Compiling ELPA on a local Ubuntu machine

Recently I needed a working local build of ELPA on Ubuntu. This post records a simple recipe that worked for me with an MPI-enabled build and a local installation prefix. The steps below are aimed at a workstation or laptop rather than a cluster module environment.

Install the required packages

On Ubuntu, the main dependencies are a Fortran compiler, OpenMPI, OpenBLAS, and ScaLAPACK. I also install the usual Autotools utilities in case the build system needs to regenerate files.

sudo apt update
sudo apt install -y \
  build-essential \
  gfortran \
  openmpi-bin \
  libopenmpi-dev \
  libopenblas-dev \
  libscalapack-openmpi-dev \
  autoconf \
  automake \
  libtool \
  pkg-config \
  m4 \
  perl

The important point is to keep the MPI stack consistent. On Ubuntu that usually means using the OpenMPI-flavored ScaLAPACK package together with the OpenMPI compiler wrappers.

Configure and build ELPA

Assuming you have already unpacked the ELPA source tarball, I recommend building in a separate build/ directory:

mkdir -p build
cd build

../configure \
  FC=mpifort \
  CC=mpicc \
  CXX=mpicxx \
  --prefix="$HOME/local/elpa/2025.06.001" \
  --libdir="$HOME/local/elpa/2025.06.001/lib" \
  --includedir="$HOME/local/elpa/2025.06.001/include" \
  --enable-openmp \
  --enable-shared \
  CFLAGS='-O2 -march=native' \
  FCFLAGS='-O2 -march=native' \
  CXXFLAGS='-O2 -march=native' \
  SCALAPACK_LDFLAGS='-lscalapack-openmpi' \
  BLAS_LDFLAGS='-lopenblas' \
  --disable-avx512-kernels

make -j"$(nproc)"
make install

This installs ELPA into $HOME/local/elpa/2025.06.001 instead of writing into a system path such as /usr/local.

Why these flags matter

There are three choices above that solved the build reliably for me:

1. FC=mpifort, CC=mpicc, and CXX=mpicxx make sure ELPA is built against the MPI toolchain you actually intend to use.
2. SCALAPACK_LDFLAGS='-lscalapack-openmpi' and BLAS_LDFLAGS='-lopenblas' force the linker to use the Ubuntu package names for ScaLAPACK and BLAS/LAPACK.
3. --disable-avx512-kernels avoids configure-time problems on systems where ELPA enables aggressive SIMD kernels but the local compiler flags or CPU support do not line up cleanly.

I also found it helpful to keep

CFLAGS='-O2 -march=native'
FCFLAGS='-O2 -march=native'
CXXFLAGS='-O2 -march=native'

so the compiler targets the local machine cleanly.

Check the installation

After make install, I usually verify that the shared library and pkg-config metadata are present:

ls $HOME/local/elpa/2025.06.001/lib/libelpa*.so*
pkg-config --variable=libdir elpa

If pkg-config cannot find ELPA, add the install location to your environment:

export PKG_CONFIG_PATH=$HOME/local/elpa/2025.06.001/lib/pkgconfig:$PKG_CONFIG_PATH
export LD_LIBRARY_PATH=$HOME/local/elpa/2025.06.001/lib:$LD_LIBRARY_PATH

These two variables are especially useful when the installation lives under your home directory rather than a system library path.

A note for ABACUS users

If you are compiling ELPA for use with ABACUS, keep OpenMP enabled and prefer a shared-library build. In that case, it is also a good idea to use the MPI compiler wrappers explicitly and make sure the OpenMP ELPA library variant is installed, since ABACUS typically links against -lelpa_openmp.

Final remarks

The main trap was not missing BLAS or ScaLAPACK packages. The more subtle issue was ELPA’s SIMD kernel selection during configure. Once I used the MPI wrappers, pointed the linker to Ubuntu’s OpenMPI math libraries, and disabled AVX512 kernels, the build was straightforward.