Gareth Callanan's PhD defence

Thesis title: Dataflow Actor Networks: Representations, Compilation and MLIR Integration

Author: Gareth Callanan, Department of Computer Science, Lund University

Faculty opponent: Assistant Professor Lana Josipovic, ETH Zurich, Switzerland

Examination Committee:

• Professor Tobias Wrigstad, Uppsala University
• Associate Professor Marc Geilen, Eindhoven University of Technology, The Netherlands
• Professor Ahmed Hemani, KTH - Royal Institute of Technology
• Deputy: Professor Johan Eker, Lund University

Session chair: Professor Görel Hedin, Lund University

Supervisors:

• Main supervisor: Senior lecturer Flavius Gruian, Lund University
• Professor Liang Liu, Lund University

Location: E:1406, E-huset, Klas Anshelms väg 10/Ole Römers väg 3, Lund

Here is a link to download the thesis at LU Research Portal

Abstract

Software applications can be described using a computational model. These Models of Computation (MoCs) define rules governing application behaviour and properties that are enforced during execution. This thesis focuses on an actor-based MoC, known as dataflow-with-firing, where applications are modelled as actors connected by buffered channels. This model yields a concurrent application description that maps well to parallel hardware and allows buffer size and scheduling optimisations.

The actor model has been used successfully to describe multimedia applications, notably in the MPEG-RVC standard, which was defined using the CAL Actor Language (CAL). This success has been enabled by actor-based tools that synthesise these applications into hardware or map them onto targeted architectures.

Despite its use in multimedia systems, this model has seen limited adoption in other domains. This is partly due to the restricted expressiveness of CAL, which can only describe simple networks and is confined to primitive data types. Because of these limitations, existing CAL compilers have only been tested on multimedia applications, leaving them without the evaluation and optimisation needed to compete in other domains.

To broaden the applicability of the actor model, this thesis addresses these limitations by extending CAL, improving its execution model, and creating a new Intermediate Representation (IR) to decouple the model from the language. The first part of this thesis (Papers I to IV) focuses on enhancing CAL and its compilation flow. In Paper I, CAL is extended with arrays of ports and action generators to enable parametric network descriptions. These parametric models reveal scaling issues when converted to the Actor Machine (AM), a state-machine model for CPU execution. Paper II resolves this by updating the AM model. Paper III then demonstrates that the updated CAL can describe parametric QR decomposition, a key operation in modern 5G/6G communications. Finally, Paper IV contributes a complete compilation flow and an evaluation of AM performance on FPGAs.

An IR based on CAL and the MLIR framework is then introduced in Paper~V. Multiple frontends can target this IR, which supports more advanced types and operations than CAL itself. This IR is also amenable to actor-based scheduling algorithms and can be compiled to both CPU and GPU platforms.

These contributions enable the actor model and associated optimisations to be used with other frontends. Applications in diverse domains can now use this representation for accelerated execution on heterogeneous platforms.