The Julia manual nicely makes the case for pre-allocating outputs:
https://docs.julialang.org/en/latest/manual/performance-tips/#Pre-allocating-outputs-1
Yet the function diff that is used extensively in scientific computing seems not to have a corresponding diff!. I searched and could not find one so I wrote one myself (with some extra features like periodic boundary conditions):
https://github.com/JeffFessler/MIRT.jl/blob/master/src/regularize/diffl.jl
Would some version of diff! be appropriate for the standard library, or is it better left to external packages? (And did I reinvent the wheel?)
1 Like
You might find this package of interest
2 Likes
Thanks so much for the pointer to that package!
I updated my version to use @inbounds and then added a comparison and found that StaticKernels and my “diff specific” code had essentially identical compute times. That is a testament to the efficiency of the more general approach in StaticKernels
https://github.com/JeffFessler/MIRT.jl/blob/master/time/diffl.jl
My advice to anyone who comes here looking for diff! is to use StaticKernels but be sure to use map! not map.
Here’s an equivalent of diff!(out, in, dims=2) for a 2D input in:
@inline kf(w) = @inbounds w[0,0] - w[0,-1] # for dims=2
k = Kernel{(0:0,-1:0)}(kf)
map!(k, out, extend(in, StaticKernels.ExtensionReplicate()))
3 Likes