Regression: LightTree failing with large number of emitters #106876

New Issue

Mark Bolstad · 2023-04-13T01:42:45+02:00

Mark Bolstad commented

2023-04-13 01:42:45 +02:00

System Information
Operating system: MacOS 13.3.1
Graphics card: AMD Radeon Vega 20

Blender Version
Broken: latest git version
Worked: 3.5.1
Most likely broken by 17c8e0d309

Short description of error
With a large number of emitters cycles crashes in LightManager::device_update_tree. From what I can see, the left and right index stacks are hard-coded to 32 ints, and the stack_id variable eventually reaches 35

Exact steps for others to reproduce the error
I'd love to send you the blend file, but the test case is 1.3GB. Let me know if you want it. But to reproduce:

Download the San Miguel test scene with materials (https://casual-effects.com/g3d/data10/research/model/San_Miguel/San_Miguel.zip)
Open the .obj and globally replace all of the material definitions to use the same material (in my scene, I globally replaced with "usemtl material_036". It doesn't matter what, you just need to have all the objects have the same material.
Import into a new blender scene
Add a default sun and camera, set renderer to cycles
Put a breakpoint in LightManager::device_update_tree, render (inline), and look at kintegrator->num_distribution. It should be 1.
Edit the unified material and set the emission color to white. Re-render. num_distribution will now be ~10million and the code with crash with a segfault and stack_id > 32.

*I'm writing a converter for cycles that takes MaterialX as input and exports OSL shaders. I'm doing something stupid that makes cycles think that every object is an emitter, and tripped this bug.

**System Information** Operating system: MacOS 13.3.1 Graphics card: AMD Radeon Vega 20 **Blender Version** Broken: latest git version Worked: 3.5.1 Most likely broken by 17c8e0d309 **Short description of error** With a large number of emitters cycles crashes in LightManager::device_update_tree. From what I can see, the left and right index stacks are hard-coded to 32 ints, and the stack_id variable eventually reaches 35 **Exact steps for others to reproduce the error** I'd love to send you the blend file, but the test case is 1.3GB. Let me know if you want it. But to reproduce: 1. Download the San Miguel test scene with materials (https://casual-effects.com/g3d/data10/research/model/San_Miguel/San_Miguel.zip) 2. Open the .obj and globally replace all of the material definitions to use the same material (in my scene, I globally replaced with "usemtl material_036". It doesn't matter what, you just need to have all the objects have the same material. 3. Import into a new blender scene 4. Add a default sun and camera, set renderer to cycles 5. Put a breakpoint in LightManager::device_update_tree, render (inline), and look at kintegrator->num_distribution. It should be 1. 6. Edit the unified material and set the emission color to white. Re-render. num_distribution will now be ~10million and the code with crash with a segfault and stack_id > 32. *I'm writing a converter for cycles that takes MaterialX as input and exports OSL shaders. I'm doing something stupid that makes cycles think that every object is an emitter, and tripped this bug.

Mark Bolstad added the

labels 2023-04-13 01:42:45 +02:00

Iliya Katushenock added the