https://github.com/Conqueror712/CUDA-Simulator

A self-developed version of the user-mode CUDA emulator project and a learning repository for Rust
https://github.com/Conqueror712/CUDA-Simulator

Last synced: 3 months ago
JSON representation

A self-developed version of the user-mode CUDA emulator project and a learning repository for Rust

• Host: GitHub
• URL: https://github.com/Conqueror712/CUDA-Simulator
• Owner: Conqueror712
• License: mit
• Created: 2023-07-02T09:02:25.000Z (over 1 year ago)
• Default Branch: main
• Last Pushed: 2023-09-22T02:34:52.000Z (over 1 year ago)
• Last Synced: 2024-08-04T02:06:48.769Z (6 months ago)
• Language: Rust
• Homepage:
• Size: 35.5 MB
• Stars: 3
• Watchers: 1
• Forks: 2
• Open Issues: 0
• Metadata Files:
  • Readme: README.md
  • License: LICENSE

Awesome Lists containing this project

• awesome-cuda-triton-hpc - Conqueror712/CUDA-Simulator - Simulator?style=social"/> : A self-developed version of the user-mode CUDA emulator project and a learning repository for Rust. (Frameworks)
• awesome-cuda-triton-hpc - Conqueror712/CUDA-Simulator - Simulator?style=social"/> : A self-developed version of the user-mode CUDA emulator project and a learning repository for Rust. (Frameworks)
• awesome-rust-list - Conqueror712/CUDA-Simulator - Simulator?style=social"/> : A self-developed version of the user-mode CUDA emulator project and a learning repository for Rust. (GPU Programming)
• awesome-rust-list - Conqueror712/CUDA-Simulator - Simulator?style=social"/> : A self-developed version of the user-mode CUDA emulator project and a learning repository for Rust. (GPU Programming)

README

        
# 一、Linux环境选择

首先准备一个带NVIDIA GPU的Linux环境；

笔者这里用的是WSL，N卡是原本Windows上带的。

*   Ubuntu 20.04 LTS

*   RTX 1650

> 尝试过的其他方法：

>

> *   Windows环境（兼容性问题）

> *   Vmware虚拟机环境（虚拟化过程繁琐）

> *   单主机双系统（切换不便，卡顿）

> *   云服务器（成本问题，有些驱动版本需要更新）

> *   Mac环境（不支持CUDA）

# 二、NVIDIA驱动安装

直接用Windows下载.exe文件安装；

另外，安装过程会伴随多次黑屏和重启，并且会需要一段时间，请耐心等待，会自动同步到WSL中；

> 如果没同步就请`apt install nvidia-utils-535-server`

值得一提的是，安装之前需要把旧版的带NVIDIA的软件全部卸载（除了NVIDIA Control Panel）

![image.png](https://p1-juejin.byteimg.com/tos-cn-i-k3u1fbpfcp/b3d5beaf549e4132992979a706d3110f~tplv-k3u1fbpfcp-watermark.image?)

作为参考，笔者的驱动版本是536.67

> check: `nvidia-smi`

>

> ![image.png](https://p3-juejin.byteimg.com/tos-cn-i-k3u1fbpfcp/e1e603154d6e442e8d38a39790e2d95f~tplv-k3u1fbpfcp-watermark.image?)

# 三、更新依赖

    sudo apt update

    sudo apt-get update

    sudo apt-get install libclang-dev

    sudo apt-get install freeglut3-dev build-essential libx11-dev libxmu-dev libxi-dev libgl1-mesa-glx libglu1-mesa libglu1-mesa-dev

获取权限（可选）：`sudo su`

# 四、CUDA 12.2下载安装：

## 方式一：

推荐基于发行版的文档进行安装，这里是Ubuntu，参考文档如下：

## 方式二：

直接安装：

    wget https://developer.download.nvidia.com/compute/cuda/repos/wsl-ubuntu/x86_64/cuda-keyring_1.1-1_all.deb

    sudo dpkg -i cuda-keyring_1.1-1_all.deb

    sudo apt-get update

    sudo apt-get -y install cuda

更新环境变量`nano ~/.bashrc` --> 以下操作 --> `source ~/.bashrc`

> 注意，这一步如果原本机器上没有旧版的CUDA就不用做

    export CUDA_HOME=/usr/local/cuda

    export PATH=$PATH:$CUDA_HOME/bin

    export LD_LIBRARY_PATH=/usr/local/cuda-12.2/lib64${LD_LIBRARY_PATH:+:${LD_LIBRARY_PATH}}

注：退出nano的方法可以是`Ctrl + X, Y, Enter`（三步）

> check: `nvcc --version`

>

> ![image.png](https://p3-juejin.byteimg.com/tos-cn-i-k3u1fbpfcp/75208dd1dbe7446196a243e56dc4e58c~tplv-k3u1fbpfcp-watermark.image?)

# 五、Rust & Cargo下载

`curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh`之后输入`1`；

随后重新加载`source $HOME/.cargo/env`；

> check: `cargo --version` && `rustc --version`

>

> ![image.png](https://p3-juejin.byteimg.com/tos-cn-i-k3u1fbpfcp/47b0d93c12c1454b9d20fda8e583febd~tplv-k3u1fbpfcp-watermark.image?)

***

# 六、克隆仓库 + 环境变量

`git clone https://github.com/Conqueror712/CUDA-Simulator.git`

配置环境变量，用nano写入`nano ~/.bashrc`，写入后保存`source ~/.bashrc`

    export PATH=/usr/local/cuda-12.2/bin${PATH:+:${PATH}}

    export LD_LIBRARY_PATH=/usr/local/cuda-12.2/lib64${LD_LIBRARY_PATH:+:${LD_LIBRARY_PATH}}

    export LD_LIBRARY_PATH=/usr/local/cuda-12.2/lib64:$LD_LIBRARY_PATH

    export PATH=/usr/local/cuda-12.2/bin:$PATH

    export LIBCLANG_PATH=/usr/lib/x86_64-linux-gnu/

    export BINDGEN_EXTRA_CLANG_ARGS="-I /usr/local/cuda-12.2/include"

# 七、生成动态链接库

于项目根目录下`cargo build --release`，生成libcuda.so文件，默认在target/release/

![image.png](https://p3-juejin.byteimg.com/tos-cn-i-k3u1fbpfcp/06e88351e7f24d39829cdadef7112cbf~tplv-k3u1fbpfcp-watermark.image?)

# 八、简单测试

1.  cd进入smoketest，用`nvcc smoketest.cu`编译得到`a.out`，再运行`./a.out`

    ![image.png](https://p9-juejin.byteimg.com/tos-cn-i-k3u1fbpfcp/bf7c94df9cd948fa84fd41f95d14b921~tplv-k3u1fbpfcp-watermark.image?)

2.  无错误之后方可`LD_PRELOAD=/home//CUDA-Simulator/cargo_demo/target/release/libcuda.so ./a.out`查看trace（已经实现的会显示..... --> CUDA\_SUCCESS）

    ![image.png](https://p3-juejin.byteimg.com/tos-cn-i-k3u1fbpfcp/304ecd2ad37e4ea79a58f14177393fab~tplv-k3u1fbpfcp-watermark.image?)

3.  `cargo doc --open`在浏览器中打开文档查看函数签名

    ![image.png](https://p1-juejin.byteimg.com/tos-cn-i-k3u1fbpfcp/eadc237df6f9445999b22761abab2b88~tplv-k3u1fbpfcp-watermark.image?)

4.  之后往`lib.rs中添加自己的实现即可`

***

FIN