CUDA 计时功能,记录GPU程序/函数耗时,cudaEventCreate,cudaEventRecord,cudaEventElapsedTime

为了测试GPU函数的耗时,可以使用 CUDA 提供的计时功能:cudaEventCreate , cudaEventRecord , 和 cudaEventElapsedTime。这些函数可以帮助你测量某个 CUDA 操作(如设置设备)所花费的时间。

一、记录耗时案例

以下是一个示例程序,它测量调用 cudaSetDevice 所花费的时间:

cpp 复制代码
#include <iostream>
#include <vector>
#include <cuda_runtime.h>

 
__global__ void dummyKernel() {

    // Dummy kernel to ensure CUDA context is initialized
}

 

int main() {

    // CUDA device IDs
    int device1 = 0;
    int numIterations = 10; // Number of times to call cudaSetDevice

 
    // Create CUDA events
    cudaEvent_t start, stop;
    cudaEventCreate(&start);
    cudaEventCreate(&stop);

    // Vector to store elapsed times
    std::vector<float> elapsedTimes(numIterations);

 
    // Set initial device (optional, but ensures a known starting state)
    cudaSetDevice(device1);

 
    // Measure time for multiple cudaSetDevice calls
    for (int i = 0; i < numIterations; ++i) {
        // Record the start event
        cudaEventRecord(start, 0);
 
        // Set the device (this is the operation we are timing)
        cudaSetDevice(device1);

        // Record the stop event
        cudaEventRecord(stop, 0);

        // Measure the elapsed time between the start and stop events
        cudaEventElapsedTime(&elapsedTimes[i], start, stop);

        // Output results
        std::cout << "Number of iterations: i " << i << std::endl;

        std::cout << " time to set device " << device1 << ": " << elapsedTimes[i] << " ms" << std::endl;

    }

 

    // Calculate statistics (e.g., average time)
    float totalTime = 0.0f;
    for (float time : elapsedTimes) {
        totalTime += time;
    }
    float averageTime = totalTime / numIterations;

 

    // Output results
    std::cout << "Number of iterations: " << numIterations << std::endl;
    std::cout << "Average time to set device " << device1 << ": " << averageTime << " ms" << std::endl;

 
    // Optionally, run a dummy kernel to ensure CUDA is initialized and ready
    dummyKernel<<<1, 1>>>();
    cudaDeviceSynchronize();
 

    // Clean up
    cudaEventDestroy(start);
    cudaEventDestroy(stop);

    return 0;
}

二、编译和运行

2.1 编译 : 使用 nvcc 编译这个 CUDA 程序。(上面程序文件铭为test_cudaSetDevice_multiple.cu)

bash 复制代码
nvcc -o test_cudaSetDevice_multiple test_cudaSetDevice_multiple.cu

2.2 运行: ,然后运行生成的可执行文件。

bash 复制代码
./test_cudaSetDevice_multiple

哈哈哈,就得到运行结果啦!

相关推荐
野蛮人6号1 小时前
如何查看自己电脑的CUDA版本?
cuda
SHIPKING39312 天前
【CUDA&cuDNN安装】深度学习基础环境搭建
人工智能·深度学习·cuda·cudnn
咩咩大主教15 天前
2025最新版使用VSCode和CMake图形化编译调试Cuda C++程序(保姆级教学)
c++·vscode·cmake·visual studio·cuda·cpp·cuda c++
量化投资和人工智能17 天前
【CUDA编程】OptionalCUDAGuard详解
c++·人工智能·python·机器学习·云计算·cuda
weiwei2284419 天前
VS2019+CUDA12.5入门
gpu·vs2019·cuda
爱串门的小马驹1 个月前
VScode编译调试debug,gpu的cuda程序,Nsight
vscode·gpu·cuda
阿巴阿阿巴巴巴巴1 个月前
【深度学习相关安装及配环境】Anaconda搭建虚拟环境并安装CUDA、cuDVV和对应版本的Pytorch,并在jupyter notebook上部署
人工智能·pytorch·python·深度学习·jupyter·cuda
r0ysue_1 个月前
5060显卡驱动PyCUDA开发环境搭建
cuda·驱动·5060ti
易·木1 个月前
在Visual Studio中进行cuda编程
visual studio·cuda
weixin_428498491 个月前
NVC++ 介绍与使用指南
c++·cuda