gistfile1.c

#include <pmmintrin.h>
double ddotp2(double *u,
              double *v,
              int n)
{
    int i;
    __m128d x = _mm_setzero_pd();

    for (i=0; i<n-1; i+=2) {
        x = _mm_add_pd(x, _mm_mul_pd(_mm_loadu_pd(&u[i]),
                                     _mm_loadu_pd(&v[i])));
    }

    double ret = _mm_cvtsd_f64(_mm_hadd_pd(x,x));

    if (n & 1) {
        return ret + (u[n-1] * v[n-1]);
    } else {
        return ret;
    }
}