代码

#include <cstdlib>
#include <cstddef>
#include <algorithm>

using uint = unsigned;

const size_t n = 1e8;

void F(uint* buc, uint* a, uint* b, auto foo) {
    auto ptr = (uint**)malloc(256 * sizeof(uint*));
    for (size_t i = 0; i < 256; i++) ptr[i] = b + buc[i];
    uint val[8], **index[8];
    for (size_t i = 0; i < 8; i++) val[i] = a[i], index[i] = &ptr[foo(val[i])];
    for (size_t is = 0; is < n - 8; is += 8) {
        #pragma GCC unroll 8
        for (size_t i = 0; i < 8; i++) {
            *(*index[i])++ = val[i];
            val[i] = a[is + i + 8];
            index[(i + 4) % 8] = &ptr[foo(val[(i + 4) % 8])];
        }
    }
    for (size_t i = 0; i < 8; i++) *ptr[foo(val[i])]++ = val[i];
}
void sort(uint* a, int __n) {
    uint buc[4][256] = {};
    auto b = (uint*)malloc(n * sizeof(uint));
    for (size_t i = 0; i < n; i++) {
        buc[0][a[i] & 255]++;
        buc[1][a[i] >> 8 & 255]++;
        buc[2][a[i] >> 16 & 255]++;
        buc[3][(a[i] >> 16) >> 8 & 255]++;
    }
    for (size_t k = 0; k < 4; k++) {
        uint offset = 0;
        for (size_t i = 0; i < 256; i++)
            std::swap(buc[k][i], offset), offset += buc[k][i];
    }
    F(buc[0], a, b, [](uint v) { return v & 255; });
    F(buc[1], b, a, [](uint v) { return v >> 8 & 255; });
    F(buc[2], a, b, [](uint v) { return v >> 16 & 255; });
    F(buc[3], b, a, [](uint v) { return v >> 24; });
}

评测结果