electronic-structure / SIRIUS

Table of contents

• Introduction
• Installation
  • Configuring SIRIUS
  • Developing and debugging SIRIUS
  • Manual installation
  • Archlinux
  • Installation on Piz Daint
• Accelerating DFT codes
  • Quantum ESPRESSO
  • CP2K
• Contacts
• Acknowledgements

Introduction

SIRIUS is a domain specific library for electronic structure calculations. It implements pseudopotential plane wave (PP-PW) and full potential linearized augmented plane wave (FP-LAPW) methods and is designed for GPU acceleration of popular community codes such as Exciting, Elk and Quantum ESPRESSO. SIRIUS is written in C++14 with MPI, OpenMP and CUDA/ROCm programming models. SIRIUS is organised as a collection of classes that abstract away the different building blocks of DFT self-consistency cycle.

The following functionality is currently implemented in SIRIUS:

• (PP-PW) Norm-conserving, ultrasoft and PAW pseudopotentials
• (PP-PW) Spin-orbit coupling
• (PP-PW) Stress tensor
• (PP-PW, FP-LAPW) Atomic forces
• (PP-PW, FP-LAPW) Collinear and non-collinear magnetism
• (FP-LAPW) APW and LAPW basis sets with arbitrary number of local orbitals
• (FP-LAPW) ZORA and IORA approximations for valence states; full relativistic Dirac equation for core states
• Symmetrization of lattice-periodic functions and on-site matrices
• Generation of irreducible k-meshes
• Python frontend

Installation

It is recommended to install SIRIUS through Spack. To set it up, use

git clone https://github.com/spack/spack.git
. spack/share/spack/setup-env.sh
spack install sirius

Configuring SIRIUS

SIRIUS has many different configurations to enable specific hardware and library support. Some common setups include:

# Use default BLAS, LAPACK, MPI and FFTW3 implementations, without GPU support, using the latest GCC 9.x
$ spack install sirius %gcc@:9

# Explicitly use the latest 3.x release of MPICH for MPI, OpenBLAS for BLAS and LAPACK, FFTW for FFTW3, without GPU support
$ spack install sirius ^mpich@:3 ^fftw ^openblas

# Enable distributed linear algebra, and use Intel MKL for BLAS, ScaLAPACK and FFTW3, without GPU support
$ spack install sirius +scalapack ^intel-mkl

# Build with CUDA support for NVIDIA GPUs
$ spack install sirius +cuda cuda_arch=75

# Build with ROCm support for AMD GPUs
$ spack install sirius +rocm amdgpu_target=gfx906

# Build with MAGMA
$ spack install sirius +cuda +magma

# Build with ELPA
$ spack install sirius +scalapack +elpa

Language interop with Fortran and Python can be enabled with +fortran and +python respectively.

See spack info sirius for the full list of support variants.

Developing and debugging SIRIUS

The recommended way to install the latest development version of SIRIUS is through spack dev-build.

As an example, the following builds SIRIUS with CUDA support in debug mode:

$ git clone --recursive -b develop https://github.com/electronic-structure/SIRIUS.git
$ cd SIRIUS
$ spack dev-build sirius@develop build_type=Debug +cuda

When more control over the build commands is necessary, use spack build-env [spec] -- [command]:

$ mkdir SIRIUS/build && cd SIRIUS/build
$ export SPEC="sirius@develop build_type=Debug +cuda"
$ spack install --only=dependencies $SPEC
$ spack build-env $SPEC -- cmake ..
$ spack build-env $SPEC -- make -j$(nproc)

Manual installation

When installing SIRIUS without Spack, make sure to install the required dependencies first:

• CMake ≥ 3.14
• C++ compiler with C++14 support
• MPI (OpenMPI or MPICH)
• BLAS/LAPACK (OpenBLAS or Intel MKL)
• GSL - GNU scientific library
• LibXC - library of exchange-correlation potentials
• HDF5
• spglib - library for finding and handling crystal symmetries
• SpFFT - domain-specific FFT library
• SPLA - domain-specific distributed GEMM library

and optionally any of the additional libraries:

• ScaLAPACK (Intel MKL or netlib scalapack)
• ELPA
• MAGMA
• CUDA/ROCm
• Boost Filesystem*

* Only required when BUILD_APPS=On and your compiler does not support std::filesystem or std::experimental::filesystem.

Clone the repository and build as follows:

git clone --recursive https://github.com/electronic-structure/SIRIUS.git
mkdir SIRIUS/build
cd SIRIUS/build
export CXX=mpicxx CC=mpicc FC=mpif90
export CMAKE_PREFIX_PATH="path/to/BLAS;path/to/GSL;path/to/LibXC;path/to/HDF5;..."
cmake -DCMAKE_INSTALL_PREFIX=$PWD/sirius
make -j install

where CMAKE_PREFIX_PATH is a list of installation paths of dependencies installed in non-standard locations.

Adding GPU support

To enable CUDA you need to pass the following options to CMake: -DUSE_CUDA=On -DCUDA_ARCH='60;70', where CUDA_ARCH is a list of NVIDIA architectures. Use 60, 61, 62 for Pascal; 70, 72 for Volta; 75 for Turing; and 80 for Ampere. If CUDA is installed in a non-standard directory, you have to pass additional parameter to cmake -DCUDA_TOOLKIT_ROOT_DIR=/path/to/cuda.

