Struct TensorRef 

pub struct TensorRef<'a, T, const N: usize>
where
    T: DeviceRepr + Copy + 'static,
{
    pub data: DeviceSlice<'a, T>,
    pub shape: [i32; N],
    pub stride: [i64; N],
}

Read-only view of a device-resident rank-N tensor.

shape[i] is the extent along axis i. stride[i] is the element stride along axis i (memory offset to advance one step along that axis). A stride of 0 along an axis signals a broadcast operand — the kernel reads the same memory cell for every step along that axis. The default row-major contiguous stride for shape [d0, d1, …, dN-1] is [d1·d2·…·dN-1, …, dN-1, 1].

T is bounded by DeviceRepr only (not by crate::Element / crate::IntElement) so the same view struct can carry any scalar payload — input, output, mask, or auxiliary buffer. The per-op element-class enforcement happens at the Plan layer.

Fields

data: DeviceSlice<'a, T>

Device-resident element storage.

shape: [i32; N]

Extent along each axis (in elements).

stride: [i64; N]

Element stride along each axis. Stride 0 marks a broadcast axis.

Implementations

impl<'a, T, const N: usize> TensorRef<'a, T, N>

where T: DeviceRepr + Copy + 'static,

pub fn numel(&self) -> i64

Total number of logical elements (product of shape).

Returns 1 for the rank-0 (scalar) case. Saturates on overflow rather than wrapping — a tensor with overflowing element count cannot fit in CUDA’s int64_t element-count surface anyway, so reporting i64::MAX is a fine sentinel.

pub fn is_contiguous(&self) -> bool

true iff stride matches the standard row-major contiguous layout (rightmost axis has stride 1; each prior axis multiplies by the extent to its right).

The rank-0 case is always contiguous. A broadcast tensor (any stride[i] == 0) is not contiguous.

Trait Implementations

impl<'a, T, const N: usize> Clone for TensorRef<'a, T, N>

where T: Clone + DeviceRepr + Copy + 'static,

fn clone(&self) -> TensorRef<'a, T, N>

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl<'a, T, const N: usize> Copy for TensorRef<'a, T, N>

where T: Copy + DeviceRepr + 'static,

impl<'a, T, const N: usize> Debug for TensorRef<'a, T, N>

where T: Debug + DeviceRepr + Copy + 'static,

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

Formats the value using the given formatter.