Skip to content

Memristor Mapper

Memristive crossbar array mapper with non-linear conductance, device aging simulation, and power/area estimation.

Quick Start

Python
from sc_neurocore.memristor.memristor_mapper import (
    CrossbarArray, ConductanceModel, AgingSimulator, CrossbarEstimator,
)

sc_neurocore.memristor.memristor_mapper

Variability-aware memristor crossbar mapper for stochastic computing.

Takes SC bitstreams + per-device conductance variability models (from real fab data distributions) and emits optimised crossbar-aware SystemVerilog with compensation LUTs or stochastic encoding that absorbs read/write noise at design time.

Supports Monte Carlo variability injection for co-simulation, enabling fab-ready crossbar SC deployment on emerging memristor arrays (Mythic, Weebit, 2D-material ReRAM, phase-change, etc.).

MemristorTechnology

Bases: Enum

Supported memristor fabrication technologies.

Source code in src/sc_neurocore/memristor/memristor_mapper.py
Python
34
35
36
37
38
39
40
41
class MemristorTechnology(Enum):
    """Supported memristor fabrication technologies."""

    RERAM_HFOX = "reram_hfox"  # HfOₓ-based ReRAM (Weebit, TSMC)
    RERAM_2D = "reram_2d"  # 2D-material ReRAM (MoS₂, h-BN)
    PCM = "pcm"  # Phase-change memory (GST-based)
    MYTHIC_AMP = "mythic_amp"  # Mythic analog matrix processor
    GENERIC = "generic"

ConductanceModel dataclass

Per-device conductance variability model.

Models both device-to-device (D2D) fabrication variability and cycle-to-cycle (C2C) read/write noise as Gaussian distributions.

Source code in src/sc_neurocore/memristor/memristor_mapper.py
Python
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
@dataclass
class ConductanceModel:
    """Per-device conductance variability model.

    Models both device-to-device (D2D) fabrication variability and
    cycle-to-cycle (C2C) read/write noise as Gaussian distributions.
    """

    technology: MemristorTechnology
    g_on: float = 0.0
    g_off: float = 0.0
    sigma_g: float = 0.0
    sigma_rw: float = 0.0
    num_levels: int = 0

    def __post_init__(self) -> None:
        """Fill any unset conductance parameters from the technology presets."""
        params = _TECH_PARAMS[self.technology]
        if self.g_on <= 0:
            self.g_on = params["g_on"]
        if self.g_off <= 0:
            self.g_off = params["g_off"]
        if self.sigma_g <= 0:
            self.sigma_g = params["sigma_g"]
        if self.sigma_rw <= 0:
            self.sigma_rw = params["sigma_rw"]
        if self.num_levels <= 0:
            self.num_levels = int(params["num_levels"])

    @property
    def dynamic_range(self) -> float:
        """ON/OFF conductance ratio."""
        return self.g_on / self.g_off if self.g_off > 0 else float("inf")

    @property
    def level_step(self) -> float:
        """Conductance step between adjacent levels."""
        return (self.g_on - self.g_off) / max(1, self.num_levels - 1)

    def target_conductance(self, level: int) -> float:
        """Nominal conductance for a given quantisation level."""
        level = max(0, min(self.num_levels - 1, level))
        return self.g_off + level * self.level_step

    def sample_d2d(self, level: int, rng: np.random.Generator) -> float:
        """Sample actual conductance with device-to-device variability."""
        nominal = self.target_conductance(level)
        return float(rng.normal(nominal, nominal * self.sigma_g))

    def sample_rw(self, conductance: float, rng: np.random.Generator) -> float:
        """Apply read/write noise to a conductance value."""
        return float(rng.normal(conductance, abs(conductance) * self.sigma_rw))

    def drift(self, conductance: float, elapsed_s: float, alpha: float = 0.1) -> float:
        """Model conductance drift over time.

        Uses power-law drift: G(t) = G₀ × (t/t₀)^(-α)
        """
        t0 = 1.0
        if elapsed_s <= t0:
            return conductance
        drifted: float = conductance * (elapsed_s / t0) ** (-alpha)
        return drifted

    def thermal_shift(self, conductance: float, temp_c: float, ref_c: float = 25.0) -> float:
        """Temperature-dependent conductance shift.

        Linear TC model: ΔG/G ≈ tc_ppm × ΔT × 1e-6.
        """
        tc_ppm = 1500.0  # typical for metal-oxide ReRAM
        delta_t = temp_c - ref_c
        return conductance * (1.0 + tc_ppm * delta_t * 1e-6)

dynamic_range property

ON/OFF conductance ratio.

level_step property

Conductance step between adjacent levels.

__post_init__()

Fill any unset conductance parameters from the technology presets.

Source code in src/sc_neurocore/memristor/memristor_mapper.py
Python
111
112
113
114
115
116
117
118
119
120
121
122
123
def __post_init__(self) -> None:
    """Fill any unset conductance parameters from the technology presets."""
    params = _TECH_PARAMS[self.technology]
    if self.g_on <= 0:
        self.g_on = params["g_on"]
    if self.g_off <= 0:
        self.g_off = params["g_off"]
    if self.sigma_g <= 0:
        self.sigma_g = params["sigma_g"]
    if self.sigma_rw <= 0:
        self.sigma_rw = params["sigma_rw"]
    if self.num_levels <= 0:
        self.num_levels = int(params["num_levels"])

target_conductance(level)

Nominal conductance for a given quantisation level.

Source code in src/sc_neurocore/memristor/memristor_mapper.py
Python
135
136
137
138
def target_conductance(self, level: int) -> float:
    """Nominal conductance for a given quantisation level."""
    level = max(0, min(self.num_levels - 1, level))
    return self.g_off + level * self.level_step

sample_d2d(level, rng)

Sample actual conductance with device-to-device variability.

Source code in src/sc_neurocore/memristor/memristor_mapper.py
Python
140
141
142
143
def sample_d2d(self, level: int, rng: np.random.Generator) -> float:
    """Sample actual conductance with device-to-device variability."""
    nominal = self.target_conductance(level)
    return float(rng.normal(nominal, nominal * self.sigma_g))

sample_rw(conductance, rng)

Apply read/write noise to a conductance value.

Source code in src/sc_neurocore/memristor/memristor_mapper.py
Python
145
146
147
def sample_rw(self, conductance: float, rng: np.random.Generator) -> float:
    """Apply read/write noise to a conductance value."""
    return float(rng.normal(conductance, abs(conductance) * self.sigma_rw))

drift(conductance, elapsed_s, alpha=0.1)

Model conductance drift over time.

Uses power-law drift: G(t) = G₀ × (t/t₀)^(-α)

Source code in src/sc_neurocore/memristor/memristor_mapper.py
Python
149
150
151
152
153
154
155
156
157
158
def drift(self, conductance: float, elapsed_s: float, alpha: float = 0.1) -> float:
    """Model conductance drift over time.

    Uses power-law drift: G(t) = G₀ × (t/t₀)^(-α)
    """
    t0 = 1.0
    if elapsed_s <= t0:
        return conductance
    drifted: float = conductance * (elapsed_s / t0) ** (-alpha)
    return drifted

thermal_shift(conductance, temp_c, ref_c=25.0)

Temperature-dependent conductance shift.

