Skip to content

Adaptive Precision

Per-layer adaptive bitstream length and per-synapse bit-width planning for mixed-precision SC networks.

  • assign_lengths — auto-select bitstream length per layer from Hoeffding or sensitivity bounds.
  • assign_synapse_precisions — choose integer bit widths and stochastic bitstream lengths per synapse.
  • auto_tune_synapse_precisions — emit the public precision-plan manifest for a percent error target.
  • write_precision_formal_evidence_bundle — write bounded SVA/SBY/JSON evidence scaffolding for a precision plan.
Python
from sc_neurocore.compiler.adaptive_precision import assign_lengths

sc_neurocore.compiler.adaptive_precision is in the scoped public-docstring policy. Its dedicated adaptive-precision tests are strict typed and cover layer-length assignment, sensitivity analysis, per-synapse planning, manifest stability, row validation, and formal-evidence bundle writing at 100% isolated facade coverage.

LayerPrecision and SynapsePrecision are validated row contracts behind the public manifest. Layer rows reject negative indices, empty names, non-positive or non-power-of-two bitstream lengths, and non-finite or negative error and sensitivity values. Synapse rows reject negative coordinates, empty layer names, non-positive bit widths or bitstream lengths, invalid component bounds, and total error bounds that do not cover the quantisation plus stochastic components.

The layer and synapse planners now fail closed before row emission. assign_lengths rejects unsupported methods, mismatched layer_names, invalid target/length bounds, empty layers, non-finite weights, and tensors outside the documented 1D/2D contract. assign_synapse_precisions rejects invalid target, bit-width, length, and confidence bounds, mismatched names or sensitivity maps, empty or non-finite weight layers, and non-finite or negative sensitivity maps. One dimensional layer and sensitivity vectors remain supported and are serialized as single-output rows.

The polyglot adaptive-precision mirrors are validated with the Python row contract:

  • Rust: src/sc_neurocore/accel/rust/safety/adaptive_precision.rs compiles as a Rust test binary and exercises layer/synapse contract checks.
  • Julia: src/sc_neurocore/accel/julia/compiler/adaptive_precision.jl exposes LayerPrecisionState, SynapsePrecisionState, to_dict, and validation helpers.
  • Mojo: src/sc_neurocore/accel/mojo/kernels/adaptive_precision.mojo runs the same layer and synapse validation smoke path.

Local non-isolated benchmark and validation evidence is stored in benchmarks/results/bench_adaptive_precision_rows.json. On this workstation, the 2026-06-27 artifact records 10,000 Python LayerPrecision row constructor/serialization calls in 0.031964 s (312849.473 calls/s), 10,000 SynapsePrecision row constructor/serialization calls in 0.10749 s (93031.712 calls/s), Rust compile status pass, Rust tests status pass, Julia validation status pass, and Mojo validation status pass. The artifact sets production_benchmark_claim to false; it is regression evidence, not a production performance claim.

The 2026-06-27 planner validation slice did not change numeric planner objectives or add a new backend benchmark. It added fail-closed validation and a dedicated coverage gate: tests/.coveragerc_adaptive_precision_planners, covering length_planner.py and synapse_planner.py at 100% with the adaptive precision planner tests.

2026-04-30 per-synapse precision plan

The adaptive precision module now includes a conservative per-synapse planner for the roadmap auto-adaptive precision optimiser. It assigns integer bit_width, SC bitstream_length, sensitivity, quantisation-error bound, stochastic-error bound, and total bound for each synapse:

Python
import numpy as np

from sc_neurocore.compiler.adaptive_precision import (
    assign_synapse_precisions,
    precision_plan_manifest,
)

weights = [np.array([[0.1, 0.8], [0.0, 0.4]])]
plan = assign_synapse_precisions(weights, target_error=0.05)
manifest = precision_plan_manifest(plan)

This is a deterministic planning surface, not a training-result claim. Bounds are intentionally conservative: quantisation is bounded by half an integer step scaled by sensitivity, and stochastic sampling uses the existing Hoeffding bitstream-length helper. Custom sensitivity maps can be supplied after an external sensitivity-analysis pass.

Adaptive runtime precision BFP metadata

compile_adaptive_precision(...) accepts fixed Q-format strings such as Q8.8/Q16.16 and block-floating strings such as BFP16E3X32. When a block-floating precision is supplied with lp_parameter_count or hp_parameter_count, the generated manifest records the exact block_exponent_layout: flattened row-major parameter count, block size, exponent-vector length, and final partial-block size. Invalid negative parameter counts fail before RTL emission.

This metadata is a compiler contract for downstream emitters. The generated adaptive wrapper still emits fixed mantissa-width datapaths; shared exponents remain explicit metadata until the target-specific BFP datapath is selected. Every adaptive manifest carries adaptive_precision_emitter.v1, explicit emitted_datapath_width, emitted_datapath_fraction, exponent_stream_width, and exponent_vector_width fields. Fixed Q-format paths set the exponent widths to zero and reject accidental *_parameter_count inputs so a block-exponent layout cannot be silently dropped before HDL/Rust emitter handoff.

sc_neurocore.compiler.adaptive_precision

Per-layer adaptive bitstream length and per-synapse bit width facade.

LayerPrecision dataclass

Bitstream length assignment for one layer.

Parameters

layer_index: Zero-based layer index in the adaptive-precision plan. name: Non-empty layer name used in reports and manifests. bitstream_length: Positive stochastic-computing bitstream length. Layer-level planners round this value to a power of two. error_bound: Finite non-negative per-layer stochastic error bound. sensitivity: Finite non-negative sensitivity score used for budget allocation.

Source code in src/sc_neurocore/compiler/layer_precision.py
Python
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
@dataclass
class LayerPrecision:
    """Bitstream length assignment for one layer.

    Parameters
    ----------
    layer_index:
        Zero-based layer index in the adaptive-precision plan.
    name:
        Non-empty layer name used in reports and manifests.
    bitstream_length:
        Positive stochastic-computing bitstream length. Layer-level planners
        round this value to a power of two.
    error_bound:
        Finite non-negative per-layer stochastic error bound.
    sensitivity:
        Finite non-negative sensitivity score used for budget allocation.
    """

    layer_index: int
    name: str
    bitstream_length: int
    error_bound: float
    sensitivity: float

    def __post_init__(self) -> None:
        """Validate adaptive-precision row invariants."""
        _validate_non_negative_int(self.layer_index, "layer_index")
        if not isinstance(self.name, str) or not self.name:
            raise ValueError("name must be a non-empty string")
        _validate_positive_int(self.bitstream_length, "bitstream_length")
        if self.bitstream_length & (self.bitstream_length - 1) != 0:
            raise ValueError("bitstream_length must be a power of two")
        _validate_non_negative_float(self.error_bound, "error_bound")
        _validate_non_negative_float(self.sensitivity, "sensitivity")

    def to_dict(self) -> dict[str, int | float | str]:
        """Return a JSON-serializable adaptive-precision manifest row."""
        return {
            "layer_index": self.layer_index,
            "name": self.name,
            "bitstream_length": self.bitstream_length,
            "error_bound": self.error_bound,
            "sensitivity": self.sensitivity,
        }

__post_init__()

Validate adaptive-precision row invariants.

Source code in src/sc_neurocore/compiler/layer_precision.py
Python
42
43
44
45
46
47
48
49
50
51
def __post_init__(self) -> None:
    """Validate adaptive-precision row invariants."""
    _validate_non_negative_int(self.layer_index, "layer_index")
    if not isinstance(self.name, str) or not self.name:
        raise ValueError("name must be a non-empty string")
    _validate_positive_int(self.bitstream_length, "bitstream_length")
    if self.bitstream_length & (self.bitstream_length - 1) != 0:
        raise ValueError("bitstream_length must be a power of two")
    _validate_non_negative_float(self.error_bound, "error_bound")
    _validate_non_negative_float(self.sensitivity, "sensitivity")

to_dict()

Return a JSON-serializable adaptive-precision manifest row.

Source code in src/sc_neurocore/compiler/layer_precision.py
Python
53
54
55
56
57
58
59
60
61
def to_dict(self) -> dict[str, int | float | str]:
    """Return a JSON-serializable adaptive-precision manifest row."""
    return {
        "layer_index": self.layer_index,
        "name": self.name,
        "bitstream_length": self.bitstream_length,
        "error_bound": self.error_bound,
        "sensitivity": self.sensitivity,
    }

SynapsePrecision dataclass

Precision assignment and conservative error bound for one synapse.

