Question

我试图将n * n矩阵划分为p行，n可能无法被p整除。所以我需要划分成不同大小的行，最简单的方法是向每个处理器发送n / p行，除了最后一行需要n / p + n％p。

这是我的代码：

using namespace std;
int main(int argc, char* argv[])
{
    int my_rank = 0;
    int comm_size = 0;

    MPI_Init(&argc, &argv);

    MPI_Comm_rank(MPI_COMM_WORLD, &my_rank);
    MPI_Comm_size(MPI_COMM_WORLD, &comm_size);

    double *Adata;
    double **adjArray;
    int n;


    if (my_rank == 0){
        n=6;
        Adata = (double *)malloc(sizeof(double)*n*n);
        adjArray = (double **)malloc(sizeof(double*) * n);

        for(int i = 0; i < n; i++) {
            adjArray[i] = &(Adata[i*n]);
        }

        int k=0;
        for (int i=0; i<n; i++) {
            for (int j=0; j<n; j++) {
                adjArray[i][j]=k;
                k++;
            }
        }

        cout<<"---Adjacancy Matrix:"<<endl;

        for (int i=0; i<n; i++) {
            for (int j=0; j<n; j++) {
                if(adjArray[i][j]==INT_MAX)
                {
                    cout<< " - ";
                }else
                {
                    cout<< adjArray[i][j]<<" ";
                }
            }
            cout<<endl;
        }
        cout<<"----------------------------------------------------"<<endl;
    }


    //---------------------------------------------------------
    // Broadcasting the data among the processors.

    MPI_Bcast( &n,1,MPI_INT,0,MPI_COMM_WORLD);

    //---------------------------------------------------------
    // Scatter the rows to each processor

    int rem = 0; // elements remaining after division among processes
    int sum = 0; // Sum of counts. Used to calculate displacements
    if(my_rank==comm_size-1) rem=n%comm_size;

    int *displs = (int *)malloc(comm_size*sizeof(int));
    int *sendcounts = (int *)malloc(comm_size*sizeof(int));
    int numPerProc=n/comm_size;
    int receive_buffer[numPerProc+rem];

    for (int i=0; i<comm_size-1; i++) {
        sendcounts[i]=(n)/comm_size;
        displs[i] = sum;
        sum += sendcounts[i];
    }
    sendcounts[comm_size-1]=n/comm_size+rem;
    displs[comm_size-1]=sum;

    MPI_Datatype strip;
    /* defining a datatype for sub-matrix */
    MPI_Type_vector(numPerProc, n, n, MPI_DOUBLE, &strip);
    MPI_Type_commit(&strip);

    double **strip_A,*stripdata;

    stripdata = (double *)malloc(sizeof(double)*numPerProc*n);
    strip_A = (double **)malloc(sizeof(double*)*numPerProc);
    for(int i= 0; i< numPerProc+rem; i++) {
        strip_A[i] = &(stripdata[i*n]);
    }

    MPI_Scatterv(Adata, sendcounts, displs, strip, &(strip_A[0][0]), sendcounts[my_rank], strip, 0, MPI_COMM_WORLD);


    for(int i = 0; i < sendcounts[my_rank]; i++) {
        if(i == 0) {
            printf("rank = %d\n", my_rank);
        }
        for(int j = 0; j < n; j++) {

            if(strip_A[i][j]==INT_MAX)
            {
                cout<< " - ";
            }else
            {
                cout<< strip_A[i][j]<<" ";
            }
        }
        printf("\n");
    }

    MPI_Finalize();

    return 0;
}

不幸的是，一旦n不等于p，它就不起作用。例如，一旦我尝试p = 4，输出为：

[warn] kq_init: detected broken kqueue; not using.: No such file or directory
[warn] kq_init: detected broken kqueue; not using.: No such file or directory
[warn] kq_init: detected broken kqueue; not using.: No such file or directory
[warn] kq_init: detected broken kqueue; not using.: No such file or directory
[warn] kq_init: detected broken kqueue; not using.: No such file or directory
[warn] kq_init: detected broken kqueue; not using.: No such file or directory
[warn] kq_init: detected broken kqueue; not using.: No such file or directory
[warn] kq_init: detected broken kqueue; not using.: No such file or directory
[warn] kq_init: detected broken kqueue; not using.: No such file or directory
---Adjacancy Matrix:
0 1 2 3 4 5 
6 7 8 9 10 11 
12 13 14 15 16 17 
18 19 20 21 22 23 
24 25 26 27 28 29 
30 31 32 33 34 35 
----------------------------------------------------
rank = 0
0 1 2 3 4 5 
rank = 2
12 13 14 15 16 17 
rank = 1
6 7 8 9 10 11 
rank = 3
18 19 20 21 22 23 
6.95287e-310 6.95287e-310 6.95287e-310 1.99804e+161 8.11662e+217 3.25585e-86 
1.94101e-80 2.68185e-80 4.81827e+151 1.39957e-306 2.33584e-314 6.95287e-310

任何帮助表示赞赏！谢谢！

Answer 1

一行的派生数据类型应该像这样构建（音符计数为1而不是numPerProc）

MPI_Type_vector(1, n, n, MPI_DOUBLE, &strip);

注意一个更简单的选项是

MPI_Type_contiguous(n, MPI_DOUBLE, &strip);

还有其他问题

sendcounts和displs仅与排名0相关，sendcounts[comm_size-1]在该排名上不正确
stripdata和strip_A在最后一个排名上的大小不正确（例如，您分配了numPerProc行，但访问了numPerProc+rem行。

分区矩阵为p行

1 个答案: