Repository navigation
np.finfo(np.clongdouble).precision wrong on powerpc #3125
Description
Hello,
I'm forwarding here the Debian bug http://bugs.debian.org/702169 :
Version: 1:1.6.2-1
some scipy tests fail on powerpc, this is because
import numpy as np;
np.finfo(np.clongdouble).precision
returns 75 on powerpc but it should return not more than about 18.
see e.g. the test_signaltools.TestCorrelateComplex256
https://buildd.debian.org/status/fetch.php?pkg=python-scipy&arch=powerpc&ver=0.11.0%2Bdfsg1-1&stamp=1362143334
the decimal is obtained in scipy/signal/tests/test_signaltools.py:553 via:
for datatype in [np.csingle, np.cdouble, np.clongdouble]:
cls = _get_testcorrelate_class(datatype, _TestCorrelateComplex)
cls.decimal = int(2 * np.finfo(datatype).precision / 3)
globals()[cls.__name__] = cls
Thanks,
Sandro
Activity
charris commentedon Mar 6, 2013
Yes, long double on power PC isn't handled correctly due to the fact that it isn't an ieee type, but is implemented as two doubles. I think long doubles need to be special cased but I don't know what current power PC's do or exactly what precision the two double representation actaully has. We need someone who has one to track these things down and test them.
jdgleeson commentedon Mar 14, 2013
The recommended value for the precision of the PPC double-double representation is 106 bits. I can test on Darwin + gcc.
charris commentedon Feb 21, 2014
Closing in favor of #2669.