Mars: Open-Source Neural Network Runtime for Ingenic T41

Hardware-accelerated inference on IP cameras using the MXUv3 SIMD unit and ORAM.

🚀 Performance

Metric	Before (Scalar)	After (Mars)	Speedup
Inference Time	35 seconds	1.75 seconds	20x
Memory Read	41 MB/s (DDR)	314 MB/s (ORAM)	7.6x
Memory Write	77 MB/s (DDR)	1578 MB/s (ORAM)	20.6x

Overview

Mars is an open-source neural network runtime for the Ingenic T41 SoC, reverse-engineered from Ingenic's proprietary Venus SDK. It enables hardware-accelerated inference on IP cameras running Thingino firmware.

Why Mars?

• 🔓 Open Source: No proprietary SDKs or closed toolchains
• 🐧 musl Compatible: Works with Thingino's musl libc (Venus requires glibc)
• ⚡ Hardware Accelerated: Uses MXUv3 SIMD (512-bit) and on-chip ORAM
• 🎯 Purpose-Built: Custom TinyDet model for security camera use cases

Features

• ✅ MXUv3 SIMD acceleration (16 floats per instruction)
• ✅ ORAM weight staging (640KB on-chip, 7.6x faster than DDR)
• ✅ Conv2D, ReLU, MaxPool, Add, Concat operations
• ✅ NHWC tensor format (optimized for T41 memory access)
• ✅ Custom .mars model format
• ✅ ONNX → Mars compiler (Python + Rust)

Quick Start

Building

# Set cross-compiler (adjust path to your toolchain)
export CROSS_COMPILE=/path/to/mipsel-linux-

# Build runtime library and tools
make

# Output:
#   build/lib/libmars.so    - Runtime library
#   build/bin/mars_detect   - Detection CLI tool

Running on Device

# Copy to camera
scp build/bin/mars_detect build/lib/libmars.so root@camera:/opt/

# Run detection
ssh root@camera
cd /opt
LD_LIBRARY_PATH=/opt ./mars_detect model.mars input.jpg output.jpg

Compiling Models

cd mars-compiler

# Stage 1: ONNX → JSON + weights
python3 onnx2mars.py model.onnx -o model

# Stage 2: JSON → .mars binary
cargo run -- -i model.json -o model.mars --float32

Architecture

thingino-accel/
├── src/mars/              # Mars runtime (C)
│   ├── mars_runtime.c     # Model loader and executor
│   ├── mxu_conv.c         # MXUv3 convolution kernels
│   └── mars_nn_hw.c       # Hardware initialization (ORAM, MXU)
├── mars-compiler/         # ONNX → Mars compiler
│   ├── onnx2mars.py       # Python: ONNX → JSON extraction
│   └── src/               # Rust: JSON → .mars binary
├── training/              # TinyDet model training
│   ├── tinydet.py         # Model architecture
│   └── train_*.py         # Training scripts
├── include/               # Public headers
└── docs/                  # Documentation
    └── MARS_PROJECT_WRITEUP.md  # Full research paper

TinyDet: Custom Detection Model

We trained a purpose-built 4-class detector optimized for security cameras:

Class	Description
Person	Human detection
Vehicle	Cars, trucks
Cat	Feline pets
Dog	Canine pets

Model specs:

• Input: 320×192 RGB (NHWC)
• Parameters: ~202K
• Architecture: Anchor-free, single-stage
• Training: COCO + Oxford Pets datasets

Hardware Details

Ingenic T41 SoC

• CPU: Dual XBurst2 @ 1.5GHz (MIPS)
• MXUv3: 512-bit SIMD, 32 VPR registers
• ORAM: 640KB @ 0x12640000 (on-chip SRAM)
• NNA: Neural Network Accelerator with NNDMA

Memory Performance

Region	Bandwidth	Latency
DDR	41 MB/s read	High
ORAM	314 MB/s read	Low

Weights are staged to ORAM before convolution for maximum throughput.

Documentation

• 📄 Full Project Writeup - Research paper covering reverse engineering, MXUv3 discovery, and model design
• 📘 Mars Runtime README - Runtime architecture and API
• 🎓 Training README - Model training guide

Comparison with Venus SDK

Feature	Venus (OEM)	Mars
License	Proprietary	GPLv3
C Library	glibc only	musl/glibc
Model Format	.mgk (closed)	.mars (open)
Source Code	No	Yes
Compiler	Closed	Python + Rust

Contributing

Contributions welcome! See the project writeup for technical background.

License

GPLv3 - See LICENSE for details.

Acknowledgments

• Thingino - Open-source IP camera firmware
• OpenSensor - Project home