Hiya

Just want to clear up something I'm unsure about.

Does the new Golbal Pressure Solve in distributed sims apply to Flip only? If the new functionality is on the Gas Project nodes then hopefully this would work on sliced PyroFX too but the docs don't specify if they do or not.

Any clarification would be appreciated.

Cheers
Dan

The global pressure solve is for the Gas Project Non Divergent Variational node.

Thus it works for Flip and the Particle-level-set fluids.

Pyro uses Gas Project Non Divergent Multigrid, however, which does not have a distributed pressure solve. It is possible to rebuild it with the variational node; but since multigrid is 10x faster, the gains of distribution wouldn't be sufficient.

With H15, the resize dynamic node now supports distribution, however. So sliced PyroFX can now distribute and resize, just not with the global pressure solve.

Thanks for the info Jeff, much appreciated.

Does the variational node dynamically copy across as much information as it needs for the pressure solve or is there a distance limit on the number of voxels accross the boundary?

Any plans for multigrid to support collisions too?

on a side note, I compiled openvdb 3.1 for houdini to try the PCG/possion solver. It ended up crashing which I'm trying to figure out over on the openvdb forum. But since we can now advect arbitrary volumes on a vdb grid and VOPs can operate on them, are there any plans to build a solver from vdb nodes? I guess 3.1 will be coming at some point officially for houdini (maybe in an upcoming build pretty please?).

I'm also wondering how distribution would work with openvdb. The data structure would have to support that itself im guessing or have multiple grids and a complicated higher level manager.

My own thoughts are that having a data structure made up of tiles that are indexed with a concurrent hash table means that you can copy tiles between machines to make pressure solves easier.

Anyway the future is exciting whatever happens!

Daniel Elliott
Thanks for the info Jeff, much appreciated.

Does the variational node dynamically copy across as much information as it needs for the pressure solve or is there a distance limit on the number of voxels accross the boundary?

The distance limit is for the advection step. You also want at least a voxel bandwidth of particles to ensure the surface field matches on both sides. But the pressure projection itself only needs the immediate faces, which it transfers as it needs it.

My own thoughts are that having a data structure made up of tiles that are indexed with a concurrent hash table means that you can copy tiles between machines to make pressure solves easier.

Dneg had a cool presentation on how they did this. A big issue is making sure your tiles stay contiguous, or you soon have way too much information to move every iteration. We handled this by making the slicing explicit, having multiple grids, and the higher level manager.