C++ 如何在MPI-2+中复制MPI#u累积的功能；

我正在学习MPI-2/MPI-3中介绍的MPI单边通信，并且遇到了关于

MPI\u累积的以下内容：

MPI_累计允许调用者合并移动到
已存在数据的目标进程，例如
在目标进程上求和。同样的功能也可以通过以下方式实现：
使用MPI_Get检索数据（然后同步）；
在调用者处执行求和操作；然后使用MPI_Put发送
将更新的数据返回到目标进程。累积简化
这种混乱

但是，只有有限数量的操作允许与
MPI_累计
（最大、最小、总和、乘积等）一起使用，并且不允许使用用户定义的操作。我想知道如何使用
MPI\u Get
、sync、op和
MPI\u Put
实现上述messiness。是否有C/C++中的教程或工作代码示例

谢谢


为了进行测试，我修改了其中的一段代码，其中使用单边通信来创建一个整数计数器，该计数器在MPI进程之间保持同步。使用MPI_累计的目标问题行已标记

代码按原样编译，大约15秒后返回。但是，当我试图用问题行后面的注释块中显示的等效基本操作序列替换
MPI\u acculate
时，编译的程序无限期挂起

谁能帮我解释一下出了什么问题
在这种情况下，替换MPI_累计的正确方法是什么

顺便说一句，我用

g++ -std=c++11 -I..   mpistest.cpp -lmpi

并用

mpiexec -n 4 a.exe


代码：

使用get和put实现累加确实会非常麻烦，尤其是当您必须处理派生数据类型等时。但是，假设您正在对单个整数进行累加，并且只想将本地值求和到远程缓冲区中，则可以执行以下操作（仅限伪代码）：

您的代码不正确，因为您在GET之后放弃了独占锁，这会在两个进程同时尝试求和数据时导致原子性问题。
锁的粒度比fences（集体）更细，因此最好在可以的地方使用它们。但是要小心嵌套同步！你在一个已经被锁定的区域内设置了围栏，所以你被吊死了。还要注意的是，在那个问题/答案之后出现的MPI-3大大改进了片面的例程和语义。@JonathanDursi谢谢。我试过上锁/开锁。它仍然没有给出正确的结果。请参阅编辑。MPI-3.0标准改进了例程和语义。现有的实现没有：）如果这种情况太复杂而无法解释，我会问一个更简单的问题，在这种情况下，只有一个“全局”值用于记录进程中的“最小”值。即使在更简单的情况下，我也有一些同步问题。
//adpated from https://stackoverflow.com/questions/4948788/
#include <mpi.h>
#include <stdlib.h>
#include <stdio.h>
#include <thread>
#include <chrono>

struct mpi_counter_t {
  MPI_Win win;
  int  hostrank;  //id of the process that host values to be exposed to all processes
  int  rank;    //process id
  int  size;     //number of processes
  int  val;      
  int  *hostvals; 
};

struct mpi_counter_t *create_counter(int hostrank) {
    struct mpi_counter_t *count;

    count = (struct mpi_counter_t *)malloc(sizeof(struct mpi_counter_t));
    count->hostrank = hostrank;
    MPI_Comm_rank(MPI_COMM_WORLD, &(count->rank));
    MPI_Comm_size(MPI_COMM_WORLD, &(count->size));

    if (count->rank == hostrank) {
        MPI_Alloc_mem(count->size * sizeof(int), MPI_INFO_NULL, &(count->hostvals));
        for (int i=0; i<count->size; i++) count->hostvals[i] = 0;
        MPI_Win_create(count->hostvals, count->size * sizeof(int), sizeof(int),
                       MPI_INFO_NULL, MPI_COMM_WORLD, &(count->win));
    } 
    else {
        count->hostvals = NULL;
        MPI_Win_create(count->hostvals, 0, 1,
                       MPI_INFO_NULL, MPI_COMM_WORLD, &(count->win));
    }
    count -> val = 0;

    return count;
}

int increment_counter(struct mpi_counter_t *count, int increment) {
    int *vals = (int *)malloc( count->size * sizeof(int) );
    int val;

