Feature Request Soft floating points calculations and determinism

Hello Unity

I'm very interested in Unity DOTS physics, but I'm facing a problem where cross-platform simulation is not deterministic.
This seems a bit obvious since floating point calculations are quite hardware bound.
the fix has already been implemented by someone on the internet and posted a modified version of the package using soft floating point which seems to work as expected.
the problem here is that it needs to be officially supported, so we don't have to modify the full package by replacing floats with soft floats on every update.

it will be amazing if unity team can integrate it and give us the option to enable it using something like define symbols.

this is a video showing the difference betwee using standard floats and soft floats

modified physics engine repo:
https://github.com/Kimbatt/unity-deterministic-physics

Thanks!

 

I think that's a compromise some people are willing to accept, because determinism is key for some projects.

 

thank you for your reactivity

 

This is interesting, the burst compiler seems to be pretty stable across arm and x86 chips when comes to simulation. Currently, the problem is sometimes we need to pass in values from mono behaviour(Input). this breaks certain system which requires withoutburst().run(). Thus breaking the deterministic simulation.

Wouldn't it more feasible just have the mathematics floating library can switch in and out between softfloat and regular float library?

 

Did you try adding the input as a singleton entity with a dynamic buffer? And set the dynamic buffer from a separate default world. You should be able to burst all your sim code. For networking just pass bytes around.

 

Your approach sounds feasible. This requires all logic code done inside ECS. I was wondering if it is possible for certain system withoutBurst().Run() be calculated in soft floats, then when reading and writing, I will cast between soft float and float.

 

Your entity components would store floats or soft floats? Since soft floats are software, casting would probably be much more expensive than traditional float/double/int conversion. I haven't measured anything though, so maybe I'm off.

EDIT:
If you want simulation determinism with soft float, you need to convert the user inputs to soft floats before sending them over the network / into your simulation. That way, all players have their simulation 100% in soft float.

 

No Need for Soft float on my parts, Arm8 and x86 follows IEEE 754 standards, with burst compiler keeping everything in check, Currently I do get Determinstic simulation. I just need to follow what TheOtherMonarch said about the dynamic buffer.