Cycles: NaN pixels in Glossy and Transmissive Passes with roughness values near 0.0266 #125750

New Issue

AgentA1cr · 2024-08-01T04:35:01+02:00

AgentA1cr commented

2024-08-01 04:35:01 +02:00

System Information
Operating system: Windows-10-10.0.19045-SP0 64 Bits
Graphics card: Quadro K5100M/PCIe/SSE2 NVIDIA Corporation 4.6.0 NVIDIA 426.78

Blender Version
Broken: version: 4.2.0, branch: blender-v4.2-release, commit date: 2024-07-16 06:20, hash: a51f293548ad
Worked: Unknown. I've checked as far back as 3.6

Short description of error
In the Cycles engine, Roughness values between 0.026592 and 0.026679 cause Not a Number pixels to appear in the Direct and Indirect Glossy and Transmissive passes.
Similar to blender/blender#108125. Unlike that bug, this one appears in both CPU and GPU rendering. It appears in the Glossy passes with or without Transmission>0, but only appears in the Transmission passes when Transmission>0.

Exact steps for others to reproduce the error
Open attached .blend file. (Glossy Glitch Test.blend)
Render Image
Open Compositing tab
A large number of pixels are NaN, as shown by the node network.

**System Information** Operating system: Windows-10-10.0.19045-SP0 64 Bits Graphics card: Quadro K5100M/PCIe/SSE2 NVIDIA Corporation 4.6.0 NVIDIA 426.78 **Blender Version** Broken: version: 4.2.0, branch: blender-v4.2-release, commit date: 2024-07-16 06:20, hash: `a51f293548ad` Worked: Unknown. I've checked as far back as 3.6 **Short description of error** In the Cycles engine, Roughness values between 0.026592 and 0.026679 cause Not a Number pixels to appear in the Direct and Indirect Glossy and Transmissive passes. Similar to blender/blender#108125. Unlike that bug, this one appears in both CPU and GPU rendering. It appears in the Glossy passes with or without Transmission>0, but only appears in the Transmission passes when Transmission>0. <img width="750" alt="image" src="attachments/724d52ab-4ace-46d1-8775-27d6f6f86bba"> <img width="842" alt="image" src="attachments/3cbaa076-dd5d-4996-beeb-8cadfb0446f8"> **Exact steps for others to reproduce the error** Open attached .blend file. ([Glossy Glitch Test.blend](/attachments/53fc87e6-cc93-4f0c-8011-cb179b6bb5a5)) Render Image Open Compositing tab A large number of pixels are NaN, as shown by the node network.

image.png

344 KiB

Glossy Glitch Test.blend

1.0 MiB

image.png

76 KiB

image.png

316 KiB

AgentA1cr added the

labels 2024-08-01 04:35:02 +02:00

Alaska commented

2024-08-01 05:39:21 +02:00

I tested a few different configurations and here are my observation.

CPU Rendering on Windows with a AMD Ryzen 5 5600: Has the issue
GPU Rendering on Windows with a AMD RX 7800XT: Has the issue

CPU Rendering on Linux with a AMD Ryzen 5 5600: No issue
CPU Rendering on macOS with a M1 Pro: No issue
GPU Rendering on macOS with a M1 Pro: No issue
GPU rendering on Windows with a NVIDIA RTX 4090: No issue
GPU Rendering on Windows with a Intel Arc A750: No issue

If there is a configuration that could be tested but that isn't listed (E.g. GPU rendering on Linux, it just means I didn't test it)

I tested a few different configurations and here are my observation. CPU Rendering on Windows with a AMD Ryzen 5 5600: Has the issue GPU Rendering on Windows with a AMD RX 7800XT: Has the issue CPU Rendering on Linux with a AMD Ryzen 5 5600: No issue CPU Rendering on macOS with a M1 Pro: No issue GPU Rendering on macOS with a M1 Pro: No issue GPU rendering on Windows with a NVIDIA RTX 4090: No issue GPU Rendering on Windows with a Intel Arc A750: No issue If there is a configuration that could be tested but that isn't listed (E.g. GPU rendering on Linux, it just means I didn't test it)

Alaska added

and removed

labels 2024-08-01 05:39:33 +02:00

Alaska commented

2024-08-01 07:24:43 +02:00

I've done some debugging into this.

