MiniNN

Build a deep learning inference framework from scratch

Quick Start

envirooment requirement：

• conda，python=3.11
• mininn_test.gynn download

conda create -n mininn python=3.11
conda activate mininn
pip install mininn
python python/test/pip_test.py

Build from source

Environment Version

Host	Target	Tool	Compiler	Backend	Note
Windows	Windows	cmake 3.26.4	msvc 14.44	cuda✅ opencl✅ avx✅ sse✅ mkl✅	use VS shell
			clang 20.1.5	cuda❌ opencl✅ avx✅ sse✅ mkl✅
			g++ 13.1.0	cuda❌ opencl✅ avx✅ sse✅ mkl✅
		bazel 7.3.1	msvc 14.44	cuda✅ opencl✅ avx✅ sse✅ mkl✅	default msvc
			clang 20.1.5
			g++ 13.1.0
Linux	Linux	cmake 3.16.3	clang 10.0.0	cuda✅ opencl✅ avx✅ sse✅ mkl✅
			g++ 9.4.0	cuda✅ opencl✅ avx✅ sse✅ mkl✅
		bazel 8.0.0	clang 10.0.0
			g++ 9.4.0	cuda✅ opencl✅ avx✅ sse✅ mkl❌	default g++

Download

# https
git clone --recursive https://github.com/masteryi-0018/MiniNN.git

# ssh
git clone --recursive [email protected]:masteryi-0018/MiniNN.git

Build

# cmake
python build.py

# bazel
python build.py --tool bazel

# if you want to identify generator, add this flag
python build.py --tool bazel --generator ninja

# if you want to identify compiler, add this flag
python build.py --tool bazel --compiler clang

# if you want build python wheel, add this flag
python build.py --tool bazel --wheel

Unit Test

windows

# cmake
.\build\mininn\test\gtest-main.exe

# bazel
.\bazel-bin\mininn\gtest-main.exe

linux

# cmake
./build/mininn/test/gtest-main

# bazel
./bazel-bin/mininn/gtest-main

TODO

1. mininn convertor

• 支持将 onnx 模型转换为 gynn 格式
• 支持将 pytorch 模型转换为 gynn 格式
• 支持将 tensorflow 模型转换为 gynn 格式

2. mininn IR

• 支持多算子构图

3. mininn kernel

• 增加 opencl 后端
• 增加 cuda 后端
• 增加 avx 后端
• 增加 sse 后端
• 增加 mkl 后端

4. mininn release

• demo
• cpp sdk
• pip package

5. mininn build

• cmake
• bazel
• clang++
• g++
• msvc