YOLO26 Inference (ARM64 Assembly + C)

YOLO26 object detection inference written from scratch in ARM64 Assembly and C, optimized for the Raspberry Pi 4 (ARM Cortex-A72).

Bachelor's Final Project
Author: Mohammad Ghaderi
Advisor: Prof. Saleh Yousefi

---

Project Overview

This project implements the inference stage of the YOLO26n object detection model entirely from scratch using C and ARM64 Assembly. The main focus is on understanding and reproducing the low-level optimizations employed by modern ARM inference engines.

Key optimization areas include:

• ARM NEON SIMD vectorization
• Cache-aware computation
• Winograd convolution
• Optimized GEMM kernels
• Operator fusion
• ARM64 micro-kernel design

---

YOLO26 Architecture

The implementation follows the YOLO26 architecture and focuses on efficient execution of convolutional layers on ARM Cortex-A72 processors.

---

Features

• ARM64 (AArch64) optimized
• ARM64 Assembly + C
• ARM NEON vectorization
• Float32 inference
• SiLU/sigmoid/softmax activation ^{(NCNN-inspired implementation - implementing exp)}
• Winograd F(2×3) transform ^{(using NCNN transform matrix values)}
• GEMM convolution
• Pointwise & Depthwise convolution
• Cache-aware tiling
• Custom ARM micro-kernels
• Operator fusion
• Multiple kernel and tile-size experiments
• Attention
• Components (C3K2, SPPF, Conv, C2PSA, Detect, BottleNeck, PSA)
• read jpg and save result as png with bounding boxes

---

Performance

All benchmarks were performed on:

• Raspberry Pi 4 Model B / ARM Cortex-A72 /ARMv8-A (64-bit) / Linux

Example

_{This is the output of the implementation. It surprised me—you can test it too.}

Build & Run

Compile the project : ^{( same for arm or x86 on QEMU emulator)}

If ARM64 hardware is not available, the project can be executed through QEMU. The provided Makefile supports both native ARM64 execution and QEMU-based execution using the same commands. Go To Install QEMU emulator

First, clone the repository:

git clone https://github.com/mohammad-ghaderi/YOLO26.git
cd YOLO26

Build

make

Run inference on an image: _{input image should be 640*640 jpg}

make run IMG=path/to/image.jpg

Example:

make run IMG=images/img1.jpg

Install QEMU Emulator

Ubuntu / Debian:

sudo apt update
sudo apt install qemu-user qemu-user-static

Arch Linux:

sudo pacman -S qemu-user

Fedora:

sudo dnf install qemu-user-static

Acknowledgments

• During development, several experimental implementations were tested. Some of these (e.g., QS8, FP16, etc.) remain in the repository but are not part of the final inference pipeline.
• The parameters from the official YOLO26n pretrained model were extracted, restructured, and repacked into a custom binary format (param.bin). The ordering and layout were redesigned to match the optimized memory access pattern of this ARM64 inference engine.

---