Particle with fluid physics unstable #39180

New Issue

Simeon Conzendorf · 2014-03-14T10:25:04+01:00

Simeon Conzendorf commented

2014-03-14 10:25:04 +01:00

System Information
Operating system and graphics card

Windows 7 64bit
Geforce GTX 660

Blender Version
Broken: 2.70 2525f9c = official RC2
Worked: 2.69 r60995 (also the example below doesn't crash on load)

Short description of error
The particle system with fluid physics seem to be very unstable, especially if you want to use interaction between 2 of them.

Exact steps for others to reproduce the error

default Blender scene
give the standard cube a particle system
select "fluid" physic panel
duplicate cube and position it beside the first
make the particle system single user
alt + a (play)

CRASH or sometimes one of the systems get's a very strange force.

it get's even worser if you set up interaction: select cube1, in particle/physic/fluid set interaction to cube2 and vice versa. I hardly managed it to set up a particle fluid system with interaction.

If I save the blend file before pressing play and try to reload it, blender crashes.

sim10.blend

**System Information** Operating system and graphics card - Windows 7 64bit - Geforce GTX 660 **Blender Version** Broken: 2.70 2525f9c = official RC2 Worked: 2.69 r60995 (also the example below doesn't crash on load) **Short description of error** The particle system with fluid physics seem to be very unstable, especially if you want to use interaction between 2 of them. **Exact steps for others to reproduce the error** 01. default Blender scene 02. give the standard cube a particle system 03. select "fluid" physic panel 04. duplicate cube and position it beside the first 05. make the particle system single user 06. alt + a (play) CRASH or sometimes one of the systems get's a very strange force. 07. it get's even worser if you set up interaction: select cube1, in particle/physic/fluid set interaction to cube2 and vice versa. I hardly managed it to set up a particle fluid system with interaction. If I save the blend file before pressing play and try to reload it, blender crashes. [sim10.blend](https://archive.blender.org/developer/F81089/sim10.blend)

Simeon Conzendorf commented

2014-03-14 10:25:04 +01:00

Changed status to: 'Open'

Lukas Tönne was assigned by Simeon Conzendorf

2014-03-14 10:25:04 +01:00

Simeon Conzendorf commented

2014-03-14 10:25:04 +01:00

Added subscriber: @SimeonConzendorf

Sebastian Koenig commented

2014-03-14 10:28:01 +01:00

Added subscriber: @sebastian_k

Sebastian Koenig commented

2014-03-14 10:28:01 +01:00

Crashes also on OSX.

blender-admin commented

2014-03-14 11:03:04 +01:00

This issue was referenced by blender/blender-addons-contrib@ade6646755

This issue was referenced by blender/blender-addons-contrib@ade6646755224ce9178d280738727c1a968b28b6

blender-admin commented

2014-03-14 11:03:04 +01:00

This issue was referenced by ade6646755

This issue was referenced by ade6646755224ce9178d280738727c1a968b28b6

Lukas Tönne commented

2014-03-14 11:11:18 +01:00

Changed status from 'Open' to: 'Resolved'

Lukas Tönne closed this issue

2014-03-14 11:11:18 +01:00

Lukas Tönne commented

2014-03-14 11:11:18 +01:00

Closed by commit ade6646755.

Closed by commit ade6646755.

Simeon Conzendorf commented

2014-03-15 18:53:54 +01:00

Changed status from 'Resolved' to: 'Open'

Simeon Conzendorf reopened this issue

2014-03-15 18:53:54 +01:00

Simeon Conzendorf commented

2014-03-15 18:53:54 +01:00

Hello lukas,

thank you for fixing it so fast, but I think there could be more bugs related to my topic, i hope it's ok, that I reopen it.

