How can @tensor macro be used in variable length systems?

[billysrh]

I'm used to write code in ncon fashion, and it's necessary because I have to deal with arbitrary number of tensors. But I know @ncon is slower than @tensor. How can I keep the notion of ncon, and get the speed of @tensor at the same time?

[Jutho]

There is some speed advantage with using @tensor, but it is not that substantial. It arises specifically from the fact that @tensor only works if the contraction is exactly known at compile time, and the different steps can therefore be hardcoded. That's what the macro does, it does some code rewriting at compile time.

If the contraction requires dynamic information, which is not known at compile time, you cannot use @tensor for that.

[sents]

If you really need the speed and you don't have a lot of different contractions you can call the macro from a generated function. The generated function can dynamically dispatch by handing it a Val type.

Generated functions have to compile a function for every call of different type they get. This means that each contraction with a different signature would have to be recompiled. So using this approach with rapidly changing contraction orders will introduce a lot of compilation overhead.

Here is an example, handing the operators the indices (as Tuples of Ints):

using TensorOperations
@generated function contract_operator!(S,
                                       A,
                                       B,
                                       op,
                                       order_A::Val{K}) where {K}
    leftside = Expr(:call, :*,
                    :(A[$(K.a...)]),
                    :(B[$(K.b...)]),
                    :(op[$(K.op...)]))
    return :(@tensor S[:] = $leftside)
end
S = rand(5,5,5,2,5,5,5,2)
a = rand(5,5,5,5,2)
b = rand(5,5,5,5,2)
op = rand(2,2,2,2)
contraction_order = Val((a = (1,-1,-2,-3,2),
                     b = (-5,-6,1,-7,3),
                     op = (2,3,-4,-8)))
contract_operator!(S,a,b,op, contraction_order);

[davidsmp]

Hi,

I wanted to ask, is there an alternative to the above solution? What is the best way of dealing with variable length systems and the @tensor macro?

Speed is not really an issue for me, but as I want ot use AD I cannot use the ncon function.

Are there plans to have AD support for ncon?

Best,
David

[lkdvos]

Hi!

I was not aware that ncon does not work with AD. In principle it should, so could you perhaps show us what is not working for you?
(This might be better in a new issue)

[davidsmp]

Hi, sorry, it was not an issue of ncon, indeed it works

[lkdvos]

Great! Happy that got sorted out :)