To enable MAGMA (GPU implementation of LAPACK) use -DUSE_MAGMA=On. Append MAGMA's installation directory to CMAKE_PREFIX_PATH if necessary.

Parallel eigensolvers

To compile with ScaLAPACK use -DUSE_SCALAPACK=On. To use ELPA, both -DUSE_SCALAPACK=On and -DUSE_ELPA=On are required, as we need ScaLAPACK functionality to transform the generalized eigenvalue problem to standard form, which can then be solved by ELPA. Append ScaLAPACK's and ELPA's install directory to CMAKE_PREFIX_PATH if necessary.

Python module

Use -DCREATE_PYTHON_MODULE=On to build the Python module. The SIRIUS Python module depends on mpi4py and pybind11, which need to be installed on your system.

Additional options

To link against Intel MKL use -DUSE_MKL=On. For Cray libsci use -DUSE_CRAY_LIBSCI=On. Building tests requires -DBUILD_TESTING=On.

By default example applications are built. This can be turned off via -DBUILD_APPS=Off, which is recommended when just building Fortran bindings.

Arch Linux

Arch Linux users can find SIRIUS in the AUR.

Installation on Piz Daint

Please refer to the SIRIUS wiki page and CSCS User portal for detailed instructions.

Accelerating DFT codes

Quantum ESPRESSO

Quantum ESPRESSO is a popular open source suite of computer codes for electronic-structure calculations and materials modeling at the nanoscale. It is based on DFT, plane waves, and pseudopotentials. We maintain the GPU-accelerated version of Quantum ESPRESSO with SIRIUS bindings. This version is frequently synchronised with the develop branch of the official QE repository. A typical example of using SIRIUS inside QE is listed below:

subroutine get_band_energies_from_sirius
  !
  use wvfct,    only : nbnd, et
  use klist,    only : nkstot, nks
  use lsda_mod, only : nspin
  use sirius
  !
  implicit none
  !
  integer, external :: global_kpoint_index
  !
  real(8), allocatable :: band_e(:,:)
  integer :: ik, nk, nb, nfv

  allocate(band_e(nbnd, nkstot))

  ! get band energies
  if (nspin.ne.2) then
    ! non-magnetic or non-collinear case
    do ik = 1, nkstot
      call sirius_get_band_energies(ks_handler, ik, 0, band_e(1, ik))
    end do
  else
    ! collinear magnetic case
    nk = nkstot / 2
    ! get band energies
    do ik = 1, nk
      call sirius_get_band_energies(ks_handler, ik, 0, band_e(1, ik))
      call sirius_get_band_energies(ks_handler, ik, 1, band_e(1, nk + ik))
    end do

  endif

  ! convert to Ry
  do ik = 1, nks
    et(:, ik) = 2.d0 * band_e(:, global_kpoint_index(nkstot, ik))
  enddo

  deallocate(band_e)

end subroutine get_band_energies_from_sirius

To compile Quantum ESPRESSO with SIRIUS it is easiest to use Spack. The following installs a CUDA enabled version:

spack install q-e-sirius ^sirius +shared +scalapack +cuda ~apps ^intel-mkl ^mpich

Now you can load pw.x and MPI related executables:

spack load q-e-sirius

Run pw.x using the same parameters and input files as you would with native QE. Note that you have to explicitly enable SIRIUS through the command-line option -sirius in pw.x. For instance:

# run in default mode
pw.x -i pw.in
# run with SIRIUS enabled
pw.x -i pw.in -sirius

The SIRIUS library is using OpenMP for node-level parallelization. To run QE/SIRIUS efficiently, follow these simple rules:

• always prefer k-point pool parallelization over band parallelization
• use as few MPI ranks as possible for band parallelization
• by default, use one rank per node and many OMP threads; if the calculated system is really small, try to saturate the GPU card by using more MPI ranks (e.g.: on a 12-core node, use 2-3-4 ranks with 6-4-3 OMP threads)

Benchmarks

In the following examples we compare the performance of native and SIRIUS-enabled versions of QE. CPU-only runs are executed on dual-socket multi-core nodes containing two 18-core Intel Broadwell CPUs. GPU runs are executed on hybrid nodes containing a 12-core Intel Haswell CPU and an NVIDIA Tesla P100 card:

Ground state calculation (input) of Si511Ge.

Another example is the variable cell relaxation of B6Ni8 (input). The Brillouin zone contains 204 irreducible k-points and only k-pool parallelization is used.

CP2K

CP2K uses the SIRIUS library to enable plane-wave functionality. The detailed description of the input parameters can be found here under the /CP2K_INPUT/FORCE_EVAL/PW_DFT section.

Contacts

If you have any questions, feel free to contact us:

• Anton Kozhevnikov ([email protected])
• Mathieu Taillefumier ([email protected])
• Simon Pintarelli ([email protected])

Acknowledgements

The development of the SIRIUS library would not be possible without support of the following organizations:

Logo	Name	URL
	Swiss Federal Institute of Technology in Zürich	https://www.ethz.ch/
	Swiss National Supercomputing Centre	https://www.cscs.ch/
	Platform for Advanced Scientific Computing	https://www.pasc-ch.org/
	NCCR MARVEL Centre on Computational Design and Discovery of Novel Materials	https://nccr-marvel.ch/
	MAX (MAterials design at the eXascale) European Centre of Excellence	http://www.max-centre.eu/
	Partnership for Advanced Computing in Europe	https://prace-ri.eu/