Developer Reference for Intel® oneAPI Math Kernel Library for C
Sparse BLAS Routines
Intel® oneAPI Math Kernel Library (oneMKL) provides Sparse BLAS optimized routines for linear algebra operations on sparse matrices. The routines are C APIs and we also include a Fortran 90 version so that all routines can be called from C, C++, and modern Fortran applications. The library offers two distinct Sparse BLAS solutions:
Inspector-Executor Sparse BLAS (recommended): this interface uses an approach that separates analysis (inspector) from computation (executor). The inspector stage analyzes matrix structure and caches optimization information, while the executor stage performs operations using these cached optimizations for improved performance. This API uses an opaque handle-based interface and supports standard sparse matrix storage formats including Compressed Sparse Row (CSR), Compressed Sparse Column (CSC), Coordinate (COO), and Block Sparse Row (BSR). See Inspector-Executor Sparse BLAS Routines for complete documentation.
Legacy Sparse BLAS: this is the legacy interface following the traditional NIST Sparse BLAS specification with single-stage operations and direct array-based interfaces. Level 1 routines provide sparse vector operations. The current release includes only selected Level 2 and Level 3 routines, all of which have been previously deprecated and are scheduled for removal in future releases. See Legacy Sparse BLAS Routines for details.
This section also includes Sparse QR Routines for QR factorization and least squares solvers for sparse matrices. The Sparse QR routines share the same data structures and matrix handles as the Inspector-Executor Sparse BLAS API.
For reference information on Inspector-Executor sparse matrix handles, constants, and data types, see Sparse BLAS Reference Types. For a detailed description of the supported sparse matrix formats and their mathematical definitions, see Sparse Matrix Storage Formats in the appendix.
Product and Performance Information |
|---|
Performance varies by use, configuration and other factors. Learn more at www.intel.com/PerformanceIndex. Notice revision #20201201 |