AMD GPU Support via HIP/ROCm

[ptheywood]

Adds AMD GPU support via HIP / ROCm

Caution

This is still a WIP - do not review, merge or expect CI to be happy

It will be rebased many times, and may become the base branch for other AMD/HIP/ROCm/clang related PRs so they can all be merged into master in a single go.

• Todo: Add the actual PR text
• Abstracts (most, some is not portable) HIP/CUDA macro usage into flamegpu::detail::gpu namespace.
  • This is also a step towards Manylinux compliant wheels (dlopen) #647, as dlopen'd functions should only need to be within that file (excluding vis interop, but vis is CUDA only currently)
• Closes Initial AMD GPU Support (ROCm/HIP) #1367
• Closes AMD + GLM #1376
• Closes AMD ROCTx support #1377

Decisions to make during review

• FLAMEGPU_GPU=CUDA/HIP/OFF - Best I've come up with, but not super keen on this cmake option name.
  • support lower case values as well (i.e. make it upper case before validation)?
  • BACKEND, VENDOR (prolyl not), PLATFORM, API
• FLAMEGPU_CURAND_ENGINE - deprecate and rename? Its a bit grim to do with it being a cache var not an option.
• flamegpu::detail::gpu::Stream_t is used in classes which belong in the public API (but not sure if it is actaulyl part of the true public API, i.e. users arent' supposed to provide those values themselves?) Should this be flamegpu::gpu::Stream_t instead?
• Rename CUDASimulation -> GPUSimulation etc? with a deprecated type alias?
  • This is not too bad just for class names, as using X [[deprecated(("msg")]] = Y should make that trivial, but CUDAConfig means a deprecated member is needed too to fully make the change, so is just a decent chunk more effort.
• Ensure consistency in use of new macros? i.e. defined(FLAMEGPU_USE_CUDA) vs ifdef FLAMEGPU_USE_CUDA. Maybe do HIP first and fallback to the CUDA version so intellisense doesn't get offended?
• Update Docs repo accordingly
• TODO: DO NOT MERGE in test suite, where wrapper function changes happened but I didn't update the test suite.
• Still don't understand why the wrapper helper is needed for agent functions but not for agent function conditions....
• Fully udpated changelog from this PR (more breaking changes / deprecations not yet included)
• Telemetry NVCCVersion. Or maybe use a diff field name anyway?

---

As of 2026-06-10 tests pass with:

• CUDA 13.2 on linux
• HIP 7.0 on linux
• CUDA 12.4 on windows