
Ublas : 
Subject: Re: [ublas] [bindings][lapack] Interface design
From: Thomas Klimpel (Thomas.Klimpel_at_[hidden])
Date: 20081203 19:46:38
Rutger ter Borg wrote:
> Suppose we take LAPACK's problem type as a base, not their actual driver
> routines (Thomas' suggestion of 'xxev.hpp' is along this line). See LAPACK
> User's Guide, especially tables 2.2, 2.3, 2.4, 2.5, etc.. In other words,
> we might have functions called similar to these problem types (I've
> included some suggestions after 'or'):
We should try to ensure that it's both easy to find the name of the binding for a given lapack routine, and easy to find the name of the lapack routine for a given binding. Using the LAPACK User's Guide for that purpose is sufficiently easy, as long as it's stated explicitly enough in the numeric bindings documentation where to look in the LAPACK User's Guide in order to deduce the names.
> Next, let's take the difficult case mentioned by Thomas, geevx. I think we
> can achieve an easy and consistent interface by compiletime
> typedecoration of our variables. Herewith some examples which might not be
> perfect at their first shot, it's for the idea, suppose we just say
>
> lapack::eigen( A, left_eigen_values( vl ) );
> lapack::eigen( A, right_eigen_values( vl ) );
> lapack::eigen( A, right_eigen_values( vr ), left_eigen_values( vl ),
> eigen_vectors( W ) );
The decorations are a nice idea, since they allow to state the role of certain arguments.
> lapack::eigen( symmetric_lower( A ), left_eigen_values( vl ) );
> lapack::eigen( hermitian_upper( A ), right_eigen( v ) );
symmetric_lower and hermitian_upper fall into a different category. These are some sort of typerefinement. I thought about these before in the context of interpreting a vector as a matrix, because a vector might be interpreted as a (1,n) or as a (n,1) matrix, and this sort of typerefinement allows to distinguish the two cases.
> Now, why I am happy with this?
>
> * it is very readable
Indeed, but it also raises some questions. The boost parameter library also allows to specify the role (/name) of the arguments at the calling site. For me, this raises the question whether the role should always be specified, or only in the cases where it is necessary.
> * we might dispatch not only to xgeevx: other algorithms may be selected at
> compiletime, like the divideandconquer algorithm
The selection at compiletime is a good point.
> * no expert lapackknowledge is required by the user (a part of the posts
> to this list is actually about finding the right lapack driver)
I don't like this one. From my point of view, the numeric bindings library is concerned with providing good C++ interfaces to existing external numerical libraries. So it should enable its users the application of typical C++ programming techniques like generic programming, but it should limit itself to remove the restrictions of the original interfaces caused by fortran or C. LAPACK is differnt from Matlab, and the numeric bindings library should not try to turn LAPACK into Matlab. The user of the numeric bindings library should still be willing to read the original documentation of the LAPACK routine he uses, in case he runs into problems.
> * there is zero performance penalty in spite of the clear syntax (exploit
> power of C++)
Except compile times, but C++ programmers are used to it.
> Now, if a user wants to have more flexibility and/or access to the
> implementation details, we can offer them the stuff generated by the
> bindings generator, that may be used for lowlevel C++ calls to lapack:
>
> * templated calls to all driver / computational routines, by calling
> detail::xxxxx( .... ) (or without detail::, since the actual interface
> has other names)
without detail please, the user should not try to exploit implementation details.
> * detail::optimal_workspace_xxxx( .... );
> * detail::minimal_workspace_xxxx( .... );
The workspace handling should not be done by different function names, but by a workspace argument which defaults to optimal_workspace.
> What do you think? Let's figure out a good set of names for decoration...
symmetric_lower, symmetric_upper, hermitian_lower, hermitian_upper, matrix_1_n, matrix_n_1 + other refined types for cases with insufficient information in the available types.
left_eigen_vectors, right_eigen_vectors, left_singular_vectors, right_singular_vectors, eigen_values, singular_values, right_hand_sides, permutation_matrix, index_range, value_range, abs_tolerance, transposed, reflectors ...
I'm not really sure whether this is the right way to go, but I tried to enumerate at least some of the potential roles in order to get a better feeling for this.
How are you planing to implement this? I could imagine free template functions of the form
struct left_eigen_vectors_tag;
template <class T> boost::tagged_type<T,left_eigen_vectors_tag> left_eigen_vectors(T& ref) {
return boost::tagged_type<T,left_eigen_vectors_tag>(ref);
}
, but I don't know whether there is such a thing as boost::tagged_type.
Regards,
Thomas