Cycles renders volumetrics very differently on CPU vs. GPU
Open, ConfirmedPublic

Description

I stumbled on this very inconsistency while working on some volumetrics where a notable difference in render results for CPU vs. GPU became visible. This effect appears for both Path Tracing and Branched Path Tracing for all types of lamps, and it gets more prominent on low sample rates. This difference is too strong to be acceptable when using a render farm with a mix of systems.

Following both images have been rendered with exactly the same settings on CPU top and GPU bellow for 1 SPP (so should look the same):


Here with 2 SPP:


Here is the blend file to test:

Tested with latest official build bot release on Win7 64.
Hardware: i7-980X, GTX 480

Details

Type
Bug
Kai Kostack (kaikostack) edited the task description. (Show Details)
Kai Kostack (kaikostack) edited the task description. (Show Details)
Kai Kostack (kaikostack) edited the task description. (Show Details)
Lukas Stockner (lukasstockner97) triaged this task as "Confirmed" priority.Dec 7 2016, 5:52 PM

Yes, I can reproduce the issue.
The difference is caused by decoupled sampling, which is currently enabled on CPUs and disabled on CUDA. So, there are two ways to fix this problem - either allow the user to disable it on the CPU, or enable in on CUDA.
I'm not sure why it's disabled on CUDA right now, though.

Decoupled Ray Marching for homogeneous volumes should work on GPU (when enabled via ifdefs of course), but when I tested that, it wasn't stable. For heterogenous volumes, the current code uses dynamic mem alloc, which is slow on GPU and was giving me mem errors. That was 1-2 years ago though, situation might have changed.

I can take a look and see if we can enable it.. We shouldn't add an option to disable that on CPU, that's exactly the kind of UI clutter that we should avoid.