将顺序代码集成到并行代码M​​PI的问题

时间:2011-03-23 09:06:47

标签: mpi pgm

我试图将边缘检测拉普拉斯算子集成到我以前的MPI代码中。我现在的问题是在数据散射后在1d数组中进行边缘检测。我得到了输出,但颜色从假设的图像反转。任何人都可以帮我解决这个问题。这是并行代码:

#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <mpi.h>
#include <math.h>

#define SIZE_X 640
#define SIZE_Y 480
#define   smooth  3

int mod(int z, int l);


/****************** Main Program ***********************/

int main(int argc, char **argv)
{
FILE *FR,*FW;
int ierr;
int rank, size, a[100000], sum, m, n;
int ncells;
int greys[SIZE_X][SIZE_Y];
int rows,cols, maxval;
int mystart, myend, myncells;
const int IONODE=0;
int *disps, *counts, *mydata;
int *data;
int i,j,temp1;
char dummy[50]="";
int csx,sum1,sum2,k,l,x;//w1[3][3],w2[3][3]; 

  //Laplacian Operator
static int w1[3][3]={
{0,-1,0},
{-1,4,-1},
{0,-1,0}
};

static int w2[3][3]={
{0,-1,0},
{-1,4,-1},
{0,-1,0}
};

/****************** Initialize MPI ***********************/



ierr = MPI_Init(&argc, &argv);
if (argc != 3) {
    fprintf(stderr,"Usage: %s infile outfile\n",argv[0]);
    fprintf(stderr,"outputs the negative of the input file.\n");
    return -1;
}            

ierr  = MPI_Comm_rank(MPI_COMM_WORLD, &rank);
ierr = MPI_Comm_size(MPI_COMM_WORLD, &size);
if (ierr) {
fprintf(stderr,"Catastrophic MPI problem;  exiting\n");MPI_Abort(MPI_COMM_WORLD,1);
  }




  /****************** Master open image file for read  ***********************/


if (rank == IONODE) {
    rows=SIZE_X;
    cols=SIZE_Y;
    maxval=255;
    FR=fopen(argv[1], "r+");


  /****************** Read the header part of the image ***********************/

    fgets(dummy,50,FR);
    do{  fgets(dummy,50,FR); } while(dummy[0]=='#');
    fgets(dummy,50,FR);

    for (j = 0; j <cols; j++)
        for (i = 0; i <rows; i++)
   {
       fscanf(FR,"%d",&temp1);
     greys[i][j] = temp1;
   }


 /****************** Read pixel values ***********************/

ncells = rows*cols;
disps = (int *)malloc(size * sizeof(int));
counts= (int *)malloc(size * sizeof(int));
data = &(greys[0][0]); /* we know all the data is contiguous */
}

// Time every processor 
//MPI_Barrier(MPI_COMM_WORLD);
//p_time = MPI_Wtime();


 /****************** Everyone calculate their number of cells ***********************/

ierr = MPI_Bcast(&ncells, 1, MPI_INT, IONODE, MPI_COMM_WORLD);
myncells = ncells/size;
mystart = rank*myncells;
myend   = mystart + myncells - 1;
if (rank == size-1) myend = ncells-1;
myncells = (myend-mystart)+1;
mydata = (int *)malloc(myncells * sizeof(int));


 /****************** Assemble the list of counts. Might not be equal if don't divide evenly.***********************/

ierr = MPI_Gather(&myncells, 1, MPI_INT, counts, 1, MPI_INT, IONODE, MPI_COMM_WORLD);
if (rank == IONODE) {
    disps[0] = 0;
    for (i=1; i<size; i++) {
        disps[i] = disps[i-1] + counts[i-1];
    }
}


/****************** Scatter the data to all processor ***********************/

ierr = MPI_Scatterv(data, counts, disps, MPI_INT, mydata, myncells, MPI_INT, IONODE, MPI_COMM_WORLD);


/****************** All processor do AVERAGE FILTERING ***********************/

csx=smooth/2; 
for (i=0; i<myncells; i++)
{
 sum1=0;
 sum2=0;
 for(k=0;k<smooth;k++)
 {
   for(l=0;l<smooth;l++) 
   {
     x=i+k-csx; 
     sum1+=w1[k][l]* mydata[mod(x,myncells)];
     sum2+=w2[k][l]* mydata[mod(x,myncells)];    
   }
 }

 if((abs(sum1)+abs(sum2))>125)
   mydata[i]=255;
 else
   mydata[i]=0;
}


/******************  Gather the data from all processor ***********************/

ierr = MPI_Gatherv(mydata, myncells, MPI_INT, data, counts, disps, MPI_INT, IONODE, MPI_COMM_WORLD);

