You don't get SIMD with this approach though, so it's about a quarter of the speed of using vectorized arithmetic (assuming an AVX vector of doubles).