MPI_gather中的致命错误（使用MPI进行矩阵乘法） [英] Fatal error in MPI_gather (matrix multiplication using MPI)

问题描述

您好，我最近刚刚在MPI（OpenMP）之后乱搞我的并行编程！我运行了一个代码，用于将虚拟机（Ubuntu）上的矩阵与4个内核相乘。不幸的是，我收到了一个错误：

Hello, I've just recently been messing around with my parallel programming in MPI (OpenMP after this!) and I ran a code to multiply matrices on a virtual machine (Ubuntu) with 4 cores. Unfortunately, I got an error:

PMPI_Gather中的致命错误：无效的缓冲区指针，错误堆栈：

PMPI_Gather（856）：MPI_Gather（sbuf = 0x601d20，scount = 100，MPI_FLOAT，rbuf = 0x601d20，rcount = 100，MPI_FLOAT，root = 0，MPI_COMM_WORLD）失败
PMPI_Gather（797）：缓冲区不能别名

Fatal error in PMPI_Gather: Invalid buffer pointer, error stack:
PMPI_Gather(856): MPI_Gather(sbuf=0x601d20, scount=100, MPI_FLOAT, rbuf=0x601d20, rcount=100, MPI_FLOAT, root=0, MPI_COMM_WORLD) failed
PMPI_Gather(797): Buffers must not be aliased

这是从SpeedGoComputing获得的代码：

This is the code obtained from SpeedGoComputing:

#include <mpi.h>

const int size = 20;

float a[20][20];
float b[20][20];
float c[20][20];

void multiply(int istart, int iend)
{
	int i,j,k;
    for (i = istart; i <= iend; ++i) {
        for (j = 0; j < size; ++j) {
            for (k = 0; k < size; ++k) {
                c[i][j] += a[i][k] * b[k][j];
            }
        }
    }
}

int main(int argc, char* argv[])
{
    int rank, nproc;
    int istart, iend;

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

    if (rank == 0) {
        // Initialize buffers.
		int i;
		int j;
        for (i = 0; i < size; ++i) {
            for (j = 0; j < size; ++j) {
                a[i][j] = (float)i + j;
                b[i][j] = (float)i - j;
                c[i][j] = 0.0f;
            }
        }
    }

    // Broadcast matrices to all workers.
    MPI_Bcast(a, size*size, MPI_FLOAT, 0,MPI_COMM_WORLD);
    MPI_Bcast(b, size*size, MPI_FLOAT, 0,MPI_COMM_WORLD);
    MPI_Bcast(c, size*size, MPI_FLOAT, 0,MPI_COMM_WORLD);

    // Partition work by i-for-loop.
    istart = (size / nproc) * rank;
    iend = (size / nproc) * (rank + 1) - 1;

    // Compute matrix multiplication in [istart,iend]
    // of i-for-loop.
    // C <- C + A x B
    multiply(istart, iend);

    // Gather computed results.
    MPI_Gather(c + (size/nproc*rank),
               size*size/nproc,
               MPI_FLOAT,
               c + (size/nproc*rank),
               size*size/nproc,
               MPI_FLOAT,
               0,
               MPI_COMM_WORLD);
			   



    if (rank == 0) {
        // Compute remaining multiplications
        // when size % nproc > 0.
        if (size % nproc > 0) {
            multiply((size/nproc)*nproc, size-1);
        }
    }

    MPI_Finalize();
    return 0;
}

我的尝试：



我怀疑MPI_Gather的sendbuf参数有问题，所以我将其更改为MPI_IN_PLACE，但即使这样也行不通。我也尝试了MPI_gatherV，但是没有用。



我试过在Visual Studio中调试它，它说有一个断点。在if（rank == 0）：行。也许它与指针和地址有关，我不太确定。我会读它。

如果有人可以提供帮助，请谢谢！我不是很擅长C.

What I have tried:

I suspect there's something wrong with the sendbuf argument of MPI_Gather, so I changed it to MPI_IN_PLACE but even that didn't work. I tried MPI_gatherV too and that didn't work.

I've tried debugging this in Visual Studio and it says there was a breakpoint reached. At the "if (rank == 0)": line. Perhaps it has something to do with pointers and addresses, I'm not very sure. I will be reading up on it.
If anyone could help, please and thank you! I'm not very adept at C.

推荐答案

在你的电话中打电话给

In your call to

MPI_Gather(c + (size/nproc*rank),
    size*size/nproc,
    MPI_FLOAT,
    c + (size/nproc*rank),
    size*size/nproc,
    MPI_FLOAT,
    0,
    MPI_COMM_WORLD);

发送和接收缓冲区相同，不使用 MPI_IN_PLACE [ ^ ]选项。

这篇关于MPI_gather中的致命错误（使用MPI进行矩阵乘法）的文章就介绍到这了，希望我们推荐的答案对大家有所帮助，也希望大家多多支持IT屋！