Regression: Cycles: unstable triangulation of mesh in viewport #106469

New Issue

Closed

opened 2023-04-02 22:55:43 +02:00 by Max Mustermann · 9 comments

Max Mustermann commented

2023-04-02 22:55:43 +02:00

System Information
Operating system: Fedora Linux 37 (Wayland)
Graphics card: AMD

Blender Version
Broken: 3.5.0 (flatpak)
Worked: 3.4.1 (fedora rpm)

Exposed/Caused by 511ac66dab

Short description of error

There are two observable issues:

Mesh edit mode (quad) triangulation (edit mode overlay) is flickering when previewing with Cycles.
Cycles triangulation itself doesn't seem to flicker here, but the triangulation result is the inverse of what is in object mode.

Steps to reproduce

Enable cycles.
Enable shaded viewport.
Enter edit mode.
Rotate top face of default cube by -72° on Z-Axis.
Extrude top face.
Move extruded top face.
The shape of the default cube now flickers.

Note:

When Cycles preview is showing the wrong triangulation, switching back to material/solid shading would retain the wrong triangulation, move the mesh will refresh the state and the triangulation looks stable again.

**System Information** Operating system: Fedora Linux 37 (Wayland) Graphics card: AMD **Blender Version** Broken: 3.5.0 (flatpak) Worked: 3.4.1 (fedora rpm) Exposed/Caused by 511ac66dab **Short description of error** There are two observable issues: - Mesh edit mode (quad) triangulation (edit mode overlay) is flickering when previewing with Cycles. - Cycles triangulation itself doesn't seem to flicker here, but the triangulation result is the inverse of what is in object mode. **Steps to reproduce** - Enable cycles. - Enable shaded viewport. - Enter edit mode. - Rotate top face of default cube by -72° on Z-Axis. - Extrude top face. - Move extruded top face. - The shape of the default cube now flickers. Note: - When Cycles preview is showing the wrong triangulation, switching back to material/solid shading would retain the wrong triangulation, move the mesh will refresh the state and the triangulation looks stable again.

trangulation.blend

828 KiB

Max Mustermann added the

labels 2023-04-02 22:55:44 +02:00

Iliya Katushenock added the

Render & Cycles

labels 2023-04-02 23:02:08 +02:00

Max Mustermann commented

2023-04-02 23:23:43 +02:00

Author

It is easier to see against a dark background.

It is easier to see against a dark background.

YimingWu commented

2023-04-03 11:12:16 +02:00

Member

Can confirm the issue. Might relate to threading and different triangulation algorithm used in bmesh and mesh.

I'll attach a file on the first post to show the problem.

Can confirm the issue. Might relate to threading and different triangulation algorithm used in bmesh and mesh. I'll attach a file on the first post to show the problem.

YimingWu added

Render & Cycles

and removed

labels 2023-04-03 11:20:04 +02:00

Philipp Oeser commented

2023-04-03 11:27:06 +02:00

Member

Probably also caused by 16fbadde36

We had #106134 already, but I can confirm this is still happening (even after 5efa39096c)

@HooglyBoogly : mind checking?

Probably also caused by 16fbadde363c8074ec725fa1e1079add096bd741 We had #106134 already, but I can confirm this is still happening (even after 5efa39096c8e242c9bf9ed7b7a25124692e2b9d0) @HooglyBoogly : mind checking?

YimingWu commented

2023-04-03 11:29:00 +02:00

Member

I suspect #106430 is also affected by that commit...

I suspect #106430 is also affected by that commit...

Philipp Oeser changed title from ~~Cycles: unstable triangulation of mesh in viewport~~ to Regression: Cycles: unstable triangulation of mesh in viewport

2023-04-03 11:30:08 +02:00

Philipp Oeser commented

2023-04-03 11:34:46 +02:00

Member

I suspect #106430 is also affected by that commit...

Think this is more about 6c774feba2 (but would have to check... lets continue discussion in #106430)

> I suspect #106430 is also affected by that commit... Think this is more about 6c774feba2c9a1eb5834646f597a0f2c63177914 (but would have to check... lets continue discussion in #106430)

Philipp Oeser commented

2023-04-03 11:40:03 +02:00

Member

Not caused by 16fbadde36, will bisect the real culprit

Not caused by 16fbadde36, will bisect the real culprit

Philipp Oeser commented

2023-04-03 12:23:13 +02:00

Member

Exposed/Caused by 511ac66dab

Havent checked if cycles triangulation is actually the same as viewport triangulation (or if this is another bug).

@HooglyBoogly : mind checking?

Exposed/Caused by 511ac66dabf170de2e0e445afbee261ddb944cf6 Havent checked if cycles triangulation is actually the same as viewport triangulation (or if this is another bug). @HooglyBoogly : mind checking?

Philipp Oeser added

and removed

Render & Cycles

labels 2023-04-03 12:24:07 +02:00

Hans Goudey self-assigned this 2023-04-11 21:35:48 +02:00

Hans Goudey commented

2023-04-14 22:42:21 +02:00

Member

It looks like the problem is in mesh_calc_tessellation_for_face_impl. This gives different results depending on whether face normals happen to be calculated already or not. In other words, is_quad_flip_v3_first_third_fast_with_normal is not equivalent to is_quad_flip_v3_first_third_fast. I'm surprised this was only exposed now actually, it looks like this issue has been there for a while, since 2d60c496a2. It's just visible because Cycles copies meshes that are in edit mode, which doesn't carry over any cached triangulation.

