til.jpace121/_drafts/powf_is_slow.md

title	author	date
Rust: powf is slow	James Pace	2024/01/18

I'm currently working on my motion planner on a microcontroller project and was looking at switching from libm to micromath as the library I use for standard math operations. For context, in Rust, normal math operations (like absolute value and square root for floats) are provided as part of the standard library (called std). The standard library is not available in environments that don't have a backing operating system. Crates than run in those environments thus can't use std, which is called being no_std. A lot of functions that are in std are actually in two other crates, core or alloc which can be used in no_std environments. The normal math functions aren't one of them, and therefore a different library has to be used.

My intitial implementation of the planner used libm (largely because I found it first), and when I later found micromath I wanted to do some profiling comparisons to see which one was faster. I'm going to make a separate blog post with statistics from the profiling, but I wanted to write a quick article today talking about powf.

One of the operations my planner does a lot is find the distance between two things. My initial implementation did something like

fn distance(point1: &Point, point2: &Point) -> f32 {
     let squared_dist = powf((point2.x - point1.x), 2.0) + powf((point2.y - point1.y), 2.0);
     sqrt(squared_dist)
}

which works fine.

When I was profiling I switched to:

fn distance(point1: &Point, point2: &Point) -> f32 {
     let squared_dist = powf((point2.x - point1.x), 2.0) + powf((point2.y - point1.y), 2.0);
     sqrt(squared_dist)
}

which was noticeably faster for both libm and micromath particularly on the microbit that I was using for profiling.

Does this replicate for std::powf?

Not when building with --release. ¹

How much faster?

---

1. As an aside, I think I'm going to switch to using --release when building rust code all the time. The optimizations help a ton, to the extent I would never "use" software not built with --release and extrapolating performance, even experientially, from binaries built with the dev profile is just not transferrable at all. ↩︎