0

我最初有这个功能，我正在尝试使用循环展开来进一步优化它 - 我遇到了麻烦 - 翻转 for 循环可以提高效率，以及在循环之外进行调用。然而，当我像我一样应用循环展开时，它错过了函数应该做的事情。

double
applyFilter(class Filter *filter, cs1300bmp *input, cs1300bmp *output)
{

  long long cycStart, cycStop;

  cycStart = rdtscll();

  output -> width = input -> width;
  output -> height = input -> height;


  for(int col = 1; col < (input -> width) - 1; col = col + 1) {
    for(int row = 1; row < (input -> height) - 1 ; row = row + 1) {
      for(int plane = 0; plane < 3; plane++) {

    output -> color[plane][row][col] = 0;

    for (int j = 0; j < filter -> getSize(); j++) {
      for (int i = 0; i < filter -> getSize(); i++) {   
        output -> color[plane][row][col]
          = output -> color[plane][row][col]
          + (input -> color[plane][row + i - 1][col + j - 1] 
         * filter -> get(i, j) );
      }
    }
    
    output -> color[plane][row][col] =  
      output -> color[plane][row][col] / filter -> getDivisor();

    if ( output -> color[plane][row][col]  < 0 ) {
      output -> color[plane][row][col] = 0;
    }

    if ( output -> color[plane][row][col]  > 255 ) { 
      output -> color[plane][row][col] = 255;
    }
      }
    }
  }

  cycStop = rdtscll();
  double diff = cycStop - cycStart;
  double diffPerPixel = diff / (output -> width * output -> height);
  fprintf(stderr, "Took %f cycles to process, or %f cycles per pixel\n",
      diff, diff / (output -> width * output -> height));
  return diffPerPixel;
}

这是我到达的地方，但它似乎不起作用。我会很感激解释我在循环展开部分做错了什么。

double applyFilter(class Filter *filter, cs1300bmp *input, cs1300bmp *output){

  long long cycStart, cycStop;

  cycStart = rdtscll();

//start
    
  output -> width = input -> width;
  output -> height = input -> height;
    
//function calls outside loop.
    int filterSize = filter -> getSize();
    int divisor = filter -> getDivisor();
    
//intializaions
    
    int inputHlen = input -> height - 1;
    int inputWlen = input -> width - 1;
    

// loop unrolling row + k - 1 , col + k - 1
for(int plane = 0; plane < 3; plane++){
  for(int row = 1; row + 3 < inputHlen; row += 4){
    for(int col = 1; col +3 < inputWlen; col += 4){ 
      
              
          output -> color[plane][row][col] = 0;
          output -> color[plane][row + 1][col + 1] = 0;
          output -> color[plane][row + 2][col + 2] = 0;
          output -> color[plane][row + 3][col + 3] = 0;
          
          int acc1 = output -> color[plane][row][col];
          int acc2 = output -> color[plane][row + 1][col + 1];
          int acc3 = output -> color[plane][row + 2][col + 2];
          int acc4 = output -> color[plane][row + 3][col + 3];
          
          for (int j = 0; j + 3 < filterSize; j += 4) {
              for (int i = 0; i + 3 < filterSize; i += 4){ 
                acc1 = acc1 + (input -> color[plane][row + i - 1][col + j - 1] * filter -> get(i, j) );
                acc2 = acc2 + (input -> color[plane][row + i][col + j]         * filter -> get(i + 1, j + 1) );
                acc3 = acc3 + (input -> color[plane][row + i + 1][col + j + 1] * filter -> get(i +2, j + 2) );
                acc4 = acc4 + (input -> color[plane][row + i + 2][col + j + 2] * filter -> get(i + 3, j + 3) );
              }
            }
          
          acc1 = acc1 / divisor;
          acc2 = acc2 / divisor;
          acc3 = acc3 / divisor;
          acc4 = acc4 / divisor;

          
          acc1 = (acc1 < 0) ? 0 : acc1;
          acc1 = (acc1 > 255) ? 255 : acc1;

          acc2 = (acc1 < 0) ? 0 : acc1;
          acc2 = (acc1 > 255) ? 255 : acc1;
          
          acc3 = (acc1 < 0) ? 0 : acc1;
          acc3 = (acc1 > 255) ? 255 : acc1;
          
          acc4 = (acc1 < 0) ? 0 : acc1;
          acc4 = (acc1 > 255) ? 255 : acc1;
          
          
          output -> color[plane][row][col] = acc1;
          output -> color[plane][row + 1][col + 1] = acc2;
          output -> color[plane][row + 2][col + 2] = acc3;
          output -> color[plane][row + 3][col + 3] = acc4;
          
      }
    }
  }
  
//end

  cycStop = rdtscll();
  double diff = cycStop - cycStart;
  double diffPerPixel = diff / (output -> width * output -> height);
  fprintf(stderr, "Took %f cycles to process, or %f cycles per pixel\n",
      diff, diff / (output -> width * output -> height));
  return diffPerPixel;
}
4

1 回答 1

0

一个问题是您只想展开最内层的循环 ( col)。通过尝试展开两个循环 (row和col)，您会丢失很多计算。看到这一点的一种方法是，您已将循环迭代减少了 16 倍（rowandcol循环都以 4 为单位），而循环体只完成了 4 倍的工作。

这是因为您只查看“对角线”元素：

color[plane][1][1]
color[plane][2][2]
color[plane][3][3]
color[plane][4][4]

对于第一次循环迭代，而不是

color[plane][1][1]
color[plane][1][2]
color[plane][1][3]
color[plane][1][4]

您永远不会计算其中的 12 个值。color[plane][2][1]将是其中之一。您应该在row每次迭代中保持递增 1，并且应该更新color[plane][row][...]每个循环。

于 2020-07-07T04:47:28.960 回答