Prepare three-way partition tuning policies for lookahead

[bernhardmgruber]

Fixes part of: #10075

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Enterprise

Run ID: e94aaa11-6301-415a-a9fc-e2471933d9d0

📥 Commits

Reviewing files that changed from the base of the PR and between e9a6538 and 4581da4.

📒 Files selected for processing (1)

• cub/cub/device/dispatch/tuning/tuning_three_way_partition.cuh

🚧 Files skipped from review as they are similar to previous changes (1)

• cub/cub/device/dispatch/tuning/tuning_three_way_partition.cuh

---

📝 Walkthrough

Summary by CodeRabbit

• New Features
  • Added explicit algorithm selection for three-way partition tuning, with lookback-specific configuration encapsulated as a dedicated policy.
• Bug Fixes
  • Fixed dispatch and kernel launch configuration to consistently derive threads/items per thread from the selected lookback sub-policy.
• Refactor
  • Reworked the three-way partition policy structure to use an algorithm selector plus algorithm-specific lookback tuning.
• Tests
  • Updated device partition policy tests and expected formatted output to match the revised policy representation.

Walkthrough

Changes

Three-way partition policies now include an explicit lookback algorithm selector and nested tuning configuration. Dispatch, kernels, tests, and benchmarks use the revised policy layout.

Three-way partition policy

Layer / File(s)	Summary
Policy model and selection cub/cub/device/dispatch/tuning/tuning_three_way_partition.cuh	Adds the algorithm enum and lookback sub-policy, updates formatting and comparisons, and wraps selected tuning results.
Dispatch and kernel wiring cub/cub/device/dispatch/dispatch_three_way_partition.cuh, cub/cub/device/dispatch/kernels/kernel_three_way_partition.cuh	Constructs explicit lookback policies and reads launch and agent parameters from active_policy.lookback.
Selectors, tests, and benchmarks cub/test/catch2_test_device_partition_env.cu, cub/test/catch2_test_device_partition_env_api.cu, cub/benchmarks/bench/partition/three_way.cu	Updates custom selectors, policy property expectations, string output, and benchmark initialization.

Possibly related issues

• Prepare tuning policies for lookahead implementations #10075 — Directly covers preparation of the three-way partition tuning policy for algorithm dispatch.

Possibly related PRs

• NVIDIA/cccl#10076 — Applies the same explicit algorithm discriminator and nested policy structure to related partition policies.

Suggested reviewers: gonidelis, pauleonix

---

_{Comment @coderabbitai help to get the list of available commands.}

[github-actions]

🥳 CI Workflow Results

🟩 Finished in 1h 42m: Pass: 100%/287 | Total: 5d 19h | Max: 1h 36m | Hits: 50%/432725

See results here.