Struct RepeatBackwardPlan 

pub struct RepeatBackwardPlan<T: Element, const N: usize> { /* private fields */ }

repeat backward plan.

Adjoint of crate::RepeatPlan: dx[c_in] = Σ_k dy[c_in + k · input_shape] — every dy cell that maps back to c_in under the FW’s modulo contributes. One thread per dx cell sweeps the prod(repeats[d]) contributing cells. f16 / bf16 accumulate in f32 internally; f32 / f64 accumulate in their native dtype.

When to use: BW for RepeatPlan.

Dtypes: {f32, f64, f16, bf16}.

Shape limits: rank in [1, 8]; repeats[d] ≥ 1.

Workspace: none.

Precision guarantee: deterministic (no atomics — one thread per output cell, deterministic iteration order). Conservatively reported as not bit-stable because summation order matters in FP semantics and a future refactor might reorder the inner loop.

Implementations

impl<T: Element, const N: usize> RepeatBackwardPlan<T, N>

pub fn select( _stream: &Stream, desc: &RepeatBackwardDescriptor<N>, _pref: PlanPreference, ) -> Result<Self>

Pick a kernel for desc.

pub fn can_implement(&self, args: &RepeatBackwardArgs<'_, T, N>) -> Result<()>

Validate args.

pub fn workspace_size(&self) -> usize

Workspace size in bytes. Always 0.

pub fn sku(&self) -> KernelSku

Identity of the kernel this plan picked.

pub fn precision_guarantee(&self) -> PrecisionGuarantee

Numerical guarantees.

pub fn run( &self, stream: &Stream, _workspace: Workspace<'_>, args: RepeatBackwardArgs<'_, T, N>, ) -> Result<()>

Launch.