Make stream and memory resource explicit in hyperloglog APIs

[PointKernel]

Description

This PR removes the default CUDA stream and default memory resource from cuda::experimental::cuco::hyperloglog and hyperloglog_ref, following the cuda::buffer design adopted for fixed_capacity_map in #9719:

• All hyperloglog constructors now take cuda::stream_ref as the first argument and the memory resource as the second, matching cuda::buffer's argument order. The constructors that silently picked the default memory pool of device 0 are removed.
• All host bulk APIs (clear, clear_async, add, add_async, merge, merge_async, estimate) on both hyperloglog and hyperloglog_ref take the stream as the first parameter, with no default.

It also applies the conventions established during the #7705 review to the hyperloglog headers:

• _CCCL_HOST_API/_CCCL_HOST_DEVICE_API annotations on all hyperloglog member functions, which were previously unannotated constexpr.
• _CCCL_TEMPLATE/_CCCL_REQUIRES instead of raw enable_if SFINAE for the constrained clear(_CG) and merge(_CG, ...) device overloads.
• #if !_CCCL_COMPILER(NVRTC) guard around the host-only hyperloglog class and #if _CCCL_CUDA_COMPILATION() around the device kernels.
• Granular per-symbol includes instead of the bulk <cuda/experimental/container.cuh>, <cuda/experimental/memory_resource.cuh>, and <cuda/experimental/stream.cuh> headers.
• West const and no this-> prefixes for member access.

Checklist

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

[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.

[PointKernel]

/ok to test 84e48f4

[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: 3a525955-5815-4433-8139-cbabf54ef272

📥 Commits

Reviewing files that changed from the base of the PR and between e981a78 and 669032e.

📒 Files selected for processing (2)

• cudax/include/cuda/experimental/__cuco/detail/hyperloglog/kernels.cuh
• cudax/include/cuda/experimental/__cuco/hyperloglog.cuh

🚧 Files skipped from review as they are similar to previous changes (2)

• cudax/include/cuda/experimental/__cuco/detail/hyperloglog/kernels.cuh
• cudax/include/cuda/experimental/__cuco/hyperloglog.cuh

---

📝 Walkthrough

Summary by CodeRabbit

• New Features
  • HyperLogLog (hyperloglog and hyperloglog_ref) now requires explicit CUDA stream selection for construction and stream-enqueued operations, with ::cuda::stream_ref as the first argument.
• Breaking Changes
  • Removed default-stream overloads and reordered stream parameters for clear, add, merge, and estimate; callers must pass a stream explicitly.
• Tests
  • Updated HyperLogLog tests to validate the stream- and memory-resource-aware APIs across cases (including pinned and device-side paths).

Walkthrough

HyperLogLog host and reference APIs now require explicit ::cuda::stream_ref arguments, with stream-first ordering across construction, clear/add/merge/estimate calls. The implementation and kernel headers also received small cleanup and guard changes, and the tests were updated to use stream-aware construction and memory pools.

Changes

HyperLogLog Stream API Rework

Layer / File(s)	Summary
hyperloglog_ref stream-first API and constraints cudax/include/cuda/experimental/__cuco/hyperloglog_ref.cuh	clear, add, merge, and estimate now take ::cuda::stream_ref first with no default stream arguments; cooperative-group overloads use CCCL_REQUIRES; related includes were updated.
hyperloglog host wrapper stream-first API cudax/include/cuda/experimental/__cuco/hyperloglog.cuh	Constructors and host methods now require explicit stream arguments first, accessors and sketch helpers gained host/device annotations, precision helpers were requalified, and internal include dependencies were switched.
Impl and kernel cleanup, guard adjustments cudax/include/cuda/experimental/__cuco/detail/hyperloglog/default_policy.cuh, cudax/include/cuda/experimental/__cuco/detail/hyperloglog/finalizer.cuh, cudax/include/cuda/experimental/__cuco/detail/hyperloglog/hyperloglog_impl.cuh, cudax/include/cuda/experimental/__cuco/detail/hyperloglog/kernels.cuh	Redundant this-> qualifiers are removed, one include is dropped, a local declaration style changes, and kernel preprocessor guards plus a toolkit-conditional remainder branch are restructured.
Tests updated for stream-aware estimator cudax/test/cuco/hyperloglog/test_hyperloglog.cu	Tests now construct estimators with explicit streams and memory pools, pass streams into add/clear/estimate calls, and use a shared scaled_index functor in the Spark parity case.

