Skip to main content

dae_runtime/
lib.rs

1//! Assembles a circuit's linear MNA system (via `general-mna`, including its `'D'` stamp) with
2//! PWL device segments (via `pwl-devices`) into one Linear Complementarity Problem, solves it
3//! (via `lcp-solver`), and reports the resolved DC operating point. No Newton-Raphson, no
4//! voltage limiting, anywhere in this crate — see this repository's `docs/architecture.md`.
5//!
6//! ## The generic Thevenin/LCP fold
7//!
8//! `general-mna` stamps each diode `k` as a *fixed* conductance `{k}_G` plus a per-instance
9//! current-source symbol `{k}_Ioff`. Fixing `{k}_G` at the diode's canonical reference slope
10//! `g_off` (see `pwl_devices::IdealDiode::canonical`) makes the whole linear system `A0` genuinely
11//! fixed — independent of which segment every diode ends up in — with every segment's actual
12//! nonlinearity pushed entirely into the `{k}_Ioff` terms via each diode's canonical
13//! `max(0, ...)` (`z`) decomposition. Since the system is linear in those `Ioff` terms, each
14//! diode's terminal voltage is an affine function of every diode's `z` variables:
15//!
16//! ```text
17//! x(z)   = x0 + sum_j w_j * raw_Ioff_j                  (raw_Ioff_j = delta_on_j*z2_j - delta_br_j*z1_j)
18//! V_k(z) = v0_k + sum_j gamma_kj * raw_Ioff_j            (gamma_kj = w_j[p_k] - w_j[n_k])
19//! ```
20//!
21//! where `x0 = solve(A0, u0)` is the baseline operating point with every diode's `Ioff = 0`,
22//! and `w_j = solve(A0, B[:, diode j's input column])` is how much a unit "raw Ioff" on diode
23//! `j` moves every unknown. Substituting `V_k(z)` into each diode's guard equations
24//! (`w1_k = V_k - v_breakdown_k + z1_k`, `w2_k = v_th_k - V_k + z2_k`) gives exactly the LCP
25//! `(M, q)` this crate builds and hands to `lcp_solver::solve`.
26
27mod block_graph;
28pub mod checkpoint;
29mod closed_loop;
30mod linsolve;
31mod step_control;
32mod topology;
33
34pub use block_graph::{
35    reject_sig2phys_wired_into_circuit, simulate_transient_with_blocks,
36    simulate_transient_with_blocks_capped, simulate_transient_with_blocks_checkpointed,
37    simulate_transient_with_blocks_streamed, BlockInstance, BlockKind, CheckpointControl,
38    CheckpointSink, ConstValue, GainValue, GateBinding, Phys2SigTarget, PhysicalDomain, PidClamp,
39    SampleTimeSpec, Signal, SignalValue, TransientWithBlocksStep, ADAPTIVE_STEP_HARD_CAP,
40};
41pub use closed_loop::{sawtooth_carrier, simulate_closed_loop};
42pub use step_control::{AdaptiveConfig, TimeStep};
43
44use std::collections::BTreeMap;
45
46use general_mna::{
47    BuildError, BuildOptions, EvaluationError, Expression, InitialStateError, MnaBuilder, MnaSystem,
48};
49use general_spice_core::ast::Statement;
50use general_spice_core::Dialect;
51use lcp_solver::LcpError;
52use linsolve::{dense_solve, SingularMatrix};
53use pwl_devices::{IdealDiode, IdealSwitch};
54
55pub use general_mna::SwitchState as GateState;
56pub use general_mna::TransientFunction;
57
58#[derive(Debug, Clone, PartialEq)]
59pub struct OperatingPoint {
60    pub unknowns: Vec<String>,
61    pub x: Vec<f64>,
62    /// Resolved `(z1, z2)` for each diode, in the same iteration order as the `diodes` map
63    /// passed to [`solve_dc`] (a `BTreeMap`, so alphabetical by name) — useful for reporting
64    /// which segment each diode ended up in.
65    pub diode_names: Vec<String>,
66    pub diode_z: Vec<(f64, f64)>,
67    /// Each diode's true (unscaled) resolved current `delta_on*z2 - delta_br*z1`, same order
68    /// as `diode_names`/`diode_z`. Needed as history by [`simulate_transient`]'s trapezoidal
69    /// step (see its doc comment); exposed here rather than recomputed by every caller.
70    pub diode_raw_ioff: Vec<f64>,
71}
72
73impl OperatingPoint {
74    pub fn value(&self, unknown: &str) -> Option<f64> {
75        self.unknowns
76            .iter()
77            .position(|name| name == unknown)
78            .map(|index| self.x[index])
79    }
80}
81
82#[derive(Debug, Clone, PartialEq)]
83pub enum DaeError {
84    /// Parsing/hierarchy-flattening the netlist itself failed — see
85    /// [`general_mna::parse_and_flatten`]; a `.subckt`/`X` structural error (undefined
86    /// subcircuit, node-count mismatch, a recursive instantiation, ...) surfaces here, distinct
87    /// from [`DaeError::Build`]'s "parsed fine, but this statement has no valid electrical
88    /// stamp" class of error.
89    Parse(String),
90    /// A streaming `on_step` callback's own I/O failed (writing a CSV/raw row straight to a
91    /// file/stream as it's produced — see [`block_graph::simulate_transient_with_blocks_streamed`]).
92    /// Carries the formatted underlying `std::io::Error` as a string rather than the error type
93    /// itself, since `DaeError` derives `PartialEq` and `io::Error` doesn't implement it.
94    Io(String),
95    Build(BuildError),
96    /// The netlist declares `ic=` initial conditions that could not be assigned — see
97    /// [`general_mna::MnaSystem::initial_state`], whose own error explains the two ways that
98    /// fails: conditions that contradict each other (a loop of `ic`-bearing capacitors whose
99    /// declared voltages do not sum to zero), and an assignment that violates one of the
100    /// circuit's own algebraic equations in which every unknown is already fixed by an `ic=`
101    /// (an `ic` on a capacitor across an ideal voltage source; series inductors declaring
102    /// different currents).
103    InitialCondition(InitialStateError),
104    Evaluate(EvaluationError),
105    Linear(SingularMatrix),
106    Lcp(LcpError),
107    UnknownDiodeInput(String),
108    /// A [`block_graph::BlockInstance`]'s input, or a [`block_graph::GateBinding`], names a
109    /// block that doesn't exist anywhere in the graph (a typo, or a genuinely undeclared name)
110    /// — declaration *position* is no longer a possible cause: [`block_graph::topological_order`]
111    /// derives each step's evaluation order from the `Signal::Block` dependency graph itself, so
112    /// a block may reference another declared anywhere in the same slice, before or after it.
113    UnknownBlockInput(String),
114    /// The `Signal::Block` dependency graph among a [`block_graph::BlockInstance`] slice
115    /// contains a genuine same-step cycle — computed once, upfront, by
116    /// [`block_graph::topological_order`] (a DFS-based topological sort; a cycle shows up as a
117    /// back-edge into a node still on the current recursion path) before any step is solved.
118    /// `cycle` is the exact closing path, block names in dependency order with the first name
119    /// repeated at the end to show the loop closing (e.g. `["A", "B", "C", "A"]`); a
120    /// self-referencing block reports `["A", "A"]`. Fix by inserting a `Signal::BlockPrev`
121    /// somewhere in the cycle — the one-sample delay that turns a same-step algebraic loop into
122    /// a legitimate sampled-data feedback path.
123    AlgebraicLoop {
124        cycle: Vec<String>,
125    },
126    /// Loading or resolving symbols in a [`block_graph::BlockKind::CScript`]'s shared library
127    /// failed — see [`cscript_ffi::CScriptError`].
128    CScript(cscript_ffi::CScriptError),
129    /// A netlist declares a [`block_graph::BlockKind::CScript`] block whose library doesn't
130    /// export `cscript_clone`, but the run was requested with [`TimeStep::Adaptive`], which
131    /// needs to clone every block's state before each trial step — see `cscript_ffi`'s own
132    /// module doc comment, "Adaptive step-size control and instance cloning." Pass a fixed
133    /// `TimeStep::Fixed` instead, or add `cscript_clone` to the library.
134    CScriptRequiresCloneForAdaptiveStep {
135        block_name: String,
136    },
137    /// A netlist declares a [`block_graph::BlockKind::CScript`] block with `ts=variable`
138    /// (`block_graph::SampleTimeSpec::Variable`), but its library doesn't export
139    /// `cscript_next_sample_hit` — the required function for a block that computes its own
140    /// schedule; see `cscript_ffi`'s own module doc comment, "The optional block-controlled
141    /// sample time." Checked once, up front, the same "checked once, up front" discipline
142    /// `CScriptRequiresCloneForAdaptiveStep` already uses.
143    CScriptRequiresNextSampleHitForVariableSampleTime {
144        block_name: String,
145    },
146    /// The `PyBlock` counterpart to [`Self::CScriptRequiresNextSampleHitForVariableSampleTime`]
147    /// — a `.py` file declared `ts=variable` but doesn't define `next_sample_hit`.
148    #[cfg(feature = "python")]
149    PyBlockRequiresNextSampleHitForVariableSampleTime {
150        block_name: String,
151    },
152    /// Loading, calling into, or cloning a [`block_graph::BlockKind::PyBlock`]'s own `.py` file
153    /// failed — see [`pyblock_ffi::PyBlockError`]. Unlike [`Self::CScriptRequiresCloneForAdaptiveStep`],
154    /// there is no separate "doesn't support adaptive stepping" variant for `PyBlock`: cloning
155    /// uses Python's own generic `copy.deepcopy`, which needs no author opt-in, so a `PyBlock`
156    /// is always eligible — a genuine failure to deep-copy its own state instead surfaces here,
157    /// as an ordinary [`pyblock_ffi::PyBlockError::Exception`].
158    #[cfg(feature = "python")]
159    PyBlock(pyblock_ffi::PyBlockError),
160    /// A netlist declares a [`block_graph::BlockKind::PyBlock`]/[`block_graph::BlockKind::
161    /// PyFunction`] block, but this build of `dae-runtime` was compiled without the optional
162    /// `python` feature (which pulls in `pyblock-ffi`, and through it a build-time dependency
163    /// on a discoverable Python/libpython) — unconditional, unlike the two variants above,
164    /// specifically so a build without that feature can still report a clear, actionable error
165    /// instead of failing to compile at all or panicking.
166    PythonSupportNotCompiledIn {
167        block_name: String,
168    },
169    /// Spawning the shared `octave-cli` subprocess, adding a [`block_graph::BlockKind::OctFunc`]
170    /// block's own `.m` file directory to its path, or calling into it failed — see
171    /// [`octave_ffi::OctaveError`]'s own three cases (not found on `PATH`, the process died, or
172    /// an ordinary Octave-side runtime error the session survives). Unlike
173    /// [`Self::PythonSupportNotCompiledIn`], there is no "not compiled in" counterpart for this
174    /// variant — `octave-ffi` is an unconditional dependency of this crate (a pure subprocess
175    /// manager, never linking against Octave itself), so `kind=octfunc` is always available to
176    /// parse and construct; only an actually-missing `octave-cli` binary on `PATH` at runtime
177    /// surfaces here, as [`octave_ffi::OctaveError::NotFound`].
178    Octave(octave_ffi::OctaveError),
179    /// A [`block_graph::BlockKind::OctBlock`]'s own `path=` directory is missing one of the
180    /// `.m` files its declared contract requires (see that variant's own doc comment for the
181    /// full file-per-function table) — checked once, up front at construction, the same
182    /// "checked once, up front" discipline [`Self::CScriptRequiresCloneForAdaptiveStep`]
183    /// already uses, and deliberately *before* ever spawning `octave-cli` for this block, since
184    /// the filesystem already has the answer without needing a round trip through Octave's own
185    /// "undefined function" error.
186    OctBlockMissingRequiredFile {
187        block_name: String,
188        expected_path: std::path::PathBuf,
189    },
190    /// A netlist declares a [`block_graph::BlockKind::OctBlock`] block with `ts=variable`
191    /// (`block_graph::SampleTimeSpec::Variable`), but `<function>_next_sample_hit.m` doesn't
192    /// exist in its own `path=` directory — the `OctBlock` counterpart to
193    /// [`Self::CScriptRequiresNextSampleHitForVariableSampleTime`]/
194    /// [`Self::PyBlockRequiresNextSampleHitForVariableSampleTime`].
195    OctBlockRequiresNextSampleHitForVariableSampleTime {
196        block_name: String,
197    },
198    /// A netlist declares a [`block_graph::BlockKind::OctBlock`] block, but the run was
199    /// requested with [`TimeStep::Adaptive`], which clones the whole `block_states` vector
200    /// before every trial step and simply discards a rejected trial's clone (see
201    /// [`Self::CScriptRequiresCloneForAdaptiveStep`]'s own doc comment). That protocol depends
202    /// on a block's own state living entirely inside the cloned Rust value -- true for
203    /// `CScript`/`PyBlock` (an owned, independently-cloned native/Python state), but **not**
204    /// true for `OctBlock`: its own opaque per-instance state lives inside the one shared
205    /// `octave-cli` session itself (see [`block_graph::BlockState::OctBlock`]'s own doc
206    /// comment), which is never cloned, so a rejected trial's own state-mutating calls
207    /// (`output`/`output_xc`/`update`) would silently persist in the real Octave-side state
208    /// with no way to roll them back. Unlike `CScriptRequiresCloneForAdaptiveStep` (opt-in,
209    /// satisfiable by exporting `cscript_clone`), this restriction is unconditional -- there is
210    /// no `.m`-file convention that could make Octave's own shared global-struct state
211    /// trial-cloneable. Use a fixed `TimeStep::Fixed` for any netlist using `kind=octblock`.
212    OctBlockDoesNotSupportAdaptiveStep {
213        block_name: String,
214    },
215    /// [`TimeStep::Adaptive`] took more than [`ADAPTIVE_STEP_HARD_CAP`] accepted steps without
216    /// reaching `t_final` — a hard safety ceiling, not a normal error a well-behaved run should
217    /// ever hit. Exists because the adaptive loop accumulates one row per accepted step into an
218    /// in-memory trace with no other bound, and a genuinely stalled step controller (e.g. a
219    /// near-zero-`dt` loop that never makes real forward progress — the exact class of bug
220    /// `next_pwm_edge_dt`'s own `THETA_EPS` guards against, see its doc comment) would otherwise
221    /// grow that trace without limit until the process — and, on the real run that first hit
222    /// this, the whole machine — ran out of memory. `steps_taken` is how many were accepted
223    /// before the cap tripped; `t_reached`/`t_final` show how far the run actually got.
224    AdaptiveStepStalled {
225        steps_taken: usize,
226        t_reached: f64,
227        t_final: f64,
228    },
229    /// A [`block_graph::GateBinding`] names a block that exists but isn't a `domain=voltage`
230    /// [`block_graph::BlockKind::Sig2Phys`] — the enforced physical/signal-domain boundary:
231    /// an ideal switch's gate is itself a voltage, so any signal driving it must first pass through
232    /// the same `domain=voltage` `Sig2Phys` converter a `V`-source's own magnitude uses, never a
233    /// raw `Pid`/`Vco`/`Hysteresis`/etc. block directly, and never a `domain=current` `Sig2Phys`
234    /// either. `gate` is the ideal switch this binding belongs to; `block` and `found_kind` name
235    /// the offending target and (for a human-readable message) what it actually is.
236    GateTargetNotSig2Voltage {
237        gate: String,
238        block: String,
239        found_kind: &'static str,
240    },
241    /// An independent `V`/`I` source's own literal value in the netlist is a bare symbol that
242    /// names a declared block, but that block isn't a [`block_graph::BlockKind::Sig2Phys`] of
243    /// the matching domain (`V` needs `domain=voltage`, `I` needs `domain=current`) — the
244    /// enforced Signal-to-PS boundary for driving a source's own magnitude from the block graph.
