[Opensim-dev] thread from ODE in regards to physics acceleration

[Dahlia Trimble]

It's hard to say how much complexity can fit into 1gb. I've had opensim
processes reach 3gb with around 22,000 prims but those prims were quite
complex as they were mostly cut dimpled spheres which contain many
individual triangles. I suspect that many regions approaching the 15,000
prim limit that LL imposes are mostly simple boxes scaled into different
sizes and as such will not use nearly as much memory. It's also not
currently known how effectively our memory is being used and how much may be
wasted due to the inability of the garbage collection process to reclaim
small chunks of memory that may be left after the proxy meshes are passed to
ODE. While our meshes are constructed in managed memory, they are copied to
"pinned" memory before being passed to ODE. This is necessary since ODE is
unmanaged code. Once the copying is complete the mesh objects should be
disposed but it's not known how well that memory can be reclaimed and if it
is tightly interspersed with the pinned memory which cannot be manipulated
by the garbage collector.
One possible optimization may be to page out areas of the simulation where
no collisions are taking place and replace it with large bounding boxes that
encompass many smaller proxies. I believe that LL may do something similar
to this with Havok and it helps them with their goal of fitting a full
simulator into 512 mb of ram (which i think has recently changed to 1 gb).



On Mon, Nov 3, 2008 at 6:10 AM, Eugen Leitl <eugen at leitl.org> wrote:

>
> Just FYI. Sorry for the formatting, Google Groups sucks.
>
> Eugen Leitl
> View profile
>         More options Oct 31, 1:51 pm
> From: Eugen Leitl <ele... at gmail.com>
> Date: Fri, 31 Oct 2008 05:51:40 -0700 (PDT)
> Local: Fri, Oct 31 2008 1:51 pm
> Subject: hardware acceleration
> Reply | Reply to author | Forward | Print | Individual message | Show
> original | Remove | Report this message | Find messages by this author
> Has there any work being done on accelerating some ODE features (say,
> collision detection) using ATI/nVidia GPUs? I wasn't able to find
> anything so far.
>
> Those familiar with both AMD Stream and CUDA, which architecture
> provides a better match for ODE?
>
> With 1 GByte on-card memory, how much scene complexity (e.g. in
> OpenSim) can I expect to handle without resorting swapping to main
> memory?
>
> Thanks.
>
>
> Daniel K. O.
> View profile
>         More options Nov 1, 12:37 am
> From: "Daniel K. O." <danielko.lis... at gmail.com>
> Date: Fri, 31 Oct 2008 21:37:44 -0200
> Local: Sat, Nov 1 2008 12:37 am
> Subject: Re: [ode-users] hardware acceleration
> Reply | Reply to author | Forward | Print | Individual message | Show
> original | Report this message | Find messages by this author
> Eugen Leitl escreveu:
>
> > Has there any work being done on accelerating some ODE features (say,
> > collision detection) using ATI/nVidia GPUs? I wasn't able to find
> > anything so far.
>
> I'm working on a dWorldStep implementation in CUDA, but it's not
> complete, nor optimized. Hopefully I'll have some time to finish it in
> two or three moths (hopefully with some other operations, like
> narrow-phase collision detection too).
>
> > Those familiar with both AMD Stream and CUDA, which architecture
> > provides a better match for ODE?
>
> I think NVIDIA is doing a better job promoting CUDA than AMD
> promoting... whatever it's the current name.
>
> Feature-wise they are incredibly similar. I don't have much experience
> with FireStream, but apparently it doesn't expose different types of
> memory like CUDA does, so you have to trust the platform (which IMHO you
> shouldn't for any new technology) to do the correct optimizations.
>
> --
> Daniel K. O.
> "The only way to succeed is to build success yourself"
>
>    Reply    Reply to author    Forward       Rate this post: Text for
> clearing space
>
>
> gl
> View profile
>         More options Nov 1, 1:08 am
> From: "gl" <g... at ntlworld.com>
> Date: Sat, 1 Nov 2008 00:08:35 -0000
> Local: Sat, Nov 1 2008 1:08 am
> Subject: Re: [SPAM?][ode-users] Re: hardware acceleration
> Reply | Reply to author | Forward | Print | Individual message | Show
> original | Report this message | Find messages by this author
>
> > Eugen Leitl escreveu:
> >> Has there any work being done on accelerating some ODE features (say,
> >> collision detection) using ATI/nVidia GPUs? I wasn't able to find
> >> anything so far.
>
> Apparently the Bullet guys are working on CUDA too, might be worth checking
> out:  http://www.bulletphysics.com/Bullet/wordpress/
> --
> gl
>
>    Reply    Reply to author    Forward       Rate this post:
>
>
> Erwin Coumans
> View profile
>         More options Nov 1, 1:16 am
> From: "Erwin Coumans" <erwin.coum... at gmail.com>
> Date: Fri, 31 Oct 2008 17:16:50 -0700
> Local: Sat, Nov 1 2008 1:16 am
> Subject: Re: [ode-users] Re: [SPAM?][ode-users] Re: hardware acceleration
> Reply | Reply to author | Forward | Print | Individual message | Show
> original | Report this message | Find messages by this author
>
> We are indeed working on a broadphase, narrowphase and constraint solver in
> CUDA for Bullet. This parallel constraint solver is similar to quickstep.
> Some results will be discussed at the Game Developers Conference in 2009.
>
> CUDA is probably the best way to go for now. It will likely be easy to port
> from CUDA to OpenCL in the future (both NVidia and AMD will support
> OpenCL).
> Hope this helps,
> Erwin
> _______________________________________________
> Opensim-dev mailing list
> Opensim-dev at lists.berlios.de
> https://lists.berlios.de/mailman/listinfo/opensim-dev
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://opensimulator.org/pipermail/opensim-dev/attachments/20081103/97f2897d/attachment-0001.html>