AOCL-BLAS is an optimized BLAS library that provides high-performance linear algebra operations for AMD EPYC and Ryzen processors.
The library addresses the need for CPU-specific optimization of basic linear algebra subroutines by implementing AMD-tuned versions of BLAS operations. It is developed as a fork of BLIS, retaining all upstream features while adding AMD-specific optimizations. The tool uses dynamic dispatch to select optimized code paths based on the target Zen architecture, supporting SSE2, AVX2, and AVX512 instruction sets. Level 1, Level 2, and Level 3 BLAS APIs are implemented with specialized algorithms designed for different problem sizes—small, medium, and large—to exploit the architectural characteristics of AMD processors.
Developers targeting AMD EPYC or Ryzen systems should consider this library when performance-critical linear algebra operations are central to their workload. The tool is particularly suited for high-performance computing applications where leveraging CPU-specific features matters. Since the library maintains compatibility with the upstream BLIS API and retains all its features, existing code written against BLIS can benefit from AMD optimizations with minimal changes. The project provides configuration and build instructions through the AOCL User Guide on AMD's developer portal, and the upstream BLIS repository offers additional design documentation and usage examples.
Development is maintained directly by AMD with active issue tracking on the repository. The project receives regular updates that incorporate improvements from the upstream BLIS repository, ensuring it stays current with broader BLIS development while maintaining AMD-specific enhancements.
