[NumPy-Tickets] [NumPy] #1863: Buggy results when operating on array copied with astype()

[NumPy Trac]

#1863: Buggy results when operating on array copied with astype()
------------------------+---------------------------------------------------
 Reporter:  embray      |       Owner:  somebody   
     Type:  defect      |      Status:  new        
 Priority:  normal      |   Milestone:  Unscheduled
Component:  numpy.core  |     Version:  devel      
 Keywords:              |  
------------------------+---------------------------------------------------
 I'm getting really buggy results from any mathematical operation I do on
 an array of float64 that was obtained by calling `arr.astype('float64')`
 on an existing array of float64s.

 Here's an example that illustrates the problem (the fft stuff is
 irrelevant; it just creates an array similar to the one my code is bugging
 out on):

 {{{
 #!python
 import numpy
 from numpy import fft


 shape = (32, 16)
 type = numpy.complex64
 true = numpy.arange(shape[0] * shape[1], dtype=type)
 true.shape = shape

 y = fft.fft(true, shape[0], 0)
 y = y.imag

 z = y.astype(numpy.float64)

 print (y ** 2)[3]
 print (z ** 2)[3]
 }}}

 This particular example also throws a `RuntimeWarning: overflow
 encountered in square` when squaring `z`, the copy of the original array
 `y`.

 This just started happening recently using the latest from git.  I'm
 betting it has something to do with
 [https://github.com/numpy/numpy/commit/a17a4996e4ed63d1b855a0917fb5fcdd5855a7d0
 this commit] that made changes to astype(), but I haven't tracked down the
 exact issue yet.

-- 
Ticket URL: <http://projects.scipy.org/numpy/ticket/1863>
NumPy <http://projects.scipy.org/numpy>
My example project