245    SourceNotSig2PhysicalConverter {
246        source: String,
247        block: String,
248        expected_kind: &'static str,
249        found_kind: &'static str,
250    },
251    /// A [`block_graph::BlockKind::Sig2Phys`] converter's own name is also used as an ordinary
252    /// circuit node on an element line — i.e. the converter was *wired into* the physical
253    /// network instead of being *referenced by name* from the one place that can consume it.
254    ///
255    /// A `kind=sig2phys` block has no terminals and stamps nothing into the MNA system; it is
256    /// purely a named Signal-to-PS value another statement substitutes. So a net that happens to
257    /// share its name is just an ordinary, otherwise-undriven node: its KCL row is solved
258    /// consistently at `0` (a resistor to ground is a perfectly non-singular
259    /// `G·v = 0`) and the run silently reports `V(<name>) = 0` next to the block's own,
260    /// correct, non-zero output column. That silent wrong answer is exactly what this variant
261    /// exists to turn into a build-time error, before any step is solved.
262    ///
263    /// The two legitimate ways to consume a `Sig2Phys` are both by name, never by wire:
264    /// an independent `V`/`I` source's own value field (`V1 a 0 VDRV`, see
265    /// [`DaeError::SourceNotSig2PhysicalConverter`]) and an ideal switch's `gate=`/`ctrl=` field
266    /// (see [`DaeError::GateTargetNotSig2Voltage`]).
267    ///
268    /// `block` is the converter as declared, `element` the first element line that wires it, and
269    /// `node` that element's node token as written. Node names are matched case-insensitively,
270    /// matching how the netlist grammar itself treats them (`A` and `a` are one node).
271    Sig2PhysUsedAsCircuitNode {
272        block: String,
273        element: String,
274        node: String,
275    },
276    /// Two blocks in the same slice resolve to the same name — either two
277    /// [`block_graph::BlockInstance`]s declared with the same `.name`, or one block's own
278    /// extra `output_names` entry (a [`block_graph::BlockKind::CScript`]/`CoordinateTransform`/
279    /// `Pmsm`/`Pwm`/`PhaseShiftPwm`'s secondary output alias) colliding with another block's
280    /// name. Without this check, `block_index_by_name`'s `BTreeMap::insert` would silently let
281    /// the later one win, so a `Signal::Block(name)` reference would silently resolve to the
282    /// wrong block instead of failing loudly.
283    DuplicateBlockName(String),
284    /// A checkpoint file is not one, or is a version this build does not read — see
285    /// `checkpoint`'s own module doc comment for the format.
286    CheckpointFormat(String),
287    /// Reading or writing a checkpoint file failed.
288    CheckpointIo {
289        path: std::path::PathBuf,
290        message: String,
291    },
292    /// A checkpoint was loaded into a deck it was not taken from: a different netlist, block
293    /// list, device model, or unknown ordering. `unknowns_in_checkpoint` versus the deck's own
294    /// unknowns is reported to make the usual cause (an edited netlist) visible at a glance.
295    CheckpointDeckMismatch {
296        unknowns_in_checkpoint: Vec<String>,
297        unknowns_in_deck: Vec<String>,
298    },
299    /// A block whose state this crate cannot read is in a run being checkpointed: a `cscript`
300    /// whose library exports none of the `cscript_state_size`/`_write`/`_read` triple (its
301    /// state is an opaque C heap object only that library can lay out). Refused whole, never
302    /// written partially. See `checkpoint`'s own module doc comment.
303    CheckpointUnsupportedBlock {
304        block: String,
305        kind: &'static str,
306    },
307    /// A hand-built [`block_graph::BlockInstance`] carries an `ic` whose length is not its
308    /// block's own state count. Unreachable from a netlist — `general-mna` checks the same thing
309    /// at parse time, with a line number — so this only guards callers that construct blocks
310    /// directly.
311    BlockInitialConditionLength {
312        block: String,
313        expected: usize,
314        got: usize,
315    },
316    /// A [`block_graph::BlockKind`] that requires a `SignalValue::Scalar` input (or a specific
317    /// operand shape, e.g. `Sum`/`Product`'s "all-scalar or all-vector, never mixed" rule, or a
318    /// matrix `Gain`'s own required vector length) received a `SignalValue::Vector` it can't
319    /// accept — the default, fail-closed rule for any block kind without an elementwise/
320    /// broadcast rule of its own. See `book/dev-guide/src/vector-signals.md` for exactly which
321    /// block kinds accept a `Vector` and under what rule.
322    VectorSignalNotSupported {
323        block: String,
324    },
325    /// Two (or more) `SignalValue::Vector`s feeding the same block disagree in length (e.g.
326    /// `Sum`'s own inputs, `MathFn2`'s two operands), or a `Vector`'s own length doesn't match
327    /// a fixed arity the block itself declares (a matrix `Gain`'s own column count, a
328    /// `StateSpace`'s own declared input count).
329    VectorSignalSizeMismatch {
330        block: String,
331        expected: usize,
332        got: usize,
333    },
334}
335
336/// Solves the DC operating point of a netlist containing linear devices plus any number of
337/// `D` (diode) elements, whose piecewise-linear parameters are supplied in `diodes` (keyed by
338/// element name) rather than parsed from the netlist — SPICE's `.model` syntax has no notion
339/// of this crate's breakdown/leakage/forward segments.
340pub fn solve_dc(
341    source: &str,
342    dialect: Dialect,
343    diodes: &BTreeMap<String, IdealDiode>,
344) -> Result<OperatingPoint, DaeError> {
345    let statements = general_mna::parse_and_flatten(source, dialect).map_err(DaeError::Parse)?;
346    let system = MnaBuilder::new(dialect)
347        .build_statements(&statements)
348        .map_err(DaeError::Build)?;
349    fold_and_solve(
350        &system,
351        &statements,
352        diodes,
353        &Scheme::Dc,
354        0.0,
355        &BTreeMap::new(),
356        &BTreeMap::new(),
357    )
358}
359
360/// Which implicit integration formula a given step of [`simulate_transient`] uses. Both
361/// reduce a descriptor DAE `A x + K dx/dt = B u` to the same linear-solve shape [`solve_dc`]
362/// already uses (an effective matrix/RHS folded into the diode LCP unchanged) — see
363/// `docs/architecture.md` and each variant's own derivation below.
364enum Scheme<'a> {
365    /// `A x = B u` directly (no time derivative at all) — an ordinary DC operating point.
366    Dc,
367    /// `dx/dt ~= (x_{n+1} - x_n) / dt`, giving `(A + K/dt) x_{n+1} = B u_{n+1} + (K/dt) x_n`.
368    /// First-order accurate, but tolerant of an inconsistent `x_prev` (doesn't need `x_prev`
369    /// to already satisfy the circuit's algebraic constraints exactly) — this is what makes
370    /// it the right choice for the very first transient step, and immediately after a
371    /// LCP-resolved mode change (see `Trapezoidal`'s doc comment for why trapezoidal is *not*
372    /// safe in either of those cases).
373    BackwardEuler { x_prev: &'a [f64], dt: f64 },
374    /// Second-order accurate. Derived by summing the DAE at both `t_n` and `t_{n+1}` and using
375    /// the trapezoidal relation `x'_n + x'_{n+1} = (2/dt)(x_{n+1} - x_n)` to eliminate the
376    /// derivatives entirely:
377    ///
378    /// ```text
379    /// (A/2)(x_n + x_{n+1}) + (K/dt)(x_{n+1} - x_n) = (B/2)(u_n + u_{n+1})
380    /// => (A/2 + K/dt) x_{n+1} = (B/2)(u_n + u_{n+1}) + (K/dt - A/2) x_n
381    /// ```
382    ///
383    /// Crucially, `B u` includes each diode's *nonlinear* current, not just the netlist's own
384    /// (constant, in this crate's scope) sources — true trapezoidal accuracy requires
385    /// averaging that too, not just the linear part. Since the new step's diode current
386    /// contributes `(B/2) * raw_Ioff_{n+1}` rather than the full `B * raw_Ioff_{n+1}`, the
387    /// LCP fold's per-diode coupling coefficients are scaled by `0.5` for this scheme (see
388    /// `fold_and_solve`'s `coupling_scale`) — while the *previous* step's already-known
389    /// current contributes its own `(B/2) * raw_Ioff_n` term directly to the effective RHS,
390    /// via `prev_diode_raw_ioff`.