I have some ideas I'd like to investigate next:

Check the performance of just calculating the quad normal in that function.
Copy logic from BM_face_triangulate MOD_TRIANGULATE_QUAD_BEAUTY. Check to see if the performance is much worse
Figure out the history of f088c6b9f6. Where does this flipping mechanism come from, compared to something simpler like equalizing the area of the two triangles (mentioned here)
Also, random unrelated thought: it might be good to add a runtime tag to some primitive meshes like grids to avoid this check, it's totally unnecessary when they haven't been deformed.
Figure out why the "fast" version of the function is used for Mesh evaluation when the "non-fast" version of the function is used for BMesh.

@ideasman42 From the git history it looks like you've worked on mesh triangulation in the past, do you have any thoughts about this, maybe about which of those ideas seem worth investigating?

I attached a simpler reproduction with a single face.

no poly normals	poly normals

It looks like the problem is in `mesh_calc_tessellation_for_face_impl`. This gives different results depending on whether face normals happen to be calculated already or not. In other words, `is_quad_flip_v3_first_third_fast_with_normal` is not equivalent to `is_quad_flip_v3_first_third_fast`. I'm surprised this was only exposed now actually, it looks like this issue has been there for a while, since 2d60c496a200288f100c11845ccb196f04e45ba3. It's just visible because Cycles copies meshes that are in edit mode, which doesn't carry over any cached triangulation. I have some ideas I'd like to investigate next: - Check the performance of just calculating the quad normal in that function. - Copy logic from `BM_face_triangulate` `MOD_TRIANGULATE_QUAD_BEAUTY`. Check to see if the performance is much worse - Figure out the history of f088c6b9f6d69e7f1af15a059f122589ee6a39ab. Where does this flipping mechanism come from, compared to something simpler like equalizing the area of the two triangles (mentioned [here](https://computergraphics.stackexchange.com/questions/10180/how-to-decide-which-way-to-triangulate-a-quad)) - Also, random unrelated thought: it might be good to add a runtime tag to some primitive meshes like grids to avoid this check, it's totally unnecessary when they haven't been deformed. - Figure out why the "fast" version of the function is used for `Mesh` evaluation when the "non-fast" version of the function is used for `BMesh`. @ideasman42 From the git history it looks like you've worked on mesh triangulation in the past, do you have any thoughts about this, maybe about which of those ideas seem worth investigating? I attached a simpler reproduction with a single face. | no poly normals | poly normals | | --------------- | ------------ | | ![image](/attachments/7ce6a0c9-8f05-4e64-90ca-48461bb7dd09) | ![image](/attachments/0329518b-4dc7-4567-9955-5a882052c7f4) |

Bad Triangulation.blend

100 KiB

147 KiB

138 KiB

Hans Goudey added

and removed

labels 2023-04-14 22:43:49 +02:00

Campbell Barton referenced this issue from a commit

2023-04-21 07:04:16 +02:00

Fix #106469: unstable tessellation with quad-flipping detection

Campbell Barton closed this issue

2023-04-21 07:04:17 +02:00

Blender Bot added

and removed

labels 2023-04-21 07:04:18 +02:00

Hans Goudey commented

2023-04-21 14:22:49 +02:00

Member

At least in one case (a large UV sphere primitive), this fix seems to have made triangulation a bit faster too, that's nice!

BEFORE: Timer 'looptris_calc_with_normals': (Average: 7.5 ms, Min: 7.3 ms, Last: 7.6 ms)
AFTER: Timer 'looptris_calc_with_normals': (Average: 7.1 ms, Min: 6.9 ms, Last: 7.1 ms)

At least in one case (a large UV sphere primitive), this fix seems to have made triangulation a bit faster too, that's nice! ``` BEFORE: Timer 'looptris_calc_with_normals': (Average: 7.5 ms, Min: 7.3 ms, Last: 7.6 ms) AFTER: Timer 'looptris_calc_with_normals': (Average: 7.1 ms, Min: 6.9 ms, Last: 7.1 ms) ```

Sign in to join this conversation.

No Label

Animation & Rigging

Asset Browser Project Overview

Automated Testing

Blender Asset Bundle

Dependency Graph

Development Management

EEVEE & Viewport

Images & Movies

Motion Tracking

Nodes & Physics

Pipeline, Assets & IO

Platforms, Builds & Tests

Render & Cycles

Render Pipeline

Sculpt, Paint & Texture

Video Sequencer

Virtual Reality

Blender 2.8 Project

Milestone 1: Basic, Local Asset Browser

Good First Issue

Animation & Rigging

Development Management

EEVEE & Viewport

Nodes & Physics

Pipeline, Assets & IO

Platforms, Builds & Tests

Render & Cycles

Sculpt, Paint & Texture

Needs Info from Developers

Needs Information from User

No Milestone

No project

No Assignees

4 Participants

Notifications

Due Date

The due date is invalid or out of range. Please use the format 'yyyy-mm-dd'.

No due date set.

Dependencies

No dependencies set.

Reference: blender/blender#106469

No description provided.