Julia: inference failure for Float64 constructor?

Created on 23 Jul 2020 · 5Comments · Source: JuliaLang/julia

julia> x = rand(Float16, 10_000);

julia> @btime sum($Float64, $x);
  776.911 μs (29999 allocations: 468.73 KiB)

julia> @btime sum($(y -> Float64(y)), $x);
  42.417 μs (0 allocations: 0 bytes)

performance

Source

stevengj

Most helpful comment

Actually it's a specialization issue for sum not BenchmarkTools this time. @benchmark sum(Float64, $x) gives the same result this time.

yuyichao on 23 Jul 2020

👍2

All 5 comments

https://github.com/JuliaCI/BenchmarkTools.jl/issues/71

Most likely a benchmarktools bug.

yuyichao on 23 Jul 2020

Actually it's a specialization issue for sum not BenchmarkTools this time. @benchmark sum(Float64, $x) gives the same result this time.

yuyichao on 23 Jul 2020

👍2

Yes, it's not specializing on the function argument when it's a Type. Maybe we should disable the heuristic for Types when the argument is called in the body?

JeffBezanson on 24 Jul 2020

👍1

Agreed, seems like called should apply to anything we heuristically limit.

vtjnash on 24 Jul 2020

Thanks for this. I never really understood the motivation for not specializing on Type variables, as they are usually important to infer concrete types further down the road. This fix just reduced inference time of some example code in our group from 200 seconds down to the normal inference time when you hard code the specific type.