Geometry Node: Strange overhead on access to curve normals #104045

New Issue

Iliya Katushenock · 2023-01-21T14:12:46+01:00

Iliya Katushenock commented

2023-01-21 14:12:46 +01:00

System Information
Operating system: Windows-10-10.0.19045-SP0 64 Bits
Graphics card: NVIDIA GeForce GTX 1060 6GB/PCIe/SSE2 NVIDIA Corporation 4.5.0 NVIDIA 527.56

Blender Version
Broken: version: 3.3.2 Release Candidate, branch: master, commit date: 2022-10-26 09:55, hash: 2a7b8e78e1
Worked: (newest version of Blender that worked as expected)

Short description of error
Strange overhead on access to curve normals:

This example demonstrates the big difference in the speed of the Curve Sampling node.
It appears depending on whether the normal attribute for the curve has been precalculated in any way.
As I can guess, these are lazy normal calculations.

Save normals as named attribute	No

Exact steps for others to reproduce the error

Create geometry node tree.
Some curve primitive node and duplicate node. Create a lot of curves.
Sample Curve node (Normal input). Sample Index (for sampled normal) on some other primitive. As index use single value.
Now, if you save Normal attrinute as named on curve before sampling, you will see x8 time difference in tree timings.

Example file:
Performance5.blend

**System Information** Operating system: Windows-10-10.0.19045-SP0 64 Bits Graphics card: NVIDIA GeForce GTX 1060 6GB/PCIe/SSE2 NVIDIA Corporation 4.5.0 NVIDIA 527.56 **Blender Version** Broken: version: 3.3.2 Release Candidate, branch: master, commit date: 2022-10-26 09:55, hash: `2a7b8e78e1` Worked: (newest version of Blender that worked as expected) **Short description of error** Strange overhead on access to curve normals: This example demonstrates the big difference in the speed of the Curve Sampling node. It appears depending on whether the normal attribute for the curve has been precalculated in any way. As I can guess, these are lazy normal calculations. | Save normals as named attribute | No | | -- | -- | | ![image.png](https://archive.blender.org/developer/F14181309/image.png) | ![image.png](https://archive.blender.org/developer/F14181307/image.png) | **Exact steps for others to reproduce the error** 1. Create geometry node tree. 2. Some curve primitive node and duplicate node. Create a lot of curves. 3. `Sample Curve` node (`Normal` input). `Sample Index` (for sampled normal) on some other primitive. As index use single value. Now, if you save `Normal` attrinute as named on curve before sampling, you will see x8 time difference in tree timings. Example file: [Performance5.blend](https://archive.blender.org/developer/F14181316/Performance5.blend)

Iliya Katushenock commented

2023-01-21 14:12:46 +01:00

Added subscriber: @mod_moder

Iliya Katushenock commented

2023-01-21 14:13:00 +01:00

Changed status from 'Needs Triage' to: 'Needs Developer To Reproduce'

Gerstmann Bradley commented

2023-01-21 15:24:51 +01:00

Added subscriber: @Bradley_G

Gerstmann Bradley commented

2023-01-21 15:24:51 +01:00

Performance5_original.blend
The original file from simplification reported to Iliya.

There are several problems in the node tree.

The Compare node is not and should never result in "True". This means the first switch should always use "Sample Curve-Normal", which makes the second/framed switch result irrelevant.
The second/framed Switch node has a huge performance difference. 600 ms (False) Vs. 2500 ms (True). Note that less nodes in "True" node tree rather takes much more time for the performance.
The first switch is the cause of this performance difference. If it uses "Normal", it will take less time while giving the wrong results; if it uses "Sample Curve-Normal", it will take more time. However according to 1., this node tree should never use "Normal".
The "False" node tree is functioning well in non-simplified file. It means practically the node tree is using "Sample Curve-Normal", but it's faster than the node tree with just "Sample Curve-Normal".

