With ever more demand for compute performance and large dataset processing for machine learning, high performance compute and other domains, efficiently integrating sophisticated, custom accelerators into established large application frameworks that initially were focusing on host CPU processing becomes ever more challenging.

The presentation outlines existing challenges and outlines ongoing work in the AMD Research and Advanced Development work on explorations to simplify integration of scalable custom domain accelerator functionality into well-established, low overhead application programming models.

Mr. Paul Blinzer is a Fellow at AMD in the Platform Software Group, has contributed on a wide range of technologies over his 24 years at AMD and holds over 40 issued patents. His most recent area of work leverages the capabilities of advanced data fabrics, accelerators and CPUs into application and operating system programming models. He is participating in workgroups of several standards organizations and was serving as the Chairperson of the System Architecture Workgroup of the HSA Foundation.

Mr. Blinzer lives in the Seattle/WA area, enjoys working on the forefront of technology and many outdoor activities, in accordance to the assumptions about the general lifestyle in the Pacific Northwest of the United States.

This short Compilers Design and Implementation course is intended to give the students an overview of the organization of traditional compilers for modern computer programming languages. This course briefly covers the classical phases of parsing, internal representation, code generation as well as optimization, for speed and energy-efficiency, and time permitting also describing the basics of advanced topics such as data-flow analysis and control-flow analysis, and optimization. Students are expected to know basic concepts of imperative programming languages such as syntax, scoping and semantics as well as understand basic computer architecture and organization. This short course is structured as five 2-hour lectures with daily take-home review exercises with solutions to illustrate the application of various analysis and compiler algorithms and implementations.

Starting with mathematical models for human cognition, we consider mathematical structures connected to logic, with references to Aristotle, Boole, Birkhoff, and von Neumann. Based on this, we characterize probabilistic extensions of classical logics and contrast them with fuzzy logic. The logic of quantum mechanics provides a structure where all probabilistic logics are substructures or special cases, and which, moreover, contains extensions which open the way to further applications. We indicate some applications of quantum-inspired logic in the context of databases and psychology. Finally, we present quantum-inspired aspects of signal representation and vector symbolic architectures.

In 2017, Bulatov and Zhuk independently proved the Feder-Vardi dichtomy conjecture: every constraint satisfaction problem (CSP) with a finite domain is in P or NP-complete. What remains open is a classifictation of
which finite-domain CSPs are in the complexity classes L, NL, NC, and which are P-complete. These questions are open even if the template structure is an orientation of a tree.

In this talk I will present results about CSPs of orientations of trees that were obtained with the help of computer experiments. We computed the smallest tree that is NL-hard (assuming L is not NL), the smallest tree that cannot be solved by arc consistency, and the smallest tree that cannot be solved by Datalog. Our experimental results also support a conjecture of Bulin concerning a question of Hell, Nesetril and Zhu, namely that "easy trees lack the ability to count". To obtain these results we make use of the most recent universal-algebraic theory about the complexity of finite-domain CSPs. However, further ideas are required because of the huge number of orientations of trees. In particular, we use the well-known fact that it suffices to study orientations of trees that are cores and show how to efficiently decide whether a given orientation of a tree is a core using the arc-consistency procedure. Moreover, we present a method to generate orientations of trees that are cores that works well in practice. In this way we found interesting examples for the open research problem to classify finite-domain CSPs in NL.

Joint work with Jakub Bulin, Florian Starke, and Michael Wernthaler.