Struct SortPlan 

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

sort plan.

sort(x, dim=last, descending) — returns sorted values AND sorted indices along the last axis (PyTorch torch.sort).

When to use: forward row-wise sort. Pair with SortBackwardPlan for autograd; for indices-only output use ArgsortPlan.

Dtypes: {f32, f64, i32, i64}; indices always i32.

Shape limits: rank-2 [batch, row_len]; row_len ≤ 1024 (one CUDA block per row, bitonic network in shared memory). Larger rows return Unsupported — tile-radix follow-up reserved.

Workspace: none.

Precision guarantee: deterministic, bit-stable. Block-bitonic is a fixed comparator network — no atomics, no reductions.

Saved-indices contract: FW emits both values and indices in a single launch. BW reads the saved indices verbatim; callers must retain indices for autograd.

Implementations

impl<T: Element> SortPlan<T>

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

Pick a kernel for desc.

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

Validate args.

pub fn workspace_size(&self) -> usize

Workspace size in bytes.

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: SortArgs<'_, T>, ) -> Result<()>

Launch.