Remove most CCCL_GRID_CONSTANT annotations

[bernhardmgruber]

QA reported several regressions caused by the introduction of _CCCL_GRID_CONSTANT in #6642. The regressions include:

• select_flagged regresses in CCCL 3.4 on H100 #9763
• select_if regresses in CCCL 3.4 #9765
• select_unique regresses in CCCL 3.4 on Blackwell and Ampere #9766
• radix_sort_keys regresses in CCCL 3.4 #9768
• reduce_by_key regressed in CCCL 3.4 on Blackwell, Hopper and Ada #9822

The cause of the regression is probably a compiler bug, reported in NVBug 6448961. When the fix is available, we will be able to bring the _CCCL_GRID_CONSTANT annotations back, but this time after carefully benchmarking their impact :)

This PR mainly reverts #6642, but since a lot of time has passed, I needed to do a lot of manual editing as well.

This PR supersedes:

• Remove _CCCL_GRID_CONSTANT from DeviceReduce::ReduceByKey kernels #9852
• Remove _CCCL_GRID_CONSTANT from scan kernel #9854
• Fix radix sort regression #9816

Previous PRs removing annotations:

• Remove _CCCL_GRID_CONSTANT from merge sort kernel parameters #9829
• Remove _CCCL_GRID_CONSTANT from DeviceSelectSweepKernel parameters #9795

This PR will not remove annotations from these kernels, because they have not been added by #6642 and QA did not report any regressions:

• device_partition_find_bound_sorted_values_kernel
• device_find_bound_sorted_values_kernel
• __copy_contiguous_kernel
• __copy_optimized_kernel
• __copy_shared_mem_kernel
• loop (in parallel_for_scope.cuh in STF)
• __kernel_launcher_with_block_size (part of cuda::launch)
• benchmark_kernel (part of nvbench_helper)

Fixes: #9768
Fixes: #9822
Fixes: #9851

[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: 3f778b8c-06e7-4aac-b3cf-2204d66109e2

📥 Commits

Reviewing files that changed from the base of the PR and between 0a21375 and 9b4fe36.

📒 Files selected for processing (1)

• cub/cub/device/dispatch/kernels/kernel_scan.cuh

💤 Files with no reviewable changes (1)

• cub/cub/device/dispatch/kernels/kernel_scan.cuh

---

📝 Walkthrough

Summary by CodeRabbit

• Bug Fixes
  • Improved CUDA kernel build compatibility by standardizing kernel parameter qualifiers across multiple CUB device algorithms (e.g., scan, reduce, sort, merge, top-k, transform, partition, histogram).
  • Preserved existing kernel and dispatch behavior while reducing the likelihood of toolchain-specific compilation issues.
• Performance
  • No runtime behavior changes expected; algorithm and dispatch logic remain the same.

Walkthrough

Changes

CUB CUDA kernel signatures remove _CCCL_GRID_CONSTANT annotations and retain equivalent standard const and pointer-const qualifiers. Kernel bodies, dispatch flow, parameter ordering, and algorithm behavior are unchanged.

CUB kernel entrypoints

Layer / File(s)	Summary
Dispatch kernel entrypoints cub/cub/device/dispatch/dispatch_*.cuh	Adjacent-difference, batch-memcpy, merge, reduce-by-key, RLE, scan-by-key, segmented-sort, and TopK kernel parameters no longer use _CCCL_GRID_CONSTANT.
General algorithm kernels cub/cub/device/dispatch/kernels/kernel_for_each.cuh, kernel_histogram.cuh, kernel_transform.cuh, kernel_unique_by_key.cuh, kernel_three_way_partition.cuh	Kernel scalar, iterator, wrapper, and pointer parameters use standard qualifiers instead of _CCCL_GRID_CONSTANT.
Sorting and selection kernels cub/cub/device/dispatch/kernels/kernel_radix_sort.cuh, kernel_segmented_radix_sort.cuh, kernel_segmented_sort.cuh	Radix-sort and segmented-sort kernel signatures remove grid-constant annotations while preserving parameter types and ordering.
Reduction and scan kernels cub/cub/device/dispatch/kernels/kernel_reduce.cuh, kernel_scan.cuh, kernel_segmented_reduce.cuh, kernel_segmented_scan.cuh	Reduction and scan kernel entrypoints replace _CCCL_GRID_CONSTANT with ordinary const and pointer-const declarations.