Parameters

layer_index: Zero-based layer index in the adaptive-precision plan. layer_name: Non-empty layer name used in reports and manifests. output_index: Zero-based output row index for the weight matrix. input_index: Zero-based input column index for the weight matrix. bit_width: Positive fixed-point bit width assigned to the synapse. bitstream_length: Positive stochastic-computing bitstream length assigned to the synapse. sensitivity: Finite non-negative synapse sensitivity score. quantization_error_bound: Finite non-negative error bound from fixed-point quantization. stochastic_error_bound: Finite non-negative Hoeffding-style stochastic error bound. total_error_bound: Finite non-negative aggregate bound that must cover both components.

Source code in src/sc_neurocore/compiler/synapse_precision.py
Python
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
@dataclass(frozen=True)
class SynapsePrecision:
    """Precision assignment and conservative error bound for one synapse.

    Parameters
    ----------
    layer_index:
        Zero-based layer index in the adaptive-precision plan.
    layer_name:
        Non-empty layer name used in reports and manifests.
    output_index:
        Zero-based output row index for the weight matrix.
    input_index:
        Zero-based input column index for the weight matrix.
    bit_width:
        Positive fixed-point bit width assigned to the synapse.
    bitstream_length:
        Positive stochastic-computing bitstream length assigned to the synapse.
    sensitivity:
        Finite non-negative synapse sensitivity score.
    quantization_error_bound:
        Finite non-negative error bound from fixed-point quantization.
    stochastic_error_bound:
        Finite non-negative Hoeffding-style stochastic error bound.
    total_error_bound:
        Finite non-negative aggregate bound that must cover both components.
    """

    layer_index: int
    layer_name: str
    output_index: int
    input_index: int
    bit_width: int
    bitstream_length: int
    sensitivity: float
    quantization_error_bound: float
    stochastic_error_bound: float
    total_error_bound: float

    def __post_init__(self) -> None:
        """Validate per-synapse precision-row invariants."""
        _validate_non_negative_int(self.layer_index, "layer_index")
        if not isinstance(self.layer_name, str) or not self.layer_name:
            raise ValueError("layer_name must be a non-empty string")
        _validate_non_negative_int(self.output_index, "output_index")
        _validate_non_negative_int(self.input_index, "input_index")
        _validate_positive_int(self.bit_width, "bit_width")
        _validate_positive_int(self.bitstream_length, "bitstream_length")
        _validate_non_negative_float(self.sensitivity, "sensitivity")
        _validate_non_negative_float(
            self.quantization_error_bound,
            "quantization_error_bound",
        )
        _validate_non_negative_float(
            self.stochastic_error_bound,
            "stochastic_error_bound",
        )
        _validate_non_negative_float(self.total_error_bound, "total_error_bound")
        component_sum = self.quantization_error_bound + self.stochastic_error_bound
        if self.total_error_bound + 1e-15 < component_sum:
            raise ValueError("total_error_bound must cover quantization and stochastic bounds")

    def to_dict(self) -> dict[str, int | float | str]:
        """Return a JSON-serialisable precision-plan row."""
        return {
            "layer_index": self.layer_index,
            "layer_name": self.layer_name,
            "output_index": self.output_index,
            "input_index": self.input_index,
            "bit_width": self.bit_width,
            "bitstream_length": self.bitstream_length,
            "sensitivity": self.sensitivity,
            "quantization_error_bound": self.quantization_error_bound,
            "stochastic_error_bound": self.stochastic_error_bound,
            "total_error_bound": self.total_error_bound,
        }

__post_init__()

Validate per-synapse precision-row invariants.

