[julia-users] use of @generated for specializing on functions

[Alireza Nejati]

Hi all,

I was wondering if this is a julian use of the @generated macro:

type Functor{Symbol} end

# A simple general product-sum operator;
# returns a[1]⊙b[1] ⊕ a[2]⊙b[2] ⊕ ...
@generated function dot{⊕,⊙,T}(::Type{Functor{⊕}}, ::Type{Functor{⊙}}, 
a::Array{T}, b::Array{T})
return quote
assert(length(a)==length(b))
p = zero(T)
for i=1:length(a)
@inbounds p=$⊕(p,$⊙(a[i],b[i]))
end
p
end
end

The idea is to produce a specialized dot() operator that can work with 
arbitrary product and sum operators yet still be computationally efficient. 
These are some examples on how to use the above function:

dot(Functor{:+},   Functor{:*}, [-1,0,1],[1,0,1])  # vector dot; returns 0
dot(Functor{:max}, Functor{:+}, [-1,0,1],[1,0,1])  # max-sum; returns 2
dot(Functor{:|},   Functor{:&}, [true,false,true],[false,true,true])  # 
constraint satisfaction; returns true

Testing shows that this is faster by about 10x than passing the functions 
directly and runs at the same speed . It also doesn't go through any memory 
allocations.

Re: [julia-users] use of @generated for specializing on functions

[Yichao Yu]

On Mon, Oct 26, 2015 at 4:27 PM, Alireza Nejati  wrote:
> Hi all,
>
> I was wondering if this is a julian use of the @generated macro:
>
> type Functor{Symbol} end
>
> # A simple general product-sum operator;
> # returns a[1]⊙b[1] ⊕ a[2]⊙b[2] ⊕ ...
> @generated function dot{⊕,⊙,T}(::Type{Functor{⊕}}, ::Type{Functor{⊙}},
> a::Array{T}, b::Array{T})
> return quote
> assert(length(a)==length(b))
> p = zero(T)
> for i=1:length(a)
> @inbounds p=$⊕(p,$⊙(a[i],b[i]))
> end
> p
> end
> end
>
> The idea is to produce a specialized dot() operator that can work with
> arbitrary product and sum operators yet still be computationally efficient.
> These are some examples on how to use the above function:
>
> dot(Functor{:+},   Functor{:*}, [-1,0,1],[1,0,1])  # vector dot; returns 0
> dot(Functor{:max}, Functor{:+}, [-1,0,1],[1,0,1])  # max-sum; returns 2
> dot(Functor{:|},   Functor{:&}, [true,false,true],[false,true,true])  #
> constraint satisfaction; returns true
>
> Testing shows that this is faster by about 10x than passing the functions
> directly and runs at the same speed . It also doesn't go through any memory
> allocations.

As a temporary personal solution before Jeff fix the actual problem,
maybe ,in general, no.

See my comment 
https://github.com/JuliaLang/julia/issues/12357#issuecomment-125791661
and https://github.com/JuliaLang/julia/pull/12322#issuecomment-126810815

These cannot be a generic API since you can't get the scope right and
it would be really confusing when it will work and when it will not.