并行程序设计基础——MPI不连续数据发送（2）

目录

一、地址函数

二、与函数类型有关的调用

三、下界标记类型和上界标记类型

四、打包与解包

五、小结

---

        上一节我们介绍了MPI不连续发送中定义不同数据类型的接口，这一节我们继续对MPI不连续数据发送的其它知识点进行介绍。开始之前，我们先以一个编程示例对上节内容进行回顾。

#include <stdio.h>
#include <stdlib.h>
#include "mpi.h"
#define NUMBER_OF_TESTS 10

int main(int argc, char *argv[])
{
    int blocklens[2];
    int rank, i, j, k, n, stride, nloop;
    double *buf, *lbuf;
    double t1, t2, tmin;
    MPI_Aint indices[2];
    MPI_Datatype old_types[2];
    MPI_Status status;
    register double *in_p, *out_p;

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

    n = 1000;
    stride = 24;
    nloop = 100000/n;
    
    buf = (double*)malloc(n*stride*sizeof(double));
    if(!buf)
    {
        printf("Could not allocate send/recv buffer of size %d.\n", n*stride);
        MPI_Abort(MPI_COMM_WORLD, 1);
    }

    lbuf = (double*)malloc(n*sizeof(double));
    if(!lbuf)
    {
        printf("Could not allocate send/recv lbuffer of size %d.\n", n);
        MPI_Abort(MPI_COMM_WORLD, 1);
    }

    if(rank == 0)
        printf("Kind\tn\tstride\ttime(sec)\tRate(MB/sec)\n");

    MPI_Type_vector(n, 1, stride, MPI_DOUBLE, &vec1);
    MPI_Type_commit(&vec1);

    tmin = 1000;
    for(k = 0; k < NUMBER_OF_TESTS; k++)
    {
        if(rank == 0)
        {
            MPI_Sendrecv(MPI_BOTTOM, 0, MPI_INT, 1, 14, MPI_BOTTOM, 0, MPI_INT, 1, 14, MPI_COMM_WORLD, &status);
            t1 = MPI_Wtime();
            for(j = 0; j < nloop; j++)
            {
                MPI_Send(buf, 1, vec1, 1, k, MPI_COMM_WORLD);
                MPI_Recv(buf, 1, vec1, 1, k, MPI_COMM_WORLD, &status);
            }
            t2 = (MPI_Wtime() - t1)/nloop;
            if(t2 < tmin) tmin = t2;
        }
        else if(rank == 1)
        {
            MPI_Sendrecv(MPI_BOTTOM, 0, MPI_INT, 0, 14, MPI_BOTTOM, 0, MPI_INT, 0, 14, MPI_COMM_WORLD, &status);
            for(j = 0; j < nloop; j++)
            {
                MPI_Recv(buf, 1, vec1, 0, k, MPI_COMM_WORLD, &status);
                MPI_Send(buf, 1, vec1, 0, k, MPI_COMM_WORLD);
            }
        }
    }

    tmin = tmin / 2.0;
    if(rank == 0)
    {
        printf("Vector\t%d\t%d\t%f\t%f\n", n, stride, tmin, n*sizeof(double)*1.0e-6/tmin);
    }

    MPI_Type_free(&vec1);

    blocklens[0] = 1;
    blocklens[1] = 1;
    indices[0] = 0;
    indices[1] = stride * sizeof(double);
    old_types[0] = MPI_DOUBLE;
    old_types[1] = MPI_UB;
    MPI_Type_struct(2, blocklens, indices, old_types, &vec_n);
    MPI_Type_commit(&vec_n):

    tmin = 1000;
    for(k = 0; k < NUMBER_OF_TESTS; k++)
    {
        if(rank == 0)
        {
            MPI_Sendrecv(MPI_BOTTOM, 0, MPI_INT, 1, 14, MPI_BOTTOM, 0, MPI_INT, 1, 14, MPI_COMM_WORLD, &status);
            t1 = MPI_Wtime();
            
            for(j = 0; j < nloop; j++)
            {
                MPI_Send(buf, n, vec_n, 1, k, MPI_COMM_WORLD);
                MPI_Recv(buf, n, vec_n, 1, k, MPI_COMM_WORLD, &status);
            }
            t2 = (MPI_Wtime() - 1)/nloop;

            if(t2 < tmin) tmin = t2;
        }
        else if(rank == 1)
        {
            MPI_Sendrecv(MPI_BOTTOM, 0, MPI_INT, 0, 14, MPI_BOTTOM, 0, MPI_INT, 0, 14, MPI_COMM_WORLD, &status);
            for(j = 0; j < nloop; j++)
            {
                MPI_Recv(buf, n, vec_n, 0, k, MPI_COMM_WORLD, &status);
                MPI_Send(buf, n, vec_n, 0, k, MPI_COMM_WORLD);
            }
        }
    }

    tmin = tmin/2.0;
    if(rank == 0)
    {
        printf("Struct \t%d\t%d\t%f\t%f\n", n, stride, tmin, n*sizeof(double)*1.0e-6/tmin);
    }

    MPI_Type_free(&vec_n);

    tmin = 1000;
    for(k = 0; k < NUMBER_OF_TESTS; k++)
    {
        if(rank == 0)
        {
            MPI_Sendrecv(MPI_BOTTOM, 0, MPI_INT, 1, 14, MPI_BOTTOM, 0, MPI_INT, 1, 14, MPI_COMM_WORLD, &status);
            t1 = MPI_Wtime();
            
            for(j = 0; j < nloop; j++)
            {
                for(i = 0; i < n; i++)
                    lbuf[i] = buf[i*stride];

                MPI_Send(lbuf, n, MPI_DOUBLE, 1, k, MPI_COMM_WORLD);
                MPI_Recv(lbuf, n, MPI_DOUBLE, 1, k, MPI_COMM_WORLD, &status);

                for(i = 0; i < n; i++)
                    buf[i*stride] = lbuf[i];
            }

            t2 = (MPI_Wtime() - t1)/nloop;
            if(t2 < tmin) tmin = t2;
        }
        else if(rank == 1)
        {
            MPI_Sendrecv(MPI_BOTTOM, 0, MPI_INT, 0, 14, MPI_BOTTOM, 0. MPI_INT, 0, 14, MPI_COMM_WORLD, &status);
            
            for(j = 0; j < nloop; j++)
            {
                MPI_Recv(lbuf, n, MPI_DOUBLE, 0, k, MPI_COMM_WORLD, &status);
                for(i = 0; i < n; i++)
                    buf[i*stride] = lbuf[i];
                for(i = 0; i < n; i++)
                    lbuf[i] = buf[i*stride];
                MPI_Send(lbuf, n, MPI_DOUBLE, 0, k, MPI_COMM_WORLD);