Source code in src/sc_neurocore/compiler/synapse_precision.py
Python
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
def __post_init__(self) -> None:
    """Validate per-synapse precision-row invariants."""
    _validate_non_negative_int(self.layer_index, "layer_index")
    if not isinstance(self.layer_name, str) or not self.layer_name:
        raise ValueError("layer_name must be a non-empty string")
    _validate_non_negative_int(self.output_index, "output_index")
    _validate_non_negative_int(self.input_index, "input_index")
    _validate_positive_int(self.bit_width, "bit_width")
    _validate_positive_int(self.bitstream_length, "bitstream_length")
    _validate_non_negative_float(self.sensitivity, "sensitivity")
    _validate_non_negative_float(
        self.quantization_error_bound,
        "quantization_error_bound",
    )
    _validate_non_negative_float(
        self.stochastic_error_bound,
        "stochastic_error_bound",
    )
    _validate_non_negative_float(self.total_error_bound, "total_error_bound")
    component_sum = self.quantization_error_bound + self.stochastic_error_bound
    if self.total_error_bound + 1e-15 < component_sum:
        raise ValueError("total_error_bound must cover quantization and stochastic bounds")

to_dict()

Return a JSON-serialisable precision-plan row.

Source code in src/sc_neurocore/compiler/synapse_precision.py
Python
79
80
81
82
83
84
85
86
87
88
89
90
91
92
def to_dict(self) -> dict[str, int | float | str]:
    """Return a JSON-serialisable precision-plan row."""
    return {
        "layer_index": self.layer_index,
        "layer_name": self.layer_name,
        "output_index": self.output_index,
        "input_index": self.input_index,
        "bit_width": self.bit_width,
        "bitstream_length": self.bitstream_length,
        "sensitivity": self.sensitivity,
        "quantization_error_bound": self.quantization_error_bound,
        "stochastic_error_bound": self.stochastic_error_bound,
        "total_error_bound": self.total_error_bound,
    }

auto_tune_synapse_precisions(layer_weights, *, layer_names=None, target_error_percent=0.1, min_bits=4, max_bits=16, min_length=32, max_length=4096, confidence=0.95)

Auto-tune per-synapse precision for an explicit percent error target.

Source code in src/sc_neurocore/compiler/auto_tune.py
Python
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
def auto_tune_synapse_precisions(
    layer_weights: list[np.ndarray[Any, Any]],
    *,
    layer_names: list[str] | None = None,
    target_error_percent: float = 0.1,
    min_bits: int = 4,
    max_bits: int = 16,
    min_length: int = 32,
    max_length: int = 4096,
    confidence: float = 0.95,
) -> dict[str, Any]:
    """Auto-tune per-synapse precision for an explicit percent error target."""
    if target_error_percent <= 0:
        raise ValueError("target_error_percent must be positive")
    target_error = target_error_percent / 100.0
    assignments = assign_synapse_precisions(
        layer_weights,
        layer_names=layer_names,
        target_error=target_error,
        min_bits=min_bits,
        max_bits=max_bits,
        min_length=min_length,
        max_length=max_length,
        confidence=confidence,
    )
    manifest = precision_plan_manifest(assignments)
    manifest["api_surface"] = {
        "action_id": "auto_tune_adaptive_precision",
        "target_error_percent": target_error_percent,
        "target_error_fraction": target_error,
        "objective": "minimal_luts_under_error_target",
        "cost_metric": "sum(bit_width * log2(bitstream_length))",
        "estimated_lut_cost": manifest["cost_summary"]["estimated_lut_cost"],
        "uniform_length_reference_cost": manifest["cost_summary"]["uniform_length_reference_cost"],
        "estimated_lut_savings_vs_uniform_length": manifest["cost_summary"][
            "estimated_lut_savings_vs_uniform_length"
        ],
    }
    return manifest

precision_plan_manifest(assignments)

Build a deterministic manifest for a per-synapse precision plan.

Source code in src/sc_neurocore/compiler/auto_tune.py
Python
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
def precision_plan_manifest(assignments: list[Any]) -> dict[str, Any]:
    """Build a deterministic manifest for a per-synapse precision plan."""
    rows = [assignment.to_dict() for assignment in assignments]
    cost_summary = _precision_cost_summary(assignments)
    return {
        "schema": "sc-neurocore.adaptive_precision_plan.v1",
        "granularity": "synapse",
        "num_synapses": len(assignments),
        "max_total_error_bound": max(
            (assignment.total_error_bound for assignment in assignments),
            default=0.0,
        ),
        "cost_summary": cost_summary,
        "assignments": rows,
    }

write_precision_formal_evidence_bundle(output_dir, assignments, *, module_name='adaptive_precision_plan')

Write a deterministic SymbiYosys evidence bundle for precision claims.

