CoreTensor is a tensor library and a sub-project of the DLVM project.
It provides:
- Shaping
- Storage
- Indexing
- Slicing
- Linear algebra shape transformations
- Broadcasting
- Collection behavior
-
CoreTensor implements completely dynamically shaped tensors. Scalars, vectors, matrices and n-D arrays are all represented as
Tensor<T>
, whereT
represents the type of each unit. EachTensor<T>
stores aTensorShape
, which wraps an array of integers representing the shape of the tensor. -
RankedTensor implements dynamically shaped but statically ranked tensors. Instead of storing
TensorShape
(which wraps a dynamically sized array), RankedTensor uses a tuple of integers to represent a shape with known rank. TypesT
,Tensor1D<T>
,Tensor2D<T>
,Tensor3D<T>
, andTensor4D<T>
represent scalars, vectors, matrices, rank-3 tensors, and rank-4 tensors, respectively.
Apache 2.0