Looks awesome. I need water physics but it doesn’t need to be too realistic, even Minecraft’s water physic is quite enough for me. Is it possible to get some cheap water simulation on huge world?
@lubor_ If you need help figuring out how to get your project integrated with ue4 in an efficient manner, then you could also post on this forum. If you were a bit more transparent about your problems and why you can’t proceed, then the people that are interested in ue4 support may do research for you for free. I’ve been coming back to this forum for a while waiting for more news. I’d pitch in if it meant finishing the ue4 integration.
@ I’d love to know how you solved or resolved the GPU-CPU-GPU transfer problem. Have you achieved direct GPU-GPU data transfer on UE4?
@lubor_ Any eta on plugin release date, i thinks there’s a lot of demand, i’ve been following it from two years and i would really like to get my hands on it.
We got it to work with 4.14 some time ago, but the solution doesn’t work with 4.21 due to compatibility issues (macros in RHI changed and there isn’t much documentation). We also got it to work with Unity (good documentation) and everything works same fast as in our demo. However, our further tests revealed some problems. There was too much overhead, when our fluid should have interacted with other physics frameworks (PhysX / Bullet), due to different data representations etc. After some analysis we concluded we can do much better:) and decided to do our own semi-ML rigid body / destruction solver. We’ve been working on it in the last half year and so far all finished components look very promising. We do not want to publicly release anything before it is all practically useful without much support from our side.
@lubor_ Thank you for your reply. So did you make a kind of your original procedural mesh component like “GPU mesh component” using some RHI functions? I’ve also read Unity has a good documentation for using GPU resource directly.
When I made a real-time marching cubes with CUDA, GPU data transfer was a big problem…
https://forums.unrealengine.com/community/work-in-progress/1379694-real-time-marching-cubes-by-cuda-on-unreal-engine-4
Looking at the paper… I am no PhD candidate so correct me please: It looks like there needs to be a training of the neural network that can take days? Is that based on the environment of the to be simulated fluid? So it has to have a static surroundings?
How big is the data set of the neural network? Is this like “prerendering” the fluid behavior and if so how does that translate to a “new* real-time*” fluid engine? Or am I totally missing the way this works?
Training takes week(s), but when it gets released, it will already contain trained model and only evaluation code. The simulation works on completely new scenes - none of the meshes or animations in the released demo has been used during training.
Any news on development, how long you expect it to be finished?
I know that i would buy it right away.
As written before, we will not release the plugin, before it is practically useful for game development.
The downloadable demo we released last year showed what we could do at the time. It worked acceptably fast only for up to 5-10 moving obstacles, at most 1-2 animated. Interface with existing rigid solvers was not really feasible.
In the last half a year we developed a new learning method for handling rigid bodies. We train a data representation, which allows calculation of operations on meshes (collision, occupancy, distance, projection, …) super fast and scales well with the number of obstacles. It supports rigged meshes (could be pre-animated, simulated or controlled for example by kinect) with almost no overhead over static meshes (demo supported only pre-processed baked animations), So far we implemented only interactions of rigid / animated objects with fluids, but it is designed to work well also for interactions between rigid bodies (not implemented yet). It looks very promising and we assume it will end up much faster than existing rigid body solvers.
Regarding fluids, we implemented adaptive-size particles, which speeds up simulation part of the algorithm (not surface generation and rendering) approximately 2x with no visible difference. There is a lot of more work to be done there.
Sounds you’ve come a long way since the demo, in the mean time i’ll keep checking this thread for any updates, so please keep us posted.
I talked to creators over the email and they said they are not gonna do UE4 plugin since it’s too complicated after Epic did some updates it’s almost impossible to make a plugin, so we can forget about seeing this in action in UE4.I dont think they gonna be able to overcome this easily since they are small team and they are not the first ones who have problems with plugins in newer UE versions.
1 Like