代码

void sort(unsigned *a, int n)
{
    unsigned b[200000000],*rs[256], t0[256], t1[256], t2[256], t3[256];
    unsigned *p;
    for (int i = 0; i < 256; ++i)
    {
        t0[i] = 0;
        t1[i] = 0;
        t2[i] = 0;
        t3[i] = 0;
    }
    for (int i = 0; i < n;i+=4)
    {
        ++t0[a[i] & 255];
        ++t1[a[i] >> 8 & 255];
        ++t2[a[i] >> 16 & 255];
        ++t3[a[i] >> 24];
        ++t0[a[i+1] & 255];
        ++t1[a[i+1] >> 8 & 255];
        ++t2[a[i+1] >> 16 & 255];
        ++t3[a[i+1] >> 24];
        ++t0[a[i+2] & 255];
        ++t1[a[i+2] >> 8 & 255];
        ++t2[a[i+2] >> 16 & 255];
        ++t3[a[i+2] >> 24];
        ++t0[a[i+3] & 255];
        ++t1[a[i+3] >> 8 & 255];
        ++t2[a[i+3] >> 16 & 255];
        ++t3[a[i+3] >> 24];
    }
    p = b - 1;
    for (int i = 0; i < 256; ++i)
        rs[i] = p, p += t0[i];
    for (int i = 0; i < n; i += 8)
    {
        p = a + i;
        *++rs[p[0] & 255] = p[0];
        *++rs[p[1] & 255] = p[1];
        *++rs[p[2] & 255] = p[2];
        *++rs[p[3] & 255] = p[3];
        *++rs[p[4] & 255] = p[4];
        *++rs[p[5] & 255] = p[5];
        *++rs[p[6] & 255] = p[6];
        *++rs[p[7] & 255] = p[7];
    };
    p = a - 1;
    for (int i = 0; i < 256; ++i)
        rs[i] = p, p += t1[i];
    for (int i = 0; i < n; i += 8)
    {
        p = b + i;
        *++rs[p[0] >> 8 & 255] = p[0];
        *++rs[p[1] >> 8 & 255] = p[1];
        *++rs[p[2] >> 8 & 255] = p[2];
        *++rs[p[3] >> 8 & 255] = p[3];
        *++rs[p[4] >> 8 & 255] = p[4];
        *++rs[p[5] >> 8 & 255] = p[5];
        *++rs[p[6] >> 8 & 255] = p[6];
        *++rs[p[7] >> 8 & 255] = p[7];
    };
    p = b - 1;
    for (int i = 0; i < 256; ++i)
        rs[i] = p, p += t2[i];
    for (int i = 0; i < n; i += 8)
    {
        p = a + i;
        *++rs[p[0] >> 16 & 255] = p[0];
        *++rs[p[1] >> 16 & 255] = p[1];
        *++rs[p[2] >> 16 & 255] = p[2];
        *++rs[p[3] >> 16 & 255] = p[3];
        *++rs[p[4] >> 16 & 255] = p[4];
        *++rs[p[5] >> 16 & 255] = p[5];
        *++rs[p[6] >> 16 & 255] = p[6];
        *++rs[p[7] >> 16 & 255] = p[7];
    };
    p = a - 1;
    for (int i = 0; i < 256; ++i)
        rs[i] = p, p += t3[i];
    for (int i = 0; i < n; i += 8)
    {
        p = b + i;
        *++rs[p[0] >> 24] = p[0];
        *++rs[p[1] >> 24] = p[1];
        *++rs[p[2] >> 24] = p[2];
        *++rs[p[3] >> 24] = p[3];
        *++rs[p[4] >> 24] = p[4];
        *++rs[p[5] >> 24] = p[5];
        *++rs[p[6] >> 24] = p[6];
        *++rs[p[7] >> 24] = p[7];
    };
}

评测结果