The issue originates from this line of code:

return alpha2 / (M_PI_F * sqr(1.0f + (alpha2 - 1.0f) * cos_NH2));

# alpha2 = 5.01428076e-07
# cos_NH2 = 1.00000048
# Which basically results in 0.0000005 / 0.0 which becomes infinity or NaN.

		intern/cycles/kernel/closure/bsdf_microfacet.h
		Line 518 in c502368141
	
				    return alpha2 / (M_PI_F * sqr(1.0f + (alpha2 - 1.0f) * cos_NH2));

The code gets there through the path of:

integrate_surface_bsdf_bssrdf_bounce()
surface_shader_bsdf_sample_closure()
bsdf_sample()
bsdf_microfacet_ggx_sample()
bsdf_microfacet_sample()
bsdf_D()

I should note that bsdf_D is only called if the material is !m_singular. And m_singular is basically alpha_2 < 5.0e-7. Considering that alpha_2 is right on the verge of this threshold, I think that increasing the threshold may be the way to fix this issue.

Edit: Increasing the threshold probably isn't the best approach. The minimum threshold I found that worked without issues was 8e-7, which is high enough to create a sudden change when transitioning from low roughness to something higher than that value.

I've done some debugging into this. The issue originates from this line of code: ``` return alpha2 / (M_PI_F * sqr(1.0f + (alpha2 - 1.0f) * cos_NH2)); # alpha2 = 5.01428076e-07 # cos_NH2 = 1.00000048 # Which basically results in 0.0000005 / 0.0 which becomes infinity or NaN. ``` https://projects.blender.org/blender/blender/src/commit/c5023681412b122f3b16624881a5d0f84f85b917/intern/cycles/kernel/closure/bsdf_microfacet.h#L518 The code gets there through the path of: ``` integrate_surface_bsdf_bssrdf_bounce() surface_shader_bsdf_sample_closure() bsdf_sample() bsdf_microfacet_ggx_sample() bsdf_microfacet_sample() bsdf_D() ``` I should note that `bsdf_D` is only called if the material is `!m_singular`. And `m_singular` is basically `alpha_2 < 5.0e-7`. Considering that `alpha_2` is right on the verge of this threshold, I think that increasing the threshold may be the way to fix this issue. Edit: Increasing the threshold probably isn't the best approach. The minimum threshold I found that worked without issues was `8e-7`, which is high enough to create a sudden change when transitioning from low roughness to something higher than that value.

Alaska referenced a pull request that will close this issue

2024-08-01 08:47:55 +02:00

Fix #125750: NaN on Glossy materials with low roughness #125756

Alaska referenced this issue

2024-08-01 08:48:40 +02:00

Fix #125750: NaN on Glossy materials with low roughness #125756

david494 commented

2024-08-01 15:13:10 +02:00

For these particular values the NaN can be avoided by rewriting that line as:

return alpha2 / (M_PI_F * sqr((1.0f - cos_NH2) + alpha2 * cos_NH2));

That looks like an overall improvement to floating point precision in any case, but I don't know if that is enough. I don't have access to a system with the issue at the moment.

For these particular values the NaN can be avoided by rewriting that line as: `return alpha2 / (M_PI_F * sqr((1.0f - cos_NH2) + alpha2 * cos_NH2));` That looks like an overall improvement to floating point precision in any case, but I don't know if that is enough. I don't have access to a system with the issue at the moment.

Alaska commented

2024-08-01 15:19:03 +02:00

@david494 Thanks for the suggestion. We've identified the main thing for the fix (cos_NH2 should be 1.0 at the maximum). Once that is merged, feel free to open a new pull request for increased precision.