Module shape 

Shape and Strides - Tensor Dimension Management

Provides types and functions for managing tensor shapes, strides, and broadcasting rules. Shapes define the dimensions of a tensor, while strides define how to traverse the underlying storage.

Key Features

• Efficient shape representation with small-vector optimization
• Stride computation for contiguous and transposed layouts
• Broadcasting support following NumPy rules
• Shape validation and manipulation

@version 0.1.0 @author AutomataNexus Development Team

Functions

broadcast_shape

Computes the broadcast shape of two shapes.

broadcast_strides

Computes broadcast strides for a shape to match a target shape.

can_broadcast

Checks if two shapes are broadcastable.

contiguous_strides

Computes row-major (C-order) strides for a shape.

is_contiguous

Checks if strides represent a contiguous memory layout.

linear_index

Computes the linear index from multi-dimensional indices.

normalize_dim

Normalizes a dimension index, supporting negative indexing.

numel

Computes the total number of elements from a shape.

reshape

Reshapes a tensor shape, validating that total elements match.

squeeze

Computes the shape after squeezing (removing dimensions of size 1).

transpose_shape

Computes the shape after transposing dimensions.

transpose_strides

Swaps two stride values.

unravel_index

Converts a linear index to multi-dimensional indices.

unsqueeze

Computes the shape after unsqueezing (adding a dimension of size 1).

validate_indices

Validates that indices are within bounds for a shape.

Type Aliases

Shape

Shape type - dimensions of a tensor. Uses SmallVec for stack allocation of small shapes (up to 6 dimensions).

Strides

Strides type - step sizes for each dimension.