High-Performance Tensor Computing
Prof. Alexander BREUER
Illustration of the Einsum Tree IR optimization procedure. Starting at the root, the procedure optimizes all tensor contractions by using the three simple transformations: swap the left and right children of a binary node, reorder the index string of an interior node, and insert a permutation node.
Prof. Alexander BREUER
Image: PrivateProf. Alexander BREUER
Email: alex.breuer@uni-jena.de
Phone: +49 3641 946371
The High-Performance Tensor Computing group develops scalable algorithms and software for computationally demanding problems in science and industry, working across the whole stack from the bare metal of modern computer architectures to fully automated production workflows.
Screenshot of a self-contained application that illustrates the Tiled Execution Intermediate Representation (TEIR). The app allows users to optimize tensor operations through IR-to-IR transformations.
Image: PrivateAt the center of the group's work are tensor computations, the high-dimensional generalizations of matrix multiplication that power much of machine learning and the computational sciences. The group designs domain-specific compilers and code generators that turn high-level tensor expressions into fast machine code. It builds high-performance kernels for the vector and matrix units of modern CPUs and GPUs, for neural-processing units, and for other emerging accelerators. A second line of work brings the same performance-engineering methods to large-scale scientific simulation, in particular to high-order methods for wave propagation. Across these directions the group pursues one goal: portable abstractions that reach a high fraction of peak performance on an increasingly diverse landscape of hardware.