Use result policies for MGMN algorithms

[Jacobfaib]

Description

Use the new result policies in MGMN algorithms. Currently only support what the original algorithm implemented, as users open issues, support for other forms may be added later on

Checklist

• New or existing tests cover these changes.
• The documentation is up to date with these changes.

[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: 44eb1e7e-7086-4123-a9db-1acc0b8e66ea

📥 Commits

Reviewing files that changed from the base of the PR and between 58ec695 and dc025ca.

📒 Files selected for processing (14)

• cudax/include/cuda/experimental/__multi_gpu/algorithm/reduce/reduce.h
• cudax/include/cuda/experimental/__multi_gpu/algorithm/scan/scan.h
• cudax/test/multi_gpu/algorithms/exclusive_scan/range_basic.cu
• cudax/test/multi_gpu/algorithms/exclusive_scan/range_defaults.cu
• cudax/test/multi_gpu/algorithms/exclusive_scan/single_comm_basic.cu
• cudax/test/multi_gpu/algorithms/exclusive_scan/single_comm_defaults.cu
• cudax/test/multi_gpu/algorithms/inclusive_scan/range_basic.cu
• cudax/test/multi_gpu/algorithms/inclusive_scan/range_defaults.cu
• cudax/test/multi_gpu/algorithms/inclusive_scan/single_comm_basic.cu
• cudax/test/multi_gpu/algorithms/inclusive_scan/single_comm_defaults.cu
• cudax/test/multi_gpu/algorithms/reduce/range_basic.cu
• cudax/test/multi_gpu/algorithms/reduce/range_defaults.cu
• cudax/test/multi_gpu/algorithms/reduce/single_comm_basic.cu
• cudax/test/multi_gpu/algorithms/reduce/single_comm_defaults.cu

---

📝 Walkthrough

Summary by CodeRabbit

• New Features
  • Multi-GPU scan operations now accept an explicit distributed execution policy for distributed scans (range-based and single-communicator overloads).
  • Multi-GPU reduce operations now accept an explicit broadcasted result policy (range-based and single-communicator overloads).
• Documentation
  • Updated API docs to describe the new policy parameter and supported modes.
• Tests
  • Adjusted multi-GPU scan and reduction test coverage and examples to use the explicit policies.

Walkthrough

Changes

Multi-GPU reduce and scan APIs now accept explicit result-policy arguments. Reduce supports broadcasted, while scan supports distributed; affected tests pass the corresponding policy tags.

Multi-GPU result-policy APIs

Layer / File(s)	Summary
Reduce policy API cudax/include/cuda/experimental/__multi_gpu/algorithm/reduce/reduce.h	Reduce overloads accept and forward result policies, with support restricted to broadcasted_t.
Scan policy API cudax/include/cuda/experimental/__multi_gpu/algorithm/scan/scan.h	Exclusive and inclusive scan overloads forward result policies to distributed scan, restricted to distributed_t.
Reduce policy call sites cudax/test/multi_gpu/algorithms/reduce/*	Reduce tests pass cudax::broadcasted across range, single-communicator, and default-argument cases.
Scan policy call sites cudax/test/multi_gpu/algorithms/{exclusive_scan,inclusive_scan}/*	Scan tests pass cudax::distributed across range, single-communicator, documentation-example, and default-argument cases.

Possibly related PRs

• NVIDIA/cccl#9796: Introduced the multi-GPU exclusive_scan overloads updated here.
• NVIDIA/cccl#9827: Added multi-GPU inclusive-scan support in the same implementation area.
• NVIDIA/cccl#9862: Introduced or revised the result-policy infrastructure used by these APIs.

Suggested reviewers: caugonnet, davebayer, pciolkosz, naderalawar

---

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

[Jacobfaib]

@coderabbitai full review

[coderabbitai]

✅ Action performed

Full review finished.

[github-actions]

🥳 CI Workflow Results

🟩 Finished in 46m 56s: Pass: 100%/57 | Total: 7h 39m | Max: 46m 45s | Hits: 87%/43680

See results here.