Open
Description
The simplest approach is to define a n-dimensional data structure, together with a few operations (access, slicing). Then 2-d (resp. 1-d) tensors can be converted to matrices (resp. vectors) without copying, and viceversa.
This already would give us the possibility move around higher-dimensional data, and then treat slices of these data as matrices and vectors to perform linear algebraic operations.
Tensor contraction (dot) can be implemented in these terms, or perhaps directly on top of gemm