Skip to content

v4.6.0

Latest

Choose a tag to compare

@jameslehoux jameslehoux released this 17 Nov 22:33
c8d71f6

Highlights

OpenImpala v4.6.0 exposes the complete C++ transport and microstructural characterisation suite in Python with full pure-Python fallback parity, adds a universal py3-none-any wheel so pip install openimpala works out of the box on macOS, Windows, and Linux ARM64, adds Python 3.13 and NVIDIA Ada Lovelace (sm_89) CUDA support, and substantially reduces both CPU and GPU wheel sizes.


Full Python API & Pure-Python Parity (#304)

  • 3-D Steady-State Concentration / Potential Fields: Pass return_field=True to oi.tortuosity() or oi.tortuosity_all() to retrieve TortuosityResult.solution_field (float64 NumPy array on $(z, y, x)$) alongside TortuosityResult.effective_diffusivity.
  • Multi-Phase Variable-Diffusivity Transport: Pass phase_diffusivities={phase_id: D_i} (dimensionless or physical SI $\text{m}^2/\text{s}$) to oi.tortuosity(), oi.tortuosity_all(), and oi.effective_diffusivity(), with scale-invariant solver and plane-flux diagnostics.
  • Full $3 \times 3$ Effective Diffusivity Tensor: New oi.effective_diffusivity() solves the three coupled periodic homogenisation cell problems and returns EffectiveDiffusivityResult (D_eff, tortuosity_tensor, principal_values, anisotropy_ratio, porosity).
  • Microstructural Metrics in Python:
    • oi.specific_surface_area() — interfacial area per unit total and phase volume (supports anisotropic voxel spacing).
    • oi.connected_components() — 6- and 26-connectivity 3-D phase labelling with .label_field(), largest-cluster fraction, and percolating cluster counts.
    • oi.particle_size_distribution() — Euclidean distance-transform local-thickness / spherical-equivalent diameter histogram ($d_{10}, d_{50}, d_{90}$).
    • oi.through_thickness_profile() — slice-by-slice phase volume fraction along x, y, or z.
    • oi.rev_study() — Representative Elementary Volume (REV) convergence sweep across increasing sub-volume fractions.
  • Non-Cubic & Physical SI Voxel Spacing: Pass voxel_size=(dz, dy, dx) (or isotropic scalar voxel_size=dx) across Python facade functions and --voxel-size DX [DY DZ] on the openimpala CLI.

Universal pip install openimpala Everywhere (#304)

  • Publishes a pure-Python fallback wheel (openimpala-4.6.0-py3-none-any.whl, 29 KB) alongside compiled manylinux_2_28_x86_64 MPI/AMReX/HYPRE wheels.
  • pip install openimpala now works out of the box with zero compiler setup on macOS (Apple Silicon & Intel), Windows, Linux ARM64, and future Python versions using the built-in NumPy/SciPy solver backend, while Linux x86_64 automatically selects the compiled AMReX/HYPRE/MPI wheel.

Python 3.13, Ada Lovelace (sm_89) CUDA Support & Slimmer Wheels (#303, #304)

  • Python 3.9 – 3.13: Pre-built openimpala (CPU) and openimpala-cuda (CUDA 12) wheels now include cp313.
  • Expanded GPU Architecture Coverage: openimpala-cuda now embeds native SASS for both NVIDIA Ampere (sm_80) and Ada Lovelace (sm_89 — RTX 40-series, L4, L40S) plus Turing (compute_75) PTX forward-compatibility.
  • Slimmer Binaries & OpenMPI 5 Compatibility:
    • CPU wheels shrunk by 34% (6.4 MB → 4.2 MB) by dropping unused Fortran and AMReX_EB dependencies and removing deprecated libmpi_cxx.so linkage.
    • CUDA GPU wheels shrunk by 56% (204 MB → 89.7 MB) by disabling redundant AMReX relocatable device code (AMReX_GPU_RDC=OFF) and compressing device fatbins.

Merged Pull Requests

  • #303 — build(v4.5.1): drop Fortran & AMReX EB, add cp313 & sm_89, skip libmpi_cxx
  • #304 — feat: full Python API, universal fallback wheel, cp313 & Ada sm_89 wheels

What's Changed

  • build(v4.5.1): drop Fortran & AMReX EB, add cp313 & sm_89 by @jameslehoux in #303
  • feat(v4.6.0): full Python API, universal pure-Python wheel, and fast wheel CI by @jameslehoux in #304

Full Changelog: v4.5.0...v4.6.0