c - 使用 MPI_Send 和 MPI_Recv 实现 MPI_Scatter 的问题

Question

我正在尝试使用 MPI_Send 和 MPI_Recv 实现 MPI 函数 MPI_Scatter。

我想使用函数的官方声明，它使用 vector/数组的指针。

MPI_Scatter(
    void* send_data,
    int send_count,
    MPI_Datatype send_datatype,
    void* recv_data,
    int recv_count,
    MPI_Datatype recv_datatype,
    int root,
    MPI_Comm communicator)

我创建了一个非常适合 MPI MPI_Scatter 的示例。它显示了正确的结果。

我必须使用函数来实现这些东西，一个用指针，一个用静态整数数组。第二个工作正常，但第一个只显示创建的数组的前三个元素。我认为这是与矩阵的分配内存相关的问题，但我看不到修复它的方法。

这是代码(MMPI_Scatter 让我出错):

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


#define ROOT 0
#define N 3

int main(int argc, char **argv) {

    // for storing this process' rank, and the number of processes
    int rank, np;  
    int *matrix;

    //MPI_Scatter
    int send_count, recv_count;
    int *recv_data;

    MPI_Status status, info;

    MPI_Init(&argc, &argv);
    MPI_Comm_size(MPI_COMM_WORLD,&np);
    MPI_Comm_rank(MPI_COMM_WORLD,&rank);


    if (rank == ROOT) {
        matrix = createMatrix(np, np);
        printArray(matrix, np * np);

    }

    send_count = np;
    recv_count = np;
    recv_data = malloc(recv_count * sizeof(int));


    //The original function provided by MPI works great!!
    MPI_Scatter(matrix, send_count, MPI_INT, recv_data, recv_count, MPI_INT, ROOT, MPI_COMM_WORLD);

    //This function just show the first three elements of the matrix
    //MMPI_Scatter(matrix, send_count, MPI_INT, recv_data, recv_count, MPI_INT, ROOT, MPI_COMM_WORLD);

    //This function works great, but it not use the official declaration of the MPI_Scatter
    //MMPI_Scatter2(matrix, send_count, MPI_INT, recv_data, recv_count, MPI_INT, ROOT, MPI_COMM_WORLD);
    printArray(recv_data , recv_count);


    MPI_Finalize();
    return 0;
}


//http://mpitutorial.com/tutorials/mpi-scatter-gather-and-allgather/
void MMPI_Scatter(void* send_data, int send_count, MPI_Datatype send_datatype, 
                  void* recv_data, int recv_count, MPI_Datatype recv_datatype, 
                  int root, MPI_Comm communicator) {

    int np, rank;
    int i;
    MPI_Status status;
    MPI_Comm_size(communicator, &np);
    MPI_Comm_rank(communicator, &rank);

    printArray(send_data, np * np);

    if (rank == ROOT) {
        for (i = 0; i < np; i++) {
            MPI_Send(send_data + (i * send_count), send_count, send_datatype, i, 0, communicator);
        }
    }
    MPI_Recv(recv_data, recv_count, recv_datatype, root, 0, communicator, &status);
    printArray(send_data, np * np);

}

//Works great, but without pointer
void MMPI_Scatter2(int send_data[], int send_count, MPI_Datatype send_datatype, 
                   int recv_data[], int recv_count, MPI_Datatype recv_datatype, 
                   int root, MPI_Comm communicator) {

    int np, rank;
    int i;
    MPI_Status status;
    MPI_Comm_size(communicator, &np);
    MPI_Comm_rank(communicator, &rank);



    if (rank == ROOT) {
        for (i = 0; i < np; i++) {
            MPI_Send(send_data + (i * send_count), send_count, send_datatype, i, 0, communicator);
        }
    }
    MPI_Recv(recv_data, recv_count, recv_datatype, root, 0, communicator, &status); 
    printArray(recv_data, np);
}


int *createMatrix(int nRows, int nCols) {

  int *matrix;

  int h, i, j;

  if ((matrix = malloc(nRows * nCols * sizeof(int))) == NULL) {
    printf("Malloc error:");
    exit(1);
  }

  //Test values 
  for (h = 0; h < nRows * nCols; h++) {
    matrix[h] = h + 1;
  }

  return matrix;
}

更新 1:

我认为它与此链接中的信息有关: https://www.mpi-forum.org/docs/mpi-1.1/mpi-11-html/node71.html#Node71

有一行:

MPI_Send(sendbuf + i*sendcount*extent(sendtype), sendcount, sendtype, i.....)

但是我不知道如何处理extend(sendtype)

更新 2:

现在它可以工作了，但目前因为我自己知道数据类型

void MMPI_Scatter(void* send_data, int send_count, MPI_Datatype send_datatype, 
                  void* recv_data, int recv_count, MPI_Datatype recv_datatype, 
                  int root, MPI_Comm communicator) {

    int np, rank;
    int i;
    int size;

    MPI_Datatype type;
    type = MPI_INT;
    MPI_Type_size(type, &size);

    MPI_Status status;
    MPI_Comm_size(communicator, &np);
    MPI_Comm_rank(communicator, &rank);



    if (rank == ROOT) {
        for (i = 0; i < np; i++) {
            MPI_Send(send_data + ((i * send_count) * size), send_count, send_datatype, i, 0, communicator);
        }
    }

    MPI_Recv(recv_data, recv_count, recv_datatype, root, 0, communicator, &status);

}

更新 3(已解决):

void MMPI_Scatter(void* send_data, int send_count, MPI_Datatype send_datatype, 
                  void* recv_data, int recv_count, MPI_Datatype recv_datatype, 
                  int root, MPI_Comm communicator) {

    int np, rank;
    int i;
    int size;

    MPI_Datatype type;
    type = send_datatype;
    MPI_Type_size(type, &size);

    MPI_Status status;
    MPI_Comm_size(communicator, &np);
    MPI_Comm_rank(communicator, &rank);



    if (rank == ROOT) {
        for (i = 0; i < np; i++) {
            MPI_Send(send_data + ((i * send_count) * size), send_count, send_datatype, i, 0, communicator);

        }
    }
    MPI_Recv(recv_data, recv_count, recv_datatype, root, 0, communicator, &status);

}

更新 4

这个函数工作正常，因为使用了 ROOT，但是一旦从集体调用，ROOT 必须被 root 替换，如下所示:

if (rank == root) {
}

Answer 1

将 send_data + (i * send_count) 更改为:

send_data + (i * send_count + MPI_Type_size(send_datatype)