Hi paulb

Just some questions and checks:

1. What resolution are you rendering in? (width*height*percentage)
2. Have you tried different tile sizes?
3. What NVidia driver version are you working on?

Best regards,
Jeroen

Hi jbakker

Thanks for your interest.

1. I'm rendering in 1080p (1920*1080*100%).
2. Yes, I tried every tile size available but they all didn't work on my machine.
3. The driver changed over the time of the tests (I have "GeForce Experience" installed so my driver is updated automatically) but at the moment I use the "GeForce R331 Game Ready Driver" Version 331.82.

One last thing: What's really strange is that when I recompose (shortcut: c) it sometimes works as expected and sometimes not (without changing anything in the scene or node-setup).

@Lukas Tönne (lukastoenne), maybe you've got clues here?

Might be related to T38011, at least this skipping of later chunks (when starting in the middle) seems to indicate that. I think we have a problem in the compo scheduler ...

@Paul B. (paulb) could you make a test file and attach it here? Neither @Sergey Sharybin (sergey) nor i could reproduce this bug so far.

I think I actually did upload the test file with the original description but here is it (again).

Thanks. I still can't reproduce the error with that file, but i did get a couple of OpenCL error messages (can't tie them to anything specific):

• "CUDA cuInit: Out of memory"
• "<N> childs per node: 0/0 = -nan" with N between 0 and 15

This looks to me as if the OpenCL device is running out of memory, but i have to investigate this more. Given that i also have a GTX 660 this is a bit strange.

Hi @Lukas Tönne (lukastoenne),
as far as I know there are different memmory configurations of the GTX 660 available.
You may have a different one than I have. I also mentioned that sometimes it works and sometimes not.
I don't know why this happens but it could be that you just had (bad-)luck an it worked.

I actually also encountered this bug. I have an AMD Radeon HD 6500M/5600/5700 Series . I suspect the defocus node is the cause, because it occurred when I used it. But it's not easily reproducible. I rendered an animation and most of the frames were fine, but several had this exact error.

I can also reproduce the error (Windows 7 64bit, NVIDIA GeForce GTX 760 2 GB, latest driver).

The culprit is here:
https://developer.blender.org/diffusion/B/browse/master/source/blender/compositor/operations/COM_VariableSizeBokehBlurOperation.cpp;7aa86ec243f92cee6f5de040f989fc7779f0e5a1$200

The sizeMemoryBuffer doesn't seem to be fully defined, it returns bogus values for the chunks in question, which leads to invalid results of the defocusKernel OpenCL code.

Have been trying all day to get my head around how the synchronization between OpenCL and CPU work packages is achieved, but this is really confusing code ...

@Jeroen Bakker (jbakker), found any time to look into this?

This bug is intriguing. It seems simple. But no one solved in 2 years

Edit: In the latest builds seems to have been SOLVED.

Edit2: No, now the image resolution must be 3840x2160 for the problem to can be seen.

Just for the clarity.

I would take effort to fix this, if I can reproduce it on one of my machines.