Linear TC model: ΔG/G ≈ tc_ppm × ΔT × 1e-6.

Source code in src/sc_neurocore/memristor/memristor_mapper.py
Python
160
161
162
163
164
165
166
167
def thermal_shift(self, conductance: float, temp_c: float, ref_c: float = 25.0) -> float:
    """Temperature-dependent conductance shift.

    Linear TC model: ΔG/G ≈ tc_ppm × ΔT × 1e-6.
    """
    tc_ppm = 1500.0  # typical for metal-oxide ReRAM
    delta_t = temp_c - ref_c
    return conductance * (1.0 + tc_ppm * delta_t * 1e-6)

SneakPathModel

Estimates sneak-path leakage in passive crossbar arrays.

In an M×N passive crossbar, selecting cell (r,c) exposes parallel leakage paths through unselected devices. Worst-case sneak current is proportional to (M+N-2) × G_off.

Source code in src/sc_neurocore/memristor/memristor_mapper.py
Python
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
class SneakPathModel:
    """Estimates sneak-path leakage in passive crossbar arrays.

    In an M×N passive crossbar, selecting cell (r,c) exposes parallel
    leakage paths through unselected devices. Worst-case sneak current
    is proportional to (M+N-2) × G_off.
    """

    @staticmethod
    def worst_case_sneak(rows: int, cols: int, g_off: float, v_read: float = 0.2) -> float:
        """Worst-case sneak current (A) through unselected paths."""
        n_paths = (rows - 1) + (cols - 1)
        return n_paths * g_off * v_read

    @staticmethod
    def signal_to_sneak_ratio(g_on: float, g_off: float, rows: int, cols: int) -> float:
        """Ratio of desired signal current to sneak current."""
        sneak = SneakPathModel.worst_case_sneak(rows, cols, g_off)
        if sneak <= 0:
            return float("inf")
        return (g_on * 0.2) / sneak  # assume 0.2V read

worst_case_sneak(rows, cols, g_off, v_read=0.2) staticmethod

Worst-case sneak current (A) through unselected paths.

Source code in src/sc_neurocore/memristor/memristor_mapper.py
Python
181
182
183
184
185
@staticmethod
def worst_case_sneak(rows: int, cols: int, g_off: float, v_read: float = 0.2) -> float:
    """Worst-case sneak current (A) through unselected paths."""
    n_paths = (rows - 1) + (cols - 1)
    return n_paths * g_off * v_read

signal_to_sneak_ratio(g_on, g_off, rows, cols) staticmethod

Ratio of desired signal current to sneak current.

Source code in src/sc_neurocore/memristor/memristor_mapper.py
Python
187
188
189
190
191
192
193
@staticmethod
def signal_to_sneak_ratio(g_on: float, g_off: float, rows: int, cols: int) -> float:
    """Ratio of desired signal current to sneak current."""
    sneak = SneakPathModel.worst_case_sneak(rows, cols, g_off)
    if sneak <= 0:
        return float("inf")
    return (g_on * 0.2) / sneak  # assume 0.2V read

IRDropModel dataclass

Models interconnect wire resistance and voltage drop.

Source code in src/sc_neurocore/memristor/memristor_mapper.py
Python
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
@dataclass
class IRDropModel:
    """Models interconnect wire resistance and voltage drop."""

    r_wire_per_cell: float = 2.5  # Ω per cell segment

    def voltage_drop(self, row: int, col: int) -> float:
        """Accumulated IR drop at cell (row, col) from corner."""
        return self.r_wire_per_cell * (row + col) * 1e-3  # mV → V approx

    def effective_conductance(
        self, g_nominal: float, row: int, col: int, v_read: float = 0.2
    ) -> float:
        """Conductance seen at read amplifier after IR drop."""
        v_drop = self.voltage_drop(row, col)
        v_eff = max(0.0, v_read - v_drop)
        return g_nominal * (v_eff / v_read) if v_read > 0 else g_nominal

voltage_drop(row, col)

Accumulated IR drop at cell (row, col) from corner.

Source code in src/sc_neurocore/memristor/memristor_mapper.py
Python
205
206
207
def voltage_drop(self, row: int, col: int) -> float:
    """Accumulated IR drop at cell (row, col) from corner."""
    return self.r_wire_per_cell * (row + col) * 1e-3  # mV → V approx

effective_conductance(g_nominal, row, col, v_read=0.2)

Conductance seen at read amplifier after IR drop.

Source code in src/sc_neurocore/memristor/memristor_mapper.py
Python
209
210
211
212
213
214
215
def effective_conductance(
    self, g_nominal: float, row: int, col: int, v_read: float = 0.2
) -> float:
    """Conductance seen at read amplifier after IR drop."""
    v_drop = self.voltage_drop(row, col)
    v_eff = max(0.0, v_read - v_drop)
    return g_nominal * (v_eff / v_read) if v_read > 0 else g_nominal

StuckFaultMap dataclass

Map of stuck-at ON/OFF devices in a crossbar.

Source code in src/sc_neurocore/memristor/memristor_mapper.py
Python
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
@dataclass
class StuckFaultMap:
    """Map of stuck-at ON/OFF devices in a crossbar."""

    rows: int
    cols: int
    stuck_on: List[Tuple[int, int]] = field(default_factory=list)
    stuck_off: List[Tuple[int, int]] = field(default_factory=list)

    @classmethod
    def generate(
        cls,
        rows: int,
        cols: int,
        fault_rate: float = 0.001,
        seed: int = 42,
    ) -> StuckFaultMap:
        """Generate random stuck faults at given rate."""
        rng = np.random.default_rng(seed)
        total = rows * cols
        n_faults = int(total * fault_rate)
        fault_idx = rng.choice(total, size=min(n_faults, total), replace=False)
        on_faults = []
        off_faults = []
        for idx in fault_idx:
            r, c = divmod(int(idx), cols)
            if rng.random() < 0.5:
                on_faults.append((r, c))
            else:
                off_faults.append((r, c))
        return cls(rows, cols, on_faults, off_faults)

    def is_stuck(self, row: int, col: int) -> Optional[str]:
        """Return 'on', 'off', or None."""
        if (row, col) in self.stuck_on:
            return "on"
        if (row, col) in self.stuck_off:
            return "off"
        return None

    @property
    def num_faults(self) -> int:
        """Return the total count of stuck-on and stuck-off devices."""
        return len(self.stuck_on) + len(self.stuck_off)

    @property
    def fault_rate(self) -> float:
        """Return the fraction of crossbar cells that are faulty."""
        total = self.rows * self.cols
        return self.num_faults / total if total > 0 else 0.0

num_faults property

Return the total count of stuck-on and stuck-off devices.

fault_rate property

Return the fraction of crossbar cells that are faulty.

generate(rows, cols, fault_rate=0.001, seed=42) classmethod

Generate random stuck faults at given rate.

Source code in src/sc_neurocore/memristor/memristor_mapper.py
Python
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
@classmethod
def generate(
    cls,
    rows: int,
    cols: int,
    fault_rate: float = 0.001,
    seed: int = 42,
) -> StuckFaultMap:
    """Generate random stuck faults at given rate."""
    rng = np.random.default_rng(seed)
    total = rows * cols
    n_faults = int(total * fault_rate)
    fault_idx = rng.choice(total, size=min(n_faults, total), replace=False)
    on_faults = []
    off_faults = []
    for idx in fault_idx:
        r, c = divmod(int(idx), cols)
        if rng.random() < 0.5:
            on_faults.append((r, c))
        else:
            off_faults.append((r, c))
    return cls(rows, cols, on_faults, off_faults)

is_stuck(row, col)

Return 'on', 'off', or None.

