void printBuff(__m128i buff)
{
    unsigned char tmp[16];
    _mm_store_si128((__m128i*)tmp, buff);
    for(int i = 0; i < 16; i++)
    {
        std::cout << (int)tmp[i] << ", ";
    }
    std::cout << std::endl;
}

void printBuff256(__m256i buff)
{
    alignas(32) unsigned char tmp[32];
    _mm256_storeu_si256((__m256i*)tmp, buff);
    for(int i = 0; i < 32; i++)
    {
        std::cout << (int)tmp[i] << ", ";
    }
    std::cout << std::endl;
}