Speed ud OpenCL kernel loading time.. for better AMD cards usage

Hi Guys, we have some machine with Amd Radeon pro duo (polaris) 32gb, and some rx 580.
Working with cycle is pretty boring cause of that infinite kernel loading time…
Lux render and Amd Prorender take 2 second to compile kernel while often cycles require more than a minute…
This is really really boring… there is any way to speed up the pain? staring with --debug-value 256 and change kernel type can help?

Or maybe compile Blender in some different way?

Thanks in advance.

1 Like

afaik the kernel building should only happen the first time.

So no way? really? please :smiley:

We’d like to work on improve the performance here, it’s a matter of finding the time.

Well, to be honest cycles calculate kernel before preview, and before render, then if had a diffuse material, and you add a new pbr, it calculate again, when you add denoise it calculate again both preview and render…
There are a lot of things that activate the kernel recalc…

Right now, in general, AMD is selling lot of cards, our Radeon Pro Duo kick out our Titan x… and dual 580 go faster than 1080 but only when Cycle finish kernel… lol.

Pleaseeee i’m sure that a lot of people can tank you so much if you find time :smiley:
Don’t let other render engine come over cycles guys.
Thanks in advance.

  • this case… where amd pro duo with 16gb x2 memory and performance of 2 GTX 1080 for price of just 1000$ acts super wird for many users. some people have it performance fast some people super slow. some people don’t work at all

AMD supported Blender in 2017 - enabled us to hire a developer 1 year on OpenCL support for Cycles. This was meant to renew in 2018 but it was postponed. Last summer discussions with AMD started up again and (it took a while but) the outlook is very good.

The main topic that will be worked on is to solve kernel compile time. Official announcements soon.


Great news, Ton, let’s hope this will be fast, thanks.

Btw, in general Amd cards render faster by far, and preview window is much more fast using nightly build. Ofc the prebuild kernel time don’t change but they overall perform great once render start. So no 2.79b with Amd cards by my experience.

1 Like

“Jeroen Bakker starts work at the Blender Institute this week, welcome! He will first focus on Cycles OpenCL optimizations, funded by AMD.”

Welcome Jeroen!! great news guys !! cmon cmon let’s pump up the opencl !!