Regression: Precision artifacts on refractive materials with Cycles MetalRT #125307

New Issue

Xhaos · 2024-07-23T13:30:32+02:00

Xhaos commented

2024-07-23 13:30:32 +02:00

System Information
Operating system: macOS Sonoma 14.5
Graphics card: M3 Max

Blender Version
Broken: (Blender 4.2 LTS)
Worked: (Blender 4.1.0, Blender 4.1.1 LTS)

Spent the whole day trying to reproduce the issue, and as of writing this I discovered that I think it has to do with how far away the object is from the world origin.

Context: Was working on a project till I noticed something off about an asset I simply appended which was this transparent cosmetic bottle. It had these cross-hatch-like lines on it that was noticeable on renders. This cosmetic bottle did not have this artifact when I modeled it. (In hindsight, it did but it wasn't apparent because it was close to the world origin)

Create mesh object (cube/cylinder)
Add loop cuts (not necessary, but makes it more apparent, though it already is apparent)
Add a glass shader (Glass BSDF works fine, but I've also tried glass shaders from Realtime Materials & Sanctus Library)
Move object away from world origin (the further, the stronger the artifact)

The artifacts can be seen at close-ups as far as 10 meters away from world origin.
It also depends on which axis. -1000 m on the Y or X produces artifacts on the sides of the object. -1000 m on the Z on produces artifacts on the top & bottom faces, see attached screenshots.

I've attached a frantic video of me reproducing the issue and a blend file.
Thank you!

**System Information** Operating system: macOS Sonoma 14.5 Graphics card: M3 Max **Blender Version** Broken: (Blender 4.2 LTS) Worked: (Blender 4.1.0, Blender 4.1.1 LTS) Spent the whole day trying to reproduce the issue, and as of writing this I discovered that I think **it has to do with how far away the object is from the world origin.** Context: Was working on a project till I noticed something off about an asset I simply appended which was this transparent cosmetic bottle. It had these cross-hatch-like lines on it that was noticeable on renders. This cosmetic bottle did not have this artifact when I modeled it. (In hindsight, it did but it wasn't apparent because it was close to the world origin) 1. Create mesh object (cube/cylinder) 2. Add loop cuts (not necessary, but makes it more apparent, though it already is apparent) 3. Add a glass shader (Glass BSDF works fine, but I've also tried glass shaders from Realtime Materials & Sanctus Library) 4. Move object away from world origin (the further, the stronger the artifact) The artifacts can be seen at close-ups as far as 10 meters away from world origin. It also depends on which axis. -1000 m on the Y or X produces artifacts on the sides of the object. -1000 m on the Z on produces artifacts on the top & bottom faces, see attached screenshots. I've attached a frantic video of me reproducing the issue and a blend file. Thank you!

no crosshatch and crosshatch.blend

1.6 MiB

crosshatch issue.mkv

57 MiB

Screenshot 2024-07-23 at 5.04.37 PM.png

970 KiB

Screenshot 2024-07-23 at 7.37.35 PM.png

4.1 MiB

Screenshot 2024-07-23 at 7.25.15 PM.png

1.5 MiB

Screenshot 2024-07-23 at 7.22.07 PM.png

1.9 MiB

Screenshot 2024-07-23 at 7.25.11 PM.png

2.8 MiB

Screenshot 2024-07-23 at 5.10.02 PM.png

1.2 MiB

Xhaos added the

labels 2024-07-23 13:30:32 +02:00

Alaska commented

2024-07-23 16:45:40 +02:00

I can confirm this issue, with the file provided I can only confirm this issue with MetalRT on (which is enabled by default on M3 or newer Macs).

And I can confirm that the artifacts are less pronounced in 4.1. I will look into bisecting when this was introduced.

I can confirm this issue, with the file provided I can only confirm this issue with MetalRT on (which is enabled by default on M3 or newer Macs). And I can confirm that the artifacts are less pronounced in 4.1. I will look into bisecting when this was introduced.

Alaska added

and removed

labels 2024-07-23 16:45:59 +02:00

Alaska changed title from ~~Cross hatching artifacts on glass BSDF~~ to Regression: Precision artifacts on refractive materials with Cycles MetalRT

2024-07-23 16:46:25 +02:00

Xhaos commented

2024-07-23 16:52:15 +02:00

I can confirm this issue, with the file provided I can only confirm this issue with MetalRT on (which is enabled by default on M3 or newer Macs).

And I can confirm that 4.1 didn't have this issue. I will look into bisecting when this was introduced.

Thanks! I turned Metal RT Off and the issue resolved as well. My workaround for now is to adjust my entire scene to make the refractive object as close as possible to the world origin.

