StructsOfArrays

A traditional Julia array of immutable objects is an array of structures. Fields of a given object are stored adjacent in memory. However, this often inhibits SIMD optimizations. StructsOfArrays implements the classic structure of arrays optimization. The contents of a given field for all objects is stored linearly in memory, and different fields are stored in different arrays. This permits SIMD optimizations in more cases and can also save a bit of memory if the object contains padding. It is especially useful for arrays of complex numbers.

You can construct a StructOfArrays directly:

using StructsOfArrays
A = StructOfArrays(Complex128, 10, 10)

or by converting an AbstractArray:

A = convert(StructOfArrays, complex(randn(10), randn(10)))

Beyond that, there's not much to say. Assignment and indexing works as with other AbstractArray types. Indexing a StructOfArrays{T} yields an object of type T, and you can assign objects of type T to a given index. The "magic" is in the optimizations that the alternative memory layout allows LLVM to perform.

While you can create a StructOfArrays of non-isbits immutables, this is probably slower than an ordinary array, since a new object must be heap allocated every time the StructOfArrays is indexed. In practice, StructsOfArrays works best with isbits immutables such as Complex{T}.

using StructsOfArrays
regular = complex(randn(1000000), randn(1000000))
soa = convert(StructOfArrays, regular)

function f(x, a)
    s = zero(eltype(x))
    @simd for i in 1:length(x)
        @inbounds s += x[i] * a
    end
    s
end

using Benchmarks
@benchmark f(regular, 0.5+0.5im)
@benchmark f(soa, 0.5+0.5im)

The time for f(regular, 0.5+0.5im) is:

Average elapsed time: 1.244 ms
  95% CI for average: [1.183 ms, 1.305 ms]
Minimum elapsed time: 1.177 ms

and for f(soa, 0.5+0.5im):

Average elapsed time: 832.198 μs
  95% CI for average: [726.349 μs, 938.048 μs]
Minimum elapsed time: 713.730 μs

In this case, StructsOfArrays are about 1.5x faster than ordinary arrays. Inspection of generated code demonstrates that f(soa, a) uses SIMD instructions, while f(regular, a) does not.