ph-curves

ph-curves turns integer observations and normalized positions into deterministic firmware values: static curves, sparse ADC-to-measurement transfers, runtime calibration, fixed-memory stabilization, hysteretic decisions, and tickless deadlines.

The host generator may model and fit with floating point, but firmware receives only static integer tables and bounded state. The runtime is always no_std and allocation-free.

Start here

Choose the path that matches the job:

Features

The features are ordered from build-time definition through the runtime measurement pipeline:

Install

Firmware normally uses the default, runtime-only crate:

[dependencies]
ph-curves = "0.3"

Host generation is opt-in:

#![no_std] is unconditional. Cargo feature unification cannot turn the firmware runtime into a std build when another crate enables a host feature.

Quick starts

Quick start: generate a curve

Create assets/curves.toml:

[curves.gamma_22]
formula = "pow(t, 2.2)"

[curves.ease_in_quad]
builtin = "ease_in_quad"

Install or run the generator:

cargo install ph-curves --version 0.3.0 --features gen-cli --bin ph-curves-gen
ph-curves-gen --input assets/curves.toml --output src/curves.rs

Use the generated constant:

use ph_curves::{Curve, MonotonicCurve};

include!("curves.rs");

let brightness: u8 = GAMMA_22.eval(input);
let input_again: u8 = GAMMA_22.inv(brightness);

Curve definitions use exactly one of builtin, formula, or points. Monotonic curves also emit inverse data. See the checked-in curve definitions for complete examples.

The curve-generation guide lists every built-in curve, the formula language, naming rules, and LUT target constraints.

Quick start: convert an ADC observation

Generate the reference NTC transfer:

ph-curves-gen --input assets/transfers.toml --output src/transfers.rs

use ph_curves::{InverseTransferFunction, TransferFunction};

include!("transfers.rs");

let milli_celsius = NTC_10K_BETA_3950.convert(adc_code)?;
let code_for_25_c = NTC_10K_BETA_3950.invert(25_000)?;

The reference uses 61 adaptive knots over a 12-bit ADC domain rather than a 4,096-entry dense table. The generator checks every integer code and fails if the requested interpolation error cannot be met within the knot budget.

Use a transfer when one monotonic u16 observation determines one signed, scaled i32 result. Formula, physical-point, and supported model fitting are host-only; generated firmware uses binary search and checked i64 interpolation. Start from the transfer examples.

The physical-transfer guide covers source selection, fitting, guards, accuracy scope, and non-goals.

Quick start: calibrate a measurement

Use AffineTransform when a measurement is already converted:

use ph_curves::AffineTransform;

// y' = (y * 1_005 - 120_000) / 1_000
let trim = AffineTransform::new(1_005, -120_000, 1_000).unwrap();
let corrected_milli_celsius = trim.apply(25_000).unwrap();
let original = trim.unapply(corrected_milli_celsius).unwrap();
assert!((original - 25_000).abs() <= 1);

Apply calibration before mutating temporal state so an affine overflow cannot insert a sample into a filter window. AffineCalibration<T> provides the same arithmetic around a TransferFunction and supports inverse conversion when the inner transfer does.

Quick start: stabilize and decide

This policy smooths already-converted unsigned measurements, independently classifies stability, and changes the latch only while stable:

use ph_curves::{
    Hysteresis, MovingAverage, Stability, StabilityDetector, TemporalFilter,
};

let mut average = MovingAverage::<u32, 4>::new();
let mut settled = StabilityDetector::<u32, 3>::new(5_000);
let mut high = Hysteresis::<u32>::new(900_000, 1_000_000);
let mut high_light = false;

for micro_lux in [
    1_010_000, 1_006_000, 1_004_000, 1_002_000, 1_001_000, 999_000,
] {
    let Some(smoothed) = average.update(micro_lux).ready() else {
        continue;
    };
    if matches!(settled.update(smoothed), Stability::Stable { .. }) {
        high_light = high.update(smoothed);
    }
}

assert!(high_light);

Warm-up, missing samples, invalid samples, reset behavior, cadence, and whether to hold or clear a decision during instability are caller policy. The filters own fixed, const-generic state only.

The measurement-pipeline guide compares the primitives, their fixed state and update cost, warm-up behavior, and caller-owned reset policy.

