TensorOperations and cuTENSOR combined
Author Jutho
11 Stars
Updated Last
11 Months Ago
Started In
March 2019


This functionality is now incorporated in TensorOperations.jl and this package has become obsolete

CuTensorOperations.jl is a Julia interface to NVidia's cuTENSOR library that enables TensorOperations.jl to work on CuArray objects. The low-level methods from cuTENSOR are wrapped by CuArrays.jl on the custom/experimental branch ksh/tensor. CuTensorOperations.jl provides a high level interface, but does not (yet) provide any fallback definitions for those cases where the cuTENSOR library does not work (e.g. certain combinations of eltypes, certain trace operations, ...). In particular, known operations that are currently not supported are outer products (contractions without any shared indices) and traces where the first index or dimension of the tensor participates in a trace.

By using CuTensorOperations, definitions are provided for a set of methods from TensorOperations.jl for CuArray objects to make @tensor and friends work with CuArray objects. This is type piracy: defining methods from another package for objects from another package. In time, CuTensorOperations.jl should become integrated into TensorOperations.jl and loaded when it is loaded together with CuArrays.jl.

Furthermore, CuTensorOperations provides a @cutensor macro, that acts like @tensor but will first transfer objects to the GPU by calling CuArray on them (the precise transformation call might change in the future). This is a no-op on existing CuArray objects that already live on the GPU, but otherwise (for arrays in the main memory) transfers them prior to doing the computation, so that NVidia's cuTENSOR library can be used for the computation. In a complicated expression with many basic operations, the transfer to GPU is only performed just before that object is required, so that some of the computation and transfer time can hopefully coincide. Newly created objects will reside on the GPU; if instead the left hand side is an existing array in the main memory, the final output will be transferred back into it.

Installation instructions

Get a working copy of Julia (v1.1 or later). Make sure CUDA is available at standard location.

Download cuTENSOR from the NVidia Developer Program and test your installation. Make sure is available at a location that is found by the system, e.g. /usr/local/lib on a MacOS X or Linux system.

Launch the Julia REPL, enter package mode by typing ] and install necessary packages

pgk> add CuArrays#ksh/tensor
pgk> add TensorOperations
pkg> add