Build cross compile environment (build on x86, run on aarch64 + GPU) to test MG C++ BFS performance

[seunghwak]

Just for archival purpose, will be closed immediately.

1. Install cross compilers

gcc-aarch64-linux-gnu
g++-aarch64-linux-gnu
gfortran-aarch64-linux-gnu (maybe unnecessary)

2. Install CUDA toolkit for cross compile.

Note that this is in-addition to the existing CUDA toolkit. Once installed, you will see new directories with sbsa-linux (e.g. /usr/local/cuda-12.8/targets/sbsa-linux).

3. Set-up a minimal conda environment (minimal_cuda-128_arch_cross_linux-aarch64.yaml). Note that the purpose of this PR is to just run MG C++ BFS tests. No need to install any packages to run python libraries.

Then, install NCCL for aarch64 to the minimal conda environment.

CONDA_SUBDIR=linux-aarch64 conda install -c conda-forge nccl

4. Build librapids_logger.so for aarch64.

git clone https://github.com/rapidsai/rapids-logger
cd rapids-logger
mkdir build
cd build
cmake .. -DCMAKE_SYSTEM_NAME=Linux -DCMAKE_SYSTEM_PROCESSOR=aarch64 -DCMAKE_INSTALL_PREFIX=$CONDA_PREFIX -DCMAKE_CUDA_HOST_COMPILER=aarch64-linux-gnu-g++ -DCMAKE_C_COMPILER=aarch64-linux-gnu-gcc -DCMAKE_CXX_COMPILER=aarch64-linux-gnu-g++ -G Ninja
ninja

Copy librapids_logger.so to the lib directory of the minimal conda environment.

5. Build OpenMPI for cross-compile

export CC=aarch64-linux-gnu-gcc
export CXX=aarch64-linux-gnu-g++
export FC=aarch64-linux-gnu-gfortran (maybe unnecessary)
./configure --host=aarch64-linux-gnu --build=x86_64-linux-gnu --prefix=/home/seunghwak/Programs/openmpi-aarch64 --disable-mpi-fortran --enable-script-wrapper-compilers
make
make install

Update MPI_ROOT in the tests/CMakeLists.txt file to use this MPI.

    set(MPI_ROOT "/home/seunghwak/Programs/openmpi-aarch64")
    set(MPI_CXX_COMPILER /home/seunghwak/Programs/openmpi-aarch64/bin/mpic++)

6. Build cugraph on x86 for aarch64.

cmake .. -DCMAKE_SYSTEM_NAME=Linux -DCMAKE_SYSTEM_PROCESSOR=aarch64 -DCMAKE_INSTALL_PREFIX=$CONDA_PREFIX -DCMAKE_CUDA_HOST_COMPILER=aarch64-linux-gnu-g++ -DCMAKE_C_COMPILER=aarch64-linux-gnu-gcc -DCMAKE_CXX_COMPILER=aarch64-linux-gnu-g++ -DCMAKE_CUDA_COMPILER=/usr/local/cuda-12.8/bin/nvcc -DCMAKE_CUDA_ARCHITECTURES=100 -DBUILD_TESTS=ON -DUSE_RAFT_STATIC=OFF -DCUGRAPH_COMPILE_RAFT_LIB=OFF -DBUILD_CUGRAPH_MG_TESTS=ON -DBUILD_CUGRAPH_MTMG_TESTS=OFF -G Ninja

7. Copy the libcugraph.so and MG_BFS_TEST to the target system.

[copy-pr-bot]

Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually.

Contributors can view more details about this message here.