代码

// v7 + 4-pass responsibility brute-force (exact, no dedup, sequential reads).
// Responsibility: q counted in smallest dim d where q is in j's partial block.
// e<d uses ordpos (position) for "complete block" check; e>d uses ordrank (first_pos).
#pragma GCC target("avx2,bmi")
#pragma GCC optimize("O3")
#include <cstring>
#include <cstdlib>
#include <algorithm>
#include <sys/mman.h>
#include <immintrin.h>
typedef unsigned u32; typedef unsigned long long u64;
static int order[4][100010];
static int first_pos[4][100010];
static int cnt[100010];
static int ordrank[4][4][100010];
static int ordpos[4][4][100010];
static int pos[4][100010];
static u32 partial[100010];
static u32 corr[100010];
static int proc[100010];
static const unsigned char pc_lut_bytes[32] = {0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4};
static inline __m256i pc_bytes(__m256i v){
    const __m256i lut = _mm256_loadu_si256((const __m256i*)pc_lut_bytes);
    const __m256i m4 = _mm256_set1_epi8(0x0F);
    __m256i lo = _mm256_and_si256(v, m4);
    __m256i hi = _mm256_and_si256(_mm256_srli_epi16(v, 4), m4);
    lo = _mm256_shuffle_epi8(lut, lo); hi = _mm256_shuffle_epi8(lut, hi);
    return _mm256_add_epi8(lo, hi);
}
static inline u32 sum4(__m256i v){ return (u32)(_mm256_extract_epi64(v,0)+_mm256_extract_epi64(v,1)+_mm256_extract_epi64(v,2)+_mm256_extract_epi64(v,3)); }
void count_4d(int n, const unsigned *x[4], unsigned *out) {
    int NW = (n + 63) >> 6;
    const int B = 512;
    int M = (n + B - 1) / B;
    int NWp = (NW + 7) & ~7;
    for (int d = 0; d < 4; d++) {
        const unsigned *xd = x[d];
        memset(cnt, 0, sizeof(int) * (n + 1));
        for (int i = 0; i < n; i++) cnt[xd[i]]++;
        int s = 0;
        for (int v = 0; v <= (int)n; v++) { int c = cnt[v]; cnt[v] = s; s += c; }
        for (int i = 0; i < n; i++) first_pos[d][i] = cnt[xd[i]];
        memset(cnt, 0, sizeof(int) * (n + 1));
        for (int i = 0; i < n; i++) cnt[xd[i]]++;
        for (int v = 1; v <= (int)n; v++) cnt[v] += cnt[v-1];
        for (int i = n - 1; i >= 0; i--) order[d][--cnt[xd[i]]] = i;
        for (int t = 0; t < n; t++) pos[d][order[d][t]] = t;
    }
    for (int d = 0; d < 4; d++)
        for (int a = 0; a < 4; a++)
            for (int t = 0; t < n; t++) {
                ordrank[d][a][t] = first_pos[a][order[d][t]];
                ordpos[d][a][t] = pos[a][order[d][t]];
            }
    size_t prefsz = (size_t)4 * (M + 1) * NWp * sizeof(u64);
    const int S = 8192;
    {
        int p = 0;
        for (int s = 0; s < n; s += S) {
            int e = s + S; if (e > n) e = n;
            for (int t = s; t < e; t++) proc[p++] = order[0][t];
            std::sort(proc + p - (e - s), proc + p, [](int a, int b){
                int a1 = first_pos[1][a], b1 = first_pos[1][b];
                if (a1 != b1) return a1 < b1;
                return first_pos[2][a] < first_pos[2][b];
            });
        }
    }
    void *rawp = malloc(prefsz + 64);
    u64 *pref = (u64*)(((size_t)rawp + 63) & ~(size_t)63);

