Embedded C99 & TinyML

Embedded C99 & TinyML

Sofia Engine provides a standalone, production-grade ISO C99 reference implementation under embedded/ for deployment on microcontrollers (MCUs), digital signal processors (DSPs), and resource-constrained edge nodes where a Python runtime is not viable.

---

1. Architectural Principles of the C99 Engine

The embedded runtime is engineered to satisfy the strict constraints of embedded systems:

1. Zero Dynamic Allocation Post-Initialization:
   • malloc() and free() are strictly forbidden during streaming execution.
   • All scratchpad buffers and state structures are caller-allocated or statically bound in sofia_features_init().
2. Fixed Memory Footprint:
   • Total RAM requirement is predictable at compile time (typically $< 32 \text{ KB}$ for 1024-sample FFT/feature buffers).
3. MISRA-C & ISO C99 Conformance:
   • No recursion, no variable-length arrays (VLAs), no unbounded loops, and all pointers are validated before dereferencing.
4. Platform Independence:
   • Zero vendor-specific peripheral code. Standard ANSI C99 library (math.h, stdint.h, stdbool.h).

---

2. Directory Layout & Key Files

embedded/
├── include/
│   ├── sofia_features.h     # Public C API for statistical & spectral extractors
│   └── sofia_fixed.h        # Q16.16 fixed-point arithmetic engine
├── src/
│   ├── sofia_features.c     # Core DSP implementation (RMS, Kurtosis, FFT, Bands)
│   └── sofia_fixed.c        # Fixed-point sqrt, log, and trigonometric primitives
├── tests/
│   └── test_sofia_features.c# Golden-vector test harness
├── Makefile                 # Standalone GCC/Clang build system
└── CMakeLists.txt           # Cross-platform CMake build configuration

---

3. Fixed-Point Arithmetic (Q16.16)

For microcontrollers lacking a hardware Floating Point Unit (FPU) (such as ARM Cortex-M0+, Cortex-M3, or low-power RISC-V cores), sofia_fixed.h provides a full Q16.16 fixed-point math library:

• Representation: 1 sign bit, 15 integer bits, 16 fractional bits.
• Resolution: $\Delta = 2^{-16} \approx 0.00001525878$.
• Dynamic Range: $[-32768.0, +32767.99998]$.

Essential Macros & Operations

#include "sofia_fixed.h"

// Float <-> Fixed conversion
q16_t q_val = SOFIA_F2Q(14.85f);
float f_val = SOFIA_Q2F(q_val);

// Multiplication with 32-bit overflow protection
q16_t prod = sofia_q16_mul(a, b);

// Fixed-point Square Root (Newton-Raphson approximation)
q16_t root = sofia_q16_sqrt(q_val);

---

4. Golden Vector Verification

To guarantee that features computed on microcontrollers match Python outputs with byte-level precision, Sofia uses a golden-vector testing strategy:

flowchart LR
    PY["Python Core\n(features/extractor.py)"] -->|Generate Reference Vectors| GV["golden_vectors.json\n(Synthetic & Edge Telemetry)"]
    GV -->|Load in CI| C99["C99 Test Suite\n(test_sofia_features.c)"]
    C99 -->|Compare Tolerances| PASS["Validation Pass\n(Floating: ε <= 1e-4\nFixed: ε <= 1e-2)"]

Loading

Running the C99 Tests

cd embedded
make test

The test runner asserts that all statistical features (RMS, Peak, Kurtosis, Crest Factor) and spectral band energies agree within strict numerical bounds.

---

5. Hardware Target & Porting Guidance

Deployment Class	Target Platform	Verification Status	Recommended Backend
Class A: Server / Gateway	x86_64 / ARM64 Linux, Raspberry Pi 4/5	VERIFIED	Full Python sofia-engine + NumPy
Class B: Industrial Edge	NXP i.MX8, STM32MP1, Yocto RTOS	PARTIALLY VERIFIED	Python core or C99 static library
Class C: Microcontroller	STM32 (Cortex-M4/M7), ESP32, nRF5340	NOT VERIFIED in CI	C99 reference runtime (embedded/src)

Note: "NOT VERIFIED in CI" indicates that while the C99 code compiles and passes host unit tests, Rootcastle does not maintain automated hardware-in-the-loop (HIL) testing rigs for every physical silicon vendor in continuous integration.