Hi,

I am writing a function that adds two stl vectors of floats with each other as fast as possible. What I currently have is as follows:

void AvxSum(vector<float> &a, vector<float> &b)
{
    count = 8
    int num = 0;
    int bound = a.size();
    for (int num1 = 0; num1 < bound; num1  = count)
    {
        __m256 vec1 = _mm256_setr_ps(a[num], a[num 1], a[num   2], a[num   3], a[num   4], a[num   5], a[num   6], a[num   7]);
        __m256 vec2 = _mm256_setr_ps(b[num], b[num   1], b[num   2], b[num   3], b[num   4], b[num   5], b[num   6], b[num   7]);
        __m256 res = _mm256_add_ps(vec1, vec2);
        float* f = (float*)&res;
        a[num] = f[num]; a[num   1] = f[num   1];
        a[num   2] = f[num   2]; a[num   3] = f[num   3];
        a[num   4] = f[num   4]; a[num   5] = f[num   5];
        a[num   6] = f[num   6]; a[num   7] = f[num   7];
    }
}

Making each vectorized addition parallel (with an OMP pragma) makes this about 10x slower (I think because there are so many small operations). Maybe we could split large vectors into fairly large chunks and do this in parallel, though.

However, I am unsure of how to further optimize this inner loop.