FlagSparse Overview

FlagSparse is part of FlagOS. FlagSparse is a computing library that provides sparse matrix operations and is oriented towards multiple chip backends. It defines core sparse operations such as SpMV, SpMM, SpGEMM, and SDDMM, supporting high-performance computing in fields such as scientific computing, engineering simulation, machine learning, and artificial intelligence.

FlagSparse is a high-performance sparse operator library implemented using the Triton programming language launched by OpenAI.

Operators

The complete operator registry is maintained at FlagSparse conf/operators.yaml.

Operator	Description	Formats
Gather	Gather values from dense tensor by index	–
Scatter	Scatter values into dense tensor by index	–
SpMV	Sparse matrix-vector multiplication	CSR, COO, COO-to-CSR
SpMM	Sparse matrix-dense matrix multiplication	CSR, COO, CSR-opt, CSR-opt-alg1, CSR-opt-alg2, AlphaSparse-alg1
SpGEMM	Sparse matrix-matrix multiplication	CSR
SDDMM	Sampled Dense-Dense Matrix Multiplication	CSR
SpSV	Sparse triangular solve (vector)	CSR, COO
SpSM	Sparse triangular solve (matrix)	CSR, COO

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

Sparse Formats

• CSR (Compressed Sparse Row) – Standard format for row-oriented sparse operations.

• COO (Coordinate format) – Triples of (row, column, value).

Architecture

• src/flagsparse/ – Core package. sparse_operations/ is emitted as several .py modules from string literals in flagsparse.py.

• tests/ – Pytest accuracy tests and benchmarks.

• benchmark/ – Performance benchmarks.