Struct BackendImpl 

pub struct BackendImpl {
    pub kernel: KernelRef,
    pub dtypes: &'static [DType],
    pub cost: fn(&[Shape], &FusedOpParams, &BackendCapabilities) -> CostEstimate,
    pub precision: PrecisionGuarantee,
    pub caps: KernelCaps,
    pub revision: KernelRevisionHash,
    pub cost_expr: Option<&'static CompiledCostExpr>,
}

Per-backend kernel implementation for one fused op. The optimizer reads this to (1) pre-resolve KernelRef for nodes it places on this backend, (2) score routes against CostEstimate, and (3) admit candidates against the per-route tolerance budget via PrecisionGuarantee.

Function-pointer composition (no trait-object indirection) — the registry stores BackendImpl values inline, the executor calls the function pointer directly.

A single (FusedOpId, BackendId) decision point may have multiple registered BackendImpls — one per dtype tuple, and potentially multiple alternatives at the same dtype (e.g. cuBLAS bias-epilogue vs CUTLASS bias-epilogue for (FUSED_LINEAR, [BF16,BF16,BF16,BF16], Cuda)). The route picker filters by Self::dtypes and ranks the remaining alternatives by cost + precision.

Fields

kernel: KernelRef

Dispatch wrapper for this backend’s kernel for this fused op. Same KernelRef signature as primitive-op kernels.

dtypes: &'static [DType]

Dtype tuple this kernel is registered for. Convention mirrors the binding-table key shape: typically [input1, input2, ..., output] per the existing register(table, op, &[dtype; N], backend, kernel) call sites. Using &'static [DType] keeps BackendImpl Copy and lets registrations declare dtypes inline as &[F32, F32, F32, F32] literals.

cost: fn(&[Shape], &FusedOpParams, &BackendCapabilities) -> CostEstimate

Cost-estimate function. Given the input shapes, the per-instance fused-op params, and the backend’s capabilities, returns a CostEstimate used for placement and route ranking.

precision: PrecisionGuarantee

Numerical precision properties of this kernel.

caps: KernelCaps

Layout / capability flags (e.g. strided_input).

revision: KernelRevisionHash

Revision hash — opaque identifier for the source-version of this kernel. Persisted optimization caches use this to detect kernel drift and invalidate stale cache entries (see docs/architecture/11-persistence.md).

cost_expr: Option<&'static CompiledCostExpr>

Declared cost AST compiled from the contract’s cost: field (FKC gap-closure §2.3, Task 2.4). None when the contract declared no cost expression, or for hand-written (non-FKC) registrations. When present, crate::fused_cost::fused_layer1_cost prefers it over the is_fused_cost_sentinel decompose-derived fallback.

Trait Implementations

impl Clone for BackendImpl

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 BackendImpl