Struct TensorView 

pub struct TensorView<'a> {
    pub data: DevicePtr,
    pub dtype: DataType,
    pub shape: &'a [usize],
    pub strides: &'a [i64],
    pub byte_offset: usize,
    pub device: DeviceId,
    pub backing: TensorBacking,
    /* private fields */
}

Immutable, non-owning view of a tensor on any device.

Fields

data: DevicePtr

dtype: DataType

shape: &'a [usize]

strides: &'a [i64]

Strides in elements (may be negative, matching DLPack).

byte_offset: usize

Offset in bytes of the element origin from data (DLPack semantics).

device: DeviceId

backing: TensorBacking

Implementations

impl<'a> TensorView<'a>

pub fn new( data: DevicePtr, dtype: DataType, shape: &'a [usize], strides: &'a [i64], device: DeviceId, ) -> Self

Construct a zero-offset view. data must remain valid for 'a.

pub fn absent(dtype: DataType) -> Self

Construct an absent view: the positional placeholder an executor passes for an omitted optional input (an ONNX empty-string input name). It carries a null pointer and an empty shape so it can never be read as a real tensor — kernels test TensorView::is_absent before touching an optional slot. This preserves positional arity so a later present input (e.g. Slice’s steps when axes is omitted) is not misread as the omitted one.

pub fn is_absent(&self) -> bool

Whether this view is the TensorView::absent placeholder for an omitted optional input (null backing pointer). Kernels with optional inputs check this to distinguish “input not supplied” from a real, possibly-empty, tensor.

pub fn with_byte_offset(self, byte_offset: usize) -> Self

Set the DLPack-style byte offset of the element origin.

pub fn with_backing(self, backing: TensorBacking) -> Self

pub fn validate(&self) -> Result<()>

Check the view’s invariants (rank, dtype, offset alignment). See [validate_view]; call this on any view imported from DLPack before use.

pub fn is_contiguous(&self) -> bool

Whether the view is contiguous row-major.

pub fn numel(&self) -> usize

Number of elements.

pub fn byte_size(&self) -> usize

Logical element byte size (dense; ignores stride gaps). Uses storage_bytes so sub-byte (int4/uint4) types are counted correctly.

pub fn data_ptr<T>(&self) -> *const T

Typed const pointer to the element origin, applying byte_offset. Computed with wrapping arithmetic (no deref) — safe to call.

Trait Implementations

impl<'a> Clone for TensorView<'a>

fn clone(&self) -> TensorView<'a>

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl<'a> Copy for TensorView<'a>