具有mpi4py的子数组数据类型

Question

我正在做一个简单的程序，我试图使用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，但我不断遇到分段错误。

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

当使用4个进程运行时，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

我对Create_subarray是否有错误的主意，否则还有什么不对的地方？

Answer 1

问题出在你的“位移”上。 您已经用'sizes'，'subsizes'和'starts'正确描述了一个子数组类型，但是当您替换该类型时，您已经将类型描述移到了实际2d数组的末尾。

替换其他类型通常不是问题，但是对于子数组，您将具有“原始数组”概念，这将带来问题。

我为VECTOR做了一个快速替换，但是仍然出现内存损坏。