本文共 1861 字，大约阅读时间需要 6 分钟。

定义流：

cudaStream_t stream;

创建流：

cudaStreamCreate(cudaStream_t *s)

销毁流：

cudaStreamDestroy(cudaStream_t s)

流的同步：显示同步

//同步单个流，等待该命令都完成cudaStreamSynchronize(cudaStream_t stream);//同步所有流，等待整个设备上的流都完成cudaDeviceSynchronize();//通过某个事件：等待某个事件结束后执行该流上的命令cudaStreamWaitEvent(cudaEvent_t stream);//查询一个流任务是否完成cudaStreamQuery(cudaStream_t stream);

流的使用

在使用GPU完成任务时，有2个地方可以使用到流

以_global_定义的kernel函数可以交给流去执行

kernel<<
   
    >>(parameters);
   

GPU和CPU之间数据传输

cudaMemcpyAsync(dst, src, copy_size, copy_direction, stream);

程序运行截图如下：

代码如下：

#include "cuda_runtime.h"#include "device_launch_parameters.h"#include 
   
    #include 
    
     using namespace std;__global__ void addKernel(int *c, int *a, int *b){	int i = threadIdx.x;	c[i] = a[i] + b[i];}void testStream() {	int p_data_a[100] = { 0 };	int p_data_b[100] = { 0 };	int p_data_c[100] = { 0 };	for (int i = 0; i < 100; i++) {		p_data_a[i] = i;		p_data_b[i] = 10 + i;		p_data_c[i] = 0;	}	cout << "begin" << endl;	for (int i = 0; i < 100; i++) {		cout << p_data_c[i] << " ";	}	cout << endl;	int *dev_a = nullptr;	int *dev_b = nullptr;	int *dev_c = nullptr;	cudaMalloc(&dev_a, sizeof(int) * 100);	cudaMalloc(&dev_b, sizeof(int) * 100);	cudaMalloc(&dev_c, sizeof(int) * 100);	//拷贝到内存	cudaMemcpy(dev_a, p_data_a, sizeof(int) * 100, cudaMemcpyHostToDevice);	cudaMemcpy(dev_b, p_data_b, sizeof(int) * 100, cudaMemcpyHostToDevice);	cudaStream_t streams[100];	for (int i = 0; i < 100; ++i)		cudaStreamCreate(streams + i);	for (int i = 0; i < 100; ++i)		addKernel << <1, 1, 0 >> > (dev_c + i, dev_a + i, dev_b + i);	cudaDeviceSynchronize();	cudaMemcpy(p_data_c, dev_c, sizeof(int) * 100, cudaMemcpyDeviceToHost);	cout << "end:" << endl;	for (int i = 0; i < 100; i++) {		cout << p_data_c[i] << " ";	}	cout << endl;	cout << "over" << endl;	cudaFree(dev_a);	cudaFree(dev_b);	cudaFree(dev_c);}int main() {	testStream();	getchar();	return 0;}
    
   

转载地址：https://it1995.blog.csdn.net/article/details/111880547 如侵犯您的版权，请留言回复原文章的地址，我们会给您删除此文章，给您带来不便请您谅解！