Module warp 

Collectives over logical warps within a native device subgroup.

All lanes of a logical warp participate in each call. width is positive, does not exceed the native subgroup width, and groups must not straddle a native subgroup. No NVIDIA warp width is assumed by these algorithms.

Modules

batched

broadcast

exchange

exclusive_scan

exclusive_unseeded

head_segmented_reduce

inclusive_scan

io

logical_lane_id

logical_warp_base_id

logical_warp_id

record

reduce

scan

sort

tail_segmented_reduce

Functions

broadcast

Broadcast from a logical lane, not a native subgroup lane.

exclusive_scan

Exclusive scan seeded by initial, with the same participation contract.

exclusive_unseeded

Unseeded exclusive scan; the first logical lane’s output is unspecified.

head_segmented_reduce

Segmented reduction whose result is valid at each segment’s head lane. The first lane implicitly starts a segment.

inclusive_scan

Inclusive scan in lane order. Values beyond valid_lanes are unspecified. valid_lanes is uniform within a logical warp and lies in 1..=width.

logical_lane_id

logical_warp_base_id

logical_warp_id

reduce

Ordered reduction, broadcast to all lanes of the logical warp.

scan

Compute both scan forms with one scan and return the unseeded aggregate to every participating logical lane.

tail_segmented_reduce

Segmented reduction whose result is valid at each segment’s head lane. A true tail marks the last lane of a segment; the final lane is implicit.