fix ROCm on Windows

[Cyberhan123]

There is no need for the m link library in the Windows environment. If we try to link, an error will occur.
I can make sure it's correct because at rwkv, that's what I do. We can also check in the rwkv release that the built library is correct.

[slaren]

Is there any reason to not add the library to the ggml-rocm target instead?

[Cyberhan123]

Is there any reason to not add the library to the ggml-rocm target instead?

On windows we need the following commands to complete the build.

mkdir build
cd build
cmake .. -G "Ninja" -DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=clang++ -DRWKV_HIPBLAS=ON -DCMAKE_BUILD_TYPE=Release -DAMDGPU_TARGETS=gfx1100
cmake --build . --config Release

We will get the following error report:

lld-link: error: could not open 'm.lib': no such file or directory
CLANG_~1: error: linker command failed with exit code 1 (use -v to see invocation)
ninja: build stopped: subcommand failed.

The conclusion from my analysis is that ROCm's clang/clang++ does not have m.lib.

[Cyberhan123]

@slaren I also searched on google and I found similar questions: conda-forge/fftw-feedstock#75,

[slaren]

I have been checking how this is done in llama.cpp, but from what I can tell llama.cpp doesn't link to m at all, or at least not explicitly. So maybe this isn't necessary? @ggerganov do you know why ggml links to m?

Otherwise, if I understand the issue, under Windows there is no need to link to the m library, and that applies both to MSVC and clang. So maybe checking for WIN32 instead of MSVC would also work:

if (WIN32)
    target_link_libraries(${TARGET} PUBLIC ${GGML_EXTRA_LIBS} ${CMAKE_THREAD_LIBS_INIT})
else()
    target_link_libraries(${TARGET} PUBLIC m ${GGML_EXTRA_LIBS} ${CMAKE_THREAD_LIBS_INIT})
endif()

[ggerganov]

I have forgotten why this if exists. Let's change it to never link m and merge it. If something breaks we will bring it back

[Cyberhan123]

I also added some settings, which are the same as llama.cpp.

[Cyberhan123]

ok

[Cyberhan123]

It looks like ci failed

[slaren]

I have changed it to find_library as suggested in https://cliutils.gitlab.io/modern-cmake/chapters/features/small.html
Can you test if this still works with hip?

[slaren]

I suspect the reason why it is not necessary to link with it explicitly in llama.cpp is because there are always C++ sources in the target.

[Cyberhan123]

it is working