ZLUDA丨如何在Intel的GPU上运行CUDA代码？

0270

零、前言：

此demo的Github仓库（持续更新）：github.com/Conqueror71…

?欢迎访问：

个人博客：conqueror712.github.io/

知乎：www.zhihu.com/people/soeu…

Bilibili：space.bilibili.com/57089326

掘金：juejin.cn/user/129787…

有任何疏忽和错误欢迎各位读者指出！

一、一切从C++开始！

首先我们要得到一份C++代码，功能很简单，就是做加法：

注意这是Linux版本

#include <iostream>
#include <math.h>
#include <sys/time.h>
​
// function to add the elements of two arrays
void add(int n, float *x, float *y){
  for (int i = 0; i < n; i++)
      y[i] = x[i] + y[i];
}
​
int main(){
  int N = 1<<25; // 30M elements
​
  float *x = new float[N];
  float *y = new float[N];
  // initialize x and y arrays on the host
  for (int i = 0; i < N; i++) {
    x[i] = 1.0f;
    y[i] = 2.0f;
  }
​
  struct timeval t1,t2;
  double timeuse;
  gettimeofday(&t1,NULL);
  // Run kernel on 30M elements on the CPU
  void add(N, x, y);
  gettimeofday(&t2,NULL);
  timeuse = (t2.tv_sec - t1.tv_sec) + (double)(t2.tv_usec - t1.tv_usec)/1000.0;
​
  std::cout << "add(int, float*, float*) time: " << timeuse << "ms" << std::endl;
  // Check for errors (all values should be 3.0f)
  float maxError = 0.0f;
  for (int i = 0; i < N; i++)
    maxError = fmax(maxError, fabs(y[i]-3.0f));
  std::cout << "Max error: " << maxError << std::endl;
​
  // Free memory
  delete [] x;
  delete [] y;
  
  return 0;
}
#include <iostream>


#include <math.h>


#include <sys/time.h>

​


// function to add the elements of two arrays

void add(int n, float *x, float *y){
  for (int i = 0; i < n; i++)
      y[i] = x[i] + y[i];
}
​
int main(){
  int N = 1<<25; // 30M elements
​
  float *x = new float[N];
  float *y = new float[N];
  // initialize x and y arrays on the host
  for (int i = 0; i < N; i++) {
    x[i] = 1.0f;
    y[i] = 2.0f;
  }
​
  struct timeval t1,t2;
  double timeuse;
  gettimeofday(&t1,NULL);
  // Run kernel on 30M elements on the CPU
  void add(N, x, y);
  gettimeofday(&t2,NULL);
  timeuse = (t2.tv_sec - t1.tv_sec) + (double)(t2.tv_usec - t1.tv_usec)/1000.0;
​
  std::cout << "add(int, float*, float*) time: " << timeuse << "ms" << std::endl;
  // Check for errors (all values should be 3.0f)
  float maxError = 0.0f;
  for (int i = 0; i < N; i++)
    maxError = fmax(maxError, fabs(y[i]-3.0f));
  std::cout << "Max error: " << maxError << std::endl;
​
  // Free memory
  delete [] x;
  delete [] y;
  
  return 0;
}
#include <iostream>


#include <math.h>


#include <sys/time.h>

​


// function to add the elements of two arrays

void add(int n, float *x, float *y){
  for (int i = 0; i < n; i++)
      y[i] = x[i] + y[i];
}
​
int main(){
  int N = 1<<25; // 30M elements
​
  float *x = new float[N];
  float *y = new float[N];
  // initialize x and y arrays on the host
  for (int i = 0; i < N; i++) {
    x[i] = 1.0f;
    y[i] = 2.0f;
  }
​
  struct timeval t1,t2;
  double timeuse;
  gettimeofday(&t1,NULL);
  // Run kernel on 30M elements on the CPU
  void add(N, x, y);
  gettimeofday(&t2,NULL);
  timeuse = (t2.tv_sec - t1.tv_sec) + (double)(t2.tv_usec - t1.tv_usec)/1000.0;
​
  std::cout << "add(int, float*, float*) time: " << timeuse << "ms" << std::endl;
  // Check for errors (all values should be 3.0f)
  float maxError = 0.0f;
  for (int i = 0; i < N; i++)
    maxError = fmax(maxError, fabs(y[i]-3.0f));
  std::cout << "Max error: " << maxError << std::endl;
​
  // Free memory
  delete [] x;
  delete [] y;
  
  return 0;
}

添加一些有趣的东西就可以变成CUDA代码：

  • __global__
  • cudaMallocManaged(&x, N*sizeof(float)); cudaMallocManaged(&y, N*sizeof(float));
  • add<<<1, 1>>>(N, x, y);
  • cudaFree(x); cudaFree(y);