Source code in src/sc_neurocore/compiler/formal_evidence.py
Python
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
def write_precision_formal_evidence_bundle(
    output_dir: str | Path,
    assignments: list[SynapsePrecision],
    *,
    module_name: str = "adaptive_precision_plan",
) -> dict[str, Any]:
    """Write a deterministic SymbiYosys evidence bundle for precision claims."""
    if not assignments:
        raise ValueError("assignments must not be empty")
    out = Path(output_dir)
    out.mkdir(parents=True, exist_ok=True)

    max_error = max(item.total_error_bound for item in assignments)
    max_bits = max(item.bit_width for item in assignments)
    max_length = max(item.bitstream_length for item in assignments)

    rtl_file = f"{module_name}.v"
    sva_file = f"{module_name}_sva.sv"
    sby_file = f"{module_name}.sby"
    report_file = f"{module_name}_formal_report.json"

    (out / sva_file).write_text(
        _adaptive_precision_sva(
            module_name=module_name,
            max_total_error_bound=max_error,
            max_bit_width=max_bits,
            max_bitstream_length=max_length,
        ),
        encoding="utf-8",
    )

    from .deployment import generate_sby_script

    (out / sby_file).write_text(
        generate_sby_script(module_name, sva_file=sva_file, mode="prove", depth=32),
        encoding="utf-8",
    )

    manifest = {
        "schema_version": "sc-neurocore.adaptive-precision-formal-bundle.v1",
        "evidence_classification": "compile",
        "status": "completed",
        "module_name": module_name,
        "evidence_boundary": ("bundle_generation_only_no_symbiyosys_execution_no_silicon_claim"),
        "assignments_count": len(assignments),
        "formal_claim": {
            "max_total_error_bound": max_error,
            "max_bit_width": max_bits,
            "max_bitstream_length": max_length,
            "symbiyosys_executed": False,
            "formal_proof_passed": False,
            "hardware_measurement_claimed": False,
        },
        "artifacts": {
            "rtl": rtl_file,
            "sva": sva_file,
            "sby": sby_file,
            "report": report_file,
        },
    }
    (out / f"{module_name}_formal_manifest.json").write_text(
        json.dumps(manifest, indent=2, sort_keys=True) + "\n",
        encoding="utf-8",
    )
    return manifest

assign_lengths(layer_weights, layer_names=None, total_budget=None, min_length=32, max_length=1024, target_error=0.01, method='hoeffding')

Assign per-layer bitstream lengths under a target error budget.

Parameters

layer_weights: One- or two-dimensional finite weight tensors, one tensor per layer. One-dimensional tensors are treated as single-output layers. layer_names: Optional non-empty layer names. When provided, the list length must match layer_weights exactly. total_budget: Optional aggregate bitstream-length budget for sensitivity planning. When omitted, sensitivity planning uses max_length * n_layers. min_length: Minimum bitstream length assigned to any layer. max_length: Maximum bitstream length assigned to any layer. target_error: Positive per-layer target error used by the Hoeffding planner. method: Planning method. hoeffding uses analytic Hoeffding lengths; sensitivity and proportional allocate from sensitivity scores.

Returns

list[LayerPrecision] Validated layer precision rows in input-layer order.

Raises

ValueError If planner bounds, method, names, or weight tensors are invalid.

Source code in src/sc_neurocore/compiler/length_planner.py
Python
 23
 24
 25
 26
 27
 28
 29
 30
 31
 32
 33
 34
 35
 36
 37
 38
 39
 40
 41
 42
 43
 44
 45
 46
 47
 48
 49
 50
 51
 52
 53
 54
 55
 56
 57
 58
 59
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 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
def assign_lengths(
    layer_weights: list[np.ndarray[Any, Any]],
    layer_names: list[str] | None = None,
    total_budget: int | None = None,
    min_length: int = 32,
    max_length: int = 1024,
    target_error: float = 0.01,
    method: str = "hoeffding",
) -> list[LayerPrecision]:
    """Assign per-layer bitstream lengths under a target error budget.

