[FEA]: Add LRB segmented-sort kernels using the runtime-width group merge primitive #10946

Is this a duplicate?

• I confirmed there appear to be no duplicate issues for this request and that I agree to the Code of Conduct

Area

CUB

Is your feature request related to a problem? Please describe.

Fine-grained segmented sort needs kernels for sub-warp/warp, small-block, and large-block work-plan tiers. These kernels have similar bodies and use the same runtime-width group merge-sort primitive, but require different execution and resource configurations.

Describe the solution you'd like

Track the three tier-specific kernels:

• [FEA]: Add a sub-warp/warp segmented-sort kernel using the runtime-width group merge primitive #10947: sub-warp/warp kernel.
• [FEA]: Add a small-block segmented-sort kernel using the runtime-width group merge primitive #10948: small-block kernel.
• [FEA]: Add a large-block segmented-sort kernel using the runtime-width group merge primitive #10949: large-block kernel for segments that fit within one CTA-scale tile.

Each kernel should consume its work-plan tier and invoke the runtime-width group merge-sort primitive tracked in #10944.

This issue can be closed once optimized segmented sort algorithms can be tested and benchmarked but not reachable by user.
cub::detail::lrb_based_segmented_sort

Describe alternatives you've considered

A single kernel for all tiers was considered, but the tiers require different execution and resource configurations.

Additional context