Use native 8-bit PTX atomic loads and stores.

[griwes]

Description

Resolves #10783.

This PR changes the PTX codegen of 8-bit atomics to directly load and store with the native 8 bit atomic load and store instructions.

Checklist

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

[coderabbitai]

📝 Walkthrough

Summary by CodeRabbit

• New Features

  • Added support for 8-bit atomic loads and stores across supported memory scopes and ordering modes.
  • Added 8-bit load/store support for generated CUDA operations.

• Bug Fixes

  • Improved handling of small operand sizes, including values up to 16 bits.
  • Corrected conversion behavior when transferring 8-bit values through supported hardware instructions.
  • Refined floating-point restrictions for small-width operations.

Walkthrough

The codegen adds 8-bit load and store templates using 16-bit temporaries. Generated CUDA PTX overloads now cover 8-bit atomic loads and stores across memory orders, volatility, MMIO, and thread scopes.

Changes

8-bit atomic support

Layer / File(s)	Summary
8-bit codegen templates libcudacxx/codegen/generators/definitions.h, libcudacxx/codegen/generators/ld_st.h	The generators accept 8-bit operands, use the "h" constraint for sizes up to 16 bits, and route 8-bit loads and stores through dedicated templates with uint16_t temporaries.
Generated atomic overloads libcudacxx/include/cuda/std/__atomic/functions/cuda_ptx_generated.h	Generated overloads implement 8-bit atomic loads and stores for supported memory orders, volatility, MMIO, and thread scopes.

Assessment against linked issues

Objective	Addressed	Explanation
Use native PTX atomic loads and stores for 8-bit types [#10783]	✅

Possibly related PRs

• NVIDIA/cccl#10722: Shares the same 8-bit atomic code-generation files and changes.

Suggested reviewers: pciolkosz, gevtushenko, davebayer

Mergeability Score: 🔵 Low · up to 237a9

The PR changes generated 8-bit atomic load/store declarations, but some signatures do not follow the required API qualifier, noexcept, and const-reference conventions. This creates a bounded risk for generated interface and compilation consistency, so the PR is mergeable with explicit owner follow-up to correct and regenerate those declarations.

---

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

[coderabbitai]

Actionable comments posted: 1

🧹 Nitpick comments (1)

libcudacxx/include/cuda/std/__atomic/functions/cuda_ptx_generated.h (1)

125-436: 🗄️ Data Integrity & Integration | 🔵 Trivial | ⚡ Quick win

important: Add SASS checks for small owning and non-owning atomics. cuda::atomic<T> with sizeof(T) < 4 must use its 32-bit proxy and emit b32; cuda::atomic_ref<T> must emit b8 for 1-byte operations. Add load and store checks for both paths.

Source: Learnings

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Enterprise

Run ID: 7c502c9b-a16a-4a93-a9ef-8ff4493eb48f

📥 Commits

Reviewing files that changed from the base of the PR and between a5f6c04 and 237a9cf.

📒 Files selected for processing (4)

• libcudacxx/codegen/generators/definitions.h
• libcudacxx/codegen/generators/ld_st.h
• libcudacxx/include/cuda/std/__atomic/functions/cuda_ptx_derived.h
• libcudacxx/include/cuda/std/__atomic/functions/cuda_ptx_generated.h

💤 Files with no reviewable changes (1)

• libcudacxx/include/cuda/std/__atomic/functions/cuda_ptx_derived.h

[github-actions]

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

Result: 0 regression row(s), 2 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	0.675741	-0.675741	5.407366	4.731625	Processing Header File: libcudacxx/include/cuda/std/__cccl/prologue.h	552
2	0.228610	-0.228610	1.682696	1.454086	Processing Header File: libcudacxx/include/cuda/std/__cccl/epilogue.h	552

[github-actions]

🥳 CI Workflow Results

🟩 Finished in 1h 36m: Pass: 100%/117 | Total: 2d 20h | Max: 1h 29m | Hits: 58%/657994

See results here.