SFDB Reproducibility Packages for Structured Multi-Constraint Consistency Correction for Sliding DFT: Numerical, Conservative-Design, Holdout, and FPGA Studies
Description
This dataset provides the reproducibility material accompanying the manuscript “Consistency-Projected Sliding DFT: Finite-Precision Analysis and Structured Multi-Constraint Correction.” It contains three frozen reproducibility packages. The SFDB Numerical Reproducibility Package v2.2.1 contains MATLAB code, locked campaign configuration, compact results, and validation material for the 70,500-trial fixed-point campaign comprising 141 arithmetic settings, five signal classes, ten methods, and ten common seeds. The principal recursive-method comparison reports a 98.5993% paired win rate and a geometric MSE ratio of 6.9791363 for SFDB-p4-neareven-FQ against Classical, mSDFT, oSDFT-Res, and oSDFT-Mod. The SFDB Conservative Design and Holdout Package v1.0.0 contains the operation-graph-based conservative rank/precision certificate, the frozen 648-candidate design campaign, and a separate predeclared holdout of 36 designs evaluated over 468 bounded-input trials. All 468 holdout trials respect their computed error bounds and are overflow-free. The largest observed per-frame RMS error is 5.4811848e-4 and the largest observed-to-bound ratio is 0.6303922. The certificate provides sufficient, not necessary, design conditions. The SFDB Practical FPGA Reproducibility Package v1.0.0 contains MATLAB, SystemVerilog, ModelSim, Quartus, reporting material, raw synthesis/STA evidence, and matched-format numerical results for the N=32 implementation study targeting the Intel Cyclone 10 GX 10CX105YF672E5G. It includes the seven-method common-60-MHz comparison, three fitter seeds, multicorner timing evidence, and 350 matched-format numerical trials. No physical FPGA board was used, and external I/O paths were not constrained with board-level input/output delays. Each archive includes package documentation, citation metadata, version information, SHA-256 manifests, source material, and approved result files. The packages are intended to reproduce and audit the numerical, conservative-design, holdout, and synthesis-level results reported in the associated paper.
Files
Institutions
- Budapest University of Technology and EconomicsBudapest, Budapest