1. Linux版本：

#include <iostream>
#include <math.h>
#include <sys/time.h>
​
// function to add the elemsents of two arrays
__global__
void add(int n, float *x, float *y){
  for (int i = 0; i < n; i++)
      y[i] = x[i] + y[i];
}
​
int main(){
  int N = 1<<25; // 30M elements
​
  float *x = new float[N];
  float *y = new float[N];
  cudaMallocManaged(&x, N*sizeof(float));
  cudaMallocManaged(&y, N*sizeof(float));
  // initialize x and y arrays on the host
  for (int i = 0; i < N; i++) {
    x[i] = 1.0f;
    y[i] = 2.0f;
  }
​
  struct timeval t1,t2;
  double timeuse;
  gettimeofday(&t1,NULL);
  // Run kernel on 30M elements on the CPU
  add<<<1, 1>>>(N, x, y);
  gettimeofday(&t2,NULL);
  timeuse = (t2.tv_sec - t1.tv_sec) + (double)(t2.tv_usec - t1.tv_usec)/1000.0;
​
  std::cout << "add(int, float*, float*) time: " << timeuse << "ms" << std::endl;
  // Check for errors (all values should be 3.0f)
  float maxError = 0.0f;
  for (int i = 0; i < N; i++)
    maxError = fmax(maxError, fabs(y[i]-3.0f));
  std::cout << "Max error: " << maxError << std::endl;
​
  // Free memory
  delete [] x;
  delete [] y;
​
  cudaFree(x);
  cudaFree(y);
  
  return 0;
}
#include <iostream>


#include <math.h>


#include <sys/time.h>

​


// function to add the elemsents of two arrays
__global__

void add(int n, float *x, float *y){

  for (int i = 0; i < n; i++)

      y[i] = x[i] + y[i];

}

​

int main(){

  int N = 1<<25; // 30M elements

​

  float *x = new float[N];

  float *y = new float[N];

  cudaMallocManaged(&x, N*sizeof(float));

  cudaMallocManaged(&y, N*sizeof(float));

  // initialize x and y arrays on the host

  for (int i = 0; i < N; i++) {

    x[i] = 1.0f;

    y[i] = 2.0f;

  }

​

  struct timeval t1,t2;
  double timeuse;

  gettimeofday(&t1,NULL);
  // Run kernel on 30M elements on the CPU
  add<<<1, 1>>>(N, x, y);
  gettimeofday(&t2,NULL);
  timeuse = (t2.tv_sec - t1.tv_sec) + (double)(t2.tv_usec - t1.tv_usec)/1000.0;
​
  std::cout << "add(int, float*, float*) time: " << timeuse << "ms" << std::endl;
  // Check for errors (all values should be 3.0f)
  float maxError = 0.0f;
  for (int i = 0; i < N; i++)
    maxError = fmax(maxError, fabs(y[i]-3.0f));
  std::cout << "Max error: " << maxError << std::endl;
​
  // Free memory
  delete [] x;
  delete [] y;
​
  cudaFree(x);
  cudaFree(y);
  
  return 0;
}
#include <iostream>


#include <math.h>


#include <sys/time.h>

​


// function to add the elemsents of two arrays
__global__

void add(int n, float *x, float *y){

  for (int i = 0; i < n; i++)

      y[i] = x[i] + y[i];

}

​

int main(){

  int N = 1<<25; // 30M elements

​

  float *x = new float[N];

  float *y = new float[N];

  cudaMallocManaged(&x, N*sizeof(float));

  cudaMallocManaged(&y, N*sizeof(float));

  // initialize x and y arrays on the host

  for (int i = 0; i < N; i++) {

    x[i] = 1.0f;

    y[i] = 2.0f;