Source code in src/sc_neurocore/memristor/memristor_mapper.py
Python
253
254
255
256
257
258
259
def is_stuck(self, row: int, col: int) -> Optional[str]:
    """Return 'on', 'off', or None."""
    if (row, col) in self.stuck_on:
        return "on"
    if (row, col) in self.stuck_off:
        return "off"
    return None

AgingReport dataclass

Results of aging simulation.

Source code in src/sc_neurocore/memristor/memristor_mapper.py
Python
276
277
278
279
280
281
282
283
@dataclass
class AgingReport:
    """Results of aging simulation."""

    elapsed_s: float
    mean_drift_fraction: float
    max_drift_fraction: float
    levels_shifted: int

AgingSimulator

Simulates conductance drift over device lifetime.

Source code in src/sc_neurocore/memristor/memristor_mapper.py
Python
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
class AgingSimulator:
    """Simulates conductance drift over device lifetime."""

    def __init__(self, model: ConductanceModel, alpha: float = 0.1) -> None:
        self.model = model
        self.alpha = alpha

    def simulate(
        self, conductances: np.ndarray[Any, Any], elapsed_s: float
    ) -> Tuple[np.ndarray[Any, Any], AgingReport]:
        """Apply drift to all conductances, return (drifted, report)."""
        drifted = np.zeros_like(conductances)
        for idx in np.ndindex(conductances.shape):
            drifted[idx] = self.model.drift(float(conductances[idx]), elapsed_s, self.alpha)
        abs_drift = np.abs(drifted - conductances)
        rel_drift = abs_drift / np.maximum(np.abs(conductances), 1e-15)
        step = self.model.level_step
        levels_shifted = int(np.sum(abs_drift > step)) if step > 0 else 0
        return drifted, AgingReport(
            elapsed_s=elapsed_s,
            mean_drift_fraction=float(np.mean(rel_drift)),
            max_drift_fraction=float(np.max(rel_drift)),
            levels_shifted=levels_shifted,
        )

simulate(conductances, elapsed_s)

Apply drift to all conductances, return (drifted, report).

Source code in src/sc_neurocore/memristor/memristor_mapper.py
Python
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
def simulate(
    self, conductances: np.ndarray[Any, Any], elapsed_s: float
) -> Tuple[np.ndarray[Any, Any], AgingReport]:
    """Apply drift to all conductances, return (drifted, report)."""
    drifted = np.zeros_like(conductances)
    for idx in np.ndindex(conductances.shape):
        drifted[idx] = self.model.drift(float(conductances[idx]), elapsed_s, self.alpha)
    abs_drift = np.abs(drifted - conductances)
    rel_drift = abs_drift / np.maximum(np.abs(conductances), 1e-15)
    step = self.model.level_step
    levels_shifted = int(np.sum(abs_drift > step)) if step > 0 else 0
    return drifted, AgingReport(
        elapsed_s=elapsed_s,
        mean_drift_fraction=float(np.mean(rel_drift)),
        max_drift_fraction=float(np.max(rel_drift)),
        levels_shifted=levels_shifted,
    )

SCAbsorbEncoder

Adjusts SC encoding thresholds to absorb known device variability.

Instead of compensating post-silicon, pre-distorts the bitstream encoding thresholds so that the effective computation matches ideal.

Source code in src/sc_neurocore/memristor/memristor_mapper.py
Python
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
class SCAbsorbEncoder:
    """Adjusts SC encoding thresholds to absorb known device variability.

    Instead of compensating post-silicon, pre-distorts the bitstream
    encoding thresholds so that the effective computation matches ideal.
    """

    @staticmethod
    def compute_adjusted_thresholds(
        ideal_weights: np.ndarray[Any, Any],
        actual_conductances: np.ndarray[Any, Any],
        model: ConductanceModel,
        q_bits: int = 8,
    ) -> np.ndarray[Any, Any]:
        """Return Q8.8 adjusted thresholds that absorb device error."""
        levels_ideal = np.clip(
            np.round(ideal_weights * (model.num_levels - 1)).astype(int),
            0,
            model.num_levels - 1,
        )
        g_ideal = np.array(
            [
                [
                    model.target_conductance(int(levels_ideal[i, j]))
                    for j in range(ideal_weights.shape[1])
                ]
                for i in range(ideal_weights.shape[0])
            ]
        )
        ratio = np.where(
            np.abs(actual_conductances) > 1e-15,
            g_ideal / actual_conductances,
            1.0,
        )
        scale = 1 << q_bits
        return np.clip(np.round(ratio * scale).astype(np.int32), 0, 65535)

compute_adjusted_thresholds(ideal_weights, actual_conductances, model, q_bits=8) staticmethod

Return Q8.8 adjusted thresholds that absorb device error.

Source code in src/sc_neurocore/memristor/memristor_mapper.py
Python
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
@staticmethod
def compute_adjusted_thresholds(
    ideal_weights: np.ndarray[Any, Any],
    actual_conductances: np.ndarray[Any, Any],
    model: ConductanceModel,
    q_bits: int = 8,
) -> np.ndarray[Any, Any]:
    """Return Q8.8 adjusted thresholds that absorb device error."""
    levels_ideal = np.clip(
        np.round(ideal_weights * (model.num_levels - 1)).astype(int),
        0,
        model.num_levels - 1,
    )
    g_ideal = np.array(
        [
            [
                model.target_conductance(int(levels_ideal[i, j]))
                for j in range(ideal_weights.shape[1])
            ]
            for i in range(ideal_weights.shape[0])
        ]
    )
    ratio = np.where(
        np.abs(actual_conductances) > 1e-15,
        g_ideal / actual_conductances,
        1.0,
    )
    scale = 1 << q_bits
    return np.clip(np.round(ratio * scale).astype(np.int32), 0, 65535)

WriteVerifyResult dataclass

Outcome of iterative write-verify programming.

Source code in src/sc_neurocore/memristor/memristor_mapper.py
Python
356
357
358
359
360
361
362
363
364
@dataclass
class WriteVerifyResult:
    """Outcome of iterative write-verify programming."""

    target_level: int
    target_g: float
    achieved_g: float
    iterations: int
    converged: bool

WriteVerifyProtocol

Iterative program-verify loop for memristor cells.

Source code in src/sc_neurocore/memristor/memristor_mapper.py
Python
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
class WriteVerifyProtocol:
    """Iterative program-verify loop for memristor cells."""

    def __init__(
        self,
        model: ConductanceModel,
        max_iterations: int = 10,
        tolerance: float = 0.02,
        seed: int = 42,
    ) -> None:
        self.model = model
        self.max_iter = max_iterations
        self.tolerance = tolerance
        self.rng = np.random.default_rng(seed)

    def program_cell(self, target_level: int) -> WriteVerifyResult:
        """Program a single cell to target level with verify."""
        target_g = self.model.target_conductance(target_level)
        g_current = self.model.sample_d2d(target_level, self.rng)

        for i in range(self.max_iter):
            err = abs(g_current - target_g) / max(abs(target_g), 1e-15)
            if err <= self.tolerance:
                return WriteVerifyResult(target_level, target_g, g_current, i + 1, True)
            correction = (target_g - g_current) * 0.5
            g_current += correction
            g_current = self.model.sample_rw(g_current, self.rng)

        return WriteVerifyResult(target_level, target_g, g_current, self.max_iter, False)

program_cell(target_level)

Program a single cell to target level with verify.