I downloaded latest built from http://builder.blender.org/download/ - Hash: 2aaf243 from Sat Mar 15 02:26:16 2014
(I'm not able to compile blender)

Exact steps for others to reproduce the error (same as above)

default Blender scene
give the standard cube a particle system
select "fluid" physic panel
duplicate cube and position it beside the first
make the particle system single user
alt + a (play)

WORKS

select cube1, in particle/physic/fluid set interaction to cube2 and vice versa
alt + a (play)

CRASH

In 2.69 it works.

sim12.blend

Hello lukas, thank you for fixing it so fast, but I think there could be more bugs related to my topic, i hope it's ok, that I reopen it. I downloaded latest built from http://builder.blender.org/download/ - Hash: **2aaf243** from Sat Mar 15 02:26:16 2014 (I'm not able to compile blender) **Exact steps for others to reproduce the error (same as above)** 01. default Blender scene 02. give the standard cube a particle system 03. select "fluid" physic panel 04. duplicate cube and position it beside the first 05. make the particle system single user 06. alt + a (play) WORKS 07. select cube1, in particle/physic/fluid set interaction to cube2 and vice versa 08. alt + a (play) CRASH In 2.69 it works. [sim12.blend](https://archive.blender.org/developer/F81291/sim12.blend)

Lukas Tönne commented

2014-03-16 16:48:37 +01:00

I'm afraid there is no proper solution this time ...

The fluid interaction code totally relies on other object's data being available, it contains a function call to write the particle bvh tree data of the target object:
https://developer.blender.org/diffusion/B/browse/master/source/blender/blenkernel/intern/particle_system.c;caf8684b5066720d4f4124092e23857dd462cb8b$4302

This is extremely bad, because it totally breaks with threaded depsgraph evaluation. Until the depsgraph can fully handle such object data <-> object data interaction (and the particle system implements it correctly) there is no way to make it work really. I would rather disable this feature entirely until we can do it right.

@SimeonConzendorf: Only thing you could try is to run Blender with 1 thread only (blender -t 1). You will still get errors due to unsynchronized particle systems, but with a bit of luck this won't be noticeable.

I'm afraid there is no proper solution this time ... The fluid interaction code totally relies on other object's data being available, it contains a function call to **write** the particle bvh tree data of the **target object**: https://developer.blender.org/diffusion/B/browse/master/source/blender/blenkernel/intern/particle_system.c;caf8684b5066720d4f4124092e23857dd462cb8b$4302 This is extremely bad, because it totally breaks with threaded depsgraph evaluation. Until the depsgraph can fully handle such object data <-> object data interaction (and the particle system implements it correctly) there is no way to make it work really. I would rather disable this feature entirely until we can do it right. @SimeonConzendorf: Only thing you could try is to run Blender with 1 thread only (blender -t 1). You will still get errors due to unsynchronized particle systems, but with a bit of luck this won't be noticeable.

ronan ducluzeau commented

2014-03-16 17:41:05 +01:00

Added subscriber: @zeauro

ronan ducluzeau commented

2014-03-16 17:41:05 +01:00

It seems to work fine with only one thread. No crash.
@ Lukas : is it possible to make the feature a single threaded task until you find a solution ?

It seems to work fine with only one thread. No crash. @ Lukas : is it possible to make the feature a single threaded task until you find a solution ?

Simeon Conzendorf commented

2014-03-17 09:52:07 +01:00

Runing Blender with 1 thread only works for me too... this is a good workaround until there's a solution. Thank you so far.

Sergey Sharybin commented

2014-03-18 09:39:21 +01:00

Added subscribers: @LukasTonne, @Sergey

Sergey Sharybin commented

2014-03-18 09:39:21 +01:00

@LukasTonne, we can add a mutex around bvh write?

Lukas Tönne commented

2014-03-18 10:06:34 +01:00

Yes, that's what i'd do here. Is a crappy workound for bad code but would prevent crash ...

Can store a mutex in every psys for this purpose, doesn't have to be global.

Yes, that's what i'd do here. Is a crappy workound for bad code but would prevent crash ... Can store a mutex in every psys for this purpose, doesn't have to be global.

Sergey Sharybin commented

2014-03-18 10:38:14 +01:00

You could have global rw mutex. Would be better than polluting storage with workarounds.

Lukas Tönne commented

2014-03-18 10:49:21 +01:00

Particle storage is a mess anyway, at least a local mutex would avoid adding static vars and/or core init/free calls and keep this stuff inside psys. Also it would still allow some measure of threading of unrelated psys objects (although that probably won't matter much).