Ubuntu下的CUDA编程（五）——使用pt…

本次测试目的仅仅为了尝试能否使用pthread实现多线程对CPU和GPU分别进行控制：

文件结构如下：

main.cc：控制线程的主函数

vector.cu：实现对核函数进行调用的任务函数

vector.h：任务函数声明

vector_kernel.cu：核函数

vector_kernel.h：核函数声明

生成文件如下：

vector：最终生成的可执行文件

main.o：生成的函数体obj

vector.o：生成的任务函数obj

vector_kernel.o：生成的核函数obj

功能描述：

线程1：执行GPU代码（向量加法）

线程2：执行CPU代码（循环打印数据）

线程3：执行CPU代码（循环打印数据）

线程4：执行CPU代码（循环打印数据）

文件代码：

main.cc：

#include <stdio.h>

#include <pthread.h>

#include <sys/time.h>

#include <stdlib.h>

#include <string.h>

#include "vector.h"

pthread_t thread[4];

void *thread1(void *)

{

//do GPU task

float time, start;

start = clock();

printf("thread1:I'm thread1\n");

int i, n = 100;

float *a, *b, *c;

a = (float *)malloc(n * sizeof(float));

b = (float *)malloc(n * sizeof(float));

c = (float *)malloc(n * sizeof(float));

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

{

a[i] = 1.0f;

b[i] = 1.0f;

}

for(i = 0; i < 100; i++)

{

vectorAdd(a, b, c, n);

}

printf("thread1:c[%d] = %f\n", 0, c[0]);

free(a);

free(b);

free(c);

time = clock() - start;

printf("thread1: task was finished!\ncostTime1 : %f\n", time / CLOCKS_PER_SEC);

pthread_exit(NULL);

}

void *thread2(void *)

{

//do CPU task

float time, start;

start = clock();

printf("thread2:I'm thread2\n");

int i, j, k = 1;

for(i = 0; i < 1000; i++)

{

for(j = 0; j < 1000; j++)

{

printf("thread2:k = %d\n", k);

k++;

}

}

time  = clock() - start;

printf("thread2: task was finished!\ncostTime2 : %f\n", time / CLOCKS_PER_SEC);

pthread_exit(NULL);

}

void *thread3(void *)

{

//do CPU task

float time, start;

start = clock();

printf("thread3:I'm thread3\n");

int i, j, k = 1;

for(i = 0; i < 1000; i++)

{

for(j = 0; j < 1000; j++)

{

printf("thread3:k = %d\n", k);

k++;

}

}

time  = clock() - start;

printf("thread3: task was finished!\ncostTime3 : %f\n", time / CLOCKS_PER_SEC);

pthread_exit(NULL);

}

void *thread4(void *)

{

//do CPU task

float time, start;

start = clock();

printf("thread4:I'm thread4\n");

int i, j, k = 1;

for(i = 0; i < 1000; i++)

{

for(j = 0; j < 1000; j++)

{

printf("thread4:k = %d\n", k);

k++;

}

}

time  = clock() - start;

printf("thread4: task was finished!\ncostTime4 : %f\n", time / CLOCKS_PER_SEC);

pthread_exit(NULL);

}

void thread_create()

{

int temp;

memset(&thread, 0, sizeof(thread));

if((temp = pthread_create(&thread[0], NULL, thread1, NULL)) != 0)

printf("线程1创建失败！\n");

else

printf("线程1被创建！\n");

if((temp = pthread_create(&thread[1], NULL, thread2, NULL)) != 0)

printf("线程2创建失败！\n");

else

printf("线程2被创建！\n");

if((temp = pthread_create(&thread[2], NULL, thread3, NULL)) != 0)

printf("线程3创建失败！\n");

else

printf("线程3被创建！\n");

if((temp = pthread_create(&thread[3], NULL, thread4, NULL)) != 0)

printf("线程4创建失败！\n");

else

printf("线程4被创建！\n");

}

void thread_wait()

{

if(thread[0] != 0)

{

pthread_join(thread[0], NULL);

printf("线程1已经结束\n");

}

if(thread[1] != 0)

{

pthread_join(thread[1], NULL);

printf("线程2已经结束\n");

}

if(thread[2] != 0)

{

pthread_join(thread[2], NULL);

printf("线程3已经结束\n");

}

if(thread[3] != 0)

{

pthread_join(thread[3], NULL);

printf("线程4已经结束\n");

}

}

int main()

{

float time, start;

printf("我是主函数，正在创建线程\n");

start = clock();

thread_create();

printf("我是主函数，正在等待线程完成任务\n");

thread_wait();

time = clock() - start;

printf("costTime0 : %f\n", time / CLOCKS_PER_SEC);

return 0;

}

vector.cu：

#include <stdio.h>

#include <stdlib.h>

#include <cuda.h>

#include "vector.h"

#include "vector_kernel.h"

void vectorAdd(float *a, float *b, float *c, int n)

{

float *d_a, *d_b, *d_c;

cudaMalloc((void **)&d_a, n * sizeof(float));

cudaMemcpy(d_a, a, n * sizeof(float), cudaMemcpyHostToDevice);

cudaMalloc((void **)&d_b, n * sizeof(float));

cudaMemcpy(d_b, b, n * sizeof(float), cudaMemcpyHostToDevice);

cudaMalloc((void **)&d_c, n * sizeof(float));

cudaMemcpy(d_b, b, n * sizeof(float), cudaMemcpyHostToDevice);

vectorAddKernel<<<1, n>>>(d_a, d_b, d_c, n);

cudaMemcpy(c, d_c, n * sizeof(float), cudaMemcpyDeviceToHost);

cudaFree(d_a);

cudaFree(d_b);

cudaFree(d_c);

}

vector.h：

#include <stdio.h>

#include <stdlib.h>

void vectorAdd(float *a, float *b, float *c, int n);

vector_kernel.cu：

#include <stdio.h>

#include <stdlib.h>

#include <cuda.h>

#include "vector_kernel.h"

__global__ void vectorAddKernel(float *a, float *b, float *c, int n)

{

int tid;

tid = threadIdx.x;

if(tid < n)

c[tid] = a[tid] + b[tid];

}

vector_kernel.h：

#include <stdio.h>

#include <cuda.h>

#include <stdlib.h>

__global__ void vectorAddKernel(float *a, float *b, float *c, int n);

而最主要的是makefile的内容，由于nvcc会将.cu生成obj默认为c++方式，所以，需要将主函数改为.cc文件，不然将会编译出错!

makefile：

vector : main.o vector.o vector_kernel.o

nvcc -o vector main.o vector.o vector_kernel.o

vector_kernel.o : vector_kernel.cu vector_kernel.h

nvcc -c vector_kernel.cu

vector.o : vector.cu vector.h vector_kernel.h

nvcc -c vector.cu

main.o : main.cc vector.h

cc -lpthread -c main.cc

通过测试时间可以知道几个线程之间的执行关系，大家自己测试啦～

以下是一个运行结果：