doublerc.cu

extern "C"                                                                                                              
__global__ void uvmer(const double* a)                                                                       
{                                                                                                                                                                                                                                            
    size_t id = blockIdx.x * blockDim.x + threadIdx.x;
    // 1-4
    double rc[TASKS_PER_THREAD];                                                                                        
    #pragma unroll                                                                                                      
    for (; id < (blockDim.x * gridDim.x * TASKS_PER_THREAD); id += blockDim.x * gridDim.x)                              
    {                                                                                                                   
        rc[id / (blockDim.x * gridDim.x)] = a[id * PSIZE];                                                              
    }
  // this loop does nothing but ensure compiler does not remove rc
    for (int i = 0; i < TASKS_PER_THREAD; i++)                                                                          
        if (rc[i] != 0)                                                                                                 
            continue;                                                                                                   
        else                                                                                                            
            printf("ok %lf\n", rc[i]);                                                                                  
}