CODEWITHSUNDEEP
juliavectorization
MikeRand
MikeRand

Reputation: 4828

Julia dot vectorization - how deep into call chain do I need to vectorize?

I'm trying to grok Julia's dot vectorization syntax with a very simple example of a linear interpolating spline.

function createlookup(xk)
    min_i = 1
    max_i = size(xk, 1) - 1
    function inner(x)
        max(min(searchsorted(xk, x).stop, max_i), min_i)
    end
    inner
end


function linear_dydx(x, y)
    diff(y, dims=1) ./ diff(x, dims=1)
end


function linear(xₖ, yₖ)
    segment = createlookup(xₖ)
    dydxₖ = linear_dydx(xₖ, yₖ)
    function inner(x)
        i = segment(x)
        y = yₖ[i] + dydxₖ[i] * (x - xₖ[i])
    end
    inner
end


function main()
    xₖ = [0.0, 1.0, 2.0, 4.0]
    yₖ = [0.0, 2.0, 10.0, 13.0]
    interpolator = linear(xₖ, yₖ)

    x = [i/2.0 for i=0:10]

    println(interpolator(2.43))
    println(interpolator.(x))
end

main()

How deep in the call chain do I need to use the dot notation to make sure my main() call to interpolator.(x) is fully vectorized by the interpreter?

I find myself throwing dots everywhere and nowhere and am not quite sure how to check whether I'm capturing everything while still preserving the broadcasting flexibility that dot notation is supposed to preserve.

Upvotes: 2

Views: 172

Answers (1)

Oscar Smith
Oscar Smith

Reputation: 6378

The general strategy to keep in mind is to write all your functions as scalar functions if that makes sense. Once you've written your scalar operations, you can then apply them to the array by using . once.

Upvotes: 3

Related Questions