Search Issue Tracker

Postponed means that the issue was either a feature request or something that requires major refactoring on our side. Since that makes the issue not actionable in the close future we choose to close it as Postponed and add it on our internal roadmaps and technical debt pages instead.

Postponed

Votes

0

Found in

2020.2

2020.2.0b2

2021.1

Issue ID

1282179

Regression

No

[GPU PLM] Changing scene during a bake causes large memory spikes which cause GPULM to fail / fall-back

Progressive Lightmapper

-

Repro:
1. Sync the repro project at the revision listed above. Open and import the project
2. Open the Scene MultiBounce_WIP01
3. Ensure Progressive Updates are OFF and Auto Generate is OFF
4. Set the lightmapper to Progressive GPU and choose sample counts 1000, 1000, 1000
5. Set the lightmap resolution to 60 texels, which will necessitate using 7x1024 lightmap sets
6. Generate Lighting
7. While the bake is running, select the Light "DoorForeground_Area"
8. Adjust the colour of the light. This will cause the bake to restart.

Observe:
Using GPUZ to track, watch the memory usage during the bake. In the test Scene using the settings above, we're using around 5.5GB of video memory for Windows+Editor+Bake. When the Light is modified, dirtying the Scene, the bake will restart. For a brief moment, memory usage will spike to around 11GB. If this exceeds the available video memory, this will cause the lightmapper to fallback to CPU with the errors:

OpenCL Error. Falling back to CPU lightmapper. Error callback from context: CL_MEM_OBJECT_ALLOCATION_FAILURE error executing CL_COMMAND_NDRANGE_KERNEL on TITAN V (Device 0).

OpenCL Error. Falling back to CPU lightmapper. Error callback from context: CL_MEM_OBJECT_ALLOCATION_FAILURE

Expected:
There should not be any large spikes in video memory caused by restarting the bake.

  1. Resolution Note (2021.1.X):

    We will be revisiting this bug as part of an upcoming suite of improvements to the GPU lightmapper. If you are still encountering difficulties and can't find a suitable workaround, please feel free to reopen the case.

Add comment

Log in to post comment

All about bugs

View bugs we have successfully reproduced, and vote for the bugs you want to see fixed most urgently.