Hair system, pick random from dupligroup leads to a huge viewport performance drop #50587

New Issue

Yegor · 2017-02-03T16:13:36+01:00

Yegor commented

2017-02-03 16:13:36 +01:00

System Information
CentOS, Nvidia GTX Titan

Blender Version
Broken: 2.78a official

Short description of error
Pick random from dupligroup leads to a huge viewport performance drop

Exact steps for others to reproduce the error
Create two or more objects with a subsurf mod on them
Add them to a group
Create a hair host object (plane)
Make a hair system
Use that group as a dupligroup
Check "pick random" -> huge performance drop

File attached
pick_random_performance_drop.blend

**System Information** CentOS, Nvidia GTX Titan **Blender Version** Broken: 2.78a official **Short description of error** Pick random from dupligroup leads to a huge viewport performance drop **Exact steps for others to reproduce the error** Create two or more objects with a subsurf mod on them Add them to a group Create a hair host object (plane) Make a hair system Use that group as a dupligroup Check "pick random" -> huge performance drop File attached [pick_random_performance_drop.blend](https://archive.blender.org/developer/F445077/pick_random_performance_drop.blend)

Yegor commented

2017-02-03 16:13:36 +01:00

Changed status to: 'Open'

Yegor commented

2017-02-03 16:13:36 +01:00

Added subscriber: @YegorSmirnov

Bastien Montagne commented

2017-02-04 10:50:27 +01:00

Added subscribers: @MikeErwin, @Sergey, @dfelinto, @mont29

Bastien Montagne commented

2017-02-04 10:50:27 +01:00

Ok, sooooo… This is OpenGL drawing issue.

in draw_dupli_objects_color(), line 2145 and following, code decides whether it generates a displist or not for the current dupliobject, among other things based on the fact that next dupliobject to be drawn is the same as current one or not.

In our case, this is catastrophic, because random picking among a group of few objects will often gives sequences of two or three times the same object. So the displist will be generated, but not used that much - and clearly, generating a displist is not a small process! Without random picking, since each object is picked on its turn, you never have two in a row, so displist never gets generated.

(the part of dupli object drawing involved goes from 43us to 480us average per dupli, here on a debug build.)

@Sergey, @dfelinto, @MikeErwin, not sure we consider this a bug worth fixing right now? Drawing code is being fully rewritten for 2.8 anyway…

Ok, sooooo… This is OpenGL drawing issue. in `draw_dupli_objects_color()`, line 2145 and following, code decides whether it generates a displist or not for the current dupliobject, among other things based on the fact that next dupliobject to be drawn is the same as current one or not. In our case, this is catastrophic, because random picking among a group of few objects will often gives sequences of two or three times the same object. So the displist will be generated, but not used that much - and clearly, generating a displist is not a small process! Without random picking, since each object is picked on its turn, you never have two in a row, so displist never gets generated. (the part of dupli object drawing involved goes from 43us to 480us average per dupli, here on a debug build.) @Sergey, @dfelinto, @MikeErwin, not sure we consider this a bug worth fixing right now? Drawing code is being fully rewritten for 2.8 anyway…

Clément Foucault was assigned by Dalai Felinto

2017-02-06 10:54:54 +01:00

Dalai Felinto commented

2017-02-06 10:54:54 +01:00

Assigning to Clement since he is the one that handled the new version of the "code". So at least he should be aware of this problem for the new viewport.