CUDA tips

__syncthreads的用法，防止wait forever的情况

• http://www.cnblogs.com/dwdxdy/p/3215136.html

一个自己写的Bug。。。感觉应该还蛮常见的

• __shared__ int s_data[BLOCKWIDTH][BLOCKWIDTH];
  if(val == 0) return;  // make sure maybe its connected
  
  int label = Block1DAddress;
  
  while(1)
  {
    // Pass 1 of the CCL algorithm
    s_labels[Block1DAddress] = label;
    if(Block1DAddress == 0) s_changed[0] = 0;
    __syncthreads();
    ...
    if(s_changed[0] == 0) break;
  }
  

  上述程序，我让val == 0的线程全部return了，但是下面while里面写了如果Block1DAddress==0的话才break，这样导致了如果一个block里面的第0个thread的值是0，s_changed就永远无法初始化为0

MEM issue

• http://stackoverflow.com/questions/7903566/how-is-2d-shared-memory-arranged-in-cuda