[Numpy-tickets] [NumPy] #716: Inconsistent behavior, i686 vs x86_64

[NumPy]

#716: Inconsistent behavior, i686 vs x86_64
---------------------------------+------------------------------------------
 Reporter:  jamie@goonathon.net  |       Owner:  somebody
     Type:  defect               |      Status:  new     
 Priority:  normal               |   Milestone:  1.0.5   
Component:  Other                |     Version:  none    
 Severity:  normal               |    Keywords:          
---------------------------------+------------------------------------------
 Gentoo, numpy-1.0.4-r2.

 Possible related to http://scipy.org/scipy/numpy/ticket/669

 Simple example,
 {{{
 [either i686 or x86_64]
 >>> numpy.uint64(418465321) / 8928000
 46.871115703405017
 >>> numpy.uint32(418465321) / 8928000
 46
 }}}

 Example that actually bit me,
 {{{
 [i686]
 >>> A = numpy.array([1, 2, 3, 4], numpy.uint8)
 >>> B = numpy.array([2, 4, 6, 8], numpy.uint8)
 >>> C = B - A
 >>> S = C.sum()
 >>> type(S)
 <type 'numpy.uint32'>
 >>> L = len(C)
 >>> type(L)
 <type 'int'>
 >>> S/L
 2

 [x86_64]
 >>> A = numpy.array([1, 2, 3, 4], numpy.uint8)
 >>> B = numpy.array([2, 4, 6, 8], numpy.uint8)
 >>> C = B - A
 >>> S = C.sum()
 >>> type(S)
 <type 'numpy.uint64'>
 >>> type(L)
 <type 'int'>
 >>> S/L
 2.5
 }}}

 Comments:

 1) numpy.uint64 / int -> float
    numpy.uint32 / int -> int

    It seems like these should at least evaluate to the same type.

 2) Should .sum() return a numpy.uint64, even on 32-bit platforms,
    for the sake of consistency?

-- 
Ticket URL: <http://scipy.org/scipy/numpy/ticket/716>
NumPy <http://projects.scipy.org/scipy/numpy>
The fundamental package needed for scientific computing with Python.