CUDA 事件计时 - 技术栈

CUDA 事件可以为主机代码和设备代码计时。

基本的语法

cpp 复制代码

// 定义事件变量
cudaEvent_t start, stop;
// 初始化
cudaEventCreate(&start);
cudaEventCreate(&stop);
// 记录代表时间开始的事件，注意不是地址
cudaEventRecord(start);
// 在TCC的驱动下可以省略，在WDDM驱动模式下必须保留，所以默认保留
// 不可以使用错误检测函数，默认返回值是错误的
cudaEventQuery(start);


/**
* code
**/

// 记录代表时间结束的事件
cudaEventRecord(stop);
// 事件同步函数，等待事件记录结束
cudaEventSynchronize(stop);
// 计算时间差
float elapsed_time;
cudaEventElapsedTime(&elapsed_time, start, stop);

// 销毁变量
cudaEventDestroy(start);
cudaEventDestroy(stop);

使用示例：

计算两个数组的和

cpp 复制代码

#include <cuda_runtime.h>
#include <device_launch_parameters.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <iostream>

#define NUM_REPEATS 10

static void CheckCudaErrorAux(const char*, unsigned, const char*, cudaError_t);
#define CUDA_CHECK_RETURN(value) CheckCudaErrorAux(__FILE__,__LINE__, #value, value)


// 设备函数
__device__ float add(const float x, const float y)
{
	return x + y;
}

__global__ void addFromGPU(float* A, float* B, float* C, const int N)
{
	int blockId = blockIdx.x;
	int id = blockId * blockDim.x + threadIdx.x;
	if (id >= N)
	{
		return;
	}

	C[id] = add(A[id], B[id]);
}

void initialData(float* addr, int nCount)
{
	for (size_t i = 0; i < nCount; i++)
	{
		addr[i] = (float)(rand() & 0xFFF) / 100.f;
	}
}

int main()
{
	int iElemntCount = 4096*10;
	size_t stBytesCount = iElemntCount * sizeof(float); // 字节数

	// 分配主机内存和设备内存并初始化
	float* fpHost_A = new float[iElemntCount];
	float* fpHost_B = new float[iElemntCount];
	float* fpHost_C = new float[iElemntCount];
	memset(fpHost_A, 0, stBytesCount);
	memset(fpHost_B, 0, stBytesCount);
	memset(fpHost_C, 0, stBytesCount);
	float* fpDevice_A, * fpDevice_B, * fpDevice_C;
	CUDA_CHECK_RETURN(cudaMalloc((void**)&fpDevice_A, stBytesCount));
	CUDA_CHECK_RETURN(cudaMalloc((void**)&fpDevice_B, stBytesCount));
	CUDA_CHECK_RETURN(cudaMalloc((void**)&fpDevice_C, stBytesCount));

	CUDA_CHECK_RETURN(cudaMemset(fpDevice_C, 0,stBytesCount));

	srand(666);
	initialData(fpHost_A, iElemntCount);
	initialData(fpHost_B, iElemntCount);

	CUDA_CHECK_RETURN(cudaMemcpy(fpDevice_A, fpHost_A, stBytesCount, cudaMemcpyHostToDevice));
	CUDA_CHECK_RETURN(cudaMemcpy(fpDevice_B, fpHost_B, stBytesCount, cudaMemcpyHostToDevice));


	dim3 block(32);
	dim3 grid((iElemntCount + block.x - 1)/ block.x);

	cudaEvent_t start, stop;

	for (int i = 0; i < NUM_REPEATS; i++)
	{
		CUDA_CHECK_RETURN(cudaEventCreate(&start));
		CUDA_CHECK_RETURN(cudaEventCreate(&stop));
		CUDA_CHECK_RETURN(cudaEventRecord(start));
		cudaEventQuery(start);

		addFromGPU <<<grid, block >>> (fpDevice_A, fpDevice_B, fpDevice_C, iElemntCount);

		CUDA_CHECK_RETURN(cudaEventRecord(stop));
		CUDA_CHECK_RETURN(cudaEventSynchronize(stop));
		float elapsed_time = 0.0f;
		CUDA_CHECK_RETURN(cudaEventElapsedTime(&elapsed_time, start, stop));
		CUDA_CHECK_RETURN(cudaEventDestroy(start));
		CUDA_CHECK_RETURN(cudaEventDestroy(stop));

		printf("%d \t elapsed_time = %.2f \n", i, elapsed_time);

		//CUDA_CHECK_RETURN(cudaMemcpy(fpHost_C, fpDevice_C, stBytesCount, cudaMemcpyDeviceToHost));
		//for (size_t j = 0; j < iElemntCount; j++)
		//{
		//	printf("%.2f + %.2f = %.2f \n", fpHost_A[j],  fpHost_B[j], fpHost_C[j]);
		//}
	}

	delete[]fpHost_A;
	delete[]fpHost_B;
	delete[]fpHost_C;

	fpHost_A = nullptr;
	fpHost_B = nullptr;
	fpHost_C = nullptr;

	return 0;
}

static void CheckCudaErrorAux(const char* file, unsigned line, const char* statement, cudaError_t err)
{
	if (err == cudaSuccess)
		return;
	std::cerr << statement << " returned: "<< cudaGetErrorName(err) << "  \t : " << cudaGetErrorString(err) << "(" << err << ") at " << file << ":" << line << std::endl;
	exit(1);
}

结果：

可以看出第一次调用的时候最费时