Bug#845082: numexpr FTBFS on ppc64el: test failures
X-Debbugs-Cc: gin...@debian.org, bren...@br.ibm.com On 11/29/2016 03:18 PM, Fernando Seiti Furusato wrote: > Also I had to consider a ci failure on appveyor, that is why I also > included a verification for msvc version on the PR. By the way, it has been merged upstream. signature.asc Description: OpenPGP digital signature
Bug#845082: numexpr FTBFS on ppc64el: test failures
Hello. At first I tried to fix that with round(). Problem is round() returns the same type as the argument so the cast happens anyway. I would recommend using the latest version, with llround(), which returns long long. Also I had to consider a ci failure on appveyor, that is why I also included a verification for msvc version on the PR. Regards. -- Fernando Seiti Furusato IBM Linux Technology Center signature.asc Description: OpenPGP digital signature
Bug#845082: numexpr FTBFS on ppc64el: test failures
tags 845082 + pending thanks -- Antonio Valentino
Bug#845082: numexpr FTBFS on ppc64el: test failures
Hello.
It is noticeable that it fails on a 3^3 array operation.
On "scalar" operations, it doesn't show the same behavior.
Like so:
$ python3.5
Python 3.5.2+ (default, Nov 3 2016, 11:10:16)
[GCC 6.2.0 20161027] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> import numexpr as ne
>>> a=[3]
>>> b=[3]
>>> ne.evaluate("a ** b") # array operation
array([26], dtype=int64)
>>> ne.evaluate("3 ** 3") # scalar operation
array(27, dtype=int32)
$ python
Python 2.7.12+ (default, Nov 4 2016, 17:04:30)
[GCC 6.2.0 20161027] on linux2
Type "help", "copyright", "credits" or "license" for more information.
>>> import numexpr as ne
>>> a=[3]
>>> b=[3]
>>> ne.evaluate("a ** b")
array([26])
>>> ne.evaluate("3 ** 3")
array(27, dtype=int32)
I am not sure if it fails with other values as well.
I will keep looking.
Regards
--
Fernando Seiti Furusato
IBM Linux Technology Center
signature.asc
Description: OpenPGP digital signature
Bug#845082: numexpr FTBFS on ppc64el: test failures
On 11/20/2016 07:41 AM, Adrian Bunk wrote: > > Lots of failures like: > Yes. I just tried it here, and more than 40 tests failed. It is usually off by 1, and I am wondering if we are being bite by a similar issue I found on OpenJDK, where there are math inconsistency when using optimization higher than O1. https://gcc.gnu.org/bugzilla/show_bug.cgi?id=78386 Anyway, we are looking at this problem.
Bug#845082: numexpr FTBFS on ppc64el: test failures
Source: numexpr Version: 2.6.1-1 Severity: serious https://buildd.debian.org/status/fetch.php?pkg=numexpr=ppc64el=2.6.1-1=1479607449 Lots of failures like: ... == FAIL: test_scalar0_int_aggressive_OPERATIONS_0309 (__main__.TestExpressions) -- Traceback (most recent call last): File "numexpr/tests/test_numexpr.py", line 1011, in method return func() File "numexpr/tests/test_numexpr.py", line 583, in method neval, type(neval), shape(neval)) AssertionError: '3 ** (b+3)' (test_scalar=0, dtype='int', optimization='aggressive', exact=False, npval=array([27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27]) ( - (100,)) neval=array([26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26]) ( - (100,))) == ...
