Traditional scientific and machine learning high-performance computing software is often cast in terms of a set of fundamental operations, including the linear algebra functionality that underlies many applications. For this reason, research into and development of open-source linear algebra software libraries has been a science infrastructure priority for decades. An emerging trend that has disrupted this field is the recognition that scientific discovery can be made faster and/or more cost efficient by lowering the precision of computations, utilizing non-standard data types, and developing custom computational kernels. The project will leverage insights into how to structure the required software so that the combinatorial explosion in software complexity remains manageable. The outcome of the project will be a modern linear algebra software framework and application-focused libraries that will support future generations of computational applications in academia, at the national labs, and in industry. In addition, the project will impact the training of the next generation of high-performance computing professions and help remove barriers into the field for members of traditionally underrepresented groups.
The proposed work will build on previous NSF-sponsored research in order to address the implementation of expanded precision (EP), mixed precision (MP), and mixed domain (MD) algorithms simultaneously in a single software solution. Insights gained from a recent demonstration of MP/MD matrix multiplication will be extended by adding low precision types like float16 and bfloat16 and extended precision types like double-double. The target Basic Linear Algebra Subprograms (BLAS) functionality will be expanded to all level-1, level-2, and level-3 operations which in turn will support new research on how best to exploit MP/MD for LAPACK functionality. The new BLAS-like Library Instantiation Software (BLIS) framework will also be updated to provide the flexibility required to integrate extended dense linear algebra (DLA) operations. This flexible DLA framework will then be used to implement key functionality in computational and data science: tensor contraction and factorization operations important to quantum chemistry (QC) and high-performance primitives for machine learning. As a demonstration, these capabilities will be used to build state-of-the-art QC codes to perform coupled cluster polarization propagator and tensor-factorized coupled cluster calculations with full EP/MP/MD functionality, and the machine learning kernels will be integrated into computer vision and image recognition workflows.
This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.
