Re: [eigen] Support for true Array

[Rohit Garg <rpg.314@xxxxxxxxx>]

On Wed, Nov 25, 2009 at 6:18 PM, Benoit Jacob <jacob.benoit.1@xxxxxxxxx> wrote:
> 2009/11/25 Rohit Garg <rpg.314@xxxxxxxxx>:
>> 1) Are there any other benefits to having an array class apart from
>> getting rid of .cwise() to do element wise multiplications?
>
> Well, that in itself is already a big advantage, as many people would
> like to use Eigen's xpr templates and vectorization on plain arrays,
> using only cwise ops.
>
> But to answer your question, there is another benefit, as we mentioned
> above: such an Array class could be used as Scalar type, so a
> Matrix<Array...,3,1> would behave like an array of Vector3's but each
> scalar operation would become an array operation. So if you are
> interested in SIMD on Vector3's and you have a large number of them,
> this approach allows you to get it!
>
>> 2) I haven't understood how this would help the quaternion class and
>> the related things.
>
> Currently the question is: how to add xpr templates to Quaternion and
> eventually DualQuaternion without rewriting all the xpr classes? Gael
> had to solve the same problem for Array and his solution was to have a
> class that has an array interface but can wrap an arbitrary xpr. Then
> the operator+ between arrays can just return a usual 'sum' xpr, but
> wrapper like that in an array interface! Thus, only operators/methods
> need to be rewritten for Array, but the underlying xpr classes that
> they return don't have to be rewritten.
>
> Then I thought: why not do the same in Quaternion? We could have a
> QuaternionWrapper class that wraps any MatrixBase for its array of 4
> coefficients; then the operator+ between quaternions returns a
> QuaternionWrapper<ordinary sum xpr>....
>
>> 3) A million Vector3f's would still not be vectorized as they are not
>> a multiple of packet size, or will they? Is there a possibility to
>> automagically handle the SOA form instead of the usual AOS form with
>> this proposal?
>
> Yes, this is exactly what I meant. (Dont quite remember what AOS
> stands for but that seems to be it).

Array Of Structures - AoS
===================
struct float2{
float x,y};

float2 arr[N];

Structure Of Arrays - SoA
===================
struct float2
{
float x[N],y[N];}

SOA form has *big* advantages for simd ops over aos.
>
> This is about creating an Array class that can be used as a scalar
> type, and on which the arithmetic ops are vectorized. So when you have
>   Matrix<Array,3,1> u,v,w;
> and you do:
>   u = v+w;
> this expands (thanks to Eigen xpr templates and unrollers) to:
>   u[0] = v[0]+w[0];
>   u[1] = v[1]+w[1];
>   u[2] = v[2]+w[2];
> Each of these 3 lines is now an operation on Array objects and expands
> to a vectorized loop doing the addition.
>
> Benoit
>
>
>



-- 
Rohit Garg

http://rpg-314.blogspot.com/

Senior Undergraduate
Department of Physics
Indian Institute of Technology
Bombay