[Tile] Mark more fpemu as host device only

[miscco]

We cannot use asm in tile mode

[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: cc127bb6-c3ea-4e79-933c-de51e8ffba3d

📥 Commits

Reviewing files that changed from the base of the PR and between 10d2278 and 8faecdb.

📒 Files selected for processing (25)

• libcudacxx/include/cuda/__fp/fpmp.h
• libcudacxx/include/cuda/__fp/fpmp_impl.h
• libcudacxx/include/cuda/__fp/fpmp_impl_cmp.h
• libcudacxx/include/cuda/__fp/fpmp_impl_cvt.h
• libcudacxx/include/cuda/__fp/fpmp_impl_divsqrt.h
• libcudacxx/include/cuda/__fp/fpmp_impl_muladd.h
• libcudacxx/include/cuda/__fp/fpmp_limits.h
• libcudacxx/test/libcudacxx/cuda/fp/units/fpmp/accuracy.pass.cpp
• libcudacxx/test/libcudacxx/cuda/fp/units/fpmp/accuracy_conv.pass.cpp
• libcudacxx/test/libcudacxx/cuda/fp/units/fpmp/accuracy_mixed.pass.cpp
• libcudacxx/test/libcudacxx/cuda/fp/units/fpmp/api.pass.cpp
• libcudacxx/test/libcudacxx/cuda/fp/units/fpmp/atomic_dd.pass.cpp
• libcudacxx/test/libcudacxx/cuda/fp/units/fpmp/atomic_ff.pass.cpp
• libcudacxx/test/libcudacxx/cuda/fp/units/fpmp/cast.pass.cpp
• libcudacxx/test/libcudacxx/cuda/fp/units/fpmp/charbool.pass.cpp
• libcudacxx/test/libcudacxx/cuda/fp/units/fpmp/copyable.pass.cpp
• libcudacxx/test/libcudacxx/cuda/fp/units/fpmp/cross_prec.pass.cpp
• libcudacxx/test/libcudacxx/cuda/fp/units/fpmp/fpemu_interop.pass.cpp
• libcudacxx/test/libcudacxx/cuda/fp/units/fpmp/int.pass.cpp
• libcudacxx/test/libcudacxx/cuda/fp/units/fpmp/limits.pass.cpp
• libcudacxx/test/libcudacxx/cuda/fp/units/fpmp/lut.pass.cpp
• libcudacxx/test/libcudacxx/cuda/fp/units/fpmp/reduce.pass.cpp
• libcudacxx/test/libcudacxx/cuda/fp/units/fpmp/shfl.pass.cpp
• libcudacxx/test/libcudacxx/cuda/fp/units/fpmp/unary.pass.cpp
• libcudacxx/test/libcudacxx/cuda/fp/units/fpmp/volatile.pass.cpp

---

📝 Walkthrough

Summary by CodeRabbit

• Enhancements

  • Expanded host and device compilation support across extended-precision floating-point constructors, conversions, arithmetic, comparisons, math operations, atomics, shuffles, and numeric limits.
  • Existing function signatures, calculations, and conversion behavior remain unchanged.

• Tests

  • Updated validation coverage to exercise functionality in both host and device contexts.
  • Clarified unsupported tile-based compilation scenarios and associated diagnostics.

Walkthrough

Changes

fpmp public and internal declarations now use _CCCL_HOST_DEVICE_API. fpmp tests now instantiate host/device paths and mark incompatible force-tile configurations as unsupported.

fpmp API migration

Layer / File(s)	Summary
Public fpmp2 contracts libcudacxx/include/cuda/__fp/fpmp.h	Constructors, conversions, operators, arithmetic helpers, math functions, and forwarding overloads use host/device API decorators.
Internal arithmetic and conversion helpers libcudacxx/include/cuda/__fp/fpmp_impl*.h, libcudacxx/include/cuda/__fp/fpmp_limits.h	Internal fp32/fp64 helpers, comparisons, conversions, division, square-root, multiply-add operations, and numeric limits use host/device API decorators.
Host/device test entry points libcudacxx/test/libcudacxx/cuda/fp/units/fpmp/*.pass.cpp	fpmp test helpers and entry points use host/device test annotations.
Force-tile test exclusions libcudacxx/test/libcudacxx/cuda/fp/units/fpmp/*.pass.cpp	Tests document and exclude force-tile configurations that reject host/device calls from tile code.

Possibly related PRs

• NVIDIA/cccl#10543: Retags separate CUDA atomic functions and tests for host/device execution.

Suggested reviewers: gevtushenko, jacobfaib, griwes

---

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

[github-actions]

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

Result: 0 regression row(s), 7 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

TU total compilation

-f total-compilation inclusive --sort total

🟢 TU total compilation — Improvements

Rank	Improvement impact	Selected Δ	Baseline	Current	Event	Matched traces
1	3.045204	-3.045204	9.244492	6.199288	Total Compilation Time: cub/block/specializations/block_reduce_warp_reductions.cuh	1

Direct file processing

-f file-processing exclusive --sort total

🟢 Direct file processing — Improvements

Rank	Improvement impact	Selected Δ	Baseline	Current	Event	Matched traces
1	1.023092	-1.023092	9.537031	8.513939	Processing Header File: libcudacxx/include/cuda/std/__cccl/prologue.h	550
2	0.877442	-0.877442	20.339446	19.462004	Processing Header File: libcudacxx/include/cuda/__device/physical_device.h	94
3	0.567388	-0.567388	3.420610	2.853222	Processing Header File: libcudacxx/include/cuda/std/__cccl/epilogue.h	550
4	0.526731	-0.526731	11.017288	10.490557	Processing Header File: libcudacxx/include/cuda/std/__iterator/concepts.h	437
5	0.310734	-0.310734	5.668859	5.358125	Processing Header File: libcudacxx/include/cuda/std/__iterator/iterator_traits.h	464
6	0.292581	-0.292581	11.335623	11.043042	Processing Header File: libcudacxx/include/cuda/std/__type_traits/type_list.h	477

[github-actions]

🥳 CI Workflow Results

🟩 Finished in 1h 36m: Pass: 100%/115 | Total: 2d 16h | Max: 1h 36m | Hits: 95%/345800

See results here.