Recent advances in AI accelerators have enabled efficient inference of deep learning models at the edge. These models rely on structured loop-based operations that must be mapped onto constrained hardware in an energy- and latency-efficient way. While state-of-the-art tools provide reasonable mapping support for single layers, they fall short when it comes to modeling and optimizing more advanced execution strategies such as layer fusion and execution on heterogeneous multi-core systems. This thesis addresses these limitations by: 1) developing new optimization techniques for temporal loop mapping within single-core accelerators; 2) introducing analytical models for layer fusion, allowing fine-grained parts of different layers to be scheduled jointly and evaluated for energy and latency benefits; 3) generalizing these models to multi-core systems with heterogeneous cores, memory hierarchies, and communication topologies; and 4) applying these insights to emerging neural network workloads such as state space models and early-exiting neural networks, demonstrating the flexibility and future relevance of the proposed approach.
10/10/2025 17:00 - 19:00
ESAT Aula C, B91.300