391    ///
392    /// This derivation assumes `x_n` already satisfies the circuit's algebraic constraints
393    /// exactly (so that summing the DAE at `t_n` contributes nothing extra beyond
394    /// `B u_n - A x_n = 0`) — true right after a [`Scheme::BackwardEuler`] or [`Scheme::Dc`]
395    /// step, not true in general starting from an arbitrary `x_initial`. This is the DAE
396    /// analog of "trapezoidal needs consistent initial conditions," and is why
397    /// [`simulate_transient`] always starts with backward Euler and falls back to it whenever
398    /// a diode's resolved segment changes between consecutive steps.
399    ///
400    /// That "constant, in this crate's scope" assumption about the netlist's own linear
401    /// sources held until [`TransientFunction`](general_mna::TransientFunction) support was
402    /// added: for a genuinely time-varying `V`/`I` source, `u_n != u_{n+1}` for that source's
403    /// own column too, not just for diode currents — `fold_and_solve` evaluates `numeric0.u`
404    /// (i.e. `B u_{n+1}`) a second time at `t_n = t_{n+1} - dt` and averages the two, exactly
405    /// mirroring the diode-current averaging already described above, rather than silently
406    /// reusing the same "constant source" shortcut for a source that no longer is one.
407    Trapezoidal {
408        x_prev: &'a [f64],
409        dt: f64,
410        prev_diode_raw_ioff: &'a BTreeMap<String, f64>,
411    },
412}
413
414/// Runs a transient simulation of a netlist containing linear devices (including storage —
415/// `C`, `L`) plus any number of `D` diodes, from `t = 0` to `t_final`, starting from
416/// `x_initial` (all zero if `None` — the usual "circuit at rest, then a step/DC source turns
417/// on at `t = 0`" case used by this crate's own tests). `step` picks fixed or adaptive timing
418/// — see [`TimeStep`]/[`AdaptiveConfig`]; `TimeStep::Fixed(dt)` reproduces exactly the fixed-`dt`
419/// behavior this function had before `TimeStep` existed.
420///
421/// Uses [`Scheme::Trapezoidal`] (second-order accurate, matching Xyce/SPICE's own default) for
422/// most steps, but falls back to [`Scheme::BackwardEuler`] for the very first step (an
423/// arbitrary `x_initial` is not guaranteed to satisfy the circuit's algebraic constraints —
424/// trapezoidal needs that) and for any step where a diode's resolved segment (breakdown /
425/// leakage / forward) differs from the *previous* step's — the architecture's own sanctioned
426/// policy for staying robust across an LCP-resolved mode change (see `docs/architecture.md`).
427/// This is a lagging, not predictive, policy: it reacts to the mode change already observed in
428/// the previous step's result, rather than trying to foresee this step's mode before solving
429/// it — simpler, and sufficient for the well-behaved circuits this crate targets so far.
430///
431/// No ideal switch/PWM support yet in the transient loop — see this crate's journal for why.
432/// The `x` a transient run starts from, in `system.unknowns` order.
433///
434/// An explicit `x_initial` from the caller always wins — it is the more specific instruction,
435/// and several of this crate's own tests hand-build one. Failing that, the netlist's own `ic=`
436/// values are honored via [`general_mna::MnaSystem::initial_state`], which *assigns* the
437/// declared states — an `ic`-bearing inductor's own branch current, an `ic`-bearing capacitor's
438/// node-voltage difference — and leaves every other unknown at rest, without touching the
439/// circuit's equations. Failing *that* — the overwhelmingly common case of a netlist with no
440/// `ic=` anywhere — the circuit starts from rest, exactly as before.
441///
442/// A seeded vector is therefore **not** a consistent operating point: the other node voltages
443/// and every source's branch current start at zero even where the circuit's algebraic
444/// constraints say otherwise. That is fine here, and worth saying out loud rather than leaving
445/// for the next reader to rediscover: [`simulate_transient`] already takes its first step with
446/// backward Euler precisely because an arbitrary `x_initial` need not satisfy those constraints,
447/// and that first step is what re-imposes them. The one thing an inconsistent seed must not do
448/// is pass silently, and it does not — `initial_state` reports an assignment that contradicts an
449/// equation nothing is left free to satisfy, as [`DaeError::InitialCondition`].
450///
451/// `values` is built at `t = 0` with every diode at its own off-segment reference slope and zero
452/// Norton current, matching how the first real step's `base_values` is built in `solve_step`, so
453/// that the consistency check `initial_state` runs sees the same circuit the first step will.
454pub(crate) fn resolve_initial_state(
455    system: &MnaSystem,
456    diodes: &BTreeMap<String, IdealDiode>,
457    x_initial: Option<&[f64]>,
458) -> Result<Vec<f64>, DaeError> {
459    if let Some(x) = x_initial {
460        return Ok(x.to_vec());
461    }
462    let mut values = BTreeMap::new();
463    for (name, diode) in diodes {
464        values.insert(format!("{name}_G"), diode.g_off);
465        values.insert(format!("{name}_Ioff"), 0.0);
466    }
467    for (name, transient_fn) in &system.transient_sources {
468        values.insert(name.clone(), transient_fn.value_at(0.0));
469    }
470    match system
471        .initial_state(&values, general_mna::DEFAULT_INITIAL_STATE_TOLERANCE)
472        .map_err(DaeError::InitialCondition)?
