cuda并行归约算法（不断优化过程）

归约

方法

#include<cuda_runtime_api.h>
#include<cuda_runtime.h>
#include<device_launch_parameters.h>
#include<iostream>
#include<device_functions.h>
using namespace std;
__global__ void jia(int a[101], int b[101]) {
   
    __shared__ int sdata[101];
    int idx = threadIdx.x;
    int x = blockIdx.x * blockDim.x + threadIdx.x;
    sdata[idx] = a[x];
    __syncthreads();
    for (int s = 1; s < blockDim.x; s *= 2) {
   
        if (idx % (s * 2) == 0) {
   
            sdata[idx] += sdata[idx + s];
        }
        __syncthreads();
    }
    if (idx == 0)b[0] = sdata[0];
}
int main() {
   
	int ha[101];
    int *da,*db;
    int hb[101];
    for(int i = 0; i <= 100; i++)
            ha[i] = i;


    cudaMalloc((void**)&da, sizeof(int) * 101);
    cudaMalloc((void**)&db, sizeof(int) * 101);
    cudaMemcpy(da, ha, sizeof(int) * 101, cudaMemcpyHostToDevice);
    dim3 Block(101);
    jia << <1, Block >> > (da,db);
    cudaMemcpy(hb, db, sizeof(int), cudaMemcpyDeviceToHost);
    cout << hb[0] << endl;
    cudaFree(da);
    cudaFree(db);
}

优化一

在上一方法中对于线程的使用由于间隔使用会造成很大的资源浪费

#include<cuda_runtime_api.h>
#include<cuda_runtime.h>
#include<device_launch_parameters.h>
#include<iostream>
#include<device_functions.h>
using namespace std;
__global__ void jia(int a[101], int b[101]) {
   
    __shared__ int sdata[101];
    int idx = threadIdx.x;
    int x = blockIdx.x * blockDim.x + threadIdx.x;
    sdata[idx] = a[x];
    __syncthreads();
    f