godweiyang/NN-CUDA-Example
godweiyang/NN-CUDA-Example has 1,535 GitHub stars and 206 forks. GitHot measured 0 stars in 24 hours, -2 in 7 days, and -2 in 15 days. Last checked 2026-11-21 20:44 UTC. Last push 2021-04-29.
計測したスター増加
- 24時間
- 0
- 7日間
- -2
- 15日間
- -2
- 1か月
- 履歴を収集中
期間ごとの純増数です。期間は重複しており、日別の値ではありません。未計測の値は空欄です。スター数は注目度の指標であり、品質の保証ではありません。
概要
Several simple examples for popular neural network toolkits calling custom CUDA operators.
機械学習README.md
オリジナルNeural Network CUDA Example

Several simple examples for neural network toolkits (PyTorch, TensorFlow, etc.) calling custom CUDA operators.
We provide several ways to compile the CUDA kernels and their cpp wrappers, including jit, setuptools and cmake.
We also provide several python codes to call the CUDA kernels, including kernel time statistics and model training.
For more accurate time statistics, you'd best use nvprof or nsys to run the code.
Environments
- NVIDIA Driver: 418.116.00
- CUDA: 11.0
- Python: 3.7.3
- PyTorch: 1.7.0+cu110
- TensorFlow: 2.4.1
- CMake: 3.16.3
- Ninja: 1.10.0
- GCC: 8.3.0
Cannot ensure successful running in other environments.
Code structure
├── include
│ └── add2.h # header file of add2 cuda kernel
├── kernel
│ └── add2_kernel.cu # add2 cuda kernel
├── pytorch
│ ├── add2_ops.cpp # torch wrapper of add2 cuda kernel
│ ├── time.py # time comparison of cuda kernel and torch
│ ├── train.py # training using custom cuda kernel
│ ├── setup.py
│ └── CMakeLists.txt
├── tensorflow
│ ├── add2_ops.cpp # tensorflow wrapper of add2 cuda kernel
│ ├── time.py # time comparison of cuda kernel and tensorflow
│ ├── train.py # training using custom cuda kernel
│ └── CMakeLists.txt
├── LICENSE
└── README.md
PyTorch
Compile cpp and cuda
JIT
Directly run the python code.
Setuptools
python3 pytorch/setup.py install
CMake
mkdir build
cd build
cmake ../pytorch
make
Run python
Compare kernel running time
python3 pytorch/time.py --compiler jit
python3 pytorch/time.py --compiler setup
python3 pytorch/time.py --compiler cmake
Train model
python3 pytorch/train.py --compiler jit
python3 pytorch/train.py --compiler setup
python3 pytorch/train.py --compiler cmake
TensorFlow
Compile cpp and cuda
CMake
mkdir build
cd build
cmake ../tensorflow
make
Run python
Compare kernel running time
python3 tensorflow/time.py --compiler cmake
Train model
python3 tensorflow/train.py --compiler cmake
Implementation details (in Chinese)
PyTorch自定义CUDA算子教程与运行时间分析
详解PyTorch编译并调用自定义CUDA算子的三种方式
三分钟教你如何PyTorch自定义反向传播
F.A.Q
Q. ImportError: libc10.so: cannot open shared object file: No such file or directory
A. You must doimport torchbeforeimport add2.
Q. tensorflow.python.framework.errors_impl.NotFoundError: build/libadd2.so: undefined symbol: _ZTIN10tensorflow8OpKernelE
A. Check if${TF_LFLAGS}inCmakeLists.txtis correct.
関連プロジェクトを探す
読み込み中…