[0z5a]

closes #12605

struct __fpemu_uint128 in cuda/__fp/fpemu_impl.h declared its default constructor as ::cuda::__fpemu_uint128() = default;. A qualified name is not valid for a constructor declared inside its class, so the emulated-int128 branch does not compile. NVRTC takes that branch whenever -device-int128 is not passed, which breaks #include <cuda/fpemu> with default NVRTC options. The qualifier came in with the promotion from cuda::experimental in #11956.

This PR drops the qualifier. It also adds fpemu/emulated_int128.pass.cpp, which defines _CCCL_FPEMU_FORCE_EMULATED_INT128 and runs the core.pass.cpp operations (mul, add, div, sub, fma). That way the emulated branch is built in CI; the NVRTC lane always passes -device-int128, so it never takes this branch today.

Checked locally:

• The issue's snippet compiles with NVRTC 13.0.88, 13.3.33 and 13.5.37 without -device-int128 (it failed on all three before).
• With nvcc 13.3 + GCC and _CCCL_FPEMU_FORCE_EMULATED_INT128, all tests in test/libcudacxx/cuda/fp/units/fpemu/ build and pass on the host.
• The new test fails to compile before the change and passes after. I built it by hand with nvcc rather than through lit, and did not run the emulated path on a device.

Checklist:

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

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[coderabbitai]

No actionable comments were generated in the recent review.

Summary

Summary by CodeRabbit

• Tests
  • Added coverage for emulated floating-point arithmetic, checking multiplication, addition, division, subtraction, and fused multiply-add results against native results.
  • The tests also verify that calculated values stay within a small tolerance.

The emulated __fpemu_uint128 default constructor no longer uses a namespace-qualified name. A host/device test forces emulated 128-bit integer support and compares five fp64 arithmetic results with native double results.

Changes: fp64 emulation with emulated integers

Layer / File(s)	Summary
Constructor fix and arithmetic test libcudacxx/include/cuda/__fp/fpemu_impl.h, libcudacxx/test/libcudacxx/cuda/fp/units/fpemu/emulated_int128.pass.cpp	The default constructor declaration drops the ::cuda:: qualification. The test checks multiplication, addition, division, subtraction, and fused multiply-add against native double results when emulated 128-bit integers are forced.

Assessment against linked issues

Objective	Addressed	Explanation
Fix <cuda/fpemu> compilation with NVRTC default options when emulated 128-bit integers are selected [#12605]	✅

Suggested reviewers: akolesov-nvidia

Change: Bug fix · Severity of issue fixed: Medium

Merge Risk: Minimal · up to 52656

This fixes compilation of the cuda/fpemu header under NVRTC without -device-int128 and adds a test for the emulated path. No merge-blocking risk was found.

- [ ] Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

Recent review info

Run configuration

• Source mode: Auto
• Repository: NVIDIA/cccl
• Branch: main
• File: .coderabbit.yaml
• Commit: 5d63af4
• Layers: Repository YAML
• Review profile: CHILL
• Plan: Enterprise
• Run ID: 5508f4ac-2784-4e41-b782-e5cde6209bac

Commits

Reviewing files that changed from the base of the PR and between 5d63af4 and 52656d2.

Files selected for processing: 2

• libcudacxx/include/cuda/__fp/fpemu_impl.h
• libcudacxx/test/libcudacxx/cuda/fp/units/fpemu/emulated_int128.pass.cpp

Included review availability: This review used your included allowance. Your plan provides up to 12 included reviews per hour; 11 remain after this review.