cuda归约求和最终版

#include "device_functions.h"
#include "cuda_runtime.h"
#include "device_launch_parameters.h"
#include "stdlib.h"
#include <iostream>
using namespace std;

#define LENG 117//数组长度

//返回thread和block
int getThreadNum()
{
	cudaDeviceProp prop;//cudaDeviceProp的一个对象
	int count = 0;//GPU的个数
	cudaGetDeviceCount(&count);
	std::cout << "gpu 的个数：" << count << '\n';

	cudaGetDeviceProperties(&prop, 0);//第二参数为那个gpu
	cout << "最大线程数：" << prop.maxThreadsPerBlock << endl;
	cout << "最大网格类型：" << prop.maxGridSize[0] << '\t' << prop.maxGridSize[1] << '\t' << prop.maxGridSize[2] << endl;
	return prop.maxThreadsPerBlock;
}
//形参：枚举类型
//判断是否成功分配内存
void GetCudaCalError(cudaError err)
{
	if (err != cudaSuccess)
	{
		cout << "分配内存失败！程序结束！";
	}
	return;
}

//这个函数有局限性：不能放到多个block上运行
__global__ void sumReduced(float*aGpu, float*sumGpu, int countNum)
{
	const int id = threadIdx.x;
	//定义一个共享内存
	__shared__ float sData[LENG];
	//为其赋值
	sData[id] = aGpu[id];
	//等待每个线程赋值完成
	__syncthreads();

	//实现归约求和
	/*
	1、每个数和当前索引加上总数一半的值相加，如果超出最大索引数就加0
	2、等到所有线程计算完毕
	3、线程数减半，直到减到1
	*/
	int leng = LENG;
	for (int i = countNum / 2.0 + 0.5; i > 1; i = i / 2.0 + 0.5)
	{
		if (id < i)
		{

			if (id + i < leng)
			{
				sData[id] += sData[id + i];
			}
			//sData[id] += sData[id + i];
		}
		__syncthreads();
		leng = leng / 2.0 + 0.5;
	}
	if (id == 0)
	{
		sumGpu[0] = sData[0]+ sData[1];
	}
}



int main()
{
	float a[LENG];
	float asum = 0;
	for (int i = 0; i < LENG; ++i)
	{
		a[i] = i * (i + 1);
		//cout << a[i] << '\t';
	}

	//定义Device上的内存
	float *aGpu = 0;
	float *sumGpu = 0;
	//为其开辟内存
	GetCudaCalError(cudaMalloc(&aGpu, LENG * sizeof(float)));
	GetCudaCalError(cudaMalloc(&sumGpu, 1 * sizeof(float)));
	//给aGpu 赋值
	cudaMemcpy(aGpu, a, LENG * sizeof(float), cudaMemcpyHostToDevice);
	//开一个block，每个block里面有16个thread
	sumReduced << <1, LENG >> > (aGpu, sumGpu, LENG);
	//将结果传回host
	cudaMemcpy(&asum, sumGpu, 1 * sizeof(float), cudaMemcpyDeviceToHost);
	cout << "cuda数组和为：" << asum << endl;
	cudaFree(aGpu);
	cudaFree(sumGpu);
	float testSum = 0;
	for (int i = 0; i < LENG; ++i)
	{
		testSum += a[i];
	}
	cout << "for数组和为：" << testSum << endl;


}