Dr. Peter Thoman

Projects

• Celerity - High Level C++ for Accelerator Clusters

  [code]

  

  Celerity is a high-level C++ API, closely based on the Khronos SYCL standard, which allows targeting GPU compute clusters rather than just individual GPUs. The only additional requirement compared to a normal SYCL program is providing a functor which maps a kernel’s compute range to its data requirements. In the background, Celerity maintains data residence information, and constructs and optimizes task and command graphs for distributing the work in the cluster.
• LIGATE
  

  LIGATE is EuroHPC joint undertaking with the central goal of creating and validating a leading application solution for drug discovery in High Performance Computing systems up to the exascale level.
• Insieme Compiler and Runtime System

  [code]

  

  In the Insieme project, we explored the possibilities afforded by using a high-level intermediate compiler representation inherently designed for parallelism, INSPIRE. Insieme supports C/C++ source code and several parallelism APIs including OpenMP and Cilk as input, and generates code targeting its high-performance runtime system.

Teaching

• Advanced C++

  [LFU] [git]
  A graduate-level course on modern, advanced C++. Tries to tie together language theory and background with practical advice for working in large-scale C++ software projects.
  Prerequisites: Good understanding of C, familiarity with object-oriented programming.
• Performance-oriented Computing

  [LFU] [git]
  A combined lecture and lab with the goal of covering the most important topics that are relevant for practical performance engineering. This includes profiling and correct experimental setups, memory hierarchy optimiaztions, the performance implications of common data structures, and more. Prerequisites: Good understanding of C, and the basics of computer architecture
• GPU Computing

  [LFU]
  An introductory course to GPU computing. Introduces the differences between GPU and CPU hardware, GPU paradigms and APIs, and discusses how several algorithms can be efficiently parallelized on GPUs. Prerequisites: Basic C knowledge.

For more in-depth information about my current courses, please visit the course repository of the University of Innsbruck.
If you have any teaching-related questions, I am available consultation either online or offline. Please contact me by email to schedule a meeting.

Selected Recent Publications

• Command Horizons: Coalescing Data Dependencies While Maintaining Asynchronicity,
  WAMTA 2023: Asynchronous Many-Task Systems and Applications.

  [code]
  Peter Thoman, Philip Salzmann
• Multi‐GPU room response simulation with hardware raytracing,
  Concurrency and Computation: Practice and Experience 34 (4), 2022.
  Peter Thoman, Markus Wippler, Robert Hranitzky, Philipp Gschwandtner, Thomas Fahringer
• ndzip-gpu: Efficient Lossless Compression of Scientific Floating-Point Data on GPUs,
  Supercomputing, 2021.

  [code]
  Peter Thoman, Markus Wippler, Robert Hranitzky, Thomas Fahringer
• The Allscale Framework Architecture,
  Parallel Computing, Elsevier, May 2020.
  Herbert Jordan, Philipp Gschwandtner, Peter Thoman, Peter Zangerl, Alexander Hirsch, Thomas Fahringer, Thomas Heller, Dietmar Fey
• A Taxonomy of Task-based Parallel Programming Technologies for High-performance Computing,
  The Journal of Supercomputing, Springer, April 2018.
  Peter Thoman, Kiril Dichev, Thomas Heller, Roman Iakymchuk, Xavier Aguilar, et al.

For a complete list of publications, consult my Google Scholar profile.