Question

I am multiplying constant vector<bool> on different vector<double> many times. I wonder how fast is that, wouldn’t it be faster to convert it first to vector<double>, so that sse can be used?

    void applyMask(std::vector<double>& frame, const std::vector<bool>& mask)
    {
        std::transform(frame.begin(), frame.end(), mask.begin(), frame.begin(), [](const double& x, const bool& m)->double{ return x*m;});
    }

Answer 1

It seems like you’re trying to zero parts of a vector<double> using a mask of vector<bool>.

As it stands right now, it’s not vectorizable. Furthermore, the vector<bool> template specialization is going to hinder the compiler’s ability to do any sort of auto-vectorization.

So you basically have two options:

The easy way is to indeed convert the vector<bool> to a vector<double> of corresponding zeros and ones. Then the problem reduces to simply vector-to-vector multiplication of the same datatype, which is completely vectorizable. (even auto-vectorizable)

The harder way (which might be faster), is to play some hacks with the _mm_and_pd or _mm_blendv_pd() intrinsics/instructions. But that requires a lot more work since you have to manually vectorize the code.

---

I suggest you just go with the easy way. No need to dive into manual vectorization unless you really need to.