    MPI_Win_lock(MPI_LOCK_EXCLUSIVE, count->hostrank, 0, count->win);

    for (int i=0; i<count->size; i++) {

        if (i == count->rank) {
        MPI_Accumulate(&increment, 1, MPI_INT, 0, i, 1, MPI_INT, MPI_SUM,count->win); //Problem line: increment hostvals[i] on host
        /* //Question: How to correctly replace the above MPI_Accumulate call with the following sequence? Currently, the following causes the program to hang.
            MPI_Get(&vals[i], 1, MPI_INT, 0, i, 1, MPI_INT, count->win);
            MPI_Win_fence(0,count->win);
            vals[i] += increment;
            MPI_Put(&vals[i], 1, MPI_INT, 0, i, 1, MPI_INT, count->win);
            MPI_Win_fence(0,count->win);
        //*/
        } else {
            MPI_Get(&vals[i], 1, MPI_INT, 0, i, 1, MPI_INT, count->win);
        }
    }

    MPI_Win_unlock(0, count->win);

    //do op part of MPI_Accumulate's work on count->rank
    count->val += increment; 
    vals[count->rank] = count->val; 

    //return the sum of vals
    val = 0;
    for (int i=0; i<count->size; i++)
        val += vals[i];

    free(vals);
    return val;
}

void delete_counter(struct mpi_counter_t **count) {
    if ((*count)->rank == (*count)->hostrank) {
        MPI_Free_mem((*count)->hostvals);
    }
    MPI_Win_free(&((*count)->win));
    free((*count));
    *count = NULL;

    return;
}

void print_counter(struct mpi_counter_t *count) {
    if (count->rank == count->hostrank) {
        for (int i=0; i<count->size; i++) {
            printf("%2d ", count->hostvals[i]);
        }
        puts("");
    }
}


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

    MPI_Init(&argc, &argv);

    const int WORKITEMS=50;

    struct mpi_counter_t *c;
    int rank;
    int result = 0;

    c = create_counter(0);

    MPI_Comm_rank(MPI_COMM_WORLD, &rank);
    srand(rank);

    while (result < WORKITEMS) {
        result = increment_counter(c, 1);
        if (result <= WORKITEMS) {
             printf("%d working on item %d...\n", rank, result);
         std::this_thread::sleep_for (std::chrono::seconds(rand()%2));
         } else {
             printf("%d done\n", rank);
         }
    }

    MPI_Barrier(MPI_COMM_WORLD);
    print_counter(c);
    delete_counter(&c);


    MPI_Finalize();
    return 0;
}

int increment_counter(struct mpi_counter_t *count, int increment) {
    int *vals = (int *)malloc( count->size * sizeof(int) );
    int val;

    MPI_Win_lock(MPI_LOCK_EXCLUSIVE, count->hostrank, 0, count->win);

    for (int i=0; i<count->size; i++) {

        if (i == count->rank) {
            //MPI_Accumulate(&increment, 1, MPI_INT, 0, i, 1, MPI_INT, MPI_SUM,count->win); //Problem line: increment hostvals[i] on host
            ///* //Question: How to correctly replace the above MPI_Accumulate call with the following sequence? reports that 0 does all the work
            MPI_Get(&vals[i], 1, MPI_INT, 0, i, 1, MPI_INT, count->win);
            MPI_Win_unlock(0, count->win);
            vals[i] += increment;
            MPI_Put(&vals[i], 1, MPI_INT, 0, i, 1, MPI_INT, count->win);
            MPI_Win_lock(MPI_LOCK_EXCLUSIVE, count->hostrank, 0, count->win);
        //*/
        } else {
            MPI_Get(&vals[i], 1, MPI_INT, 0, i, 1, MPI_INT, count->win);
        }
    }

    MPI_Win_unlock(0, count->win);

    //do op part of MPI_Accumulate's work on count->rank
    count->val += increment; 
    vals[count->rank] = count->val; 

    //return the sum of vals
    val = 0;
    for (int i=0; i<count->size; i++)
        val += vals[i];

    free(vals);
    return val;
}

MPI_Win_lock(EXCLUSIVE);  /* exclusive needed for accumulate atomicity constraints */
MPI_Get(&remote_data);
MPI_Win_flush(win);  /* make sure GET has completed */
new = local_data + remote_data;
MPI_Put(&new);
MPI_Win_unlock();