[ODE] Particle/cloth/skeleton (was: Re: Choleski factorization)

[david@csworkbench.com]

The problem is, the worst parts of the current implementation is the
multiplication of an arbitrarily large matrix.  In the iterated solution,
it's multiplication of up to a 6x6 matrix with padding to 6x8.  And you
may need to add or subtract the results from a vector instead of replacing
them in some cases.... Yes, there are a few cases where generic D3D
functions might work, but for the worst internal stuff, you'll need more
robust functions.  So what I'm thinking is a couple of nasm-based files,
one implementation for each of 3dnow, sse (I can do these 2), sse2 (No can
do until Hammer arrives/is affordable), and a fourth in c for none of the
above, and either statically link them into the program at compile time or
get really fancy and dynamically link them at runtime.... but then we
start running into OS issues.  Either way, NASM is cross-platform and open
source, so the only issue there is that it adds a dependency other than
just a c++ compiler.  But it sure beats writing a different inline version
for gcc, vc++ (are even 6 and .NET the same?), borland, mingw....  4
compilers and 3 processor technologies and your already up to 12 different
implementations.

Just my 2 cents (and the way I'll probably implement it unless someone has
a better cross-platform open source idea),
David

> gl wrote:
>
>> > Be nice for all us win guys, that's for sure : ).  There is some
>> processor specific optimisation going on in D3DX (the DX utility
>> library that
>> provides
>> > quat, matrix and various vector classes/functions, amongst others),
>> and
>> yes,
>> > you do get it transparently and its updated over time.  Using that
>> library doesn't tie you to any other parts of DX in any way (eg.
>> there are no rendering dependencies).
>> >
>> > It's not clear though exactly what they've optimised - their vec3
>> equivialent doesn't use padding, so they probably only accelerate a
>> few complex operations.
>
> There's a 4-D vector in D3D as well though.
>
>> >  Matrices should be (you can use an aligned version),
>> > not sure about quats.  They do intend to keep adding optimisations
>> in key areas though.
>
>> I should also mention that you would also automatically get support
>> for AMD's 3DNow.
>
> And at *runtime*, not compile time.  This idea sounds better all the
> time ;-).
>
> --
> Gary R. Van Sickle
> Brewer.  Patriot.
>
>
> _______________________________________________
> ODE mailing list
> ODE@q12.org
> http://q12.org/mailman/listinfo/ode