2.93 - Cycles Volumetrics - GPU and CPU results look inverted (Distance vs Equiangular sampling) #89353

New Issue

Carlo Andreacchio · 2021-06-22T10:33:18+02:00

Carlo Andreacchio commented

2021-06-22 10:33:18 +02:00

System Information
Operating system: Ubuntu 20.04
Graphics card: GTX 1080

Blender Version
Broken: baa7a53974

Short description of error
Volumetrics - GPU and CPU results look inverted

Exact steps for others to reproduce the error

Open attached blend file
Set render device to CPU
Render
Change render slots
Change render device to GPU
Render
Toggle back and forth.

I have tried to simplify it as much as possible from our node network, the map range node is specifically to enhance the contrast (it previously was a colour ramp into a multiply node)

volumetricBug_v001.blend

**System Information** Operating system: Ubuntu 20.04 Graphics card: GTX 1080 **Blender Version** Broken: baa7a53974 **Short description of error** Volumetrics - GPU and CPU results look inverted **Exact steps for others to reproduce the error** 1. Open attached blend file 2. Set render device to CPU 3. Render 4. Change render slots 5. Change render device to GPU 6. Render 7. Toggle back and forth. I have tried to simplify it as much as possible from our node network, the map range node is specifically to enhance the contrast (it previously was a colour ramp into a multiply node) [volumetricBug_v001.blend](https://archive.blender.org/developer/F10187633/volumetricBug_v001.blend) ![image.png](https://archive.blender.org/developer/F10187634/image.png) ![image.png](https://archive.blender.org/developer/F10187636/image.png)

Carlo Andreacchio commented

2021-06-22 10:33:18 +02:00

Added subscriber: @candreacchio

Alaska commented

2021-06-22 11:21:20 +02:00

Added subscriber: @Alaska

Alaska commented

2021-06-22 11:21:20 +02:00

The rendering differences come from the fact CPUs and GPUs render volumes differently in Cycles. CPUs have access to Multi-importance sampling which dynamically adjusts the sampling method of volumes between Distance sampling and Equiangular Sampling depending on what Cycles believes is best. GPU rendering on the other hand only supports Distance sampling.

However, the only rendering difference should be the noise, not the actual look of the volume, something else might be going on here. Might be related to {#80267}?

The rendering differences come from the fact CPUs and GPUs render volumes differently in Cycles. CPUs have access to `Multi-importance sampling` which dynamically adjusts the sampling method of volumes between `Distance sampling` and `Equiangular Sampling` depending on what Cycles believes is best. GPU rendering on the other hand only supports `Distance sampling`. However, the only rendering difference should be the noise, not the actual look of the volume, something else might be going on here. Might be related to {#80267}?

Richard Antalik commented

2021-06-22 15:37:49 +02:00

Added subscriber: @iss

Richard Antalik commented

2021-06-22 15:37:49 +02:00

I can't reproduce this, because my GPU is not capable of rendering volumes, but when I toggle between multiple importance property I did not see any difference in render, so not sure if this plays any role in this case.

Alaska commented

2021-06-22 23:10:25 +02:00

I'm personally able to reproduce the bug. Will wait for someone else's input before confirming the task or merging it.
2021-06-23 09-07-20.mp4

System Information
Operating system: Linux-5.10.0-7-amd64-x86_64-with-glibc2.31 64 Bits
Graphics card: NVIDIA GeForce RTX 3070/PCIe/SSE2 NVIDIA Corporation 4.5.0 NVIDIA 465.31

Blender Version: 3.0.0 Alpha, branch: master, commit date: 2021-06-22 08:32, hash: ad9fd47d7b

I'm personally able to reproduce the bug. Will wait for someone else's input before confirming the task or merging it. [2021-06-23 09-07-20.mp4](https://archive.blender.org/developer/F10188507/2021-06-23_09-07-20.mp4) **System Information** Operating system: Linux-5.10.0-7-amd64-x86_64-with-glibc2.31 64 Bits Graphics card: NVIDIA GeForce RTX 3070/PCIe/SSE2 NVIDIA Corporation 4.5.0 NVIDIA 465.31 **Blender Version:** 3.0.0 Alpha, branch: master, commit date: 2021-06-22 08:32, hash: `ad9fd47d7b`

Carlo Andreacchio commented

2021-06-23 00:19:51 +02:00

Glad someone is able to confirm im not going crazy...