473    {
474        Some(x) => Ok(x),
475        None => Ok(vec![0.0; system.order()]),
476    }
477}
478
479pub fn simulate_transient(
480    source: &str,
481    dialect: Dialect,
482    diodes: &BTreeMap<String, IdealDiode>,
483    x_initial: Option<&[f64]>,
484    t_final: f64,
485    step: TimeStep,
486) -> Result<Vec<(f64, OperatingPoint)>, DaeError> {
487    let statements = general_mna::parse_and_flatten(source, dialect).map_err(DaeError::Parse)?;
488    let system = MnaBuilder::new(dialect)
489        .build_statements(&statements)
490        .map_err(DaeError::Build)?;
491
492    let mut x_prev = resolve_initial_state(&system, diodes, x_initial)?;
493    let mut x_prev_prev: Option<Vec<f64>> = None;
494    let mut prev_diode_raw_ioff: BTreeMap<String, f64> = BTreeMap::new();
495    let mut prev_segments: Option<Vec<Segment>> = None;
496    let mut ringing_cooldown: u32 = 0;
497
498    let mut trace = Vec::new();
499    let mut t = 0.0;
500
501    match step {
502        TimeStep::Fixed(dt) => {
503            let steps = (t_final / dt).round() as usize;
504            trace.reserve(steps);
505            for step_index in 0..steps {
506                t += dt;
507                let forced = step_index == 0 || ringing_cooldown > 0;
508                let (point, used_backward_euler) = step_with_fallback(
509                    &system,
510                    &statements,
511                    diodes,
512                    x_prev_prev.as_deref(),
513                    &x_prev,
514                    dt,
515                    &prev_diode_raw_ioff,
516                    &prev_segments,
517                    forced,
518                    t,
519                    &BTreeMap::new(),
520                    &BTreeMap::new(),
521                )?;
522                if used_backward_euler && !forced {
523                    ringing_cooldown = RINGING_COOLDOWN_STEPS;
524                } else if ringing_cooldown > 0 {
525                    ringing_cooldown = ringing_cooldown.saturating_sub(1);
526                }
527
528                prev_segments = Some(classify_segments(&point));
529                prev_diode_raw_ioff = point
530                    .diode_names
531                    .iter()
532                    .cloned()
533                    .zip(point.diode_raw_ioff.iter().copied())
534                    .collect();
535                x_prev_prev = Some(std::mem::replace(&mut x_prev, point.x.clone()));
536                trace.push((t, point));
537            }
538        }
539        TimeStep::Adaptive(config) => {
540            let mut dt_next = config.dt_init;
541            let mut step_index = 0usize;
542            while t < t_final {
543                let dt_trial = dt_next.min(t_final - t);
544                let forced = step_index == 0 || ringing_cooldown > 0;
545                let (point, used_backward_euler, dt_used, next_dt) = step_control::adaptive_step(
546                    &system,
547                    &statements,
548                    diodes,
549                    x_prev_prev.as_deref(),
550                    &x_prev,
551                    dt_trial,
552                    &prev_diode_raw_ioff,
553                    &prev_segments,
554                    forced,
555                    &config,
556                    t,
557                )?;
558                if used_backward_euler && !forced {
559                    ringing_cooldown = RINGING_COOLDOWN_STEPS;
560                } else if ringing_cooldown > 0 {
561                    ringing_cooldown = ringing_cooldown.saturating_sub(1);
562                }
563
564                t += dt_used;
565                dt_next = next_dt;
566                prev_segments = Some(classify_segments(&point));
567                prev_diode_raw_ioff = point
568                    .diode_names
569                    .iter()
570                    .cloned()
571                    .zip(point.diode_raw_ioff.iter().copied())
572                    .collect();
573                x_prev_prev = Some(std::mem::replace(&mut x_prev, point.x.clone()));
574                trace.push((t, point));
575                step_index += 1;
576            }
577        }
578    }
579    Ok(trace)
580}
581
582#[derive(Debug, Clone, Copy, PartialEq, Eq)]
583pub(crate) enum Segment {
584    Breakdown,
585    Leakage,
586    Forward,
587}
588
589fn classify_segments(point: &OperatingPoint) -> Vec<Segment> {
590    point
591        .diode_z
592        .iter()
593        .map(|&(z1, z2)| {
594            if z1 > 1e-9 {
595                Segment::Breakdown
596            } else if z2 > 1e-9 {
597                Segment::Forward
598            } else {
599                Segment::Leakage
600            }
601        })
602        .collect()
603}
604
605/// Detects trapezoidal "ringing": trapezoidal integration is A-stable (bounded) but not
606/// L-stable, so a very lightly damped mode (e.g. a switch node left nearly floating during
607/// ideal switch dead time, with only tiny leakage conductance) doesn't decay under it — it
608/// oscillates at the Nyquist frequency (sign-flipping every single step) with roughly
609/// constant amplitude instead. Detected as three consecutive values of the same unknown
610/// alternating in sign (`x_prev_prev`, `x_prev`, `trial` share the outer two signs, opposite
611/// the middle one) with the magnitude not shrinking step to step — the direct signature of a
612/// sustained oscillation, as opposed to a legitimate sign change while settling (which
613/// shrinks, not sustains, in magnitude) or ordinary numerical noise near zero (excluded by
614/// the noise floor). Found and root-caused via `general-simulator`'s LLC validation
615/// (`crates/dae-runtime/examples/llc_validation.rs`): the switching node's voltage oscillated
616/// between roughly +/-3000V for the entire ~200ns ideal switch dead-time window every period, while
617/// every other tracked quantity (notably the actual circuit output) stayed smooth and
618/// physically reasonable throughout — falling back to backward Euler (which has no such
619/// weakness) the instant this is detected fixes it, consistent with backward Euler already
620/// being this crate's answer to every other kind of transient stiffness.
621fn is_ringing(x_prev_prev: &[f64], x_prev: &[f64], trial: &[f64]) -> bool {
622    const NOISE_FLOOR: f64 = 1e-9;
623    x_prev_prev
624        .iter()
625        .zip(x_prev)
626        .zip(trial)
627        .any(|((&a, &b), &c)| {
628            if a.abs() < NOISE_FLOOR || b.abs() < NOISE_FLOOR || c.abs() < NOISE_FLOOR {
629                return false;
630            }
631            let alternating = (a > 0.0) == (c > 0.0) && (a > 0.0) != (b > 0.0);
632            alternating && c.abs() >= b.abs() * 0.9
633        })
634}
635
636/// How many additional steps to keep using backward Euler after a ringing catch, before
637/// trusting trapezoidal again. A single corrective backward-Euler step pulls the ringing
638/// mode's *value* back to something reasonable, but doesn't fully re-establish it as "settled"
639/// in the sense trapezoidal's own derivation needs (see [`Scheme::Trapezoidal`]'s doc comment
640/// on consistent initial conditions) — resuming trapezoidal immediately was observed to let a
641/// smaller residual oscillation resume and drift to a stale-but-stable wrong plateau, rather
642/// than fully recovering. A short cooldown of plain, unconditionally-stable backward-Euler
643/// steps lets the correction actually settle first.
644const RINGING_COOLDOWN_STEPS: u32 = 3;
645
646/// Solves one timestep, choosing between [`Scheme::Trapezoidal`] and [`Scheme::BackwardEuler`]
647/// the same lagging way [`simulate_transient`] and [`simulate_transient_with_ideal_switches`] both
648/// need: `force_backward_euler` covers reasons known *before* solving (the very first step, a
649/// gate state that just changed, or an active [`RINGING_COOLDOWN_STEPS`] cooldown); a diode
650/// segment change and trapezoidal ringing (see [`is_ringing`]) are only detectable *after*
651/// trying trapezoidal, so either one redoes the step with backward Euler. `x_prev_prev` (the
652/// state two steps back) is `None` for the first two steps of a run, when ringing can't yet be
653/// detected — the existing `force_backward_euler` on the very first step, plus ringing needing
654/// at least one full oscillation to show its signature, means this is never a gap in practice.
655///
656/// Returns `(point, used_backward_euler)` — callers use the flag to drive a
657/// [`RINGING_COOLDOWN_STEPS`]-step cooldown after any backward-Euler fallback (not just a
658/// ringing-triggered one — a segment or gate change deserves the same settling courtesy).
659#[allow(clippy::too_many_arguments)]
660fn step_with_fallback(
661    system: &MnaSystem,
662    statements: &[Statement],
663    diodes: &BTreeMap<String, IdealDiode>,
664    x_prev_prev: Option<&[f64]>,
665    x_prev: &[f64],
666    dt: f64,
667    prev_diode_raw_ioff: &BTreeMap<String, f64>,
668    prev_segments: &Option<Vec<Segment>>,
669    force_backward_euler: bool,
670    t: f64,
671    extra_values: &BTreeMap<String, f64>,
672    extra_values_prev: &BTreeMap<String, f64>,
673) -> Result<(OperatingPoint, bool), DaeError> {
674    if force_backward_euler {
675        return fold_and_solve(
676            system,
677            statements,
678            diodes,
679            &Scheme::BackwardEuler { x_prev, dt },
680            t,
681            extra_values,
682            extra_values_prev,
683        )
684        .map(|point| (point, true));
685    }
686    let trial = fold_and_solve(
687        system,
688        statements,
689        diodes,
690        &Scheme::Trapezoidal {
691            x_prev,
692            dt,
693            prev_diode_raw_ioff,
694        },
695        t,
696        extra_values,
697        extra_values_prev,
698    )?;
699    let trial_segments = classify_segments(&trial);
700    let segments_changed = Some(&trial_segments) != prev_segments.as_ref();
701    let ringing = x_prev_prev.is_some_and(|xpp| is_ringing(xpp, x_prev, &trial.x));
702    if segments_changed || ringing {
703        fold_and_solve(
704            system,
705            statements,
706            diodes,
707            &Scheme::BackwardEuler { x_prev, dt },
708            t,
709            extra_values,
710            extra_values_prev,
711        )
712        .map(|point| (point, true))
713    } else {
714        Ok((trial, false))
715    }
716}
717
718/// Solves the DC operating point of a netlist containing linear devices, ordinary `D` diodes
719/// (`diodes`), and any number of `IdealSwitch` instances (`ideal_switches`), each with its own known gate
720/// state (see [`GateState`], a re-export of `general_mna::SwitchState` — the same concept: an
721/// exogenous, externally-decided mode, not something the LCP resolves).