Source code in src/sc_neurocore/memristor/memristor_mapper.py
Python
382
383
384
385
386
387
388
389
390
391
392
393
394
395
def program_cell(self, target_level: int) -> WriteVerifyResult:
    """Program a single cell to target level with verify."""
    target_g = self.model.target_conductance(target_level)
    g_current = self.model.sample_d2d(target_level, self.rng)

    for i in range(self.max_iter):
        err = abs(g_current - target_g) / max(abs(target_g), 1e-15)
        if err <= self.tolerance:
            return WriteVerifyResult(target_level, target_g, g_current, i + 1, True)
        correction = (target_g - g_current) * 0.5
        g_current += correction
        g_current = self.model.sample_rw(g_current, self.rng)

    return WriteVerifyResult(target_level, target_g, g_current, self.max_iter, False)

CrossbarPowerEstimate dataclass

Power and performance estimate for a crossbar tile.

Source code in src/sc_neurocore/memristor/memristor_mapper.py
Python
401
402
403
404
405
406
407
408
409
410
411
@dataclass
class CrossbarPowerEstimate:
    """Power and performance estimate for a crossbar tile."""

    rows: int
    cols: int
    read_power_uw: float
    write_power_uw: float
    read_latency_ns: float
    write_latency_ns: float
    area_um2: float

CrossbarEstimator

Estimates power, latency, and area for crossbar arrays.

Source code in src/sc_neurocore/memristor/memristor_mapper.py
Python
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
class CrossbarEstimator:
    """Estimates power, latency, and area for crossbar arrays."""

    TECH_POWER = {
        MemristorTechnology.RERAM_HFOX: {
            "read_pw": 0.1,
            "write_pw": 10.0,
            "read_ns": 10,
            "write_ns": 100,
            "cell_um2": 0.04,
        },
        MemristorTechnology.RERAM_2D: {
            "read_pw": 0.05,
            "write_pw": 5.0,
            "read_ns": 5,
            "write_ns": 50,
            "cell_um2": 0.01,
        },
        MemristorTechnology.PCM: {
            "read_pw": 0.2,
            "write_pw": 50.0,
            "read_ns": 20,
            "write_ns": 200,
            "cell_um2": 0.06,
        },
        MemristorTechnology.MYTHIC_AMP: {
            "read_pw": 0.08,
            "write_pw": 8.0,
            "read_ns": 8,
            "write_ns": 80,
            "cell_um2": 0.03,
        },
        MemristorTechnology.GENERIC: {
            "read_pw": 0.15,
            "write_pw": 15.0,
            "read_ns": 15,
            "write_ns": 150,
            "cell_um2": 0.05,
        },
    }

    @classmethod
    def estimate(cls, crossbar: CrossbarArray) -> CrossbarPowerEstimate:
        """Estimate read/write power, latency and area for a crossbar array."""
        p = cls.TECH_POWER[crossbar.technology]
        n = crossbar.num_devices
        return CrossbarPowerEstimate(
            rows=crossbar.rows,
            cols=crossbar.cols,
            read_power_uw=p["read_pw"] * n,
            write_power_uw=p["write_pw"] * n,
            read_latency_ns=p["read_ns"],
            write_latency_ns=p["write_ns"],
            area_um2=p["cell_um2"] * n,
        )

estimate(crossbar) classmethod

Estimate read/write power, latency and area for a crossbar array.

Source code in src/sc_neurocore/memristor/memristor_mapper.py
Python
455
456
457
458
459
460
461
462
463
464
465
466
467
468
@classmethod
def estimate(cls, crossbar: CrossbarArray) -> CrossbarPowerEstimate:
    """Estimate read/write power, latency and area for a crossbar array."""
    p = cls.TECH_POWER[crossbar.technology]
    n = crossbar.num_devices
    return CrossbarPowerEstimate(
        rows=crossbar.rows,
        cols=crossbar.cols,
        read_power_uw=p["read_pw"] * n,
        write_power_uw=p["write_pw"] * n,
        read_latency_ns=p["read_ns"],
        write_latency_ns=p["write_ns"],
        area_um2=p["cell_um2"] * n,
    )

CrossbarTopology

Bases: Enum

Physical wiring topology of a memristor crossbar array.

Source code in src/sc_neurocore/memristor/memristor_mapper.py
Python
474
475
476
477
478
479
480
class CrossbarTopology(Enum):
    """Physical wiring topology of a memristor crossbar array."""

    STANDARD = "standard"  # M×N passive crossbar
    ONE_T_ONE_R = "1t1r"  # 1-transistor-1-resistor
    ONE_S_ONE_R = "1s1r"  # 1-selector-1-resistor
    DIFFERENTIAL = "differential"  # Differential pair crossbar

CrossbarArray dataclass

Physical crossbar array specification.

Source code in src/sc_neurocore/memristor/memristor_mapper.py
Python
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
@dataclass
class CrossbarArray:
    """Physical crossbar array specification."""

    rows: int
    cols: int
    topology: CrossbarTopology = CrossbarTopology.STANDARD
    technology: MemristorTechnology = MemristorTechnology.GENERIC

    @property
    def num_devices(self) -> int:
        """Return the device count, doubling for differential topologies."""
        if self.topology == CrossbarTopology.DIFFERENTIAL:
            return self.rows * self.cols * 2
        return self.rows * self.cols

    @property
    def conductance_model(self) -> ConductanceModel:
        """Return the conductance model for this array's technology."""
        return ConductanceModel(technology=self.technology)

num_devices property

Return the device count, doubling for differential topologies.

conductance_model property

Return the conductance model for this array's technology.

VariabilityInjector

Injects fab-realistic conductance variability into weight matrices.

Source code in src/sc_neurocore/memristor/memristor_mapper.py
Python
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
class VariabilityInjector:
    """Injects fab-realistic conductance variability into weight matrices."""

