I managed to reproduce the problem!
It starts to appear after 32 frames.
Here the file that I used to test:

• Removed unnecessary raw step access of length_attr_step
• Fixed compile exception with suggestions from Clement
• HairLengthBuffer is now accessed via the stand id instead of the hair id
• Added last missing uniform texture assignment

According to the module wiki, Render & Cycles is responsible for Freestyle nowadays.

• Dereferencing pointer instead of array stile access to length vertbuf

• Merge Master

CC @Kévin Dietrich (kevindietrich) I believe this may be related to rBbbe6d4492823 Not 100% sure.

I just wanted to add, that upon further investigation this issue only seems to affect OptiX for me. CUDA and CPU rendering seem to be un-affected.

Reaffirming the issue here:

• Added rough an not working prototype of length vert buffer implementation for eevee shader

Tried with Radeon-Software-Adrenalin-2020-21.2.3-Win10-64bit-Feb22-Non-WHQL and viewport rendering is still failing. (Blender 2.91.2)

This feature, above all, is the most important. Its one of the most used tools in 3DS (vray, corona, fstorm...) when trying to find the perfect lighting solution.

In D10153#264739, @Brecht Van Lommel (brecht) wrote:

I tried to test this but could not get it working. What I did:

• Ensured I had a recent enough driver.
• Got the latest release from here: https://github.com/GPUOpen-LibrariesAndSDKs/RadeonImageFilter
• Hacked getRadeonSoftwareInstallLocation() to point to a custom location, since I don't have the registry entries. Is there already an installer that sets these? We'll also need a URL for that when we commit this, to mention in the release notes.

I got "Failed to load RadeonImageFilters.dll". The dll exists at the specified location and GetLastError() returns 0, so I'm not immediately sure what to try next, how to get a proper error message.

• Is there a reason this is not available for viewport denoising? That's the main use case I think.
• For the label in the menu, maybe "Radeon Image Filter" is better than "RIF"? At least on AMD websites the abbreviation does not seem to be used prominently.
• It would be good if we could only show the menu entry when actual RIF support is available, rather than just an OpenCL device.

I tried to test this but could not get it working. What I did:

• Ensured I had a recent enough driver.
• Got the latest release from here: https://github.com/GPUOpen-LibrariesAndSDKs/RadeonImageFilter
• Hacked getRadeonSoftwareInstallLocation() to point to a custom location, since I don't have the registry entries. Is there already an installer that sets these? We'll also need a URL for that when we commit this, to mention in the release notes.

Build without requiring AMD RIF install.

Added separator between render and viewport denoising settings and rebased on master

@Christoph Werner (Taros) The test with the AMD 1950X 16 core and GTX 1080Ti shows the following result:

@Christoph Werner (Taros) I made an extreme test with tiles of 512x512. This gives me the suspicion that the GPU takes over the tiles but renders them from scratch because the render time in this case with different Adaptive Sampling settings for a faster result:

Thanks! I'll fix the Eevee implementation as well in the commit.

Render times

In T82351#1117105, @Christoph Werner (Taros) wrote:

In T82351#1117092, @Michael Klein (Renderbricks) wrote:

In T82351#1117081, @Christoph Werner (Taros) wrote:

Excuse my question, but which version is needed to test the tile stealing feature?
I've a nvidia RTX 3090 here and an AMD 3970X CPU with 32 cores. In my tests the "Candidate" version of Blender 2.92 has nearly no difference in rendering speed, or just marginal better.

Is the tile stealing feature implemented in the current 2.92 Candidate version or do I expect too much of this option?

From what I watch here 2.92 RC is using tile stealing. When the GPU has finished it will definitely take over the CPU tiles. From my logic it should be like this: If you have a slow CPU but a very fast GPU there is maybe not much of a difference between GPU only and hybrid mode. Because the GPU might solve the whole render while the CPU is still working on the first tiles. But if you have a very fast CPU there should be a difference. How fast is the AMD 3970X compared to the RTX 3090? I have a 16 core Threadripper and GTX 1080Ti which are equal in speed. I will do a test.

My AMD Ryzen Threadripper 3790X is slower then the RTX 3090. I guess there is no "consumer" CPU, that renders tiles faster than a nvidia RTX 3090 currently...
I have an old workstation with a 1080Ti and will make some tests there, soon. Currently my old machine is not ready to go.

I'd rather not have that description. I can see how it's helpful, but we generally don't do this in the Blender UI and I think it gets in the way too much once it has been read.

In T82351#1117092, @Michael Klein (Renderbricks) wrote:

In T82351#1117081, @Christoph Werner (Taros) wrote:

Excuse my question, but which version is needed to test the tile stealing feature?
I've a nvidia RTX 3090 here and an AMD 3970X CPU with 32 cores. In my tests the "Candidate" version of Blender 2.92 has nearly no difference in rendering speed, or just marginal better.

Is the tile stealing feature implemented in the current 2.92 Candidate version or do I expect too much of this option?

From what I watch here 2.92 RC is using tile stealing. When the GPU has finished it will definitely take over the CPU tiles. From my logic it should be like this: If you have a slow CPU but a very fast GPU there is maybe not much of a difference between GPU only and hybrid mode. Because the GPU might solve the whole render while the CPU is still working on the first tiles. But if you have a very fast CPU there should be a difference. How fast is the AMD 3970X compared to the RTX 3090? I have a 16 core Threadripper and GTX 1080Ti which are equal in speed. I will do a test.

In T82351#1117081, @Christoph Werner (Taros) wrote:

Excuse my question, but which version is needed to test the tile stealing feature?
I've a nvidia RTX 3090 here and an AMD 3970X CPU with 32 cores. In my tests the "Candidate" version of Blender 2.92 has nearly no difference in rendering speed, or just marginal better.

Is the tile stealing feature implemented in the current 2.92 Candidate version or do I expect too much of this option?

Excuse my question, but which version is needed to test the tile stealing feature?
I've a nvidia RTX 3090 here and an AMD 3970X CPU with 32 cores. In my tests the "Candidate" version of Blender 2.92 had nearly no difference in rendering speed, or just marginal better.

Setting to high as it is a recent regression.

This is a must have thing. I signed up to give a token to this feature. That's all I'm going to say:D

Clément Foucault Input

Hello.
Now with Adrenalin 2020 Edition 21.2.2 driver 95% of the bugs has been fixed!
I tested the new driver with an Rx 470 4GB and it worked really well.
How to install it:
-Search on internet "Adrenalin 2020 Edition 21.2.2 driver" and download it.
-Until you install the new driver you have to download AMD Cleanup Utility (You have to delete all the files from current driver)
-Run AMD Cleanup Utility and delete your current driver.
-Install your new 21.2.2 Driver and restart your PC.
-Go in C:\Users\Caesar\AppData\Roaming\Blender Foundation\Blender
-Select your Blender Version ( 2.81 / 2.90 / 2.91) > cache > DELETE kernels folder

I so need this feature to be able to even consider switching to blender.. Without it most of my product shots would be just too time consuming to do.