Skip to content

Latest commit

 

History

History
216 lines (175 loc) · 4.84 KB

BUILD.md

File metadata and controls

216 lines (175 loc) · 4.84 KB

Build instructions

Below are the instructions to build Cosyr and its dependencies.
They are assumed to be installed in a directory dependencies.

Kokkos
git clone git@github.com:kokkos/kokkos.git
cd kokkos
mkdir build
cd build

cmake -Wno-dev \
  -DCMAKE_BUILD_TYPE=Release \
  -DCMAKE_INSTALL_PREFIX="${dependencies}/kokkos" \
  -DKokkos_ENABLE_TESTS=Off \
  -DKokkos_ENABLE_SERIAL=On \
  -DKokkos_ENABLE_OPENMP=On \
  -DKokkos_ENABLE_CUDA=[On|Off] \
  -DKokkos_ARCH_KNL=[On|Off] \
  ..
make -j 8
make install
cd ../..
Cabana
git clone git@github.com:ECP-copa/Cabana.git cabana
cd cabana
mkdir build
cd build

cmake -Wno-dev \
  -DCMAKE_BUILD_TYPE="Release" \
  -DCMAKE_PREFIX_PATH="${dependencies}/kokkos" \
  -DCMAKE_INSTALL_PREFIX="${dependencies}/cabana" \
  -DCabana_REQUIRE_OPENMP=ON \
  -DCabana_ENABLE_MPI=ON \
  -DCabana_ENABLE_EXAMPLES=ON \
  -DCabana_ENABLE_TESTING=OFF \
  -DCabana_ENABLE_PERFORMANCE_TESTING=OFF \
  -DCabana_ENABLE_CAJITA=ON \
  ..
make -j 8
make install
cd ../..
Python
# install locally
mkdir python
cd python
wget https://www.python.org/ftp/python/3.8.5/Python-3.8.5.tgz
tar xzvf Python-3.8.5.tgz
find . -type d | xargs chmod 0755
cd Python-${VERSION}
./configure --prefix="/path/to/python/install" \
            --enable-shared \
            --enable-optimizations
make -j install

# packages
wget --no-check-certificate https://bootstrap.pypa.io/get-pip.py -O - | python3 - --user
pip3 install wheel numpy matplotlib mpi4py

# in ~/.bashrc
export PATH=/path/to/python/install/bin:~/.local/bin:${PATH}
export PYTHONPATH="/path/to/python/install"
export LD_LIBRARY_PATH="${PYTHONPATH}/lib":${LD_LIBRARY_PATH}
export PYTHON_LD_LIBRARY_PATH="${PYTHONPATH}/lib"
Pybind
git clone git@github.com:pybind/pybind11.git pybind
cd pybind
mkdir build
cd build

cmake -Wno-dev \
  -DCMAKE_BUILD_TYPE=Release \
  -DCMAKE_PREFIX_PATH="${PYTHONPATH}" \
  -DCMAKE_INSTALL_PREFIX="${dependencies}/pybind" \
  -DPYBIND11_TEST=Off \
  ..
make -j 8
make install
cd ../..
Thrust
git clone git@github.com:NVIDIA/thrust.git "${dependencies}/thrust"
Lapacke
git clone git@github.com:Reference-LAPACK/lapack-release.git lapack
cd lapack
mkdir build
cd build

cmake -Wno-dev \
  -DCMAKE_INSTALL_PREFIX="${dependencies}/lapack" \
  -DCMAKE_BUILD_TYPE=Release \
  -DBUILD_SHARED_LIBS=ON \
  -DLAPACKE=ON \
  ..

make -j 8
make install
cd ../..
Wonton
git clone --recursive git@github.com:laristra/wonton.git
cd wonton
mkdir build
cd build

cmake -Wno-dev \
  -DCMAKE_BUILD_TYPE=Release \
  -DCMAKE_INSTALL_PREFIX="${dependencies}/wonton" \
  -DENABLE_UNIT_TESTS=False \
  -DWONTON_ENABLE_MPI=True \
  -DWONTON_ENABLE_THRUST=True \
  -DWONTON_ENABLE_LAPACKE=True \
  -DLAPACKE_ROOT="${dependencies}/lapack" \
  -DTHRUST_ROOT="${dependencies}/thrust" \
  ..

make -j 8
make install
cd ../..
Portage
git clone --recursive git@github.com:laristra/portage.git
cd portage
mkdir build
cd build

cmake -Wno-dev \
  -DCMAKE_BUILD_TYPE=Release \
  -DCMAKE_INSTALL_PREFIX="${dependencies}/portage" \
  -DENABLE_UNIT_TESTS=False \
  -DPORTAGE_ENABLE_MPI=True \
  -DPORTAGE_ENABLE_Jali=False \
  -DPORTAGE_ENABLE_FleCSI=False \
  -DPORTAGE_ENABLE_THRUST=True \
  -DWONTON_ROOT="${dependencies}/wonton" \
  ..

make -j 8
make install
cd ../..
Cosyr
cd "/path/to/cosyr"
mkdir build
cd build

cmake -Wno-dev \
  -DCMAKE_BUILD_TYPE=Release \
  -DPORTAGE_DIR="${dependencies}/portage" \
  -DKokkos_DIR="${dependencies}/kokkos" \
  -Dpybind11_DIR="${dependencies}/pybind" \
  -DCabana_DIR="${dependencies}/cabana" \
  ..
make -j 8

To verify that it runs fine:

export PYTHONPATH="/path/to/cosyr" \
       OMP_PROC_BIND=close \
       OMP_PLACES=cores
mpiexec -np 1 ./cosyr /path/to/cosyr/input/test_beam_g10_l200.py

A set of Jupyter notebooks are available for post-processing.

Notes

Python

Cosyr relies on scripts to handle simulation parameters and data output.
For that purpose, a python environment is required.
It is recommended to install mpi4py for faster subcycle wavelets loading.
To enable HDF5 output, the following packages are required:

  • hdf5 library (see additional instructions here)
  • h5py to use it through the python interface.
  • mpi4py to enable parallel output.
Incompatibility

It is not currently possible to run both the fields kernels using CUDA backend and the interpolation using OpenMP. There is a known bug in Thrust making it impossible to use other device backends when compiled with nvcc.