3], row_prod_inv_high *= row_pprod[2];
        permanent += row_prod + row_prod_high + row_prod_inv + row_prod_inv_high;
    }

    permanent *= sign;

    if (permanent < 0)
        printf("-"), permanent *= -1;

    int32_t output[5], print = 0;

    output[0] = permanent % BILLION, permanent /= BILLION;
    output[1] = permanent % BILLION, permanent /= BILLION;
    output[2] = permanent % BILLION, permanent /= BILLION;
    output[3] = permanent % BILLION, permanent /= BILLION;
    output[4] = permanent % BILLION;

    if (output[4])
        printf("%u", output[4]), print = 1;
    if (print)
        printf("%09u", output[3]);
    else if (output[3])
        printf("%u", output[3]), print = 1;
    if (print)
        printf("%09u", output[2]);
    else if (output[2])
        printf("%u", output[2]), print = 1;
    if (print)
        printf("%09u", output[1]);
    else if (output[1])
        printf("%u", output[1]), print = 1;
    if (print)
        printf("%09u\n", output[0]);
    else
        printf("%u\n", output[0]);
}

目前输入有点麻烦；它以行作为命令行参数，其中每个条目均以其符号表示，即+表示1且-表示-1。

测试运行

$ gcc -Wall -std=c99 -march=native -Ofast -fopenmp -fwrapv -o permanent permanent.c
$ ./permanent +--+ ---+ -+-+ +--+
-4
$ ./permanent ---- -+-- +--- +-+-
0
$ ./permanent +-+----- --++-++- +----+++ ---+-+++ +--++++- -+-+-++- +-+-+-+- --+-++++
192
$ ./permanent +-+--+++----++++-++- +-+++++-+--+++--+++- --+++----+-+++---+-- ---++-++++++------+- -+++-+++---+-+-+++++ +-++--+-++++-++-+--- +--+---+-++++---+++- +--+-++-+++-+-+++-++ +-----+++-----++-++- --+-+-++-+-++++++-++ -------+----++++---- ++---++--+-++-++++++ -++-----++++-----+-+ ++---+-+----+-++-+-+ +++++---+++-+-+++-++ +--+----+--++-+----- -+++-++--+++--++--++ ++--++-++-+++-++-+-+ +++---+--++---+----+ -+++-------++-++-+--
1021509632
$ time ./permanent +++++++++++++++++++++++++++++++{,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,}     # 31
8222838654177922817725562880000000

real    0m8.365s
user    1m6.504s
sys     0m0.000s
$ time ./permanent ++++++++++++++++++++++++++++++++{,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,}   # 32
263130836933693530167218012160000000

real    0m17.013s
user    2m15.226s
sys     0m0.001s
$ time ./permanent +++++++++++++++++++++++++++++++++{,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,} # 33
8683317618811886495518194401280000000

real    0m34.592s
user    4m35.354s
sys     0m0.001s

Python 2，n≈28

from operator import mul

def fast_glynn_perm(M):
    row_comb = [sum(c) for c in zip(*M)]
    n=len(M)

    total = 0
    old_grey = 0 
    sign = +1

    binary_power_dict = {2**i:i for i in range(n)}
    num_loops = 2**(n-1)

    for bin_index in xrange(1, num_loops + 1):  
        total += sign * reduce(mul, row_comb)

        new_grey = bin_index^(bin_index/2)
        grey_diff = old_grey ^ new_grey
        grey_diff_index = binary_power_dict[grey_diff]

        new_vector = M[grey_diff_index]
        direction = 2 * cmp(old_grey,new_grey)      

        for i in range(n):
            row_comb[i] += new_vector[i] * direction

        sign = -sign
        old_grey = new_grey

    return total/num_loops

使用带有灰色代码的Glynn公式进行更新。一分钟即可在我的计算机上运行。当然可以用一种更快的语言和更好的数据结构来更好地实现这一点。这不使用矩阵为±1值。n=23

