v26.02.00

Changes in this release:

• Support DUNDER methods for addition, subtraction and negation.
• Introduce generalized element-wise addition of sparse CSR matrices -- Check geam method in csr_array class.
• Support block_array -- sub-blocks of sparse CSR matrices can now be used to create a new bigger CSR matrix.
• Support eigsh -- Finds k eigenvalues and eigenvectors of the real symmetric square matrix or complex Hermitian matrix A.
• Support spsolve -- Sparse direct linear solver
• Update examples
• Improved Typing support

The documentation can be accessed at https://nv-legate.github.io/legate-sparse/.

Conda packages for v26.02 are available here: https://anaconda.org/legate/legate-sparse.

v25.07.00

This release improves the documentation of Legate Sparse. Documentation now includes:

• Docstrings for classes, methods, examples and tests.
• API reference for supported APIs and sparse matrix formats
• Installation, User guide, Best Practices and Developer Guide

The documentation can be accessed at https://nv-legate.github.io/legate-sparse/.

Conda packages for v25.07 are available here: https://anaconda.org/legate/legate-sparse.

v25.03.00

Conda packages for v25.03 are available here: https://anaconda.org/legate/legate-sparse.

Notes

Features:

• Nonzero API for CSR matrices is now supported
• Setting a CSR matrix with a boolean CSR matrix to a scalar value is now supported
• Comparison of a CSR matrix to a scalar returning a boolean CSR matrix is now supported
• spGEMM for matrices with integer and boolean datatypes is supported for OMP and CPU processor variants

Bug Fixes:

• An error while instantiating a CSR matrix with empty rows, cols, and vals is now resolved.

v25.01.00

This release supports the unification of two different memory pools in Legate - eager and deferred. As a result of this unification of memory pools, the option --eager-alloc-percentage is no longer supported. Misalignment errors for sparse matrix-matrix multiplication have been resolved.

V24.11.00

v24.11.00 is an alpha release of Legate Sparse and now uses the C++ Legate core.

Conda packages are available here: https://anaconda.org/legate/legate-sparse. Note that the link to legate sparse homepage is incorrect on the conda page.

Notes

• The implementation of spGEMM has computational hints that prevent copying over data from GPUs to CPU and vice-versa for partitioning. Multi-GPU runs that use spGEMM may benefit from this improvement

• Support for two different spGEMM algorithms from cuSparse is now included. A fast but memory hungry version can be turned on by setting the environment variable LEGATE_SPARSE_FAST_SPGEMM to 1

• Support for all matrix formats other than CSR has been dropped in this release but might be included over the next several releases

SC 2023

Experiments for SC 2023.