Other settins are that was provided by OP bmw27_gpu.blend inside cycles_benchmark_20160228.zip archive
Changing sample size will result different output. For BMW the default is 35.
I adjusted tiles settings because in 2.78 the default tilse size is 256 x 256 for me what is not fine for the required resolution (non-square tiles remain), so I adjusted the tiles to fit.
I found the reason of the OpenCL error in viewport mode; it appears when I have more than about 60 of 4K textures.
wait what?.. This is just how cycles works? No. When rendering with Cuda GPU or CPU, the image in the tile is updated at a pretty constant rate. This is how Cycles normally works. What these guys are experiencing is no updating at all during render until it’s done. Right? This is not normal or expected behavior for Cycles. Maybe you are saying that Open CL works so differently from the other methods and that it prevents it from working as expected? Otherwise, something is broken and a bug report should be filed.
I think we should first wait for kernel split to be coded for tile updates, so it’s not bug.
This is how Cycles work in OpenCL mode yes. What I meant is that it is not a bug but an expected behavior. It’s just how it is coded and it’s not due to OpenCL itself as a language as it is possible to have the same behavior as in CUDA. So again, speak to the devs if you are not happy with it.
But seriously, why everytime someone comes to gives you improvement for free (and a 2x speedup for micro displacement is not going to happen every day) some of you feel the need to explain that person everything they are not happy with. Stop the noise! Their are definitely many ways to make Blender and cycles better, but here the goal of this thread is to gather benchmark results and bugs brought by this patch. So please stay on topic and show that you are happy to get an improvement.
Cycles network rendering behaves the same. Uninterruptable tiles that don’t update until the end.
This is why a lot of developers rarely come to blenderartists… nothing is good enough.
Good work getting a patch up on the developer portal bliblubli 
will be following its progress
How do i make build with this patch? Sorry for noob question. Tried google but didn’t found quick instructions.
Thanks for the support Doublebishop. Could you give some benchmark results with RX480 and 1080 with custom/master kernel to get an updated picture of the new generation performance? If I remember correctly, you have both and you also do your own optimized kernel, so it would be perfect for a comparison
Install arcanist https://wiki.blender.org/index.php/Dev:Doc/Tools/Code_Review
when you run “arc install-certificate”, it will give you a link to copy in your browser to get your key to get access to the blender server (you need an account on developer.blender.org)
then type arc patch D2254 and follow build instructions here : https://wiki.blender.org/index.php/Dev:Doc/Building_Blender/Linux/Ubuntu/CMake
Once you got a working build, you can also download the raw diff and apply it as a patch:
patch -b < rawfile
Did this get added to 2.78 Master? or do we have to wait
It’s still being reviewed. I may do a build when I have time.
Cheers, And if your up for a new challenge my friend maybe :RocknRoll: you could look ay getting opecl GPU to work with the blender cycles bake tool, Even better if micro displace could be included.
I know it’s a lot to ask, But as my Mother used to say “If you dont ask, You dont get”.
Either way great work, and ignore any of the bullshit aggro people have been aiming at devs recently, Some people just dont know their born.