How can I improve this? I want to keep the Union{A,B,C,D} but make it faster. JET says c is captured and also there is runtime dispatch. abstract type T end struct A <: T x::Int end struct B <: T x::Int end struct C <: T x::Int end struct D <: T x::Int end f() = rand((A,B,C,D))(1) function g(n) …

I feel like there are quite a few ways to make this faster — mostly by rewriting completely differently I’m assuming this is a toy example. could you share a few more details about the real use case?

[image] jar1: local c::Int = 0 foreach(1:n) do _ .... c += 1 Reassigning c captured by the do block closure is responsible for the boxing. The annotation does introduce typeasserts to restore type inference immediately after. Wholly hypothetically, I would guess that a typed box w…

Replacing foreach with for does cure the boxing. As for types, the red bit is y::Union{A,B,C,D}. Replacing f() = (rand((A,B,C,D))(1))::Union{A,B,C,D} with f() = (rand((A,B,C))(1))::Union{A,B,C} seems to solve that problem, but if there are 4 things it’s not really a good solution. abstract type…

It’s becuase you weite c += 1 inside the closure body. The rebinding requires c to be boxed. You can instead write function g(n) local c = Ref{Int}(0) foreach(1:n) do _ y = f()::Union{A,B,C,D} if y isa A c[] += 1 elseif y isa B c[] += 2 …

Putting c in a Ref removes the y problem but I don’t understand why. abstract type T end struct A <: T x::Int end struct B <: T x::Int end struct C <: T x::Int end struct D <: T x::Int end f() = (rand((A,B,C,D))(1))::Union{A,B,C,D} function g(n) local c = Ref{Int}(0) foreach(1:n) do _ …

[image] jar1: As for types, the red bit is y::Union{A,B,C,D}. The red there isnt a problem. So long as the union is fully split, the compiler knows what’s going on and there’s no performance penalty or dynamism.

This is almost 20x faster than the code posted by @Mason , but it cheats :smiley: using MixedStructTypes abstract type T end @compact_structs S <: T begin struct A x::Int end struct B x::Int end struct C x::Int end struct D x::Int end end function h(t) if t === :A A(1)…

[image] jar1: Putting c in a Ref removes the type instability on y but I don’t understand why. Because c isn’t being reassigned anymore, you’re just mutating the assigned Ref instance. [image] Mason: The red there isnt a problem. So long as the union is fully split I think the red hin…

wops, I was wrong, also this seems fine, it is only a tiny bit slower than the other implementation I proposed: abstract type T end struct A <: T x::Int end struct B <: T x::Int end struct C <: T x::Int end struct D <: T x::Int end function h(t) if t === :A A(1) elseif t === :B …

Slow code with Union and Box

General Usage

jling March 6, 2024, 1:02am 8

hmm, we (as a community) have been here: Allocation and slow down when # of types involved increase (slower than C++ virtual methods) - #58 by peremato

Topic		Replies	Views
Union splitting vs C++ Performance	22	4437	July 9, 2021
Performance drawback with subtyping Performance	34	3505	August 26, 2021
[ANN]: WrappedUnions.jl: Wrap a Union for enhanced type-stability Package Announcements package , announcement , type , type-stability , union	11	831	March 24, 2026
Speed of multiple dispatch Performance	14	1314	November 27, 2023
JET: are union incurred runtime dispatches really happening? Performance performance , jet	10	382	July 5, 2024

Slow code with Union and Box

Related topics