A Dataset of SA Neighbor Allocation Feasibility Records for Voltage- and Reliability-Aware High-Level Synthesis

This dataset contains feasibility records collected during Simulated Annealing (SA) exploration of the voltage- and reliability-aware High-Level Synthesis (HLS) design space. Each row corresponds to a candidate neighbor allocation evaluated during the SA search and includes 46 input features spanning contextual information (benchmark identity, constraint parameters), SA search state (iteration count, temperature ratio), schedule structure (latency, area, slack, pressure, mobility statistics), move-specific information (move type, resource and voltage changes, latency/area delta contributions), and derived signals (scheduler-free naive latency/area predictions and violation flags, exact post-move energy and reliability values, GroupSwap-specific statistics). Each record also reports the actual post-move latency and area produced by the list scheduler (latency_new, area_new), a binary feasibility label, and, for infeasible records, the constraint that was violated (latency, area, or both).

The dataset was generated by running the SA engine in data-collection mode across eight HLS benchmark DAGs (DES, EWF, FIR, AR, Lattice, FEWF, Intaux, and Matmul) under a range of latency and area constraint combinations and under three objective weighting modes corresponding to energy minimization, reliability maximization, and balanced bi-objective optimization. A stratified sampling policy was applied: infeasible transitions, cost-worsening acceptances, and global-best improvements are always retained, while routine feasible records are subsampled to control dataset size and class imbalance.

This dataset supports the training and evaluation of FICS (Feasibility Inference Classifier for Scheduling), a LightGBM-based classifier that predicts scheduling feasibility before invoking a constrained list scheduler, accelerating SA-based design space exploration in HLS. It may also be of independent interest for research on learned surrogate models for combinatorial optimization, feasibility prediction in constrained search, or machine-learning-assisted electronic design automation (EDA).