GENESIS 2.1.6.1 with GPU support

Webpage

https://mdgenesis.org/
https://github.com/genesis-release-r-ccs/genesis

Version

2.1.6.1

Build Environment

  • GCC 15.1.1 (gcc-toolset-15)
  • MKL 2025.3 (oneAPI 2025.3.1)
  • Intel MPI 2021.17.2 (oneAPI 2025.3.1)
  • CUDA 13.1 Update 1

Files Required

  • genesis-2.1.6.1.tar.gz

Build Procedure

Built and tested on ccgpu.

#!/bin/sh

VERSION=2.1.6.1
BASEDIR=/home/users/${USER}/Software/GENESIS/${VERSION}
TARBALL=${BASEDIR}/genesis-${VERSION}.tar.gz

INSTALL_DIR=/apl/genesis/${VERSION}-CUDA

WORKDIR=/gwork/users/${USER}

PARALLEL_TESTS=8

# --------------------------------------------------------------------------
umask 0022

module -s purge
module -s load gcc-toolset/15
module -s load mkl/2025.3
module -s load intelmpi/2021.17.2
module -s load cuda/13.1u1

export LANG=C
export LC_ALL=C
export OMP_NUM_THREADS=1
MKL_LIBS="$(mkl_link_tool --quiet -libs -c gnu_f -o gomp)"
ulimit -s unlimited

cd ${WORKDIR}
if [ -d genesis-${VERSION} ]; then
 mv genesis-${VERSION} genesis-erase
 rm -rf genesis-erase &
fi

tar xf ${TARBALL}
cd genesis-${VERSION}

sed -i -e "s|nvToolsExt.h|nvtx3/nvToolsExt.h|" src/spdyn/gpu_sp_energy.cu

autoreconf
LAPACK_LIBS="${MKL_LIBS}" FC=mpif90 CC=mpicc \
 ./configure --prefix=${INSTALL_DIR} \
             --enable-gpu \
             --enable-single \
             --with-cuda=/apl/cuda/13.1u1 \
             --with-gpuarch=sm_80
make && make install

# tests
SPDYN=${INSTALL_DIR}/bin/spdyn

cd tests/regression_test
for f in test.py test_remd.py test_fep.py test_rpath.py test_gamd.py; do
 sed -i -e "s/env python/env python3/" $f
done

# spdyn tests
./test.py           "mpirun -np ${PARALLEL_TESTS} $SPDYN"
./test_remd.py      "mpirun -np ${PARALLEL_TESTS} $SPDYN"
./test_rpath.py     "mpirun -np ${PARALLEL_TESTS} $SPDYN"
./test_gamd.py      "mpirun -np ${PARALLEL_TESTS} $SPDYN"
./test_fep.py       "mpirun -np ${PARALLEL_TESTS} $SPDYN"

Tests

Tests using CUTOFF have aborted (as expected). All the other tests have passed.

メモ

  • Intel Compiler is not so good for GPU version unlike CPU-only version.
  • Fix for nvToolsExt.h is necessary for CUDA 12.9 or later.
  • --with-gpuarch=sm_80 (for Ampere) was necessary for this build. (Flag for compute capability 6.0 caused an error.)