Advanced

Tensor classes

• Expression

The library use expression template which is a useful technique used by some numerical linear algbera softwares in C++. tensors are also defined as expression. All the operations on the tensors are expressions as well. This with combination of lazy evaluation allow expression matching and fuse operations.

• Storage order

By default a tensor is colum major but row major tensors can be created using the ranked_tensor class.

using RowTensor = ten::tensor<float, 3, ten::storage_order::colum_major>;
RowTensor x({2, 3, 4});

using StaticRowTensor = ten::ranked_tensor<float, ten::shape<2, 3, 4>, ten::storage_order::row_major>;
StaticRowTensor x;

• Storage and Allocator

Storage of tensors in memory can have huge impact in performances. That’s why a tensor has its elements stored in a contiguous vector. Different storage classes are defined and can be used to store all the elements of a tensor. For example a dense storage class is used for dense tensors. By default all elements are allocated using std::allocator<T>. One can define their own allocator and plug it in the tensor class and have access to all operations.

• Indexing and slicing

Slicing operations are supported by using ten::seq and ten::mdseq. They are limited to tensors of up to 5 dimensions.

using ten::seq;
using ten::last;
ten::tensor<float, 3> x({2, 3, 4});
// Slicing using seq
auto slice = x(seq(0, 1), seq(0, 1), seq(0, last))
// Assign to slices
slice = 1.0f;