Ryser公式的实现非常相似，对系数的所有0/1向量求和，而不是对±1向量求和。它花费了Glynn公式大约两倍的时间，因为将所有2 ^ n个这样的向量相加，而Glynn将对称使用的半数仅加到了以开头的半数+1。

from operator import mul

def fast_ryser_perm(M):
    n=len(M)
    row_comb = [0] * n

    total = 0
    old_grey = 0 
    sign = +1

    binary_power_dict = {2**i:i for i in range(n)}
    num_loops = 2**n

    for bin_index in range(1, num_loops) + [0]: 
        total += sign * reduce(mul, row_comb)

        new_grey = bin_index^(bin_index/2)
        grey_diff = old_grey ^ new_grey
        grey_diff_index = binary_power_dict[grey_diff]

        new_vector = M[grey_diff_index]
        direction = cmp(old_grey, new_grey)

        for i in range(n):
            row_comb[i] += new_vector[i] * direction

        sign = -sign
        old_grey = new_grey

    return total * (-1)**n

锈+极品

这种带有格雷码实现的简单Ryser 在笔记本电脑上运行n = 31 大约需要65 90秒。我想您的机器将在60秒钟内到达目标。我正在使用extprim 1.1.1 i128。

我从未使用过Rust，也不知道我在做什么。除其他cargo build --release功能外，没有编译器选项。评论/建议/优化表示赞赏。

调用与Dennis的程序相同。

use std::env;
use std::thread;
use std::sync::Arc;
use std::sync::mpsc;

extern crate extprim;
use extprim::i128::i128;

static THREADS : i64 = 8; // keep this a power of 2.

fn main() {
  // Read command line args for the matrix, specified like
  // "++- --- -+-" for [[1, 1, -1], [-1, -1, -1], [-1, 1, -1]].
  let mut args = env::args();
  args.next();

  let mat : Arc<Vec<Vec<i64>>> = Arc::new(args.map( |ss|
    ss.trim().bytes().map( |cc| if cc == b'+' {1} else {-1}).collect()
  ).collect());

  // Figure how many iterations each thread has to do.
  let size = 2i64.pow(mat.len() as u32);
  let slice_size = size / THREADS; // Assumes divisibility.

  let mut accumulator : i128;
  if slice_size >= 4 { // permanent() requires 4 divides slice_size
    let (tx, rx) = mpsc::channel();

    // Launch threads.
    for ii in 0..THREADS {
      let mat = mat.clone();
      let tx = tx.clone();
      thread::spawn(move ||
        tx.send(permanent(&mat, ii * slice_size, (ii+1) * slice_size))
      );
    }

    // Accumulate results.
    accumulator = extprim::i128::ZERO;
    for _ in 0..THREADS {
      accumulator += rx.recv().unwrap();
    }
  }
  else { // Small matrix, don't bother threading.
    accumulator = permanent(&mat, 0, size);
  }
  println!("{}", accumulator);
}

fn permanent(mat: &Vec<Vec<i64>>, start: i64, end: i64) -> i128 {
  let size = mat.len();
  let sentinel = std::i64::MAX / size as i64;

  let mut bits : Vec<bool> = Vec::with_capacity(size);
  let mut sums : Vec<i64> = Vec::with_capacity(size);

  // Initialize gray code bits.
  let gray_number = start ^ (start / 2);

  for row in 0..size {
    bits.push((gray_number >> row) % 2 == 1);
    sums.push(0);
  }

  // Initialize column sums
  for row in 0..size {
    if bits[row] {
      for column in 0..size {
        sums[column] += mat[row][column];
      }
    }
  }

  // Do first two iterations with initial sums
  let mut total = product(&sums, sentinel);
  for column in 0..size {
    sums[column] += mat[0][column];
  }
  bits[0] = true;

  total -= product(&sums, sentinel);

