Re: which ATLAS version to use ?

[Sergei Steshenko]

----- Original Message -----
> From: Michael Goffioul <address@hidden>
> To: Sergei Steshenko <address@hidden>
> Cc: marco atzeri <address@hidden>; "address@hidden" <address@hidden>
> Sent: Saturday, March 31, 2012 11:55 AM
> Subject: Re: which ATLAS version to use ?
> 
> On Sat, Mar 31, 2012 at 9:52 AM, Sergei Steshenko <address@hidden> 
> wrote:
>> 
>> 
>> 
>> 
>>  ----- Original Message -----
>>>  From: marco atzeri <address@hidden>
>>>  To: address@hidden
>>>  Cc:
>>>  Sent: Saturday, March 31, 2012 10:58 AM
>>>  Subject: Re: which ATLAS version to use ?
>>> 
>>>  On 3/31/2012 9:41 AM, Sergei Steshenko wrote:
>>>>   Hello,
>>>> 
>>>>   I know that at the moment of writing this the last stable ATLAS 
> version is
>>>  3.8.4.
>>>> 
>>>>   But I vaguely remember that for Windows a newer version (from the 
> unstable
>>>  branch) was recommended in this list - I don't remember which 
> version
>>>  though.
>>>> 
>>>> 
>>>>   So, which ATLAS version should better be used from the point of 
> view of
>>>  being as bug-free as possible ?
>>>> 
>>>>   Thanks,
>>>>      Sergei.
>>> 
>>>  Hi Sergei
>>>  the last I built on cygwin atlas-3.9.68 seems to work fine.
>>>  All octave test are passed
>>> 
>>>  Regards
>>>  Marco
>> 
>>  Thanks.
>> 
>>  It's not that Octave built-in tests fail in my case - are there bugs 
> fixed in atlas-3.9.68, but still present in atlas-3.8.4 ?
> 
> From a stability point of view, it looks like it's the other way around:
> https://mailman.cae.wisc.edu/pipermail/octave-maintainers/2012-March/027510.html
> 
> I'm using 3.8.4 and didn't notice any problem.
> 
> Michael.

Well, then I'll stick to 3.8.4 until new stable release.
...
OTOH see http://wiki.scilab.org/Dependencies%20of%20Scilab%205.X :

"

Please notice that Atlas 3.8.0 have a bug causing wrong answer for row-major 
gemm C=A*A' or A'A (See bug 3624, bug 3640) 
Atlas 3.8.2 have a bug causing a crash (See bug 3767, bug 3827) 
Atlas 3.9.4 fix all these problems (used on Windows). 
Other implementations will be searched on the system (MKL lib, PhiPACK, CXML, 
Sun Performance library, SCSL, SGIMATH, IBM ESSL ...) but the 
operational team has not (and probably won't) test them in depth. 

"

- do they really mean 3.9.4 or it's a typo and they actually mean 3.8.4 ?


Thanks,
  Sergei.




>