722///
723/// A gated-on ideal switch is stamped as a plain `r_on` switch (reusing `general-mna`'s existing
724/// switch mechanism — every gated-on ideal switch in one call shares `shared_r_on`, matching that
725/// mechanism's own single-shared-resistance design; per-instance `Ron` is a possible future
726/// extension, not needed yet). A gated-off ideal switch is folded into the LCP exactly like an
727/// ordinary diode, using [`IdealSwitch::body_diode_for_drain_source_stamping`] (not `body_diode`
728/// directly) — see that method's own doc comment for why: the netlist declares an ideal switch's two
729/// terminals in plain SPICE-conventional `(drain, source)` order, and the mirrored curve is
730/// what makes evaluating it against those nodes as-declared give the physically correct
731/// result. Each ideal switch element must still use device letter `'D'` in the netlist text (not
732/// `'M'`) — see this crate's `docs`/journal for why.
733pub fn solve_dc_with_ideal_switches(
734    source: &str,
735    dialect: Dialect,
736    diodes: &BTreeMap<String, IdealDiode>,
737    ideal_switches: &BTreeMap<String, (IdealSwitch, GateState)>,
738    shared_r_on: f64,
739) -> Result<OperatingPoint, DaeError> {
740    let statements = general_mna::parse_and_flatten(source, dialect).map_err(DaeError::Parse)?;
741    let (system, all_diodes) =
742        build_with_ideal_switches(&statements, dialect, diodes, ideal_switches, shared_r_on)?;
743    fold_and_solve(
744        &system,
745        &statements,
746        &all_diodes,
747        &Scheme::Dc,
748        0.0,
749        &BTreeMap::new(),
750        &BTreeMap::new(),
751    )
752}
753
754/// Runs a transient simulation of a netlist containing linear devices, ordinary `D` diodes,
755/// and any number of `IdealSwitch` instances whose gate state can vary over time — a PWM driver,
756/// unlike [`solve_dc_with_ideal_switches`]'s fixed per-call state. `gate_signal(name, t)` is called
757/// once per ideal switch per timestep to get that instance's [`GateState`] at time `t`; the caller
758/// owns the PWM logic entirely (duty cycle, frequency, phase — this crate has no opinion).
759///
760/// Because a gate state change means the *symbolic* `general-mna` system itself must be
761/// rebuilt (a switch and a `'D'`-stamped diode are structurally different stamps, not just
762/// different numeric values — see `docs/architecture.md`), this rebuilds the system every
763/// timestep, unlike [`simulate_transient`]'s single build reused throughout. A documented
764/// scope choice, not an oversight: fine for the small circuits this crate targets so far: see
765/// this crate's journal for the reasoning and what a future optimization would look like.
766///
767/// Uses the same trapezoidal-with-backward-Euler-fallback policy as [`simulate_transient`],
768/// extended to also force backward Euler on any step where *any* ideal switch's gate state just
769/// changed from the previous step — a gate transition is exactly the kind of topology change
770/// the architecture's "backward Euler after a mode change" policy exists for, arguably more so
771/// than a diode's segment change, since it's a real switching event a PWM converter circuit
772/// will trigger constantly.
773#[allow(clippy::too_many_arguments)]
774pub fn simulate_transient_with_ideal_switches(
775    source: &str,
776    dialect: Dialect,
777    diodes: &BTreeMap<String, IdealDiode>,
778    ideal_switches: &BTreeMap<String, IdealSwitch>,
779    gate_signal: impl Fn(&str, f64) -> GateState,
780    shared_r_on: f64,
781    x_initial: Option<&[f64]>,
782    t_final: f64,
783    dt: f64,
784) -> Result<Vec<(f64, OperatingPoint)>, DaeError> {
785    let states_at = |t: f64| -> BTreeMap<String, (IdealSwitch, GateState)> {
786        ideal_switches
787            .iter()
788            .map(|(name, m)| (name.clone(), (*m, gate_signal(name, t))))
789            .collect()
790    };
791
792    let statements = general_mna::parse_and_flatten(source, dialect).map_err(DaeError::Parse)?;
793    let (system0, _) =
794        build_with_ideal_switches(&statements, dialect, diodes, &states_at(0.0), shared_r_on)?;
795    let mut x_prev = resolve_initial_state(&system0, diodes, x_initial)?;
796    let mut x_prev_prev: Option<Vec<f64>> = None;
797    let mut prev_diode_raw_ioff: BTreeMap<String, f64> = BTreeMap::new();
798    let mut prev_segments: Option<Vec<Segment>> = None;
799    let mut prev_gate_states: Option<BTreeMap<String, GateState>> = None;
800    let mut ringing_cooldown: u32 = 0;
801
802    let steps = (t_final / dt).round() as usize;
803    let mut trace = Vec::with_capacity(steps);
804    let mut t = 0.0;
805    for step_index in 0..steps {
806        t += dt;
807        let states = states_at(t);
808        let gate_states: BTreeMap<String, GateState> = states
809            .iter()
810            .map(|(name, (_, state))| (name.clone(), *state))
811            .collect();
812        let gate_changed = prev_gate_states.as_ref() != Some(&gate_states);
813        let forced = step_index == 0 || gate_changed || ringing_cooldown > 0;
814
815        let (system, all_diodes) =
816            build_with_ideal_switches(&statements, dialect, diodes, &states, shared_r_on)?;
817        let (point, used_backward_euler) = step_with_fallback(
818            &system,
819            &statements,
820            &all_diodes,
821            x_prev_prev.as_deref(),
822            &x_prev,
823            dt,
824            &prev_diode_raw_ioff,
825            &prev_segments,
826            forced,
827            t,
828            &BTreeMap::new(),
829            &BTreeMap::new(),
830        )?;
831        if used_backward_euler && !forced {
832            ringing_cooldown = RINGING_COOLDOWN_STEPS;
833        } else if ringing_cooldown > 0 {
834            ringing_cooldown = ringing_cooldown.saturating_sub(1);
835        }
836
837        prev_segments = Some(classify_segments(&point));
838        prev_gate_states = Some(gate_states);
839        prev_diode_raw_ioff = point
840            .diode_names
841            .iter()
842            .cloned()
843            .zip(point.diode_raw_ioff.iter().copied())
844            .collect();
845        x_prev_prev = Some(std::mem::replace(&mut x_prev, point.x.clone()));
846        trace.push((t, point));
847    }
848    Ok(trace)
849}
850
851fn build_with_ideal_switches(
852    statements: &[Statement],
853    dialect: Dialect,
854    diodes: &BTreeMap<String, IdealDiode>,
855    ideal_switches: &BTreeMap<String, (IdealSwitch, GateState)>,
856    shared_r_on: f64,
857) -> Result<(MnaSystem, BTreeMap<String, IdealDiode>), DaeError> {
858    let mut options = BuildOptions {
859        on_resistance: Expression::Constant(shared_r_on),
860        ..BuildOptions::default()
861    };
862
863    let mut all_diodes = diodes.clone();
864    for (name, (switch, state)) in ideal_switches {
865        match state {
866            GateState::On => options.set_switch(name, GateState::On),
867            GateState::Off => {
868                all_diodes.insert(name.clone(), switch.body_diode_for_drain_source_stamping());
869            }
870        }
871    }
872
873    let system = MnaBuilder::with_options(dialect, options)
874        .build_statements(statements)
875        .map_err(DaeError::Build)?;
876    Ok((system, all_diodes))
877}
878
879#[allow(clippy::too_many_arguments)]
880fn fold_and_solve(
881    system: &MnaSystem,
882    statements: &[Statement],
883    diodes: &BTreeMap<String, IdealDiode>,
884    scheme: &Scheme,
885    t: f64,
886    extra_values: &BTreeMap<String, f64>,
887    extra_values_prev: &BTreeMap<String, f64>,
888) -> Result<OperatingPoint, DaeError> {
889    let nodes = topology::diode_nodes(statements);
890    let order = system.order();
891
892    // Fix every diode's conductance at its canonical reference slope and its Norton current at
893    // zero: this is the "A0"/"u0" the whole LCP fold above is built on. Also supply every
894    // time-varying V/I source's own current value at `t` -- `system.transient_sources` is
895    // empty for a netlist with no `SIN`/`PULSE`/`EXP`/`PWL`/`SFFM` source, so this is a no-op
896    // for every existing (plain, static-source) netlist. `t=0.0` for `Scheme::Dc` matches
897    // standard SPICE convention (a DC operating point uses a transient source's own value at
898    // time zero, not its DC-only fallback, when no separate `DC` value was given). `extra_values`
899    // is this step's full block-graph output map (empty for every caller with no block graph at
900    // all) -- `system.evaluate` only ever substitutes a symbol actually present in an
901    // expression tree, so passing every block's output here unconditionally is harmless; the
902    // *only* symbols that can legitimately appear this way are a `BlockKind::Sig2Phys`
903    // converter's own name, since `simulate_transient_with_blocks` already
904    // rejects any other block name appearing as a bare V/I source literal before any step runs
905    // (see `DaeError::SourceNotSig2PhysicalConverter`).
