Refactor DeviceReduce dispatch logic

[bernhardmgruber]

No description provided.

[coderabbitai]

📝 Walkthrough

Summary by CodeRabbit

• Refactor
  • Reworked internal device reduction logic to centralize min/max handling, standardize determinism fallbacks for sum/min/max operations, and simplify dispatch paths.
  • No public API changes; results are now more consistent and reliability of reduction operations under different environment settings has been improved.

suggestion:

Walkthrough

This PR refactors determinism dispatch for DeviceReduce: reduce_impl now uses compile-time determinism branches that call detail::dispatch_with_env_and_tuning; determinism fallback computation was moved into __transform_reduce; min/max behavior was centralized into __minmax_reduce; Sum/Min/Max env overloads now delegate to those helpers.

Changes

Device reduce determinism dispatch refactoring

Layer / File(s)	Summary
reduce_impl determinism dispatch cub/cub/device/device_reduce.cuh	reduce_impl dispatch switched from lambda-based tuning/query to explicit if constexpr branches on determinism tags (gpu_to_gpu, not_guaranteed, default), each calling the corresponding dispatch routine via detail::dispatch_with_env_and_tuning.
Transform-reduce determinism handling cub/cub/device/device_reduce.cuh	__transform_reduce now computes default_determinism_t from environment requirements, defines determinism-dependent fallback predicates for integral types and float/double plus/min/max operators, and refines the "4B or greater" condition to use sizeof(AccumT).
Min/max reduce helper introduction cub/cub/device/device_reduce.cuh	New private __minmax_reduce helper centralizes min/max env determinism handling: rejects gpu_to_gpu, forces run_to_run, computes output limits, and forwards to reduce_impl with identity transform.
Public API rewiring cub/cub/device/device_reduce.cuh	Sum env overload simplified to call __transform_reduce with plus/identity. Min and Max env overloads compute OutputT and limits, then delegate to __minmax_reduce, removing duplicated determinism bodies.

Suggested reviewers

• shwina
• pauleonix

---

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

[coderabbitai]

Actionable comments posted: 2

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Enterprise

Run ID: 306f06d3-aefb-42db-b33e-5527015ec795

📥 Commits

Reviewing files that changed from the base of the PR and between 8b47911 and b295493.

📒 Files selected for processing (1)

• cub/cub/device/device_reduce.cuh

[coderabbitai]

🧹 Nitpick comments (1)

cub/cub/device/device_reduce.cuh (1)

114-161: important: This changes dispatch and determinism selection on a Device* algorithm path. Please do the required before/after SASS comparison and run the CUB benchmarks before merge; otherwise codegen and throughput regressions in the new dispatch split stay unvalidated.

As per coding guidelines, **/cub/**/device*.{cu,cuh,h}: Verify no SASS code generation changes occur for Device* algorithms in CUB by comparing generated SASS output before and after changes; Run benchmark tests using the CUB Benchmarks framework when modifying Device* algorithms to verify no performance regressions occur.

Also applies to: 173-240, 244-270, 549-551

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Enterprise

Run ID: 6fde9ca1-11b8-4a40-aa81-d6719380fcf5

📥 Commits

Reviewing files that changed from the base of the PR and between b295493 and 96c47e0.

📒 Files selected for processing (1)

• cub/cub/device/device_reduce.cuh

[github-actions]

🥳 CI Workflow Results

🟩 Finished in 2h 33m: Pass: 100%/285 | Total: 11d 04h | Max: 2h 32m | Hits: 21%/914257

See results here.

[tpn]

Reviewed the DeviceReduce dispatch refactor. No blocking issues from me; the CI summary is green.

[NaderAlAwar]

Critical: this breaks Sum(uint8_t*, uint8_t*, ...) so that it doesn't compile anymore. Prior to this change, we fell back to run to run determinism for output types < 4 bytes. But __transform_reduce checks AccumT instead of OutputT. For uint8_t, AccumT is int due to integer promotion, so the kernel will attempt to do atomicAdd(unsigned char*, int) which doesn't compile

[NaderAlAwar]

This issue already existed for Reduce() before this PR, but sum had its own detection logic

Thinking about this some more, it is probably okay to merge this PR and I'll create a fix for __transform_reduce to check OutputT instead of AccumT, or you can driveby fix it if you want, but the fix is slightly involved and I would like to add tests to verify the behavior. We basically need to check two things:

1. OutputT == AccumT
2. atomicAdd() is supported for that type. This is mostly handled by is_4b_or_greater but it occurs to me now that there is no overload for long long.

I'm going to approve, let me know if you prefer fixing it here or I can fix it after this gets merged