Re: [eigen] Transform class performance and inconsistencies

[Christoph Hertzberg <chtz@xxxxxxxxxxxxxxxxxxxxxxxx>]

On 17.12.2012 12:24, Jakob Schwendner wrote:
> ok, thanks for confirming this. The other strange case, that you can see in my
> benchmark is:

For the record, I can't reproduce this using g++-4.7.1 (on 32bit linux). 
I get the following timings (using Jacob's benchmark):

> $ g++ benchGeometry.cpp -O2 -DNDEBUG -march=native -I .. -lrt  -mfpmath=sse && ./a.out
> vec = trans * vec
> float	 Isometry	 AutoAlign 3 0.0267s
> float	 Isometry	 DontAlign 3 0.0284s
> float	 Isometry	 AutoAlign 4 0.0189s
> float	 Isometry	 DontAlign 4 0.0208s
> float	 Projective	 AutoAlign 4 0.0111s
> float	 Projective	 DontAlign 4 0.0318s
> double	 Isometry	 AutoAlign 3 0.0257s
> double	 Isometry	 DontAlign 3 0.0407s
> double	 Isometry	 AutoAlign 4 0.0196s
> double	 Isometry	 DontAlign 4 0.0206s
> double	 Projective	 AutoAlign 4 0.0186s
> double	 Projective	 DontAlign 4 0.0258s
> vec = trans.matrix() * vec
> float	 Isometry	 AutoAlign 4 0.0109s
> float	 Isometry	 DontAlign 4 0.0172s
> double	 Isometry	 AutoAlign 4 0.0186s
> double	 Isometry	 DontAlign 4 0.0272s
> trans = trans1 * trans
> float	 Isometry	 AutoAlign   0.0619s
> float	 Isometry	 DontAlign   0.0765s
> double	 Isometry	 AutoAlign   0.0636s
> double	 Isometry	 DontAlign   0.0955s
> float	 Projective	 AutoAlign   0.0279s
> float	 Projective	 DontAlign   0.1029s
> double	 Projective	 AutoAlign   0.0565s
> double	 Projective	 DontAlign   0.1035s
> trans = trans1.matrix() * trans.matrix()
> float	 Isometry	 AutoAlign   0.0246s
> float	 Isometry	 DontAlign   0.0998s
> double	 Isometry	 AutoAlign   0.0569s
> double	 Isometry	 DontAlign   0.0941s

I also did not see any strange copies for the vector variants, the 
aligned variants all end with sth similar to
> 	mulps	%xmm2, %xmm0
> 	addps	%xmm1, %xmm0
> 	movaps	%xmm0, (%eax)
> 	ret	$4

However, all variants where the result does not fit into a register lead 
to a copy of the result (with aligned moves where possible, though), so 
either RVO is not optimal there, or not possible due to aliasing. And 
maybe that does not occur on 64bit machines, due to the higher number of 
registers.

Still, the timings show that e.g. aligned Isometry * X could be 
optimized to Projective * X in most cases (maybe unless the last line of 
the Isometry is not stored explicitly)


As for the non-compiling vec3 = Projective * vec3 from your first mail, 
I don't think this should be allowed, since for arbitrary Projective 
matrices you can't make any assumptions about the last component of the 
result.


Regards,
Christoph

--
----------------------------------------------
Dipl.-Inf., Dipl.-Math. Christoph Hertzberg
Cartesium 0.049
Universität Bremen
Enrique-Schmidt-Straße 5
28359 Bremen

Tel: +49 (421) 218-64252
----------------------------------------------