Sparse_Numba

A lightweight, Numba-compatible sparse linear solver designed for efficient parallel computations in Python.

Why Sparse_Numba?

Existing sparse linear solvers in Python (e.g., SciPy, CVXOPT, KVXOPT) work well for single-task scenarios but face performance bottlenecks due to frequent data exchanges and the Global Interpreter Lock (GIL). Sparse_Numba provides a sparse linear solver fully compatible with Numba's JIT compilation, allowing computationally intensive tasks to run in parallel by bypassing the GIL.

Installation

pip install sparse-numba

Due to licensing, UMFPACK DLLs are not bundled. To use the UMFPACK solver, install SuiteSparse separately and add the DLLs to the system path or place them under:

.venv/site-packages/sparse_numba/vendor/suitesparse/bin

Platform Compatibility

Note: Starting from v0.1.11, macOS pre-built wheels target Apple Silicon (ARM64). Intel Mac users can install from source: pip install sparse-numba --no-binary=sparse-numba. Python 3.8 support has been dropped to align with NumPy and Numba compatibility.

Building from Source (Windows)

1. Install MinGW-w64 (x86_64-posix-seh) and add its bin directory to PATH.
2. Create %USERPROFILE%\.distutils.cfg:

   [build]
   compiler=mingw32
   

   Note: The setting is mingw32 even for 64-bit builds.
3. Build and install:

   python -m build --wheel
   pip install dist/sparse_numba-<VERSION>.whl
   

To build only the SuperLU extension (skip UMFPACK if headers are unavailable):

SPARSE_NUMBA_SKIP_UMFPACK=1 python setup.py build_ext --inplace

Detailed installation information: Installation Guide.

Troubleshooting: SIZEOF_VOID_P Compilation Error (Windows)

When building from source on Windows with MinGW and Microsoft Store Python, you may encounter:

error: enumerator value for '__pyx_check_sizeof_voidp' is not an integer constant

This happens because the MS Store Python's pyconfig.h defines SIZEOF_VOID_P=4 (32-bit) while MinGW compiles for 64-bit. Two fixes are needed:

1. In the generated Cython C files (cy_superlu_wrapper.c, cy_umfpack_wrapper.c), find:

   enum { __pyx_check_sizeof_voidp = 1 / (int)(SIZEOF_VOID_P == sizeof(void*)) };
   

   Replace with:

   //    enum { __pyx_check_sizeof_voidp = 1 / (int)(SIZEOF_VOID_P == sizeof(void*)) };
   enum { __pyx_check_sizeof_voidp = 1 };
   

2. In setup.py, the Windows extra_compile_args already includes -DSIZEOF_VOID_P=8. If building manually, add this flag to your gcc command.

Note: The pre-built wheels on PyPI do not have this issue.

API Reference

Solver Functions (Combined Factorize + Solve)

All functions are @njit(nogil=True) compatible for use inside Numba-compiled code.

Return: (x: float64[:], info: int) where info=0 is success.

Pre-Factorization API (Factorize Once, Solve Many Times)

For systems where the matrix A stays constant across many solves (e.g., linear ODE integration), pre-factorization avoids redundant LU decomposition. Factorize once, then solve with different right-hand side vectors.

Note: The handle is an opaque int64 value. Each handle is independent and thread-safe. The user must call free_factors() when done.

Sparse Utilities

Usage Examples

Basic Solve (Combined Factorize + Solve)

import numpy as np
from sparse_numba.sparse_superlu.superlu_numba_interface import superlu_solve_csc

# CSC format sparse matrix
indptr = np.array([0, 2, 3, 6], dtype=np.int32)
indices = np.array([0, 2, 2, 0, 1, 2], dtype=np.int32)
data = np.array([1.0, 2.0, 3.0, 4.0, 5.0, 6.0])
b = np.array([1.0, 2.0, 3.0])

# Solve Ax = b
x, info = superlu_solve_csc(data, indices, indptr, b)

Pre-Factorization (Factorize Once, Solve Many)

from sparse_numba.sparse_superlu.superlu_numba_interface import (
    superlu_factorize_csc, superlu_solve_factored, superlu_free_factors
)

# Factorize once
handle, info = superlu_factorize_csc(A_csc.data, A_csc.indices, A_csc.indptr)
assert info == 0

# Solve many times with different RHS vectors
for b in list_of_rhs_vectors:
    x, info = superlu_solve_factored(handle, b)