    def __init__(self, model: ConductanceModel, seed: int = 42) -> None:
        self.model = model
        self.rng = np.random.default_rng(seed)

    def quantize_weights(self, weights: np.ndarray[Any, Any]) -> np.ndarray[Any, Any]:
        """Map floating-point weights [0, 1] to conductance levels."""
        levels = np.clip(
            np.round(weights * (self.model.num_levels - 1)).astype(int),
            0,
            self.model.num_levels - 1,
        )
        return levels

    def inject_d2d(self, levels: np.ndarray[Any, Any]) -> np.ndarray[Any, Any]:
        """Apply device-to-device variability to quantised levels."""
        result = np.zeros_like(levels, dtype=np.float64)
        for idx in np.ndindex(levels.shape):
            result[idx] = self.model.sample_d2d(int(levels[idx]), self.rng)
        return result

    def inject_rw(self, conductances: np.ndarray[Any, Any]) -> np.ndarray[Any, Any]:
        """Apply read/write noise to conductance values."""
        result = np.zeros_like(conductances, dtype=np.float64)
        for idx in np.ndindex(conductances.shape):
            result[idx] = self.model.sample_rw(float(conductances[idx]), self.rng)
        return result

    def inject_full(
        self, weights: np.ndarray[Any, Any]
    ) -> Tuple[np.ndarray[Any, Any], np.ndarray[Any, Any]]:
        """Full pipeline: quantise → D2D → R/W noise. Returns (levels, conductances)."""
        levels = self.quantize_weights(weights)
        g_d2d = self.inject_d2d(levels)
        g_final = self.inject_rw(g_d2d)
        return levels, g_final

    def compute_error(
        self, weights: np.ndarray[Any, Any], conductances: np.ndarray[Any, Any]
    ) -> Dict[str, float]:
        """Compute variability-induced error statistics."""
        levels = self.quantize_weights(weights)
        ideal = np.array(
            [[self.model.target_conductance(int(levels[idx])) for idx in np.ndindex(levels.shape)]]
        ).reshape(levels.shape)
        abs_err = np.abs(conductances - ideal)
        rel_err = abs_err / np.maximum(np.abs(ideal), 1e-15)
        return {
            "mae": float(np.mean(abs_err)),
            "max_abs_err": float(np.max(abs_err)),
            "mean_rel_err": float(np.mean(rel_err)),
            "max_rel_err": float(np.max(rel_err)),
        }

quantize_weights(weights)

Map floating-point weights [0, 1] to conductance levels.

Source code in src/sc_neurocore/memristor/memristor_mapper.py
Python
515
516
517
518
519
520
521
522
def quantize_weights(self, weights: np.ndarray[Any, Any]) -> np.ndarray[Any, Any]:
    """Map floating-point weights [0, 1] to conductance levels."""
    levels = np.clip(
        np.round(weights * (self.model.num_levels - 1)).astype(int),
        0,
        self.model.num_levels - 1,
    )
    return levels

inject_d2d(levels)

Apply device-to-device variability to quantised levels.

Source code in src/sc_neurocore/memristor/memristor_mapper.py
Python
524
525
526
527
528
529
def inject_d2d(self, levels: np.ndarray[Any, Any]) -> np.ndarray[Any, Any]:
    """Apply device-to-device variability to quantised levels."""
    result = np.zeros_like(levels, dtype=np.float64)
    for idx in np.ndindex(levels.shape):
        result[idx] = self.model.sample_d2d(int(levels[idx]), self.rng)
    return result

inject_rw(conductances)

Apply read/write noise to conductance values.

Source code in src/sc_neurocore/memristor/memristor_mapper.py
Python
531
532
533
534
535
536
def inject_rw(self, conductances: np.ndarray[Any, Any]) -> np.ndarray[Any, Any]:
    """Apply read/write noise to conductance values."""
    result = np.zeros_like(conductances, dtype=np.float64)
    for idx in np.ndindex(conductances.shape):
        result[idx] = self.model.sample_rw(float(conductances[idx]), self.rng)
    return result

inject_full(weights)

Full pipeline: quantise → D2D → R/W noise. Returns (levels, conductances).

Source code in src/sc_neurocore/memristor/memristor_mapper.py
Python
538
539
540
541
542
543
544
545
def inject_full(
    self, weights: np.ndarray[Any, Any]
) -> Tuple[np.ndarray[Any, Any], np.ndarray[Any, Any]]:
    """Full pipeline: quantise → D2D → R/W noise. Returns (levels, conductances)."""
    levels = self.quantize_weights(weights)
    g_d2d = self.inject_d2d(levels)
    g_final = self.inject_rw(g_d2d)
    return levels, g_final

compute_error(weights, conductances)

Compute variability-induced error statistics.

Source code in src/sc_neurocore/memristor/memristor_mapper.py
Python
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
def compute_error(
    self, weights: np.ndarray[Any, Any], conductances: np.ndarray[Any, Any]
) -> Dict[str, float]:
    """Compute variability-induced error statistics."""
    levels = self.quantize_weights(weights)
    ideal = np.array(
        [[self.model.target_conductance(int(levels[idx])) for idx in np.ndindex(levels.shape)]]
    ).reshape(levels.shape)
    abs_err = np.abs(conductances - ideal)
    rel_err = abs_err / np.maximum(np.abs(ideal), 1e-15)
    return {
        "mae": float(np.mean(abs_err)),
        "max_abs_err": float(np.max(abs_err)),
        "mean_rel_err": float(np.mean(rel_err)),
        "max_rel_err": float(np.max(rel_err)),
    }

CompensationStrategy

Bases: Enum

Strategy for compensating memristor conductance non-idealities.

Source code in src/sc_neurocore/memristor/memristor_mapper.py
Python
568
569
570
571
572
573
574
class CompensationStrategy(Enum):
    """Strategy for compensating memristor conductance non-idealities."""

    NONE = "none"
    LUT = "lut"  # per-device compensation lookup table
    STOCHASTIC_ENCODING = "sc_absorb"  # adjust SC encoding thresholds
    HYBRID = "hybrid"

CompensationLUT dataclass

Per-device compensation lookup table.

Maps nominal level → compensated threshold to absorb known D2D variability at design time (program-verify or digital pre-distortion).

Source code in src/sc_neurocore/memristor/memristor_mapper.py
Python
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
@dataclass
class CompensationLUT:
    """Per-device compensation lookup table.

    Maps nominal level → compensated threshold to absorb known D2D
    variability at design time (program-verify or digital pre-distortion).
    """

    device_id: Tuple[int, int]
    nominal_levels: np.ndarray[Any, Any]  # shape (num_levels,)
    compensated_thresholds: np.ndarray[Any, Any]  # Q8.8 fixed-point thresholds

    @classmethod
    def build(
        cls,
        device_id: Tuple[int, int],
        model: ConductanceModel,
        measured_g: Optional[np.ndarray[Any, Any]] = None,
    ) -> CompensationLUT:
        """Build compensation LUT from measured or modelled conductances.

        If measured_g is None, generates from the model's nominal values.
        """
        nominal = np.array([model.target_conductance(i) for i in range(model.num_levels)])
        if measured_g is not None and len(measured_g) == model.num_levels:
            ratio = nominal / np.maximum(measured_g, 1e-15)
        else:
            ratio = np.ones(model.num_levels)
        # Q8.8 fixed-point: multiply by 256, round to int
        thresholds = np.clip(np.round(ratio * 256).astype(np.int32), 0, 65535)
        return cls(
            device_id=device_id,
            nominal_levels=np.arange(model.num_levels),
            compensated_thresholds=thresholds,
        )

    @property
    def max_compensation(self) -> float:
        """Maximum compensation ratio (deviation from 1.0)."""
        ratios = self.compensated_thresholds.astype(np.float64) / 256.0
        return float(np.max(np.abs(ratios - 1.0)))

max_compensation property

Maximum compensation ratio (deviation from 1.0).

build(device_id, model, measured_g=None) classmethod

Build compensation LUT from measured or modelled conductances.

If measured_g is None, generates from the model's nominal values.

Source code in src/sc_neurocore/memristor/memristor_mapper.py
Python
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
@classmethod
def build(
    cls,
    device_id: Tuple[int, int],
    model: ConductanceModel,
    measured_g: Optional[np.ndarray[Any, Any]] = None,
) -> CompensationLUT:
    """Build compensation LUT from measured or modelled conductances.

