GPU Physics fun: rigid body, fluids and 1.000.000 particles

[Erwin Coumans]

His GPU method is not just particles, the rigid bodies allow for stacking:

I put a new demo on my page. Bricks are stacked in this demo.
http://www.iii.u-tokyo.ac.jp/~takahiroh ... dBody.html

Harada-san is visiting GDC and we are meeting up again. Hope he will disclose/publish more of his approach soon.

Enjoy,
Erwin

[KenB]

Yes, true, the methods have broad phase and partially narrow phase in common, and since the rigid body implementation most likely is a relaxational method, they also have the pairwise iterative term in common (though they look different in the two different models of course, and SPH typically just computes a single pairwise instance for the force).
My experience is though, that it is easier to simulate 10.000 rigid bodies in real-time than 10.000 sph particles in real-time.
A fluid must be fairly incompressible, which is terribly expensive on the timestep, but it must still have dynamics gouverned by the SPH-many-body potential that _stores_ energy (i.e. it can't be hardcore all the way).
The rigid body pair potential need only be rigid, and doesn't need to store potential energy beyond the constraint stabilization, and this is not so demanding on the step size.

His GPU method is not just particles, the rigid bodies allow for stacking:

I put a new demo on my page. Bricks are stacked in this demo.
http://www.iii.u-tokyo.ac.jp/~takahiroh ... dBody.html

Harada-san is visiting GDC and we are meeting up again. Hope he will disclose/publish more of his approach soon.

Enjoy,
Erwin

[Dirk Gregorius]

The rigid body pair potential need only be rigid, and doesn't need to store potential energy beyond the constraint stabilization, and this is not so demanding on the step size.

Could you explain a little bit more what you exactly mean by this? How are the potential energy and stabilization connected (because it might add energy). And how does it effect the stepping?

For my understanding I linearize the non-linear position constraint for stabilization which leads of course to instabilities.

[Etherton]

Wow I just run the demo, and I have to say that this is without a doubt that fastest and the more stable staking demonstration I had ever seen.
I mean those pyramid not more a pixel while they are stacked, until they are touched by the box projectile, when they collapse to the floor.
It seems his method is faster at stacking rigid bodies than it is at simulating single bodies, I do not know how he does it, but it seems very impressive.

My experience is though, that it is easier to simulate 10.000 rigid bodies in real-time than 10.000 sph particles in real-time.

10,000 in real time easy or easier? wow that's some serious simulation there? what kind of physic libraries are doing that?

[KenB]

Oh, C-code looking at me from Visual Studio...
Far off from the real useful physics library. Just testing steppers, constraint stabilization and convergence of pairwise relaxation methods.

Wow I just run the demo, and I have to say that this is without a doubt that fastest and the more stable staking demonstration I had ever seen.
I mean those pyramid not more a pixel while they are stacked, until they are touched by the box projectile, when they collapse to the floor.
It seems his method is faster at stacking rigid bodies than it is at simulating single bodies, I do not know how he does it, but it seems very impressive.

My experience is though, that it is easier to simulate 10.000 rigid bodies in real-time than 10.000 sph particles in real-time.

10,000 in real time easy or easier? wow that's some serious simulation there? what kind of physic libraries are doing that?

[coderchris]

So, has he released any more details as to how he may be doing his broadphase?

[Erwin Coumans]

So, has he released any more details as to how he may be doing his broadphase?

I met Takahiro Harada again last week at SIGGRAPH. He wrote an article on his GPU rigid body dynamics approach for GPU Gems 3 (Addison Wesley), which just got published.

I highly recommend ordering GPU Gems 3, it also has other useful articles, including an article by Havok on implementing LCP on GPU for convex collision detection. Therefore, I won't spoil the facts, sorry

Thanks,
Erwin

[Erwin Coumans]

Takahiro Harada's latest work support over 1.000.000 particles for a dual SLI GPU configuration, using CUDA.

Related work will be shown at upcoming, from Prometech:
http://www.prometech.co.jp/

Some English translation will follow.
Erwin