//MPI_Barrier(MPI_COMM_WORLD);  
//p2_time = MPI_Wtime();
//printf("\nProcessor %d = %g microseconds\n", rank, (p2_time*1000000)-(p_time*1000000));       


/******************  Master open image file for write ***********************/  

if (rank == IONODE){
    FW=fopen(argv[2], "w");
    fprintf(FW,"P2\n%d %d\n255\n",rows,cols);    
    for(j=0;j<cols;j++)
        for(i=0;i<rows;i++)
            fprintf(FW,"%d ", greys[i][j]);

    }

free(mydata);
if (rank == IONODE) {
free(counts);
free(disps);

}

if (rank == IONODE) {
    fclose(FR);
    fclose(FW);

    }



MPI_Finalize();

return 0;

 }


//Sub routine

 /* periodic extension (outside of the  image frame) */
  int mod(int z, int l)
  {
  if( z >= 0 && z < l ) return z;
  else
  if( z < 0) return (z+l);
  else
    if( z > (l-1)) return (z-l);
   return 0;
  }

这是我需要集成到上面的程序

的顺序代码
#include <stdio.h>
#include <stdlib.h>
#include<conio.h>
#include<time.h>
 #include<math.h>

#define   size_x  203   
#define   size_y  152


typedef struct

{

  int imagesize_x, imagesize_y;

  int **pixel;

}image_t;

image_t allocate_image(const int imagesize_x, const int imagesize_y);

int mod(int z, int l);



void main()

{

image_t   image_in,image_out;



 int m,n, temp;

 int smooth,csx,csy; 

 int k,l,x,y,sum1,sum2;



 FILE *cpp1,*cpp2;



 char dummy[50]="";



 //Laplacian Operator



 static int w1[3][3]={

  {0,-1,0},

  {-1,4,-1},

  {0,-1,0}

  };



 static int w2[3][3]={

  {0,-1,0},

  {-1,4,-1},

  {0,-1,0}

  };





 cpp1=fopen("e:\\input_image\\A.pgm", "r+");

 cpp2=fopen("e:\\output_image\\edge_lap.pgm", "w+");







 fgets(dummy,50,cpp1);           

 do{

  fgets(dummy,50,cpp1);         

 }while(dummy[0]=='#');

fgets(dummy,50,cpp1);           





 fprintf(cpp2,"P2\n%d %d\n255\n",(size_x),(size_y));





image_in =  allocate_image(size_x,size_y);

 image_out = allocate_image(size_x,size_y);





                            //Reading Input Image



 for (n = 0; n < size_y; n++)

 {

  for (m = 0; m <size_x; m++)

   {



    fscanf(cpp1,"%d",&temp);



      image_in.pixel[m][n] = temp;





    }

   }


                        /* Edge Detection  */



 smooth=3;

 csx=smooth/2; csy=smooth/2;



                         //Edge detection

 for (n = 0; n < size_y; n++) {

  for (m = 0; m < size_x; m++) { 



 sum1=0;sum2=0;



 for(k=0;k<smooth;k++){

   for(l=0;l<smooth;l++) {



 x=m+k-csx; y=n+l-csy;    



     sum1+=w1[k][l]* image_in.pixel[mod(x,size_x)][mod(y,size_y)];

     sum2+=w2[k][l]* image_in.pixel[mod(x,size_x)][mod(y,size_y)];



   }

 }

 if((fabs(sum1)+fabs(sum2))>125)

   image_out.pixel[m][n]=255;

 else

   image_out.pixel[m][n]=0;

   }

  }



                        //Writing Edge Detected Image



   for (n = 0; n < size_y; n++)

   {

for (m = 0; m <size_x; m++)

 {



    fprintf(cpp2,"%d ",image_out.pixel[m][n]);       



   }

    }

   }



    image_t allocate_image(const int imagesize_x, const int imagesize_y)

     {

  image_t result;

   int x =  0, y = 0;



   result.imagesize_x = imagesize_x;

     result.imagesize_y = imagesize_y;



      result.pixel =(int **) calloc(imagesize_x, sizeof(int*));



   for(x = 0; x < imagesize_x; x++)

     {

     result.pixel[x] =(int*) calloc(imagesize_y, sizeof(int));



     for(y = 0; y < imagesize_y; y++)

