[Issue]: shared libhsakmt

[trixirt]

Problem Description

Some issues with the transition of hsakmt in creating a shared library on Fedora

https://github.com/ROCm/ROCR-Runtime/blob/amd-staging/libhsakmt/CMakeLists.txt#L140
Hardcodes 'Static'

https://github.com/ROCm/ROCR-Runtime/blob/amd-staging/libhsakmt/src/libhsakmt.ver
Is missing
hsaKmtCreateQueueExt;
hsaKmtRegisterGraphicsHandleToNodesExt;

Operating System

Fedora Rawhide

CPU

ALL

GPU

ALL

ROCm Version

ROCm 6.3.0

ROCm Component

No response

Steps to Reproduce

No response

(Optional for Linux users) Output of /opt/rocm/bin/rocminfo --support

No response

Additional Information

No response

[ppanchad-amd]

Hi @trixirt. Internal ticket has been created to investigate your issue. Thanks!

[Mystro256]

@trixirt starting with 6.3, hsakmt is expected not to be interfaced publicly anymore. It's meant to be a static lib within rocr now. Perhaps the cmake logic needs to be fixed a bit, but the intention is for all public hsakmt interfaces to go through libhsa-runtime64 now.

[trixirt]

A separate delivered static library is a flaw, distros are very strict about no static libraries.

[Mystro256]

A separate delivered static library is a flaw, distros are very strict about no static libraries.

Yeah I was speaking with @kentrussell , I suspect this is a bug.
These lines should be dropped to be consistent with the design:
https://github.com/ROCm/ROCR-Runtime/blob/rocm-6.3.0/libhsakmt/CMakeLists.txt#L194-L200

Need to confirm, but I was told libhsakmt was supposed to be dropped for 6.3

[nstomlinson]

It would be helpful to us to retain the ability to build HSAKMT as a separate shared library, as we make direct use of some of its functions that aren't available when statically linked to HSA to implement some of the stuff in our SCALE library.

[trixirt]

It would be helpful to us to retain the ability to build HSAKMT as a separate shared library, as we make direct use of some of its functions that aren't available when statically linked to HSA to implement some of the stuff in our SCALE library.

Can you share a URL for SCALE ? I am not familiar with this library.

[nstomlinson]

Sure! https://docs.scale-lang.com/

[AngryLoki]

Update for 6.4.1:

https://github.com/ROCm/ROCR-Runtime/blob/amd-staging/libhsakmt/CMakeLists.txt#L140
Hardcodes 'Static'

Still applicable (patched in-place in Gentoo).

https://github.com/ROCm/ROCR-Runtime/blob/amd-staging/libhsakmt/src/libhsakmt.ver
Is missing
hsaKmtCreateQueueExt;
hsaKmtRegisterGraphicsHandleToNodesExt;

Was applicable previously due to usage in kfdtest. Not applicable in rocm-6.4.1 branch. Might be applicable to rocm-staging - see

ROCR-Runtime/libhsakmt/tests/kfdtest/src/BaseQueue.cpp

Line 60 in 167602e

status = hsaKmtCreateQueueExt(NodeId,

Sorry, the above is wrong, with shared libhsakmt, we have this:

$ rocminfo
rocminfo: symbol lookup error: /usr/lib64/libhsa-runtime64.so.1: undefined symbol: hsaKmtCreateQueueExt

$ nm -gD /usr/lib64/libhsa-runtime64.so.1 | grep hsaKmtCreateQueueExt
                 U hsaKmtCreateQueueExt

[kentrussell]

Thanks for the help and update. We're currently working on fixing the whole static thunk thing entirely, instead of just patching together fixes to handle the issues from the somewhat incomplete move-to-static task. Hopefully we'll have that ready for the next major ROCm release, and we can avoid the repeated libhsakmt.ver updates

[systems-assistant]

This issue has been migrated to: ROCm/rocm-systems#284