2023-03-07 Render & Cycles Meeting

Attendees

  • Brecht Van Lommel (Blender)
  • Nikita Sirgienko (Intel)
  • Stefan Werner (Intel)
  • Xavier Hallade (Intel)
  • Patrick Mours (NVIDIA)
  • Christophe Hery (Meta)
  • William Leeson
  • Michael Jones (Apple)
  • Brian Savery (AMD)

Notes

  • Cluster Render: William has a WIP patch for optimizing primarily GPU transfers of geometry, so that it can be done in parallel for multiple GPUs and asynchronosuly. Brecht will review when it gets closer to ready.

  • OpenImageDenoise: Stefan is working on integrating v2 with GPU support. This will involve some refactoring to deduplicate code with the OptiX GPU denoising backend. A question is when to enable GPU denoising, since this does use additional GPU memory and it can run out. Solution will most likely be to add OIDN CPU and GPU as separate options in the denoisers menu, defaulting to GPU for the viewport.

  • HIP-RT: Brian notes Windows driver with all the required functionality was released, which will help testing the patch. Brecht will review the latest iteration of the patch when it gets uploaded to projects.blender.org.

  • CPU + GPU: Michael will revive the patch to improve performance of this, since it simplifies the code and may also help with multi GPU this seems worth doing.

  • GPU Preferences: hardware ray-tracing for Intel and AMD should become checkboxes similar to Metal. For Blender 4.0 we may refactor this so it’s the same for OptiX, and perhaps remove the device selection entirely to give one list of GPUs instead.

  • Microfacet Hair: Christophe ask about status of hair orientation for elliptical hair, where there was some plan to compute this in geometry nodes. Check with Weizhen about the status, but most likely not implement yet.

  • Embree: Nikita plans to submit patches, including v4 support as well as optimizations that also help with v3.

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.

18 Likes

HIP Ray Tracing patch for those interested: #105538 - WIP: Cycles HIP-RT for hardware ray-tracing - blender - Blender Projects

4 Likes

Perhaps it could always try to use GPU (unless the user has specifically chosen CPU), but swap to CPU automatically if there’s insufficient VRAM when the denoising process begins, for example if persistent data is enabled and it’s a large scene.

For the denoise node in the compositor, I guess if persistent data is not enabled, then the VRAM for the just rendered frame should always be emptied prior to the compositor beginning? Making sure plenty of space is available for the GPU denoising.