Skip to content

vasp-aurora.md

Latest commit

 

History

History
106 lines (77 loc) · 3.94 KB

vasp-aurora.md

File metadata and controls

106 lines (77 loc) · 3.94 KB
Raw

VASP on Aurora (Intel Data Center GPU Max / PVC)

This page records which VASP 6.6.0 makefile was used for the Aurora build in this repository without committing the proprietary VASP source tree.

Repository policy

  • The VASP source tree under external/vasp6/src/vasp.6.6.0/ is gitignored.
  • The original archive external/vasp6/src/vasp.6.6.0.tgz is gitignored.
  • Only build notes, provenance, and non-proprietary configuration summaries are committed.

Makefile provenance

  • VASP version: 6.6.0
  • Upstream template selected: arch/makefile.include.oneapi_omp_off
  • Local working file used for the Aurora build: external/vasp6/src/vasp.6.6.0/makefile.include
  • Aurora build script: scripts/setup/aurora/build-vasp-gpu-aurora.sh

The Aurora build started from the VASP-provided Intel oneAPI OpenMP offload template and was then adjusted locally for Aurora's module stack and Intel PVC GPUs.

The script builds the three standard executables in one invocation by running the VASP targets std, gam, and ncl sequentially.

Aurora-specific changes applied

The committed record here is intentionally descriptive rather than a verbatim copy of the proprietary makefile.

  • Switched the Fortran and C MPI entry points to Aurora's wrappers: mpifort and mpicc
  • Kept Intel compilers for non-wrapper C/C++ compilation where needed: icx and icpx
  • Retained OpenMP offload for Intel GPUs with PVC device targeting
  • Added Level Zero linkage needed by the oneAPI offload stack on Aurora
  • Kept Intel MKL for BLAS, LAPACK, FFT, and ScaLAPACK linkage
  • Used host-specific CPU tuning for the build host via -xHOST
  • Preserved the OMP_OFFLOAD preprocessor path required by the GPU build

Build command

Run from the repository root:

bash scripts/setup/aurora/build-vasp-gpu-aurora.sh

Useful overrides:

NCORES=32 bash scripts/setup/aurora/build-vasp-gpu-aurora.sh
CLEAN_BUILD=1 bash scripts/setup/aurora/build-vasp-gpu-aurora.sh
VASP_TARGET=gam bash scripts/setup/aurora/build-vasp-gpu-aurora.sh
PREFIX=build-aurora bash scripts/setup/aurora/build-vasp-gpu-aurora.sh

Validated build state

The Aurora tuning above was used for the successful GPU-enabled build of:

  • bin/vasp_std

The scripted default build target produces all three executables:

  • bin/vasp_std
  • bin/vasp_gam
  • bin/vasp_ncl

If additional VASP targets are built later, extend this page with the date, target name, and any makefile deltas beyond the base Aurora configuration.

Rebuild checklist

When reproducing the Aurora build, confirm these points before compiling:

  1. Start from arch/makefile.include.oneapi_omp_off in the VASP 6.6.0 source.
  2. Re-apply the Aurora wrapper/compiler adjustments summarized above.
  3. Keep the edited file only as the local external/vasp6/src/vasp.6.6.0/makefile.include working copy.
  4. Do not commit the VASP source tree, generated objects, or the vendor tarball.

Related repository docs

  • Aurora setup overview: scripts/setup/aurora/README.md
  • HPC index: docs/hpc-platforms.md
  • Top-level project guide: README.md

Driving VASP from Python

Two paths are exposed:

  • Active-learning labelling (single-point SCF, no ionic motion): driven by src/matsim_agents/active_learning/backends/vasp.py using a user-supplied INCAR.template. This is what matsim-agents al run invokes when dft.backend: vasp is set in the AL config.
  • Standalone DFT relaxations (scf / relax / vc-relax / vc-relax-shape): driven by src/matsim_agents/tools/vasp_relax.py, the symmetric pair to tools/qe_relax.py. It writes a self-contained INCAR (no template required for the common case), composition-aware ENCUT/ISMEAR/SIGMA/KSPACING, and parses vasprun.xml+OUTCAR for the full per-ionic-step trajectory. Set MATSIM_VASP_LAUNCHER to point at a wrapper script (analogue of the QE launcher pattern).