[amdsmi] Drop patches, bump submodule and enable testing

[dmitrii-galantsev]

Motivation

1. Combine Remove amdsmi patches after applying to base/amdsmi #1969 and [amdsmi] Fix amdsmitst's sourcing of libamd_smi.so #1869
2. Bump amdsmi submodule (it includes the gtest, libamd_smi.so, and librocm_core.so patches)

[marbre]

As discussed offline, the patches now upstreamed need to specify the exact soname, including the version as the rocm-core package does not provide symlinks and only ships versioned shared objects.

[dmitrii-galantsev]

bumped smi to ROCm/amdsmi@2bccfd1 after merging
ROCm/amdsmi@8bdf951

[araravik-psd]

We are seeing multiple .so file imports while running "rocm-sdk test". Removed skipping amdsmi tests from core tests, devel tests and library tests in rocm_sdk tests

Gist file showing the import failures:

https://gist.github.com/araravik-psd/b16048ba563d6a3c2eab5dbc2b7fae21

[araravik-psd]

Tried adding the below .so file path imports in dist_info files for rocm_sdk, rocm_core and rocm_devel packages but dont see them taking effect

LibraryEntry(
"LLVMOffload",
"core",
"libLLVMOffload.so*",
"",
"lib/llvm/lib",
)

LibraryEntry("hipsolver_fortran", "core", "libhipsolver_fortran.so*", "")

[jayhawk-commits]

Tried adding the below .so file path imports in dist_info files for rocm_sdk, rocm_core and rocm_devel packages but dont see them taking effect

LibraryEntry( "LLVMOffload", "core", "libLLVMOffload.so*", "", "lib/llvm/lib", )

LibraryEntry("hipsolver_fortran", "core", "libhipsolver_fortran.so*", "")

Based on the gist provided, it is erroring out on treating libLLVMOffload.so.22.0git as a .so file. The wild cards in the LibraryEntry are pulling in this file. Please take a look at the packages from a nightly run to confirm the actual name of the .so file and search specifically for that and not this git file.

[jayhawk-commits]

I downloaded a nightly build on an Ubuntu system and those .so*git files are the actual .so files and ldd command works on them. So that theory was off.

[jayhawk-commits]

Taking a look at the gist log again, this path below does not exist.

  testSharedLibrariesLoad (rocm_sdk.tests.devel_test.ROCmDevelTest.testSharedLibrariesLoad) [Check shared library loads] (so_path=PosixPath('/workspace/TheRock/.venv/lib/python3.12/site-packages/_rocm_sdk_devel/lib/python3.12/site-packages/rocpd/libpyrocpd.cpython-312-x86_64-linux-gnu.so')) ... ERROR

[jayhawk-commits]

rocprofiler-sdk had a PR involving rocpd recently here #2384

I also see amd_smi in the gist being in the wrong path. The .so files are not expected to be in /share/ but in /lib/ .

  testSharedLibrariesLoad (rocm_sdk.tests.devel_test.ROCmDevelTest.testSharedLibrariesLoad) [Check shared library loads] (so_path=PosixPath('/workspace/TheRock/.venv/lib/python3.12/site-packages/_rocm_sdk_devel/share/amd_smi/amdsmi/libamd_smi.so')) ... ERROR

[jayhawk-commits]

Resolved a merge conflict.

[araravik-psd]

rocprofiler-sdk had a PR involving rocpd recently here #2384

I also see amd_smi in the gist being in the wrong path. The .so files are not expected to be in /share/ but in /lib/ .

  testSharedLibrariesLoad (rocm_sdk.tests.devel_test.ROCmDevelTest.testSharedLibrariesLoad) [Check shared library loads] (so_path=PosixPath('/workspace/TheRock/.venv/lib/python3.12/site-packages/_rocm_sdk_devel/share/amd_smi/amdsmi/libamd_smi.so')) ... ERROR

Thanks @jayhawk-commits , I picked up the latest version of rocm with these changes after rebase and tested it on a shark machine.

I see the rocpd and roctx and a few more packages are not picked up properly, see the files are available and named fine. Wil to add these library imports to dist_info like before to check if this can be fixed.

ERROR: testSharedLibrariesLoad (rocm_sdk.tests.devel_test.ROCmDevelTest.testSharedLibrariesLoad) [Check shared library loads] (so_path=PosixPath('/home/tester/TheRock/.venv/lib/python3.12/site-packages/_rocm_sdk_devel/lib/python3.11/site-packages/rocpd/libpyrocpd.cpython-311-x86_64-linux-gnu.so'))

ERROR: testSharedLibrariesLoad (rocm_sdk.tests.devel_test.ROCmDevelTest.testSharedLibrariesLoad) [Check shared library loads] (so_path=PosixPath('/home/tester/TheRock/.venv/lib/python3.12/site-packages/_rocm_sdk_devel/lib/python3.13/site-packages/roctx/libpyroctx.cpython-313-x86_64-linux-gnu.so'))

Full logs:

rocm_sdk_test_logs.txt

[Copilot]

Copilot encountered an error and was unable to review this pull request. You can try again by re-requesting a review.

[ScottTodd]

@araravik-psd please be careful with git pushes. This branch now contains additional commits.

[jayhawk-commits]

Was CLI used to rebase or the 'update branch' button on GitHub?

[jayhawk-commits]

#2470 bumps amdsmi to a newer SHA than this PR, also drops the patches, but is missing enabling the tests and the post-hook. Due to the bad state this PR is now in, I suggest we proceed to merge in #2470 as that has passed CI except for the rocwmma failure that is already pre-existing and unrelated. The enablement of tests with the post-hook can proceed on top after.

[araravik-psd]

#2470 bumps amdsmi to a newer SHA than this PR, also drops the patches, but is missing enabling the tests and the post-hook. Due to the bad state this PR is now in, I suggest we proceed to merge in #2470 as that has passed CI except for the rocwmma failure that is already pre-existing and unrelated. The enablement of tests with the post-hook can proceed on top after.

Makes sense we can proceed with #2470 first and add the post-hook script on top of that. I was trying to use the CLI method for the rebase as update branch with rebase was disabled due to conflicts.

[jayhawk-commits]

#2486 should also address the shared library load test failures that were from unrelated libraries.

[ScottTodd]

What is the status of this PR? Are you waiting on a review? I see some merge conflicts.

[araravik-psd]

This commits in the PR were cherry picked into

90a1ae6

Closing this PR.