无效的设备符号 cudaMemcpyFromSymbol CUDA

我想计算 CUDA 中数组所有元素的总和。我想出了这段代码。它编译没有任何错误。但结果始终为零。我收到了无效的设备符号cudaMemcpyFromSymbol。我无法使用 Thrust 或 Cublas 等任何库。

#define TRIALS_PER_THREAD 4096
#define NUM_BLOCKS 256
#define NUM_THREADS 256
double *dev;
__device__ volatile double pi_gpu = 0;

__global__ void ArraySum(double *array)

{
unsigned int tid = threadIdx.x + blockDim.x * blockIdx.x;
pi_gpu = pi_gpu + array[tid];
__syncthreads();
}

int main (int argc, char *argv[]) {
cudaMalloc((void **) &dev, NUM_BLOCKS * NUM_THREADS * sizeof(double));
    double pi_gpu_h;

ArraySum<<<NUM_BLOCKS, NUM_THREADS>>>(dev);
cudaDeviceSynchronize();
cudaError err = cudaMemcpyFromSymbol(&pi_gpu_h, &pi_gpu, sizeof(double), cudaMemcpyDeviceToHost);
if( cudaSuccess != err )
{
    fprintf( stderr, "cudaMemcpyFromSymbolfailed : %s\n", cudaGetErrorString( err ) );
    exit( -1 );
}
return pi_gpu_h; // this is always zero!!!
}

符号调用副本中的符号参数不正确。它应该看起来像这样：

cudaMemcpyFromSymbol(&pi_gpu_h, pi_gpu, sizeof(double), 0, cudaMemcpyDeviceToHost)