Layout: Optimized memory layout using struct of array, Data-oriented design in Rust.

This crate is a hard fork of soa-derive with no_std support and additional features.

This crate provides a custom derive (#[derive(SOA)]) to automatically generate code from a given struct T that allow to replace Vec<T> with a struct of arrays. For example, the following code

#[derive(SOA)]
pub struct Cheese {
    pub smell: f64,
    pub color: (f64, f64, f64),
    pub with_mushrooms: bool,
    pub name: String,
}

will generate a CheeseVec struct that looks like this:

pub struct CheeseVec {
    pub smell: Vec<f64>,
    pub color: Vec<(f64, f64, f64)>,
    pub with_mushrooms: Vec<bool>,
    pub name: Vec<String>,
}

It will also generate the same functions that a Vec<Cheese> would have, and a few helper structs: CheeseSlice, CheeseSliceMut, CheeseRef and CheeseRefMut corresponding respectivly to &[Cheese], &mut [Cheese], &Cheese and &mut Cheese.

Any struct derived by SOA will auto impl trait SOA. You can use <Cheese as SOA>::Type instead of the explicitly named type CheeseVec.

How to use it

Add #[derive(SOA)] to each struct you want to derive a struct of array version. If you need the helper structs to derive additional traits (such as Debug or PartialEq), you can add an attribute #[layout(Debug, PartialEq)] to the struct declaration.

#[derive(Debug, PartialEq, SOA)]
#[layout(Debug, PartialEq)]
pub struct Cheese {
    pub smell: f64,
    pub color: (f64, f64, f64),
    pub with_mushrooms: bool,
    pub name: String,
}

If you want to add attribute to a specific generated struct(such as #[cfg_attr(test, derive(PartialEq))] on CheeseVec), you can add an attribute #[soa_attr(Vec, cfg_attr(test, derive(PartialEq)))] to the struct declaration.

#[derive(Debug, PartialEq, SOA)]
#[soa_attr(Vec, cfg_attr(test, derive(PartialEq)))]
pub struct Cheese {
    pub smell: f64,
    pub color: (f64, f64, f64),
    pub with_mushrooms: bool,
    pub name: String,
}

Mappings for first argument of soa_attr to the generated struct for Cheese:

• Vec => CheeseVec
• Slice => CheeseSlice
• SliceMut => CheeseSliceMut
• Ref => CheeseRef
• RefMut => CheeseRefMut
• Ptr => CheesePtr
• PtrMut => CheesePtrMut

Usage and API

All the generated code have some generated documentation with it, so you should be able to use cargo doc on your crate and see the documentation for all the generated structs and functions. Most of the time, you should be able to replace Vec<Cheese> by CheeseVec, with exception of code using direct indexing in the vector and a few other caveats listed below.

Caveats and limitations

Vec<T> functionalities rely a lot on references and automatic deref feature, for getting function from [T] and indexing. But the SoA vector (let's call it CheeseVec, generated from the Cheese struct) generated by this crate can not implement Deref<Target=CheeseSlice>, because Deref is required to return a reference, and CheeseSlice is not a reference. The same applies to Index and IndexMut trait, that can not return CheeseRef/CheeseRefMut. This means that the we can not index into a CheeseVec, and that a few functions are duplicated, or require a call to as_ref()/as_mut() to change the type used.

Iteration

It is possible to iterate over the values in a CheeseVec

let mut vec = CheeseVec::new();
vec.push(Cheese::new("stilton"));
vec.push(Cheese::new("brie"));

for cheese in vec.iter() {
    // when iterating over a CheeseVec, we load all members from memory
    // in a CheeseRef
    let typeof_cheese: CheeseRef = cheese;
    println!("this is {}, with a smell power of {}", cheese.name, cheese.smell);
}

Note that using iter() is nearly as fast as manually selecting fields (see below), as LLVM optimizes away unused field references in release builds.

One of the main advantage of the SoA layout is to be able to only load some fields from memory when iterating over the vector. In order to do so, one can manually pick the needed fields:

for name in &vec.name {
    // We get referenes to the names
    let typeof_name: &String = name;
    println!("got cheese {}", name);
}

In order to iterate over multiple fields at the same time, one can use the soa_zip! macro.

for (name, smell, color) in soa_zip!(vec, [name, mut smell, color]) {
    println!("this is {}, with color {:#?}", name, color);
    // smell is a mutable reference
    *smell += 1.0;
}

Nested Struct of Arrays

In order to nest a struct of arrays inside another struct of arrays, one can use the #[nested_soa] attribute.

For example, the following code

#[derive(SOA)]
pub struct Point {
    x: f32,
    y: f32,
}
#[derive(SOA)]
pub struct Particle {
    #[nested_soa]
    point: Point,
    mass: f32,
}

will generate structs that looks like this:

pub struct PointVec {
    x: Vec<f32>,
    y: Vec<f32>,
}
pub struct ParticleVec {
    point: PointVec, // rather than Vec<Point>
    mass: Vec<f32>
}

All helper structs will be also nested, for example PointSlice will be nested in ParticleSlice.

Benchmarks

The benchmarks compare two different data layouts:

• AoS (Array of Structures): A simple Rust Vec<Structure> where each element is a complete structure.
• SoA (Structure of Arrays): Uses the struct-of-arrays layout provided by this crate, where each field is stored in a separate contiguous array.

Results: The SoA implementation shows up to 3x faster performance for read operations compared to the traditional AoS approach.

test aos_big_do_work_100k        ... bench:     161,151 ns/iter (+/- 57,573)
test aos_big_do_work_10k         ... bench:       6,979 ns/iter (+/- 158)
test aos_big_push                ... bench:          58 ns/iter (+/- 27)
test aos_small_do_work_100k      ... bench:      66,672 ns/iter (+/- 599)
test aos_small_push              ... bench:          16 ns/iter (+/- 7)
test soa_big_do_work_100k        ... bench:      69,611 ns/iter (+/- 2,165)
test soa_big_do_work_10k         ... bench:       6,708 ns/iter (+/- 117)
test soa_big_do_work_simple_100k ... bench:      76,656 ns/iter (+/- 1,675)
test soa_big_push                ... bench:          42 ns/iter (+/- 4)
test soa_small_do_work_100k      ... bench:      66,586 ns/iter (+/- 1,238)
test soa_small_push              ... bench:           6 ns/iter (+/- 3)

Benchmarks tests exist for soa (struct of array) and aos (array of struct) versions of the same code, using a small (24 bytes) and a big (240 bytes) struct.

You can run the same benchmarks on your own system by cloning this repository and running cargo bench.

Licensing and contributions

This crate distributed under either the MIT or the Apache license, at your choice. Contributions are welcome, please open an issue before to discuss your changes !

Thanks to @maikklein for the initial idea: https://maikklein.github.io/soa-rust/