Linking Error with HIP SDK 5.7 (Windows)

[blinkfrog]

Hello,

I recently updated the HIP SDK from version 5.5 to 5.7 and am experiencing a build error when trying to compile SYCL programs targeting AMD devices with AdaptiveCpp. Here's the error I encounter during the build process of a test project (and any other project results in the same errors too):

[1/2] Building CXX object CMakeFiles/Blur.dir/blur.cpp.obj
FAILED: CMakeFiles/Blur.dir/blur.cpp.obj
python C:/AdaptiveCpp/lib/cmake/AdaptiveCpp/syclcc-launcher --launcher-cxx-compiler=C:/llvm/bin/clang++.exe --launcher-syclcc="python*C:/AdaptiveCpp/bin/acpp"  --acpp-targets="omp;hip.explicit-multipass:gfx1030,gfx1031,gfx1032,gfx1100,gfx1101,gfx1102" C:\llvm\bin\clang++.exe  -IC:/Temp/sycltest/SyclCmakeProject2.AMD/build -IC:/Temp/sycltest/SyclCmakeProject2.AMD -isystem C:/AdaptiveCpp/include -isystem C:/AdaptiveCpp/include/AdaptiveCpp -O3 -DNDEBUG -D_DLL -D_MT -Xclang --dependent-lib=msvcrt "--rocm-path=C:/Program Files/AMD/ROCm/5.7" -MD -MT CMakeFiles/Blur.dir/blur.cpp.obj -MF CMakeFiles\Blur.dir\blur.cpp.obj.d -o CMakeFiles/Blur.dir/blur.cpp.obj -c C:/Temp/sycltest/SyclCmakeProject2.AMD/blur.cpp
fatal error: cannot link module '': linking module flags 'amdgpu_code_object_version': IDs have conflicting values in 'C:\Program Files\AMD\ROCm/5.7\/amdgcn/bitcode\hip.bc' and 'C:/Temp/sycltest/SyclCmakeProject2.AMD/blur.cpp'
1 error generated when compiling for gfx1030.
ninja: build stopped: subcommand failed.

The error seems to involve conflicting module flags related to the amdgpu_code_object_version.
There was no such error when I used HIP SDK 5.5.

I've rebuilt LLVM and AdaptiveCpp, ensuring that it correctly finds HIP SDK 5.7. Unfortunately, this didn't resolve the issue. I have a limited background in LLVM and in using cross-platform open source software in overall, so I may be missing something obvious. Has anyone here encountered a similar problem or could provide guidance on resolving these linking module flag conflicts? Any help or pointers would be greatly appreciated.

Thank you!

[illuhad]

This is probably because your LLVM defaults to an earlier version of amdgpu code object version than what is used in your ROCm. Try upgrading your LLVM. You can also try targeting a newer code object version, e.g. -mlink-builtin-bitcode "/opt/rocm/amdgcn/bitcode/oclc_abi_version_<version>.bc", but I'm not sure if just linking the newer builtin bitcode is sufficient.

[blinkfrog]

Thank you. I've checked, and both ROCm SDK 5.5 and 5.7 use LLVM 17.0.0, while I am using 17.0.4, so this probably isn't the case, if my understanding of how it all works is correct.