Soapy

Soapy makes it simple to work with structure-of-arrays memory layout. What Vec<T> is to array-of-structures (AoS), Soa<T> is to structure-of-arrays (SoA).

Example

 use soapy::{Soa, Soapy};
[derive(Soapy, Debug, Clone, Copy, PartialEq)]
struct Example {
    foo: u8,
    bar: u16,
}

let elements = [Example { foo: 1, bar: 2 }, Example { foo: 3, bar: 4 }];
let mut soa: Soa<_> = elements.into_iter().collect();

// The index operator is not possible, but we can use nth:
*soa.nth_mut(0).foo += 10;

// We can get the fields as slices as well:
let slices = soa.slices();
assert_eq!(slices.foo, &[11, 3][..]);
assert_eq!(slices.bar, &[2, 4][..]);

for (actual, expected) in soa.iter().zip(elements.iter()) {
    assert_eq!(&expected.bar, actual.bar);
}

What is SoA?

The following types illustrate the difference between AoS and Soa:

[(u8,   u64)] // AoS
([u8], [u64]) // Soa

Whereas AoS stores all the fields of a type in each element of the array, SoA splits each field into its own array. This has several benefits:

• There is no padding required between instances of the same type. In the above example, each AoS element requires 128 bits to satisfy memory alignment requirements, whereas each SoA element only takes 72. This can mean better cache locality and lower memory usage.
• SoA can be more amenable to vectorization. With SoA, multiple values can be direcly loaded into SIMD registers in bulk, as opposed to shuffling struct fields into and out of different SIMD registers.

SoA is a popular technique in data-oriented design. Andrew Kelley gives a wonderful talk describing how SoA and other data-oriented design patterns earned him a 39% reduction in wall clock time in the Zig compiler.

Note that SoA does not offer performance wins in all cases. SoA is most appropriate when either

• Sequential access is the common access pattern
• You are frequently accessing or modifying only a subset of the fields

As always, it is best to profile both for your use case.

Derive

Soapy provides the [Soapy] derive macro to generate SoA compatibility for structs automatically. When deriving Soapy, several new structs are created. Because of the way SoA data is stored, iterators and getters often yield these types instead of the original struct. If each field of some struct Example has type F, our new structs have the same fields but different types:

These types are included as associated types on the [Soapy] trait as well. Generally, you won't need to think about these them as [Soa] picks them up automatically. However, since they inherit the visibility of the derived struct, you should consider whether to include them in the pub items of your module.

Comparison

soa_derive

soa_derive makes each field its own Vec. Because of this, each field's length, capacity, and allocation are managed separately. In contrast, Soapy manages a single allocation for each Soa. This uses less space and allows the collection to grow and shrink more efficiently. soa_derive also generates a new collection type for every struct, whereas Soapy generates a minimal, low-level interface and uses the generic Soa type for the majority of the implementation. This provides more type system flexibility, less code generation, and more accessible documentation.

soa-vec

Whereas soa-vec only compiles on nightly, Soapy also compiles on stable. Rather than using derive macros, soa-vec instead uses macros to generate eight static copies of their SoA type with fixed tuple sizes.

Progress

Soa

• depup / dedup_by / dedup_by_key
• drain
• extend_from_slice / extend_from_within
• extract_if
• leak
• retain
• try_reserve / try_reserve_exact
• dedup_by / dedup_by_key
• resize / resize_with
• splice
• split_off

SoaSlice

• select_nth_unstable / select_nth_unstable_by / select_nth_unstable_by_key
• sort / sort_by / sort_by_key / sort_by_cached_key
• sort_unstable / sort_unstable_by / sort_unstable_by_key / sort_unstable_by_cached_key
• binary_search / binary_search_by / binary_search_by_key
• is_sorted / is_sorted_by / is_sorted_by_key
• chunks / rchunks
• chunks_exact / rchunks_exact
• first / last
• rotate_left / rotate_right
• split / rsplit / splitn
• split_at / split_first / split_last
• swap
• swap_with_slice
• group_by
• contains
• copy_within
• fill / fill_with
• repeat
• reverse