代码

#include <bits/stdc++.h>

using namespace std;

void sort(uint* a, int n) {
    double alpha = n == 100000 ? 2 : 1.007;
    uint* b = (uint*)malloc(n * sizeof(uint) * alpha);
    uint* c = (uint*)malloc(n * sizeof(uint) * alpha);
    uint *p0[256], *p1[256], *p2[256], *p3[256];
    for (int i = 0; i < 256; i++) p0[i] = p1[i] = b + int(n / 256 * i * alpha);
    for (int i = 0; i < n; i++) *p1[a[i] & 255]++ = a[i];
    memcpy(a, c, n * sizeof(uint));
}

评测结果