[Openmp-dev] Building Clang+OpenMP on Power9

[Itaru Kitayama via Openmp-dev]

I'm on a POWER8 machine at JSC and am able to build the toolchain with GCC
7.2.0.
The machine is backed by CentOS Linux release 7.6.1810.

On Wed, Oct 16, 2019 at 2:06 AM Jacob Lambert via Openmp-dev <
openmp-dev at lists.llvm.org> wrote:

> I'm having trouble building clang+OpenMP on a Power9 system with support
> of offloading to an attached Nvidia GPU (V100). Is this kind of build
> currently supported? I was able to successfully build clang, which I'm now
> attempting to use as the C/C++ compiler for a clang+openMP build.
>
>  Here are my Cmake/make commands:
> ----------------------------------------
> /home/username/misc/bin/cmake -G "Unix Makefiles" ../llvm \
>   -DLLVM_ENABLE_PROJECTS="clang;openmp" \
>   -DCMAKE_BUILD_TYPE=Release \
>
> -DCMAKE_INSTALL_PREFIX=/home/username/compilers/clang-build-RELEASE_900-rc6_pwr9
> \
>
> -DCMAKE_C_COMPILER=/home/username/git/llvm-project/install_pwr9/bin/clang \
>   -DCMAKE_C_FLAGS="--gcc-toolchain=/auto/software/gcc/ppc64le/gcc-8.3.0" \
>
> -DCMAKE_CXX_COMPILER=/home/username/git/llvm-project/install_pwr9/bin/clang++
> \
>   -DCMAKE_CXX_FLAGS="--gcc-toolchain=/auto/software/gcc/ppc64le/gcc-8.3.0"
> \
>   -DCLANG_OPENMP_NVPTX_DEFAULT_ARCH=sm_70 \
>   -DLIBOMPTARGET_NVPTX_COMPUTE_CAPABILITIES=70
>
> make VERBOSE=1 -j8
> ----------------------------------------
>
> Here's the error I'm getting:
> ----------------------------------------
> /usr/local/cuda-10.1/bin/nvcc
> /home/username/git/llvm-project/openmp/libomptarget/deviceRTLs/nvptx/src/
> omp_data.cu -dc -o
> /home/username/git/llvm-project/build_pwr9/projects/openmp/libomptarget/deviceRTLs/nvptx/CMakeFiles/omptarget-nvptx.dir/src/./omptarget-nvptx_generated_omp_data.cu.o
> -ccbin /auto/software/gcc/ppc64le/gcc-8.3.0/bin/gcc -m64 -D_GNU_SOURCE
> -D__STDC_CONSTANT_MACROS -D__STDC_FORMAT_MACROS -D__STDC_LIMIT_MACROS
> -gencode arch=compute_35,code=sm_35 -DNVCC -I/usr/local/cuda-10.1/include
> -I/usr/include/libxml2 -I/home/username/git/llvm-project/build_pwr9/include
> -I/home/username/git/llvm-project/llvm/include
> -I/home/username/git/llvm-project/openmp/libomptarget/include
> /auto/software/gcc/ppc64le/gcc-8.3.0/include/c++/8.3.0/type_traits(335):
> error: identifier "__ieee128" is undefined
> /auto/software/gcc/ppc64le/gcc-8.3.0/include/c++/8.3.0/bits/std_abs.h(101):
> error: identifier "__ieee128" is undefined
> /auto/software/gcc/ppc64le/gcc-8.3.0/include/c++/8.3.0/bits/std_abs.h(102):
> error: identifier "__ieee128" is undefined
> 3 errors detected in the compilation of
> "/tmp/tmpxft_0000bd6b_00000000-6_omp_data.cpp1.ii".
> ----------------------------------------
>
> Finally, here are the offending lines from std_abs.h:
> ----------------------------------------
>  99 #if !defined(__STRICT_ANSI__) && defined(_GLIBCXX_USE_FLOAT128)
> 100   inline _GLIBCXX_CONSTEXPR
> 101   __float128
> 102   abs(__float128 __x)
> 103   { return __x < 0 ? -__x : __x; }
> 104 #endif
>
> I haven't been able to find any reference of a variable substitution/macro
> between __float128 and __ieee128, I'm not sure where/when this takes place.
> ----------------------------------------
>
> Any ideas? I'm not sure if the errors stem from my clang build process, my
> gcc installation, my cuda/nvcc installlation, Unix Makefiles vs Ninja, etc.
>
> Jacob
> _______________________________________________
> Openmp-dev mailing list
> Openmp-dev at lists.llvm.org
> https://lists.llvm.org/cgi-bin/mailman/listinfo/openmp-dev
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/openmp-dev/attachments/20191016/5221f689/attachment.html>