I render on Linux most of the time, so I don’t remember the switch, but there are some registry keys to avoid that the driver think it crashed. Rendering on Linux is faster anyway. try to google it or ask here on the forums. I tested on windows with the gooseberry scene (victor) which is the heaviest of the benchmarks and it rendered correctly (driver 16.8.1)
Thanks. I wrote on the bug report it is an extreme file what I’m working on; it is heavier than anything I worked on the last 20 years and I count very huge archviz scenes to that:))))
Anyway, I hope I will find the problematic part sooner or later as it could be a problem for me to order additional AMD cards.
Also I hope the rendering process will be improved soon (remaining time, showing rendering process).
Maybe stupid question, but why rendered image/tile cannot be seen during rendering in OpenCL?
it was a generic question for anyone who could answer it; I simple do not understand why OpenCL rendering does not show the tile content during rendering.
Why opencl tile didn’t show content while rendering ?
Why we can’t also stop rendering while tile are already launch.
please can someone give an appropriate answer to this good question ?
Because the way it was coded, there are multiple samples being rendered at the same time, for more efficient rendering… so there is no current sample it is rendering as its currently rendering all of them. Only once the tile has finished does it send back the resultant image to blender
Thanks. As for OpenCL single tile is the most efficient method currently, does it mean that this function will never be implemented in a more handy way? Progressive rendering is also slower, but not seeing what is rendering is a huge NO GO for me with AMD cards in the ‘creation’ phase. For final renders it is ok.
in creation phase, are you able to use viewport rendering fine?
Just tested the viewport rendering with this build and it is 3x faster in the scene I’m working on than latest master 
Holla, some questions:
Is there a Linux build for these improvements, cause the first post is a windows version,right?
Can I use Cuda and OpenCL parallel on a Linux system?
Is Cuda faster on a Geforce card than OpenCL (with your improvements), what does it depend on?
That should do it.
you have checked under General settings –> Thread Display Options –> Visible Post Elements –> Show Signatures
and then, when you post something also check “Show your signature”
i’ve corrupted hdri display with tis build in the viewport ! will do deeper test !
For me it works, tested with many hdri’s from hdrlabs without problem.
No, it results “ghosting” images randomly about parts of the scene (once it completely freezed my PC)
Also the reason behind seeing tiles during the render is that I usually render in bigger resolution than the viewport has and it is easier to discover an artifact while I could see tiles.
Can’t reproduce the bug with own hdris, can you send me an example file with your specs?
It’s just how it is designed in Cycles. Luxcore shows each new sample pass and make it “cancel-able”. I think Lukas is working on a better time estimation for one-tile rendering. But as I already said, if you are unhappy with some designs, talk to the paid devs. I use viewport for preview and actually work in preview mode for fine tuning, so F12 for me is really just about rendering as fast as I can. I sometime render with 4 samples to check everything is right before rendering videos or very long single frames.
The Linux build will be possible with the patch I’ll submit this WE, I have no idea how to statically link under Linux, so if you know tell me and I’ll do one build. I need benchmark results guys.
@all I need benchmark results and bug reports (with reproducible steps and files), not wish list for new features.
2.78 needs it, please make it happen as i just bought RX 470 8GB… Ton please open emergency hatch for this code 
Potska, I’m really interested in the rendering performance of RX 470.
Would you please check my results, tile settings on 2.78 RC II ans when you will have time, run the same benchmarks?
Thank you very much.
Thats not how blender works. it needs to go through code review and make sure that its up to quality levels, so that people can maintain it in the future.
bliblubli: just tested your build simple way with BMW scene.
1.49.38 - 2.78 patch hash 16ed49b
2.05.43 - 2.78rc2 hash 9d70344
Samples i turned down to 15 from 35
autorun disabled from blender
image 960x540
In short, yes you patch is faster.
it would be nice to see with my settings the results, adjusting tile size and keeping samples. Thanks.