MArray is slower than Array in reduction

[YingboMa]

MWE:

julia> using BenchmarkTools

julia> a = @MArray rand(4);

julia> @btime sum( $(Array(a)) )
  3.896 ns (0 allocations: 0 bytes)
1.5713808992800153

julia> @btime sum( $a )
  30.021 ns (1 allocation: 16 bytes)
1.5713808992800153

[c42f]

This is still the case on the latest julia. Ultimately it may be due to keyword functions (in this case mapreduce) not getting inlined.

[tkoolen]

Yeah, in some of the Base reduce/mapreduce code, NamedTuples are sometimes passed around as positional argument instead of using keyword arguments, probably for this reason.

https://github.com/JuliaLang/julia/blob/cce5a6b3d7303a4ca74a7f630b452d29b06e5903/base/reduce.jl#L39

[mateuszbaran]

Unfortunately, passing keyword arguments around is relatively slow in Julia (it requires an invoke). The easiest solution is to define a bunch of methods like

@inline sum(a::StaticArray{<:Tuple,T}) where {T} = reduce(+, a)

to avoid passing keyword arguments. I can make a patch if you want.

[c42f]

I can make a patch if you want.

Yes please!

Did you read the code @tkoolen pointed to above? It seemed to me that might be a way to avoid some of the cost.

[mateuszbaran]

Did you read the code @tkoolen pointed to above? It seemed to me that might be a way to avoid some of the cost.

Yes, I'm trying to invent a less tedious solution.

[c42f]

Fixed in #659 (thanks to @mateuszbaran :-) )