  // Do rest of iterations updating gray code bits incrementally
  let mut gray_bit : usize;
  let mut idx = start + 2;
  while idx < end {
    gray_bit = idx.trailing_zeros() as usize;

    if bits[gray_bit] {
      for column in 0..size {
        sums[column] -= mat[gray_bit][column];
      }
      bits[gray_bit] = false;
    }
    else {
      for column in 0..size {
        sums[column] += mat[gray_bit][column];
      }
      bits[gray_bit] = true;
    }

    total += product(&sums, sentinel);

    if bits[0] {
      for column in 0..size {
        sums[column] -= mat[0][column];
      }
      bits[0] = false;
    }
    else {
      for column in 0..size {
        sums[column] += mat[0][column];
      }
      bits[0] = true;
    }

    total -= product(&sums, sentinel);
    idx += 2;
  }
  return if size % 2 == 0 {total} else {-total};
}

#[inline]
fn product(sums : &Vec<i64>, sentinel : i64) -> i128 {
  let mut ret : Option<i128> = None;
  let mut tally = sums[0];
  for ii in 1..sums.len() {
    if tally.abs() >= sentinel {
      ret = Some(ret.map_or(i128::new(tally), |n| n * i128::new(tally)));
      tally = sums[ii];
    }
    else {
      tally *= sums[ii];
    }
  }
  if ret.is_none() {
    return i128::new(tally);
  }
  return ret.unwrap() * i128::new(tally);
}

Haskell，n = 31（54秒）

@Angs做出了许多宝贵贡献：use Vector，使用短路产品，看一下奇数n。

import Control.Parallel.Strategies
import qualified Data.Vector.Unboxed as V
import Data.Int

type Row = V.Vector Int8

x :: Row -> [Row] -> Integer -> Int -> Integer
x p (v:vs) m c = let c' = c - 1
                     r = if c>0 then parTuple2 rseq rseq else r0
                     (a,b) = ( x p                  vs m    c' ,
                               x (V.zipWith(-) p v) vs (-m) c' )
                             `using` r
                 in a+b
x p _      m _ = prod m p

prod :: Integer -> Row -> Integer
prod m p = if 0 `V.elem` p then 0 
                           else V.foldl' (\a b->a*fromIntegral b) m p

p, pt :: [Row] -> Integer
p (v:vs) = x (foldl (V.zipWith (+)) v vs) (map (V.map (2*)) vs) 1 11
           `div` 2^(length vs)
p [] = 1 -- handle 0x0 matrices too  :-)

pt (v:vs) | even (length vs) = p ((V.map (2*) v) : vs ) `div` 2
pt mat                       = p mat

main = getContents >>= print . pt . map V.fromList . read

我在Haskell中首次尝试并行化。您可以在修订历史记录中看到许多优化步骤。令人惊讶的是，这些更改几乎都是很小的。该代码基于Wikipedia文章中关于计算永久物的 “ Balasubramanian-Bax / Franklin-Glynn公式”部分中的公式。

p计算永久物。称为，通过pt它将始终有效的方式转换矩阵，但对我们在此获得的矩阵特别有用。

用编译ghc -O2 -threaded -fllvm -feager-blackholing -o <name> <name>.hs。要使用并行化运行，请为其提供运行时参数，例如：./<name> +RTS -N。输入来自标准输入，带有嵌套逗号分隔列表的方括号，[[1,2],[3,4]]如上一个示例所示（到处都允许换行）。

Mathematica，n≈20

p[m_] := Last[Fold[Take[ListConvolve[##, {1, -1}, 0], 2^Length[m]]&,
  Table[If[IntegerQ[Log2[k]], m[[j, Log2[k] + 1]], 0], {j, n}, {k, 0, 2^Length[m] - 1}]]]

使用该Timing命令，在我的系统上20x20矩阵大约需要48秒。这并不是完全一样高效，因为它依赖于以下事实：可以将永久物作为矩阵每一行的多项式乘积的系数来找到。通过创建系数列表并使用进行卷积，可以执行有效的多项式乘法ListConvolve。假设使用快速傅立叶变换或类似方法执行卷积，则需要O（n log n）时间，这大约需要O（2 ⁿ n ²）时间。

Python 2，n = 22 [参考]

这是我昨天与Lembik共享的“参考”实现，它错过n=23了在他的计算机上花几秒钟的时间，而在我的计算机上大约52秒就完成了。为了达到这些速度，您需要通过PyPy运行它。

