Re: tolerance in binopdf.m

[Ben Abbott]

On Sep 21, 2011, at 11:01 AM, Marco atzeri wrote:

> On 9/21/2011 3:07 PM, Ben Abbott wrote:
>> 
>> On Sep 21, 2011, at 1:59 AM, Jordi Gutiérrez Hermoso wrote:
>> 
>>> Just for fun, I asked someone to run this program on a Macintosh:
>>> 
>>>    #include<stdio.h>
>>>    #include<math.h>
>>> 
>>>    int main()
>>>    {
>>>      printf("%0.100f\n",lgamma(3.0));
>>>    }
>>> 
>>> It turns out it does run, i.e. it does have an lgamma implementation.
>>> However, the exact value it outputs is 390207173010335/2^49, while on
>>> my Debian system, the exact value it outputs is
>>> 31216573840826795/2^52. It seems that somewhere along the way, 3 bits
>>> of precision were lost on the Macintosh.
>>> 
>>> So I don't think there's anything we can do other than increase the
>>> tolerance by 3 bits to account for this. We already increase tolerance
>>> slightly for other systems for other tests.
>>> 
>>> HTH,
>>> - Jordi G. H.
>> 
>> Thanks for all the quick replies. From config.log I see ...
>> 
>> | #define HAVE_LGAMMA 1
>> | #define HAVE_LGAMMAF 1
>> | #define HAVE_LGAMMA_R 1
>> | #define HAVE_LGAMMAF_R 1
>> 
>> Thus, it looks to me as if Apple has a different implementation of lgamma 
>> (?).
>> 
>> As this is not a bug in Octave, I'm inclined to add a tolerance for MaOS.
>> 
>> However, I'm curious about what Apple did.
>> 
>> I'm using Xcode 4.1. The sources are below.
>> 
>>      http://www.opensource.apple.com/
>> 
>> Does anyone have an idea of where to look to find the sources for lgamma?
>> 
>> Ben
>> 
> 
> Could be here ?
> 
> http://opensource.apple.com/source/Libm/
> http://opensource.apple.com/source/Libm/Libm-2026/Source/Intel/xmm_erfgamma.c
> 
> Regards
> Marco

Yes, that is it!

Ben