GAPMA: Graph-Attention Cell Usage Prediction with Polarity-Minimized NPN Aggregation for Design-Specific Cell Composition (MLCAD 2026 Artifact)

Le résumé fourni par la source

Artifact for MLCAD 2026 paper 179, GAPMA: Graph-Attention Cell Usage Prediction with Polarity-Minimized NPN Aggregation for Design-Specific Cell Composition. The archive is the training-data-generation, GATv2-training, and iterative candidate-cell-selection pipeline described in the paper. It selects a small set of custom standard cells to add to a minimal baseline library for a given RTL design, before any cell layout exists, and measures the resulting post-route cell-area reduction. Contents. Ten Verilog RTL benchmarks (five training, five evaluation), a curated 41 MB subset of the SO3 academic 3 nm PDK, the mining and feature-extraction and GATv2 demand-prediction code with a pretrained checkpoint, and three top-level entry points: flow/datagen.sh (training-data generation), flow/train_gat.sh (GATv2 training), and flow/autorun.sh (the D0 baseline through D3 iterative candidate-cell-selection loop). Dependencies. The full pipeline invokes Synopsys Fusion Compiler for synthesis and place-and-route, and Cadence Pegasus, Quantus, and Liberate for the characterization and extraction steps. These are commercial, license-gated EDA tools not included here. The ILP-based cell-layout-generation step needs a Gurobi license, for which academic users can use the free academic tier. Reviewers without EDA-tool access can still inspect the full source, the curated PDK, and the pretrained model, and can run the Python-only mining and GATv2 inference code. Submodules. The three public submodules (Yosys, OpenDB, SO3-Cell) are present as empty directories. Their upstream URLs are in .gitmodules. No confidential credentials, license files, or full commercial PDK contents are included. Corresponds to GAPMA repository commit f5615008652163648ae1c24c8df2d012e851bce5.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.