Color management difference when using multiple and different GPUs together
Closed, ResolvedPublic

Description

System Information
Windows 10 64-bits
Intel Core i7-2600K
CUDA Rendering with GTX 560 Ti (sm_21) and GTX 780 (sm_35)

Blender Version
Broken: latest builds from Blender Buildbot, GraphicAll.org, and my own compiled builds (all of them 2.78.5)
Worked: official 2.78.1 release

Short description of error

I tried to use my older GTX 560 Ti with my GTX 780 to get faster renders, but I encountered this strange behavior when combining them: Picture here, Another example
I thought that the different CUDA kernels (sm_35 and sm_21) were the cause of this weird tiling effect. However, this does not happen when I use only one GPU at a time (and I get exactly the same results for both GPUs, no problem).

May have something to do with color management when using multiple GPUs, the problem disappears when I mute the Image texture node here:

Original post @Blender StackExchange

Maxime Michel (maximemichel) renamed this task from Hue difference when using multiple and different GPUs together to Color management difference when using multiple and different GPUs together.

@Stefan Werner (swerner), I'm guessing this is due to rBec25060a05e3: Unlimited number of textures for Cycles, mixing cases with bindless texture support and not. Not sure if there's a good way to make this work.

Brecht Van Lommel (brecht) triaged this task as Normal priority.Jul 18 2017, 2:40 PM

The attached scene is missing its textures. Can you please pack them and upload it again?

@Brecht Van Lommel (brecht) You're probably right. I think it needs to use the same texture IDs for all CUDA architectures and then map back to SM_2x on the kernel side. I'll go dig out my GTX 460...

Sorry I forgot to pack the textures indeed, should be okay now:

@Maxime Michel (maximemichel) if you could please verify that this is fixed with the upcoming automated builds.

I can't get the kernel to be compiled :

ptxas fatal   : More than 128 textures used in entry function 'kernel_cuda_branched_path_trace'

Full log here

I can't get the kernel to be compiled :

ptxas fatal   : More than 128 textures used in entry function 'kernel_cuda_branched_path_trace'

Same here. Win10 x64

Thanks, will take another look. Give me a couple of minutes.

Will test your new commit ASAP ;)

Should be working in c1ca3c8.

Seems to be working fine now! Thanks a lot ;)