2024-09-03 Render & Cycles Meeting

Meetings
1

Attendees

  • Christophe Hery (Meta)
  • Nikita Sirgienko (Intel)
  • Patrick Mours (NVIDIA)
  • Sahar A. Kashi (AMD)
  • Sergey Sharybin (Blender)
  • Weizhen Huang (Blender)

Notes

  • For 4.3
    • Metallic BSDF has landed
    • Artisitc Conductor Fresnel type to Metallic BSDF needs to be looked at
    • HIP-RT is getting closer to be ready
      • It works pretty well on the AMD test machine at the studio, main concerns from Cycles side, and from CI/CD are resolved
      • Waiting for platform maintainers to finish review
  • Nikita was looking into failing tests with oneAPI
    • The Light Linking tests seems to fail due to noise difference
      • Maybe we already diverting from the “clean” result on other platforms, but just managing to stay within the threshold?
      • Usual way of either increasing sample count or test tolerance seems to be the way to go.
      • Weizhen will help Nikitra to cover more platforms
    • Difference in rendering hair on different platforms
      • To keep GPU rendering fast, it seems that the practical solution seems to be align CPU to GPU
      • Not high priority issue
  • Weizhen is working on volume algorithms
    • Ray marching is the old method, slower but lower variance
    • The new tracking based methods are faster, unbiased, but higher variance
  • OptiX 7.4
    • Christophe runs into some weird and wonderful issues. Might need some changes from the Cycles side, not clear yet.

Practical Info

This is a weekly video chat meeting for planning and discussion of Blender rendering development. Any contributor (developer, UI/UX designer, writer, …) working on rendering in Blender is welcome to join and add proposed items to the agenda.

For users and other interested parties, we ask to read the meeting notes instead so that the meeting can remain focused.

16 Likes
2

There are quite a few tests like this. For example a lot of light tree tests are no longer clean after the bounding sphere change, but they still pass so they weren’t updated. I believe the bevel bake test isn’t clean. And there’s probably a few others.

3

Then, perhaps, we should try updating the reference images of failing tests and see if that helps.
Not sure we should update all of the files yet.

4

I hope these volume algorithms can also be used for an SDF modeling toolset in Blender. It’s about time for Blender to jump onto the SDF train.

Have a look at this thread and these videos for SDF modeling info and impressions.

To me personally, since I started SDF-modeling, polygon subdivision modeling feels like a previous chapter.

6 Likes
5

Finally getting unbiased volume rendering would be huge

1 Like
6

I am sure there is a good technical reason for this but out of pure curiosity… Why not going straight to Optix 8.0?

7

Increasing the OptiX library version typically comes with a increase in GPU driver requirements for the end user.
Applications compiled with OptiX 7.4 need driver 495 or newer, while applications compiled with 8.0 need 535 or newer.

For most users a increase in driver requirements isn’t a issue, but for some it’s a hassle. Examples of why it might be a hassle:

  • You’re on a LTS Linux Distrobution that doesn’t offer newer GPU drivers.
  • You work at a company where stability and reliability is important, so updating a GPU driver requires extensive testing from the IT department before you get permission to update.

Updating OptiX versions can also drop support for older generations of GPUs that the Cycles team want to support. So they only try to update when neccesary. This is already a issue with CUDA. CUDA 12.X drops support for some of the GPUs the Cycles team wants to support, so they use up to 3 different versions of CUDA (10.X, 11.X, 12.X) to compile GPU kernels for all the devices they want to support.

8 Likes
8

Thanks for the informative reply.

I am used to how development on Davinci Resolve is moving forward as i am colaborating with the team there. It is almost the exact opposite in philosophy. They usually will not require the absolute latest drivers but every 1 to 1.5 year they will add something that will need a driver update for sure. So far it has worked to our benefit although i as well had to upgrade my ancient GPU. (i am on 555.99 now)

My initial curiosity was because Optix 8.0 had a whole year to mature and maybe issues that are present in 7.4 for the Cycles team are already solved on 8.0.

Again, your explanation was clear and i totaly get why it is the way it is. Thanks!

2 Likes
9

Thanks for your input, but I’d prefer meeting notes threads not being converted to, essentially, feature requests.

2 Likes
10

I understand. Sorry for that.