so it is about 25% faster and uses a lot less memory.
In order to use it, you first have to install TBLIS.jl (and right now, to get commands like TBLIS.set_num_threads(4), you need to use my fork: https://github.com/mtfishman/TBLIS.jl).
Note that it is only limited to real contractions right now, since the TBLIS support for complex tensors is currently limited (https://github.com/devinamatthews/tblis/issues/18). I've also found that it doesn't help with the DMRG calculations I have tried (specifically, it doesn't give a speedup, and sometimes is even a bit slower). My guess is that many DMRG contractions can map directly to BLAS calls without permutations, and in those cases TBLIS is slower than MKL. It may also have to do with the tensor dimensions involved.
This adds support for using TBLIS as a backend for (real) tensor contractions. Here is the interface:
This outputs something like the following on my laptop:
so it is about 25% faster and uses a lot less memory.
In order to use it, you first have to install TBLIS.jl (and right now, to get commands like
TBLIS.set_num_threads(4)
, you need to use my fork: https://github.com/mtfishman/TBLIS.jl).Note that it is only limited to real contractions right now, since the TBLIS support for complex tensors is currently limited (https://github.com/devinamatthews/tblis/issues/18). I've also found that it doesn't help with the DMRG calculations I have tried (specifically, it doesn't give a speedup, and sometimes is even a bit slower). My guess is that many DMRG contractions can map directly to BLAS calls without permutations, and in those cases TBLIS is slower than MKL. It may also have to do with the tensor dimensions involved.