EVAS — Event-driven Verilog-A Simulator

A lightweight behavioral simulator for digital/mixed-signal Verilog-A models. EVAS runs supported event-driven designs through the fail-closed EVAS2/Rust backend. No ngspice, no KCL/KVL solver.

Docs: evas.tokenzhang.com

What EVAS does

EVAS simulates voltage-mode, event-driven Verilog-A behavioral models. You provide:

1. A .va file — your behavioral model (comparator, DAC, SAR logic, DWA controller, …)
2. A .scs testbench netlist — voltage sources, ahdl_include, and a tran statement
3. Run evas simulate — get tran.csv waveforms and optional plots

The main simulator is transient-first. Compiled models also expose lightweight behavioral helpers for analysis("ac"), ac_stim(), and Verilog-A noise source functions so source-style models can run AC/noise sweeps from Python without claiming SPICE-style linearized circuit analysis.

The bundled examples are a compact smoke-test set. For your own design, copy the closest example directory, swap in your .va, adjust the stimulus sources and save list, and run. The larger verification example library belongs with the agent workflow in veriloga-skills/evas-sim.

Installation

pip install evas-sim
evas list        # verify install — prints bundled example groups

If evas is not on PATH, use python -m evas.

The production engine is EVAS2/Rust. Compatible Linux wheels include the evas-rust shared library, and source installs build it with cargo. A missing, unloadable, or ABI-incompatible Rust core is a hard error; EVAS never falls back to the Python simulation engine. The legacy evas2 and rust2 selectors remain accepted as time-bounded input aliases and are recorded as canonical evas-rust in run metadata.

Verify an installation and capture its machine-readable build identity:

evas --version
evas --version --format json > evas-identity.json

The JSON includes the package/CLI version, Rust core version and ABI, build revision when available, and whether the native core is present and loadable. Benchmark images should capture this output alongside the image digest. Every simulation also writes the same provenance to <output>/evas_identity.json.

Simulating your own design

evas simulate path/to/tb.scs -o output/mydesign
evas simulate path/to/tb.scs -o output/mydesign --ahdllint
evas simulate path/to/tb.scs -o output/mydesign --spectre-strict

Output in -o dir: tran.csv (waveforms), strobe.txt (log messages), evas_identity.json (build provenance), and .png plots. --ahdllint runs EVAS lint as a non-blocking simulation preflight and writes diagnostics into the simulation log before model compilation. Netlists may also request this with simulatorOptions options ahdllint=true. --spectre-strict runs the same lint preflight in blocking strict standalone Spectre mode, rejecting EVAS extension syntax before compilation. Netlists may also request this with simulatorOptions options spectre_strict=true; CI and benchmark wrappers can enable the same behavior globally with EVAS_SPECTRE_STRICT=1.

Before a full simulation, you can run a Spectre/AHDL-style static lint pass:

evas lint path/to/model.va
evas lint path/to/tb.scs --format json
evas lint path/to/model.va --spectre-strict

evas lint follows ahdl_include statements in .scs files and reports two classes of issues: compat-error for EVAS/Spectre subset problems that should block a candidate, and warning diagnostics for AHDL-style modeling risks such as discrete signals directly driving analog contributions or suspicious transition usage. Lint warnings do not change simulation pass/fail status. Current warning coverage includes a Cadence AHDL-inspired subset for transition timing and simple continuous-input dataflow, conditional potential contributions, case defaults, exact branch equality tests, floor/ceil contribution discontinuities, gnd node portability, discrete function arguments, implicit integer casts, and simulator-stop tasks inside loops. Each lint diagnostic is backed by a small rule registry that records its code, severity, canonical rule name, phase, source category, and related Cadence/Spectre identifiers when known. Diagnostics produced from parsed Verilog-A nodes include source line and column coordinates when available, so repair tools can point users back to the triggering statement or expression. Compatibility diagnostics include Cadence/Spectre-aligned cases such as conditionally executed analog operators (transition, slew, idt) and discipline vector ranges that depend on runtime variables instead of numeric or parameter constant expressions. Strict Spectre mode additionally rejects EVAS extension syntax that standalone Spectre rejects in the current compatibility target, including AMS bridge constructs (logic, wreal, continuous assign, always), task/do-while extensions, runtime electrical-node indexing, selected version-gated random distributions, seeded $rdist_* distributions whose Spectre PRNG sequence parity is not certified, integer bit/part-select concatenation gaps, generate, specify, connectmodule, and connectrules. The lint regression suite also keeps a small set of public oracle fixtures under tests/fixtures/lint_oracle_cases. These cases record distilled expected EVAS diagnostic codes only; raw Cadence/Spectre logs and generated certification reports are not committed.