  }

​

  struct timeval t1,t2;
  double timeuse;

  gettimeofday(&t1,NULL);
  // Run kernel on 30M elements on the CPU
  add<<<1, 1>>>(N, x, y);
  gettimeofday(&t2,NULL);
  timeuse = (t2.tv_sec - t1.tv_sec) + (double)(t2.tv_usec - t1.tv_usec)/1000.0;
​
  std::cout << "add(int, float*, float*) time: " << timeuse << "ms" << std::endl;
  // Check for errors (all values should be 3.0f)
  float maxError = 0.0f;
  for (int i = 0; i < N; i++)
    maxError = fmax(maxError, fabs(y[i]-3.0f));
  std::cout << "Max error: " << maxError << std::endl;
​
  // Free memory
  delete [] x;
  delete [] y;
​
  cudaFree(x);
  cudaFree(y);
  
  return 0;
}

2. Windows版本：

#include <iostream>
#include <math.h>
#include <windows.h>
​
// function to add the elements of two arrays
__global__
void add(int n, float *x, float *y){
  for (int i = 0; i < n; i++)
      y[i] = x[i] + y[i];
}
​
int main(){
  int N = 1<<25; // 30M elements
​
  float *x = new float[N];
  float *y = new float[N];
  cudaMallocManaged(&x, N*sizeof(float));
  cudaMallocManaged(&y, N*sizeof(float));
  // initialize x and y arrays on the host
  for (int i = 0; i < N; i++) {
    x[i] = 1.0f;
    y[i] = 2.0f;
  }
​
  ULONGLONG t1, t2, freq;
  double timeuse;
  QueryPerformanceFrequency((LARGE_INTEGER*)&freq);
  QueryPerformanceCounter((LARGE_INTEGER*)&t1);
  // Run kernel on 30M elements on the CPU
  add<<<1, 1>>>(N, x, y);
  QueryPerformanceCounter((LARGE_INTEGER*)&t2);
  timeuse = (double)(t2 - t1) / (double)freq * 1000.0;
​
  std::cout << "add(int, float*, float*) time: " << timeuse << "ms" << std::endl;
  // Check for errors (all values should be 3.0f)
  float maxError = 0.0f;
  for (int i = 0; i < N; i++)
    maxError = fmax(maxError, fabs(y[i]-3.0f));
  std::cout << "Max error: " << maxError << std::endl;
​
  // Free memory
  delete [] x;
  delete [] y;
​
  cudaFree(x);
  cudaFree(y);
  
  return 0;
}
#include <iostream>


#include <math.h>


#include <windows.h>
​


// function to add the elements of two arrays

__global__

void add(int n, float *x, float *y){

  for (int i = 0; i < n; i++)

      y[i] = x[i] + y[i];

}

​

int main(){

  int N = 1<<25; // 30M elements

​

  float *x = new float[N];

  float *y = new float[N];

  cudaMallocManaged(&x, N*sizeof(float));

  cudaMallocManaged(&y, N*sizeof(float));

  // initialize x and y arrays on the host

  for (int i = 0; i < N; i++) {

    x[i] = 1.0f;

    y[i] = 2.0f;

  }

​

  ULONGLONG t1, t2, freq;
  double timeuse;

  QueryPerformanceFrequency((LARGE_INTEGER*)&freq);
  QueryPerformanceCounter((LARGE_INTEGER*)&t1);
  // Run kernel on 30M elements on the CPU
  add<<<1, 1>>>(N, x, y);
  QueryPerformanceCounter((LARGE_INTEGER*)&t2);
  timeuse = (double)(t2 - t1) / (double)freq * 1000.0;
​
  std::cout << "add(int, float*, float*) time: " << timeuse << "ms" << std::endl;
  // Check for errors (all values should be 3.0f)
  float maxError = 0.0f;
  for (int i = 0; i < N; i++)
    maxError = fmax(maxError, fabs(y[i]-3.0f));
  std::cout << "Max error: " << maxError << std::endl;
​
  // Free memory
  delete [] x;
  delete [] y;
​
  cudaFree(x);
  cudaFree(y);
  
  return 0;
}
#include <iostream>


#include <math.h>


#include <windows.h>
​


// function to add the elements of two arrays

__global__

void add(int n, float *x, float *y){

  for (int i = 0; i < n; i++)

      y[i] = x[i] + y[i];

}

​

int main(){

  int N = 1<<25; // 30M elements

​

  float *x = new float[N];

  float *y = new float[N];

  cudaMallocManaged(&x, N*sizeof(float));

  cudaMallocManaged(&y, N*sizeof(float));

  // initialize x and y arrays on the host

  for (int i = 0; i < N; i++) {

    x[i] = 1.0f;

    y[i] = 2.0f;

