[NumPy-Tickets] [NumPy] #1574: 1.5.0b1 umath complex failures on Windows

[NumPy Trac]

#1574: 1.5.0b1 umath complex failures on Windows
------------------------+---------------------------------------------------
 Reporter:  rgommers    |       Owner:  somebody
     Type:  defect      |      Status:  new     
 Priority:  normal      |   Milestone:  1.5.0   
Component:  numpy.core  |     Version:  devel   
 Keywords:              |  
------------------------+---------------------------------------------------

Comment(by pv):

 Yep, `HAVE_CSQRT`, `HAVE_CEXP`, et al. are defined, so it's using the
 platform-provided routines. And those appear to be buggy on that platform.

 It's not so clear how to fix this:

 1) It's a bug, but,

 2) it's not a bug in Numpy,

 3) Probably, most users won't care if `exp(inf+0j) == inf+0j` or
 `inf+nanj`.

 One could re-check that the results in `test_umath_complex.py` are
 specified by C99 as mandatory and not as optional, though.

 If we want to provide reliable corner-case behavior on all platforms, we
 should check some simple cases on compile time.

 But if we don't really care, we could as well add a check in
 `test_umath_complex.py` that sets
 {{{
 functions_seem_flaky = (np.exp(complex(np.inf, 0)).imag != 0)
 skip_complex_tests = have_platform_functions and functions_seem_flaky

 def platform_skip(func):
     return dec.skipif(skip_complex_tests,
         "Numpy is using complex functions (e.g. sqrt) provided by your
 platform's C library. "
         "However, they do not seem to behave according to C99 -- so C99
 tests are skipped.")(func)
 }}}
 and sprinkle `platform_skip` to all C99-specific tests. I'm not sure how
 to best check for the platform functions though.

 I think just skipping these tests is a sensible option.

-- 
Ticket URL: <http://projects.scipy.org/numpy/ticket/1574#comment:1>
NumPy <http://projects.scipy.org/numpy>
My example project