Struct Layout 

pub struct Layout { /* private fields */ }

Describes the memory layout of a tensor: shape, strides, and start offset.

Strides are in units of elements (not bytes) and are signed — metadata-only view ops (notably Op::Flip) reverse iteration along a dim by negating that dim’s stride and adjusting start_offset, so a stride can be negative even though sizes can’t. A contiguous (row-major) tensor of shape [a, b, c] has strides [b*c, c, 1] (all positive); a Flip-on-dim-0 of that has stride [-b*c, c, 1] with start_offset set to the byte offset of the (originally) last element along dim 0.

start_offset is the byte offset of the layout’s first element in iteration order — by construction always non-negative, even for negative-stride layouts.

Implementations

impl Layout

pub fn new(shape: Shape, stride: StrideVec, start_offset: usize) -> Self

Creates a layout with explicit shape, strides, and start offset.

pub fn contiguous_with_offset<S: Into<Shape>>( shape: S, start_offset: usize, ) -> Self

Creates a contiguous (row-major) layout with the given start offset.

pub fn contiguous<S: Into<Shape>>(shape: S) -> Self

Creates a contiguous (row-major) layout with zero start offset.

pub fn dims(&self) -> &[usize]

Returns the dimension sizes as a slice.

pub fn dim<D: Dim>(&self, dim: D) -> Result<usize>

Returns the size of a specific dimension.

pub fn shape(&self) -> &Shape

Returns a reference to the Shape.

pub fn has_dynamic(&self) -> bool

Whether the layout’s shape has any bounded-symbolic (Range) axis. Phase D step 1b.

pub fn resolve(&self, env: &SymEnv) -> Result<Layout>

Resolve the layout’s shape through env (every dynamic axis → concrete), keeping strides and offset unchanged: the physical walk of a fixed-capacity buffer is the same; only the logical live extent becomes concrete. Errors if a dynamic axis’s symbol is unbound. Phase D step 1b.

pub fn stride(&self) -> &[isize]

Returns the strides as a slice (in elements, not bytes). Signed — see struct-level docs.

pub fn stride_unsigned(&self) -> DimVec

Returns strides cast to usize, debug-asserting non-negative. Convenience for kernels that haven’t been updated to handle signed strides yet (most of them, since auto-Contiguize guarantees positive strides on kernel input today). One stack allocation via DimVec; effectively zero cost for rank ≤ 6.

Don’t use this if your kernel needs to support negative strides (e.g. is the consumer of an Op::Flip view): iterate via crate::StridedIndex which handles the sign correctly.

pub fn start_offset(&self) -> usize

Returns the start offset into the underlying storage.

pub fn contiguous_offsets(&self) -> Option<(usize, usize)>

Returns (start, end) offsets if the layout is contiguous, None otherwise.

pub fn is_contiguous(&self) -> bool

Returns true if the layout is C-contiguous (row-major).

This does not require the start offset to be 0.

pub fn is_fortran_contiguous(&self) -> bool

Returns true if the layout is Fortran-contiguous (column-major).

pub fn narrow(&self, dim: usize, start: usize, len: usize) -> Result<Self>

Returns a narrowed layout selecting len elements starting at start on dimension dim.

pub fn transpose(&self, dim1: usize, dim2: usize) -> Result<Self>

Returns a layout with two dimensions swapped.

pub fn permute(&self, idxs: &[usize]) -> Result<Self>

Returns a layout with dimensions reordered according to idxs.

pub fn unsqueeze(&self, dim: usize) -> Result<Self>

Returns a layout with a size-1 dimension inserted at position dim. dim must be in 0..=rank — dim == rank appends to the end. Pure metadata: the underlying bytes are unchanged.

Stride for the inserted size-1 axis is set to 0. This stride value never affects indexing (the axis only ever has coordinate 0, and 0 * stride == 0), so any value would be correct. Choosing 0 matches the convention used by broadcast_as for size-1 axes broadcast to larger sizes — keeps the convention uniform and Shape::is_contiguous stays unaffected (it already skips dims of size 1).

pub fn squeeze(&self, dim: usize) -> Result<Self>

Drop a size-1 dimension at position dim. Inverse of Self::unsqueeze. The pruned dim’s stride is discarded — it’s irrelevant for indexing because the dim has size 1. Errors if dim is out of bounds or its size isn’t 1.

pub fn flip(&self, dim: usize) -> Result<Self>

Returns a layout that iterates dim in reverse — pure metadata, zero copy. The stride at dim is negated; start_offset is shifted forward by (size-1) * |stride| so iteration starts at what was the last element along dim.

Size-0 and size-1 axes are no-ops (no element to flip; the new start_offset shift is 0). Successive flip calls on the same dim are involutive: stride re-negates, offset cancels, you get the original layout back.

pub fn broadcast_as<S: Into<Shape>>(&self, shape: S) -> Result<Self>

Returns a layout broadcast to the given shape.

Dimensions of size 1 in self are stretched to match the target shape.

pub fn strided_index(&self) -> StridedIndex<'_>

pub fn strided_blocks(&self) -> StridedBlocks<'_>

pub fn offsets_b(&self) -> Option<ContiguousOffsetsWithBroadcast>

Trait Implementations

impl Clone for Layout

fn clone(&self) -> Self

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for Layout

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.

impl Eq for Layout

impl PartialEq for Layout

fn eq(&self, other: &Self) -> bool

Equality operator ==.

fn ne(&self, other: &Rhs) -> bool

Inequality operator !=.

impl StructuralPartialEq for Layout