Roughness below 0.01 treated as mirror by PATH engine

[Martini]

I have gathered some evidence... see the following render from Blender 2.80, LuxCore 2.2rc1:

12,288spp rendered at 4K but scaled to 1080p for upload

I checked the output .scn file as well, the exported materials look correct.

Material for roughness 0.01:

Code: Select all

scene.textures.2441963438968Fresnel.type = "fresnelconst"
scene.textures.2441963438968Fresnel.n = 0.0981203318 0.373658836 1.57057941
scene.textures.2441963438968Fresnel.k = 3.35364223 2.52603722 1.76780081
scene.textures.Implicit-ConstFloatTexture-2e7883ad-e292-43a8-a2d3-a4b9df7771f3.type = "constfloat1"
scene.textures.Implicit-ConstFloatTexture-2e7883ad-e292-43a8-a2d3-a4b9df7771f3.value = 0.00999999978
scene.textures.Implicit-ConstFloatTexture-cf3860dd-d6b1-41e7-9c39-d391e48fc320.type = "constfloat1"
scene.textures.Implicit-ConstFloatTexture-cf3860dd-d6b1-41e7-9c39-d391e48fc320.value = 0.00999999978
scene.materials.MetallicPointOhOne2440874536696.type = "metal2"
scene.materials.MetallicPointOhOne2440874536696.fresnel = "2441963438968Fresnel"
scene.materials.MetallicPointOhOne2440874536696.uroughness = "Implicit-ConstFloatTexture-2e7883ad-e292-43a8-a2d3-a4b9df7771f3"
scene.materials.MetallicPointOhOne2440874536696.vroughness = "Implicit-ConstFloatTexture-cf3860dd-d6b1-41e7-9c39-d391e48fc320"
scene.materials.MetallicPointOhOne2440874536696.id = 13398304
scene.materials.MetallicPointOhOne2440874536696.emission.gain = 1 1 1
scene.materials.MetallicPointOhOne2440874536696.emission.power = 0
scene.materials.MetallicPointOhOne2440874536696.emission.efficency = 0
scene.materials.MetallicPointOhOne2440874536696.emission.theta = 90
scene.materials.MetallicPointOhOne2440874536696.emission.id = 0
scene.materials.MetallicPointOhOne2440874536696.emission.importance = 1
scene.materials.MetallicPointOhOne2440874536696.emission.directlightsampling.type = "AUTO"
scene.materials.MetallicPointOhOne2440874536696.visibility.indirect.diffuse.enable = 1
scene.materials.MetallicPointOhOne2440874536696.visibility.indirect.glossy.enable = 1
scene.materials.MetallicPointOhOne2440874536696.visibility.indirect.specular.enable = 1
scene.materials.MetallicPointOhOne2440874536696.shadowcatcher.enable = 0
scene.materials.MetallicPointOhOne2440874536696.shadowcatcher.onlyinfinitelights = 0

Material for roughness 0.00999:

Code: Select all

scene.textures.2441963428056Fresnel.type = "fresnelconst"
scene.textures.2441963428056Fresnel.n = 0.0981203318 0.373658836 1.57057941
scene.textures.2441963428056Fresnel.k = 3.35364223 2.52603722 1.76780081
scene.textures.Implicit-ConstFloatTexture-cc8d6cb8-5537-4475-bdc2-5332e2e51548.type = "constfloat1"
scene.textures.Implicit-ConstFloatTexture-cc8d6cb8-5537-4475-bdc2-5332e2e51548.value = 0.00999000017
scene.textures.Implicit-ConstFloatTexture-61406085-9f72-4eaa-adc4-6d216b01460c.type = "constfloat1"
scene.textures.Implicit-ConstFloatTexture-61406085-9f72-4eaa-adc4-6d216b01460c.value = 0.00999000017
scene.materials.MetallicPointOhOhNines2440874528968.type = "metal2"
scene.materials.MetallicPointOhOhNines2440874528968.fresnel = "2441963428056Fresnel"
scene.materials.MetallicPointOhOhNines2440874528968.uroughness = "Implicit-ConstFloatTexture-cc8d6cb8-5537-4475-bdc2-5332e2e51548"
scene.materials.MetallicPointOhOhNines2440874528968.vroughness = "Implicit-ConstFloatTexture-61406085-9f72-4eaa-adc4-6d216b01460c"
scene.materials.MetallicPointOhOhNines2440874528968.id = 706207
scene.materials.MetallicPointOhOhNines2440874528968.emission.gain = 1 1 1
scene.materials.MetallicPointOhOhNines2440874528968.emission.power = 0
scene.materials.MetallicPointOhOhNines2440874528968.emission.efficency = 0
scene.materials.MetallicPointOhOhNines2440874528968.emission.theta = 90
scene.materials.MetallicPointOhOhNines2440874528968.emission.id = 0
scene.materials.MetallicPointOhOhNines2440874528968.emission.importance = 1
scene.materials.MetallicPointOhOhNines2440874528968.emission.directlightsampling.type = "AUTO"
scene.materials.MetallicPointOhOhNines2440874528968.visibility.indirect.diffuse.enable = 1
scene.materials.MetallicPointOhOhNines2440874528968.visibility.indirect.glossy.enable = 1
scene.materials.MetallicPointOhOhNines2440874528968.visibility.indirect.specular.enable = 1
scene.materials.MetallicPointOhOhNines2440874528968.shadowcatcher.enable = 0
scene.materials.MetallicPointOhOhNines2440874528968.shadowcatcher.onlyinfinitelights = 0