906    let mut base_values = BTreeMap::new();
907    for (name, diode) in diodes {
908        base_values.insert(format!("{name}_G"), diode.g_off);
909        base_values.insert(format!("{name}_Ioff"), 0.0);
910    }
911    for (name, transient_fn) in &system.transient_sources {
912        base_values.insert(name.clone(), transient_fn.value_at(t));
913    }
914    for (name, value) in extra_values {
915        base_values.insert(name.clone(), *value);
916    }
917    let numeric0 = system.evaluate(&base_values).map_err(DaeError::Evaluate)?;
918
919    // a_eff/u_eff_base: the effective matrix and the part of the effective RHS that doesn't
920    // depend on this step's own diode currents (those are added per-diode below, since
921    // Trapezoidal's history term needs each diode's own B column). coupling_scale multiplies
922    // how much *this* step's resolved diode current contributes to x/V — see Scheme::Trapezoidal.
923    let (a_eff, u_eff_base, coupling_scale): (general_mna::Matrix<f64>, Vec<f64>, f64) =
924        match scheme {
925            Scheme::Dc => (numeric0.a.clone(), numeric0.u.clone(), 1.0),
926            Scheme::BackwardEuler { x_prev, dt } => {
927                let mut a_eff = numeric0.a.clone();
928                for row in 0..order {
929                    for col in 0..order {
930                        a_eff[(row, col)] += numeric0.k[(row, col)] / dt;
931                    }
932                }
933                let u_eff: Vec<f64> = (0..order)
934                    .map(|row| {
935                        let k_x: f64 = (0..order)
936                            .map(|col| numeric0.k[(row, col)] * x_prev[col])
937                            .sum();
938                        numeric0.u[row] + k_x / dt
939                    })
940                    .collect();
941                (a_eff, u_eff, 1.0)
942            }
943            Scheme::Trapezoidal { x_prev, dt, .. } => {
944                let mut a_eff = numeric0.a.clone();
945                for row in 0..order {
946                    for col in 0..order {
947                        a_eff[(row, col)] = a_eff[(row, col)] / 2.0 + numeric0.k[(row, col)] / dt;
948                    }
949                }
950                // True trapezoidal RHS uses (B/2)*(u_n + u_{n+1}), not B*u_{n+1} alone (see
951                // Scheme::Trapezoidal's own doc comment) -- correct as-is only because every
952                // netlist source was constant (u_n == u_{n+1}) until time-varying sources
953                // existed. `u_prev` re-evaluates the netlist's own linear sources at `t - dt`
954                // for a `TransientFunction` source (an analytic, re-evaluatable function of
955                // time), and substitutes `extra_values_prev` (the *previous* step's already-
956                // computed block outputs -- a block-driven source has no analytic form to
957                // re-evaluate, only the discrete sample already used last step, zero-order-held
958                // for that step, exactly the sampled-data convention this whole block graph
959                // uses) for a `Sig2Phys`-driven one. Skipped (and `numeric0.u`
960                // cloned directly, exactly the previous behavior) only when there are neither,
961                // since u_n == u_{n+1} trivially and a second `evaluate` call would be pure
962                // overhead.
963                let u_prev: Vec<f64> =
964                    if system.transient_sources.is_empty() && extra_values_prev.is_empty() {
965                        numeric0.u.clone()
966                    } else {
967                        let mut base_values_prev = base_values.clone();
968                        for (name, transient_fn) in &system.transient_sources {
969                            base_values_prev.insert(name.clone(), transient_fn.value_at(t - dt));
970                        }
971                        for (name, value) in extra_values_prev {
972                            base_values_prev.insert(name.clone(), *value);
973                        }
974                        system
975                            .evaluate(&base_values_prev)
976                            .map_err(DaeError::Evaluate)?
977                            .u
978                    };
979                let u_eff: Vec<f64> = (0..order)
980                    .map(|row| {
981                        let k_x: f64 = (0..order)
982                            .map(|col| numeric0.k[(row, col)] * x_prev[col])
983                            .sum();
984                        let a_x: f64 = (0..order)
985                            .map(|col| numeric0.a[(row, col)] * x_prev[col])
986                            .sum();
987                        (numeric0.u[row] + u_prev[row]) / 2.0 + k_x / dt - a_x / 2.0
988                    })
989                    .collect();
990                (a_eff, u_eff, 0.5)
991            }
992        };
993
994    struct DiodeInfo {
995        name: String,
996        canonical: pwl_devices::IdealDiodeCanonical,
997        p: Option<usize>,
998        n: Option<usize>,
999        w: Vec<f64>,
1000    }
1001
1002    let mut u_eff = u_eff_base;
1003    let mut infos = Vec::with_capacity(diodes.len());
1004    for (name, diode) in diodes {
1005        let column = system
1006            .inputs
1007            .iter()
1008            .position(|input| input == name)
1009            .ok_or_else(|| DaeError::UnknownDiodeInput(name.clone()))?;
1010        let b_col: Vec<f64> = (0..order).map(|row| numeric0.b[(row, column)]).collect();
1011
1012        // Trapezoidal history: the previous step's already-known diode current contributes
1013        // its own (B/2)*raw_Ioff_n term to this step's effective RHS directly (only this
1014        // step's *new* current needs the LCP-fold coupling_scale treatment below).
1015        if let Scheme::Trapezoidal {
1016            prev_diode_raw_ioff,
1017            ..
1018        } = scheme
1019        {
1020            if let Some(&prev_ioff) = prev_diode_raw_ioff.get(name) {
1021                for (row, &b) in b_col.iter().enumerate() {
1022                    u_eff[row] += 0.5 * b * prev_ioff;
1023                }
1024            }
1025        }
1026
1027        let w = dense_solve(&a_eff, &b_col).map_err(DaeError::Linear)?;
1028
1029        let terminals = &nodes[name];
1030        let p = topology::node_index(&system.unknowns, &terminals.positive);
1031        let n = topology::node_index(&system.unknowns, &terminals.negative);
1032
1033        infos.push(DiodeInfo {
1034            name: name.clone(),
1035            canonical: diode.canonical(),
1036            p,
1037            n,
1038            w,
1039        });
1040    }
1041
1042    let x0 = dense_solve(&a_eff, &u_eff).map_err(DaeError::Linear)?;
1043
1044    let get = |x: &[f64], index: Option<usize>| index.map(|i| x[i]).unwrap_or(0.0);
1045
1046    let n_diodes = infos.len();
1047    let v0: Vec<f64> = infos
1048        .iter()
1049        .map(|d| get(&x0, d.p) - get(&x0, d.n))
1050        .collect();
1051    // gamma[k][j] = w_j[p_k] - w_j[n_k]: how much diode j's unit raw current moves diode k's
1052    // own terminal voltage (the diagonal, gamma[k][k], is the self term; everything else is
1053    // genuine cross-coupling through the shared linear network).
