使用动态分配的 Bcast、分散和收集时出现 MPI 隔离错误

我正在使用一些代码，在从 concole 输入实现矩阵 mult 的表大小时遇到了一些问题。

第一个版本适用于：

 const int size = 1000;
 int mat_a[size][size], mat_b[size][size], mat_c[size][size];

要使用控制台参数，我发现需要实现动态数组分配。不幸的是，我遇到了一个问题：

 *** Process received signal ***
 Signal: Segmentation fault (11)
 Signal code: Address not mapped (1)
 Failing at address: 0x7ffd955237f8

我想这个问题可能出在 MPI 函数中，如 Bcast、Scatter 和 Gather。我在 stackovf 中搜索了类似的案例，但我看不到它。

这是代码：

#include <mpi.h>
#include <stdio.h>
#include <iostream>
#include <cstdlib>
#include <cmath>
#include <math.h>
int main(int argc, char *argv[])
{
  int taskid, ntasks, mat_start, mat_end, i, j, k;
double start_time; //hold start time
double end_time; // hold end time  
  MPI_Init (&argc, &argv);
  MPI_Comm_rank(MPI_COMM_WORLD, &taskid);   
  MPI_Comm_size(MPI_COMM_WORLD, &ntasks); 
        int size = 0;
        if (argc != 2) {
                printf("No arguments");
                exit(-1);
        }
        size = atoi(argv[1]);
        if (size < 0 ) {
                printf("SIZE: %2d n", size);
                exit(-1);
        }
int **mat_a = (int **)malloc(sizeof(int *)*size);
int **mat_b = (int **)malloc(sizeof(int *)*size);  
int **mat_c = (int **)malloc(sizeof(int *)*size);
for (int z = 0 ; z < size ; z++){
     mat_a[z] = (int *)malloc(sizeof(int)*size);
     mat_b[z] = (int *)malloc(sizeof(int)*size);
     mat_c[z] = (int *)malloc(sizeof(int)*size);
}
  mat_start = taskid * size/ntasks;
  mat_end = (taskid+1) * size/ntasks;
  if (taskid==0) {
    for (i = 0; i < size; i++) {
        for (j = 0; j < size; j++) {
            mat_a[i][j] = (int)(sin(i) * i * j) % 10;
        }
    }
    for (i = 0; i < size; i++) {
        for (j = 0; j < size; j++) {
            mat_b[i][j] = (int)(cos(j) * (i + j)) % 10;
        }
    }
  }
start_time = MPI_Wtime();
  MPI_Bcast (&mat_b, size*size, MPI_INT, 0, MPI_COMM_WORLD);
  MPI_Scatter (&mat_a, size*size/ntasks, MPI_INT, mat_a[mat_start], size*size/ntasks, MPI_INT, 0, MPI_COMM_WORLD);
  printf("computing slice %d (from row %d to %d)n", taskid, mat_start, mat_end-1);
  for (i=mat_start; i<mat_end; i++) 
    for (j=0; j<size; j++) {
      mat_c[i][j]=0;
      for (k=0; k<size; k++)
    mat_c[i][j] += mat_a[i][k]*mat_b[k][j];
    }
  MPI_Gather (mat_c[mat_start], size*size/ntasks, MPI_INT, mat_c, size*size/ntasks, MPI_INT, 0, MPI_COMM_WORLD);
        end_time = MPI_Wtime();
        printf("nRunning Time = %fnn", end_time - start_time);
  MPI_Finalize();
  return 0;
}

谁能告诉我出了什么问题？

编辑：

谢谢你的回答。我试图实施您的解决方案建议，但没有取得好结果。我将代码的某些部分更改为如下所示：

int **mat_a=(int **)malloc(size*sizeof(int *));
        int **mat_b=(int **)malloc(size*sizeof(int *));
        int **mat_c=(int **)malloc(size*sizeof(int *));
    if(mat_a==NULL){fprintf(stderr,"malloc failedn");exit(1);}
    if(mat_b==NULL){fprintf(stderr,"malloc failedn");exit(1);}
    if(mat_c==NULL){fprintf(stderr,"malloc failedn");exit(1);}
    mat_a[0]=(int*)malloc(size*size*sizeof(int));
 mat_b[0]=(int*)malloc(size*size*sizeof(int));
mat_c[0]=(int*)malloc(size*size*sizeof(int));
   if(mat_a[0]==NULL){fprintf(stderr,"malloc failedn");exit(1);}
if(mat_b[0]==NULL){fprintf(stderr,"malloc failedn");exit(1);}
if(mat_c[0]==NULL){fprintf(stderr,"malloc failedn");exit(1);}
  int ti;
    for(ti=1;ti<size;ti++){
        mat_a[ti]=&mat_a[0][size*ti];
mat_b[ti]=&mat_a[0][size*ti];
mat_c[ti]=&mat_a[0][size*ti];
 }
  mat_start = taskid * size/ntasks;
  mat_end = (taskid+1) * size/ntasks;
//populating the array ......
start_time = MPI_Wtime();
  MPI_Bcast(mat_a[0],size*size, MPI_INT,0,MPI_COMM_WORLD); 
 //MPI_Bcast (&mat_b, size*size, MPI_INT, 0, MPI_COMM_WORLD);
//  MPI_Scatter (&mat_b, size*size/ntasks, MPI_INT, mat_a[mat_start], size*size/ntasks, MPI_INT, 0, MPI_COMM_WORLD);
  MPI_Scatter (mat_b[0], size*size/ntasks, MPI_INT, mat_a[mat_start], size*size/ntasks, MPI_INT, 0, MPI_COMM_WORLD);
  printf("computing slice %d (from row %d to %d)n", taskid, mat_start, mat_end-1);
  for (i=mat_start; i<mat_end; i++) 
    for (j=0; j<size; j++) {
      mat_c[i][j]=0;
      for (k=0; k<size; k++)
    mat_c[i][j] += mat_a[i][k]*mat_b[k][j];
    }
  MPI_Gather (mat_c[mat_start], size*size/ntasks, MPI_INT, mat_c, size*size/ntasks, MPI_INT, 0, MPI_COMM_WORLD);
        end_time = MPI_Wtime();
        printf("nRunning Time = %fnn", end_time - start_time);
  MPI_Finalize();
  return 0;
}

程序开始运行，甚至打印mat_a（添加打印时），但经过一段时间的延迟，我得到了这个：

[cuda:05167] *** Process received signal ***
[cuda:05167] Signal: Segmentation fault (11)
[cuda:05167] Signal code:  (128)
[cuda:05167] Failing at address: (nil)

散而聚可能是抢劫？如何改变，让它最终发挥作用？