Enum ShapeLayoutKind 

#[non_exhaustive]
#[repr(u16)]
pub enum ShapeLayoutKind {
    Pad = 0,
    Concat = 1,
    Permute = 2,
    Repeat = 3,
    Flip = 4,
    Roll = 5,
    Meshgrid = 6,
    Fill = 7,
    WriteSlice = 8,
    Contiguize = 9,
    Triu = 10,
    Tril = 11,
}

Shape / layout op discriminant — Category N.

Tags the kernel SKU for telemetry / autotuner-cache keys. Each variant has its own Plan type today (PadPlan, ConcatPlan, …) because their descriptor / args shapes differ enough that one ShapeLayoutPlan<T, N> doesn’t fit. The enum exists so all of them populate KernelSku::op from a shared discriminant space.

Variants (Non-exhaustive)

Non-exhaustive enums could have additional variants added in future. Therefore, when matching against variants of non-exhaustive enums, an extra wildcard arm must be added to account for any future variants.

Pad = 0

F.pad(x, pad, mode='constant', value=v) — Phase 3 trailblazer.

Concat = 1

torch.cat(tensors, dim) — variable-arity input. Reserved.

Permute = 2

Materialized torch.permute(x, dims) (strided-view materialization when needed). Reserved.

Repeat = 3

x.repeat(...) / torch.tile(x, ...). Reserved.

Flip = 4

torch.flip(x, dims) — reverse along axes. Reserved.

Roll = 5

torch.roll(x, shifts, dims) — shift along axes. Reserved.

Meshgrid = 6

torch.meshgrid(*tensors) — N rank-1 → N rank-N. Reserved.

Fill = 7

torch.full(shape, value) / Tensor.fill_(value) — fill every element of an output tensor with a scalar constant. Wired from fuel-cuda-kernels/fill.cu.

WriteSlice = 8

dest[start_0..end_0, ..., start_{N-1}..end_{N-1}] = source (assign, not accumulate). Per-axis range write. Phase 13.1 trailblazer — driven by Fuel team’s persistent KV-cache append (autoregressive decoding). See baracuda_kernels::WriteSlicePlan.

Contiguize = 9

Strided→contiguous materialization (torch.Tensor.contiguous). Phase 13.2: closes the D2H→CPU contiguize→H2D fallback cliff for non-contiguous CUDA inputs. Byte-level dtype-agnostic (sizeof-templated kernel) covering every byte-aligned dtype; nibble (S4 / U4) shipped behind a documented innermost-stride constraint. See baracuda_kernels::ContiguizePlan.

Triu = 10

torch.triu(input, diagonal) — keep upper triangular part of the last two dims of input; zero everything below the diagonal-th diagonal. Batch dims (anything before the last two) are independently masked. Phase 13.4 trailblazer — driven by Fuel team’s CPU-only triu/tril gap. See baracuda_kernels::TriuPlan.

Tril = 11

torch.tril(input, diagonal) — keep lower triangular part of the last two dims of input; zero everything above the diagonal-th diagonal. Sibling of Self::Triu with the predicate flipped. See baracuda_kernels::TrilPlan.

Trait Implementations

impl Clone for ShapeLayoutKind

fn clone(&self) -> ShapeLayoutKind

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Copy for ShapeLayoutKind

impl Debug for ShapeLayoutKind

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

Formats the value using the given formatter.

impl Eq for ShapeLayoutKind

impl Hash for ShapeLayoutKind

fn hash<__H>(&self, state: &mut __H)

where __H: Hasher,

Feeds this value into the given Hasher.

fn hash_slice<H>(data: &[Self], state: &mut H)

where H: Hasher, Self: Sized,

Feeds a slice of this type into the given Hasher.

impl PartialEq for ShapeLayoutKind

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

Tests for self and other values to be equal, and is used by ==.

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

Tests for !=. The default implementation is almost always sufficient, and should not be overridden without very good reason.

impl StructuralPartialEq for ShapeLayoutKind