find no HIP devices while excuate acpp-info on Ubuntu22.04 with rocm-5.7.1 and AMD Readon 7900xtx device

[jiuxiaocloud]

find no HIP devices while excuate acpp-info on Ubuntu22.04 with rocm-5.7.1 and AMD Readon 7900xtx device

[jiuxiaocloud]

Is rocm-5.7.1 with llvm-17.0 not supportted now ?

[gogo2]

What a timing.. I've just encountered the same problem, but in slightly different conditions.

What's different in my case:

• My environment is in an apptainer container (based on official rocm container with ubuntu 22.04 and rocm 5.4.2)
• I use LLVM 16
• Device not found is MI100 (listed correctly by rocm-smi)

[illuhad]

Hi @jiuxiaocloud - please validate that a regular HIP program works. In your case, the backend was loaded correctly but loads no devices. This means that most likely hipGetDeviceCount() returned 0. In that case, there is an issue at the HIP or driver level.

There is no inherent reason why LLVM 17 or ROCm 5.7.1 won't work. But please note that we only provide limited support for building AdaptiveCpp against the LLVM shipped with ROCm, as ROCm LLVM is known to be unstable for anything but HIP compilation, and is a moving target that typically sits in between LLVM versions and thus difficult to support. We discourage such configurations in our documentation.
In your case, the issue is most likely a runtime and not a compiler issue, so this won't matter.

@gogo2 Please validate whether a) the HIP backend plugin was loaded successfully (you'll see a warning if it could not be loaded) and if so, b) check whether a regular HIP program can see the devices.

[gogo2]

Nevermind, in my case, it was a permissions issue on the cluster.

[illuhad]

Right, for ROCm to work you need to be in the render and/or video group. @jiuxiaocloud - please also check that :)