Question

我正在制作一个简单的程序，我正在尝试使用mpi4py从8x8矩阵发送4 x 4矩阵子矩阵。我的方法是使用子数组数据类型，但我不断得到分段错误。

我的代码如下：

import numpy as np
from mpi4py import MPI

comm = MPI.COMM_WORLD
rank = comm.Get_rank()

sizes = (8,8)    # Matrix dimensions
subsizes = (4,4) # Sub-matrix dimensions
displs = (0, 4, 32, 36)

counts = (1, 1, 1, 1) 
starts = (0,0)

mynum = counts[rank]
glob = np.empty((8,8), dtype = 'f')
local = np.empty((4,4), dtype = 'f')

if rank == 0:  # Fill each quadrant with 0-3

    for row in range(0,4):
        for col in range(0,4):
            glob[row, col] = 0
    for row in range(0,4):
        for col in range(4,8):
            glob[row, col] = 1
    for row in range(4,8):
        for col in range(0,4):
            glob[row, col] = 2
    for row in range(4,8):
        for col in range(4,8):
            glob[row, col] = 3
    print glob

submatrixType = MPI.DOUBLE.Create_subarray(sizes, subsizes, starts, order =    MPI.ORDER_C)
submatrixType.Commit()

sendbuf = [glob, counts, displs, submatrixType]
recvbuf = [local, 16, MPI.DOUBLE]

comm.Scatterv(sendbuf, recvbuf, root = 0)
# or use
#comm.Scatterv(sendbuf, recvbuf, MPI.DOUBLE)

print rank, " " ,local

运行4个进程时，我的结果如下：

Proc 0: [[0. 0. 0. 0.] 
[1. 1. 1. 1.]
[0. 0. 0. 0.]
[1. 1. 1. 1.]]

Traceback (most recent call last):
Traceback (most recent call last):
  File "main.py", line 41, in <module>
      File "main.py", line 41, in <module>
    comm.Scatterv(sendbuf, local, root = 0)
comm.Scatterv(sendbuf, local, root = 0)
  File "Comm.pyx", line 454, in mpi4py.MPI.Comm.Scatterv (src/mpi4py.MPI.c:67458)
  File "Comm.pyx", line 454, in mpi4py.MPI.Comm.Scatterv (src/mpi4py.MPI.c:67458)
mpi4py.MPImpi4py.MPI.Exception: .Exception: MPI_ERR_TRUNCATE: message truncated
MPI_ERR_TRUNCATE: message truncated
Traceback (most recent call last):
  File "main.py", line 41, in <module>
    comm.Scatterv(sendbuf, local, root = 0)
  File "Comm.pyx", line 454, in mpi4py.MPI.Comm.Scatterv (src/mpi4py.MPI.c:67458)
mpi4py.MPI.Exception: MPI_ERR_TRUNCATE: message truncated

我是否对Create_subarray有错误的想法或者其他什么可能是错的？

Answer 1

问题在于您的排名＆＃39;。您已经正确地描述了一种子阵列类型，其中包含＆＃39;尺寸＆＃39;，＆＃39; subsizes＆＃39;和＆＃39;启动＆＃39;但是当您取代此类型时，您可以使用＆＃39; ve将类型描述从实际的2d数组的末尾移开。

取代其他类型通常不是问题，但是对于子阵列，你有这个＆＃34;原始阵列＆＃34;这个概念会产生问题。

我为VECTOR做了快速替换，但我仍然有内存损坏。

带有mpi4py的子数组数据类型

1 个答案: