Here it is, it is maintained by Sebastián Barschkis
Fluid Mantaflow GPU
2 Likes
This is so frustrating. Mantaflow is in a completely broken, unusable state due to completely broken update system, where many changes do not properly update simulation, yet we are getting a whole new GPU support with last commit Sep 13?
It totally escapes me how a developer can keep working on a giant new features to a system and at the same time constantly refuse to fix bugs which make the said system literally unusable:
⚓ T77170 Mantaflow smoke preview will not update till domain settings are changed ← This has been reported almost year and a half ago!
I honestly think Mantaflow should be flat out removed from Blender if such conduct continues, because right now it serves as a trap for users who may assume that Blender actually has fluid simulation capabilities, while in reality that can not be used properly.
This would be okay if this was someone’s personal Blender branch, but we are talking about a feature bundled in official Blender releases, which should be production ready (There’s no beta sticker on it). One thing is if the developer would not be aware of the issues. But actual active refusal to fix issues of this extent, which invalidate the entire feature, is just crazy.
14 Likes
YeaH, I think we should just make some noise about this. Because it is very frustrating this issue is still around.
The whole everything nodes with the focus on simulation is the perfect opportunity to make noise about Mantaflow’s issues
yeah, there’s not even a mantaflow user feedback post
I think I may have some good news coming to You soon 
Still an early alpha and the gvdb library had to be a bit rewritten to make it work for me but I have some initial tests with good results so I will show it the near time. Not in blender yet but built on top of gvdb voxels, however with less limitations, better performance, and more functionality, so stay tuned 
25 Likes
23 Likes
You re simulating and meshing 80m particles in real time ? 
1 Like
Yup 
I could actually do more but at this moment it’s eating too much GPU memory and 80M is the maximum I can fit entirely on VRAM.
BUT
I’ve managed to overcome the limitation by using mapped host memory (regular RAM) for the particles that cannot fit on the GPU so I can simulate over 150M particles, however at about 0.25-0.0.5fps because of the memory bus slowdown
10 Likes
those are nice numbers !
1 Like
I’m working on it in my free time (it’s my project). But I’m alone, not the smartest and without budget so it’s not the fastest development ever but I’m trying my best. So I don’t think it’s possible for blender 3.1 or 3.2 but for 3.3 maybe I could manage to make the early beta
16 Likes
Hello! I don’t understand much about these things but will this paper be useful to you:
https://www.preprints.org/manuscript/202105.0762/v1/download
This seems to be an open paper.
amd hip can translate cuda, did experiments on cyclesx experience much better than the optimization of many years of opencl, do not know if hip will reduce the workload than opencl
I have been beating my head against the wall about mantaflow being broken and yes there is basically no movement on making it work. To the point where I have been asked why I was trying to do a high res ( 400 on the resolution slider ) sim. This is not a question to ask when being handed a bug report.
Mantaflow in blender needs to be fixed. Even now with 3.0 its a broken mess
4 Likes
While I understand your (and ludovicKoutny’s) frustration, I think it’s unfair to vent it in this topic over here. This topic is from a developer who is trying to make something useful/better/different. He is afaik not the one who did the mantaflow stuff in blender. It would be (way) more useful to praise michal1000w and help him get this working than to complain about some other developers code which happens to work in the same domain (fluidsim). I agree stuff needs to happen in mataflow (I can’t really get it to work reliably either) but complaining on some unrelated developer’s feedback thread is not too useful.
This is a completely different developer.
My post was a response to this post:
Not the OP.
I fully support any effort towards making something to replace Mantaflow.
i seen the NVIDIA Introduce nanoVDB for support OPENVDB on realtime with support GPU
i think is time to blender for update openvdb on GPU BASED COMPUTE AND LEAVE THE CPU COMPUTE ! real time is now 
2 Likes
Any new updates for the software? and integration of gpu smoke and liquid simulators
?