在 CUDA 内核中访问 2D 数据

Posted 2023-03-23

技术标签:

【中文标题】在 CUDA 内核中访问 2D 数据【英文标题】：Accessing 2D data in a CUDA kernel [duplicate] 【发布时间】：2021-08-12 17:59:09 【问题描述】：

我正在为我的大学做作业，主要想法是将 CUDA 数据并行性与 CUDA 任务并行性进行比较。我想出了一个想法来并行化康威的生活游戏。问题是，我无法弄清楚如何在 CUDA 中以多个方向导航二维数组，即上/下/右/左以及内核评估的单元格周围的角。

到目前为止，我想出了以下几点：

第一个内核代码

//determines the alive cell and save value of each cell into an array
__global__ void numAliveAround(int *oldBoard, int *newBoard, int xSize, int ySize, size_t pitchOld, size_t pitchNew)

   int x = (blockIdx.x * blockDim.x) + threadIdx.x;
   int y = (blockIdx.y * blockDim.y) + threadIdx.y;

   if(x < xSize && y < ySize)
   
      //cell above
      //xMod is to make sure the number wraps when it overflows the board
      xMod = ((x + 1) % xSize + xSize) % xSize;
      //idx calculation
      idx = xMod * xSize + y;
      outputNumber += board[idx];
      
      //more of the same code, just for cell under, left, right, and corners
      
      newBoard[x * xSize + y] = outputNumber;

第二个内核代码

//sets new cell status according to the number of alive cells around
__global__ void  determineNextState(int *board, int *newBoard, int xSize, int ySize, size_t pitchOld, size_t pitchNew)

   //getting threads
   int x = (blockIdx.x * blockDim.x) + threadIdx.x;
   int y = (blockIdx.y * blockDim.y) + threadIdx.y;

   if (x < xSize && y < ySize)
   
      int idxNew = x * xSize + y;
      int idxOld = x * xSize + y;
      int state = board[idxOld];

      //ALIVE = 1, DEAD = 0;      

      int output = DEAD;

      //checking if any alive condition is met
      if (state == ALIVE)
      
         if ((newBoard[idxNew] == 2 || newBoard[idxNew] == 3))
         
            output = ALIVE;
         
      
      else
      
         if (newBoard[idxNew] == 3)
         
            output = ALIVE;
         
      

      newBoard[idxNew] = output;

内核调用函数

void SendToCUDA(int oldBoard[COLUMNS][ROWS], int newBoard[COLUMNS][ROWS])

   //CUDA pointers
   int *d_oldBoard;
   int *d_newBoard;

   size_t pitchOld;
   size_t pitchNew;

   cudaMallocPitch(&d_oldBoard, &pitchOld, COLUMNS * sizeof(int), ROWS);
   cudaMallocPitch(&d_newBoard, &pitchNew, COLUMNS * sizeof(int), ROWS);

   cudaMemcpy2D(d_oldBoard, pitchOld, oldBoard, COLUMNS * sizeof(int), COLUMNS * sizeof(int), ROWS, cudaMemcpyHostToDevice);

   dim3 grid(divideAndRound(COLUMNS, BLOCKSIZE_X), divideAndRound(ROWS, BLOCKSIZE_Y));
   dim3 block(BLOCKSIZE_Y, BLOCKSIZE_X);

   printf("counting \n");
   numberAliveAround <<<block, grid>>> (d_oldBoard, d_newBoard, COLUMNS, ROWS, pitchOld, pitchNew);
   cudaDeviceSynchronize();
   printf("determining \n");
   determineNextState <<<block, grid>>> (d_oldBoard, d_newBoard, COLUMNS, ROWS, pitchOld, pitchNew);
   cudaDeviceSynchronize();

   //using newBoard later (outside the function) to display the Board
   cudaMemcpy2D(newBoard, COLUMNS * sizeof(int), d_newBoard, pitchNew, COLUMNS * sizeof(int), ROWS, cudaMemcpyDeviceToHost);

   cudaFree(d_oldBoard);
   cudaFree(d_newBoard);

我发现了多种访问扁平二维数组的方法，其中一些相互矛盾，例如：

//what is usually used as an exmplanation
idx = x * widht + y;
//sometimes x and y are swapped
idx = y * width + x;

//what works with simple access
int *value = (int *)((char *)(d_matrix + y * pitch)) + x;
//or
idx = x * xDim + y + pitch;

有趣的是，当我只访问数组中的一个点（例如将其中的所有值增加 1）时，后面的 2 个可以工作，但完全不适用于更复杂的导航。在这一点上，我已经坐在这个问题上很长一段时间了。因此，任何形式的见解都会非常有帮助。

【问题讨论】：

【参考方案1】：

我确实想出了答案，即在cudaMalloc2D之后访问二维数组的正确方法是：