AgX on the GPU displays incorrect Object texture coordinates #113220

New Issue

Jesse Yurkovich · 2023-10-03T21:13:38+02:00

Jesse Yurkovich commented

2023-10-03 21:13:38 +02:00

System Information
Operating system: Windows-10-10.0.22621-SP0 64 Bits
Graphics card: NVIDIA GeForce RTX 4090/PCIe/SSE2 NVIDIA Corporation 4.6.0 NVIDIA 528.24

Blender Version
Broken: version: 4.1.0 Alpha, branch: main, commit date: 2023-10-02 08:29, hash: 7f7e1ab08842

Short description of error
AgX on the GPU differs from the CPU implementation for Object coordinates. This seems to only impact the visual, not the actual coordinates, but is fairly extreme.

Exact steps for others to reproduce the error

Load attached .blend
Switch to Viewport Render mode
Try to view the Object Texture Coordinates using Cycles CPU [1]
Now switch to Cycles GPU [2]

Addendum: An odd thing happens with Final renders. When trying F12 render in the attached file you'll get both outputs. The "Noisy Image" will be skewed, but the "Combined" image will be ok.

[1]

[2]

**System Information** Operating system: Windows-10-10.0.22621-SP0 64 Bits Graphics card: NVIDIA GeForce RTX 4090/PCIe/SSE2 NVIDIA Corporation 4.6.0 NVIDIA 528.24 **Blender Version** Broken: version: 4.1.0 Alpha, branch: main, commit date: 2023-10-02 08:29, hash: `7f7e1ab08842` **Short description of error** AgX on the GPU differs from the CPU implementation for Object coordinates. This seems to only impact the visual, not the actual coordinates, but is fairly extreme. **Exact steps for others to reproduce the error** - Load attached .blend - Switch to Viewport Render mode - Try to view the Object Texture Coordinates using Cycles CPU [1] - Now switch to Cycles GPU [2] Addendum: An odd thing happens with Final renders. When trying F12 render in the attached file you'll get both outputs. The "Noisy Image" will be skewed, but the "Combined" image will be ok. [1] ![image](/attachments/a140ec1a-c5bf-42a8-8d56-10fdb435b756) [2] ![image](/attachments/667aa0cb-4d40-418f-acd6-f5c7c99aa710)

image.png

134 KiB

agx-objectcoords.blend

923 KiB

image.png

132 KiB

Jesse Yurkovich added the

labels 2023-10-03 21:13:39 +02:00

YimingWu commented

2023-10-04 03:55:27 +02:00

I think this is mostly due to it's using LUTs to display data channels... But it should not differ on CPU and GPU. Could you check if you could click on rendered image and see if the same position has the same rgb values?

Jesse Yurkovich commented

2023-10-04 04:07:41 +02:00

For final renders, where I can view the rgb values, the result is the same for CPU and GPU there; but as noted in the description things are still skewed for "Noisy Image" but seem ok for the "Combined" image.

Pratik Borhade commented

2023-10-04 06:33:36 +02:00

I can confirm.
cc @Eary @nathanvegdahl

I can confirm. cc @Eary @nathanvegdahl

Pratik Borhade added

and removed

labels 2023-10-04 06:34:09 +02:00

Zijun Zhou commented

2023-10-04 07:22:02 +02:00

This is not a bug. First of all texture coordinate is not color, they are Generic Data, A.K.A Non-Color data. They have blue sockets, meaning they are vectors not colors. They are, as their name suggests, coordinate data to map a texture onto an object.

What happens if you send unbounded vectors to AgX? It thought they are wider gamut values! What else would AgX think they are?

EEVEE clips negative values all the time. Cycles CPU viewport has always clipped wider gamut values (this holds true even in Spectral branch), but end render after finished will display unclipped result. The denoiser will clip wider gamut values, it has always been the case, even in Spectral Cycles branch.

Here is what the denoiser has done to the above data:

So this is expected behavior of AgX treating negative values as wider gamut. And you just start noticing negative values getting clipped by different places of Blender, which we need to eventually deal with after Spectral Cycles becomes ready.

Not a bug here.

This is not a bug. First of all texture coordinate is not color, they are Generic Data, A.K.A Non-Color data. They have blue sockets, meaning they are vectors not colors. They are, as their name suggests, coordinate data to map a texture onto an object. What happens if you send unbounded vectors to AgX? It thought they are wider gamut values! What else would AgX think they are? ![image](/attachments/a5f562d2-7780-4a4e-a57a-19a0b5fda326) EEVEE clips negative values all the time. Cycles CPU viewport has always clipped wider gamut values (this holds true even in Spectral branch), but end render after finished will display unclipped result. The denoiser will clip wider gamut values, it has always been the case, even in Spectral Cycles branch. Here is what the denoiser has done to the above data: ![image](/attachments/c1cc7f78-c971-4750-b888-2d79f7549afe) So this is expected behavior of AgX treating negative values as wider gamut. And you just start noticing negative values getting clipped by different places of Blender, which we need to eventually deal with after Spectral Cycles becomes ready. Not a bug here.

image.png

702 KiB

image.png

551 KiB

👍 1

Jesse Yurkovich commented

2023-10-04 20:22:11 +02:00

Why the difference between Cycles CPU viewport and Cycles GPU viewport?

All in all this is good information to put in the Documentation - about the clamping and the negative values and the denoisers etc. You can even use the texture coordinate as a visual example to demonstrate. A half dozen people occupied their time on this issue the other day before I filed it on their behalf. It would be good to just have something written down that would describe the behavior they're seeing.

Why the difference between Cycles CPU viewport and Cycles GPU viewport? All in all this is good information to put in the Documentation - about the clamping and the negative values and the denoisers etc. You can even use the texture coordinate as a visual example to demonstrate. A half dozen people occupied their time on this issue the other day before I filed it on their behalf. It would be good to just have something written down that would describe the behavior they're seeing.

Miguel Pozo referenced this issue

2023-11-17 17:44:36 +01:00

GPU: Sanitize closure nodes inputs #115059

Miguel Pozo referenced this issue from a commit

2023-11-21 20:16:08 +01:00

GPU: Sanitize closure nodes inputs

Zijun Zhou referenced this issue

2024-01-19 06:54:47 +01:00

AgX Implementation doesn't seem right on Images/UV Editors #117271

Richard Antalik referenced this issue

2024-01-19 21:08:40 +01:00

AgX Implementation doesn't seem right on Images/UV Editors #117271