As you can see it's virtually identical. Exported roughness values are 0.00999999978 vs 0.00999000017, yet somehow the second one gets rendered as a perfect mirror (as though roughness were 0). Also you may note that Glossy material is also affected, and the roughness on the Disney material is not working at all.

It also renders the same way from within LuxCoreUI:

Rendered in LuxCoreUI

I have tested with Light Tracing on and off, makes no difference. CPU vs GPU makes no difference.

The only thing that does make a difference is BIDIRCPU. Roughness works just fine with BiDir engine!

Rendered in LuxCoreUI with BIDIRCPU

It seems reasonable to me to conclude that the bug may lie within the PATH LuxCore engine.

[Dade]

I'm seeing the problem only with OpenCL, not with CPU rendering. Are you sure to have the same problem with CPU rendering (don't forget to disable hybrid CPU/GPU rendering when doing the test) ?

[Martini]

I'm seeing the problem only with OpenCL, not with CPU rendering. Are you sure to have the same problem with CPU rendering (don't forget to disable hybrid CPU/GPU rendering when doing the test) ?

Well, you are right of course. Apologies, I could've sworn I tested that!

Ah yes, so it seems limited to OpenCL... well, hopefully that can help narrow down possible locations for the cause even more, right?

[Dade]

Ok, it was intentional, there was lower cap at sqrt(.0001): https://github.com/LuxCoreRender/LuxCor ... ric.cl#L68

Code: Select all

// Values of roughness < .0001f seems to trigger some kind of exceptions with
// AMD OpenCL on GPUs. The result is a nearly freeze of the PC.

I have remove the work around and the bug in AMD drivers seems to have been fixed.

About Disney roughness and Glossy/Metal roughness: as far as I remember Disney roughness is equal to the square root of Glossy/Metal roughness because "(Disney) roughness = (Glossy/Metal) uroughness * vroughness".

P.S. sqrt(.0001) = 0.001

[Dade]

Side note: Disney material glossiness is now computed as the square root of the previous value to match Glossy2/Metal roughness (for PhtonGI and Hybrid rendering).

[lacilaci]

Side note: Disney material glossiness is now computed as the square root of the previous value to match Glossy2/Metal roughness (for PhtonGI and Hybrid rendering).

oh no... Disney values are more intuitive to use. Why should I type stupid values like 0.0whatever and see some big changes in roughness?

[B.Y.O.B.]

oh no... Disney values are more intuitive to use. Why should I type stupid values like 0.0whatever and see some big changes in roughness?

It's only about the glossiness threshold for PhotonGI and light tracing.

[lacilaci]

oh no... Disney values are more intuitive to use. Why should I type stupid values like 0.0whatever and see some big changes in roughness?

It's only about the glossiness threshold for PhotonGI and light tracing.

ohh. So no reason to panic then thanks for explanation

[Martini]

I have remove the work around and the bug in AMD drivers seems to have been fixed.

Thanks Dade!

BTW, had a look at your commit and although I don't understand what the u0 value is, it does seem there is no longer anything stopping roughness == 0, which may cause a divide-by-zero error in the case where u0 == 0 also. Was that the original cause of the crash? I know BlendLuxCore would prevent you from entering roughness less than 0.001 but you could do it by hand...

Is there a less restrictive sensible minimum? .0001f was not enough due to coming from input 0.01 squared, maybe something like .000001f (which is input 0.001 squared). Or just set cos2Theta to 1.f when roughness == 0.f?

Sorry these are just ideas and it is probably clear by now that although I'm a programmer, I don't know how the maths behind ray tracing actually works so take these suggestions with due caution.

[Dade]

BTW, had a look at your commit and although I don't understand what the u0 value is, it does seem there is no longer anything stopping roughness == 0, which may cause a divide-by-zero error in the case where u0 == 0 also. Was that the original cause of the crash? I know BlendLuxCore would prevent you from entering roughness less than 0.001 but you could do it by hand...

Is there a less restrictive sensible minimum? .0001f was not enough due to coming from input 0.01 squared, maybe something like .000001f (which is input 0.001 squared). Or just set cos2Theta to 1.f when roughness == 0.f?

Sorry these are just ideas and it is probably clear by now that although I'm a programmer, I don't know how the maths behind ray tracing actually works so take these suggestions with due caution.

"u0" is a sampling variable, it is a "random" value between 0.0 and 1.0. U/V roughness is clamped between 1e-9 and 1.0 so roughness can not be 0.0: https://github.com/LuxCoreRender/LuxCor ... 2.cpp#L162