CUDA to AMD / software compatibility / 14 November 2026

HIPIFY #2811: warning earlier about unsupported CUDA headers

Locally validated candidate source-level repair. Not yet merged upstream and not a claim that unsupported CUDA primitives now run on AMD.

Problem

HIPIFY could preserve includes for cuda/barrier, cuda/pipeline, cuda/semaphore and cuda/ptx without warning that AMD libhipcxx does not support these features. This hides a known incompatibility until downstream compilation.

What was changed

A small candidate patch adds conversion-time warnings in both hipify-perl and hipify-clang. Includes are deliberately left unchanged. This diagnoses a problem; it does not implement replacement CUDA APIs.

Verified results

The original baseline-versus-patch reproducer cases show the expected warning in the candidate and no warning in the baseline. Supported-header controls remain warning-free.

Filtered LLVM lit suite: candidate 150/150 passing; original upstream baseline 149/149 passing. The unfiltered suites each retain an identical, pre-existing cuDNN synthetic test failure (candidate 150/151; baseline 149/150). We do not present the filtered run as an unconditional full-suite pass.

Test environment and limits

Developed on an AMD Ryzen AI MAX+ 395 Strix Halo with Radeon 8060S, running Linux in Docker/WSL2, CUDA 12.8.93 and LLVM 20.1.2. Tests validate host-side source conversion, not GPU kernel execution. The reporter's original ROCm 7.1.0, LLVM 20.1.8 and gfx942 target configuration has not been independently reproduced.

Publication status

Locally qualified candidate: the portable source patch and hashed evidence logs are preserved. An upstream pull request and independent external reproduction remain outstanding. The full white paper is now published on Zenodo (DOI 10.5281/zenodo.24230766).

Original upstream GitHub issue #2811 · AMD libhipcxx API limitations

Author: Nigel Hutchinson / NH Applied. Report: version 1.0, 14 November 2026.