LazyBroadcast.jl

LazyBroadcast.jl provides a macro, LazyBroadcast.@lazy_broadcasted to transform a given Julia broadcast expression into a Base.Broadcast.Broadcasted object, without materializing it.

For more information about Julia broadcasting, please see https://docs.julialang.org/en/v1/manual/arrays/#Broadcasting.

This utility is useful in a few situations:

• Debugging broadcast machinery
• Fusing operations in multiple broadcast expressions (e.g., see MultiBroadcastFusion.jl, which has lots of overlapping infrastructure)
• Delaying execution of a broadcast expression

For not-in-place expressions, @lazy_broadcasted simply returns the broadcasted object, via Base.Broadcast.broadcasted of the right-hand-side:

using Test
import LazyBroadcast: @lazy_broadcasted

a = rand(3,3)
b = rand(3,3)

@testset "lazy_broadcasted" begin
    bc = @lazy_broadcasted @. a + b # get the broadcasted object
    @test Base.broadcasted(+, a, b) == @lazy_broadcasted @. a + b
    @test Base.Broadcast.materialize(bc) == @. a + b # materialize the broadcasted object
end

For in-place expressions, @lazy_broadcasted will strip out the left-hand-side, and still return return the broadcasted object, via Base.Broadcast.broadcasted, of the right-hand-side:

import LazyBroadcast: @lazy_broadcasted

a = rand(3,3)
b = rand(3,3)
L = rand(3,3)

bc = @lazy_broadcasted @. L = a + b # get the broadcasted object for
@test Base.broadcasted(+, a, b) == @lazy_broadcasted @. L = a + b
@test Base.Broadcast.materialize!(L, bc) == @. L = a + b # materialize the broadcasted object

To eagerly execute expressions, you may simply immediately materialize the returned broadcasted object:

import LazyBroadcast: @lazy_broadcasted

a = rand(3,3)
b = rand(3,3)
L = rand(3,3)

@. L = a + b

# is exactly equivalent to

bc = @lazy_broadcasted @. L = a + b
Base.Broadcast.materialize!(L, bc)