第一个函数通过遍历每个子矩阵直到您剩下一个2x2的基数来计算永久性，类似于如何计算行列式，您可以应用基本规则。这是非常慢的。

第二个函数是实现Ryser函数的函数（维基百科中列出的第二个等式）。该集合S本质上是数字的幂集{1,...,n}（s_list代码中的变量）。

from random import *
from time import time
from itertools import*

def perm(a): # naive method, recurses over submatrices, slow 
    if len(a) == 1:
        return a[0][0]
    elif len(a) == 2:
        return a[0][0]*a[1][1]+a[1][0]*a[0][1]
    else:
        tsum = 0
        for i in xrange(len(a)):
            transposed = [zip(*a)[j] for j in xrange(len(a)) if j != i]
            tsum += a[0][i] * perm(zip(*transposed)[1:])
        return tsum

def perm_ryser(a): # Ryser's formula, using matrix entries
    maxn = len(a)
    n_list = range(1,maxn+1)
    s_list = chain.from_iterable(combinations(n_list,i) for i in range(maxn+1))
    total = 0
    for st in s_list:
        stotal = (-1)**len(st)
        for i in xrange(maxn):
            stotal *= sum(a[i][j-1] for j in st)
        total += stotal
    return total*((-1)**maxn)


def genmatrix(d):
    mat = []
    for x in xrange(d):
        row = []
        for y in xrange(d):
            row.append([-1,1][randrange(0,2)])
        mat.append(row)
    return mat

def main():
    for i in xrange(1,24):
        k = genmatrix(i)
        print 'Matrix: (%dx%d)'%(i,i)
        print '\n'.join('['+', '.join(`j`.rjust(2) for j in a)+']' for a in k)
        print 'Permanent:',
        t = time()
        p = perm_ryser(k)
        print p,'(took',time()-t,'seconds)'

if __name__ == '__main__':
    main()

RPython 5.4.1，n≈32（37秒）

from rpython.rlib.rtime import time
from rpython.rlib.rarithmetic import r_int, r_uint
from rpython.rlib.rrandom import Random
from rpython.rlib.rposix import pipe, close, read, write, fork, waitpid
from rpython.rlib.rbigint import rbigint

from math import log, ceil
from struct import pack

bitsize = len(pack('l', 1)) * 8 - 1

bitcounts = bytearray([0])
for i in range(16):
  b = bytearray([j+1 for j in bitcounts])
  bitcounts += b


def bitcount(n):
  bits = 0
  while n:
    bits += bitcounts[n & 65535]
    n >>= 16
  return bits


def main(argv):
  if len(argv) < 2:
    write(2, 'Usage: %s NUM_THREADS [N]'%argv[0])
    return 1
  threads = int(argv[1])

  if len(argv) > 2:
    n = int(argv[2])
    rnd = Random(r_uint(time()*1000))
    m = []
    for i in range(n):
      row = []
      for j in range(n):
        row.append(1 - r_int(rnd.genrand32() & 2))
      m.append(row)
  else:
    m = []
    strm = ""
    while True:
      buf = read(0, 4096)
      if len(buf) == 0:
        break
      strm += buf
    rows = strm.split("\n")
    for row in rows:
      r = []
      for val in row.split(' '):
        r.append(int(val))
      m.append(r)
    n = len(m)

  a = []
  for row in m:
    val = 0
    for v in row:
      val = (val << 1) | -(v >> 1)
    a.append(val)

  batches = int(ceil(n * log(n) / (bitsize * log(2))))

  pids = []
  handles = []
  total = rbigint.fromint(0)
  for i in range(threads):
    r, w = pipe()
    pid = fork()
    if pid:
      close(w)
      pids.append(pid)
      handles.append(r)
    else:
      close(r)
      total = run(n, a, i, threads, batches)
      write(w, total.str())
      close(w)
      return 0

  for pid in pids:
    waitpid(pid, 0)

  for handle in handles:
    strval = read(handle, 256)
    total = total.add(rbigint.fromdecimalstr(strval))
    close(handle)

  print total.rshift(n-1).str()

  return 0


def run(n, a, mynum, threads, batches):
  start = (1 << n-1) * mynum / threads
  end = (1 << n-1) * (mynum+1) / threads

