[ODE] Faster ODE

[Henri Hakl]

Hey! :)

you're right... I found the missing comma - the file on the website has been
updated to reflect that as well.

Please check the accuracy and speed using the testsuite provided with ODE.
In the fbuild subdirectory there is a "test_ldlt.cpp" - compile it and run
with parameter f  ("test_ldlt f"). This should test both the speed and the
accuracy of the factorizer. Try to compare both the original and the
replacement version... - just because I think it is faster doesn't mean
much - there may be compiler optimizations that I'm not aware of and other
subtle issues that I'm not taking into account.

thanks :)
  Henri

----- Original Message -----
From: "Niko Nevatie" <nnevatie@welho.com>
To: <ode@q12.org>
Sent: Thursday, November 21, 2002 8:59 PM
Subject: Re: [ODE] Faster ODE


> I tried out the code in 'fastldlt_henri.c'. It works as expected. There
was
> a small typo somewhere in the beginning of one function, but it caused no
> probs. I cannot say anything about the speed of the code, as I haven't
> benchmarked anything yet. But hey, if you say it's faster than the
original,
> I'll take your word on it :)
>
> ----- Original Message -----
> From: Henri Hakl
> To: ode@q12.org
> Sent: Thursday, November 21, 2002 11:49 AM
> Subject: [ODE] Faster ODE
>
>
> Hi ODE :)
>
> Please can somebody look at this and experiment with the file?
>
> LDLT decomposition is one of the primary computational activities in ODE's
> physics engine. LDLT = Lower Diagonal Lower Transposed, meaning a
constraint
> matrix A (that describes the physics of a world) can be decomposed into
two
> matrices (L and D, where L is a lower triangular matrix (all entries above
> the diagonal are zero) and D is a diagonal matrix (all entries except the
> diagonal are zero)). Now  A = L D L'  (A equals the lower matrix times the
> diagonal matrix times the transposed lower matrix.) The A matrix varies in
> size and is typically between 10x10 and 500x500 for ODE (depending on
> how many constraints/collisions/etc need to be solved for a given
simulation
> step).
>
> This LDLT decomposition needs to be solved with every step and is quite
> computationally intensive. I've looked at the stldlt.c file that is part
of
> the ODE source and it looks to me like it could be improved upon.
> Unforetunately I use Delphi, so I cannot test my suggested modifications
> directly (though the Delphi version seems fine and considerably faster
then
> the original).
>
> The file http://www.cs.sun.ac.za/~henri/fastldlt_henri.c is a replacement
> for fastldlt.c and should (barring errors) work flawlessly and faster then
> the original.
>
> Let me know if it works, thanks :)
>   Henri
>
> _______________________________________________
> ODE mailing list
> ODE@q12.org
> http://q12.org/mailman/listinfo/ode