#include <iostream>

#include <cstdlib>
#include <cstring>
#include <ctime>

#include <x86intrin.h>

typedef union
{
    __m128 a;
    float b[4];
} float4;

int rand_int(int limit)
{
    return rand() % limit;
}

int main(int argc, char *argv[])
{
    if(argc < 2)
    {
        std::cout << "Missing parameter" << std::endl;
        return 1;
    }

    if(std::atoi(argv[1]) == 0)
    {
        std::cout << "Invalid parameter" << std::endl;
        return 1;
    }

    int vector_size = std::atoi(argv[1]);

    // Pad vector to bit width divisible by 128
    vector_size += (vector_size % 4 == 0) ? vector_size : 4 - (vector_size % 4);

    float* a = new float[vector_size];
    float* b = new float[vector_size];
    float* r_std = new float[vector_size];
    float* r_sse = new float[vector_size];

    srand(time(NULL));

    // Fill vectors with random ints
    for(int i = 0; i < vector_size; i++)
    {
        a[i] = rand_int(5);
        b[i] = rand_int(5);
    }

    // Print input arrays
    std::cout << "a    : ";
    for(int i = 0; i < vector_size; i++)
    {
        std::cout << a[i] << ", ";
    }
    std::cout << std::endl;

    std::cout << "b    : ";
    for(int i = 0; i < vector_size; i++)
    {
        std::cout << b[i] << ", ";
    }
    std::cout << std::endl;

    // Add both vectors normal way
    for(int i = 0; i < vector_size; i++)
    {
        r_std[i] = a[i] + b[i];
    }

    // Add vectors with sse instructions
    float4 c;
    float4 d;
    float4 r;
    for(int i = 0; i < vector_size; i+=4)
    {
        for(int j = 0; j < 4; j++)
        {
            c.b[j] = a[i+j];
            d.b[j] = b[i+j];
        }

        //SSE magic
        r.a = _mm_add_ps(c.a, d.a);

        for(int j = 0; j < 4; j++)
        {
            r_sse[i+j] = r.b[j];
        }
    }

    std::cout << std::endl;
    std::cout << std::endl;

    // Print output arrays
    std::cout << "r_std: ";
    for(int i = 0; i < vector_size; i++)
    {
        std::cout << r_std[i] << ", ";
    }
    std::cout << std::endl;

    std::cout << "r_sse: ";
    for(int i = 0; i < vector_size; i++)
    {
        std::cout << r_sse[i] << ", ";
    }
    std::cout << std::endl;
}