Re: [ublas] vector of matrices

From: Michael Stevens (m-stevens_at_[hidden])
Date: 2005-01-22 12:06:26

On Friday 21 January 2005 20:10, Toon Knapen wrote:
> Ralf Denzer wrote:
> > Hi Michael,
> >
> >> Does that fix you problem?
> >
> > Unfortunately, no.

I asked because the sematics have changed from Boost_1_32_0 and the current 
development version which is based on uBLAS_pure. In uBLAS_pure this is not 
supposed to work! Then the correct solution as Gunter suggested is to make 
the object be constructable from zero.

> patched. Actually the resize zero-initialised the new elements. This
> does not work of course for value_type's that have no constructor that
> takes an int.

Toon,

I think is incorrect for the current development HEAD. It should be committed 
only to the Boost_1_32_0 branch were the value_type() semantics are specified 
in documentation for resize. Even then such ad-hoc patching is not good. 
There are several more locations in vector.hpp the should be fixed for 
consistency! If I get some time I see I can come up with a consistent patch.

uBLAS_pure specificly made the distinction by requiring elements be 
constructible from zero. This also means that the hundred or so location in 
uBLAS where this is required are easy to find! I think at the uBLAS meeting 
we discussed more user friendly solutions but they require consistent 
application and detail thought about the semantics.

Personally I think the zero construable requirement has the advantage that it 
is simple and requires no template magic external to the class type. Also 
specifying element requirements and semantics that are consistent and 
efficient for both dense and sparse containers is not easy!

Hopefully this argument makes sense. Can we revert the patch to HEAD and apply 
it to the Branch? Please!

Michael
-- 
___________________________________
Michael Stevens Systems Engineering
Navigation Systems, Estimation  and
                 Bayesian Filtering
    http://bayesclasses.sf.net
___________________________________