Strange rendering bug

Question

Strange rendering bug

straaljager opened this issue 5 years ago · comments

Hi there,

firstly, thanks a lot for releasing such an awesome framework! This might become a very useful addition to our visualization pipeline.

I made a few tests and found that some areas which should be darker (occluded by ambient or directional lights), show up very brightly. Also, some occluded areas converge very slowly (like the parts inside the mouth of the dragon and shadows on the floor in the attached screenshot).

(Btw, I tested with a Quadro M2000, CUDA 9.1 and Nvidia driver v390 on Ubuntu 18.04)

Any idea what might be causing this?

Cheers,
Sam

Tim Biedert · Answer 1 · Sat Apr 06 2019 20:48:23 GMT+0800 (China Standard Time)

Hey Sam,
Thanks a lot for this immensely valuable feedback!

Can you please provide some details on the materials/parameters you are using?
If you can easily post/send me a reproducer, that would be great as well.

Maybe you can give some other less reflective material a try to compare convergence and AO behavior.

How many samples/frames did you render for above screenshot?

If your scene contains multiple lights, there is also an option “Sample all lights”, which enables next-event estimation for every light in the scene. If that option is disabled, one light is randomly selected every frame, which increases rendering time at the expense of convergence speed, esp. for large light counts.

We’d be excited to learn more about the use cases and needs of the visualization pipeline you are working on! Feel free to drop me a message at tbiedert@nvidia.com. Thanks!

Sam Lapere · Answer 2 · Sat Apr 06 2019 22:26:13 GMT+0800 (China Standard Time)

Hi Tim,

thanks for the quick reply and tips. I will send you the scene and scene description when I'm back at my desk in the office.

Cheers,
Sam

Sam Lapere · Answer 3 · Tue Apr 09 2019 02:55:44 GMT+0800 (China Standard Time)

Hi Tim,

I could reproduce the strange rendering behaviour with the included SampleLightTest. I'm using the default parameters and materials that come with the scene. The shadow from the green spherelight is initially brighter than the surrounding, but when resizing the window it actually becomes dark. It also seems that the buffer is not cleared, resulting in some ghosting when moving the camera.
Enabling "Sample all lights" didn't make a difference. I tested with Optix 5.1 (Cuda 9.1, Nvidia driver 390) on a Quadro M2000).

Hope this helps,

Cheers

Tim Biedert · Answer 4 · Tue Apr 09 2019 17:56:58 GMT+0800 (China Standard Time)

Hey Sam,

this must be somehow related to the OptiX/driver/CUDA version. I'll try to reproduce it.

I highly recommend upgrading to OptiX 6, which also works on your Quadro M2000. Even on non-RTcore hardware, it should provide a small performance boost.

For reference, this is what it looks like on my machine with OptiX 6 / CUDA 10.1 / driver 419.17 on 2x Quadro RTX 8000.

Sam Lapere · Answer 5 · Tue Apr 09 2019 20:45:33 GMT+0800 (China Standard Time)

Thanks, I will give that a try.

Tim Biedert · Answer 6 · Tue Apr 09 2019 20:53:20 GMT+0800 (China Standard Time)

I was able to reproduce the issue with OptiX 5.1.1 (and CUDA 10.1 + driver 419.17).

It was a stack overflow, which I have fixed in the recent 0.1.2 release (7ac5d25).
For future reference, enable VISRTX_USE_DEBUG_EXCEPTIONS=ON in CMake in case of weird rendering artifacts.

Thanks a lot for catching this!

Anyway, I still recommend upgrading to OptiX 6 :-)

Sam Lapere · Answer 7 · Tue Apr 09 2019 21:07:29 GMT+0800 (China Standard Time)

Awesome! Thanks a lot. Optix 6 upgrade is on the todo list for sure.