Ok, I reduced the max. path depth to 64 and worked fine with GPU and Path and Sobol sampler.
I got curious, set it to METROPOLIS sampler and worked fine with this settings:
Total Path Depth = 64
Diffused = 64
Glossy = 64
Specular = 64
Would be nice for max settings not able to go past 64 in those option boxes as you said.
Crash with OpenCL
Forum rules
Please upload a testscene that allows developers to reproduce the problem, and attach some images.
Please upload a testscene that allows developers to reproduce the problem, and attach some images.
- infinityeternity
- Posts: 8
- Joined: Fri Feb 15, 2019 11:19 pm
Re: Crash with OpenCL
I also have a problem with OpenCL rendering (viewport preview rendering). It freezes. I suppose this is due to CL_DEVICE_MAX_MEM_ALLOC_SIZE=536870912. But why is it only ~500 MB while my card has 2GB dedicated GPU memory?
Last edited by infinityeternity on Sun Feb 17, 2019 11:31 am, edited 2 times in total.
- alpistinho
- Developer
- Posts: 198
- Joined: Thu Jul 05, 2018 11:38 pm
- Location: Rio de Janeiro
- infinityeternity
- Posts: 8
- Joined: Fri Feb 15, 2019 11:19 pm
Re: Crash with OpenCL
Is there a way to increase this value? It seems that CUDA is less limited in this sense.
Re: Crash with OpenCL
It was an hardware limit on some old AMD GPU too: the HD5870 wasn't able to allocate a single block of memory larger of 256MB even if it had 4GB of ram. It was happening because the memory access was handled like a texture map read (so there was limits like 16384x16384 max. texture map size).alpistinho wrote: ↑Sun Feb 17, 2019 11:27 am https://devtalk.nvidia.com/default/topi ... n-nvidia-/
I would say it is a nvidia limitation
Re: Crash with OpenCL
Yes, there is a LuxCore parameter to reduce the number of threads used (and so the amount of memory required for all GPUTask buffers). The name of the property is "opencl.task.count" and the value should be something like 65536. BlendLuxCore probably doesn't exposes the parameter so you should export the scene in text format, add the following line to .cfg file:infinityeternity wrote: ↑Sun Feb 17, 2019 11:50 am Is there a way to increase this value? It seems that CUDA is less limited in this sense.
Code: Select all
opencl.task.count = 65536
- infinityeternity
- Posts: 8
- Joined: Fri Feb 15, 2019 11:19 pm
Re: Crash with OpenCL
Yes, depend on the GPU but reducing the number of threads will slightly reduce the performance due to some additional CPU <=> GPU communication.