[Performance5_original.blend](https://archive.blender.org/developer/F14181414/Performance5_original.blend) The original file from simplification reported to Iliya. ![image.png](https://archive.blender.org/developer/F14181415/image.png) There are several problems in the node tree. 1. The Compare node is not and should never result in "True". This means the first switch should always use "Sample Curve-Normal", which makes the second/framed switch result irrelevant. 2. The second/framed Switch node has a huge performance difference. 600 ms (False) Vs. 2500 ms (True). Note that less nodes in "True" node tree rather takes much more time for the performance. The first switch is the cause of this performance difference. If it uses "Normal", it will take less time while giving the wrong results; if it uses "Sample Curve-Normal", it will take more time. However according to 1., this node tree should never use "Normal". 3. The "False" node tree is functioning well in non-simplified file. It means practically the node tree is using "Sample Curve-Normal", but it's faster than the node tree with just "Sample Curve-Normal".

Hans Goudey commented

2023-01-26 23:53:35 +01:00

Added subscriber: @HooglyBoogly

Hans Goudey commented

2023-01-26 23:53:35 +01:00

Changed status from 'Needs Developer To Reproduce' to: 'Archived'

Hans Goudey closed this issue

2023-01-26 23:53:35 +01:00

Hans Goudey commented

2023-01-26 23:53:35 +01:00

Thanks for the report, but I don't think this is a bug. When the normals are calculated, they are cached on the curve. If you calculate them earlier, the price will be payed earlier in the node tree. There are other changes we could make to improve performance in this file, but we don't generally handle that with reports, other than specific todos.

Iliya Katushenock commented

2023-01-27 00:31:03 +01:00

I drew attention to this due to the fact that the wound several times. Rather, it looks like problems with locks when accessing the normal (or calculating it).

Gerstmann Bradley commented

2023-01-27 00:37:44 +01:00

This isn't necessarily a bug report, but rather a bizarre phenomenon I find and I can't explain it at all. That's why I talked to others including Iliya before filing a report.
Regardless of reason, I don't think my questions/confusions are really being answered. and especially as an end result, the performance difference is dramatic.

One difference between my file and Iliya's file is that, I have a Switch node that should have never evaluated "Normal" in the node tree.
But somehow it's still improving the performance with a correct result, while "it end up taking more unnecessary nodes".

This isn't necessarily a bug report, but rather a bizarre phenomenon I find and I can't explain it at all. That's why I talked to others including Iliya before filing a report. Regardless of reason, I don't think my questions/confusions are really being answered. and especially as an end result, the performance difference is dramatic. One difference between my file and Iliya's file is that, I have a Switch node that should have never evaluated "Normal" in the node tree. But somehow it's still improving the performance with a correct result, while "it end up taking more unnecessary nodes".

Hans Goudey commented

2023-01-27 00:52:08 +01:00

Changed status from 'Archived' to: 'Confirmed'

Hans Goudey reopened this issue

2023-01-27 00:52:08 +01:00

Hans Goudey commented

2023-01-27 00:52:08 +01:00

Okay, I'll keep this open for now. I don't have time to look into this much now, but D17134 might help a bit. Or the related CoW patch. Not sure.

Okay, I'll keep this open for now. I don't have time to look into this much now, but [D17134](https://archive.blender.org/developer/D17134) might help a bit. Or the related CoW patch. Not sure.

blender-admin commented

2023-02-03 18:14:15 +01:00

This issue was referenced by 7f958217ad

This issue was referenced by 7f958217ada294a3aa19116aebb7bef832b9c648

Hans Goudey was assigned by Iliya Katushenock

2023-02-03 19:40:41 +01:00

Iliya Katushenock commented

2023-02-03 19:40:41 +01:00

Changed status from 'Confirmed' to: 'Resolved'

Iliya Katushenock closed this issue

2023-02-03 19:40:41 +01:00

Iliya Katushenock commented

2023-02-03 19:42:32 +01:00

| ![image.png](https://archive.blender.org/developer/F14237808/image.png) | ![image.png](https://archive.blender.org/developer/F14237810/image.png) | | -- | -- |

Sign in to join this conversation.

No Label

Download

What's New

Blender Studio

Manual

Developers Blog

Documentation

Benchmark

Blender Conference

Development Fund

One-time Donations

Geometry Node: Strange overhead on access to curve normals #104045