From: 2024-01-09 08:30 to 09:15
Place: E:2116
Contact: birger [dot] swahn [at] cs [dot] lth [dot] se

One Computer Science MSc thesis to be presented on 9 January

Tuesday, 9 January there will be a master thesis presentation in Computer Science at Lund University, Faculty of Engineering.

The presentation will take place in E:2116.

Note to potential opponents: Register as an opponent to the presentation of your choice by sending an email to the examiner for that presentation (firstname.lastname@cs.lth.se. Do not forget to specify the presentation you register for! Note that the number of opponents may be limited (often to two), so you might be forced to choose another presentation if you register too late. Registrations are individual, just as the oppositions are! More instructions are found on this page.)

08:30-09:15 in E:2116

Presenter: Johanna Gustafson
Title: MLIR-based Code Generation for High-Performance Machine Learning on AArch64
Examiner: Flavius Gruian
Supervisors: Jonas Skeppstedt (LTH), Fredrik Knutsson (Arm Sweden AB), Per Åstrand (Arm Sweden AB)

Given the growing complexity of machine learning architectures driving the state-of-the-art, inference acceleration is an important aspect to consider in the development of machine learning systems. Machine learning frameworks often address this challenge using hand-crafted compute libraries optimized for a narrow range of hardware devices — often implemented by the hardware vendors themselves — meaning significant engineering effort is needed to extend to new platforms. For that reason, we propose a modular approach in the design of machine learning systems, utilizing the MLIR (Multi-Level Intermediate Representation) compiler framework of the LLVM project. This notion is strengthened through a demonstration of an MLIR-based implementation of an optimized double-precision general matrix multiply routine, which is shown to reach a perform of 86% of the theoretical machine peak on a single Neoverse-N1 core. Given the reusable and extensible nature of MLIR, we believe that there is value in developing MLIR-based compilers for machine learning software.

Link to popular science summary: https://fileadmin.cs.lth.se/cs/Education/Examensarbete/Popsci/240109_08Gustafson.pdf