floating-distance

Crates.iofloating-distance
lib.rsfloating-distance
version0.3.1
sourcesrc
created_at2023-09-09 17:19:55.404348
updated_at2023-10-02 04:59:57.966005
descriptionMeasure distance between floating-point vectors in Rust
homepage
repositoryhttps://github.com/AsherJingkongChen/floating-distance.git
max_upload_size
id968260
size42,308
ChenYiran (AsherJingkongChen)

documentation

https://docs.rs/floating-distance

README

Measure distance between floating-point vectors in Rust

crates.io docs.rs coverage GitHub Actions

Hello Rustaceans!

This is a Rust library. Here you can do:

  1. Add this crate as a dependency of your Rust project:
cargo add floating-distance
  1. Check out the documentation and source codes (click the badges above for more information)
  2. Clone the GitHub repository and run the examples:
cargo run --example default

Examples

  1. Measure the cosine similarity between two vectors v0 and v1
use floating_distance::*;

fn main() {
  let v0: &[f32] = &[1.0, 2.0, 2.0, 1.0, 2.0, 1.0, 1.0];
  let v1: &[f32] = &[2.0, 1.0, 1.0, 1.0, 2.0, 1.0, 2.0];
  let metric = Metric::Cosine;
  let result = metric.measure::<f32>(v0, v1);
  let expectation: f32 = 14.0 / (4.0 * 4.0);

  assert_eq!(result, expectation);
}

feature = ["simd"]

What's special about SIMD?

SIMD is the acronym for Single Instruction Multiple Data

Modern CPUs have special instructions. We can use them to accelerate vector computations!

How to enable SIMD?

You can enable simd feature in this crate by the following steps:

  1. Specify features = ["simd"] in Cargo.toml manifest:
[dependencies]
floating-distance = { version = "*", features = ["simd"] }
  1. Compile with Rust nightly version. You can add this to rust-toolchain.toml, which is next to Cargo.toml:
[toolchain]
channel = "nightly"
  1. (Optional) Choose the SIMD instruction sets which are supported by the target architecture. You can use RUSTFLAGS environment variable and -C target-feature compiler option like these:
RUSTFLAGS="-C target-feature=+ssse3" cargo build
RUSTFLAGS="-C target-feature=+avx,+sse3" cargo build --release

You can find all target features of Rust by this:

rustc --print target-features

How powerful is SIMD?

I have run a simple benchmark on my laptop (architecture: aarch64-apple-darwin). The SIMD instruction set is NEON.

Let's check out the results first!

  • SIMD vs No SIMD (single-precision floating-point type):
no_simd: 198,306 ns/iter (+/- 5,173)
simd:    18,430 ns/iter (+/- 655)
Type Avarage time (ns/iter) Rate (relative)
No SIMD 198306 1.00
SIMD 18430 10.76

As the data shown, we can see that SIMD can improve the performance significantly!

You can also benchmark it by repeating the following steps:

  1. Clone the repository and change it to the current directory
  2. Check target features in .cargo/config.toml
  3. Run this command:
(cargo +nightly bench -p benchmarks-no-simd &&
 cargo +nightly bench -p benchmarks-simd) 2> /dev/null

Note for SIMD feature

  1. This feature is built by experimental features of Rust standard library, portable-simd.
  2. If a program is built with target features which are not supported by the target architecture, it may lead to runtime errors.
Commit count: 34

cargo fmt