Python bindings for sparse-ir-capi

This is a low-level binding for the sparse-ir-capi Rust library.

Requirements

• Python >= 3.10
• Rust toolchain (for building the Rust library)
• uv (for building and managing Python dependencies)
• numpy >= 1.26.4
• scipy

BLAS Support

This package automatically uses SciPy's BLAS backend for optimal performance. No additional BLAS installation is required - SciPy will provide the necessary BLAS functionality.

Build

Install Dependencies and Build

# Build the package (Rust library will be built automatically)
cd python
uv build

This will:

• Automatically build the Rust sparse-ir-capi library using Cargo (via CMake)
• Copy the built library and header files to the Python package
• Create both source distribution (sdist) and wheel packages

Development Build

For development:

# Install in development mode (will auto-prepare if needed)
uv sync --locked

Note for CI/CD: The Rust library is built automatically during the Python package build. No separate build step is needed:

# In CI/CD scripts
cd python
uv build

See .github-workflows-example.yml for a complete GitHub Actions example.

BLAS Configuration

The package automatically uses SciPy's BLAS backend, which provides optimized BLAS operations without requiring separate BLAS installation. The build system is configured to use SciPy's BLAS functions directly.

Clean Build Artifacts

To remove build artifacts and files copied from the parent directory:

uv run clean

This will remove:

• Build directories: build/, dist/, *.egg-info
• Compiled libraries: pylibsparseir/*.so, pylibsparseir/*.dylib, pylibsparseir/*.dll
• Cache directories: pylibsparseir/__pycache__

Build Process Overview

The build process works as follows:

1. CMake Configuration: scikit-build-core invokes CMake, which:

   • Finds the Cargo executable
   • Sets up build targets for the Rust library

2. Rust Library Build: CMake calls Cargo to build sparse-ir-capi:

   • Compiles the Rust library to a shared library (.so, .dylib, or .dll)
   • Generates C header file (sparseir.h) using cbindgen (via build.rs)
   • Copies the library and header to the pylibsparseir directory

3. Python Package Building: uv build or uv sync --locked:

   • Packages everything into distributable wheels and source distributions

4. Installation: The built package includes the compiled shared library and Python bindings

Conda Build

This package can also be built and distributed via conda-forge. The conda recipe is located in conda-recipe/ and supports multiple platforms and Python versions.

Building conda packages locally:

# Install conda-build
conda install conda-build

# Build the conda package
cd python
conda build conda-recipe

# Build for specific platforms
conda build conda-recipe --platform linux-64
conda build conda-recipe --platform osx-64
conda build conda-recipe --platform osx-arm64

Supported platforms:

• Linux x86_64
• macOS Intel (x86_64)
• macOS Apple Silicon (ARM64)

Supported Python versions:

• Python 3.11, 3.12, 3.13, 3.14

Supported NumPy versions:

• NumPy 2.1, 2.2, 2.3

The conda build automatically:

• Uses SciPy's BLAS backend for optimal performance
• Cleans up old shared libraries before building
• Builds platform-specific packages with proper dependencies

Handle Ownership

Every function in pylibsparseir.core that creates a C object returns an owning handle instead of a raw ctypes pointer: KernelHandle, SVEResultHandle, BasisHandle, FuncsHandle, SamplingHandle, or GemmBackendHandle (the default BLAS backend returned by get_default_blas_backend()).

• A handle is released exactly once with the matching spir_*_release: when it becomes unreachable, when close() is called (also on leaving a with handle: block), or at interpreter exit. Closing it again does nothing.
• An open handle can be passed directly to any _lib.spir_* function and is truthy. A released handle is falsy, and passing it to C raises ctypes.ArgumentError instead of handing C a freed pointer.
• Do not release an owned handle through _lib.spir_*_release. Those entry points refuse owned handles with ctypes.ArgumentError, because the owner would free them a second time. Handles created directly through _lib remain the caller's to release, or can be handed over to an owner, e.g. FuncsHandle(_lib.spir_funcs_clone(funcs)).
• C handles never depend on the handles they were created from (spir_basis_new copies the kernel and the SVE result; funcs and samplings own their data), so handles can be released in any order.

Performance Notes

BLAS Support

This package automatically uses SciPy's optimized BLAS backend for improved linear algebra performance:

• Automatic BLAS: Uses SciPy's BLAS functions for optimal performance
• No additional setup: SciPy provides all necessary BLAS functionality

The build system automatically configures BLAS support through SciPy. You can verify BLAS support with debug output enabled:

export SPARSEIR_DEBUG=1
python -c "import pylibsparseir"

This will show:

[core.py] Created SciPy BLAS backend handle
[core.py] Registered SciPy BLAS dgemm @ 0x...
[core.py] Registered SciPy BLAS zgemm @ 0x...

Debug Output

The SPARSEIR_DEBUG environment variable enables debug output only when it is set to 1, true, yes or on, in any letter case. Any other value, including 0, false or an empty string, disables it, as does leaving it unset. The same rule applies to both layers:

• pylibsparseir prints its messages to stdout. It checks the variable when it is imported.
• The Rust library prints [SPARSEIR DEBUG], [SPARSEIR DEBUG ERROR] and [SPARSEIR WARN] lines to stderr. It checks the variable each time it would print one.

Troubleshooting

Build fails with missing Cargo:

# Make sure Rust toolchain is installed
# Install from https://rustup.rs/
# Then retry:
cd python
uv build

Clean rebuild:

# Remove all build artifacts
uv run clean
cd ../sparse-ir-capi
cargo clean
cd ../python
uv build