    If measured_g is None, generates from the model's nominal values.
    """
    nominal = np.array([model.target_conductance(i) for i in range(model.num_levels)])
    if measured_g is not None and len(measured_g) == model.num_levels:
        ratio = nominal / np.maximum(measured_g, 1e-15)
    else:
        ratio = np.ones(model.num_levels)
    # Q8.8 fixed-point: multiply by 256, round to int
    thresholds = np.clip(np.round(ratio * 256).astype(np.int32), 0, 65535)
    return cls(
        device_id=device_id,
        nominal_levels=np.arange(model.num_levels),
        compensated_thresholds=thresholds,
    )

CrossbarMapping dataclass

Mapping of a weight matrix to a crossbar array.

Source code in src/sc_neurocore/memristor/memristor_mapper.py
Python
623
624
625
626
627
628
629
630
631
@dataclass
class CrossbarMapping:
    """Mapping of a weight matrix to a crossbar array."""

    crossbar: CrossbarArray
    weight_levels: np.ndarray[Any, Any]  # shape (rows, cols), quantised levels
    conductances: np.ndarray[Any, Any]  # shape (rows, cols), actual conductances
    compensation_luts: List[CompensationLUT] = field(default_factory=list)
    error_stats: Dict[str, float] = field(default_factory=dict)

MappingResult dataclass

Full result of a memristor mapping pass.

Source code in src/sc_neurocore/memristor/memristor_mapper.py
Python
634
635
636
637
638
639
640
641
642
643
@dataclass
class MappingResult:
    """Full result of a memristor mapping pass."""

    mappings: List[CrossbarMapping]
    total_devices: int
    total_crossbars: int
    mean_rel_error: float
    max_rel_error: float
    compensation_strategy: CompensationStrategy

MemristorMapper

Maps SC network weight matrices to physical crossbar arrays.

Source code in src/sc_neurocore/memristor/memristor_mapper.py
Python
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
class MemristorMapper:
    """Maps SC network weight matrices to physical crossbar arrays."""

    def __init__(
        self,
        technology: MemristorTechnology = MemristorTechnology.GENERIC,
        topology: CrossbarTopology = CrossbarTopology.STANDARD,
        max_crossbar_size: int = 256,
        compensation: CompensationStrategy = CompensationStrategy.LUT,
        seed: int = 42,
    ) -> None:
        self.technology = technology
        self.topology = topology
        self.max_size = max_crossbar_size
        self.compensation = compensation
        self.model = ConductanceModel(technology=technology)
        self.injector = VariabilityInjector(self.model, seed)

    def map_weights(self, weights: np.ndarray[Any, Any]) -> MappingResult:
        """Map a weight matrix (or list of matrices) to crossbar arrays."""
        if weights.ndim == 1:
            weights = weights.reshape(1, -1)

        rows, cols = weights.shape
        tile_rows = min(rows, self.max_size)
        tile_cols = min(cols, self.max_size)

        mappings = []
        for r0 in range(0, rows, tile_rows):
            for c0 in range(0, cols, tile_cols):
                tile = weights[r0 : r0 + tile_rows, c0 : c0 + tile_cols]
                tr, tc = tile.shape
                xbar = CrossbarArray(tr, tc, self.topology, self.technology)
                levels, conductances = self.injector.inject_full(tile)
                err = self.injector.compute_error(tile, conductances)

                luts = []
                if self.compensation in (CompensationStrategy.LUT, CompensationStrategy.HYBRID):
                    for i in range(tr):
                        for j in range(tc):
                            measured = np.array(
                                [
                                    self.model.sample_d2d(lv, self.injector.rng)
                                    for lv in range(self.model.num_levels)
                                ]
                            )
                            lut = CompensationLUT.build((r0 + i, c0 + j), self.model, measured)
                            luts.append(lut)

                mappings.append(
                    CrossbarMapping(
                        crossbar=xbar,
                        weight_levels=levels,
                        conductances=conductances,
                        compensation_luts=luts,
                        error_stats=err,
                    )
                )

        total_dev = sum(m.crossbar.num_devices for m in mappings)
        rel_errs = [m.error_stats.get("mean_rel_err", 0) for m in mappings]
        max_errs = [m.error_stats.get("max_rel_err", 0) for m in mappings]

        return MappingResult(
            mappings=mappings,
            total_devices=total_dev,
            total_crossbars=len(mappings),
            mean_rel_error=float(np.mean(rel_errs)) if rel_errs else 0.0,
            max_rel_error=float(np.max(max_errs)) if max_errs else 0.0,
            compensation_strategy=self.compensation,
        )

map_weights(weights)

Map a weight matrix (or list of matrices) to crossbar arrays.

Source code in src/sc_neurocore/memristor/memristor_mapper.py
Python
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
def map_weights(self, weights: np.ndarray[Any, Any]) -> MappingResult:
    """Map a weight matrix (or list of matrices) to crossbar arrays."""
    if weights.ndim == 1:
        weights = weights.reshape(1, -1)

    rows, cols = weights.shape
    tile_rows = min(rows, self.max_size)
    tile_cols = min(cols, self.max_size)

    mappings = []
    for r0 in range(0, rows, tile_rows):
        for c0 in range(0, cols, tile_cols):
            tile = weights[r0 : r0 + tile_rows, c0 : c0 + tile_cols]
            tr, tc = tile.shape
            xbar = CrossbarArray(tr, tc, self.topology, self.technology)
            levels, conductances = self.injector.inject_full(tile)
            err = self.injector.compute_error(tile, conductances)

            luts = []
            if self.compensation in (CompensationStrategy.LUT, CompensationStrategy.HYBRID):
                for i in range(tr):
                    for j in range(tc):
                        measured = np.array(
                            [
                                self.model.sample_d2d(lv, self.injector.rng)
                                for lv in range(self.model.num_levels)
                            ]
                        )
                        lut = CompensationLUT.build((r0 + i, c0 + j), self.model, measured)
                        luts.append(lut)

            mappings.append(
                CrossbarMapping(
                    crossbar=xbar,
                    weight_levels=levels,
                    conductances=conductances,
                    compensation_luts=luts,
                    error_stats=err,
                )
            )

    total_dev = sum(m.crossbar.num_devices for m in mappings)
    rel_errs = [m.error_stats.get("mean_rel_err", 0) for m in mappings]
    max_errs = [m.error_stats.get("max_rel_err", 0) for m in mappings]

    return MappingResult(
        mappings=mappings,
        total_devices=total_dev,
        total_crossbars=len(mappings),
        mean_rel_error=float(np.mean(rel_errs)) if rel_errs else 0.0,
        max_rel_error=float(np.max(max_errs)) if max_errs else 0.0,
        compensation_strategy=self.compensation,
    )

MonteCarloReport dataclass

Results from Monte Carlo variability co-simulation.

Source code in src/sc_neurocore/memristor/memristor_mapper.py
Python
722
723
724
725
726
727
728
729
730
731
732
@dataclass
class MonteCarloReport:
    """Results from Monte Carlo variability co-simulation."""

    num_trials: int
    mean_output_error: float
    std_output_error: float
    max_output_error: float
    yield_fraction: float  # fraction of trials within tolerance
    output_distribution: np.ndarray[Any, Any]
    error_histogram: np.ndarray[Any, Any]

MonteCarloSimulator

Monte Carlo co-simulation with variability injection.

