If you spend any time in the Clojure community, I think you'll come to an appreciation that practicality and performance are key principles of its design. The goal is to obtain both high levels of abstraction and performance.
But performance means different things to different people.
If you're writing code where dot products are not something you do a lot of, then keeping a clean, idiomatic functional implementation is probably fine.
If you're writing code where dot products happen a lot, such as on real-time three-dimensional games (which happens to by my oeuvre), then the performance increase of having specific, unrolled implementations is huge.
It depends on what the language and compiler do for you, the elegant form may actually be faster than your supposed uglier-but-faster optimization. As always, time-test before and (if you've decided it's necessary) after optimizing. I remember reading a warning somewhere in the Clojure docs about how using (get x idx) is much faster than (nth x idx), I don't remember why though.
This is of course true; ALWAYS profile the code before and after any optimizations, it is quite common for things not to act the way you'd expect (either because of the hardware, the compiler, or even other bits of code you're not aware of, which perhaps relies on certain assumptions).
In clojure, you can solve this by using a variadic macro (instead of a function) to generate the fast, unrolled version. I think pg had a similar example in On Lisp for generating fast code to find points on a bezier curve.
