Re: [eigen] Small dynamic matrix optimization

[Christoph Hertzberg <chtz@xxxxxxxxxxxxxxxxxxxxxxxx>]

If you can ensure that the allocated memory is aligned, you can use 
`Map<Matrix<...>, Aligned>`.  Also, we do allow vectorization of 
non-aligned data (with newer CPUs there is barely a performance 
difference, except when loads/store are split between cache-lines).

Btw: There is a macro `ei_declare_aligned_stack_constructed_variable` in 
Eigen/src/Core/util/Memory.h which already does most of what you need to 
do (grep the source code to see where/how it is used).

Cheers,
Christoph


On 21/10/2019 16.21, Rob Conde wrote:
> Thanks...I will think about this and maybe try to prototype something.
>
> Using the Map approach...I think this would disable SIMD right? Because you don't know if the mapped data is aligned?
>
> Rob
> ________________________________
> From: Christoph Hertzberg <chtz@xxxxxxxxxxxxxxxxxxxxxxxx>
> Sent: Monday, October 21, 2019 10:15 AM
> To: eigen@xxxxxxxxxxxxxxxxxxx <eigen@xxxxxxxxxxxxxxxxxxx>
> Subject: Re: [eigen] Small dynamic matrix optimization
>
> This is already done for some temporaries (like blocks in
> cache-optimized GEMM). Doing that for custom types would be interesting,
> but not trivial (e.g., move-behavior is completely different, if you
> want to move the data to an object higher in the stack).
>
> If you are ok with managing data yourself, you can manually achieve this
> by alloca/allocate-ing data yourself and use an Eigen::Map to refer to
> the data.
>
> If you have a proof-of-concept how this could be integrated into the
> Eigen API, we are open for suggestions!
>
> Another idea would be to allow custom allocators for matrices -- in that
> case the custom allocator could very easily be a stack-allocator (which
> would not even need to use the "main" stack). But that's also not too
> easy to integrate without breaking the current API (adding a new type
> would of course be an option for that).
>
> Cheers,
> Christoph
>
>
>
> On 21/10/2019 15.39, Rob Conde wrote:
> > I'm wondering if a small dynamic matrix optimization has been considered. In other words, allowing a stack allocation for matrices of unknown size up to a certain limit and using a heap allocation beyond that?
> >
> > Rob Conde
>
> --
>    Dr.-Ing. Christoph Hertzberg
>
>    Besuchsadresse der Nebengeschäftsstelle:
>    DFKI GmbH
>    Robotics Innovation Center
>    Robert-Hooke-Straße 5
>    28359 Bremen, Germany
>
>    Postadresse der Hauptgeschäftsstelle Standort Bremen:
>    DFKI GmbH
>    Robotics Innovation Center
>    Robert-Hooke-Straße 1
>    28359 Bremen, Germany
>
>    Tel.:     +49 421 178 45-4021
>    Zentrale: +49 421 178 45-0
>    E-Mail:   christoph.hertzberg@xxxxxxx
>
>    Weitere Informationen: http://www.dfki.de/robotik
>     -------------------------------------------------------------
>     Deutsches Forschungszentrum für Künstliche Intelligenz GmbH
>     Trippstadter Strasse 122, D-67663 Kaiserslautern, Germany
>
>     Geschäftsführung:
>     Prof. Dr. Jana Koehler (Vorsitzende)
>     Dr. Walter Olthoff
>
>     Vorsitzender des Aufsichtsrats:
>     Prof. Dr. h.c. Hans A. Aukes
>     Amtsgericht Kaiserslautern, HRB 2313
>     -------------------------------------------------------------

--
 Dr.-Ing. Christoph Hertzberg

 Besuchsadresse der Nebengeschäftsstelle:
 DFKI GmbH
 Robotics Innovation Center
 Robert-Hooke-Straße 5
 28359 Bremen, Germany

 Postadresse der Hauptgeschäftsstelle Standort Bremen:
 DFKI GmbH
 Robotics Innovation Center
 Robert-Hooke-Straße 1
 28359 Bremen, Germany

 Tel.:     +49 421 178 45-4021
 Zentrale: +49 421 178 45-0
 E-Mail:   christoph.hertzberg@xxxxxxx

 Weitere Informationen: http://www.dfki.de/robotik
  -------------------------------------------------------------
  Deutsches Forschungszentrum für Künstliche Intelligenz GmbH
  Trippstadter Strasse 122, D-67663 Kaiserslautern, Germany

  Geschäftsführung:
  Prof. Dr. Jana Koehler (Vorsitzende)
  Dr. Walter Olthoff

  Vorsitzender des Aufsichtsrats:
  Prof. Dr. h.c. Hans A. Aukes
  Amtsgericht Kaiserslautern, HRB 2313
  -------------------------------------------------------------