Re: [julia-users] Reductions sometimes not typestable?

[Yichao Yu]

On Sat, Oct 15, 2016 at 6:21 PM, jw3126  wrote:

>
>
> On Sunday, October 16, 2016 at 12:12:14 AM UTC+2, Yichao Yu wrote:
>>
>>
>>
>> 2016-10-15 18:06 GMT-04:00 jw3126 :
>>
>>> myop(::Int16, ::Int16) = Int32(1)
>>> myop(::Int16, ::Int32) = Int64(1)
>>> myop(::Int16, ::Int64) = Int128(1)
>>> myop(::Int16, ::Int128) = Int128(1)
>>>
>>> foldr(myop, Int16[1]) |> typeof |> println
>>> foldr(myop, Int16[1,1]) |> typeof |> println
>>> foldr(myop, Int16[1,1,1]) |> typeof |> println
>>> foldr(myop, Int16[1,1,1,1]) |> typeof |> println
>>>
>>>
>>> gives
>>>
>>>
>>> Int32
>>> Int64
>>> Int128
>>> Int128
>>>
>>> Would it be better if the answer was typestable (always Int128)? See
>>> also here .
>>>
>>
>> Yes it would be better if you implement your operation that way and no
>> the compiler cannot do this.
>>
>
> One could change the promotion in foldr to use something like the
> following:
>
> accumulate_eltype(op, T) = accumulate_eltype(op, T, T)
>

The `promote_op` business is messed up enough. I don't think we need yet
another promotion system to make it harder to write code in most cases.


>
> function accumulate_eltype(op, T, S)
> S_next = promote_op(op, T, S)
> if S == S_next
> return S
> else
> return accumulate_eltype(op, T, S_next)
> end
> end
>
>>
>>
>>

Re: [julia-users] Reductions sometimes not typestable?

[jw3126]

On Sunday, October 16, 2016 at 12:12:14 AM UTC+2, Yichao Yu wrote:
>
>
>
> 2016-10-15 18:06 GMT-04:00 jw3126 >:
>
>> myop(::Int16, ::Int16) = Int32(1)
>> myop(::Int16, ::Int32) = Int64(1)
>> myop(::Int16, ::Int64) = Int128(1)
>> myop(::Int16, ::Int128) = Int128(1)
>>
>> foldr(myop, Int16[1]) |> typeof |> println
>> foldr(myop, Int16[1,1]) |> typeof |> println
>> foldr(myop, Int16[1,1,1]) |> typeof |> println
>> foldr(myop, Int16[1,1,1,1]) |> typeof |> println
>>  
>>
>> gives
>>
>>
>> Int32
>> Int64
>> Int128
>> Int128
>>
>> Would it be better if the answer was typestable (always Int128)? See 
>> also here .
>>
>
> Yes it would be better if you implement your operation that way and no the 
> compiler cannot do this.
>

One could change the promotion in foldr to use something like the following:

accumulate_eltype(op, T) = accumulate_eltype(op, T, T)

function accumulate_eltype(op, T, S)
S_next = promote_op(op, T, S)
if S == S_next
return S
else
return accumulate_eltype(op, T, S_next) 
end
end 

>  
>
>

Re: [julia-users] Reductions sometimes not typestable?

[Yichao Yu]

2016-10-15 18:06 GMT-04:00 jw3126 :

> myop(::Int16, ::Int16) = Int32(1)
> myop(::Int16, ::Int32) = Int64(1)
> myop(::Int16, ::Int64) = Int128(1)
> myop(::Int16, ::Int128) = Int128(1)
>
> foldr(myop, Int16[1]) |> typeof |> println
> foldr(myop, Int16[1,1]) |> typeof |> println
> foldr(myop, Int16[1,1,1]) |> typeof |> println
> foldr(myop, Int16[1,1,1,1]) |> typeof |> println
>
>
> gives
>
>
> Int32
> Int64
> Int128
> Int128
>
> Would it be better if the answer was typestable (always Int128)? See also
> here .
>

Yes it would be better if you implement your operation that way and no the
compiler cannot do this.