SI2:SSE: MAtrix, TEnsor, and Deep-Learning Optimized Routines (MATEDOR)

University Of Tennessee Knoxville, Knoxville TN

Investigators

Abstract

A number of scientific software applications from important fields, including applications in deep learning, data mining, astrophysics, image and signal processing, hydrodynamics, and more, do many computations on small matrices (also known as "tensors") and using widely available standard linear-algebra software libraries. Scientists are trying to make these applications run faster by running them on advanced high performance computing (HPC) systems, that are heterogeneous systems that use processors of many different types, such as "accelerators" - that use of specialized computer hardware to perform some functions more efficiently than standard, general-purpose processors - and "co-processors" - that can run certain specialized functions in parallel with the central processor. However, standard linear algebra software libraries cannot make use of these specialized hardware components, and so the scientific applications mentioned above do not become much faster. Many existing linear algebra libraries, including libraries supplied by commercial vendors of computing technology have been tried to no avail. This issue is now critical because advancements in science from important fields are being held back due to the lack of progress in speeding up software. This project will address this through research and development that will create efficient software that can repetitively execute tensor operations grouped together in "batches" and which can be written to run very efficiently and quickly on the types of hardware components that exist in HPC systems. In addition to the research and development, several students will be engaged in the project, thus helping develop a critically needed component of the U.S. workforce. The trend in high performance computing (HPC) toward large-scale, heterogeneous systems with GPU accelerators and coprocessors has made the near total absence of linear algebra software for small matrix or tensor operations especially noticeable. Given the fundamental importance of numerical libraries to science and engineering applications of all types, the need for libraries that can perform batched operations on small matrices or tensors has become acute. This MAtrix, TEnsor, and Deep-learning Optimized Routines (MATEDOR) project seeks to provide a solution to this problem by developing a sustainable and portable library for such small computations. Future releases of MATEDOR are expected to have a significant impact on application areas that use small matrices and tensors and need to exploit the power of advanced computing architectures. Such application areas include deep-learning, data mining, metabolic networks, computational fluid dynamics, direct and multi-frontal solvers, image and signal processing, and many more. This team has a proven record of providing software infrastructure that is widely adopted and used, that supports ongoing community contributions, and that becomes incorporated in vendor libraries (e.g., Intel's MKL and NVIDIA's CUBLAS) and other software tools and frameworks (e.g., MATLAB and R). Students will be regularly integrated into the project activities, and this group of PIs has an exceptionally strong record of community outreach, having given numerous performance optimization and software tutorials at conferences and Users Group meetings. 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.