Struct UniquePlan 

pub struct UniquePlan<T: Element> { /* private fields */ }

unique plan.

Sort-then-dedup composition (PyTorch torch.unique(x, sorted=True)). At the plan layer chains SortPlan into a caller-supplied scratch buffer, then UniqueConsecutivePlan to collapse runs.

When to use: per-row distinct-value extraction. Set-valued — no BW (output dimensionality is data-dependent).

Dtypes: {f32, f64, i32}.

Shape limits: input [batch, row_len]; row_len ≤ 1024 (sort cap). Outputs [batch, max_unique]; the caller’s max_unique bounds the output; rows with more uniques have the overflow dropped (the counter[] reports the actual count).

Workspace: zero in Workspace; caller supplies sorted_scratch, sorted_idx_scratch, and counter in UniqueArgs.

Precision guarantee: deterministic, bit-stable.

Implementations

impl<T: Element> UniquePlan<T>

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

Pick a kernel for desc.

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

Validate args.

pub fn workspace_size(&self) -> usize

Workspace size in bytes (the sorted-scratch buffers are caller- supplied as Args fields, so the plan reports 0 here).

pub fn sku(&self) -> KernelSku

Identity of the kernel this plan picked.

pub fn precision_guarantee(&self) -> PrecisionGuarantee

Numerical guarantees for this plan’s kernel.

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

Launch — sort, then dedup.