cuda共享内存-不一致的结果

我正在尝试进行并行归约，以求和CUDA中的数组。目前，我通过一个数组，用于在每个块中存储元素的总和。这是我的代码：

#include <cstdlib>
#include <iostream>
#include <cuda.h>
#include <cuda_runtime_api.h>
#include <helper_cuda.h>
#include <host_config.h>
#define THREADS_PER_BLOCK 256
#define CUDA_ERROR_CHECK(ans) { gpuAssert((ans), __FILE__, __LINE__); }

using namespace std;

inline void gpuAssert(cudaError_t code, char *file, int line, bool abort=true)
{
   if (code != cudaSuccess) 
   {
      fprintf(stderr,"GPUassert: %s %s %d\n", cudaGetErrorString(code), file, line);
      if (abort) exit(code);
   }
}

struct double3c {
    double x; 
    double y;
    double z;
    __host__ __device__ double3c() : x(0), y(0), z(0) {}
    __host__ __device__ double3c(int x_, int y_, int z_) : x(x_), y(y_), z(z_) {}
    __host__ __device__ double3c& operator+=(const double3c& rhs) { x += rhs.x; y += rhs.y; z += rhs.z;}
    __host__ __device__ double3c& operator/=(const double& rhs) { x /= rhs; y /= rhs; z /= rhs;}

};

class VectorField {
public:
    double3c *data;
    int size_x, size_y, size_z;
    bool is_copy;  

    __host__ VectorField () {}

    __host__ VectorField (int x, int y, int z) {
        size_x = x; size_y = y; size_z = z;
        is_copy = false;
        CUDA_ERROR_CHECK (cudaMalloc(&data, x * y * z * sizeof(double3c))); 
    }

    __host__ VectorField (const VectorField& other) {
        size_x = other.size_x; size_y = other.size_y; size_z = other.size_z;
        this->data = other.data;
        is_copy = true;
    }

    __host__ ~VectorField() {     
        if (!is_copy) CUDA_ERROR_CHECK (cudaFree(data));
    }
};

__global__ void KernelCalculateMeanFieldBlock (VectorField m, double3c* result) {
    __shared__ double3c blockmean[THREADS_PER_BLOCK];    
    int index = threadIdx.x + blockIdx.x * blockDim.x;
    if (index < m.size_x * m.size_y * m.size_z) blockmean[threadIdx.x] = m.data[index] = double3c(0, 1, 0);
    else blockmean[threadIdx.x] = double3c(0,0,0);
    __syncthreads();
    for(int s = THREADS_PER_BLOCK / 2; s > 0; s /= 2) {
        if (threadIdx.x < s) blockmean[threadIdx.x] += blockmean[threadIdx.x + s];
        __syncthreads();
    }


    if(threadIdx.x == 0) result[blockIdx.x] = blockmean[0];   
}

double3c CalculateMeanField (VectorField& m) { 
    int blocknum = (m.size_x * m.size_y * m.size_z - 1) / THREADS_PER_BLOCK + 1;
    double3c *mean = new double3c[blocknum]();
    double3c *cu_mean;
    CUDA_ERROR_CHECK (cudaMalloc(&cu_mean, sizeof(double3c) * blocknum));
    CUDA_ERROR_CHECK (cudaMemset (cu_mean, 0, sizeof(double3c) * blocknum));

        KernelCalculateMeanFieldBlock <<<blocknum, THREADS_PER_BLOCK>>> (m, cu_mean);
        CUDA_ERROR_CHECK (cudaPeekAtLastError());
        CUDA_ERROR_CHECK (cudaDeviceSynchronize());
        CUDA_ERROR_CHECK (cudaMemcpy(mean, cu_mean, sizeof(double3c) * blocknum, cudaMemcpyDeviceToHost));

    CUDA_ERROR_CHECK (cudaFree(cu_mean));
    for (int i = 1; i < blocknum; i++) {mean[0] += mean[i];}
    mean[0] /= m.size_x * m.size_y * m.size_z;
    double3c aux = mean[0];
    delete[] mean;
    return aux;
}



int main() {
    VectorField m(100,100,100);
    double3c sum = CalculateMeanField (m);
    cout <<  sum.x << '\t' << sum.y << '\t' <<sum.z;  


    return 0;
}

编辑

发布了功能代码。VectorField用10x10x10元素构造a可以很好地工作，并且给出平均值1，但是用100x100x100元素构造它给出的平均值约为0.97（每次运行都不同）。这是进行并行缩减的正确方法，还是应该坚持每个块一个内核启动？