[SciPy-Dev] arccosh/arctanh test precision

Ralf Gommers ralf.gommers@googlemail....
Thu Jul 8 07:17:20 CDT 2010


The arccosh test precision is set a bit too high for real input - it's kept
at the default level which is 5*eps. Setting it to 5e-14 solves the issue on
32-bit Windows, which is OK I think. On 64-bit Windows the results are
further off (see below) which could be considered an actual problem. The
arctanh result also looks bad. So I propose to change the 32-bit test
precision only. For the curious, it's line 34 of special/tests/test_data.py.

Is the above fine? Should I open a ticket for 64-bit?

Thanks,
Ralf



32-bit-Windows:
========================================
FAIL: test_data.test_boost(<Data for arccosh: acosh_data_ipp-acosh_data>,)
----------------------------------------------------------------------
<...>
Max |adiff|: 1.77636e-15
Max |rdiff|: 2.44233e-14



64-bit Windows (from the binaries by Christoph Gohlke, built with msvc9):
========================================
FAIL: test_data.test_boost(<Data for arccosh: acosh_data_ipp-acosh_data>,)
----------------------------------------------------------------------
<...>
AssertionError:
Max |adiff|: 1.77636e-15
Max |rdiff|: 1.09352e-13
<...>

======================================================================
FAIL: test_data.test_boost(<Data for arctanh: atanh_data_ipp-atanh_data>,)
----------------------------------------------------------------------
<...>
AssertionError:
Max |adiff|: 6.39488e-12
Max |rdiff|: 1.01982e-12
<...>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://mail.scipy.org/pipermail/scipy-dev/attachments/20100708/f03085d5/attachment.html 


More information about the SciPy-Dev mailing list