    Parameters
    ----------
    layer_weights:
        One- or two-dimensional finite weight tensors, one tensor per layer.
        One-dimensional tensors are treated as single-output layers.
    layer_names:
        Optional non-empty layer names. When provided, the list length must match
        `layer_weights` exactly.
    total_budget:
        Optional aggregate bitstream-length budget for sensitivity planning.
        When omitted, sensitivity planning uses `max_length * n_layers`.
    min_length:
        Minimum bitstream length assigned to any layer.
    max_length:
        Maximum bitstream length assigned to any layer.
    target_error:
        Positive per-layer target error used by the Hoeffding planner.
    method:
        Planning method. `hoeffding` uses analytic Hoeffding lengths;
        `sensitivity` and `proportional` allocate from sensitivity scores.

    Returns
    -------
    list[LayerPrecision]
        Validated layer precision rows in input-layer order.

    Raises
    ------
    ValueError
        If planner bounds, method, names, or weight tensors are invalid.
    """
    _validate_planner_bounds(min_length, max_length, target_error)
    if method not in {"hoeffding", "sensitivity", "proportional"}:
        raise ValueError("method must be one of: hoeffding, sensitivity, proportional")

    validated_weights = [_as_weight_array(weights) for weights in layer_weights]
    n_layers = len(validated_weights)
    if layer_names is None:
        layer_names = [f"layer_{i}" for i in range(n_layers)]
    elif len(layer_names) != n_layers:
        raise ValueError("layer_names length must match layer_weights")

    if method == "hoeffding":
        assignments: list[LayerPrecision] = []
        for i, (w, name) in enumerate(zip(validated_weights, layer_names)):
            fan_in = w.shape[1] if w.ndim == 2 else 1
            per_syn_eps = target_error / max(1, np.sqrt(fan_in))
            L = adaptive_length(p=0.5, epsilon=per_syn_eps, confidence=0.95)
            L = int(np.clip(L, min_length, max_length))
            L = int(2 ** np.ceil(np.log2(max(L, min_length))))
            L = min(L, max_length)
            bound = 0.5 / np.sqrt(L) if L > 0 else 1.0
            assignments.append(
                LayerPrecision(
                    layer_index=i,
                    name=name,
                    bitstream_length=L,
                    error_bound=bound,
                    sensitivity=0.0,
                )
            )
        return assignments

    sensitivities = analyze_sensitivity(validated_weights)
    total_sens = sum(sensitivities) or 1.0

    if total_budget is None:
        total_budget = max_length * n_layers
    elif total_budget <= 0:
        raise ValueError("total_budget must be positive when provided")

    assignments = []
    for i, (w, name, sens) in enumerate(zip(validated_weights, layer_names, sensitivities)):
        fraction = sens / total_sens
        L = int(fraction * total_budget / n_layers * n_layers)
        L = int(np.clip(L, min_length, max_length))
        L = int(2 ** np.ceil(np.log2(max(L, min_length))))
        L = min(L, max_length)
        bound = 0.5 / np.sqrt(L) if L > 0 else 1.0
        assignments.append(
            LayerPrecision(
                layer_index=i,
                name=name,
                bitstream_length=L,
                error_bound=bound,
                sensitivity=sens,
            )
        )

    return assignments

analyze_sensitivity(layer_weights, lengths=None, n_trials=100, seed=42)

Measure per-layer sensitivity to bitstream length reduction.

Parameters

layer_weights: One-dimensional or two-dimensional layer weight arrays. Vector weights are treated as a single-output dense layer. lengths: Candidate stochastic bitstream lengths to sample. When omitted, the estimator uses the default production planning ladder. n_trials: Number of independent input-vector samples per layer. seed: Deterministic NumPy random seed used for reproducible planning.

Returns

list[float] One non-negative sensitivity score for each supplied layer.

Raises

ValueError If trial count, candidate lengths, or layer weight arrays are invalid.

Source code in src/sc_neurocore/compiler/sensitivity_analysis.py
Python
 21
 22
 23
 24
 25
 26
 27
 28
 29
 30
 31
 32
 33
 34
 35
 36
 37
 38
 39
 40
 41
 42
 43
 44
 45
 46
 47
 48
 49
 50
 51
 52
 53
 54
 55
 56
 57
 58
 59
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
def analyze_sensitivity(
    layer_weights: list[np.ndarray[Any, Any]],
    lengths: list[int] | None = None,
    n_trials: int = 100,
    seed: int = 42,
) -> list[float]:
    """Measure per-layer sensitivity to bitstream length reduction.

