TensorMap constructor is slow when applied to an array

[ebelnikola]

Hi!

I noticed that TensorMap constructor is extraordinarily slow when applied to an array. Here is a minimal example:

space=Z2Space(0=>10,1=>10);

@time A=TensorMap(randn,space⊗space←space⊗space); # takes time of roder 1e-4s
A_array=convert(Array,A);

@time A_copy_1=TensorMap(A_array,space⊗space←space⊗space); # takes time of order 10s
println(A_copy_1==A) # prints true

At the same time, the following code takes time of order 1e-4s:

@time begin
    dom=space⊗space
    codom=dom
    A_copy_2=TensorMap(zeros,codom←dom)

    for (f1,f2) in fusiontrees(A_copy_2)
        a3,a4=axes(dom,f2.uncoupled)
        a1,a2=axes(dom,f1.uncoupled)
        A_copy_2[f1,f2]=A_array[a1,a2,a3,a4]
    end
end
println(A_copy_1==A)

[Jutho]

The code for this in the latest tagged version is indeed quite suboptimal, because this is not something that is expected to be used in computation-heavy parts of the code. We figured this would be used only to initialise some small tensors such as single-site operators or two-site hamiltonians. However, on the development version that we are currently working on, this has been improved quite dramatically, and this specifica example currently takes only order 2*10^-3 seconds on my computer.

[ebelnikola]

I see, thanks!