Re: [eigen] Eigen2 to Eigen3 Migration Path

On 01/13/2011 09:06 PM, Benoit Jacob wrote:

2011/1/13 Hauke Heibel <hauke.heibel@xxxxxxxxxxxxxx>:

On Thu, Jan 13, 2011 at 12:13 AM, Gael Guennebaud
<gael.guennebaud@xxxxxxxxx> wrote:

The EIGEN2_SUPPORT way seems much more powerful. Let's be specific. If
project A wants to use Eigen3 as well as a lib B which relies on
Eigen2 then there are two cases:

1 - lib B is a pure template lib (or at least the parts which rely on
Eigen2) then EIGEN2_SUPPORT should do the job (we can make it better
if needed).

2 - lib B is a binary library with a public API involving Eigen2
objects. Then there is still the possibility to recompile B with
Eigen3+EIGEN2_SUPPORT.

3 - lib B is a binary library without Eigen2 in the public API.. Again, there
exists the possibility to recombile B with Eigen3+EIGEN2_SUPPORT.

Asumming for a moment we had Eigen:: and Eigen3:: namespaces. This would
imply the following things.

OK, so let's focus for now on Gael's solution (tell people to use
Eigen3 + EIGEN2_SUPPORT) plus my suggestion to make EIGEN2_SUPPORT as
complete as possible and validated by the actual eigen2 test suite. At
least it doesn't hurt anyone.

It's true that if Eigen3 + EIGEN2_SUPPORT were 100% API compatible
with Eigen2 then the lack of ABI compatibility wouldn't be a big
problem anymore since switching would be virtually costless.

I believe that if we do have full coverage of the Eigen2 API with Eigen3+EIGEN2_SUPPORT this would be fine for our use cases. We would just compile all libraries against EIgen3+EIGEN2_SUPPORT instead of Eigen2.

Tully

 Regarding kdl (http://www.orocos.org/kdl), I tried to compile it
against Eigen3+EIGEN2_SUPPORT and the major hickup I am seeing is caused by
Map's const-correctness. We should be able to fix that through the plugin by
re-enabling the non-const constructor of Maps though I did not properly
think this over.

Ah, ok. It should be quite easy to allow non-const-correct Map
constructors with EIGEN2_SUPPORT.

Benoit