- [🐛 bug report] Optix render

[IvanVishnevskiy]

Summary

I'm trying to render my scene with the following command: .\mitsuba.exe -m gpu_autodiff_spectral ..\Glass3.xml but it doesn't work

System configuration

For bug report, please enter information regarding your system configuration<- [remove this]

• Platform: Windows 10 2004
• Compiler: CMAKE 3.19
• Python version: 2.7.7
• Mitsuba 2 version: the lastest as of today
• Compiled variants:
  • scalar_rgb
  • scalar_spectral
  • gpu_autodiff_spectral
• GPU: GTX 1070
• CUDA: 10.2

Description

I have exported a simple scene from blender with a glass + some whiskey (60k tris) and a specular plane + hdri It works fine if I run it with scalar_spectral but when I try to do gpu rendering it tells me **Caught a critical exception: cuda_malloc(): out of memory!**

I have read all the issues and I did drop my sample count to 16 (it was set to 256) as well as resolution from 1080p to 1/3 of that and now the error is as such: **Caught a critical exception: cuda_trace_append(): arithmetic involving uninitialized variable!**

Here's a link to my final xml file: https://drive.google.com/file/d/1tWdCn5RJ_9wAh1KbRBcpv9mK55Sr_GpT/view?usp=sharing

I can also say that Cycles in the OPTIX mode works just fine as well as Octane/Luxcore, allowing me to render 2x the 1080p res with undefined amount of samples.

Steps to reproduce

1. Compile gpu_autodiff_spectral
2. run .\mitsuba.exe -m gpu_autodiff_spectral ..\Glass3.xml from the dist with my xml sitting one directory below

[Speierers]

Hi @IvanVishnevskiy ,

We are well aware that the current version of Mitsuba 2 on GPU quickly runs out of memory. This issue will be addressed in the upcoming major update.

In order to reproduce you issue, I will need access to the geometry files as well. Could you please package all the necessary files in a ZIP folder and share it here so I can take a look.

In the meantime, you could also try the following steps for debugging:

• reduce max_depth in the integrator. Try with 2, 4, 6, 8, 12, 16 to see if the error only occurs after a specific number of bounces
• switch every shape to diffuse to see if the issue comes from the materials
• switch ON/OFF the emitters one by one to see if the issue comes from there

[IvanVishnevskiy]

iv_meshes_textures.zip Oh, right, I forgot about these. Thank you for your answer, I will investigate your steps tomorrow

[vucenovic]

Hello,

I am running into the same problems, and am only able to render a resolution of 512px with a sample size of 10 per pixel. When is the new release update planned, that tackles this issue?

[Speierers]

Hi @vucenovic,

We haven't decided on a specific date yet, but we are making good progress. Hopefully in the coming month.

[vucenovic]

Thanks for the information, and good luck with the further implementation :)