    for (int d = 0; d < 4; d++) {
        u64 *p = pref + (size_t)d * (M + 1) * NWp;
        memset(p, 0, (size_t)NWp * sizeof(u64));
        for (int b = 0; b < M; b++) {
            u64 *cur = p + (size_t)(b + 1) * NWp;
            u64 *prev = p + (size_t)b * NWp;
            memcpy(cur, prev, (size_t)NWp * sizeof(u64));
            int e = (b + 1) * B; if (e > n) e = n;
            for (int t = b * B; t < e; t++) {
                int pt = order[d][t];
                cur[pt >> 6] |= 1ULL << (pt & 63);
            }
        }
    }
    u64 *p0 = pref;
    u64 *p1 = pref + (size_t)(M + 1) * NWp;
    u64 *p2 = pref + (size_t)2 * (M + 1) * NWp;
    u64 *p3 = pref + (size_t)3 * (M + 1) * NWp;
    for (int pos = 0; pos < n; pos += 2) {
        int j0 = proc[pos];
        int j1 = (pos + 1 < n) ? proc[pos + 1] : -1;
        const u64 *q0a = p0 + (size_t)(first_pos[0][j0] / B) * NWp;
        const u64 *q1a = p1 + (size_t)(first_pos[1][j0] / B) * NWp;
        const u64 *q2a = p2 + (size_t)(first_pos[2][j0] / B) * NWp;
        const u64 *q3a = p3 + (size_t)(first_pos[3][j0] / B) * NWp;
        const u64 *q0b = q0a, *q1b = q1a, *q2b = q2a, *q3b = q3a;
        if (j1 >= 0) {
            q0b = p0 + (size_t)(first_pos[0][j1] / B) * NWp;
            q1b = p1 + (size_t)(first_pos[1][j1] / B) * NWp;
            q2b = p2 + (size_t)(first_pos[2][j1] / B) * NWp;
            q3b = p3 + (size_t)(first_pos[3][j1] / B) * NWp;
        }
        __m256i acc0 = _mm256_setzero_si256();
        __m256i acc1 = _mm256_setzero_si256();
        int w = 0;
        for (; w + 4 <= NW; w += 4) {
            __m256i a0 = _mm256_load_si256((const __m256i*)(q0a + w));
            __m256i a1 = _mm256_load_si256((const __m256i*)(q1a + w));
            __m256i a2 = _mm256_load_si256((const __m256i*)(q2a + w));
            __m256i a3 = _mm256_stream_load_si256((__m256i*)(q3a + w));
            __m256i b0 = _mm256_load_si256((const __m256i*)(q0b + w));
            __m256i b1 = _mm256_load_si256((const __m256i*)(q1b + w));
            __m256i b2 = _mm256_load_si256((const __m256i*)(q2b + w));
            __m256i b3 = _mm256_stream_load_si256((__m256i*)(q3b + w));
            __m256i ra = _mm256_and_si256(_mm256_and_si256(a0,a1), _mm256_and_si256(a2,a3));
            __m256i rb = _mm256_and_si256(_mm256_and_si256(b0,b1), _mm256_and_si256(b2,b3));
            acc0 = _mm256_add_epi64(acc0, _mm256_sad_epu8(pc_bytes(ra), _mm256_setzero_si256()));
            acc1 = _mm256_add_epi64(acc1, _mm256_sad_epu8(pc_bytes(rb), _mm256_setzero_si256()));
        }
        u32 c0 = sum4(acc0), c1 = sum4(acc1);
        for (; w < NW; w++) {
            c0 += (u32)__builtin_popcountll(q0a[w] & q1a[w] & q2a[w] & q3a[w]);
            if (j1 >= 0) c1 += (u32)__builtin_popcountll(q0b[w] & q1b[w] & q2b[w] & q3b[w]);
        }
        partial[j0] = c0;
        if (j1 >= 0) partial[j1] = c1;
    }
    // 4-pass responsibility brute force
    memset(corr, 0, sizeof(u32) * n);
    for (int d = 0; d < 4; d++) {
        int e0 = (d == 0) ? 1 : 0;
        int e1 = (d <= 1) ? 2 : 1;
        int e2 = 3 - d;
        if (d == 0) { e0 = 1; e1 = 2; e2 = 3; }
        else if (d == 1) { e0 = 0; e1 = 2; e2 = 3; }
        else if (d == 2) { e0 = 0; e1 = 1; e2 = 3; }
        else { e0 = 0; e1 = 1; e2 = 2; }
        // select candidate arrays: e<d -> ordpos, e>d -> ordrank
        int* A = (e0 < d) ? ordpos[d][e0] : ordrank[d][e0];
        int* B_ = (e1 < d) ? ordpos[d][e1] : ordrank[d][e1];
        int* C = (e2 < d) ? ordpos[d][e2] : ordrank[d][e2];
        int* Dsel = ordrank[d][d];
        for (int t = 0; t < n; t++) {
            int j = order[d][t];
            int f = Dsel[t];
            int bs = (f / B) * B;
            int ta = (e0 < d) ? ((ordrank[d][e0][t] / B) * B) : ordrank[d][e0][t];
            int tb = (e1 < d) ? ((ordrank[d][e1][t] / B) * B) : ordrank[d][e1][t];
            int tc = (e2 < d) ? ((ordrank[d][e2][t] / B) * B) : ordrank[d][e2][t];
            __m256i vta = _mm256_set1_epi32(ta);
            __m256i vtb = _mm256_set1_epi32(tb);
            __m256i vtc = _mm256_set1_epi32(tc);
            int p = bs;
            int cntm = 0;
            for (; p + 16 <= f; p += 16) {
                __m256i a1 = _mm256_loadu_si256((const __m256i*)(A + p));
                __m256i a2 = _mm256_loadu_si256((const __m256i*)(A + p + 8));
                __m256i b1 = _mm256_loadu_si256((const __m256i*)(B_ + p));
                __m256i b2 = _mm256_loadu_si256((const __m256i*)(B_ + p + 8));
                __m256i c1 = _mm256_loadu_si256((const __m256i*)(C + p));
                __m256i c2 = _mm256_loadu_si256((const __m256i*)(C + p + 8));
                __m256i ma = _mm256_and_si256(_mm256_and_si256(_mm256_cmpgt_epi32(vta, a1), _mm256_cmpgt_epi32(vtb, b1)), _mm256_cmpgt_epi32(vtc, c1));
                __m256i mb = _mm256_and_si256(_mm256_and_si256(_mm256_cmpgt_epi32(vta, a2), _mm256_cmpgt_epi32(vtb, b2)), _mm256_cmpgt_epi32(vtc, c2));
                int mask = _mm256_movemask_ps(_mm256_castsi256_ps(ma)) | (_mm256_movemask_ps(_mm256_castsi256_ps(mb)) << 8);
                cntm += __builtin_popcount(mask);
            }
            for (; p < f; p++) {
                if (A[p] < ta && B_[p] < tb && C[p] < tc) cntm++;
            }
            corr[j] += cntm;
        }
    }
    for (int i = 0; i < n; i++) out[i] = partial[i] + corr[i];
    free(rawp);
}

评测结果