In image processing pipelines, the complex interaction between operations and their hyperparameters creates an exponential design space where manual optimization is often infeasible. This challenge is magnified in Field-Programmable Gate Array (FPGA) implementations, where achieving an optimal trade-off between conflicting metrics — specifically output quality and hardware resource utilization (LUTs, Flip-Flops, and BRAM) — is critical for system integration. Standard heuristics often fail to navigate these non-linear dependencies effectively due to the structural complexity of the computational graphs. This paper introduces GENIO, a framework designed for High-Level Synthesis (HLS)-based pipelines. GENIO employs a structured representation of pipeline configurations that captures both topology and parameterization, enabling flexible exploration of compositional design spaces under hardware constraints. Within this framework, a Structural Genetic Programming (SGP) strategy is proposed to efficiently navigate the search space. A comprehensive comparative evaluation against NSGA-II, MOEA/D, Particle Swarm Optimization, and reinforcement learning–based strategies demonstrates that the proposed structural evolutionary approach attains competitive Pareto-front quality while exhibiting markedly faster convergence under identical synthesis budgets. In particular, near-optimal configurations are identified approximately 3.2 times earlier than with dominance-based multi-objective algorithms, highlighting its practical advantages in synthesis-driven optimization environments.

GENIO: A Genetic Programming framework for structural design space exploration of HLS image pipelines

Antonio Caruso;
2026-01-01

Abstract

In image processing pipelines, the complex interaction between operations and their hyperparameters creates an exponential design space where manual optimization is often infeasible. This challenge is magnified in Field-Programmable Gate Array (FPGA) implementations, where achieving an optimal trade-off between conflicting metrics — specifically output quality and hardware resource utilization (LUTs, Flip-Flops, and BRAM) — is critical for system integration. Standard heuristics often fail to navigate these non-linear dependencies effectively due to the structural complexity of the computational graphs. This paper introduces GENIO, a framework designed for High-Level Synthesis (HLS)-based pipelines. GENIO employs a structured representation of pipeline configurations that captures both topology and parameterization, enabling flexible exploration of compositional design spaces under hardware constraints. Within this framework, a Structural Genetic Programming (SGP) strategy is proposed to efficiently navigate the search space. A comprehensive comparative evaluation against NSGA-II, MOEA/D, Particle Swarm Optimization, and reinforcement learning–based strategies demonstrates that the proposed structural evolutionary approach attains competitive Pareto-front quality while exhibiting markedly faster convergence under identical synthesis budgets. In particular, near-optimal configurations are identified approximately 3.2 times earlier than with dominance-based multi-objective algorithms, highlighting its practical advantages in synthesis-driven optimization environments.
File in questo prodotto:
File Dimensione Formato  
1-s2.0-S0141933126001080-main.pdf

accesso aperto

Tipologia: Versione editoriale
Licenza: Creative commons
Dimensione 5.33 MB
Formato Adobe PDF
5.33 MB Adobe PDF Visualizza/Apri

I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.

Utilizza questo identificativo per citare o creare un link a questo documento: https://hdl.handle.net/11587/586208
Citazioni
  • ???jsp.display-item.citation.pmc??? ND
  • Scopus 0
  • ???jsp.display-item.citation.isi??? 0
social impact