ShortVectors.jl

This package provides the type ShortVector, which stores short vectors locally (i.e. on the stack) and reverts to heap storage for longer vectors. This can result in significant performance gains when working with lots of short vectors, while retaining the flexbility to accommodate longer vectors when necessary. This package is inspired by Rust's smallvec type and a related Julia Discourse discussion..

This is currently an experimental package: only the most basic functionality is provided, and little effort (so far) has been made to test and improve performance. This package may eventually merge with similar efforts such as SmallCollections.jl and/or ImmutableVectors.jl, which (as of 11/15/2024) do not allow for vectors longer than some chosen upper bound.

Usage

struct ShortVector{N,T} <: DenseVector{T}

Immutable vector with local storage for up to N elements of type T. If there are more than N elements they are stored on the heap. For best performance N should be large enough to accommodate the majority of likely vectors, but smaller than the size at which tuples yield poor performance (say, 10). If the element type T is not specified at construction it is inferred.

> using ShortVectors
> x = ShortVector{4}(10:10:30)		# stored locally
[10, 20, 30]ˢᵛ

> length(x)
3

x[2]
20

> y = ShortVector{4}(10:10:50)		 # stored on the heap
[10, 20, 30, 40, 50]ˢᵛ

> y[2:4]
[20, 30, 40]ˢᵛ 

Under the Hood

ShortVector is wrapper for what is essentially a Union{NTuple{N,T}, Vector{T}}.

Open Issues

ShortVector is not performing as well as I had hoped; while its performance is better than Vector in some cases, it still triggers allocations sometimes. There is some indication that the internal Union may not be inlined.