cuda歸約求和最終版

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

#define LENG 117//數組長度

//返回thread和block
int getThreadNum()
{
	cudaDeviceProp prop;//cudaDeviceProp的一個對象
	int count = 0;//GPU的個數
	cudaGetDeviceCount(&count);
	std::cout << "gpu 的個數：" << count << '\n';

	cudaGetDeviceProperties(&prop, 0);//第二參數爲那個gpu
	cout << "最大線程數：" << prop.maxThreadsPerBlock << endl;
	cout << "最大網格類型：" << prop.maxGridSize[0] << '\t' << prop.maxGridSize[1] << '\t' << prop.maxGridSize[2] << endl;
	return prop.maxThreadsPerBlock;
}
//形參：枚舉類型
//判斷是否成功分配內存
void GetCudaCalError(cudaError err)
{
	if (err != cudaSuccess)
	{
		cout << "分配內存失敗！程序結束！";
	}
	return;
}

//這個函數有侷限性：不能放到多個block上運行
__global__ void sumReduced(float*aGpu, float*sumGpu, int countNum)
{
	const int id = threadIdx.x;
	//定義一個共享內存
	__shared__ float sData[LENG];
	//爲其賦值
	sData[id] = aGpu[id];
	//等待每個線程賦值完成
	__syncthreads();

	//實現歸約求和
	/*
	1、每個數和當前索引加上總數一半的值相加，如果超出最大索引數就加0
	2、等到所有線程計算完畢
	3、線程數減半，直到減到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';
	}

	//定義Device上的內存
	float *aGpu = 0;
	float *sumGpu = 0;
	//爲其開闢內存
	GetCudaCalError(cudaMalloc(&aGpu, LENG * sizeof(float)));
	GetCudaCalError(cudaMalloc(&sumGpu, 1 * sizeof(float)));
	//給aGpu 賦值
	cudaMemcpy(aGpu, a, LENG * sizeof(float), cudaMemcpyHostToDevice);
	//開一個block，每個block裏面有16個thread
	sumReduced << <1, LENG >> > (aGpu, sumGpu, LENG);
	//將結果傳回host
	cudaMemcpy(&asum, sumGpu, 1 * sizeof(float), cudaMemcpyDeviceToHost);
	cout << "cuda數組和爲：" << asum << endl;
	cudaFree(aGpu);
	cudaFree(sumGpu);
	float testSum = 0;
	for (int i = 0; i < LENG; ++i)
	{
		testSum += a[i];
	}
	cout << "for數組和爲：" << testSum << endl;


}