Struct KernelCaps 

pub struct KernelCaps {
    pub strided_input: bool,
    pub requires_broadcast: bool,
}

Per-binding capability flags. Today carries one flag — strided_input — that signals “this kernel walks input strides explicitly and so can consume non-contiguous input layouts (including stride-0 broadcast axes) without auto-Contiguize materializing them first.” The executor’s contiguize gate consults this to skip the materialize step for capable kernels, so broadcast/transpose/slice layouts can reach the kernel as metadata-only views.

Default is the conservative all-false; binding sites opt in via KernelBindingTable::register_with_caps. Forward-extensible by adding fields (no enum/bitflags churn).

Fields

strided_input: bool

Kernel handles non-contiguous input layouts directly (consumes the strides side-channel from KernelRef::layouts). When true, the executor passes non-contiguous inputs through unchanged instead of running auto-Contiguize first. Inputs with non-zero start_offset still go through auto-Contiguize today (slicing the device buffer to honor offset is a separate concern).

requires_broadcast: bool

This kernel BAKES a specific broadcast pattern into its body — a broadcast_stride0: required operand it reads as stride-0 (hoisting in[0] / dropping the bcast-axis stride terms). It is therefore NOT a generic dense-safe kernel: a dense operand routed here would be read WRONG (element 0 repeatedly). The realize pick (KernelBindingTable::lookup_with_caps) SKIPS such a sibling so a dense operand is never routed into a broadcast-baked kernel — the exclusion is order-independent (it does not rest on the generic contract being registered first). Selecting it CORRECTLY, for an operand genuinely stride-0 on exactly the required axes, is the layout-aware follow-on (path 1b) that also teaches the contiguize gate; today the safe behavior is to exclude it, so a required-broadcast contract is a present-but-never-mis-selected sibling. Defaults false — every existing binding is unaffected (the pick stays first-registered).

Implementations

impl KernelCaps

pub const fn empty() -> Self

All flags off. Equivalent to Default::default(); provided as a const for use in const-context registration tables.

pub const fn strided_input() -> Self

Just strided_input on. Ergonomic for binary/unary registrations that opt in to the wrapper-side broadcast path.

Trait Implementations

impl Clone for KernelCaps

fn clone(&self) -> Self

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Copy for KernelCaps

impl Debug for KernelCaps

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

Formats the value using the given formatter.

impl Default for KernelCaps

fn default() -> Self

Returns the “default value” for a type.

impl Eq for KernelCaps

impl PartialEq for KernelCaps

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

Equality operator ==.

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

Inequality operator !=.

impl StructuralPartialEq for KernelCaps