libcu++: respect atomic scopes in pointer arithmetic operations.

[griwes]

Description

It seems that we've had pointer arithmetic go through the system scope. This PR fixes that.

Resolves #10808.

Checklist

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

[coderabbitai]

📝 Walkthrough

Summary by CodeRabbit

• Bug Fixes

  • Fixed pointer atomic additions and subtractions so they honor the requested thread scope instead of always using system scope.
  • Improved correctness for block-, device-, and system-scoped operations.

• Tests

  • Added coverage for non-volatile pointer atomic addition and subtraction across supported thread scopes.
  • Added verification that relaxed operations generate the appropriate scope-specific GPU instructions.

Walkthrough

Pointer atomic fetch_add and fetch_sub now forward the atomic scope to dispatch. New CUDA kernels and PTX checks validate block, device, and system scope code generation for non-volatile pointer operations.

Changes

Pointer atomic scope handling

Layer / File(s)	Summary
Scope-aware pointer arithmetic dispatch libcudacxx/include/cuda/std/__atomic/api/common.h	Pointer atomic addition and subtraction dispatch use _Sco instead of a fixed system-scope tag.
Scope-specific pointer atomic code generation libcudacxx/test/atomic_codegen/atomic_add_non_volatile.cu, libcudacxx/test/atomic_codegen/atomic_sub_non_volatile.cu	New relaxed pointer arithmetic kernels and PTX checks cover CTA, GPU, and system scopes.

Assessment against linked issues

Objective	Addressed	Explanation
Pointer arithmetic atomics use the scope of the atomic instead of always using system scope [#10808]	✅

Possibly related PRs

• NVIDIA/cccl#10722: Adds related pointer atomic scope handling and code-generation test infrastructure.

Suggested reviewers: pciolkosz, wmaxey

Mergeability Score: ⚪ Minimal · up to d5304

The PR updates atomic pointer arithmetic to honor the intended atomic scope; the remaining test-only const-correctness cleanup does not affect runtime behavior. No actionable merge-blocking risk remains after normal checks.

---

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

[coderabbitai]

Actionable comments posted: 1

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Enterprise

Run ID: bf464693-4e73-4a77-8eb8-dbc4b288cf2a

📥 Commits

Reviewing files that changed from the base of the PR and between 48c8219 and d530404.

📒 Files selected for processing (3)

• libcudacxx/include/cuda/std/__atomic/api/common.h
• libcudacxx/test/atomic_codegen/atomic_add_non_volatile.cu
• libcudacxx/test/atomic_codegen/atomic_sub_non_volatile.cu

[github-actions]

⏱️ CCCL compile-time benchmark comparison: Public headers compile-time bench

Result: 0 regression row(s), 1 improvement row(s) above threshold.

Run	Value
Config	public-headers-gcc13
Baseline	origin/main
Preset	all-dev
Targets	cub.headers.base, thrust.cpp.cuda.headers.base, libcudacxx.test.public_headers
GPU / launch args	rtx2080 / --cuda 13.3 --host gcc13

Artifacts: reports and traces

Direct file processing

-f file-processing exclusive --sort total

🟢 Direct file processing — Improvements

Rank	Improvement impact	Selected Δ	Baseline	Current	Event	Matched traces
1	0.625751	-0.625751	5.407631	4.781880	Processing Header File: libcudacxx/include/cuda/std/__cccl/prologue.h	552

[github-actions]

🥳 CI Workflow Results

🟩 Finished in 2h 02m: Pass: 100%/117 | Total: 2d 18h | Max: 1h 29m | Hits: 59%/636436

See results here.