I can confirm that changing the Volume Sampling to 'Distance', makes the CPU render like the GPU render.

However, when I was simplifying the scene, I was getting the same results when i was changing the step rate / max steps. As a result, I dont believe it is linked to #80267. I may be wrong though!

Glad someone is able to confirm im not going crazy... I can confirm that changing the Volume Sampling to 'Distance', makes the CPU render like the GPU render. However, when I was simplifying the scene, I was getting the same results when i was changing the step rate / max steps. As a result, I dont believe it is linked to #80267. I may be wrong though!

Alaska commented

2021-06-23 03:09:10 +02:00

In #89353#1181097, @candreacchio wrote:
Glad someone is able to confirm im not going crazy...

I can confirm that changing the Volume Sampling to 'Distance', makes the CPU render like the GPU render.

However, when I was simplifying the scene, I was getting the same results when i was changing the step rate / max steps. As a result, I dont believe it is linked to #80267. I may be wrong though!

I was able to reduce the rendering difference by changing the step rate and/or scaling down the object. However, I had to change the step rate to a value lower than the soft limit set by Blender (E.G. 0.01) to reduce the rendering difference without scaling the object. I don't really have much else to add. I'll let someone else decide what to do with this report (confirm it as a bug, merge it with another report, or something else)

Maybe Cycles-X will handle volumes in such a way this issue is resolved? I don't know.

> In #89353#1181097, @candreacchio wrote: > Glad someone is able to confirm im not going crazy... > > I can confirm that changing the Volume Sampling to 'Distance', makes the CPU render like the GPU render. > > However, when I was simplifying the scene, I was getting the same results when i was changing the step rate / max steps. As a result, I dont believe it is linked to #80267. I may be wrong though! I was able to reduce the rendering difference by changing the step rate and/or scaling down the object. However, I had to change the step rate to a value lower than the soft limit set by Blender (E.G. 0.01) to reduce the rendering difference without scaling the object. I don't really have much else to add. I'll let someone else decide what to do with this report (confirm it as a bug, merge it with another report, or something else) Maybe Cycles-X will handle volumes in such a way this issue is resolved? I don't know.

Philipp Oeser commented

2021-08-10 16:48:14 +02:00

Added subscriber: @lichtwerk

Philipp Oeser commented

2021-08-10 16:48:14 +02:00

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

Philipp Oeser commented

2021-08-10 16:48:14 +02:00

Can also confirm the difference.

I think at this point it would be good to have a rendering dev comment on this situation (since I would also expect the sampling method to only affect noise/performance -- and not the final values?)

Can also confirm the difference. I think at this point it would be good to have a rendering dev comment on this situation (since I would also expect the sampling method to only affect noise/performance -- and not the final values?)

Philipp Oeser changed title from ~~2.93 - Cycles Volumetrics - GPU and CPU results look inverted~~ to 2.93 - Cycles Volumetrics - GPU and CPU results look inverted (Distance vs Equiangular sampling)

2021-08-10 16:48:44 +02:00

William Leeson commented

2021-09-15 16:13:03 +02:00

Added subscriber: @leesonw

William Leeson commented

2021-09-15 16:13:03 +02:00

I could not get the images to differ with the latest cycles-x build from the HEAD commit 58ab616d4c6e5d65d9415bb54845eddb0acf54e7 (origin/cycles-x) Have you tested this recently?

Philipp Oeser commented

2021-09-15 16:34:05 +02:00

In #89353#1220546, @leesonw wrote:
I could not get the images to differ with the latest cycles-x build from the HEAD commit 58ab616d4c6e5d65d9415bb54845eddb0acf54e7 (origin/cycles-x) Have you tested this recently?

Can you repro in master though?

> In #89353#1220546, @leesonw wrote: > I could not get the images to differ with the latest cycles-x build from the HEAD commit 58ab616d4c6e5d65d9415bb54845eddb0acf54e7 (origin/cycles-x) Have you tested this recently? Can you repro in master though?

William Leeson commented

2021-09-16 14:01:57 +02:00

Yes I can repo this on the master branch.