Year: 2024
Type: dissertation
Abstract: This habilitation manuscript presents a decade of research aimed at taming the design complexity of high-performance embedded systems, with contributions along three main research axes.The first research axis explored is the use of dataflow models of computation to program heterogeneous multicore chips. Contributions include extensions to dataflow semantics and model-based optimizations to automate the deployment of large applications on high-performance computing systems.The second research axis explores the use of approximations as a tool to reduce the computational complexity of complex systems. Proposed contributions include generic methods for rapidly characterizing hardware and software errors caused by approximations, and techniques for exploiting the resilience of computer vision algorithms to artifacts introduced by lossy image compression.The third research axis explored concerns the development of frugal artificial agents based on Tangled Program Graphs (TPGs). Contributions in this area aim at accelerating the training of TPGs through parallelization, and at accelerating inference for low-power devices through the generation of dedicated inference code.Research perspectives that follow the ongoing shift towards multi-objective and multi-constraint design space exploration are presented in the final chapter of the manuscript.
Source: HAL
Author: Desnos, Karol
Language: English
Subfield: Hardware and Architecture
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W4403926508