From Tool Invocation to Source-Mechanism Exploration: Protected White-Box DSE for Open-Source EDA

Jul 16, 2026

Open-source EDA tools enable design-space exploration to move beyond public knobs and into bounded source-level mechanisms inside staged optimizers. ReviewDSE proposes a protected white-box DSE framework that evaluates complete source candidates under a protected evaluator, while storing reviewed mechanism-level evidence for reuse. It builds method evidence from calibration designs, then warm-starts target exploration using Teacher review and full-flow validation. On OpenROAD detailed placement, it reduces final post-DPL HPWL by 1.78% on average versus 0.38% for public-knob black-box DSE, with runtime-aware selection and failure-aware repairs.

LanguageEnglish

CategoryTechnology

FormatPDF

Pages7

Reading Time18 min

Abstract

Introduction

Design-space exploration and black-box boundaries

White-box DSE in open-source EDA

Challenges in source-level exploration

ReviewDSE framework overview

Two-level workflow and instantiation

From Tool Invocation to Source-Mechanism Exploration: Protected White-Box DSE for Open-Source EDA Zhiyu Zheng Fudan University Shanghai, China [zyzheng24@m.fudan.edu.cn](zyzheng24@m.fudan.edu.cn) Yiming Du Fudan University Shanghai, China [iiamadoga@163.com](iiamadoga@163.com) Ziyi Wang The Chinese University of Hong Kong Hong Kong, China [zeayw1@gmail.com](zeayw1@gmail.com) Zhiang Wang Fudan University Shanghai, China [zhiangwang@fudan.edu.cn](zhiangwang@fudan.edu.cn) arXiv :2607 . 11294v1 [ cs .AR] 13 Jul 2026 Abstract Open-source EDA tools allow design-space exploration (DSE) to move beyond public knobs and into bounded source-level mechanisms inside staged optimizers. We present ReviewDSE, a protected white-box DSE framework that explores such mechanisms for a target design. ReviewDSE evaluates complete source candidates under a protected evaluator and records reusable search knowledge as reviewed mechanism-level evidence. It first constructs method evidence and source-start branches from calibration designs, then uses these fixed warm-start products to initialize target-case exploration under Teacher review and full-flow validation. We instantiate ReviewDSE on OpenROAD detailed placement as a representative staged open-source EDA optimizer. Across nine target tasks, ReviewDSE reduces final post-DPL half-perimeter wirelength (HPWL) by 1.78% on average under a 2× runtime gate, compared with 0.38% for public-knob black-box DSE. A runtime-aware ReviewDSE selection retains a 1 .68% reduction at 1 . 11× runtime, and full-flow review exposes stage-composability failures while source-mechanism exploration repairs hard cut-row legality failures. 1 Introduction Design-space exploration (DSE) is a standard methodology for improving electronic-design-automation (EDA) flows [1] . Conventional DSE treats each tool as a fixed executable and searches its public control surface, such as command options, flow scripts, or other exposed knobs, to find configurations with better quality of results [2] . This black-box formulation is practical and broadly applicable, especially for closed-source tools, but it leaves quality-critical implementation choices inside large EDA codebases outside the search space. Open-source EDA changes where the DSE boundary can be drawn. In an open codebase such as OpenROAD [3], an optimizer is not only invocable but also inspectable and modifiable. This enables white-box DSE for open-source EDA: instead of searching only over public knobs ofa fixed executable, the search can enter selected parts of the optimizer implementation and explore bounded source-level mechanisms. In EDA flows, this mechanism space is naturally organized by stages and stage interfaces. A placement flow, for example, may include legalization policies, local improvement kernels, objective scoring, displacement control, rollback logic, and state passed from legalization to detailed-placement optimization. White-box DSE therefore asks which internal mechanisms should be changed, where they should be inserted, and how they should be configured for a target design. The expanded mechanism space is difficult to explore manually. Source-level choices are numerous, implementation-dependent, and strongly coupled across stages. A local improvement in one stage (a) Conventional Black-box DSE (b) White-box DSE Figure 1: Search boundary for one target: black-box DSE tunes public knobs; white-box DSE explores bounded internal mechanisms under the protected evaluator. may not survive the downstream flow; for example, a legalizer that reduces post-legalization HPWL can still produce a placement that is harder for detailed-placement optimization (DPO) to improve. Recent coding-evolution systems suggest a possible path forward: languagemodel agents can modify programs using evaluator feedback [4– 6], and related ideas have begun to appear in EDA source-level evolution [7–10] . These systems show that agents can generate use

FAQs

What problem does ReviewDSE address in open-source EDA design-space exploration?

ReviewDSE tackles the limitation of black-box DSE that only searches public knobs, leaving implementation choices inside EDA codebases outside the search space. It also addresses the difficulty of reliably exploring a large, tightly coupled source-level mechanism space across optimizer stages.

How does ReviewDSE keep white-box source exploration reliable?

ReviewDSE uses a protected evaluator to validate complete source candidates, and it records review unit evidence finer than whole-diff acceptance. The evidence supports liveness, compatibility, and failure diagnosis, reducing risks from corrupted evaluator paths or invalid bypasses.

What results does ReviewDSE achieve on OpenROAD detailed placement?

Across nine target tasks, ReviewDSE reduces final post-DPL half-perimeter wirelength (HPWL) by an average of 1.78% under a 2× runtime gate, compared with 0.38% for public-knob black-box DSE. Runtime-aware selection retains a 1.68% reduction at 1.11× runtime, and full-flow review identifies stage-composability failures and repairs hard legality issues.

Jul 16, 2026

Upload by Clementine