Page MenuHome

Cloth simulation + motion blur = rendering slower 100x and driver crush.
Closed, ArchivedPublic

Description

Windows 7 64bit, i5 4460, GTX 970, 16GB DDR3

2.76 48f7dd6

Problem makes baked cloth simulation and motion blur. Problem exists, when baked simulation is shorter than sceene. For example: sceene has 190 frames, simulation is ending on 180 frame.
After end of simulation (in that case on 181 frame) rendering on CPU is slower 100 times and rendering on CUDA is crushed by drivers' error.

Download file

Go to 179 frame and render (you can use CPU or GPU)
Rendering should takes several seconds (4s on GTX 970)
Go to frame 180 or later and try to render - it will be much slower
If you are using CUDA, then driver will crush and send to http://nvidia.custhelp.com/app/answers/detail/a_id/3633
If you are using CPU, then it will be 100 times slower - you must close Blender

Reload file and select red rag
See, that cloth simulation is baked only to 180 frame and there is 190 frames
Delete red rag and render 179 or 180 or later - everything will work fine

Reload file and turn-off motion blur
Render 179 and 180 (without deleting red rag) - everything will work fine

BTW - sometimes on rendered image appears random distortion. Not everytime so I can't prepare file to show the problem.
If you want to see them in 3D_view change shading to 'Rendered' and then press F12. Sometimes "shadows" appers, someteimes not.
<a href='http://wstaw.org/w/3MOI/'><img src='http://wstaw.org/m/2016/02/04/randomshadow_01_jpg_300x300_q85.jpg'></a>
<a href='http://wstaw.org/w/3MOJ/'><img src='http://wstaw.org/m/2016/02/04/randomshadow_02_jpg_300x300_q85.jpg'></a>

Details

Type
Bug

Event Timeline

Sergey Sharybin (sergey) closed this task as Archived.
Sergey Sharybin (sergey) claimed this task.

This is a subframe access issue with simulated physics -- basically the subframe state of the cloth at the end of the simulated region jumps back to the original object state, making it really long trail of the object, which slows down ray shooting a lot.

Just bake the full range of frames, that's recommended for simulations anyway.

Improvements here are possible, but would consider it a TODO for a bigger picture reconsideration of how dependency graph works together with physics. Thanks for the report anyway.