    Parameters
    ----------
    layer_weights:
        One-dimensional or two-dimensional layer weight arrays. Vector weights
        are treated as a single-output dense layer.
    lengths:
        Candidate stochastic bitstream lengths to sample. When omitted, the
        estimator uses the default production planning ladder.
    n_trials:
        Number of independent input-vector samples per layer.
    seed:
        Deterministic NumPy random seed used for reproducible planning.

    Returns
    -------
    list[float]
        One non-negative sensitivity score for each supplied layer.

    Raises
    ------
    ValueError
        If trial count, candidate lengths, or layer weight arrays are invalid.
    """
    candidate_lengths = _validate_lengths(lengths)
    if n_trials <= 0:
        raise ValueError("n_trials must be positive")

    rng = np.random.RandomState(seed)
    sensitivities: list[float] = []

    for raw_weights in layer_weights:
        weights = _as_weight_matrix(raw_weights)
        n_outputs, n_inputs = weights.shape
        clipped_weights = np.clip(weights, 0.0, 1.0)
        errors: list[float] = []

        for _ in range(n_trials):
            input_probabilities = rng.random_sample(n_inputs).astype(np.float64)
            exact = weights @ input_probabilities
            target = np.clip(exact, 0.0, None)

            length_errors: list[float] = []
            for bitstream_length in candidate_lengths:
                sc_results: list[_FloatArray] = []
                for _ in range(5):
                    bits_x = (
                        rng.random_sample((bitstream_length, n_inputs)) < input_probabilities
                    ).astype(np.float64)
                    bits_w = (
                        rng.random_sample((bitstream_length, n_outputs, n_inputs))
                        < clipped_weights[np.newaxis, :, :]
                    ).astype(np.float64)
                    and_result = bits_x[:, np.newaxis, :] * bits_w
                    counts = np.sum(
                        and_result,
                        axis=(0, 2),
                        dtype=np.float64,
                        initial=0.0,
                    )
                    sc_results.append(
                        np.asarray(counts / float(bitstream_length), dtype=np.float64)
                    )

                sc_mean = np.mean(np.stack(sc_results, axis=0), axis=0)
                err = np.mean(np.abs(sc_mean - target))
                length_errors.append(float(err))

            sensitivity = max(length_errors) - min(length_errors) if length_errors else 0.0
            errors.append(sensitivity)

        sensitivities.append(float(np.mean(errors)))

    return sensitivities

assign_synapse_precisions(layer_weights, layer_names=None, sensitivity_maps=None, target_error=0.01, min_bits=4, max_bits=16, min_length=32, max_length=4096, confidence=0.95)

Assign per-synapse bit widths and SC lengths with error bounds.

Parameters

layer_weights: One- or two-dimensional finite weight tensors, one tensor per layer. One-dimensional tensors are treated as single-output layers. layer_names: Optional non-empty layer names. When provided, the list length must match layer_weights exactly. sensitivity_maps: Optional finite non-negative sensitivity maps with shapes matching each corresponding weight tensor. target_error: Positive aggregate target error fraction. min_bits: Minimum fixed-point bit width assigned to any synapse. max_bits: Maximum fixed-point bit width assigned to any synapse. min_length: Minimum stochastic bitstream length assigned to any synapse. max_length: Maximum stochastic bitstream length assigned to any synapse. confidence: Hoeffding confidence in the open interval (0, 1).

Returns

list[SynapsePrecision] Validated per-synapse precision rows in layer/output/input order.

Raises

ValueError If planner bounds, names, sensitivity maps, or weight tensors are invalid.

Source code in src/sc_neurocore/compiler/synapse_planner.py
Python
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 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
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
def assign_synapse_precisions(
    layer_weights: list[np.ndarray[Any, Any]],
    layer_names: list[str] | None = None,
    sensitivity_maps: list[np.ndarray[Any, Any]] | None = None,
    target_error: float = 0.01,
    min_bits: int = 4,
    max_bits: int = 16,
    min_length: int = 32,
    max_length: int = 4096,
    confidence: float = 0.95,
) -> list[SynapsePrecision]:
    """Assign per-synapse bit widths and SC lengths with error bounds.