  dtotal = rbigint.fromint(0)
  for delta in range(start, end):
    pdelta = rbigint.fromint(1 - ((bitcount(delta) & 1) << 1))
    for i in range(batches):
      pbatch = 1
      for j in range(i, n, batches):
        pbatch *= n - (bitcount(delta ^ a[j]) << 1)
      pdelta = pdelta.int_mul(pbatch)
    dtotal = dtotal.add(pdelta)

  return dtotal


def target(*args):
  return main

要进行编译，请下载最新的PyPy源，然后执行以下命令：

pypy /path/to/pypy-src/rpython/bin/rpython matrix-permanent.py

生成的可执行文件将matrix-permanent-c在当前工作目录中被命名或类似。

从PyPy 5.0开始，RPython的线程原语比以前少了很多。新产生的线程需要GIL，它对于并行计算几乎没有用。我fork改用了它，因此尽管我没有测试过编译失败（unresolved external symbol _fork），但它可能无法在Windows上按预期工作。

该可执行文件最多接受两个命令行参数。第一个是线程数，第二个可选参数是n。如果提供，将生成一个随机矩阵，否则将从标准输入读取。每行必须用换行符分隔（没有尾随换行符），并且每个值空间都必须分隔。第三个示例输入为：

1 -1 1 -1 -1 1 1 1 -1 -1 -1 -1 1 1 1 1 -1 1 1 -1
1 -1 1 1 1 1 1 -1 1 -1 -1 1 1 1 -1 -1 1 1 1 -1
-1 -1 1 1 1 -1 -1 -1 -1 1 -1 1 1 1 -1 -1 -1 1 -1 -1
-1 -1 -1 1 1 -1 1 1 1 1 1 1 -1 -1 -1 -1 -1 -1 1 -1
-1 1 1 1 -1 1 1 1 -1 -1 -1 1 -1 1 -1 1 1 1 1 1
1 -1 1 1 -1 -1 1 -1 1 1 1 1 -1 1 1 -1 1 -1 -1 -1
1 -1 -1 1 -1 -1 -1 1 -1 1 1 1 1 -1 -1 -1 1 1 1 -1
1 -1 -1 1 -1 1 1 -1 1 1 1 -1 1 -1 1 1 1 -1 1 1
1 -1 -1 -1 -1 -1 1 1 1 -1 -1 -1 -1 -1 1 1 -1 1 1 -1
-1 -1 1 -1 1 -1 1 1 -1 1 -1 1 1 1 1 1 1 -1 1 1
-1 -1 -1 -1 -1 -1 -1 1 -1 -1 -1 -1 1 1 1 1 -1 -1 -1 -1
1 1 -1 -1 -1 1 1 -1 -1 1 -1 1 1 -1 1 1 1 1 1 1
-1 1 1 -1 -1 -1 -1 -1 1 1 1 1 -1 -1 -1 -1 -1 1 -1 1
1 1 -1 -1 -1 1 -1 1 -1 -1 -1 -1 1 -1 1 1 -1 1 -1 1
1 1 1 1 1 -1 -1 -1 1 1 1 -1 1 -1 1 1 1 -1 1 1
1 -1 -1 1 -1 -1 -1 -1 1 -1 -1 1 1 -1 1 -1 -1 -1 -1 -1
-1 1 1 1 -1 1 1 -1 -1 1 1 1 -1 -1 1 1 -1 -1 1 1
1 1 -1 -1 1 1 -1 1 1 -1 1 1 1 -1 1 1 -1 1 -1 1
1 1 1 -1 -1 -1 1 -1 -1 1 1 -1 -1 -1 1 -1 -1 -1 -1 1
-1 1 1 1 -1 -1 -1 -1 -1 -1 -1 1 1 -1 1 1 -1 1 -1 -1

样品用量

$ time ./matrix-permanent-c 8 30
8395059644858368

real    0m8.582s
user    1m8.656s
sys     0m0.000s

方法

我使用了Balasubramanian-Bax / Franklin-Glynn公式，其运行时复杂度为O（2 ⁿ n）。但是，我没有以格雷码顺序迭代δ，而是用单个异或运算（映射（1，-1）→（0，1））代替了矢量行乘法。通过将n减去两倍的popcount，同样可以在单个操作中找到向量和。

