Tensor

Tensor

Header: include/stratax/containers/Tensor.hpp.

stratax::container::Tensor<T> owns contiguous row-major elements of a supported DType, including bool. ArrayBase supplies storage, metadata, checked flat access, and iterators. Default construction produces empty shape {0}.

Common interface and invariants

size() == shape().elements() and rank() == shape().rank() == strides().rank(). Metadata size/empty/rank/shape/strides queries take O(1). Shape and strides are read-only. operator[] is unchecked flat indexing; at(i) checks a signed flat index, including negative indices. Invalid at, front, or back access raises Exceptions::IndexError. data returns null for empty owning storage.

begin/end, const variants, and reverse iterators traverse flat row-major values; acquisition is O(1), traversal and fill are O(n). Owning iterators are contiguous. Member and ADL swap exchange storage and metadata in O(1). Pointers and iterators follow the allocation under its new owner; replacing or destroying storage invalidates its references. Slicing is a free function returning ArrayView.

Rank, construction, and moves

Shape construction accepts any rank and value-initializes elements; a second value argument fills them. Rank zero explicitly constructed with Shape{} is empty, not a scalar. The default Tensor instead has shape {0} and rank one.

Copies deeply copy elements and metadata; assignment is transactional. Copy construction costs O(n + r), and assignment also destroys old state. Noexcept moves transfer the three buffers, leaving a distinct source empty with rank zero. Move construction is O(1), assignment also destroys old state.

Multidimensional indexing

operator() accepts an integral index pack or vector of size_t indices. It checks neither rank nor bounds. Supply exactly rank nonnegative in-range components; vector indexing additionally requires positive rank.

Multidimensional at accepts signed indices as a pack or vector. It raises RankError for the wrong component count and IndexError for an out-of-range component or empty storage. An empty signed-index vector on a rank-zero Tensor raises IndexError; a rank mismatch still takes precedence. A single scalar at argument selects the inherited flat overload. Multidimensional access takes O(r).

#include <stratax.h>
#include <cassert>
#include <utility>
 
int main() {
    stratax::container::Tensor<double> values(stratax::core::Shape{2, 2, 2}, 0.0);
    values(1, 0, 1) = 5.0;
    assert(values.at(-1, 0, -1) == 5.0);
    auto moved = std::move(values);
    assert(values.rank() == 0 && values.empty());
    assert(moved.shape() == stratax::core::Shape({2, 2, 2}));
}

Construction costs and failures

Allocation failures propagate. Shape/count/stride overflow raises DimensionError; byte-count overflow may raise std::bad_array_new_length. Construction copies or initializes n elements plus rank metadata. Matrix nested-list validation also visits each row, including empty rows. Invalid fixed-rank shapes are rejected by the container; see the generated header reference for exact constructors.

See ArrayBase, Shape, and ArrayView.