#include <iostream>
#include <fstream>
#include <chrono>

using namespace std;
using namespace std::chrono;

#define SSE_WIDTH	4

#define ALIGNED		__attribute__((aligned(16)))

float
SimdMulSum( float *a, float *b, int len )
{
    float sum[4] = { 0., 0., 0., 0. };
    int limit = ( len/SSE_WIDTH ) * SSE_WIDTH;

    {
        float *atmp = a, *btmp = b;
        register __int128 tmp __asm ("xmm2") = 0;
        for (int i = 0; i < limit; i += SSE_WIDTH) {
            __asm (
                ".att_syntax\n\t"
                "movups	(%0), %%xmm0\n\t"	// load the first sse register
                "movups	(%1), %%xmm1\n\t"	// load the second sse register
                "mulps	%%xmm1, %%xmm0\n\t"	// do the multiply
                "addps	%%xmm0, %2\n\t"	// do the add
                "addq $16, %0\n\t"
                "addq $16, %1\n\t"
                : /* outputs */ "+r" (atmp), "+r" (btmp), "+x" (tmp)
                : /* inputs */ "m" (*atmp), "m" (*btmp)
                : /* clobbers */ "xmm0", "xmm1"
            );
        }
        __asm ("movups %1,%0" : "=m" (sum) : "x" (tmp) : /*no clobbers*/);
    }

    for( int i = limit; i < len; i++ )
    {
        sum[0] += a[i] * b[i];
    }

    return sum[0] + sum[1] + sum[2] + sum[3];
}


int main()
{
    ifstream in_file; // An input file stream variable
    int size = 0; // A number of elements in a signal series
    // Open the file stream
    in_file.open("signal.txt");
    // Check that the file was opened
    if (!in_file)
    {
        cerr << "Unable to open file signal.txt";
        exit(1);   // call system to stop
    }
    in_file >> size;
    cout << "Number of elements: " << size << endl;
    float *A = new float[ 2 * size ];
    float *Sums = new float[ 1 * size ];
    cout << "Reading series...";
    for (int i = 0; i < size; i++)
    {
        float value = 0.0;
        in_file >> value;
        A[i] = value;
        A[i + size] = value;
    }
    cout << "Done." << endl;
    // Serial implementation of the autocorrelation sum
    cout << "Calculating sums using SIMD...";
    high_resolution_clock::time_point t1 = high_resolution_clock::now();
    for( int shift = 0; shift < size; shift++ )
    {
        Sums[shift] = SimdMulSum( &A[0], &A[0+shift], size );
    }
    high_resolution_clock::time_point t2 = high_resolution_clock::now();
    duration<double> time_span = t2 - t1;
    cout << "Done in " << time_span.count() << " seconds." << endl;
    // Close the input stream
    in_file.close();
    delete [] A;
    delete [] Sums;
    return 0;
}