球拍84字节

以下简单函数适用于较小的矩阵，但对于较大的矩阵却挂在我的机器上：

(for/sum((p(permutations(range(length l)))))(for/product((k l)(c p))(list-ref k c)))

取消高尔夫：

(define (f ll) 
  (for/sum ((p (permutations (range (length ll))))) 
    (for/product ((l ll)(c p)) 
      (list-ref l c))))

可以很容易地将代码修改为行数和列数不相等。

测试：

(f '[[ 1 -1 -1  1]
     [-1 -1 -1  1]
     [-1  1 -1  1]
     [ 1 -1 -1  1]])

(f '[[ 1 -1  1 -1 -1 -1 -1 -1]
 [-1 -1  1  1 -1  1  1 -1]
 [ 1 -1 -1 -1 -1  1  1  1]
 [-1 -1 -1  1 -1  1  1  1]
 [ 1 -1 -1  1  1  1  1 -1]
 [-1  1 -1  1 -1  1  1 -1]
 [ 1 -1  1 -1  1 -1  1 -1]
 [-1 -1  1 -1  1  1  1  1]])

输出：

-4
192

正如我上面提到的，它挂在以下测试上：

(f '[[1 -1 1 -1 -1 1 1 1 -1 -1 -1 -1 1 1 1 1 -1 1 1 -1]
 [1 -1 1 1 1 1 1 -1 1 -1 -1 1 1 1 -1 -1 1 1 1 -1]
 [-1 -1 1 1 1 -1 -1 -1 -1 1 -1 1 1 1 -1 -1 -1 1 -1 -1]
 [-1 -1 -1 1 1 -1 1 1 1 1 1 1 -1 -1 -1 -1 -1 -1 1 -1]
 [-1 1 1 1 -1 1 1 1 -1 -1 -1 1 -1 1 -1 1 1 1 1 1]
 [1 -1 1 1 -1 -1 1 -1 1 1 1 1 -1 1 1 -1 1 -1 -1 -1]
 [1 -1 -1 1 -1 -1 -1 1 -1 1 1 1 1 -1 -1 -1 1 1 1 -1]
 [1 -1 -1 1 -1 1 1 -1 1 1 1 -1 1 -1 1 1 1 -1 1 1]
 [1 -1 -1 -1 -1 -1 1 1 1 -1 -1 -1 -1 -1 1 1 -1 1 1 -1]
 [-1 -1 1 -1 1 -1 1 1 -1 1 -1 1 1 1 1 1 1 -1 1 1]
 [-1 -1 -1 -1 -1 -1 -1 1 -1 -1 -1 -1 1 1 1 1 -1 -1 -1 -1]
 [1 1 -1 -1 -1 1 1 -1 -1 1 -1 1 1 -1 1 1 1 1 1 1]
 [-1 1 1 -1 -1 -1 -1 -1 1 1 1 1 -1 -1 -1 -1 -1 1 -1 1]
 [1 1 -1 -1 -1 1 -1 1 -1 -1 -1 -1 1 -1 1 1 -1 1 -1 1]
 [1 1 1 1 1 -1 -1 -1 1 1 1 -1 1 -1 1 1 1 -1 1 1]
 [1 -1 -1 1 -1 -1 -1 -1 1 -1 -1 1 1 -1 1 -1 -1 -1 -1 -1]
 [-1 1 1 1 -1 1 1 -1 -1 1 1 1 -1 -1 1 1 -1 -1 1 1]
 [1 1 -1 -1 1 1 -1 1 1 -1 1 1 1 -1 1 1 -1 1 -1 1]
 [1 1 1 -1 -1 -1 1 -1 -1 1 1 -1 -1 -1 1 -1 -1 -1 -1 1]
 [-1 1 1 1 -1 -1 -1 -1 -1 -1 -1 1 1 -1 1 1 -1 1 -1 -1]])