Runs N trials of a crossbar multiply-accumulate operation with independently sampled D2D + R/W variability to estimate output error distributions and yield.

Source code in src/sc_neurocore/memristor/memristor_mapper.py
Python
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
class MonteCarloSimulator:
    """Monte Carlo co-simulation with variability injection.

    Runs N trials of a crossbar multiply-accumulate operation with
    independently sampled D2D + R/W variability to estimate output
    error distributions and yield.
    """

    def __init__(
        self,
        model: ConductanceModel,
        num_trials: int = 1000,
        tolerance: float = 0.05,
        seed: int = 42,
    ) -> None:
        self.model = model
        self.num_trials = num_trials
        self.tolerance = tolerance
        self.rng = np.random.default_rng(seed)

    def simulate_mac(
        self,
        weights: np.ndarray[Any, Any],
        inputs: np.ndarray[Any, Any],
    ) -> MonteCarloReport:
        """Simulate multiply-accumulate with variability.

        Parameters
        ----------
        weights : ndarray, shape (M, N)
            Ideal weight matrix in [0, 1].
        inputs : ndarray, shape (N,)
            Input vector in [0, 1].
        """
        ideal_out = weights @ inputs

        outputs = np.zeros((self.num_trials, len(ideal_out)))
        for trial in range(self.num_trials):
            injector = VariabilityInjector(self.model, seed=int(self.rng.integers(0, 2**31)))
            levels, g_actual = injector.inject_full(weights)

            g_ideal = np.array(
                [
                    [
                        self.model.target_conductance(int(levels[i, j]))
                        for j in range(weights.shape[1])
                    ]
                    for i in range(weights.shape[0])
                ]
            )
            scale = np.where(np.abs(g_ideal) > 1e-15, g_actual / g_ideal, 1.0)
            effective_weights = weights * scale
            outputs[trial] = effective_weights @ inputs

        errors = np.abs(outputs - ideal_out[np.newaxis, :])
        mean_err = float(np.mean(errors))
        rel_errors = errors / np.maximum(np.abs(ideal_out[np.newaxis, :]), 1e-15)
        within_tol = np.all(rel_errors < self.tolerance, axis=1)
        yield_frac = float(np.mean(within_tol))

        err_flat = errors.flatten()
        hist, _ = np.histogram(err_flat, bins=50)

        return MonteCarloReport(
            num_trials=self.num_trials,
            mean_output_error=mean_err,
            std_output_error=float(np.std(errors)),
            max_output_error=float(np.max(errors)),
            yield_fraction=yield_frac,
            output_distribution=np.mean(outputs, axis=0),
            error_histogram=hist,
        )

simulate_mac(weights, inputs)

Simulate multiply-accumulate with variability.

Parameters

weights : ndarray, shape (M, N) Ideal weight matrix in [0, 1]. inputs : ndarray, shape (N,) Input vector in [0, 1].

Source code in src/sc_neurocore/memristor/memristor_mapper.py
Python
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
def simulate_mac(
    self,
    weights: np.ndarray[Any, Any],
    inputs: np.ndarray[Any, Any],
) -> MonteCarloReport:
    """Simulate multiply-accumulate with variability.

    Parameters
    ----------
    weights : ndarray, shape (M, N)
        Ideal weight matrix in [0, 1].
    inputs : ndarray, shape (N,)
        Input vector in [0, 1].
    """
    ideal_out = weights @ inputs

    outputs = np.zeros((self.num_trials, len(ideal_out)))
    for trial in range(self.num_trials):
        injector = VariabilityInjector(self.model, seed=int(self.rng.integers(0, 2**31)))
        levels, g_actual = injector.inject_full(weights)

        g_ideal = np.array(
            [
                [
                    self.model.target_conductance(int(levels[i, j]))
                    for j in range(weights.shape[1])
                ]
                for i in range(weights.shape[0])
            ]
        )
        scale = np.where(np.abs(g_ideal) > 1e-15, g_actual / g_ideal, 1.0)
        effective_weights = weights * scale
        outputs[trial] = effective_weights @ inputs

    errors = np.abs(outputs - ideal_out[np.newaxis, :])
    mean_err = float(np.mean(errors))
    rel_errors = errors / np.maximum(np.abs(ideal_out[np.newaxis, :]), 1e-15)
    within_tol = np.all(rel_errors < self.tolerance, axis=1)
    yield_frac = float(np.mean(within_tol))

    err_flat = errors.flatten()
    hist, _ = np.histogram(err_flat, bins=50)

    return MonteCarloReport(
        num_trials=self.num_trials,
        mean_output_error=mean_err,
        std_output_error=float(np.std(errors)),
        max_output_error=float(np.max(errors)),
        yield_fraction=yield_frac,
        output_distribution=np.mean(outputs, axis=0),
        error_histogram=hist,
    )

VerilogEmitter

Generates crossbar-aware SystemVerilog with compensation LUTs.

Source code in src/sc_neurocore/memristor/memristor_mapper.py
Python
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
class VerilogEmitter:
    """Generates crossbar-aware SystemVerilog with compensation LUTs."""

    def __init__(self, bit_width: int = 16, frac_bits: int = 8) -> None:
        self.bw = bit_width
        self.frac = frac_bits

    def emit_crossbar(
        self,
        mapping: CrossbarMapping,
        module_name: str = "sc_memristor_crossbar",
    ) -> str:
        """Generate SystemVerilog for a single crossbar tile."""
        r, c = mapping.crossbar.rows, mapping.crossbar.cols
        bw = self.bw

        # Build weight parameter block
        weight_params = []
        for i in range(r):
            for j in range(c):
                lvl = int(mapping.weight_levels[i, j])
                weight_params.append(f"    localparam [{bw - 1}:0] W_{i}_{j} = {bw}'d{lvl};")
        weight_block = "\n".join(weight_params)

        # Compensation LUT (if present)
        comp_block = ""
        if mapping.compensation_luts:
            num_levels = mapping.compensation_luts[0].nominal_levels.shape[0]
            comp_lines = [f"    // Compensation LUT ({num_levels} levels)"]
            comp_lines.append(f"    logic [{bw - 1}:0] comp_lut [0:{num_levels - 1}];")
            comp_lines.append("    initial begin")
            lut = mapping.compensation_luts[0]
            for k in range(num_levels):
                val = int(lut.compensated_thresholds[k])
                comp_lines.append(f"        comp_lut[{k}] = {bw}'d{val};")
            comp_lines.append("    end")
            comp_block = "\n".join(comp_lines)

        # MAC accumulator
        mac_lines = []
        for i in range(r):
            terms = " + ".join(f"(i_bitstream[{j}] & W_{i}_{j}[0])" for j in range(c))
            mac_lines.append(f"            o_mac[{i}] <= {terms};")
        mac_block = "\n".join(mac_lines)

        return textwrap.dedent(f"""\
            // SPDX-License-Identifier: AGPL-3.0-or-later
            // Commercial license available
            // © Concepts 1996–2026 Miroslav Šotek. All rights reserved.
            // © Code 2020–2026 Miroslav Šotek. All rights reserved.
            // ORCID: 0009-0009-3560-0851
            // Contact: www.anulum.li | protoscience@anulum.li
            // SC-NeuroCore — Memristor Crossbar ({mapping.crossbar.technology.value})

            module {module_name} #(
                parameter ROWS = {r},
                parameter COLS = {c},
                parameter BW = {bw}
            )(
                input  logic clk,
                input  logic rst_n,
                input  logic [COLS-1:0] i_bitstream,
                output logic [{bw - 1}:0] o_mac [0:ROWS-1],
                output logic o_valid
            );

