On 2/17/10 8:26 PM, Raymond Toy wrote:
> On 2/17/10 5:31 PM, Dieter Kaiser wrote:
>
>> Am Mittwoch, den 17.02.2010, 16:21 -0500 schrieb Raymond Toy:
>>
>>
>>> I'm trying to run the testsuite with the current CVS code. I'm getting
>>> quite a few failures now, especially with bigfloat accuracy of special
>>> functions like gamma_incomplete, Fresnel functions and erf. The
>>> tolerance is a little too tight now.
>>>
>>> Before I look further into this, is any one else having problems? I see
>>> the same issue with clisp, cmucl, and ecl.
>>>
>>>
>> I have no problems with the testsuite and sbcl 1.0.29.11 or clisp 2.44.1
>> on Linux.
>>
>>
>>
> Hmm. This is clisp 2.47 and cmucl (cvs) on Mac OS X. rtest_gamma, test
> 273 fails on both with an error of 2.476b-30 instead of the expected
> 1b-30. There are other failures like this where the expected accuracy
> is not achieved. Then there are a few failures in rtest_expintegral
> where expintegral_e continued fractions failed.
>
> I don't understand what's going on. Since bfloat arithmetic ought to be
> the same everywhere, I don't understand why you pass but I do not. More
> investigation needed....
>
Recompiling the world seems to have taken care of most issues. The
testsuite takes a very long time to run, however. Ecl takes a good 45
minutes on a 2 GHz macmini, and it seems a large fraction of the time is
spent on rtest_integrate. rtest_gamma and rtest_expintegral also takes
quite a bit of time.
CMUCL takes 700 sec, which is reasonable. But I'm also seen some kind
of intermittent error where $dependencies is somehow set to ((mlist
simp) 0 . 0). Kill doesn't like this.
I haven't tried clisp yet.
Ray