FlagSparse Release Notes

Contents

FlagSparse Release Notes#

v0.2.0#

Note

This is a preview release. The version number shown is a pre-release identifier and may change upon final release. Content in this preview is for reference only and does not constitute a commitment or warranty for the final product.

  • Added Features

    • SpMM Operators — SpMM COO, SpMM CSR-opt, SpMM CSR-opt-alg2, AlphaSparse SpMM alg1 (with TLE and TLE-opt variants).

      • Supported on NVIDIA.

    • SpGEMM Operator — Sparse matrix-sparse matrix multiplication for CSR inputs.

      • Supported on NVIDIA.

    • SDDMM Operator — Sampled dense-dense matrix multiplication on CSR sparsity pattern.

      • Supported on NVIDIA.

    • SpSM Operator — Sparse triangular solve with dense matrix right-hand sides (CSR and COO).

      • Supported on NVIDIA.

    • SpSV Descriptor API — Full descriptor, buffer-size, analysis, preprocess, and solve API for SpSV workflows.

      • Supported on NVIDIA.

    • Sparse Format Constructors — create_csr_matrix, create_coo_matrix, create_csc_matrix, create_bsr_matrix, create_sell_matrix, create_blocked_ell_matrix, coo_to_csr, coo_to_csc, coo_to_bsr, coo_to_sell, coo_to_blocked_ell, generate_random_sparse_matrix, read_mtx_file.

      • Supported on NVIDIA.

    • Operator Registry — Added conf/operators.yaml with full operator metadata and ops_support.csv / ops_support.py for tracking backend support.

      • Supported on NVIDIA.

    • CI/CD Pipeline — GPU CI workflow, nightly CPU checks, release drafter, pre-commit config, and Triton smoke tests.

      • Supported on NVIDIA.

  • Enhanced Features

    • SpMV CSR and COO operators underwent significant optimization and accuracy improvements.

    • SpSV CSR and COO solvers expanded with comprehensive test coverage and descriptor-based API.

    • Gather/Scatter operators enhanced with int64 support.

    • Benchmark framework unified with standardized shape configurations and performance utilities.

    • pytest accuracy suite expanded with dedicated test files for all operators.

    • CI pipeline added GPU benchmark workflow, packaging metadata tests, and public API validation.

v0.1.0#

Initial release of FlagSparse.

  • Added Features

    • GPU sparse operations package with SpMV, SpMM, SpGEMM, SDDMM, gather, and scatter operators.

    • CSR and COO sparse format support.

    • SpSV and SpSM triangular solve operators.

    • FlagGems-style operator interface registry (conf/operators.yaml).

    • pytest accuracy suite with CPU-FP64 golden reference comparison.

    • Performance benchmark framework with two-level speedup reporting.

    • Native CLI --help and README command snippets in CI smoke tests.

Contents