Scientific discovery now often involves computer simulation in addition to, or instead of, laboratory experimentation. This can accelerate, improve, and/or expand scientific insight, often at a great reduction in cost. Many such computer simulations spend much or most of their time solving linear algebra (matrix) problems. For these simulations, linear algebra problems constitute the most basic building blocks of the computation. As a result, software libraries (bundles of specialized code) that efficiently solve linear algebra problems fundamentally support sustained innovation in science. The project aims to create a next generation of software libraries for this domain and will make these libraries available to the scientific community as open source software that can be easily ported to current and future computer architectures. This will directly and indirectly impact discovery in academia, at the national labs, and in industry. The project will also impact affordable education through open course ware that is expected to reach a broad audience. The involvement of undergraduate and graduate students will strengthen the pool of qualified individuals trained to support scientific computing. The project involves research staff and students who are members of traditionally underrepresented groups.
The BLAS (Basic Linear Algebra Subprograms) are well-known routines that provide standard building blocks for performing basic vector and matrix operations. The Level 1 BLAS perform scalar, vector and vector-vector operations, the Level 2 BLAS perform matrix-vector operations, and the Level 3 BLAS perform matrix-matrix operations. Because the BLAS are efficient, portable, and widely available, they are commonly used in the development of high quality linear algebra software, such as the well-known Linear Algebra PACKage (LAPACK), as an example. However, the BLAS libraries that exist today have not evolved to new computing architectures, and hence do not perform as well as they could. The technical goal and scope of this project, therefore, is to develop a new high-performance dense linear algebra library with broad functionality that can be easily ported to current and future multi-core and many-core processors. The project builds on the BLAS-like Library Instantiation Software (BLIS) effort that has exposed low-level primitives that facilitate the high-performance implementation of BLAS. By implementing the higher-level dense linear algebra functionality in terms of these low-level primitives, portable high performance will be achieved for higher-level functionality needed by many scientific computing applications. Contributions will include the to-be developed techniques for implementing such software, the resulting open source software, and pedagogical artifacts that will include open course ware.