Suggested reviewers: bernhardmgruber

---

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

[coderabbitai]

Caution

Some comments are outside the diff and can’t be posted inline due to platform limitations.

⚠️ Outside diff range comments (1)

cudax/include/cuda/experimental/__cuco/hyperloglog.cuh (1)

154-169: 🎯 Functional Correctness | 🟠 Major | ⚡ Quick win

important: constructor still returns before the initial clear completes The three constructors enqueue clear_async(__stream) and return without synchronizing, so the @note claiming stream sync is stale. Either call clear(__stream) here or update the note to say construction is asynchronous and the caller must synchronize before using the object.

🧹 Nitpick comments (3)

cudax/test/cuco/hyperloglog/test_hyperloglog.cu (1)

79-88: 📐 Maintainability & Code Quality | 🔵 Trivial | 💤 Low value

suggestion: the ::cuda::stream{device_ref{0}} + device_default_memory_pool(device_ref{0}) setup is repeated identically across all five tests. Could extract into a shared fixture/helper to reduce duplication, but not blocking.

Also applies to: 122-146, 189-199, 224-230, 251-258

cudax/include/cuda/experimental/__cuco/hyperloglog_ref.cuh (1)

231-231: 📐 Maintainability & Code Quality | 🔵 Trivial | ⚡ Quick win

suggestion: Stray // Review: note sits inside the //! doxygen block and will bleed into the generated docs / is a leftover artifact. Drop it (or convert to a tracked TODO).

-  //! `@throw` If sketch_bytes() != __other.sketch_bytes()
-  //!
-  // Review: For host-side merges, consider validating `__other` is on the same device.
-  //! `@tparam` _OtherScope Thread scope of `other` estimator
+  //! `@throw` If sketch_bytes() != __other.sketch_bytes()
+  //!
+  //! `@tparam` _OtherScope Thread scope of `other` estimator

cudax/include/cuda/experimental/__cuco/hyperloglog.cuh (1)

159-166: 📐 Maintainability & Code Quality | 🔵 Trivial | 💤 Low value

suggestion: add #include <cuda/__utility/no_init.h> here; ::cuda::no_init is used directly and shouldn’t depend on device_memory_pool.h for availability.

Source: Coding guidelines

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Enterprise

Run ID: 87f29846-e10b-4203-be12-9e09a637769c

📥 Commits

Reviewing files that changed from the base of the PR and between 48ff8d1 and 84e48f4.

📒 Files selected for processing (7)

• cudax/include/cuda/experimental/__cuco/detail/hyperloglog/default_policy.cuh
• cudax/include/cuda/experimental/__cuco/detail/hyperloglog/finalizer.cuh
• cudax/include/cuda/experimental/__cuco/detail/hyperloglog/hyperloglog_impl.cuh
• cudax/include/cuda/experimental/__cuco/detail/hyperloglog/kernels.cuh
• cudax/include/cuda/experimental/__cuco/hyperloglog.cuh
• cudax/include/cuda/experimental/__cuco/hyperloglog_ref.cuh
• cudax/test/cuco/hyperloglog/test_hyperloglog.cu

[github-actions]

😬 CI Workflow Results

🟥 Finished in 1h 09m: Pass: 98%/55 | Total: 5h 20m | Max: 1h 09m | Hits: 93%/38334

See results here.