openmx3.9.9_gpu

What does this code do?

This is a GPU-accelerated version of OpenMX, a first-principles calculation code based on numerical atomic orbitals (NAO). Currently, only certain processes (matrix multiplication and eigenvalue problem processing) for band calculations (collinear and non-collinear) and cluster calculations (collinear and non-collinear) are GPU-accelerated.

Code author

Hiroyuki Kawai (Niigata Univ.)
X account: @dc1394

How to enable GPU acceleration

To enable GPU acceleration, you must specify "cusolver" for "scf.eigen.lib" in the input file (⚠️ GPU acceleration is disabled by default!). For example:

scf.XcType                  GGA-PBE    # LDA|LSDA-CA|LSDA-PW|GGA-PBE
scf.SpinPolarization        off        # On|Off|NC
scf.ElectronicTemperature  300.0       # default=300 (K)
scf.energycutoff           150.0       # default=150 (Ry)
scf.maxIter                 40         # default=40
scf.EigenvalueSolver       band        # DC|GDC|Cluster|Band
scf.Kgrid                  9 9 9       # means n1 x n2 x n3
scf.Mixing.Type           rmm-diisk    # Simple|Rmm-Diis|Gr-Pulay|Kerker|Rmm-Diisk
scf.Init.Mixing.Weight     0.30        # default=0.30
scf.Min.Mixing.Weight      0.001       # default=0.001 
scf.Max.Mixing.Weight      0.700       # default=0.40 
scf.Mixing.History          7          # default=5
scf.Mixing.StartPulay       5          # default=6
scf.criterion             1.0e-10      # default=1.0e-6 (Hartree) 
scf.eigen.lib             cusolver     # default=elpa1

Set as shown above.

Build and install

Building and installing is more difficult than with standard OpenMX. The build requires the NVIDIA HPC SDK and OpenMPI. The Makefile contains build examples for several supercomputer systems; please refer to them. If you're unsure about the build and installation process, feel free to ask in English via GitHub issues or my X account (Japanese is also acceptable on my X account). I'll assist you as much as I can.

Docker image

I have released the OpenMX 3.9.9 GPU Docker image. You can easily try OpenMX 3.9.9 GPU on computers equipped with NVIDIA GPUs. The steps are as follows:

1. Install Docker.
2. Install the NVIDIA Container Toolkit.
3. Run: docker run --gpus all --shm-size=4gb --rm -it -v /path/to/inputs:/work dc1394/openmx3.9.9-gpu-ubuntu24.04:0.1. Ensure /path/to/inputs is created beforehand.
4. Run tests with cd openmx_work and mpirun -np 4 ./openmx -runtest.

This should yield results like the following:

  1  input_example/Benzene.dat        Elapsed time(s)=    6.43  diff Utot= 0.000000000044  diff Force= 0.000000000008
  2  input_example/C60.dat            Elapsed time(s)=  49.17  diff Utot= 0.000000000002  diff Force= 0.000000000001
  3  input_example/CO.dat            Elapsed time(s)=    9.99  diff Utot= 0.000000000072  diff Force= 0.000000001358
  4  input_example/Cr2.dat            Elapsed time(s)=    8.04  diff Utot= 0.000000000439  diff Force= 0.000000000049
  5  input_example/Crys-MnO.dat      Elapsed time(s)=  80.04  diff Utot= 0.000000000038  diff Force= 0.000000002306
  6  input_example/GaAs.dat          Elapsed time(s)=  101.35  diff Utot= 0.000000000021  diff Force= 0.000000000003
  7  input_example/Glycine.dat        Elapsed time(s)=    5.52  diff Utot= 0.000000000001  diff Force= 0.000000000001
  8  input_example/Graphite4.dat      Elapsed time(s)=    7.52  diff Utot= 0.000000000019  diff Force= 0.000000000005
  9  input_example/H2O-EF.dat        Elapsed time(s)=    5.11  diff Utot= 0.000000000105  diff Force= 0.000000000002
  10  input_example/H2O.dat            Elapsed time(s)=    4.96  diff Utot= 0.000000000102  diff Force= 0.000000001624
  11  input_example/HMn.dat            Elapsed time(s)=  12.61  diff Utot= 0.000000000345  diff Force= 0.000000000011
  12  input_example/Methane.dat        Elapsed time(s)=    4.35  diff Utot= 0.000000000006  diff Force= 0.000000000001
  13  input_example/Mol_MnO.dat        Elapsed time(s)=    8.89  diff Utot= 0.000000000584  diff Force= 0.000000000068
  14  input_example/Ndia2.dat          Elapsed time(s)=  23.17  diff Utot= 0.000000000000  diff Force= 0.000000000001


Total elapsed time (s)      327.16

You can verify that the calculation is correct.

Benchmarks

For benchmarks of GPU-accelerated OpenMX, please refer to the following literature. https://journals.jps.jp/doi/10.7566/JPSJ.94.124003

However, the current version offers improved performance compared to the version described in this paper.

Important notes

At present, GPU-accelerated OpenMX performs faster than standard OpenMX for calculations involving systems containing hundreds of atoms. For calculations involving systems with fewer than a hundred atoms, standard OpenMX should be used. Please use with caution as it may contain bugs.

About bug reports

I would appreciate it if you could actively report any bugs. Please report them via GitHub issues or send them to my X account. Bug reports sent to my X account can be in English.