OpenBLAS 32-bit Integer (LP64) Provider for SciPy/NumPy

Warnings

• breaking Do NOT add a direct runtime dependency on `scipy-openblas32` if your project is not NumPy or SciPy. This is explicitly unsupported and highly likely to lead to breakage or symbol conflicts due to internal changes in this repository or due to NumPy/SciPy starting to depend on a different version. This package is solely for use by SciPy and NumPy.
  Fix: Avoid adding `scipy-openblas32` as a direct runtime dependency unless you are a maintainer of NumPy or SciPy. Use environment-specific solutions (e.g., Conda) or vendor OpenBLAS directly if you require a specific BLAS implementation for your own project.
• gotcha The `scipy-openblas32` package provides the LP64 ABI (32-bit integer size for array indexing in BLAS functions). This is distinct from `scipy-openblas64`, which provides the ILP64 ABI (64-bit integer size). Both packages can be installed on 64-bit systems; the '32' or '64' refers to the integer size used by the BLAS library internally, not the system architecture. SciPy currently *always requires LP64* but can optionally be built with ILP64 support as well.
  Fix: Understand that '32' refers to LP64 ABI, not system architecture. If building SciPy/NumPy from source with ILP64 support, you may also need `scipy-openblas64` alongside `scipy-openblas32` for dual-ABI support. Refer to SciPy's BLAS/LAPACK documentation for detailed build instructions.
• gotcha There is no strong guarantee of backwards compatibility for the symbol names or the small Python API (`get_include_dir`, etc.). Projects using `scipy-openblas32` as a build dependency should pin exact versions (e.g., `==`) in their CI/lock files to prevent unexpected breakage.
  Fix: Strictly pin the version of `scipy-openblas32` in your build environment to ensure stability.
• gotcha On Windows, dynamic linking of OpenBLAS DLLs can be problematic due to the lack of RPATH support. If not vendored directly into a wheel (as SciPy does), you may need to manually manage the DLL search path using `os.add_dll_directory` or tools like `delvewheel` during custom builds.
  Fix: Ensure OpenBLAS DLLs are correctly located by the system or explicitly add their directory to the DLL search path at runtime or build time. When building SciPy/NumPy from source on Windows, consult their specific build documentation.
• gotcha Mixing OpenBLAS builds (e.g., vendored `pthreads`-based OpenBLAS in wheels with system-installed `OpenMP`-based OpenBLAS) can lead to 'oversubscription' issues, where multiple parallelization mechanisms conflict, potentially degrading performance or causing crashes.
  Fix: Be mindful of the OpenMP/pthreads configuration of BLAS libraries in your environment. Where possible, standardize on one parallelization backend or ensure that only compatible BLAS implementations are used. If custom building, configure OpenBLAS with the desired threading model.
• deprecated The SciPy documentation indicates that the requirement for SciPy to *always* require the LP64 ABI (32-bit integer) is "very likely changing for the 1.18.0 release." This could imply a shift in default ABI preference for SciPy's internal builds, potentially affecting how `scipy-openblas32` is prioritized or used in the future if ILP64 becomes more central.
  Fix: Monitor SciPy's release notes for version 1.18.0 and beyond to understand changes in its default BLAS/LAPACK ABI requirements and adjust build strategies accordingly.

Install

• pip install scipy-openblas32 Install via pip

Imports

• get_include_dir

  import scipy_openblas32 as so32; so32.get_include_dir()

  This library is primarily for build-time integration by other scientific Python packages (e.g., SciPy, NumPy) and not intended for direct runtime imports by end-user applications. Its Python API exposes functions for build systems to locate the bundled OpenBLAS components.
• get_lib_dir

  import scipy_openblas32 as so32; so32.get_lib_dir()

  Primarily for build-time use by packages like SciPy/NumPy.
• get_library

  import scipy_openblas32 as so32; so32.get_library()

  Primarily for build-time use by packages like SciPy/NumPy.
• get_pkg_config

  import scipy_openblas32 as so32; so32.get_pkg_config()

  Primarily for build-time use by packages like SciPy/NumPy to generate pkg-config files for Meson or CMake. This is not for direct use in typical application code.

Quickstart

This quickstart demonstrates how build systems (e.g., for SciPy or NumPy) would query `scipy-openblas32` to find the necessary OpenBLAS components (include paths, library paths, and pkg-config information) for linking during compilation. It is not intended for direct use in typical end-user Python applications, as `scipy-openblas32` is a low-level dependency provider.

import scipy_openblas32 as so32

# This library is typically used by build systems of other packages
# to locate the bundled OpenBLAS libraries for compilation.
# For example, to generate a pkg-config file for a build system:
# (This code snippet demonstrates the *intended interaction* for other libraries)
# In a real build process, the output would be redirected to a file, e.g., scipy-openblas32.pc

print("OpenBLAS include directory:", so32.get_include_dir())
print("OpenBLAS library directory:", so32.get_lib_dir())
print("OpenBLAS pkg-config info:\n", so32.get_pkg_config())