Minimal testbench template (.scs):

simulator lang=spectre
global 0

ahdl_include "my_module.va"

Vvdd (vdd 0) vsource type=dc dc=1.8
Vclk (clk 0) vsource type=pulse val0=0 val1=1.8 period=10n rise=0.1n fall=0.1n width=4.9n

IDUT (clk vdd out) my_module vdd=1.8

tran tran stop=200n maxstep=0.1n
save clk:2e out:6f

Testbench structure reference

The bundled examples contain five small groups and their testbenches. They are the in-package reference for common wiring patterns:

Supported Verilog-A

Support tiers

EVAS reports unsupported-feature diagnostics with a stable [support-tier: <name>] suffix in text output and a support_tier field in JSON lint output. Benchmark reports should use these tiers instead of treating all Verilog-A/AMS failures as one flat simulator target.

Certification boundary: current EVAS benchmark PASS claims apply to behavioral-event designs, plus explicitly documented behavioral helper subsets. Rows that require full ams-digital or conservative-current-kcl semantics must be reported separately from supported EVAS bugs.

Continuous-time policy: ddt(), idt(), idtmod(), laplace_nd(), laplace_np(), laplace_zd(), laplace_zp(), zi_nd(), zi_np(), zi_zd(), zi_zp(), and limexp() are supported as voltage-domain behavioral transient approximations in ordinary analog statements. EVAS rejects Spectre-illegal conditional/event-body placements through lint or compile-time diagnostics. Unsupported continuous-time operators such as absdelay() are reported as behavioral-continuous-time, not as generic parser gaps.

Noise and stochastic policy: in ordinary transient analysis, white_noise(), flicker_noise(), and noise_table() contribute zero instantaneous random voltage and expose their PSD through evaluate_noise(), noise_spectrum(), and integrated_noise(). Explicit $random, $dist_*(), and $rdist_*() calls produce deterministic fixed-seed behavioral draws; the same seed and draw order reproduce the same sequence. Supported distribution helpers include uniform, normal, exponential, poisson, chi-square, t, and erlang forms. Unsupported distribution names produce an EVAS-COMP-EUNSUPPORTED diagnostic in the behavioral-event tier.

Subprogram policy: EVAS supports Spectre-style old-form functions/tasks (input x; real x;), ANSI-style task/function arguments, local variables, multi-argument calls, bounded recursion, and task/function calls from event bodies with normal state-update ordering.

Vector and indexing policy: pure expression bit-select, part-select, concatenation, replication, reduction, packed logic vectors, and integer/real state arrays are separate from electrical topology. Static generate/genvar elaboration is supported for a limited behavioral/continuous-assign subset. Runtime voltage-domain electrical indexing such as V(bus[i]) and V(bus[i]) <+ ... is supported by dynamic node resolution; dynamic current indexing remains part of the unsupported conservative-current-kcl tier.

Accuracy Profiles

You can set simulatorOptions options evas_profile=<mode> in .scs:

• fast: lower refinement (refine_factor=8, refine_steps=4) for faster runtime
• balanced: default EVAS behavior (16, 8)
• precision: higher refinement (32, 16) for tighter event/cross timing

errpreset=conservative/liberal is still respected; evas_profile applies an explicit EVAS-side override when set. Practical correspondence: fast ≈ liberal, balanced ≈ moderate, precision ≈ conservative (guidance only, not solver-equivalence claim).

CSV output format

The save statement accepts per-signal format hints:

save vin:10e vout:6f clk:2e dout:d

Bundled examples (reference only)

Five groups ship with the PyPI package for install verification, CLI sanity checks, and small starting templates. The full workflow-oriented example set is maintained outside this simulator package in veriloga-skills/evas-sim.

Contributing

git clone https://github.com/Arcadia-1/EVAS.git
cd EVAS
pip install -e ".[dev]"
pytest tests/ -v

License

MIT — see LICENSE.