nncc 2.0.3

Neural Network to C Compiler

nncc - Neural Network to C Compiler

A minimal, dependency-free compiler that converts PyTorch models into portable C99 code for inference.

Features

• Python-First Workflow: Compile directly from your Python training script using nncc.compile().
• Zero-Dependency Inference: Generated C code only requires <math.h> and <stdlib.h>.
• Embedded Weights: Weights are embedded directly into the generated C header/source, making it a single-file solution.
• Dynamic & Portable: Supports arbitrary model architectures and runs on any C99-compliant compiler (GCC, Clang, MSVC).

Installation

pip install nncc

Quick Start

1. Compile directly from PyTorch

import torch
import torch.nn as nn
import nncc

# 1. Define your model
model = nn.Sequential(
    nn.Linear(10, 64),
    nn.ReLU(),
    nn.Linear(64, 2),
    nn.Softmax(dim=-1)
)

# 2. Compile to C source code
# We pass a sample input to infer shapes
c_code = nncc.compile(model, torch.randn(1, 10), name="my_model")

# 3. Save to file
with open("model.c", "w") as f:
    f.write(c_code)

print("Saved model.c!")

2. Use the generated C code

The generated file is a single-header library. You can include it directly in your C application.

// main.c
#define MODEL_IMPLEMENTATION // Define this in exactly one .c file to implement functions
#include "model.c"

#include <stdio.h>

int main() {
    // Input matches the shape provided during compilation (1, 10)
    float input[10] = {0.5f, -0.2f, /* ... */};
    float output[2];
    
    // Run inference
    my_model_inference(input, output);
    
    printf("Class 0: %f\n", output[0]);
    printf("Class 1: %f\n", output[1]);
    
    return 0;
}

Compile it with standard tools:

gcc -O3 main.c -o app -lm

Supported Layers

Category	Operations
Linear	Linear (GEMM)
Conv	Conv2d
Pool	MaxPool2d, AvgPool2d, AdaptiveAvgPool2d (Global)
Act	ReLU, ReLU6,, Sigmoid, Tanh, Softmax
Norm	BatchNorm2d
Shape	Flatten, Dropout (No-op)

Advanced / Manual Usage

CLI Tool

If you prefer the command-line interface or want to work with raw files:

1. Export Model: Use scripts/export_model.py to generate .safetensors and .nnmodel files.
2. Compile:

   nncc model.safetensors model.nnmodel -o model.c
   

Building from Source

To build the nncc compiler and shared library from source:

Windows (MinGW/Git Bash):

make

Windows (MSVC):

cl /O2 /Fe:nncc.exe src/main.c src/model.c src/loader.c src/codegen.c

Linux/macOS:

make

Project Structure

nncc/
├── nncc/              # Python package
│   ├── __init__.py
│   ├── compiler.py    # Python API & Ctypes bindings
│   └── nncc.dll       # Pre-compiled core
├── src/               # C Compiler Core
│   ├── libnncc.c      # Shared library entry point
│   ├── main.c         # CLI entry point
│   ├── model.c        # Graph IR
│   └── codegen.c      # C99 Generator
├── examples/          # Usage examples
└── setup.py           # Packaging config

License

MIT License