Re: [Help-glpk] [Fwd: Interface for exact solutions]

[Nico Del Piano]

Hi Heinrich,

thanks a lot for the code review.

> so your current changeset is in branch get_exact_values.
> https://github.com/nicodelpiano/glpk/commits/get_exact_values

Yes, I will be working there for this enhancement.

> Do we need three different scmp structures and three initalization routines?

No, of course. I just did it that way as I was trying to not break
anything from the previous functionality. Now I have integrated them
into the `glp_smcp` structure.

> You should move the check to before malloc.

Sure, I don't know why I did it before the NULL check.

> Why do you malloc at all? You are calling mpz_get_str with str = NULL.

You are right. I have changed that and works fine for the exact
rational results. For the fixed precision case, if I do not malloc
here 
https://github.com/nicodelpiano/glpk/commit/693d02b81fc2b7608cf1087cd31bbeb56f6390ea#diff-152a5a13f47885a5d7b50b4ed1db48a8R309
then the statement
https://github.com/nicodelpiano/glpk/commit/693d02b81fc2b7608cf1087cd31bbeb56f6390ea#diff-152a5a13f47885a5d7b50b4ed1db48a8R470
 will not work.

> Can't the result transfer be integrated in the existing glp_exact() function?

Yes, definitely, I have put them into that function.

> mpz_get_str should be called with a custom allocator using xmalloc and
> xfree. Use mp_set_memory_functions to set custome allocation function.

I tried to do this, but I couldn't make it work. I guess this
customization should be done before any GMP allocation (when there are
no active GMP objects) and that is why I am getting errors. Is it too
bad to leave the default allocation instead? If you have any
suggestions please let me know.

> smcp->exact_size is only used in malloc statements. That is strange.
> Shouldn't the following code consider the value you have chosen for
> exact_size?

I have changed the things a bit. I put three parameters in `glp_smcp`
structure that I think are the ones we need:
- num_states: this specifies the number of states of the solutions
- precision: an integer that specifies the precision in bits
- n_digits: number of decimal digits (this is used here
https://github.com/nicodelpiano/glpk/commit/693d02b81fc2b7608cf1087cd31bbeb56f6390ea#diff-152a5a13f47885a5d7b50b4ed1db48a8R470
we can get rid of this parameter)

Another thing I changed are the calls to `set_d_eps`: as I needed the
exact results I put `mpq_set_d` when loading data, for example here
https://github.com/nicodelpiano/glpk/commit/693d02b81fc2b7608cf1087cd31bbeb56f6390ea#diff-152a5a13f47885a5d7b50b4ed1db48a8R165
If GMP is not active we should put this between HAVE_GMP flag, am I right?

Sorry for the excessive amount of links, if something is not clear
enough, let me know.

Thanks!

Nico.

2016-07-13 17:19 GMT-03:00 Heinrich Schuchardt <address@hidden>:
> Hello Nico,
>
> so your current changeset is in branch get_exact_values.
> https://github.com/nicodelpiano/glpk/commits/get_exact_values
>
> Thank you for your contribution.
>
> I have started reviewing the changes. Some parts of the code may have to
> be adjusted before Andrew can integrate them:
>
> Do we need three different scmp structures and three initalization
> routines? I suggest to add the additional parameters to the existing
> smcp structure.
>
> smcp->exact_size is only used in malloc statements. That is strange.
> Shouldn't the following code consider the value you have chosen for
> exact_size?
>
> You dereference parm for malloc. Afterwards you check if it is NULL.
> You should move the check to before malloc.
>
> Why do you malloc at all? You are calling mpz_get_str with str = NULL.
> So you request allocation of memory here.
>
> smcp is a parameter structure. You should not put results there.
> Why can't you use the glp_prob structure for transfering the results?
>
> Do we really need glp_exact_sol()? Can't the result transfer be
> integrated in the existing glp_exact() function?
>
> mpz_get_str should be called with a custom allocator using xmalloc and
> xfree. Use mp_set_memory_functions to set custome allocation function.
> This will make it easier to identify mermory leaks.
>
> When I checked out the branch it contained a superfluous file "prec".
>
> Best regards
>
> Heinrich Schuchardt
>
>
> On 07/12/2016 03:39 PM, Nico Del Piano wrote:
>> Hello Heinrich,
>>
>> thanks for your fast reply and sorry for the code, yes, it was a complete 
>> mess!
>>
>> I created a new cleaned branch, this is the specific commit
>> https://github.com/nicodelpiano/glpk/commit/da11fac3996d9ed1012f1c50e76971e20f2b343e
>>
>> I am using gmp, so I first configure the project with --with-gmp.
>>
>> Best regards,
>>
>> Nico.
>>
>> 2016-07-12 9:48 GMT-03:00 Heinrich Schuchardt <address@hidden>:
>>> Hello Nico,
>>>
>>> your mail had to be forwarded manually because you're not subscribed to the 
>>> GLPK Help list.
>>>
>>> In GitHub the history of the exactsolver branch is messed up.
>>>
>>> You have been committing all sorts of files which are not source like 
>>> configure.log, •.o, ...
>>>
>>> I cannot find any meaningful commit messages.
>>>
>>> Please, create a clean branch.
>>>
>>> Best regards
>>>
>>> Heinrich Schuchardt
>>>
>>> http://www.xypron.de
>>>
>>> Am 12.07.16 um 14:35 schrieb Andrew Makhorin
>>>
>>>> -------- Forwarded Message --------
>>>> From: Nico Del Piano <address@hidden>
>>>> To: address@hidden
>>>> Subject: Interface for exact solutions
>>>> Date: Tue, 12 Jul 2016 08:44:39 -0300
>>>>
>>>> Hi all,
>>>>
>>>> I'm using the exact simplex solver provided by glpk to compute exact
>>>> solutions from a project written in Java (using the Java bindings
>>>> java-libglpk) and I haven't found a way to retrieve those exact
>>>> values: I'm getting double precision solutions.
>>>>
>>>> To overcome this, I defined a not-so-pretty interface to store these
>>>> exact values (I provided two ways to do this, both for rational
>>>> precision and fixed precision) here [1]. Basically, I just copied the
>>>> `glp_smcp` structure and added some parameters to properly store the
>>>> solutions in the desired precision. I also added two solvers
>>>> `glp_exact_sol` (for rationals) and `glp_fixed_sol` (for user-defined
>>>> precision). For example, if we use the `glp_fixed_sol` solver, we can
>>>> specify in the structure the precision we want, and store there the
>>>> solutions, in order to get those values later on.
>>>>
>>>> Obviously, this doesn't seem to be the best approach, so the purpose
>>>> of this email is to discuss how to improve this interface, or, even
>>>> better, how to add this feature in glpk. Furthermore, if you think
>>>> this can be done in glpk and I'm just ignoring it, please let me know.
>>>>
>>>> Thanks,
>>>>
>>>> Nico.
>>>>
>>>> [1]: https://github.com/nicodelpiano/glpk/tree/fixed_precision
>>>>
>>>>
>>>>
>>>>
>>>> _______________________________________________
>>>> Help-glpk mailing list
>>>> address@hidden
>>>> https://lists.gnu.org/mailman/listinfo/help-glpk
>>