Module sort 

Sorting / order-statistics op family — Category O.

Phase 9 of the baracuda-kernels comprehensive plan. Ships the block-bitonic trailblazer family:

• SortPlan / SortBackwardPlan / ArgsortPlan / MsortPlan / MsortBackwardPlan — block-bitonic sort, one CUDA block per row. Trailblazer cap: row_len ≤ 1024. Larger rows are reserved for a future tile-radix follow-up; the Plan returns Unsupported for row_len > 1024.
• TopkPlan / TopkBackwardPlan / KthvaluePlan / KthvalueBackwardPlan — block-bitonic select; trailblazer cap: k ≤ 64 (LLM-inference range).
• UniquePlan / UniqueConsecutivePlan — set-valued, no BW. unique chains sort + consecutive-dedup at the plan layer.
• HistogramPlan / HistogramddPlan / BincountPlan — atomic-bin accumulation; FW only. histogramdd returns Unsupported for ndim > 1 in the trailblazer.
• SearchsortedPlan — per-query binary search; FW only.

Saved-indices contract for sort / msort / topk / kthvalue BW. The FW emits both sorted values AND sorted indices in a single launch (FW Args carry values and indices as required outputs). BW Args receive the saved indices verbatim — no recomputation at BW time. The Plan’s selector pegs the indices dtype to i32 across every kernel SKU in this family.

Re-exports

pub use argsort::ArgsortArgs;

pub use argsort::ArgsortDescriptor;

pub use argsort::ArgsortPlan;

pub use bincount::BincountArgs;

pub use bincount::BincountDescriptor;

pub use bincount::BincountPlan;

pub use histogram::HistogramArgs;

pub use histogram::HistogramDescriptor;

pub use histogram::HistogramPlan;

pub use histogramdd::HistogramddArgs;

pub use histogramdd::HistogramddDescriptor;

pub use histogramdd::HistogramddPlan;

pub use kthvalue::KthvalueArgs;

pub use kthvalue::KthvalueDescriptor;

pub use kthvalue::KthvaluePlan;

pub use kthvalue_backward::KthvalueBackwardArgs;

pub use kthvalue_backward::KthvalueBackwardDescriptor;

pub use kthvalue_backward::KthvalueBackwardPlan;

pub use msort::MsortArgs;

pub use msort::MsortBackwardArgs;

pub use msort::MsortBackwardDescriptor;

pub use msort::MsortBackwardPlan;

pub use msort::MsortDescriptor;

pub use msort::MsortPlan;

pub use searchsorted::SearchsortedArgs;

pub use searchsorted::SearchsortedDescriptor;

pub use searchsorted::SearchsortedPlan;

pub use sort::SortArgs;

pub use sort::SortDescriptor;

pub use sort::SortPlan;

pub use sort_backward::SortBackwardArgs;

pub use sort_backward::SortBackwardDescriptor;

pub use sort_backward::SortBackwardPlan;

pub use topk::TopkArgs;

pub use topk::TopkDescriptor;

pub use topk::TopkPlan;

pub use topk_backward::TopkBackwardArgs;

pub use topk_backward::TopkBackwardDescriptor;

pub use topk_backward::TopkBackwardPlan;

pub use unique::UniqueArgs;

pub use unique::UniqueDescriptor;

pub use unique::UniquePlan;

pub use unique_consecutive::UniqueConsecutiveArgs;

pub use unique_consecutive::UniqueConsecutiveDescriptor;

pub use unique_consecutive::UniqueConsecutivePlan;

Modules

argsort

argsort plan — sorted indices only (no values output).

bincount

bincount plan — count occurrences of each integer in x.

histogram

histogram plan — 1-D uniform-bin atomic-accumulating histogram.

histogramdd

histogramdd plan — N-D histogram. Reserved for follow-up.

kthvalue

kthvalue plan — returns the k-th smallest value + its index along the last dimension.

kthvalue_backward

kthvalue_backward plan — scatter the scalar dy[batch] back to the saved-index position in dx[batch, row_len].

msort

msort (stable sort) plan + BW.

searchsorted

searchsorted plan — per-query binary search in a 1-D sorted array.

sort

sort plan — Category O trailblazer.

sort_backward

sort_backward plan — scatter dy via the saved indices.

topk

topk plan — block-bitonic top-k select.

topk_backward

topk_backward plan — scatter k-wide dy into row_len-wide dx via the saved indices. Launcher zeros dx first.

unique

unique plan — sort + consecutive-dedup composition.

unique_consecutive

unique_consecutive plan — emit one cell per run-start in each row.

Constants

SORT_MAX_ROW

Maximum supported row_len in the block-bitonic trailblazer. Must match MAX_ROW in baracuda_sort.cuh.

TOPK_MAX_K

Maximum supported k in the block-bitonic topk trailblazer. Must match MAX_K in baracuda_topk.cuh.