Maniphest T81473

Render artefacts with GPU / motion blur
Needs Information from Developers, Normal

Assigned To
None
Authored By
Steffen Dünner (SteffenD)
Oct 5 2020, 11:38 PM
Tags
  • BF Blender
  • Render & Cycles
Subscribers
Henrik Dick (weasel)
Philipp Oeser (lichtwerk)
Steffen Dünner (SteffenD)

Description

System Information
Operating system: Linux-5.4.0-48-generic-x86_64-with-debian-bullseye-sid 64 Bits
Graphics card: GeForce GTX 1080 Ti/PCIe/SSE2 NVIDIA Corporation 4.5.0 NVIDIA 450.80.02

Blender Version
Broken: version: 2.91.0 Alpha, branch: master, commit date: 2020-10-05 20:26, hash: rB7951312d4b5c + official releases and master from build bot.
Worked: none that I tried so far

Short description of error
I get weird rendering artefacts where some buckets render different than others. This only seems to happen when:
Rendering with motion blur AND rendering with GPU AND rendering under a Linux OS AND the shutter position is set to "start" or "center on frame". "end of frame" renders fine.
It's completely weird as it doesn't happen under Win10 (at least in my case). I completely reduced a production scene to the point where some basic geometry and particles with a sky shader show the problem.

Exact steps for others to reproduce the error

  • Open the attached blend file:

  • Press F12 to render
  • If you're under Windows (and maybe Mac) or render with CPU you should be fine. Otherwise you should get something like this:

  • Switch the motion blur position from "center on frame" to "end on frame" and render again:


Everything fine

  • Now switch back to "center on frame" and switch the render engine to CPU and render again:


Fine again.

So it only affects Linux, "start or center on frame" and GPU rendering.

P.S.: Might be somehow related to https://developer.blender.org/T76792 although I can't reproduce his problems. But the description sounds similar and he's on Linux, too.

Related Objects

Mentioned Here
rB659f81f1cbc3: Correct last commit (where_is_temp refactor)
T76792: Motion Blur issue with GPU on individual frames in Cycles

Event Timeline

Steffen Dünner (SteffenD) created this task.Oct 5 2020, 11:38 PM
Henrik Dick (weasel) added a subscriber: Henrik Dick (weasel).EditedOct 6 2020, 11:23 PM

Can not confirm even though I have a very similar setup to yours. Is it consistent for you even when restarting blender? Did you try a restart of your computer?

System Information
Operating system: Linux-5.8.0-2-amd64-x86_64-with-debian-bullseye-sid 64 Bits
Graphics card: GeForce GTX 1070/PCIe/SSE2 NVIDIA Corporation 4.5.0 NVIDIA 450.66

Blender Version: current master

Steffen Dünner (SteffenD) added a comment.Oct 6 2020, 11:29 PM

I restarted everything of course, not only Blender. Maybe it's hidden somewhere in a GPU cache or whatever. It happens with all kinds of builds (official and my own). At first I thought it might be a problem with me using CUDA 11 to compile blender, but also the official master builds show that problem.
Your setup indeed looks very similar to mine.

Philipp Oeser (lichtwerk) changed the task status from Needs Triage to Needs Information from User.Nov 19 2020, 6:24 PM
Philipp Oeser (lichtwerk) added a project: Render & Cycles.
Philipp Oeser (lichtwerk) added a subscriber: Philipp Oeser (lichtwerk).

Cannot reproduce here:

**System Information**
Operating system: Linux-5.8.18-300.fc33.x86_64-x86_64-with-fedora-33-Thirty_Three 64 Bits
Graphics card: GeForce GTX 970M/PCIe/SSE2 NVIDIA Corporation 4.5.0 NVIDIA 450.66
version: 2.92.0 Alpha, branch: master, commit date: 2020-11-18 22:29, hash: `rB9d6d5b9bebd1`

(but also checked with rB659f81f1cbc3)

Since this is lying around for so long now [with no apparent way to proceed], I would kindly ask to give it one last try:

Does this help by any chance?
Not even sure if blender uses anything from a local cuda install outside of compilation? Or if this comes with the driver? [I can see libcuda.so.450.66 in /usr/lib64 -- which looks like it comes with the driver?]

Steffen Dünner (SteffenD) added a comment.EditedNov 19 2020, 6:31 PM

Thanks!

I just gave it another try with the very latest master and the problem still looks the very same. I attached the system-info.txt from this version here.

EDIT: Same with 2.92 and 2.91 from the Buildbot and the release versions of 2.90 and 2.83.

Philipp Oeser (lichtwerk) changed the task status from Needs Information from User to Needs Information from Developers.Nov 19 2020, 7:03 PM

thx checking, afraid the cycles team needs to take over