[ODE] ODE-GIMPACT performance and advices

[Kenneth Bodin (Holmlund)]

Hmm, this comment never made it to the mailing list because
I mailed it from an address different from the subscription address.
Since I took the time to write it, it might as well resubmit it, although
it arrives very late in the discussion...
/Ken

Hi,
Agree, this ignorance has annoyed me too for quite some time.
There's a huge difference in N complexity between physics/dynamics
and traditional computer graphics (which is the origin of this 
misconception).
Physics has never had linear scaling, except in the static case...

Even "serious CG papers" report linear complexity in solvers, but
this is nonsense since what they actually find is that complexity
is N linear if they fix the number of iterations, and completely
forget about convergence and errors (and thus can only solve disconnected
blocked systems).
Even pairwise solvers are NlogN at best (temporal coherence not 
accounted for,
which is really what this discussion is about), and even over simplified 
iterative solvers
are N^3 or worse in the general, connected case -  if they produce 
anything plausible
at all without serious 1st order hacks.

Physics engines will not be able to handle the general case of "millions 
of entities"
(or even hundreds - with reasonable coulomb friction...) for many
years (lower estimate, multiply by 10-100), and therefore it is 
certainly not
efficient to model everything as being fully dynamic.

Just wanted to say this, because.... I'm an academic!
One can say funny things about game developers too :) For example,
they tend to think that methods in "game physics" were invented by "game 
developers",
but I'm not sure I know about a single original example :)

/Kenneth Bodin, Umeå university

PS Pierre (and others), speaking of blowing holes in a wall. Do
you know of any papers or other documentation about contact
generation  and CSG (constructive solid geometry)?





"Pierre Terdiman" <pierre.terdiman at novodex.com> wrote:

...

> Now it's getting interesting..... this is a recurrent topic among some
> people here. It's really typical: basically the guys with an academic
> background want to make "everything dynamic", in search of a unified
> solution, to be like "in the real world". On the other hand the guys with
> actual game experience tend to say "forget it, that's a red herring" (for
> both technical & gameplay-related reasons). It's theory vs practice.
>
> My opinion is that the static/dynamic distinction is not a thing of 
> the past
> yet. It's still there, as far as I know. People have been saying *for 
> years*
> that the next-gen PC/consoles will have enough power to make everything
> dynamic, but I didn't see it happen. Did it? We have more processing 
> power
> but we also have millions of triangles now. So, updating triangle-based
> acceleration structures is just as costly as before!






-- 
Kenneth Bodin (Holmlund)
VRlab/HPC2N, Umeå University
bodin<nospam>hpc2n.umu.se
(<nospam> = @)