> I can confirm this issue, with the file provided I can only confirm this issue with MetalRT on (which is enabled by default on M3 or newer Macs). > > And I can confirm that 4.1 didn't have this issue. I will look into bisecting when this was introduced. Thanks! I turned Metal RT Off and the issue resolved as well. My workaround for now is to adjust my entire scene to make the refractive object as close as possible to the world origin.

Alaska commented

2024-07-23 18:43:41 +02:00

Bisecting points to 5508b41a40 as the culprit. The changed that caused this issue is the changes to ray offsetting in shade_surface.h

CC @Michael-Jones Can you take a look at this?

Bisecting points to 5508b41a403d6b5b81a2d1d4c6ab4162f823bf5c as the culprit. The changed that caused this issue is the changes to ray offsetting in `shade_surface.h` CC @Michael-Jones Can you take a look at this?

Michael Jones (Apple) commented

2024-08-05 11:34:09 +02:00

Hey @Alaska. Apologies - it seems I was a little overzealous with the optimisations. I'll happily revert that ray offsetting setting change, but wondering if it's worth adding a new unit test to cover it? Also should we make the change in 4.2.x as well as main?

Alaska commented

2024-08-05 11:44:13 +02:00

@Michael-Jones You'll probably want to check with @Sergey about those, but I'll comment to the best of my understanding.

wondering if it's worth adding a new unit test to cover it?

A test for this may not work as different backends may produce different results due to their own precision differences leading to everything but the reference platform failing the test. But testing on actual hardware will reveal if this is a problem worth worrying about. And this scene may be small enough that the differences aren't large enough to be a problem. Double check with Sergey.

Also should we make the change in 4.2.x as well as main?

Now that Blender 4.2 has released, the general process is:

Make the change in Main.
If it meets certain criteria, then add it to the "to be backported" list for 4.2 #124452 and Philipp or Thomas from the Blender foundation will commit your fix to 4.2.X

Since a revert would fix the bug introduced in 4.2, and isn't risky (as it's just restoring old behaviour) I believe it's safe to put on the 4.2 backport list, but double check with Sergey.

@Michael-Jones You'll probably want to check with @Sergey about those, but I'll comment to the best of my understanding. > wondering if it's worth adding a new unit test to cover it? A test for this may not work as different backends may produce different results due to their own precision differences leading to everything but the reference platform failing the test. But testing on actual hardware will reveal if this is a problem worth worrying about. And this scene may be small enough that the differences aren't large enough to be a problem. Double check with Sergey. > Also should we make the change in 4.2.x as well as main? Now that Blender 4.2 has released, the general process is: 1. Make the change in Main. 2. If it meets certain criteria, then add it to the "to be backported" list for 4.2 #124452 and Philipp or Thomas from the Blender foundation will commit your fix to 4.2.X Since a revert would fix the bug introduced in 4.2, and isn't risky (as it's just restoring old behaviour) I believe it's safe to put on the 4.2 backport list, but double check with Sergey.

Sergey Sharybin commented

2024-08-05 14:17:23 +02:00

Alaska covered it very well, don't think I have much to add.

The idea of having tests for such shading render issues sounds very interesting, but, as Alaska said, it might come with consequences. I think we can go on case-by-case basis, and see if tests behave close enough on all supported platforms.

Maybe there are some tricks we can do with thesholding to mitigate known differences between platforms, and yet be able to catch actual regressions.

@Michael-Jones is this something you can look into?

Alaska covered it very well, don't think I have much to add. The idea of having tests for such shading render issues sounds very interesting, but, as Alaska said, it might come with consequences. I think we can go on case-by-case basis, and see if tests behave close enough on all supported platforms. Maybe there are some tricks we can do with thesholding to mitigate known differences between platforms, and yet be able to catch actual regressions. @Michael-Jones is this something you can look into?

Alaska referenced a pull request that will close this issue

2024-08-07 04:28:39 +02:00

Fix #125307: Precision issues with MetalRT #126003

Alaska commented

2024-08-07 04:29:21 +02:00

I've submitted a PR !126003 for this just so the fix can be commited quickly.

I have a wide range of hardware and can investigate whether or not this type of precision test can be safely added to the render test suite.

I've submitted a PR !126003 for this just so the fix can be commited quickly. I have a wide range of hardware and can investigate whether or not this type of precision test can be safely added to the render test suite.

Alaska referenced this issue

2024-08-07 04:32:41 +02:00

Cycles: Improve the coverage of automated regression tests #123012

Sergey Sharybin commented

2024-08-07 10:32:32 +02:00

@Alaska Thanks. That'd be lovely if you look into adding such tests.

Alaska referenced this issue

2024-08-07 15:43:20 +02:00

Cycles: Improve the coverage of automated regression tests #123012

Sergey Sharybin referenced this issue from a commit

2024-08-07 17:20:53 +02:00

Fix #125307: Precision issues with MetalRT