Physics in Half Precision

Please don't post Bullet support questions here, use the above forums instead.
Post Reply
clewin
Posts: 1
Joined: Thu Dec 26, 2013 7:52 pm

Physics in Half Precision

Post by clewin »

Hi all,

I'm a longtime lurker of this forum since it seems to be the only place on the internet where people who work on physics in games congregate.

I'm wondering if anyone here has experience simulating physical systems in reduced precision. I'm writing a mass-spring-like constrained particle simulation on the GPU. For various reasons I find myself needing to squeeze as much particle position/velocity data as possible into the 32kb of shared memory one is allowed in a DX11 compute shader. In my current implementation, all simulation is done in full precision and then updated particle positions/velocities are reduced to half precision and sent to shared memory ready for the next iteration of the solver. While storing these quantities in half precision doesn't cause anything catastrophic, there is a noticeable jitter in the resulting simulation. Obviously, this won't do.

So, do you think half precision is a dead end? What about storing as fixed-point 16 bits? Or more than that?

Best,
Chris
Post Reply