amd-device-libs: Respect ROCM_DEVICE_LIBS_BITCODE_INSTALL_LOC_NEW in AMDDeviceLibsConfig.cmake

[traversaro]

Motivation

When the ROCM_DEVICE_LIBS_BITCODE_INSTALL_LOC_NEW variable is specified, the install location of bytecode changes, but this change is not reflected in the AMDDeviceLibsConfig.cmake file, that still points to the default install location. This PR fixes this by respecting the ROCM_DEVICE_LIBS_BITCODE_INSTALL_LOC_NEW when generating the AMDDeviceLibsConfig.cmake file.

Technical Details

The Packages.cmake was redefining a install_path_suffix variable with the hardcoded install value, but a similar variable INSTALL_ROOT_SUFFIX (that respects ROCM_DEVICE_LIBS_BITCODE_INSTALL_LOC_NEW was already defined in OCL.cmake, so as OCL is included before Packages, we can just get rid of install_path_suffix and use INSTALL_ROOT_SUFFIX .

This removes the need for patches or workarounds used in downstream packaging, see:

• https://github.com/gentoo/gentoo/blob/86548b8f3404ab51d5dfc1e3e84b266d77dd77f5/dev-libs/rocm-device-libs/rocm-device-libs-6.4.3.ebuild#L61-L63
• https://github.com/conda-forge/rocm-device-libs-feedstock/blob/main/recipe/patches/0001-Use-a-folder-inside-lib-for-bitcode-instead-of-top-l.patch

Test Plan

I configured the amd/device-libs folder in standalone mode, and specified a custom install location with ROCM_DEVICE_LIBS_BITCODE_INSTALL_LOC_NEW, and I verified manually that AMDDeviceLibsConfig.cmake was correctly updated.

Test Result

See Test Plan.

Submission Checklist

• Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.

[skganesan008]

Github action triggered jenkins job: https://compiler-ci.amd.com/job/compiler-upstream-rocm-llvm-project-psdb/131

[skganesan008]

Github action triggered jenkins job: https://compiler-ci.amd.com/job/compiler-upstream-rocm-llvm-project-psdb/132

[skganesan008]

Please ignore the PSDB verification trigger. IWe are enabling PSDB for amd-staging PR's starting Oct 20. I was checking if the triggers are happening as expected on non test PR's . Sorry if this had interrupted you.

[skganesan008]

PSDB Link: https://compiler-ci.amd.com/job/compiler-psdb-amd-staging/2354

[skganesan008]

There seems to be a compilation error in the latest build. We have deployed the PSDB automation today. Please check the compilation error.

[traversaro]

There seems to be a compilation error in the latest build. We have deployed the PSDB automation today. Please check the compilation error.

Hello @skganesan008 , thanks for the feedback. The log seems at https://compiler-ci.amd.com/job/compiler-psdb-amd-staging/2354, is this accessible from outside AMD?

[lamb-j]

@traversaro not yet publicly accessible, that's a work in progress though

The failure doesn't seem related to your change though:

[2025-10-21T16:45:02.922Z] /jenkins/workspace/compiler-psdb-amd-staging/repos/llvm-project/llvm/projects/SPIRV-LLVM-Translator/lib/SPIRV/LLVMToSPIRVDbgTran.cpp:591:31: error: request for member ‘getUnversionedName’ in ‘CU->llvm::DICompileUnit::getSourceLanguage()’, which is of non-class type ‘unsigned int’

Can you rebase on top of the latest amd-staging? That may fix it

[traversaro]

Can you rebase on top of the latest amd-staging? That may fix it

Done!

[skganesan008]

PSDB Link: https://compiler-ci.amd.com/job/compiler-psdb-amd-staging/2386

[lamb-j]

lgtm, good catch

We're already setting INSTALL_ROOT_SUFFIX to amdgcn/bitcode here (as you mentioned):

llvm-project/amd/device-libs/cmake/OCL.cmake

Line 52 in 83f7f2e

set (INSTALL_ROOT_SUFFIX "amdgcn/bitcode")