    Parameters
    ----------
    layer_weights:
        One- or two-dimensional finite weight tensors, one tensor per layer.
        One-dimensional tensors are treated as single-output layers.
    layer_names:
        Optional non-empty layer names. When provided, the list length must match
        `layer_weights` exactly.
    sensitivity_maps:
        Optional finite non-negative sensitivity maps with shapes matching each
        corresponding weight tensor.
    target_error:
        Positive aggregate target error fraction.
    min_bits:
        Minimum fixed-point bit width assigned to any synapse.
    max_bits:
        Maximum fixed-point bit width assigned to any synapse.
    min_length:
        Minimum stochastic bitstream length assigned to any synapse.
    max_length:
        Maximum stochastic bitstream length assigned to any synapse.
    confidence:
        Hoeffding confidence in the open interval `(0, 1)`.

    Returns
    -------
    list[SynapsePrecision]
        Validated per-synapse precision rows in layer/output/input order.

    Raises
    ------
    ValueError
        If planner bounds, names, sensitivity maps, or weight tensors are
        invalid.
    """
    if not np.isfinite(target_error) or target_error <= 0.0:
        raise ValueError("target_error must be finite and positive")
    if min_bits < 1 or max_bits < min_bits:
        raise ValueError("bit-width bounds must satisfy 1 <= min_bits <= max_bits")
    if min_length < 1 or max_length < min_length:
        raise ValueError("length bounds must satisfy 1 <= min_length <= max_length")
    if confidence <= 0.0 or confidence >= 1.0:
        raise ValueError("confidence must satisfy 0.0 < confidence < 1.0")

    validated_weights = [_as_weight_array(weights) for weights in layer_weights]
    n_layers = len(validated_weights)
    if layer_names is None:
        layer_names = [f"layer_{i}" for i in range(n_layers)]
    if len(layer_names) != n_layers:
        raise ValueError("layer_names length must match layer_weights")
    if sensitivity_maps is not None and len(sensitivity_maps) != n_layers:
        raise ValueError("sensitivity_maps length must match layer_weights")

    total_synapses = sum(int(w.size) for w in validated_weights)
    local_target = target_error / max(1.0, float(np.sqrt(total_synapses)))
    assignments: list[SynapsePrecision] = []

    for layer_index, (w, name) in enumerate(zip(validated_weights, layer_names)):
        matrix: np.ndarray[Any, Any] = w.reshape(1, -1) if w.ndim == 1 else w

        if sensitivity_maps is None:
            max_abs = float(np.max(np.abs(matrix))) if matrix.size else 0.0
            sensitivity = np.abs(matrix) / max(max_abs, 1e-12)
        else:
            sensitivity_raw = np.asarray(sensitivity_maps[layer_index], dtype=float)
            if sensitivity_raw.shape != w.shape:
                raise ValueError("each sensitivity map must match its layer weight shape")
            sensitivity = (
                sensitivity_raw.reshape(1, -1) if sensitivity_raw.ndim == 1 else sensitivity_raw
            )
            if np.any(sensitivity < 0) or not np.all(np.isfinite(sensitivity)):
                raise ValueError("sensitivity maps must contain finite non-negative values")

        for output_index in range(matrix.shape[0]):
            for input_index in range(matrix.shape[1]):
                sens = float(sensitivity[output_index, input_index])
                bit_width = _select_bit_width(sens, local_target, min_bits, max_bits)
                quant_bound = _quantization_error_bound(sens, bit_width)
                remaining = max(local_target - quant_bound, local_target * 0.25)
                if sens <= 0:
                    length = min_length
                    stochastic_bound = 0.0
                else:
                    epsilon = max(remaining / sens, 1e-12)
                    length = adaptive_length(
                        p=0.5,
                        epsilon=epsilon,
                        confidence=confidence,
                        min_length=min_length,
                        max_length=max_length,
                    )
                    stochastic_bound = sens * _hoeffding_radius(length, confidence)

                assignments.append(
                    SynapsePrecision(
                        layer_index=layer_index,
                        layer_name=name,
                        output_index=output_index,
                        input_index=input_index,
                        bit_width=bit_width,
                        bitstream_length=length,
                        sensitivity=sens,
                        quantization_error_bound=quant_bound,
                        stochastic_error_bound=stochastic_bound,
                        total_error_bound=quant_bound + stochastic_bound,
                    )
                )

    return assignments