CUDA并行规约(相邻配对)

最新推荐文章于 2023-12-07 15:22:12 发布

武泗海

最新推荐文章于 2023-12-07 15:22:12 发布

阅读量657

点赞数

分类专栏： CUDA CUDA学习

本文链接：https://blog.csdn.net/qq_17239003/article/details/78824004

版权

CUDA 同时被 2 个专栏收录

26 篇文章 8 订阅

订阅专栏

CUDA学习

22 篇文章 2 订阅

订阅专栏

#include "cuda_runtime.h"
#include "device_launch_parameters.h"

#include <stdio.h>
#include "math.h"
#include "stdlib.h"

//错误检查的宏定义
#define CHECK(call)									\
{													\
	const cudaError_t status=call;					\
if (status!=cudaSuccess)							\
	{												\
	printf("文件:%s,函数:%s,行号:%d",__FILE__,		\
						__FUNCTION__,__LINE__);		\
	printf("%s", cudaGetErrorString(status));		\
	exit(1);										\
	}												\
}													\

//核函数
__global__ void Kernel(int *d_data, int *d_local_sum, int N)
{
	int tid = threadIdx.x;
	int index = blockIdx.x*blockDim.x + threadIdx.x;
	int *data = d_data + blockIdx.x*blockDim.x;

	if (index >= N) return;

	for (int strize = 1; strize < blockDim.x; strize *= 2)
	{
		if (tid % (2 * strize) == 0)


		{
			data[tid] += data[tid + strize];
		}
		__syncthreads();
	}

	if (tid == 0)
	{
		d_local_sum[blockIdx.x] = data[0];
	}
}

//主函数
int main()
{

	//基本参数设置
	cudaSetDevice(0);
	const int N =65536;
	int local_length =8;
	int total_sum = 0;

	dim3 grid(((N + local_length - 1) / local_length), 1);
	dim3 block(local_length, 1);

	int *h_data = nullptr;
	int *h_local_sum = nullptr;
	int *d_data = nullptr;
	int *d_local_sum = nullptr;


	//Host&Deivce内存申请及数组初始化
	h_data = (int*)malloc(N * sizeof(int));
	h_local_sum = (int*)malloc(int(grid.x) * sizeof(int));


	CHECK(cudaMalloc((void**)&d_data, N * sizeof(int)));
	
	CHECK(cudaMalloc((void**)&d_local_sum, int(grid.x) * sizeof(int)));
	

	for (int i = 0; i < N; i++)
		h_data[i] = int(10 * sin(0.02*3.14*i));//限制数组元素值，防止最终求和值超过int的范围


	//数据拷贝至Device
	CHECK(cudaMemcpy(d_data, h_data, N * sizeof(int), cudaMemcpyHostToDevice));


	//执行核函数
	Kernel << <grid, block >> > (d_data, d_local_sum, N);

	//数据拷贝至Host
	CHECK(cudaMemcpy(h_local_sum, d_local_sum, int(grid.x) * sizeof(int),
		cudaMemcpyDeviceToHost));
	
	//同步&重置设备
	CHECK(cudaDeviceSynchronize());
	CHECK(cudaDeviceReset());
	

	for (int i = 0; i < int(grid.x); i++)
	{
		total_sum += h_local_sum[i];
	}

	printf("%d \n", total_sum);

	getchar();
	return 0;

}