Dense Linear Algebra: AOCL-BLAS and AOCL-LAPACK

AOCL-BLAS

AOCL-BLAS provides a high-performance implementation of the Basic Linear Algebra Subprograms (BLAS), which provide the essential kernels of matrix and vector computation—among the most used computationally intensive operations in dense numerical linear algebra. Select kernels have been optimized for the AMD “Zen”-based processors, including AMD EPYC™, AMD Ryzen™, and AMD Ryzen™ Threadripper™ processors.

AOCL-BLAS is developed as a forked version of BLIS (https://github.com/flame/blis), originally developed by members of the Science of High-Performance Computing (SHPC) group in the Institute for Computational Engineering and Sciences at The University of Texas at Austin and other collaborators (including AMD). The library retains all known BLIS features and adds:

• Standard BLAS and CBLAS interfaces
• C++ template interfaces for BLAS functionalities 

Highlights of AOCL-BLAS 5.2

• Performance optimizations on Zen4, Zen5 architectures
  • DGEMV, DGEMM, ZGEMM, CGEMM, DCOPY
• Multi-threading and General Stride support for DGEMV
• Numerical precision improvement for ZGEMV
• Support for the complete set of GEMMTR APIs
• Build options to disable certain code paths in GEMM and TRSM
• CMake and configure options to enable DTL Logging
• Updates to aocl-gemm add-on module
  • Column Major support for BF16 and FP32
  • FP32 RD kernels for AVX512 and AVX2 ISA
  • GEMV kernel for m=1 case using AVX2 and AVX512 YMM registers
• You can find the package containing AOCL-BLAS Library binaries that includes optimizations for AMD processors, examples, and documentation in the Downloads section.

Documentation

• Source code: GitHub.

AOCL-LAPACK

AOCL-LAPACK is a high-performance implementation of Linear Algebra PACKage (LAPACK), which provides routines for solving systems of linear equations, least-squares problems, eigenvalue problems, singular value problems, and the associated matrix factorizations. Extensible, easy to use, and available under an open-source license, AOCL-LAPACK  can be utilized by applications relying on standard Netlib LAPACK interfaces with virtually no changes to their source code. AOCL-LAPACK supports C, Fortran, and C++ template interfaces (for a subset of APIs) for the LAPACK APIs.

AOCL-LAPACK is compatible with the LAPACK 3.12.0 specification. Combined with the AOCL-BLAS library, which includes optimizations for the AMD “Zen”-based processors, AOCL-LAPACK enables running high performing LAPACK functionalities on AMD platforms.

Highlights of AOCL-LAPACK 5.2

• Improved performance of the following routines
  • LU, Cholesky and QR Factorizations (DGETRF, DPOTRF & DGEQRF)
  • Symmetric Eigen Decomposition (DSYEVD)
  • Matrix Inverse routines (DGETRI & DPOTRI) for small sizes
• Build System update
  • AOCL-LAPACK now supports only CMAKE based build; autoconf based build is not supported anymore
  • Added two new modes (avx2-strict, avx512-strict) under LF_ISA_CONFIG build flag to enforce specific ISA during execution
• Test suite framework enhancements
  • Bit Reproducibility tests for test-suite supported LAPACK APIs
  • Introduced Benchmark Mode to run tests for fixed duration, displaying additional metrics, and more

Documentation

• AOCL-LAPACK API Guide
• Prior versions: AOCL-BLAS and AOCL-LAPACK Library Archive
• Source code: GitHub

Downloads