[ODE] numerical stability (sort of)

tbp ode at ompf.org
Wed May 5 14:49:48 MST 2004


I've been toying with ODE for a while and at some point got bored enough 
to profile it and fancy the idea of cutting some corners.
I thought about measuring the time it takes some system (like various 
piles of boxes on an inclined ground) to converge (everything coming to 
rest and being auto disabled).

It wasn't a good idea, but my first surprise was that contrary to gcc 
3.4 built dll, msvc 7.1 ones (my own builds) don't even allow 
reproduceable runs within the same session (that is destroying box 
bodies+geoms and simulating again gives different result). A cold run 
gives reproduceable results. Really weird.
As http://ode.org/snapshots/bin/ proved empty (i guess it's still on the 
move), i cannot check against a semi official bin.
So my first question is: is there anyone kind enough to provide me with 
a recent snapshot msvc built dll (single precision, no need for opcode)?

Then, i wonder what kind of system would allow me to have reproduceable 
behaviour between compilers & compile options (while still being large 
enough to benchmark in a useful way). Obviously stacks of boxes that 
collapse hitting each other isn't what i'm looking for :)
Maybe tons of semi isolated couple of geoms? Any idea?


Regards,
	tbp.




More information about the ODE mailing list