# Free when done
superlu_free_factors(handle)

Parallel Solving with Numba

from numba import njit, prange

@njit(parallel=True)
def solve_many_systems(A_data, A_indices, A_indptr, rhs_list, n_systems):
    solutions = np.zeros((n_systems, len(rhs_list[0])))
    for i in prange(n_systems):
        sol, info = superlu_solve_csc(A_data, A_indices, A_indptr, rhs_list[i])
        solutions[i] = sol
    return solutions

Performance

Single Problem (vs. SciPy spsolve)

Solver	Benchmark
UMFPACK
SuperLU

Multi-task Parallel (vs. sequential SciPy)

Platform	Benchmark	Speedup
Intel Ultra 7 258V
Xeon W-2255

Note: Initial JIT compilation overhead is included in single-problem benchmarks. The performance advantage is most evident in parallel multi-task scenarios.

Pre-Factorization Speedup

For a constant matrix solved repeatedly (e.g., linear ODE time-stepping), pre-factorization avoids redundant LU decomposition at each step. Benchmark on a 200x200 matrix:

Combined with parallel execution (8 threads, 8 independent systems):

To reproduce these benchmarks:

python -m sparse_numba.benchmark_prefactorize_slu

This generates three figures:

• benchmark_repeated_solve_prefactored.png — repeated solve timing and speedup
• benchmark_parallel_prefactored.png — parallel multi-system comparison
• benchmark_thread_scaling_prefactored.png — thread scaling efficiency

Architecture

User code (@njit)
    |
    v
Python/Numba layer (ctypes function pointers, @njit(nogil=True))
    |
    v
Cython layer (thin cdef api wrappers)
    |
    v
C layer (superlu_wrapper.c / umfpack_wrapper.c)
    |
    v
SuperLU / UMFPACK C libraries (vendor DLLs)

All layers release the Python GIL, enabling true parallel execution across threads.

License

BSD 3-Clause License

Third-Party Licenses

• OpenBLAS: github.com/OpenMathLib/OpenBLAS
• SuperLU: github.com/xiaoyeli/superlu
• GNU runtime libraries (libgcc_s_seh-1.dll, libgfortran-5.dll, libgomp-1.dll, libquadmath-0.dll, libwinpthread-1.dll): redistributed from the GNU toolchain

Citation

@software{hong2025sparse_numba,
  author = {Hong, Tianqi},
  title = {Sparse_Numba: A Numba-Compatible Sparse Solver},
  year = {2025},
  publisher = {GitHub},
  url = {https://github.com/th1275/sparse_numba}
}

Release Process

1. Bump version

Edit setup.py and update the version number:

version="0.1.11",  # increment as appropriate

2. Push to main (build + test only, no publish)

git add -A && git commit -m "Release v0.1.11"
git push origin main

Wait for the GitHub Actions CI to pass. The workflow builds and tests wheels on Linux, macOS, and Windows for Python 3.8–3.12. No upload to PyPI occurs on a branch push.

3. Tag and publish to PyPI

git tag v0.1.11
git push origin v0.1.11

The upload_pypi job triggers automatically on tag push — it collects all wheels and publishes via trusted publishing.

4. If CI fails on a tag push

• If the upload_pypi step never ran (build or test failed first): the version was never uploaded to PyPI. You can safely delete the tag, fix the issue, and re-tag:

  git tag -d v0.1.11              # delete local tag
  git push origin :refs/tags/v0.1.11  # delete remote tag
  # fix the issue, commit, push to main
  git tag v0.1.11                 # re-create tag
  git push origin v0.1.11         # re-push tag
  

• If the version was already uploaded to PyPI: PyPI does not allow re-uploading the same version number, ever. You must bump to v0.1.12 and tag again.

5. Recommended workflow

push to main → CI builds + tests → green? → tag vX.Y.Z → CI publishes to PyPI

Always confirm CI passes on the main push before tagging. This avoids wasting version numbers. You can also use workflow_dispatch to manually trigger a build without pushing code.

Building Windows wheels locally (alternative)

If CI Windows builds are not available or you need to build locally:

# Requires MinGW-w64 installed
python setup.py bdist_wheel
# Wheel is created in dist/ folder, named like:
# sparse_numba-0.1.11-cp311-cp311-win_amd64.whl

Build once per Python version you want to support (3.8, 3.9, 3.10, 3.11, 3.12).

Upload manually:

pip install twine
twine upload dist/sparse_numba-0.1.11-cp311-cp311-win_amd64.whl

Contributing

Contributions are welcome. Please open an issue or pull request on GitHub.