cat(a, b; dims=1) is slow compared with vcat(a, b)

[strickek]

julia> using BenchmarkTools

julia> a = [1, 2]
2-element Array{Int64,1}:
 1
 2

julia> b = [3, 4]
2-element Array{Int64,1}:
 3
 4

julia> @btime vcat(a, b)
  180.394 ns (1 allocation: 112 bytes)
4-element Array{Int64,1}:
 1
 2
 3
 4

julia> @btime cat(a, b; dims=1)
  44.690 μs (37 allocations: 1.45 KiB)
4-element Array{Int64,1}:
 1
 2
 3
 4

With hcat and dims=2 it is the same story.

[daviehh]

@code_warntype cat(a, b;dims=1) thinks it's not type-stable;

const c=[1,2]
const d=[3,4]
@btime cat(c, d; dims=1)

531.153 ns (9 allocations: 544 bytes)

bring allocations down but still higher than vcat and runs slower.

@which vcat(a,b)

vcat(arrays::Array{T,1}...) where T in Base at array.jl:1531

@which cat(c,d;dims=2)

(::getfield(Base, Symbol("#kw##cat")))(::Any, ::typeof(cat), A...) in Base

[strickek]

Maybe related: #30686

[strickek]

I checked this topic again with Julia v1.6.0-RC2:
Situation improved for cat() from 37 to 25 allocations.

julia> @btime vcat(a, b)
  50.506 ns (1 allocation: 112 bytes)
4-element Vector{Int64}:
 1
 2
 3
 4

julia> @btime cat(a, b; dims=1)
  1.630 μs (25 allocations: 704 bytes)
4-element Vector{Int64}:
 1
 2
 3
 4

faster is dims=Val(1), see: https://discourse.julialang.org/t/very-slow-loop/53608/24

julia> @btime cat(a, b; dims=Val(1))
  1.290 μs (17 allocations: 528 bytes)
4-element Vector{Int64}:
 1
 2
 3
 4

Conclusion: If possible, use vcat() or hcat()

[vtjnash]

At least the Array allocations are faster and smaller now, so it is somewhat

julia> @btime vcat(a, b)
  41.726 ns (2 allocations: 96 bytes)
4-element Vector{Int64}:
 1
 2
 3
 4

julia> @btime cat(a, b; dims=1)
  825.628 ns (12 allocations: 304 bytes)
4-element Vector{Int64}:
 1
 2
 3
 4

julia> @btime cat(a, b; dims=Val(1))
  42.858 ns (2 allocations: 96 bytes)
4-element Vector{Int64}:
 1
 2
 3
 4

Profile suggests it is due to a lot of dynamic dispatch in the dims=1 case (to compute cshp and such), which could be reduced by forcing dispatch early and converting the dims to type Val

julia> dim = 1; @btime cat(a, b; dims=Val(dim))
  120.839 ns (2 allocations: 96 bytes)
4-element Vector{Int64}:
 1
 2
 3
 4