Huge viewport performance difference between the instances from the particle system and the geometry nodes instances #92862

New Issue

kursad k · 2021-11-05T18:22:51+01:00

kursad k commented

2021-11-05 18:22:51 +01:00

System Information
Operating system: Windows 10
Graphics card: RTX 2070

Blender Version
Broken: e7e3431b29

There is a huge viewport interaction speed difference with these two types of instancing methods. The same setup from the geometry nodes setup drops the viewport interactivity speed and the playback FPS more than half.

The scene below is using the vertices as instancing locations so the number of instances should be the same between the methods.

See the video please

XZXe68juul.mp4

Exact steps for others to reproduce the error

Load the attach .blend
Select the plane
Enable those individual modifiers and compare the viewport interactivity and the playback speed

gn_05112021_1205_44.blend

**System Information** Operating system: Windows 10 Graphics card: RTX 2070 **Blender Version** Broken: e7e3431b2994 There is a huge viewport interaction speed difference with these two types of instancing methods. The same setup from the geometry nodes setup drops the viewport interactivity speed and the playback FPS more than half. The scene below is using the vertices as instancing locations so the number of instances should be the same between the methods. See the video please [XZXe68juul.mp4](https://archive.blender.org/developer/F11734272/XZXe68juul.mp4) **Exact steps for others to reproduce the error** Load the attach .blend Select the plane Enable those individual modifiers and compare the viewport interactivity and the playback speed [gn_05112021_1205_44.blend](https://archive.blender.org/developer/F11734288/gn_05112021_1205_44.blend)

kursad k commented

2021-11-05 18:22:51 +01:00

Added subscriber: @kursadk

kursad k changed title from ~~Huge viewport performance difference between instances from the particle system and from the geometry nodes instances~~ to Huge viewport performance difference between the instances from the particle system and from the geometry nodes instances

2021-11-05 18:26:41 +01:00

kursad k changed title from ~~Huge viewport performance difference between the instances from the particle system and from the geometry nodes instances~~ to Huge viewport performance difference between the instances from the particle system and the geometry nodes instances

2021-11-05 18:29:29 +01:00

Gerstmann Bradley commented

2021-11-05 20:05:42 +01:00

Added subscriber: @Bradley_G

Gerstmann Bradley commented

2021-11-05 20:05:42 +01:00

I've run a test on my side as well. Due to my laptop limitation I have to change some settings.
But I've made sure that counts of fps really matches in both GN and hair system. This is confirmed with viewport statistics and spreadsheet.
Here's my conclusion:
"As Instance" option in "object info" node is making a difference. This sounds to be weird, because instances aren't realized at the end.
[BUG TEST].mp4

I've run a test on my side as well. Due to my laptop limitation I have to change some settings. But I've made sure that counts of fps really matches in both GN and hair system. This is confirmed with viewport statistics and spreadsheet. Here's my conclusion: "As Instance" option in "object info" node is making a difference. This sounds to be weird, because instances aren't realized at the end. [[BUG TEST].mp4](https://archive.blender.org/developer/F11735781/_BUG_TEST_.mp4)

Gerstmann Bradley commented

2021-11-05 21:27:49 +01:00

I've realized that in the file the cone has 17k vertices.
that's why I can only instance about 2000 objects for blender run smoothly.
If I am using a default cone, I can instance as many as 60k instances and even more (it's hard to tell the vertices count since statistic won't show it)
In such a case, whether "As Instance" is checked or not, GN's performance is better than hair (12 fps Vs. 10 fps)

I've realized that in the file the cone has 17k vertices. that's why I can only instance about 2000 objects for blender run smoothly. If I am using a default cone, I can instance as many as 60k instances and even more (it's hard to tell the vertices count since statistic won't show it) In such a case, whether "As Instance" is checked or not, GN's performance is better than hair (12 fps Vs. 10 fps)

kursad k commented

2021-11-05 21:29:56 +01:00

I made the cone as an high poly object to demonstrate the issue, in regular cases this issue is not as visible.

Gerstmann Bradley commented

2021-11-05 21:40:56 +01:00

But in essence this issue is not relevant to difference between particle system and Geometry Nodes.
This issue stems in "As Instance", in which you just turn it on or off you can see the difference.
Considering what "Object Info" node is doing. I am not really sure if this should be called a bug or not. will need to wait dev to define it.

But in essence this issue is not relevant to difference between particle system and Geometry Nodes. This issue stems in "As Instance", in which you just turn it on or off you can see the difference. Considering what "Object Info" node is doing. I am not really sure if this should be called a bug or not. will need to wait dev to define it.

