Optimize the Julia code

[ChrisRackauckas]

The Julia code was very unoptimized so I did a few changes. BTW, why is the a_list passed in as a Vector{Any}?

[ChrisRackauckas]

BTW, welcome to the community. If you need some help getting your code optimized, do check out https://discourse.julialang.org/ and https://julialang.org/slack/ . The #helpdesk and #performance-helpdesk channels in the Slack may be particularly helpful, and feel free to ping if you need anything!

[giordano]

This line doesn't do anything, I think?

[mcabbott]

And hence make_list([]) returns a Vector{Any} which iterate_list won't accept.

[ChrisRackauckas]

It does something if a_list::Vector{Any}, which is why I'm a bit confused by what is passed in.

[giordano]

convert isn't in-place (you can't change the type of a container anyway) and the result isn't assigned to anything.

[ChrisRackauckas]

The code that is calling this isn't part of the repo, otherwise I'd change whatever is doing a_list = [] to a_list = Vector{Float64}() which would presumably solve this.

[00sapo]

convert isn't in-place (you can't change the type of a container anyway) and the result isn't assigned to anything.

Thanks, i didn't know... But with that line iterate_list accepts Vector{Vector{Float64}}, without it only accepts Vector{Any}

BTW, the Vector{Any} is decided by pyjulia+PyCall. It's not the best, but it's so for now... I opened an issue about this:

JuliaPy/pyjulia#441

[giordano]

But with that line iterate_list accepts Vector{Vector{Float64}}, without it only accepts Vector{Any}

As I already said above, convert isn't in-place (and it can't possibly be) and you aren't assigning its result to anything: this line is doing work that isn't effectively used anywhere. It has no practical effects, besides wasting CPU cycles.

[aramirezreyes]

Does this need to be initialized to zero?

[mcabbott]

It does not, but is shorter to write.

If it's legal to allocate the entire array at once, not push! every time, then you could also replace the entire function with something like this:

make_list() = [fill(0.1, 10^4) for _ in 1:10^4];

I guess it would be cheating to use fill(fill(0.1, 10^4), 10^4), although that's much faster!

[00sapo]

I'm going to change the test to avoid such suggestions. The problem is about lists, not arrays.

[mcabbott]

OK. Note that Julia doesn't really have this distinction, the length (but not the type) of any one-dimensional array can be mutated. The memory allocated need not match the length used, so that sometimes push! just writes to the next address, but when full, it re-allocates new memory (IIRC) twice as big. So a more modest change is to hint at the memory required:

               new_list = Float64[]
               sizehint!(new_list, 10^4)

This is roughly twice as fast, and 1/3 the memory.

The first line is just tidying up, as new_list::Vector{Float64} = [] allocates a new Vector{Any}, and then sees that it needs to convert it to Vector{Float64} to attach the name new_list.

f1() = new_list = Float64[];
f2() = new_list::Vector{Float64} = [];
@time f1()  # 1 allocation: 80 bytes
@time f2()  # 2 allocations: 160 bytes