Potential performance improvement?

[Staross]

Hi, I've been playing with TensorOperations and it seems pretty nice. I've noticed it's a bit slower (3-4x, Julia Version 0.4.0-dev+3052) than manual loops in some situations, probably because it allocates more memory. Maybe there's some potential to improve performances there:

http://pastebin.com/HkKU9fCp

[Jutho]

Well, there will always be many cases where manual loops for specific contractions are faster than the general purpose algorithm used by tensorcontract. It is impossible to have code that is optimal for any given contraction, unless one generates code for every given contraction. With stagedfunction in julia v0.4, this would in principle be possible, but given the exponential increase in possible number of contractions, this might not be feasible. It's something I would like to try out once I have some more time.

I will anyway be updating TensorOperations.jl to make use of some of the new features of julia v0.4 soon, and hopefully this will bring along some performance improvements.