Gerstmann Bradley commented

2021-11-06 00:14:19 +01:00

I've simplify the file and update a new demonstration now.
AS INSTANCE.blend
AS INSTANCE.mp4

I've simplify the file and update a new demonstration now. [AS INSTANCE.blend](https://archive.blender.org/developer/F11739088/AS_INSTANCE.blend) [AS INSTANCE.mp4](https://archive.blender.org/developer/F11739130/AS_INSTANCE.mp4)

michael campbell commented

2021-11-06 01:51:39 +01:00

Added subscriber: @3di

michael campbell commented

2021-11-06 01:51:39 +01:00

maybe these test files will be helpful. They indicate the issue isn't just with loading the large object each frame, as the performance is severely hit even when navigating the viewport when the timeline isn't playing:

instance performance legacy vs geo.zip

maybe these test files will be helpful. They indicate the issue isn't just with loading the large object each frame, as the performance is severely hit even when navigating the viewport when the timeline isn't playing: [instance performance legacy vs geo.zip](https://archive.blender.org/developer/F11740428/instance_performance_legacy_vs_geo.zip)

kursad k commented

2021-11-24 23:05:35 +01:00

@Bradley_G

This issue becomes a horrible mess when using the Collection Info node. It seems like that node is not outputting instances, so the speed difference between the particle system and the geometry nodes is very evident when one using collections as the instance sources.

EDIT:

The issue below seems to be a separate issue related to instancing objects with existing particle systems

Also even with enabling instancing in the object node is not a real match for particle systems when one is using very dense objects as the instancing source. Try using something like a tree object as instance type in geometry node. The particle system is at least 15-25 times faster in the viewport in real life production setups like setting up a forest or a jungle.

I have a production scene setup here (see the image). I can easily navigate around with the particle system setup. On the other hand it takes more than half a minute to rotate the viewport with the geometry nodes one.

The "as instance" option in the object node definitely makes a big difference but with low to mid-range complexity scenes.

@Bradley_G This issue becomes a horrible mess when using the Collection Info node. It seems like that node is not outputting instances, so the speed difference between the particle system and the geometry nodes is very evident when one using collections as the instance sources. ## **EDIT**: **The issue below seems to be a separate issue related to instancing objects with existing particle systems** Also even with enabling instancing in the object node is not a real match for particle systems when one is using very dense objects as the instancing source. Try using something like a tree object as instance type in geometry node. The particle system is at least 15-25 times faster in the viewport in real life production setups like setting up a forest or a jungle. I have a production scene setup here (see the image). I can easily navigate around with the particle system setup. On the other hand it takes more than half a minute to rotate the viewport with the geometry nodes one. The "as instance" option in the object node definitely makes a big difference but with low to mid-range complexity scenes. ![blender_QxtnQkV0Ag.jpg](https://archive.blender.org/developer/F12043259/blender_QxtnQkV0Ag.jpg)

Philipp Oeser commented

2022-02-16 17:12:09 +01:00

Added subscriber: @lichtwerk

Philipp Oeser commented

2022-02-16 17:12:09 +01:00

Changed status from 'Needs Triage' to: 'Needs User Info'

Philipp Oeser commented

2022-02-16 17:12:09 +01:00

Looking at older reports (sorry this has been lying around for quite some time).

testing instance performance legacy vs geo.zip in 3.2alpha
- cannot see a huge difference between legacy dupli vert instancing and geo nodes instancing (as instancing checked)
- geo nodes instancing (as instancing un-checked): yep that one is really slow, not sure this is expected
  https://docs.blender.org/manual/en/dev/modeling/geometry_nodes/input/object_info.html

Output the entire object as single instance instead of realized geometry. This allows instancing non-geometry object types, because the output will contain an instance of the object.

Is not really entirely clear to me either, since we are instancing a geometry object type in the example, this sounds like it should not make a difference?

In #92862#1259601, @kursadk wrote:
This issue becomes a horrible mess when using the Collection Info node. It seems like that node is not outputting instances, so the speed difference between the particle system and the geometry nodes is very evident when one using collections as the instance sources.

This I cannot reproduce, at least in 3.2 putting the single Suzanne in its own collection will yield similar performance.
Do you mean this is an issue with multiple objects in a collection?

Also even with enabling instancing in the object node is not a real match for particle systems when one is using very dense objects as the instancing source. Try using something like a tree object as instance type in geometry node. The particle system is at least 15-25 times faster in the viewport in real life production setups like setting up a forest or a jungle.

I have a production scene setup here (see the image). I can easily navigate around with the particle system setup. On the other hand it takes more than half a minute to rotate the viewport with the geometry nodes one.

The "as instance" option in the object node definitely makes a big difference but with low to mid-range complexity scenes.

Can you share such an example file?