Quick start: schedule without polling

Any monotonic curve can produce deadlines for its quantized output changes:

use ph_curves::{Rounding, Tickless};

include!("curves.rs");

let schedule = EASE_IN_QUAD.tickless_schedule(
    0,                 // segment start, milliseconds
    1_000,             // duration
    0,                 // start value
    255,               // end value
    10,                // output quantum
    Rounding::Nearest,
    0,                 // minimum deadline spacing
);

for deadline in schedule.iter(0) {
    set_timer(deadline.deadline_ms);
    set_output(deadline.current_val);
}

Timestamps are wrapping u32 milliseconds. Compare deadlines with wrapping remaining time rather than absolute numeric ordering across clock rollover.

The tickless-scheduling guide explains rollover-safe comparisons, duration bounds, quantization, and repeat modes.

Quick start: generate from build.rs

[build-dependencies]
ph-curves = { version = "0.3", features = ["gen-lib"] }

// build.rs
use std::{env, path::PathBuf};
use ph_curves::r#gen::{generate_to_path, GenerateOptions};

fn main() {
    let out = PathBuf::from(env::var_os("OUT_DIR").unwrap()).join("curves.rs");
    generate_to_path("assets/curves.toml", &out, &GenerateOptions::default())
        .expect("generate curves");
    println!("cargo:rerun-if-changed=assets/curves.toml");
}

// Firmware target; no host feature is enabled here.
include!(concat!(env!("OUT_DIR"), "/curves.rs"));

The library and CLI share the same parse, validation, fitting, and emission pipeline, so the same input produces the same Rust source.

Quick start: model a transfer family

Transfer families describe discrete hardware variants that share one source model without pretending selector values are interpolated dimensions. Start with the complete family acceptance document:

ph-curves-gen \
  --input assets/family-acceptance.toml \
  --output src/front_end_transfers.rs

Each expected typed-selector identity must appear exactly once as an emitted or description-only member, or as an explicit family gap with a reason. Stable emitted names and the emission manifest identify firmware symbols before fitting; structured provenance, resource budgets, and generation reports audit the fitted result before flashing.

Every family TOML document declares:

[transfers]
requires = ["transfer_families_v1"]

That marker makes released 0.2.1 readers reject the document instead of silently ignoring the family table. See the transfer-family guide and host IR guide for the schema, programmatic FamilySpec construction, overlays, reports, and provenance rules.

How the pieces compose

The common measurement path is:

u16 observation → transfer → optional affine calibration → optional filter → stability classification → optional hysteretic/debounced decision

Each stage remains independent:

• A transfer converts one observation and owns no temporal state.
• Affine calibration applies caller-provided coefficients and owns no NVM.
• A filter changes a value but does not declare it stable.
• A stability detector classifies its own recent window.
• Hysteresis and debounce change state only when the caller updates them.
• Tickless scheduling maps a time-varying curve to deadlines; it does not own a clock or timer.

Ordering is intentional. Nonlinear transfer and filtering do not commute, and even affine correction can differ across integer stages because each stage rounds. Choose the domain in which thresholds and spans should be expressed.

Guarantees and limits

• Pure runtime: no hardware access, allocation, hidden I/O, interrupts, clocks, async runtime, or device lifecycle.
• Integer firmware: floating-point formulas, models, fitting, and error analysis are host-only.
• Bounded numerical error, not sensor accuracy: separately account for sensor tolerance, ADC/reference error, self-heating, wiring, and calibration uncertainty.
• Transfer shape: one static monotonic u16 input to one i32 output. Multidimensional compensation, nonmonotonic maps, fusion, and state estimation belong in application or domain-specific crates.
• Caller-owned policy: units, acquisition, cadence, missing/invalid sample handling, reset, calibration storage, and hardware action remain outside the crate.
• Complete-domain dense LUTs: u8 uses 256 entries and u16 uses 65,536. A full u16 LUT requires a pointer width of at least 32 bits; use a sparse transfer or smaller domain on 16-bit-pointer targets.
• Fail-closed host schema: family, standalone guard, and provenance capabilities prevent older generators from silently discarding safety or identity information.

The complete migration contract is in docs/compatibility.md. The documentation map routes from each use case to the relevant guide, design record, or API reference.

API map

Everything in the runtime is re-exported at the crate root. Detailed contracts, errors, panic conditions, and compiled examples live on docs.rs.

Reference inputs

• Normalized curves
• Full-domain u16 curves
• NTC transfer
• Custom formula and point transfers
• Observation guards
• Transfer-family acceptance document

Minimum supported Rust version

Rust 1.92.0 (edition 2024).

Contributing

Contributions are welcome. Read CONTRIBUTING.md, the security policy, and the code of conduct before participating. Run the same gate as CI with:

./scripts/local-ci.ps1

License

MIT