Renesis

v92.6.

Energy-aware synthesis for adiabatic and reversible logic. Netlist in, verified circuit out, with the energy accounted for rather than asserted.

Research prototype. Every result in this repository is reproducible from the sources here, and every optimization is equivalence-checked against the original netlist before its effect on energy is measured.

export PYTHONHASHSEED=0
./renesis bench/c432.v            # Python
./csrc/renesis bench/c432.v       # C -- same configuration, same result

Both tools read the same config/ declarations and produce identical results. The C implementation is the one intended for production use; it computes its own switching statistics and depends on nothing outside this repository.

./renesis --list-tech                 # available target technologies
./renesis --show-options              # every option and its current value
./csrc/renesis --option prefix=1 bench/c432.v   # parallel-prefix re-synthesis
./renesis --tech pfal bench/c432.v    # a different target technology

PYTHONHASHSEED=0 is required and asserted. The flow is deterministic: the same input and options give the same result, on any machine.

The re-synthesis passes are the one place where the two implementations part company on TIME rather than on answers. --prefix on c432 is a minute in the C tool and the better part of an hour in Python: the passes search, and search is where a C engine earns its keep. Both return the same circuit. Reach for ./renesis to read and modify the flow, and for ./csrc/renesis to run it.

./renesis ui

Picks a free port, starts a local server and opens the browser on the energy-aware interface. On macOS, double-clicking renesis-ui.command in Finder does the same. Stop it with the page's Quit button, or Ctrl-C.

The option panel is GENERATED from config/renesis_options.json -- the same declaration the CLI and the C tool read -- so the interface cannot disagree with the command line. Two modes: Synthesis (reversible circuit; .real / .tfc / .qc / OpenQASM / quantikz out) and Technology Map (dual-rail energy-recovery network; .tgn, plus the technology-independent netlist as BLIF / Verilog / BENCH / PLA). Both offer schematic, SPICE and transistor-level exports, and a PDF of the circuit view.

To start it by hand, or on a fixed port:

export PYTHONHASHSEED=0
python3 scripts_adiabatic/adiabatic_server.py 8766   # then open http://localhost:8766
python3 scripts/revsynth_server.py                   # the general reversible UI, port 8765

Both bind 127.0.0.1 only. Do not open the .html files directly -- they are not programs; the page must be served so its requests reach the process that does the synthesis. See WEB-UI-HOWTO.md.

Cite the version you used. https://doi.org/10.5281/zenodo.21979880 is the all-versions DOI and always resolves to the newest release; it is the one in the badge above. Each release also carries its own version DOI, minted by Zenodo when the release is published and listed on the release's Zenodo record. This release, v92.5, is https://doi.org/10.5281/zenodo.22069882; the v92.4 release is v92.6, is https://doi.org/10.5281/zenodo.22135505; the v92.4 release is https://doi.org/10.5281/zenodo.21987620; the v92.3 release is https://doi.org/10.5281/zenodo.21979881.

GitHub's Cite this repository button reads CITATION.cff, which carries the same information in machine-readable form.

The accompanying paper is:

M. A. Thornton, "Renesis: Energy-Aware Synthesis of Adiabatic Logic from Irreversible Netlists," arXiv:2608.17139 [cs.AR], 2026. https://arxiv.org/abs/2608.17139

Cite the paper for the methodology, and cite a Zenodo DOI above for the specific software version you ran.

Everything marked optional degrades the same way: the run completes, the artifact that needed the tool is skipped, and the run names the tool it wanted. Nothing silently substitutes an easier answer.

"ASP-DAC" is a comparison baseline, not a tool: there is nothing to install. aspdac_baseline.py builds it, and the optimised-NOR construction needs ABC.

Two external files, both meant to be read and edited:

• config/renesis_options.json -- every option, its default, its legal range, and what it does. The CLI, the C tool and the browser interface all read this one declaration.
• config/technology/<name>.json -- one file per target technology: cell model, structural constraints, clocking discipline, and the provenance of every constant. Copy one and edit it to describe your own target.

Defaults are not chosen: every default is the value assumed by the 20-circuit validation that certifies a release. Where the underlying library's own default differs, the table records both and says so.

parse -> [structural prep] -> [re-synthesis] -> tag sweep
      -> cover + technology mapping -> buffer insertion -> energy report

renesis is a thin orchestrator: it sequences stages that remain independently callable, which is what keeps re-orchestration cheap.

