17. AOCL-Sparse - 17. AOCL-Sparse - 5.3 English - 57404

AOCL-Sparse is a library containing basic linear algebra subroutines for sparse matrices and vectors (Sparse BLAS). It is designed to be used with C, C++, and compatible languages.

The current functionality of AOCL-Sparse is organized in the following categories:

• Sparse Level 1 functions perform vector operations, such as dot product, vector additions on sparse vectors, gather, scatter and other similar operations. Fused scatter-add and scatter-subtract are available via the KT (Kernel Template) L1 APIs.

• Sparse Level 2 functions describe the operations between a matrix in sparse format and a vector in dense format, including matrix-vector product (SpMV), triangular solve (TRSV), and so on.

• Sparse Level 3 functions describe the operations between a matrix in sparse format and a matrix in dense/sparse format. The operations comprise of matrix additions (SpADD), matrix-matrix product (SpMM, SpMMd, Sp2M, Sp2Md), triangular solver with multiple right-hand sides (TRSM), symmetric product of 3 matrices (sypr, syprd) and matrix-its transpose product (syrk, syrkd). Both CSR and CSC input formats are supported for sypr, syrk, syprd, and syrkd.

• Iterative sparse solvers based on Krylov subspace methods (CG, GMRES) and preconditioners (such as SymGS, SOR and ILU0).

• Sparse format conversion functions for translating matrices in a variety of sparse storage formats.

• Auxiliary functions to allow basic operations, including create, copy, destroy, and modify matrix handles and descriptors.

Additional Highlights

• Supported data types: single, double, and the complex variants

• 0-based and 1-based indexing of sparse formats

• Hint & Optimize framework to accelerate the supported functions by a prior matrix analysis based on the users’ hints of expected operations.

Multi-Thread Support

You can set the total number of threads using the environment variables AOCLSPARSE_NUM_THREADS or OMP_NUM_THREADS. If both environment variables are set, AOCL-Sparse gives higher precedence to AOCLSPARSE_NUM_THREADS. If neither variable is set, the default number of threads is equal to the number of processors identified by the OpenMP library. The functions with multi-thread support include SpMV variants for real and complex datatypes, TRSM, SpMM, Sp2M, CSRMM and SpAdd.

The ISA code path can be selected at runtime using the AOCL_ENABLE_INSTRUCTIONS environment variable (values: GENERIC, AVX2, AVX512), selecting among the code paths compiled into the library. Requesting a higher ISA than the build supports (USE_AVX512 off) or the CPU supports silently falls back to the highest available level; no error is returned. The programmatic equivalent is aoclsparse_enable_instructions().

For more information on the AOCL-Sparse APIs, refer AOCL-Sparse_API_Guide.pdf in the docs folder (amd/aocl-sparse).