[Scilab-Dev] floating point operations in different versions of Scilab

[Jean-Pierre Dussault]

Le 2015-06-30 03:18, Clément David a écrit :
> Hello Stéphane,
>
> Sorry for the lag but I tried to investigate a little bit !
>
> Did you try to reproduce on a reduced test case which does not depends
> on fsqp ?
>
> For the 5.5.2 release, I did not notice anything that might impact the
> behavior of fsqp.

I might be slightly off topic, but floating point is sometimes 
unpredictable: scilab provides the function ieee to specify the 
behaviour of floating point operations. BUT, on my 5.5.1 under Ubuntu 14.04,

    -->ieee(0)
    -->exp(2000)
      ans  =
         Inf
    -->1/0
         !--error 27
    Division by zero...

    -->ieee(1)
    -->exp(2000)
      ans  =
         Inf
    -->1/0
      ans  =
         Inf

and according to the doc,    ieee(1) : floating point exception produces 
a warning...

The first exp(2000) should have produced an error, the second a warning 
and if using ieee(2), Inf. 1/0 should produce a warning with ieee(1) and 
correctly produces an error with ieee(0)


Unfortunately, the correct treatement of IEEE 754 floating point 
operations is very often very messy, not only in scilab. Now, what if 
the external code is compiled using different assumptions with respect 
to the desired behavior when "exeptions" or Inf are produced? An 
increased messyness!

Just my 2 cents on this topic!

JPD




>
> --
> Clément
>
> Le lundi 22 juin 2015 à 14:35 +0200, Stéphane Mottelet a écrit :
>> Hello Scilab devmasters !
>>
>> I have noticed subtle (but real) differences between successive
>> versions
>> of Scilab, here's the whole story:
>>
>> I use fsqp to solve an optimization problem where the computation of
>> the
>> cost function and its gradient uses a lot of sparse/full matrix
>> products, and I noticed different convergence behaviour of the same
>> code
>> depending on the Scilab version.
>>
>> E.g. with Scilab 5.5.0 and 5.5.1 fsqp converges within the prescribed
>>
>> tolerance (norm of the gradient), but with 5.5.2 fsqp stops its
>> iterations without reaching it. When relaxing the tolerance
>> successful
>> convergence is obtained but I got slightly different results at the
>> end.
>>
>> I am wondering about which pathway I should follow to identify the
>> problem, so my questions are the following :
>>
>> -is it possible that the scilab version which has been used to
>> compile
>> fsqp could produce such a behaviour ?
>> -did the sparse/sparse and sparse/full product routines change
>> recently ?
>>
>> Best regards,
>>
>> S.
>>
>>
> _______________________________________________
> dev mailing list
> dev at lists.scilab.org
> http://lists.scilab.org/mailman/listinfo/dev

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.scilab.org/pipermail/dev/attachments/20150701/65e635ea/attachment.htm>