Duplicate radix sort dispatch to deprecate DispatchRadixSort

[bernhardmgruber]

• No SASS changes for cub.bench.radix_sort.pairs.base on SM75;80;90;100

[copy-pr-bot]

Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually.

Contributors can view more details about this message here.

[coderabbitai]

📝 Walkthrough

Summary by CodeRabbit

• Bug Fixes

  • Improved radix sort dispatch behavior to better handle different input/output buffer scenarios.
  • Optimized the keys-only path so unnecessary value handling is removed at compile time.

• Chores

  • Marked the older radix sort dispatch interface as deprecated in favor of the newer device radix sort API.

Walkthrough

DispatchRadixSort gets its deprecation attribute enabled. InvokeCopy gains if constexpr on the values-copy path. A new pass_config struct and a new dispatch_impl struct are introduced under detail::radix_sort, replacing the fake_policy/DispatchRadixSort::__invoke forwarding path. The overwrite parameter is renamed can_overwrite_source_buffer.

Radix Sort Dispatch Refactor

Layer / File(s)	Summary
Deprecation and if constexpr fix cub/cub/device/dispatch/dispatch_radix_sort.cuh	DispatchRadixSort deprecation attribute is uncommented; values-copy guard changed to if constexpr (!KEYS_ONLY).
pass_config struct cub/cub/device/dispatch/dispatch_radix_sort.cuh	New pass_config<UpsweepKernelT, ScanKernelT, DownsweepKernelT, OffsetT> struct with init(...) that configures upsweep/scan/downsweep kernels and computes max_downsweep_grid_size and even_share.
dispatch_impl and dispatch wiring cub/cub/device/dispatch/dispatch_radix_sort.cuh	New dispatch_impl struct implements invoke_single_tile, invoke_pass/invoke_passes, invoke_onesweep, and top-level invoke(PolicyGetter). detail::radix_sort::dispatch renames the overwrite parameter to can_overwrite_source_buffer and constructs dispatch_impl directly instead of using the fake_policy path.

---

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

[coderabbitai]

🧹 Nitpick comments (1)

cub/cub/device/dispatch/dispatch_radix_sort.cuh (1)

1413-1429: 📐 Maintainability & Code Quality | 🔵 Trivial | ⚡ Quick win

suggestion: mark the new non-mutated locals const where applicable, e.g. pass_bits, pass_spine_length, max_grid_size, spine_length, num_bits, is_num_passes_odd, max_alt_passes, alt_end_bit, num_passes, num_portions, max_num_blocks, value_size, histogram_kernel, portion_num_items, num_blocks, and onesweep_kernel. As per coding guidelines, “All variables that are not modified must use const.”

Also applies to: 1588-1623, 1703-1708, 1766-1766, 1851-1858, 1897-1897

Source: Coding guidelines

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Enterprise

Run ID: 3f483d68-1ebf-4929-8706-48d14f76152a

📥 Commits

Reviewing files that changed from the base of the PR and between 47e0576 and 7f20e9a.

📒 Files selected for processing (1)

• cub/cub/device/dispatch/dispatch_radix_sort.cuh

[github-actions]

😬 CI Workflow Results

🟥 Finished in 1h 27m: Pass: 99%/287 | Total: 6d 17h | Max: 1h 05m | Hits: 46%/462539

See results here.