                // Programmed weight levels (Q{self.bw - self.frac}.{self.frac} fixed-point)
{textwrap.indent(weight_block, "    ")}

{textwrap.indent(comp_block, "    ") if comp_block else "    // No compensation LUT"}

                always_ff @(posedge clk or negedge rst_n) begin
                    if (!rst_n) begin
                        o_valid <= 1'b0;
                    end else begin
{mac_block}
                        o_valid <= 1'b1;
                    end
                end

            endmodule
        """)

    def emit_top(
        self,
        result: MappingResult,
        module_name: str = "sc_memristor_array",
    ) -> str:
        """Generate top-level module instantiating all crossbar tiles."""
        bw = self.bw
        total_rows = sum(m.crossbar.rows for m in result.mappings)
        total_cols = max((m.crossbar.cols for m in result.mappings), default=1)

        inst_lines = []
        for idx, mapping in enumerate(result.mappings):
            inst_lines.append(
                textwrap.dedent(f"""\
                sc_memristor_crossbar #(
                    .ROWS({mapping.crossbar.rows}),
                    .COLS({mapping.crossbar.cols}),
                    .BW({bw})
                ) tile_{idx} (
                    .clk(clk),
                    .rst_n(rst_n),
                    .i_bitstream(i_bitstream[{mapping.crossbar.cols - 1}:0]),
                    .o_mac(tile_{idx}_mac),
                    .o_valid(tile_{idx}_valid)
                );""")
            )

        inst_block = "\n".join(inst_lines)

        return textwrap.dedent(f"""\
            // SPDX-License-Identifier: AGPL-3.0-or-later
            // © Code 2020–2026 Miroslav Šotek. All rights reserved.
            // SC-NeuroCore — Memristor Array Top ({result.total_crossbars} tiles)

            module {module_name} (
                input  logic clk,
                input  logic rst_n,
                input  logic [{total_cols - 1}:0] i_bitstream,
                output logic o_valid
            );

            {textwrap.indent(inst_block, "    ")}

            endmodule
        """)

emit_crossbar(mapping, module_name='sc_memristor_crossbar')

Generate SystemVerilog for a single crossbar tile.

Source code in src/sc_neurocore/memristor/memristor_mapper.py
Python
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
    def emit_crossbar(
        self,
        mapping: CrossbarMapping,
        module_name: str = "sc_memristor_crossbar",
    ) -> str:
        """Generate SystemVerilog for a single crossbar tile."""
        r, c = mapping.crossbar.rows, mapping.crossbar.cols
        bw = self.bw

        # Build weight parameter block
        weight_params = []
        for i in range(r):
            for j in range(c):
                lvl = int(mapping.weight_levels[i, j])
                weight_params.append(f"    localparam [{bw - 1}:0] W_{i}_{j} = {bw}'d{lvl};")
        weight_block = "\n".join(weight_params)

        # Compensation LUT (if present)
        comp_block = ""
        if mapping.compensation_luts:
            num_levels = mapping.compensation_luts[0].nominal_levels.shape[0]
            comp_lines = [f"    // Compensation LUT ({num_levels} levels)"]
            comp_lines.append(f"    logic [{bw - 1}:0] comp_lut [0:{num_levels - 1}];")
            comp_lines.append("    initial begin")
            lut = mapping.compensation_luts[0]
            for k in range(num_levels):
                val = int(lut.compensated_thresholds[k])
                comp_lines.append(f"        comp_lut[{k}] = {bw}'d{val};")
            comp_lines.append("    end")
            comp_block = "\n".join(comp_lines)

        # MAC accumulator
        mac_lines = []
        for i in range(r):
            terms = " + ".join(f"(i_bitstream[{j}] & W_{i}_{j}[0])" for j in range(c))
            mac_lines.append(f"            o_mac[{i}] <= {terms};")
        mac_block = "\n".join(mac_lines)

        return textwrap.dedent(f"""\
            // SPDX-License-Identifier: AGPL-3.0-or-later
            // Commercial license available
            // © Concepts 1996–2026 Miroslav Šotek. All rights reserved.
            // © Code 2020–2026 Miroslav Šotek. All rights reserved.
            // ORCID: 0009-0009-3560-0851
            // Contact: www.anulum.li | protoscience@anulum.li
            // SC-NeuroCore — Memristor Crossbar ({mapping.crossbar.technology.value})

            module {module_name} #(
                parameter ROWS = {r},
                parameter COLS = {c},
                parameter BW = {bw}
            )(
                input  logic clk,
                input  logic rst_n,
                input  logic [COLS-1:0] i_bitstream,
                output logic [{bw - 1}:0] o_mac [0:ROWS-1],
                output logic o_valid
            );

                // Programmed weight levels (Q{self.bw - self.frac}.{self.frac} fixed-point)
{textwrap.indent(weight_block, "    ")}

{textwrap.indent(comp_block, "    ") if comp_block else "    // No compensation LUT"}

                always_ff @(posedge clk or negedge rst_n) begin
                    if (!rst_n) begin
                        o_valid <= 1'b0;
                    end else begin
{mac_block}
                        o_valid <= 1'b1;
                    end
                end

            endmodule
        """)

emit_top(result, module_name='sc_memristor_array')

Generate top-level module instantiating all crossbar tiles.

Source code in src/sc_neurocore/memristor/memristor_mapper.py
Python
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
def emit_top(
    self,
    result: MappingResult,
    module_name: str = "sc_memristor_array",
) -> str:
    """Generate top-level module instantiating all crossbar tiles."""
    bw = self.bw
    total_rows = sum(m.crossbar.rows for m in result.mappings)
    total_cols = max((m.crossbar.cols for m in result.mappings), default=1)

    inst_lines = []
    for idx, mapping in enumerate(result.mappings):
        inst_lines.append(
            textwrap.dedent(f"""\
            sc_memristor_crossbar #(
                .ROWS({mapping.crossbar.rows}),
                .COLS({mapping.crossbar.cols}),
                .BW({bw})
            ) tile_{idx} (
                .clk(clk),
                .rst_n(rst_n),
                .i_bitstream(i_bitstream[{mapping.crossbar.cols - 1}:0]),
                .o_mac(tile_{idx}_mac),
                .o_valid(tile_{idx}_valid)
            );""")
        )

    inst_block = "\n".join(inst_lines)

    return textwrap.dedent(f"""\
        // SPDX-License-Identifier: AGPL-3.0-or-later
        // © Code 2020–2026 Miroslav Šotek. All rights reserved.
        // SC-NeuroCore — Memristor Array Top ({result.total_crossbars} tiles)

        module {module_name} (
            input  logic clk,
            input  logic rst_n,
            input  logic [{total_cols - 1}:0] i_bitstream,
            output logic o_valid
        );

        {textwrap.indent(inst_block, "    ")}

        endmodule
    """)