Looking at older reports (sorry this has been lying around for quite some time). - testing `instance performance legacy vs geo.zip` in 3.2alpha - cannot see a huge difference between `legacy dupli vert instancing` and `geo nodes instancing (as instancing checked)` - `geo nodes instancing (as instancing un-checked)`: yep that one is really slow, not sure this is expected https://docs.blender.org/manual/en/dev/modeling/geometry_nodes/input/object_info.html > Output the entire object as single instance instead of realized geometry. This allows instancing non-geometry object types, because the output will contain an instance of the object. Is not really entirely clear to me either, since we are instancing a geometry object type in the example, this sounds like it should not make a difference? > In #92862#1259601, @kursadk wrote: > This issue becomes a horrible mess when using the Collection Info node. It seems like that node is not outputting instances, so the speed difference between the particle system and the geometry nodes is very evident when one using collections as the instance sources. This I cannot reproduce, at least in 3.2 putting the single Suzanne in its own collection will yield similar performance. Do you mean this is an issue with multiple objects in a collection? > Also even with enabling instancing in the object node is not a real match for particle systems when one is using very dense objects as the instancing source. Try using something like a tree object as instance type in geometry node. The particle system is at least 15-25 times faster in the viewport in real life production setups like setting up a forest or a jungle. > > I have a production scene setup here (see the image). I can easily navigate around with the particle system setup. On the other hand it takes more than half a minute to rotate the viewport with the geometry nodes one. > > The "as instance" option in the object node definitely makes a big difference but with low to mid-range complexity scenes. Can you share such an example file?

kursad k commented

2022-02-16 17:48:31 +01:00

@lichtwerk

The performance issue comes to surface when one instances dense objects like complicated trees (dense branches with leaves etc) over a large terrain. You won't be able to hit this issue with primitives and such.

This I cannot reproduce, at least in 3.2 putting the single Suzanne in its own collection will yield similar performance.
Do you mean this is an issue with multiple objects in a collection?

I am not sure what the collection node outputs. It seemed like it was outputting regular objects not instances when I provided a collection with mixed bunch. What is a good way to test if a node is outputting instances or not? This is easy with the regular object node since it provides an option to output an instance instead of a regular object.

@lichtwerk The performance issue comes to surface when one instances dense objects like complicated trees (dense branches with leaves etc) over a large terrain. You won't be able to hit this issue with primitives and such. >This I cannot reproduce, at least in 3.2 putting the single Suzanne in its own collection will yield similar performance. >Do you mean this is an issue with multiple objects in a collection? I am not sure what the collection node outputs. It seemed like it was outputting regular objects not instances when I provided a collection with mixed bunch. What is a good way to test if a node is outputting instances or not? This is easy with the regular object node since it provides an option to output an instance instead of a regular object.

Philipp Oeser commented

2022-03-24 14:19:07 +01:00

Changed status from 'Needs User Info' to: 'Needs Triage'

Philipp Oeser commented

2022-03-30 14:35:59 +02:00

Changed status from 'Needs Triage' to: 'Needs User Info'

Philipp Oeser commented

2022-03-30 14:35:59 +02:00

In #92862#1308812, @kursadk wrote:
@lichtwerk

The performance issue comes to surface when one instances dense objects like complicated trees (dense branches with leaves etc) over a large terrain. You won't be able to hit this issue with primitives and such.

Not sure at which point this kicks in, it would be really handy to have an example .blend file to see a clear difference of legacy/duplivert instancing vs. goenodes instancing.

I am not sure what the collection node outputs. It seemed like it was outputting regular objects not instances when I provided a collection with mixed bunch. What is a good way to test if a node is outputting instances or not? This is easy with the regular object node since it provides an option to output an instance instead of a regular object.

Think it outputs instances (single or multiple -- depending on the Separate Children option)
You can check in the spreadsheet:
{F12957014 size=full}

> In #92862#1308812, @kursadk wrote: > @lichtwerk > > > The performance issue comes to surface when one instances dense objects like complicated trees (dense branches with leaves etc) over a large terrain. You won't be able to hit this issue with primitives and such. Not sure at which point this kicks in, it would be really handy to have an example .blend file to see a clear difference of legacy/duplivert instancing vs. goenodes instancing. > I am not sure what the collection node outputs. It seemed like it was outputting regular objects not instances when I provided a collection with mixed bunch. What is a good way to test if a node is outputting instances or not? This is easy with the regular object node since it provides an option to output an instance instead of a regular object. Think it outputs instances (single or multiple -- depending on the `Separate Children` option) You can check in the spreadsheet: {[F12957014](https://archive.blender.org/developer/F12957014/image.png) size=full}

Petter Lundh commented

2022-04-05 13:15:49 +02:00

Added subscriber: @PetterLundh