  }

​

  ULONGLONG t1, t2, freq;
  double timeuse;

  QueryPerformanceFrequency((LARGE_INTEGER*)&freq);
  QueryPerformanceCounter((LARGE_INTEGER*)&t1);
  // Run kernel on 30M elements on the CPU
  add<<<1, 1>>>(N, x, y);
  QueryPerformanceCounter((LARGE_INTEGER*)&t2);
  timeuse = (double)(t2 - t1) / (double)freq * 1000.0;
​
  std::cout << "add(int, float*, float*) time: " << timeuse << "ms" << std::endl;
  // Check for errors (all values should be 3.0f)
  float maxError = 0.0f;
  for (int i = 0; i < N; i++)
    maxError = fmax(maxError, fabs(y[i]-3.0f));
  std::cout << "Max error: " << maxError << std::endl;
​
  // Free memory
  delete [] x;
  delete [] y;
​
  cudaFree(x);
  cudaFree(y);
  
  return 0;
}

二、一份CUDA代码如何在CPU上跑起来？

这个很简单，直接运行就可以了。

avatar

三、一份CUDA代码如何在NVidia-GPU上跑起来？

首先要保证电脑内的CUDA环境，这里我们只说1个Linux下配置CUDA环境的细节：

  • 下载CUDA的时候如果说有几个软件包无法下载，不妨加上-- fix-missing试一试，即sudo apt install nvidia-cuda-toolkit --fix-missing；

然后按照如图所示的方式就可以跑起来啦！可以看到明显快了很多（笔者是GTX1650，附上截图）。

avatar

avatar

四、一份CUDA代码如何在ZLUDA上跑起来？

ZLUDA仓库：github.com/vosen/ZLUDA

这个比前面的复杂一些，网上的资料也比较少，但是经过一番探索还是摸索出来了方法：

我们需要的前置环境：

  • Visual Studio 2019（对没错，必须是2017~2019的版本才行，笔者一开始下了一个2022的结果不行）
  • Rust（下载最简单的版本就可以了，可以用cargo --version和rustc --version来检查是否安装成功）
  • Visual Studio的cl.exe需要添加至环境变量，具体这个文件在哪可以使用Everything搜索
  • Clone上述ZLUDA仓库到本地并编译，编译过程下文会展示

avatar

如上这种编译错误就是没有VS工具链导致的?

1. ZLUDA前置环境如何编译？

首先要进入ZLUDA安装目录，打开终端后执行cargo build --release

2. 如何把.cu文件编译成.exe文件呢？

只需要进入我们的项目目录下执行：nvcc -o my_app.exe my_app.cu

例如笔者就是：nvcc -o hello.exe windows-hello.cu

Linux：

LD_LIBRARY_PATH=<ZLUDA_DIRECTORY> <APPLICATION> <APPLICATIONS_ARGUMENTS>

3. 如何运行我们的代码？

按理来说随便进入一个目录就行，不过如果出错的话还是在ZLUDA目录下执行如下代码：

Windows：

<ZLUDA_DIRECTORY>\zluda_with.exe -- <APPLICATION> <APPLICATIONS_ARGUMENTS>

最后一项参数实测可以为空，<APPLICATION>是你代码的.exe可执行文件，例如笔者就是：

D:\My_Files\Coding-Project-2023\OSPP\OSPP-THU-CUDA\Start\ZLUDA\ZLUDA\target\release\deps\zluda_with.exe -- hello.exe

avatar

比单纯的GPU还快！虽然这里没有控制变量，要控制的话应该在Linux下去测试，但是还是肉眼可见的快。

© 版权声明
文章版权归作者所有，未经允许请勿转载，侵权请联系 admin@trc20.tw 删除。
THE END

# C++# GPU
喜欢就支持一下吧
点赞0

Warning: mysqli_query(): (HY000/3): Error writing file '/tmp/MYykFsFt' (Errcode: 28 - No space left on device) in /www/wwwroot/583.cn/wp-includes/class-wpdb.php on line 2345
admin的头像-五八三
相关推荐
评论 抢沙发
头像
欢迎您留下宝贵的见解！
提交
头像

昵称

图形验证码
取消
昵称

请填写用户信息：

  • 昵称(必填)
  • 邮箱(必填)
代码

请输入代码：

确认
图片
文章目录
文章目录
热门推荐