[FEA]: semaphore not supported in latest release

[jamesETsmith]

Is this a duplicate?

• I confirmed there appear to be no duplicate issues for this request.

Is your feature request related to a problem? Please describe.

With the recent release, I see that semaphores are no longer supported. Can you give more details about 1) why it was removed going from v2.2.0 -> release/rocmds-25.10 and 2) when we'll get support back?

libhipcxx/include/cuda/semaphore

Lines 36 to 39 in 9ac637b

	// NOTE(HIP/AMD): semaphore is not supported on AMD hardware
	#ifdef __HIP_PLATFORM_AMD__
	#error semaphore is not supported on AMD hardware and should not be included
	#endif

Describe the solution you'd like

Implementation of <cuda/semaphore>

Describe alternatives you've considered

I'm going to try and revert to v2.2.0, but I'm worried this will cause problems with the latest versions of hipcollections

Additional context

This is needed for ROCm/dgl

[jamesETsmith]

It looks like this will cause problems with the latest hipcollections https://github.com/ROCm/hipCollections/blob/77f2e84e4b6f9d667d8733967301dbdad4e795de/CMakeLists.txt#L283-L285

[psamfass]

Hi @jamesETsmith,

Thanks for opening this issue.
As noted in the libhipcxx 2.2.x README under limitations, cuda::std::semaphore was not officially supported in 2.2.0 (see https://github.com/ROCm/libhipcxx/tree/release/2.2.x?tab=readme-ov-file#limitationsunsupported-featuresapis). In libhipcxx 2.7.0, we added explicit checks and errors to block use of these unsupported APIs.

We’d appreciate learning more about your intended use cases and the specific synchronization patterns you have in mind. That context will help us assess feasibility and prioritize a potential implementation.

[jamesETsmith]

Thanks @psamfass! I didn't go through the readme carefully enough so I missed that.

As for our use case, DGL uses cuda::std::counting_semaphore in a key part of the code: https://github.com/ROCm/dgl/blob/d9f14ea373aef0764cb91d1bb1b04a6e7beb3abe/graphbolt/src/cnumpy.h#L51.

It's used to control access to a limit number of spots in an IO queue used during data loading and preprocessing.

EDIT: After bit more reading, I'm a bit perplexed here why upstream used the cuda::std::counting_semaphore here instead of std::counting_semaphore from stl. The IO queue it's managing is just running on the CPU/host. I'll keep poking around and report back if I find anything else interesting