[Numpy-tickets] [NumPy] #913: max is bogus if nan is in the array

[NumPy]

#913: max is bogus if nan is in the array
--------------------+-------------------------------------------------------
 Reporter:  cdavid  |        Owner:  somebody
     Type:  defect  |       Status:  new     
 Priority:  normal  |    Milestone:  1.2.1   
Component:  Other   |      Version:  none    
 Severity:  major   |   Resolution:          
 Keywords:          |  
--------------------+-------------------------------------------------------
Comment (by cdavid):

 I attached a patch to handle those cases: the ufunc now put Nan in the
 output whenver NaN is encountered in any input. So in both cases (#913 and
 #914), the answer is now nan.

 Depending on the platform, the slowdown is significant for big (~1e5/1e6)
 arraygs:
  - 3x: linux 32 bits, P4, ubuntu gcc 4.2.3, glibc isnan
  - negligeable (Mac OS X; Linux 64 bits, Core 2 duo).

 I would think that the Pentium 4 is pathological, and that in most cases,
 the slowdown is limited. My computer with the P4 has a really peculiar
 FPU. I used this simple script for benchmark:

 {{{
 #!python
 from timeit import Timer

 import numpy as np
 print np.version.version
 print np.__file__

 a = np.array([5, 2, np.nan, -1, 2])
 b = np.array([5, 2, -1, 2, np.nan])
 print np.max(a)
 print np.max(b)

 setup = '''
 import numpy as np

 tests = [np.random.randn(2**n) for n in range(20)]
 '''
 timers = [Timer('np.max(tests[%d])' % i, setup=setup) for i in range(20)]

 for i in timers:
     print i.repeat(3, number=100)[0] / 100

 }}}

-- 
Ticket URL: <http://scipy.org/scipy/numpy/ticket/913#comment:1>
NumPy <http://projects.scipy.org/scipy/numpy>
The fundamental package needed for scientific computing with Python.