how to Cluster FLIP

[shenandyi]

Hi,

Actually, I want cluster the FLIP field, like vel or surface field like cluster in PYRO, some projects like the pic need optimized, in my FLIP project the particles' count is only 500,000, But the voxel count nearly 20,000,000, it's so inefficient!

Need some tips, THX!

Yi.

[edward]

Can you use VDB primitives instead?

[Shinjipierre]

Interesting problem, do you only emit that flip like a rocket trail, let's say, or does the liquid travel ? 

 

1. if it travels, you may want to try to simulate on a straight line and bend your sim later on 

2. if it's like a rocket trail, use wedges to cut the emission (for example : wedge 1 : frame 1 - 20. wedge 2 frame :21-40, wedge 3 frame :41-60

Just giving some ideas.

[shenandyi]

Can you use VDB primitives instead?

 

FLIP can use VDB instead of regular volume primitive?

I'm trying it now !

[shenandyi]

Interesting problem, do you only emit that flip like a rocket trail, let's say, or does the liquid travel ? 

 

1. if it travels, you may want to try to simulate on a straight line and bend your sim later on 

2. if it's like a rocket trail, use wedges to cut the emission (for example : wedge 1 : frame 1 - 20. wedge 2 frame :21-40, wedge 3 frame :41-60

Just giving some ideas.

 

Thank you, it's a clever solution for that.

My shot is just the two way WaterFall, emitted from the middle, so maybe VDB is my goal if it works, haha!

[Solitude]

It cannot use vdb IN the sim. You can convert the volumes to vdb before you save as bgeo, but the sim itself will only be done with regular volumes. 

 

With the shape you're getting, I don't think you'll have much luck with any sort of clustering setup. You still have a half-circle, which will cause a lot of empty space to be taken up.  I think you would have to sim maybe 3 or 4 "slices", but one by one. Each slice would have to overlap the next, and you would have to wait for slice 1 to finish before you start slice 2. The particles from slice 1 would then become the source for slice 2 where they overlap, and so on.  

 

Also, to minimize voxels, you'd maybe want to find a good way to set up non axis-aligned flip sims (the voxel grids are currently aligned to the grid only), and then go from there.

 

...might not be worth the hassle.  

Edited May 17, 2014 by Solitude

[danw]

I've actually found the pressure solve stage of Houdini FLIP fluids in 13.0 to be surprisingly tolerant of empty space.  Some part of the solve seems to discard empty regions early... I wonder if that is the point of the "preconditioner".

 

Anyway, in sims I've run, I've found that the solve can handle grid sizes up-to-and-over 100-megavoxel without becoming prohibitively slow or running out of memory, provided only a small fraction of that volume actually contains particles.

 

Ultimately, it's still more efficient to not have the excess grid there in the first place, so if there are any tricks you can use to axis-align simulations to minimize wasted domain volume, it's definitely worth using them.

 

On topic though, I'm afraid I can't offer any help on clustering, as I've yet to actually try it!

[freaq]

I still wonder why a fluid solve couldn't be done with VDB.

unless you are solving with air, you could atleast cull all the areas without fluid, (as undercurrents may be crucial)
I could see a fairly significant potential speedup.
or solve as a pointcloud with vectors instead of an actual voxel grid.(standard spaced pointcloud, not SPH)

[Solitude]

It can, they just haven't implemented it (yet?). Dreamworks' own smoke solver uses it internally from what I gather.

[shenandyi]

hope it implemented in houdini 13.1, it's a huge improvement and we won't consider to optimizing the FLIP scene.

Thank you guys!