   {

     result.pixel[x][y] = 0;

     }

      }



     return result;

  }





  int mod(int z, int l)

   {

    if( z >= 0 && z < l ) return z;

       else

    if( z < 0) return (z+l);

    else

  if( z > (l-1)) return (z-l);

     }

这是输入图像http://orion.math.iastate.edu/burkardt/data/pgm/balloons.pgm

我使用:mpirun -np 10 ./mysource balloons.pgm output.pgm

运行

1 个答案:

答案 0 :(得分:0)

因此,在您需要反转图像的previous question中,图像的二维结构无关紧要,您可以将图像分成(像素数)/(进程数)块每个都可以反转它们的像素。

但是,这里的二维结构确实重要;要应用模板,您需要拥有所有相邻像素。在图像的边缘,你需要来自另一边的数据(在物理学中,我们称之为“周期性边界条件” - 图像像环面一样环绕)。

因此,我们需要以维护数据的2d结构的方式分解数据。我们可以进行完整的2d分解;因此,如果有6个进程,则图像将被分解为

+---+---+---+
| 0 | 1 | 2 |
+---+---+---+
| 3 | 4 | 5 |
+---+---+---+

(这是2011年,我还在做ascii艺术,通过计算机进行交流。这是怎么发生的?)

但现在进行1-d分解更简单;也就是说,我们保留数据的二维结构,但我们只沿一个维度切割数据。在C中,沿着行执行此操作要容易得多,因此数据块在内存中是连续的:

-----1------
-----2------
-----3------
-----4------
-----5------
-----6------

到目前为止,最简单的方法是填充原始图像中的行数,使其在任务数之间平均分配。

现在,如果一个任务负责将模板应用于行0,它需要来自行(nrows-1)的数据,反之亦然;类似地,对于列0和列(ncols-1)。因此,我们将用两个以上的行和列填充数组,并在读取数据时将数据从列(ncols-1)复制到第0列。

因此,在这种情况下,阅读图像会改变为:

fgets(dummy,50,FR);
do{  fgets(dummy,50,FR); } while(dummy[0]=='#');
sscanf(dummy,"%d %d",&cols, &rows);
fgets(dummy,50,FR);


nrowspertask = (rows/size);
if (nrowspertask*size < rows) nrowspertask++;
int totrows = nrowspertask*size;

/* pad the number of rows so it divides evenly by # of tasks */
/* and then add 2 rows, 2 cols, for "wraparound" at edges */

image_t image;
image = allocate_image( cols+2, totrows+2 );

/****************** Read pixel values ***********************/

for (j = 0; j <cols; j++)
    for (i = 0; i <rows; i++)
    {
        fscanf(FR,"%d",&temp1);
        image.pixel[j+1][i+1] = temp1;
    }

/* copy bottom row to top, top row to bottom */
for (j=1; j<cols+1; j++) {
    image.pixel[j][0]      = image.pixel[j][rows];
    image.pixel[j][rows+1] = image.pixel[j][1];
}

/* copy leftmost col to right, rightmost col to left */
for (i=1; i<rows+1; i++) {
    image.pixel[0][i]      = image.pixel[cols][i];
    image.pixel[cols+1][i] = image.pixel[1][i];
}

现在分解了。每个任务都将负责nrowspertask行;但是他们需要上面和下面的行来完成这项工作,侧面的填充也是如此。因此,每个图片都必须从(nrowspertask+2)*(cols+2)开始接收&(image.pixel[rank*(nrowspercol)][0])个整数。我们仍然可以使用MPI_Scatterv() - 我们将计数和代价设置为

data = &(image.pixel[0][0]); /* we know all the data is contiguous */
disps = (int *)malloc(size * sizeof(int));
counts= (int *)malloc(size * sizeof(int));

for (i=0;i<size;i++) {
    counts[i]=(nrowspertask+2)*(cols+2);
    disps[i]=i*(nrowspertask)*(cols+2);
}

然后每个人都获得他们的数据:

locimage = allocate_image(cols+2, nrowspertask+2);
ierr = MPI_Scatterv(data, counts, disps, MPI_INT,&(locimage.pixel[0][0]),
                         (nrowspertask+2)*(cols+2), MPI_INT, IONODE, MPI_COMM_WORLD);

现在你的过滤器非常类似于串行情况,除了你不必担心mods,因为你已经用你需要的信息填充了数据。

你将数据收集到数组中的方式非常相似,但我会让你把它弄清楚;你只发回“真实”行,而不是填充行,否则你将覆盖数据。