Six re-synthesis optimizations, all OFF by default. Three belong to the linear-mapping family: the boundary decoder (--bdec, a global affine re-encoding of the OUTPUT space -- the matrix B is m x m over the outputs, with a decoder network at the boundary computing the inverse), single-output interior windows (--linwin), and multi-output interior windows (--mowin, which re-encode a shared-leaf region with several roots). The others are parallel-prefix re-synthesis (--prefix), which rebuilds long carry chains as a Brent-Kung all-prefix network; affine-cut extraction (--davio), which re-emits a cut whose Boolean difference is constant in every variable as an XOR tree; and bounded elimination with algebraic extraction (--elim). --optimize-all turns on davio, prefix, linwin and mowin together.

Each is gated the same way: a candidate is equivalence-checked against the ORIGINAL netlist, then priced, and accepted only if it improves one energy table and worsens neither. A pass can therefore cost runtime and return the netlist unchanged, and that is reported rather than hidden. Before a pass searches, a cheap structural test asks whether it could construct any candidate at all here; when it could not, the pass says so instead of searching. That screen never changes an answer, only the time it takes -- --no-prescreen turns it off and the release validation checks that both ways give byte-identical results.

Mapping-level levers are not re-synthesis but change every number: the shared multi-output BDD forest (auto_e2, on), fanout-one absorption (absorb_fo1, on), BDD/mux arbitration (auto_bdd, off), the block realisation route (route=auto, the one place the energy model makes a structural rather than a reporting decision), and the series cap that governs buffer insertion.

A default run is fast and tells you when a flag might do better at a runtime cost.

renesis                 the orchestration entry point (./renesis --help, ./renesis ui)
renesis-ui.command      double-click launcher for the browser interface (macOS)
csrc/                   the C implementation: renesis, rsynth, the vsim layer,
                        run_tests.sh (the build-verification suite), samples/
scripts_adiabatic/      the Python implementation: energy-aware synthesis,
                        technology mapping, the browser UI server
scripts/                the reversible-synthesis Python side and its UI server
tools/                  vendored EXORCISM (BSD-2, Bruno Schmitt/EPFL), the
                        CUDD shim, and the CUDD header; CUDD itself is built
                        separately (see MACOS-SETUP.md)
config/                 renesis_options.json -- every option, its default and
                        help text -- and one JSON per target technology
bench/                  benchmark circuits (ISCAS-85/89, EPFL, and others)
examples/               small examples, including the two custom hash
                        circuits (EightBitHashTable.pla, TwelveBitHash.pla)
spice/                  generated per-family SPICE cell decks, with README
repo_validate.sh        optional: does this checkout build and behave
                        correctly on your machine? (see below)
renesis.1               the manual page (man ./renesis.1)
RENESIS-OPTIONS-LIST.md every option, both front ends, with defaults
RENESIS-MANUAL.md       the manual: every option of both front ends
WEB-UI-HOWTO.md         the browser interface, start to finish
MACOS-SETUP.md          building on macOS, including the CUDD build

You do not have to do this. Renesis works without it. Run it when you want assurance that the tool built correctly on your machine before you trust a number it gives you, or when something looks wrong and you want to say precisely what.

One command, from the repository root:

bash repo_validate.sh

It builds from scratch, counts compiler warnings and floating-point flag coverage, runs the C suite, checks the reference answer for c17 to the last digit, checks that the Python driver and the C engine agree, and runs one re-synthesis pass end to end. Expect ALL REPO CHECKS PASSED and two to five minutes on a current laptop.

The C suite alone, if that is all you want:

bash csrc/run_tests.sh

Expect ALL TESTS PASSED: 22 stages covering both parsers, the synthesis pipelines, and an 82-pair byte-identity spot check between the Python and C implementations. If a stage reports a skip, a prerequisite is missing -- the message names it.

Please open a GitHub issue. We would rather hear about it than not.

For a build failure, a failed check, or a wrong-looking number, include the whole console output plus uname -a, cc --version and python3 -V. If repo_validate.sh failed, its output is the report.

Platform reports are especially useful. Renesis is developed on Apple silicon and tested on aarch64 Linux, x86-64 Linux, and x86-64 Windows under WSL. Anything else is new ground, and a build log from new ground is a contribution.

Two things that are worth reporting even though they are not crashes: a result that differs between the Python driver and the C engine on the same input, and any energy figure that differs from the same run on another machine. Both implementations are held to byte-identical output by design, so either one is a real defect.

Start with RENESIS-MANUAL.md (all options, both front ends, with runnable examples), then WEB-UI-HOWTO.md for the browser interface and MACOS-SETUP.md to build on a Mac. PATENTS.md states the patent position; the code is MIT-licensed (LICENSE).

This repository is the released image of a larger research bundle. Every released version is validated end to end before it ships: byte-identical outputs between the two implementations across a 237-cell parity matrix, plus the suite above. The development record (per-version checkpoints, validation procedures, and research drivers) is not part of this repository; the papers describe the methodology.