BLIS
A portable software framework for instantiating high-performance BLAS-like dense linear algebra libraries.
What is BLIS?
BLIS is a portable software framework for instantiating high-performance BLAS-like dense linear algebra libraries. It solves the problem of creating optimized BLAS implementations by isolating essential computational kernels that, when optimized, automatically enable high performance across most level-2 and level-3 operations. The framework includes BLAS compatibility and unique object-based APIs.
Developers and researchers building or optimizing dense linear algebra libraries, particularly those targeting new hardware architectures or requiring mixed-datatype operations.
BLIS offers a productivity advantage over traditional BLAS development by compartmentalizing optimization efforts into a few kernels, provides richer functionality like generalized storage and complex domain support, and enables extensibility through addons and plugins.
Overview
BLAS-like Library Instantiation Software Framework
Use Cases
Best For
- Instantiating high-performance BLAS libraries for new CPU architectures
- Implementing mixed-domain and mixed-precision matrix operations
- Developing linear algebra libraries with advanced multithreading support
- Creating custom BLAS-like operations via extensible addons and plugins
- Optimizing complex matrix multiplication using real domain kernels
- Building portable numerical software with generalized matrix storage
Not Ideal For
- End-users seeking a pre-compiled, plug-and-play BLAS replacement without any configuration or tuning
- Applications heavily dependent on level-1 BLAS operations, as BLIS's optimization focus is on level-2 and level-3 kernels
- Teams without expertise in linear algebra or kernel optimization, due to the sophisticated build and configuration system
- Projects targeting niche or unsupported hardware architectures where optimized kernels are not available, leading to generic performance
Pros & Cons
Pros
By isolating key computational kernels, optimizing them automatically boosts performance across most level-2 and level-3 operations, making it a productivity multiplier for instantiating libraries on new architectures.
Supports column-major, row-major, and general stride storage, allowing operands of different formats within the same operation—a feature omitted by traditional BLAS.
Includes complex symmetric and Hermitian forms omitted by BLAS, and implements the 1m method for complex multiplication using real kernels, reducing the need for complex microkernels.
Features addons and plugins for extending operation support and defining custom APIs without modifying core code, enabling rapid experimentation and customization.
Cons
The build system requires manual configuration for optimal performance, with detailed guides needed for architecture-specific setups, which can be daunting for newcomers.
For architectures beyond intel64, amd64, and arm64, BLIS may rely on generic kernels leading to lower performance, as noted in the External Packages section.
The framework's layered APIs and extensive documentation, while thorough, assume significant background in high-performance computing, making entry challenging for casual users.
Frequently Asked Questions
Related Projects
blazeHigh-performance C++ math library for dense and sparse arithmetic. (C++, BSD, Bitbucket)
OpenBLASOptimized BLAS library based on GotoBLAS2. (C and Assembly, BSD, GitHub)
BLASStandard building blocks for performing basic vector and matrix operations. (Fortran, public domain, GitHub)
LAPACKRoutines for solving systems of linear equations, linear least-squares, eigenvalue problems, etc. (Fortran, BSD, GitHub)
Found a gem we're missing?
Open-Awesome is built by the community, for the community. Submit a project, suggest an awesome list, or help improve the catalog on GitHub.