1054    let gamma: Vec<Vec<f64>> = infos
1055        .iter()
1056        .map(|k| {
1057            infos
1058                .iter()
1059                .map(|j| get(&j.w, k.p) - get(&j.w, k.n))
1060                .collect()
1061        })
1062        .collect();
1063
1064    let dim = 2 * n_diodes;
1065    let mut m = vec![vec![0.0; dim]; dim];
1066    let mut q = vec![0.0; dim];
1067    for k in 0..n_diodes {
1068        let row_w1 = 2 * k;
1069        q[row_w1] = v0[k] - infos[k].canonical.v_breakdown;
1070        for (j, info_j) in infos.iter().enumerate() {
1071            m[row_w1][2 * j] += -gamma[k][j] * info_j.canonical.delta_br * coupling_scale;
1072            m[row_w1][2 * j + 1] += gamma[k][j] * info_j.canonical.delta_on * coupling_scale;
1073        }
1074        m[row_w1][2 * k] += 1.0;
1075
1076        let row_w2 = 2 * k + 1;
1077        q[row_w2] = infos[k].canonical.v_th - v0[k];
1078        for (j, info_j) in infos.iter().enumerate() {
1079            m[row_w2][2 * j] += gamma[k][j] * info_j.canonical.delta_br * coupling_scale;
1080            m[row_w2][2 * j + 1] += -gamma[k][j] * info_j.canonical.delta_on * coupling_scale;
1081        }
1082        m[row_w2][2 * k + 1] += 1.0;
1083    }
1084
1085    let sol = lcp_solver::solve(&m, &q).map_err(DaeError::Lcp)?;
1086
1087    let mut x = x0;
1088    let mut diode_z = Vec::with_capacity(n_diodes);
1089    let mut diode_raw_ioff = Vec::with_capacity(n_diodes);
1090    for (k, info) in infos.iter().enumerate() {
1091        let z1 = sol.z[2 * k];
1092        let z2 = sol.z[2 * k + 1];
1093        // True (unscaled) physical current, used for reporting and as the next step's
1094        // trapezoidal history — NOT the same as how much it moves x this step (that's scaled
1095        // by coupling_scale, since only half of it entered the effective RHS above).
1096        let raw_ioff = info.canonical.delta_on * z2 - info.canonical.delta_br * z1;
1097        let x_contribution = raw_ioff * coupling_scale;
1098        for (xi, wi) in x.iter_mut().zip(info.w.iter()) {
1099            *xi += wi * x_contribution;
1100        }
1101        diode_z.push((z1, z2));
1102        diode_raw_ioff.push(raw_ioff);
1103    }
1104
1105    Ok(OperatingPoint {
1106        unknowns: system.unknowns.clone(),
1107        x,
1108        diode_names: infos.into_iter().map(|d| d.name).collect(),
1109        diode_z,
1110        diode_raw_ioff,
1111    })
1112}
1113
1114#[cfg(test)]
1115mod scheme_tests {
1116    //! `Scheme` and `fold_and_solve` are private, so the convergence-rate check that proves
1117    //! trapezoidal is actually second-order (not just "still passes at the same tolerance")
1118    //! has to live here rather than in `tests/`.
1119    use super::*;
1120
1121    enum Kind {
1122        BackwardEuler,
1123        /// Trapezoidal for every step after the first (which is always backward Euler, for a
1124        /// consistent `x0` — matching `simulate_transient`'s own policy).
1125        Trapezoidal,
1126    }
1127
1128    /// Runs the same RC circuit as `tests/transient.rs`'s
1129    /// `rc_charging_matches_closed_form_exponential` (`Vc(t) = 5*(1-e^-t)`) with a single
1130    /// scheme throughout, and returns the error against the closed-form solution at `t_final`.
1131    fn rc_final_error(kind: &Kind, dt: f64, t_final: f64) -> f64 {
1132        let source = "V1 a 0 5\nR1 a b 1\nC1 b 0 1";
1133        let statements = general_mna::parse_and_flatten(source, Dialect::Ngspice).unwrap();
1134        let system = MnaBuilder::new(Dialect::Ngspice)
1135            .build_statements(&statements)
1136            .unwrap();
1137        let diodes = BTreeMap::new();
1138        let steps = (t_final / dt).round() as usize;
1139
1140        let mut x = vec![0.0; system.order()];
1141        for step in 0..steps {
1142            let t = (step + 1) as f64 * dt;
1143            let use_be = matches!(kind, Kind::BackwardEuler) || step == 0;
1144            let point = if use_be {
1145                fold_and_solve(
1146                    &system,
1147                    &statements,
1148                    &diodes,
1149                    &Scheme::BackwardEuler { x_prev: &x, dt },
1150                    t,
1151                    &BTreeMap::new(),
1152                    &BTreeMap::new(),
1153                )
1154            } else {
1155                fold_and_solve(
1156                    &system,
1157                    &statements,
1158                    &diodes,
1159                    &Scheme::Trapezoidal {
1160                        x_prev: &x,
1161                        dt,
1162                        prev_diode_raw_ioff: &BTreeMap::new(),
1163                    },
1164                    t,
1165                    &BTreeMap::new(),
1166                    &BTreeMap::new(),
1167                )
1168            }
1169            .unwrap();
1170            x = point.x;
1171        }
1172
1173        let vb_index = system.unknowns.iter().position(|u| u == "V(b)").unwrap();
1174        let expected = 5.0 * (1.0 - (-t_final).exp());
1175        (x[vb_index] - expected).abs()
1176    }
1177
1178    /// The standard way to verify a method's convergence *order* is to halve `dt` and check
1179    /// how much the error shrinks, not to compare absolute error magnitudes at one `dt`
1180    /// against another method's (which depends on unknown, method-specific constants and can
1181    /// easily mislead). Backward Euler is first-order: halving `dt` should roughly halve the
1182    /// error. Both checked over the same halving, at deliberately coarse `dt` so the halving
1183    /// ratio is clearly visible before floating-point/near-exact-cancellation noise matters.
1184    #[test]
1185    fn backward_euler_error_roughly_halves_when_dt_halves() {
1186        let t_final = 2.0;
1187        let error_coarse = rc_final_error(&Kind::BackwardEuler, 0.1, t_final);
1188        let error_fine = rc_final_error(&Kind::BackwardEuler, 0.05, t_final);
1189        let ratio = error_coarse / error_fine;
1190        assert!(
1191            (1.7..=2.3).contains(&ratio),
1192            "backward-Euler error ratio (dt vs dt/2) = {ratio}, expected close to 2.0 (first-order)"
1193        );
1194    }
1195
1196    /// Trapezoidal is second-order: halving `dt` should quarter the error.
1197    #[test]
1198    fn trapezoidal_error_roughly_quarters_when_dt_halves() {
1199        let t_final = 2.0;
1200        let error_coarse = rc_final_error(&Kind::Trapezoidal, 0.1, t_final);
1201        let error_fine = rc_final_error(&Kind::Trapezoidal, 0.05, t_final);
1202        let ratio = error_coarse / error_fine;
1203        assert!(
1204            (3.0..=5.5).contains(&ratio),
1205            "trapezoidal error ratio (dt vs dt/2) = {ratio}, expected close to 4.0 (second-order)"
1206        );
1207    }
1208
1209    /// And, directly: at the same coarse `dt`, trapezoidal's error should be substantially
1210    /// smaller than backward Euler's (not asserting a specific factor, since that depends on
1211    /// problem-specific constants — just that it is, robustly, much better).
1212    #[test]
1213    fn trapezoidal_is_more_accurate_than_backward_euler_at_the_same_dt() {
1214        let t_final = 2.0;
1215        let dt = 0.1;
1216        let error_be = rc_final_error(&Kind::BackwardEuler, dt, t_final);
1217        let error_tr = rc_final_error(&Kind::Trapezoidal, dt, t_final);
1218        assert!(
1219            error_tr < error_be / 10.0,
1220            "trapezoidal error {error_tr} should be well under backward-Euler error {error_be} at dt={dt}"
1221        );
1222    }
1223}