Possibly related PRs

• NVIDIA/cccl#9740: Overlaps the deferred single-tile reduction kernel signature changes.
• NVIDIA/cccl#9852: Overlaps the DeviceReduceByKeyKernel qualifier changes.
• NVIDIA/cccl#9854: Overlaps the scan kernel qualifier changes.

Suggested reviewers: gevtushenko, pauleonix, srinivasyadav18, jacobfaib

---

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

[coderabbitai]

🧹 Nitpick comments (1)

cub/cub/device/dispatch/kernels/kernel_scan.cuh (1)

205-211: 📐 Maintainability & Code Quality | 🔵 Trivial | 💤 Low value

suggestion: The #if _CCCL_CUDACC_AT_LEAST(12, 8) / #endif block at Lines 206-208 is now empty — it previously guarded a _CCCL_GRID_CONSTANT annotation on init_value that has been removed. Leaving a no-op conditional is confusing dead code; consider dropping the guard entirely since it no longer wraps anything.

🧹 Proposed cleanup

   const ScanOpT scan_op,
-// nvcc 12.0 gets stuck compiling some TUs like `cub.bench.scan.exclusive.sum.base`, so only enable for newer versions
-#if _CCCL_CUDACC_AT_LEAST(12, 8)
-
-#endif // _CCCL_CUDACC_AT_LEAST(12, 8)
   const InitValueT init_value,

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Enterprise

Run ID: 525248ef-8250-440b-a044-a58937d587f4

📥 Commits

Reviewing files that changed from the base of the PR and between 5f23d49 and 0a21375.

📒 Files selected for processing (20)

• cub/cub/device/dispatch/dispatch_adjacent_difference.cuh
• cub/cub/device/dispatch/dispatch_batch_memcpy.cuh
• cub/cub/device/dispatch/dispatch_merge.cuh
• cub/cub/device/dispatch/dispatch_reduce_by_key.cuh
• cub/cub/device/dispatch/dispatch_rle.cuh
• cub/cub/device/dispatch/dispatch_scan_by_key.cuh
• cub/cub/device/dispatch/dispatch_segmented_sort.cuh
• cub/cub/device/dispatch/dispatch_topk.cuh
• cub/cub/device/dispatch/kernels/kernel_for_each.cuh
• cub/cub/device/dispatch/kernels/kernel_histogram.cuh
• cub/cub/device/dispatch/kernels/kernel_radix_sort.cuh
• cub/cub/device/dispatch/kernels/kernel_reduce.cuh
• cub/cub/device/dispatch/kernels/kernel_scan.cuh
• cub/cub/device/dispatch/kernels/kernel_segmented_radix_sort.cuh
• cub/cub/device/dispatch/kernels/kernel_segmented_reduce.cuh
• cub/cub/device/dispatch/kernels/kernel_segmented_scan.cuh
• cub/cub/device/dispatch/kernels/kernel_segmented_sort.cuh
• cub/cub/device/dispatch/kernels/kernel_three_way_partition.cuh
• cub/cub/device/dispatch/kernels/kernel_transform.cuh
• cub/cub/device/dispatch/kernels/kernel_unique_by_key.cuh

[github-actions]

🥳 CI Workflow Results

🟩 Finished in 3h 34m: Pass: 100%/287 | Total: 6d 05h | Max: 2h 12m | Hits: 43%/525623

See results here.

[github-actions]

Backport failed for branch/3.4.x, because it was unable to cherry-pick the commit(s).

Please cherry-pick the changes locally and resolve any conflicts.

git fetch origin branch/3.4.x
git worktree add -d .worktree/backport-9858-to-branch/3.4.x origin/branch/3.4.x
cd .worktree/backport-9858-to-branch/3.4.x
git switch --create backport-9858-to-branch/3.4.x
git cherry-pick -x d8fd76b44b0e4de1ae6e1b4d606fe72636b10df4