export NWCHEM_TOP=$HOME/NWChem/github
mkdir -p ${NWCHEM_TOP}
#
# TOOLCHAIN INFO - NVHPC SDK REQUIRED
#
export NVHPC_VERSION=24.7
export NVHPC_ROOT=/opt/nvidia/hpc_sdk/Linux_$(uname -m)/${NVHPC_VERSION}
#
# CHANGE THIS IF HPCX IS NOT APPROPRIATE
#
#export MPI_ROOT=${NVHPC_ROOT}/comm_libs/hpcx/latest/ompi/
export MPI_ROOT=${NVHPC_ROOT}/comm_libs/12.5/openmpi4/latest
#
# BLAS/LAPACK DEPENDENCY
#
#export USE_64TO32=y
#export BLAS_SIZE=4
#export BLAS_LIB="-L${NVHPC_ROOT}/compilers/lib -lblas_lp64"
#export LAPACK_LIB="-L${NVHPC_ROOT}/compilers/lib -llapack_lp64"
unset USE_64TO32
export BLAS_SIZE=8
export BLAS_LIB="-L${NVHPC_ROOT}/compilers/lib -lblas_ilp64"
export LAPACK_LIB="-L${NVHPC_ROOT}/compilers/lib -llapack_ilp64"
export BLASOPT="${LAPACK_LIB} ${BLAS_LIB}"
# IF NVHPC DOES NOT WORK FOR SOME REASON
#export USE_64TO32=y
#export BLAS_SIZE=4
#export BUILD_OPENBLAS=T
# IF ALL ELSE FAILS (PERFORMANCE IS NOT GOOD)
#export USE_INTERNALBLAS=y
#unset BLAS_LIB LAPACK_LIB BLASOPT
#
export PATH=${NVHPC_ROOT}/compilers/bin:${PATH}
export CC=nvc
export CXX=nvc++
export FC=nvfortran
#
export PATH=${MPI_ROOT}/bin:${PATH}
export LD_LIBRARY_PATH=${MPI_ROOT}/lib:${LD_LIBRARY_PATH}
export MPICC=${MPI_ROOT}/bin/mpicc
export MPICXX=${MPI_ROOT}/bin/mpicxx
export MPIFC=${MPI_ROOT}/bin/mpifort
#
# REMOVE "tce" TO MAKE IT COMPILE FASTER IF YOU DO NOT NEED IT
#
export NWCHEM_MODULES="smallqm moints ccsd tce xtb"
export NWCHEM_TARGET=LINUX64

export USE_OPENMP=1
export USE_MPI=y
export USE_TBLITE=1
#
# ARMCI-MPI IS MORE STABLE BUT SOMETIMES SLOWER THAN MPI-PR
#
# RUN THIS IF YOU DO NOT HAVE THESE ALREADY
#
# sudo apt-get install -y libtool automake autoconf m4
#
export EXTERNAL_ARMCI_PATH=$NWCHEM_TOP/external-armci
export ARMCI_NETWORK=ARMCI # ARMCI-MPI
#export ARMCI_NETWORK=MPI-PR
#export ARMCI_NETWORK=MPI-TS

#
# GPU AND RELATED STUFF
#
export USE_F90_ALLOCATABLE=1
export USE_OPENACC_TRPDRV=1
# use to link NWChem and to build GA
export NWCHEM_LINK_CUDA=1
# used to download GA
#export DEV_GA=1
# runtime option
#export MA_USE_CUDA_MEM=1
export CUBLAS_LOGINFO_DBG=1
export CUTENSOR_LOG_LEVEL=1
#export CUTENSOR_LOG_LEVEL=5
#
# DEBUG SYMBOLS ARE GOOD
#
export USE_DEBUG=1
#
# SET REASONABLE DEFAULTS IN CASE INPUT FILES DO NOT
#
export NWCHEM_PERMANENT_DIR=/tmp/ # MUST BE A SHARED FILESYSTEM e.g. LUSTRE
export NWCHEM_SCRATCH_DIR=/tmp/   # SHOULD BE A NODE-LOCAL FILESYSTEM
#
# 16 GB PER PROCESS IS IDEAL
#
export NWCHEM_MEMORY_TOTAL=7500000000
export NWCHEM_MEMORY_HEAP=1250000000
export NWCHEM_MEMORY_STACK=1250000000
export NWCHEM_MEMORY_GLOBAL=5000000000

export USE_SIMINT=y
export SIMINT_MAXAM=5

export CCSDTLR=1
export CCSDTQ=1