Memory allocations in for loop variable

General Usage Performance
2

This code is an excellent use case for StaticArrays.

Currently, you’re using a 2D matrix to store a list of 3-element points. That’s a very common idiom in Matlab and Numpy, but Julia can do better. If you instead store your vector of points as a Vector of 3-element static vectors, then it becomes easy to manipulate each element without allocating any memory at all.

Benchmarking your code:

julia> using BenchmarkTools

julia> @btime get_orientations!($rotvec, $ori)
  12.987 ms (700000 allocations: 41.20 MiB)

And here’s an alternative implementation using static vectors instead:

julia> using StaticArrays

julia> function orientation(r::SVector{3, Float64})
           R = norm(r)
           q0 = cos(0.5 * R)
           q1, q2, q3 = sin(0.5 * R) * (r / R)
           SVector(2*(q0*q2 + q1*q3), -2*(q0*q1 - q2*q3),
                   q0^2 - q1^2 - q2^2 + q3^2)
       end
orientation (generic function with 1 method)

julia> rotvec = [rand(SVector{3, Float64}) for i in 1:100000];

julia> ori = similar(rotvec);

julia> @btime $ori .= orientation.($rotvec)
  1.830 ms (0 allocations: 0 bytes)

That’s almost 10 times faster, uses less code, and allocates exactly zero memory.

13 Likes