CUDA实战

1.第一个程序，输出hello world，1个Block块中含有5个线程

 1 #include <stdio.h>
 2 #include "cuda_runtime.h"
 3 
 4 __global__ void hello(void)
 5 {
 6   printf("hello world from GPU!
");
 7 }
 8 int main()
 9 {
10     printf("hello world from CPU!
");
11     hello<<<1,5>>>();
12     //重置CUDA设置释放程序占用的资源
13     cudaDeviceReset();
14     return 0;
15 }

View Code

2.参数的传入，

 1 #include <stdio.h>
 2 #include "cuda_runtime.h"
 3 #include "device_launch_parameters.h"
 4 __global__ void add(int i,int j)
 5 {
 6    int count;
 7    count = i + j;
 8    printf("
 Sum is %d
",count);
 9 }
10 
11 int main()
12 {
13     add<<<1,1>>>(10,20);
14     cudaDeviceReset();
15     return 0;
16 }

View Code

3.数据的传入与传出，我们的数据要从内存copy到显存上面，然后现在又要从显存上面copy回来

 1 #include <stdio.h>
 2 #include "cuda_runtime.h"
 3 #include "device_launch_parameters.h"
 4 
 5 __global__ void decrease(int a, int b, int *c)
 6 {
 7     *c = a + b;
 8 }
 9 int main()
10 {
11     int *c=0;
12     int *dev_c=0;
13     //初始化CPU上的内存空间
14     c = (int*)malloc(sizeof(int));
15     //初始化GPU上的内存空间
16     cudaMalloc((void**)&dev_c,sizeof(int));
17     //调用内核函数
18     decrease <<<1,1>>>(15,20,dev_c);
19     //等待设备所有线程任务执行完毕
20     cudaDeviceSynchronize();
21     //将数据从device中复制到hist中
22     cudaMemcpy(c,dev_c,sizeof(int),cudaMemcpyDeviceToHost);
23     //输出
24     printf(" c = %d
",*c);
25     //释放内存
26     cudaFree(dev_c);
27     free(c);
28     return 0;
29 }

View Code

4.传入的值全改为指针类型

 1 #include <stdio.h>
 2 #include "cuda_runtime.h"
 3 #include "device_launch_parameters.h"
 4 
 5 __global__ void addCuda(int* a, int* b, int* c)
 6 {
 7     *c = *a - *b;
 8 }
 9 
10 void addWithCuda(int *c,int *a,int *b)
11 {
12     int *dev_c = 0;
13     int *dev_a = 0;
14     int *dev_b = 0;
15     
16     //初始化CUDA内存
17     cudaMalloc((void**)&dev_c,sizeof(int));
18     cudaMalloc((void**)&dev_a,sizeof(int));
19     cudaMalloc((void**)&dev_b,sizeof(int));
20     
21     //从主机复制数据复制到device上
22     cudaMemcpy(dev_a,a,sizeof(int),cudaMemcpyHostToDevice);
23     cudaMemcpy(dev_b,b,sizeof(int),cudaMemcpyHostToDevice);
24     
25     //调用内核函数
26     addCuda<<<1,1>>>(dev_a,dev_b,dev_c);
27     cudaDeviceSynchronize();
28     
29     //数据复制到host
30     cudaMemcpy(c,dev_c,sizeof(int),cudaMemcpyDeviceToHost);
31     
32     cudaFree(dev_c);
33     cudaFree(dev_a);
34     cudaFree(dev_b);
35 
36 }
37 
38 int main()
39 {
40    int a, b, c;
41    a = 30;
42    b = 15;
43    c = 10;
44    //传入参数变量（地址）
45    addWithCuda(&c,&a,&b);
46    //重置CUDA设备释放程序占用的程序
47    cudaDeviceReset();
48    printf("Value is %d
", c);
49 
50     return 0;
51 }

View Code

5.传入的值全改为指针类型

 1 #include <stdio.h>
 2 #include "cuda_runtime.h"
 3 #include "device_launch_parameters.h"
 4 
 5 __global__ void deCuda(int* a, int* b, int* c)
 6 {
 7     *c = *a - *b;
 8 }
 9 int main()
10 {
11    int *a, *b, *c;
12    a = (int*)malloc(sizeof(int));
13    b = (int*)malloc(sizeof(int));
14    c = (int*)malloc(sizeof(int));
15    *a=10;
16    *b=5;
17    *c=0;
18    int *dev_c = 0;
19    int *dev_a = 0;
20    int *dev_b = 0;
21     //3.请求CUDA设备的内存（显存），执行CUDA函数
22     cudaMalloc((void**)&dev_c, sizeof(int));
23     cudaMalloc((void**)&dev_a, sizeof(int));
24     cudaMalloc((void**)&dev_b, sizeof(int));
25     
26     cudaMemcpy(dev_a, a, sizeof(int), cudaMemcpyHostToDevice);
27     cudaMemcpy(dev_b, b, sizeof(int), cudaMemcpyHostToDevice);
28     
29     deCuda<<<1,1>>>(dev_a,dev_b,dev_c);
30     cudaMemcpy(c, dev_c, sizeof(int), cudaMemcpyDeviceToHost);
31     printf("Value is %d
", *c);
32    
33     cudaFree(dev_c);
34     cudaFree(dev_a);
35     cudaFree(dev_b);
36     free(a);
37     free(b);
38     free(c);
39    //重置CUDA设备释放程序占用的程序
40    cudaDeviceReset();
41    return 0;
42 }

View Code

6. 程序实现向量的加法操作，一个block中含有512个线程

 1 #include <stdio.h>
 2 #include <cuda_runtime.h>
 3 __global__ void add( int *dev_a, int *dev_b, int *dev_c)
 4 {
 5     int i=threadIdx.x;
 6     dev_c[i] = dev_a[i] + dev_b[i];
 7 }
 8 int main()
 9 {
10     int host_a[512], host_b[512], host_c[512];
11     for(int i = 0; i < 512; i++)
12     {
13         host_a[i] = i;
14         host_b[i] = i<<1;
15     }
16     //定义cudaError,默认为cudaSuccess
17     cudaError_t err = cudaSuccess;
18     int *dev_a, *dev_b, *dev_c;
19     err = cudaMalloc((void**)&dev_a,sizeof(int)*512);
20     err = cudaMalloc((void**)&dev_b,sizeof(int)*512);
21     err = cudaMalloc((void**)&dev_c,sizeof(int)*512);
22     
23     if(err!=cudaSuccess)
24     {
25         printf("the cuadaMalloc on GPU is failed");
26         return 1;
27     }
28     
29     printf("SUCCESS
");
30     //从host到device
31     cudaMemcpy(dev_a,host_a,sizeof(host_a),cudaMemcpyHostToDevice);
32     cudaMemcpy(dev_b,host_b,sizeof(host_b),cudaMemcpyHostToDevice);
33     
34     //调用核函数
35     add<<<1,512>>>(dev_a,dev_b,dev_c);
36     cudaMemcpy(&host_c,dev_c,sizeof(host_c),cudaMemcpyDeviceToHost);
37     for(int i=0; i<512; i++)
38     {
39         printf("host_a[%d] + host_b[%d] = %d + %d = %d
",i,i,host_a[i],host_b[i],host_c[i]);
40     }
41     
42     //释放内存
43     cudaFree(dev_c);
44     cudaFree(dev_b);
45     cudaFree(dev_a);
46     
47     return 0;
48     
49 }

View Code