On Sunday, October 11, 2015 04:12:51 AM cheng wang wrote:
> Hello everyone,
>
> I want to do an in-place tanh: y = tanh(x) assuming that y and x are
> preallocated (both are Array{Float64, 1024*1024}).
> I tested different implementations and got very different performance.
>
> Version1:
> julia> @time map!(tanh, y, x)
> 0.149988 seconds (3.15 M allocations: 48.000 MB, 1.24% gc time)
Functions as arguments have a known performance overhead. See FastAnonymous.
>
> Version2:
> julia> @time for i = 1:length(x)
> y[i] = tanh(x[i])
> end
> 0.355105 seconds (5.24 M allocations: 95.984 MB, 1.83% gc time)
Don't benchmark in global scope. Define a function that runs the loop for you.
--Tim
>
> Version3: this one is not in-place, however it's the fast one.
> julia> @time y += tanh(x)
> 0.045402 seconds (10 allocations: 16.000 MB, 6.37% gc time)
>
> It's quite counterintuitive for me that version1 and version 2 are much
> slower than version 3.
> Can someone explain this a bit?
>
> Thanks!
> Cheng
