Re: [eigen] Qt's container support

[Keir Mierle <mierle@xxxxxxxxx>]

On Tue, Jan 20, 2009 at 1:38 AM, Gael Guennebaud
<gael.guennebaud@xxxxxxxxx> wrote:
>
> Hi all,
>
> I think that being able to store Eigen's matrices in Qt's containers
> is very important (especially QVector<>), and just like std::vector,
> there are a couple of issues:
>
> Issue 1:
> QVector has a bug which avoids the use of QVector with types having a
> custom operator new
>  - problem solved in Qt 4.5
>  - for earlier versions defining
> EIGEN_WORK_AROUND_QT_BUG_CALLING_WRONG_OPERATOR_NEW_FIXED_IN_QT_4_5
> before including Eigen/Core seems to do the job.
> => PROBLEM SOLVED.
>
> Issue 2:
> the second problem is to enforce QVector to allocate aligned memory
> buffers. Fortunately, Qt's developers had the good idea to define
> wrapper functions around malloc/free/realloc called qMalloc, qFree
> etc.
> So the workaround is simply to redefine these functions to call
> ei_align_malloc. I added a Eigen/QtAlignedMalloc header which exactly
> does that (I'll commit it once the next issue will be solved).
> => PROBLEM SOLVED
>
> Issue3:
> The last issue I found is with QVector::fill(const T&,size_t) which is
> the equivalent of std::vector<T>::resize(size_t,T). Here the argument
> is passed by reference so no alignment issue. The problem is the use
> of the default ctor and operator= to initialize/copy the new elements
> while our operator= does not allow to set an uninitialized matrix.
> Solutions:
>  a) allow operator= to set/resize an uninitialized matrix

What is wrong with (a)? I'd like to have this anyway.

Consider

MatrixXd A, B, C, D;
A << some data values;

ComputeBCAndDFromA(A, &B, &C, &D); /* fails */

B.resize(...); /* resize to appropriate size */
C.resize(...); /* resize to appropriate size */
D.resize(...); /* resize to appropriate size */
ComputeBCAndDFromA(A, &B, &C, &D); /* works? */

void ComputeBCAndDFromA(...) {
  *B = ....
  *C = ...
  *D = ...
}

I realize you can use set to assign to B, C, and D inside
ComputeBCAndDfromA, but I consistently get this wrong and consider it
to be an API wart. For example, I can never remember when I have to
use .set() and when I can use operator=; especially with regards to
blocks.

What is necessary to make assigning to matrices and blocks consistent?

>  b) specialize QVector for Eigen's matrix type using a similar trick
> than for std::vector such that we only have to redefine the fill()
> method to use Matrix::set()
>  c) do nothing and enforce the user to use .resize() followed by a for
> loop instead of fill().
>
> I vote for solution b)
>
> any opinion ?
>
> Gael.
>
>